From ede8dfab3760afc8063c3418f217e52f7ec70d42 Mon Sep 17 00:00:00 2001 From: Michael Pratt Date: Thu, 30 Jan 2020 09:13:36 -0800 Subject: Enforce splice offset limits Splice must not allow negative offsets. Writes also must not allow offset + size to overflow int64. Reads are similarly broken, but not just in splice (b/148095030). Reported-by: syzbot+0e1ff0b95fb2859b4190@syzkaller.appspotmail.com PiperOrigin-RevId: 292361208 --- pkg/sentry/fs/tmpfs/inode_file.go | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/tmpfs/inode_file.go b/pkg/sentry/fs/tmpfs/inode_file.go index dabc10662..25abbc151 100644 --- a/pkg/sentry/fs/tmpfs/inode_file.go +++ b/pkg/sentry/fs/tmpfs/inode_file.go @@ -17,6 +17,7 @@ package tmpfs import ( "fmt" "io" + "math" "time" "gvisor.dev/gvisor/pkg/abi/linux" @@ -444,10 +445,15 @@ func (rw *fileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) defer rw.f.dataMu.Unlock() // Compute the range to write. - end := fs.WriteEndOffset(rw.offset, int64(srcs.NumBytes())) - if end == rw.offset { // srcs.NumBytes() == 0? + if srcs.NumBytes() == 0 { + // Nothing to do. return 0, nil } + end := fs.WriteEndOffset(rw.offset, int64(srcs.NumBytes())) + if end == math.MaxInt64 { + // Overflow. + return 0, syserror.EINVAL + } // Check if seals prevent either file growth or all writes. switch { -- cgit v1.2.3 From bc3a24d62788bd6b881afeda230f1a5550a5709a Mon Sep 17 00:00:00 2001 From: gVisor bot Date: Fri, 31 Jan 2020 11:49:37 -0800 Subject: Internal change. PiperOrigin-RevId: 292587459 --- kokoro/kythe/generate_xrefs.sh | 2 +- pkg/sentry/fs/g3doc/inotify.md | 16 ++++++++-------- pkg/sentry/mm/README.md | 8 ++++---- 3 files changed, 13 insertions(+), 13 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/kokoro/kythe/generate_xrefs.sh b/kokoro/kythe/generate_xrefs.sh index 7a0fbb3cd..323b0f77b 100644 --- a/kokoro/kythe/generate_xrefs.sh +++ b/kokoro/kythe/generate_xrefs.sh @@ -23,7 +23,7 @@ bazel version python3 -V -readonly KYTHE_VERSION='v0.0.39' +readonly KYTHE_VERSION='v0.0.41' readonly WORKDIR="$(mktemp -d)" readonly KYTHE_DIR="${WORKDIR}/kythe-${KYTHE_VERSION}" if [[ -n "$KOKORO_GIT_COMMIT" ]]; then diff --git a/pkg/sentry/fs/g3doc/inotify.md b/pkg/sentry/fs/g3doc/inotify.md index 71a577d9d..85063d4e6 100644 --- a/pkg/sentry/fs/g3doc/inotify.md +++ b/pkg/sentry/fs/g3doc/inotify.md @@ -112,11 +112,11 @@ attempts to queue a new event, it is already holding `fs.Watches.mu`. If we used `Inotify.mu` to also protect the event queue, this would violate the above lock ordering. -[dirent]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/fs/dirent.go -[event]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/fs/inotify_event.go -[fd_table]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/kernel/fd_table.go -[inode]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/fs/inode.go -[inode_watches]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/fs/inode_inotify.go -[inotify]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/fs/inotify.go -[syscall_dir]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/syscalls/linux/ -[watch]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/fs/inotify_watch.go +[dirent]: https://github.com/google/gvisor/blob/master/pkg/sentry/fs/dirent.go +[event]: https://github.com/google/gvisor/blob/master/pkg/sentry/fs/inotify_event.go +[fd_table]: https://github.com/google/gvisor/blob/master/pkg/sentry/kernel/fd_table.go +[inode]: https://github.com/google/gvisor/blob/master/pkg/sentry/fs/inode.go +[inode_watches]: https://github.com/google/gvisor/blob/master/pkg/sentry/fs/inode_inotify.go +[inotify]: https://github.com/google/gvisor/blob/master/pkg/sentry/fs/inotify.go +[syscall_dir]: https://github.com/google/gvisor/blob/master/pkg/sentry/syscalls/linux/ +[watch]: https://github.com/google/gvisor/blob/master/pkg/sentry/fs/inotify_watch.go diff --git a/pkg/sentry/mm/README.md b/pkg/sentry/mm/README.md index e1322e373..f4d43d927 100644 --- a/pkg/sentry/mm/README.md +++ b/pkg/sentry/mm/README.md @@ -274,7 +274,7 @@ In the sentry: methods [`platform.AddressSpace.MapFile` and `platform.AddressSpace.Unmap`][platform]. -[memmap]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/memmap/memmap.go -[mm]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/mm/mm.go -[pgalloc]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/pgalloc/pgalloc.go -[platform]: https://github.com/google/gvisor/blob/master/+/master/pkg/sentry/platform/platform.go +[memmap]: https://github.com/google/gvisor/blob/master/pkg/sentry/memmap/memmap.go +[mm]: https://github.com/google/gvisor/blob/master/pkg/sentry/mm/mm.go +[pgalloc]: https://github.com/google/gvisor/blob/master/pkg/sentry/pgalloc/pgalloc.go +[platform]: https://github.com/google/gvisor/blob/master/pkg/sentry/platform/platform.go -- cgit v1.2.3 From d7cd484091543827678f1548b8e5668a7a86e13f Mon Sep 17 00:00:00 2001 From: Fabricio Voznika Date: Tue, 4 Feb 2020 08:20:10 -0800 Subject: Add support for sentry internal pipe for gofer mounts Internal pipes are supported similarly to how internal UDS is done. It is also controlled by the same flag. Fixes #1102 PiperOrigin-RevId: 293150045 --- pkg/sentry/fs/gofer/BUILD | 2 + pkg/sentry/fs/gofer/cache_policy.go | 3 + pkg/sentry/fs/gofer/fifo.go | 40 ++++++++ pkg/sentry/fs/gofer/gofer_test.go | 2 +- pkg/sentry/fs/gofer/path.go | 183 +++++++++++++++++++++++------------ pkg/sentry/fs/gofer/session.go | 183 ++++++++++++++++++++++------------- pkg/sentry/fs/gofer/session_state.go | 12 +-- pkg/sentry/fs/gofer/socket.go | 8 +- test/syscalls/BUILD | 1 - 9 files changed, 293 insertions(+), 141 deletions(-) create mode 100644 pkg/sentry/fs/gofer/fifo.go (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/gofer/BUILD b/pkg/sentry/fs/gofer/BUILD index 971d3718e..fea135eea 100644 --- a/pkg/sentry/fs/gofer/BUILD +++ b/pkg/sentry/fs/gofer/BUILD @@ -9,6 +9,7 @@ go_library( "cache_policy.go", "context_file.go", "device.go", + "fifo.go", "file.go", "file_state.go", "fs.go", @@ -38,6 +39,7 @@ go_library( "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/host", "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/pipe", "//pkg/sentry/kernel/time", "//pkg/sentry/memmap", "//pkg/sentry/socket/unix/transport", diff --git a/pkg/sentry/fs/gofer/cache_policy.go b/pkg/sentry/fs/gofer/cache_policy.go index ebea03c42..07a564e92 100644 --- a/pkg/sentry/fs/gofer/cache_policy.go +++ b/pkg/sentry/fs/gofer/cache_policy.go @@ -127,6 +127,9 @@ func (cp cachePolicy) revalidate(ctx context.Context, name string, parent, child childIops, ok := child.InodeOperations.(*inodeOperations) if !ok { + if _, ok := child.InodeOperations.(*fifo); ok { + return false + } panic(fmt.Sprintf("revalidating inode operations of unknown type %T", child.InodeOperations)) } parentIops, ok := parent.InodeOperations.(*inodeOperations) diff --git a/pkg/sentry/fs/gofer/fifo.go b/pkg/sentry/fs/gofer/fifo.go new file mode 100644 index 000000000..456557058 --- /dev/null +++ b/pkg/sentry/fs/gofer/fifo.go @@ -0,0 +1,40 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" +) + +// +stateify savable +type fifo struct { + fs.InodeOperations + fileIops *inodeOperations +} + +var _ fs.InodeOperations = (*fifo)(nil) + +// Rename implements fs.InodeOperations. It forwards the call to the underlying +// file inode to handle the file rename. Note that file key remains the same +// after the rename to keep the endpoint mapping. +func (i *fifo) Rename(ctx context.Context, inode *fs.Inode, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string, replacement bool) error { + return i.fileIops.Rename(ctx, inode, oldParent, oldName, newParent, newName, replacement) +} + +// StatFS implements fs.InodeOperations. +func (i *fifo) StatFS(ctx context.Context) (fs.Info, error) { + return i.fileIops.StatFS(ctx) +} diff --git a/pkg/sentry/fs/gofer/gofer_test.go b/pkg/sentry/fs/gofer/gofer_test.go index 0c2f89ae8..2df2fe889 100644 --- a/pkg/sentry/fs/gofer/gofer_test.go +++ b/pkg/sentry/fs/gofer/gofer_test.go @@ -61,7 +61,7 @@ func rootTest(t *testing.T, name string, cp cachePolicy, fn func(context.Context ctx := contexttest.Context(t) sattr, rootInodeOperations := newInodeOperations(ctx, s, contextFile{ file: rootFile, - }, root.QID, p9.AttrMaskAll(), root.Attr, false /* socket */) + }, root.QID, p9.AttrMaskAll(), root.Attr) m := fs.NewMountSource(ctx, s, &filesystem{}, fs.MountSourceFlags{}) rootInode := fs.NewInode(ctx, rootInodeOperations, m, sattr) diff --git a/pkg/sentry/fs/gofer/path.go b/pkg/sentry/fs/gofer/path.go index 0c1be05ef..a35c3a23d 100644 --- a/pkg/sentry/fs/gofer/path.go +++ b/pkg/sentry/fs/gofer/path.go @@ -23,14 +23,24 @@ import ( "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // maxFilenameLen is the maximum length of a filename. This is dictated by 9P's // encoding of strings, which uses 2 bytes for the length prefix. const maxFilenameLen = (1 << 16) - 1 +func changeType(mode p9.FileMode, newType p9.FileMode) p9.FileMode { + if newType&^p9.FileModeMask != 0 { + panic(fmt.Sprintf("newType contained more bits than just file mode: %x", newType)) + } + clear := mode &^ p9.FileModeMask + return clear | newType +} + // Lookup loads an Inode at name into a Dirent based on the session's cache // policy. func (i *inodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string) (*fs.Dirent, error) { @@ -69,8 +79,25 @@ func (i *inodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string return nil, err } + if i.session().overrides != nil { + // Check if file belongs to a internal named pipe. Note that it doesn't need + // to check for sockets because it's done in newInodeOperations below. + deviceKey := device.MultiDeviceKey{ + Device: p9attr.RDev, + SecondaryDevice: i.session().connID, + Inode: qids[0].Path, + } + unlock := i.session().overrides.lock() + if pipeInode := i.session().overrides.getPipe(deviceKey); pipeInode != nil { + unlock() + pipeInode.IncRef() + return fs.NewDirent(ctx, pipeInode, name), nil + } + unlock() + } + // Construct the Inode operations. - sattr, node := newInodeOperations(ctx, i.fileState.s, newFile, qids[0], mask, p9attr, false) + sattr, node := newInodeOperations(ctx, i.fileState.s, newFile, qids[0], mask, p9attr) // Construct a positive Dirent. return fs.NewDirent(ctx, fs.NewInode(ctx, node, dir.MountSource, sattr), name), nil @@ -138,7 +165,7 @@ func (i *inodeOperations) Create(ctx context.Context, dir *fs.Inode, name string qid := qids[0] // Construct the InodeOperations. - sattr, iops := newInodeOperations(ctx, i.fileState.s, unopened, qid, mask, p9attr, false) + sattr, iops := newInodeOperations(ctx, i.fileState.s, unopened, qid, mask, p9attr) // Construct the positive Dirent. d := fs.NewDirent(ctx, fs.NewInode(ctx, iops, dir.MountSource, sattr), name) @@ -223,82 +250,115 @@ func (i *inodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string, return nil, syserror.ENAMETOOLONG } - if i.session().endpoints == nil { + if i.session().overrides == nil { return nil, syscall.EOPNOTSUPP } - // Create replaces the directory fid with the newly created/opened - // file, so clone this directory so it doesn't change out from under - // this node. - _, newFile, err := i.fileState.file.walk(ctx, nil) + // Stabilize the override map while creation is in progress. + unlock := i.session().overrides.lock() + defer unlock() + + sattr, iops, err := i.createEndpointFile(ctx, dir, name, perm, p9.ModeSocket) if err != nil { return nil, err } - // We're not going to use newFile after return. - defer newFile.close(ctx) - // Stabilize the endpoint map while creation is in progress. - unlock := i.session().endpoints.lock() - defer unlock() + // Construct the positive Dirent. + childDir := fs.NewDirent(ctx, fs.NewInode(ctx, iops, dir.MountSource, sattr), name) + i.session().overrides.addBoundEndpoint(iops.fileState.key, childDir, ep) + return childDir, nil +} - // Create a regular file in the gofer and then mark it as a socket by - // adding this inode key in the 'endpoints' map. - owner := fs.FileOwnerFromContext(ctx) - hostFile, err := newFile.create(ctx, name, p9.ReadWrite, p9.FileMode(perm.LinuxMode()), p9.UID(owner.UID), p9.GID(owner.GID)) - if err != nil { - return nil, err +// CreateFifo implements fs.InodeOperations.CreateFifo. +func (i *inodeOperations) CreateFifo(ctx context.Context, dir *fs.Inode, name string, perm fs.FilePermissions) error { + if len(name) > maxFilenameLen { + return syserror.ENAMETOOLONG } - // We're not going to use this file. - hostFile.Close() - i.touchModificationAndStatusChangeTime(ctx, dir) + owner := fs.FileOwnerFromContext(ctx) + mode := p9.FileMode(perm.LinuxMode()) | p9.ModeNamedPipe - // Get the attributes of the file to create inode key. - qid, mask, attr, err := getattr(ctx, newFile) - if err != nil { - return nil, err + // N.B. FIFOs use major/minor numbers 0. + if _, err := i.fileState.file.mknod(ctx, name, mode, 0, 0, p9.UID(owner.UID), p9.GID(owner.GID)); err != nil { + if i.session().overrides == nil || err != syscall.EPERM { + return err + } + // If gofer doesn't support mknod, check if we can create an internal fifo. + return i.createInternalFifo(ctx, dir, name, owner, perm) } - key := device.MultiDeviceKey{ - Device: attr.RDev, - SecondaryDevice: i.session().connID, - Inode: qid.Path, + i.touchModificationAndStatusChangeTime(ctx, dir) + return nil +} + +func (i *inodeOperations) createInternalFifo(ctx context.Context, dir *fs.Inode, name string, owner fs.FileOwner, perm fs.FilePermissions) error { + if i.session().overrides == nil { + return syserror.EPERM } - // Create child dirent. + // Stabilize the override map while creation is in progress. + unlock := i.session().overrides.lock() + defer unlock() - // Get an unopened p9.File for the file we created so that it can be - // cloned and re-opened multiple times after creation. - _, unopened, err := i.fileState.file.walk(ctx, []string{name}) + sattr, fileOps, err := i.createEndpointFile(ctx, dir, name, perm, p9.ModeNamedPipe) if err != nil { - return nil, err + return err } - // Construct the InodeOperations. - sattr, iops := newInodeOperations(ctx, i.fileState.s, unopened, qid, mask, attr, true) + // First create a pipe. + p := pipe.NewPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize) + + // Wrap the fileOps with our Fifo. + iops := &fifo{ + InodeOperations: pipe.NewInodeOperations(ctx, perm, p), + fileIops: fileOps, + } + inode := fs.NewInode(ctx, iops, dir.MountSource, sattr) // Construct the positive Dirent. childDir := fs.NewDirent(ctx, fs.NewInode(ctx, iops, dir.MountSource, sattr), name) - i.session().endpoints.add(key, childDir, ep) - return childDir, nil + i.session().overrides.addPipe(fileOps.fileState.key, childDir, inode) + return nil } -// CreateFifo implements fs.InodeOperations.CreateFifo. -func (i *inodeOperations) CreateFifo(ctx context.Context, dir *fs.Inode, name string, perm fs.FilePermissions) error { - if len(name) > maxFilenameLen { - return syserror.ENAMETOOLONG +// Caller must hold Session.endpoint lock. +func (i *inodeOperations) createEndpointFile(ctx context.Context, dir *fs.Inode, name string, perm fs.FilePermissions, fileType p9.FileMode) (fs.StableAttr, *inodeOperations, error) { + _, dirClone, err := i.fileState.file.walk(ctx, nil) + if err != nil { + return fs.StableAttr{}, nil, err } + // We're not going to use dirClone after return. + defer dirClone.close(ctx) + // Create a regular file in the gofer and then mark it as a socket by + // adding this inode key in the 'overrides' map. owner := fs.FileOwnerFromContext(ctx) - mode := p9.FileMode(perm.LinuxMode()) | p9.ModeNamedPipe - - // N.B. FIFOs use major/minor numbers 0. - if _, err := i.fileState.file.mknod(ctx, name, mode, 0, 0, p9.UID(owner.UID), p9.GID(owner.GID)); err != nil { - return err + hostFile, err := dirClone.create(ctx, name, p9.ReadWrite, p9.FileMode(perm.LinuxMode()), p9.UID(owner.UID), p9.GID(owner.GID)) + if err != nil { + return fs.StableAttr{}, nil, err } + // We're not going to use this file. + hostFile.Close() i.touchModificationAndStatusChangeTime(ctx, dir) - return nil + + // Get the attributes of the file to create inode key. + qid, mask, attr, err := getattr(ctx, dirClone) + if err != nil { + return fs.StableAttr{}, nil, err + } + + // Get an unopened p9.File for the file we created so that it can be + // cloned and re-opened multiple times after creation. + _, unopened, err := i.fileState.file.walk(ctx, []string{name}) + if err != nil { + return fs.StableAttr{}, nil, err + } + + // Construct new inode with file type overridden. + attr.Mode = changeType(attr.Mode, fileType) + sattr, iops := newInodeOperations(ctx, i.fileState.s, unopened, qid, mask, attr) + return sattr, iops, nil } // Remove implements InodeOperations.Remove. @@ -307,20 +367,23 @@ func (i *inodeOperations) Remove(ctx context.Context, dir *fs.Inode, name string return syserror.ENAMETOOLONG } - var key device.MultiDeviceKey - removeSocket := false - if i.session().endpoints != nil { - // Find out if file being deleted is a socket that needs to be + var key *device.MultiDeviceKey + if i.session().overrides != nil { + // Find out if file being deleted is a socket or pipe that needs to be // removed from endpoint map. if d, err := i.Lookup(ctx, dir, name); err == nil { defer d.DecRef() - if fs.IsSocket(d.Inode.StableAttr) { - child := d.Inode.InodeOperations.(*inodeOperations) - key = child.fileState.key - removeSocket = true - // Stabilize the endpoint map while deletion is in progress. - unlock := i.session().endpoints.lock() + if fs.IsSocket(d.Inode.StableAttr) || fs.IsPipe(d.Inode.StableAttr) { + switch iops := d.Inode.InodeOperations.(type) { + case *inodeOperations: + key = &iops.fileState.key + case *fifo: + key = &iops.fileIops.fileState.key + } + + // Stabilize the override map while deletion is in progress. + unlock := i.session().overrides.lock() defer unlock() } } @@ -329,8 +392,8 @@ func (i *inodeOperations) Remove(ctx context.Context, dir *fs.Inode, name string if err := i.fileState.file.unlinkAt(ctx, name, 0); err != nil { return err } - if removeSocket { - i.session().endpoints.remove(key) + if key != nil { + i.session().overrides.remove(*key) } i.touchModificationAndStatusChangeTime(ctx, dir) diff --git a/pkg/sentry/fs/gofer/session.go b/pkg/sentry/fs/gofer/session.go index 498c4645a..f6b3ef178 100644 --- a/pkg/sentry/fs/gofer/session.go +++ b/pkg/sentry/fs/gofer/session.go @@ -33,60 +33,107 @@ import ( var DefaultDirentCacheSize uint64 = fs.DefaultDirentCacheSize // +stateify savable -type endpointMaps struct { - // mu protexts the direntMap, the keyMap, and the pathMap below. - mu sync.RWMutex `state:"nosave"` +type overrideInfo struct { + dirent *fs.Dirent + + // endpoint is set when dirent points to a socket. inode must not be set. + endpoint transport.BoundEndpoint + + // inode is set when dirent points to a pipe. endpoint must not be set. + inode *fs.Inode +} - // direntMap links sockets to their dirents. - // It is filled concurrently with the keyMap and is stored upon save. - // Before saving, this map is used to populate the pathMap. - direntMap map[transport.BoundEndpoint]*fs.Dirent +func (l *overrideInfo) inodeType() fs.InodeType { + switch { + case l.endpoint != nil: + return fs.Socket + case l.inode != nil: + return fs.Pipe + } + panic("endpoint or node must be set") +} - // keyMap links MultiDeviceKeys (containing inode IDs) to their sockets. +// +stateify savable +type overrideMaps struct { + // mu protexts the keyMap, and the pathMap below. + mu sync.RWMutex `state:"nosave"` + + // keyMap links MultiDeviceKeys (containing inode IDs) to their sockets/pipes. // It is not stored during save because the inode ID may change upon restore. - keyMap map[device.MultiDeviceKey]transport.BoundEndpoint `state:"nosave"` + keyMap map[device.MultiDeviceKey]*overrideInfo `state:"nosave"` - // pathMap links the sockets to their paths. + // pathMap links the sockets/pipes to their paths. // It is filled before saving from the direntMap and is stored upon save. // Upon restore, this map is used to re-populate the keyMap. - pathMap map[transport.BoundEndpoint]string + pathMap map[*overrideInfo]string +} + +// addBoundEndpoint adds the bound endpoint to the map. +// A reference is taken on the dirent argument. +// +// Precondition: maps must have been locked with 'lock'. +func (e *overrideMaps) addBoundEndpoint(key device.MultiDeviceKey, d *fs.Dirent, ep transport.BoundEndpoint) { + d.IncRef() + e.keyMap[key] = &overrideInfo{dirent: d, endpoint: ep} } -// add adds the endpoint to the maps. +// addPipe adds the pipe inode to the map. // A reference is taken on the dirent argument. // // Precondition: maps must have been locked with 'lock'. -func (e *endpointMaps) add(key device.MultiDeviceKey, d *fs.Dirent, ep transport.BoundEndpoint) { - e.keyMap[key] = ep +func (e *overrideMaps) addPipe(key device.MultiDeviceKey, d *fs.Dirent, inode *fs.Inode) { d.IncRef() - e.direntMap[ep] = d + e.keyMap[key] = &overrideInfo{dirent: d, inode: inode} } // remove deletes the key from the maps. // // Precondition: maps must have been locked with 'lock'. -func (e *endpointMaps) remove(key device.MultiDeviceKey) { - endpoint := e.get(key) +func (e *overrideMaps) remove(key device.MultiDeviceKey) { + endpoint := e.keyMap[key] delete(e.keyMap, key) - - d := e.direntMap[endpoint] - d.DecRef() - delete(e.direntMap, endpoint) + endpoint.dirent.DecRef() } // lock blocks other addition and removal operations from happening while // the backing file is being created or deleted. Returns a function that unlocks // the endpoint map. -func (e *endpointMaps) lock() func() { +func (e *overrideMaps) lock() func() { e.mu.Lock() return func() { e.mu.Unlock() } } -// get returns the endpoint mapped to the given key. +// getBoundEndpoint returns the bound endpoint mapped to the given key. // -// Precondition: maps must have been locked for reading. -func (e *endpointMaps) get(key device.MultiDeviceKey) transport.BoundEndpoint { - return e.keyMap[key] +// Precondition: maps must have been locked. +func (e *overrideMaps) getBoundEndpoint(key device.MultiDeviceKey) transport.BoundEndpoint { + if v := e.keyMap[key]; v != nil { + return v.endpoint + } + return nil +} + +// getPipe returns the pipe inode mapped to the given key. +// +// Precondition: maps must have been locked. +func (e *overrideMaps) getPipe(key device.MultiDeviceKey) *fs.Inode { + if v := e.keyMap[key]; v != nil { + return v.inode + } + return nil +} + +// getType returns the inode type if there is a corresponding endpoint for the +// given key. Returns false otherwise. +func (e *overrideMaps) getType(key device.MultiDeviceKey) (fs.InodeType, bool) { + e.mu.Lock() + v := e.keyMap[key] + e.mu.Unlock() + + if v != nil { + return v.inodeType(), true + } + return 0, false } // session holds state for each 9p session established during sys_mount. @@ -137,16 +184,16 @@ type session struct { // mounter is the EUID/EGID that mounted this file system. mounter fs.FileOwner `state:"wait"` - // endpoints is used to map inodes that represent socket files to their - // corresponding endpoint. Socket files are created as regular files in the - // gofer and their presence in this map indicate that they should indeed be - // socket files. This allows unix domain sockets to be used with paths that - // belong to a gofer. + // overrides is used to map inodes that represent socket/pipes files to their + // corresponding endpoint/iops. These files are created as regular files in + // the gofer and their presence in this map indicate that they should indeed + // be socket/pipe files. This allows unix domain sockets and named pipes to + // be used with paths that belong to a gofer. // // TODO(gvisor.dev/issue/1200): there are few possible races with someone // stat'ing the file and another deleting it concurrently, where the file // will not be reported as socket file. - endpoints *endpointMaps `state:"wait"` + overrides *overrideMaps `state:"wait"` } // Destroy tears down the session. @@ -179,15 +226,21 @@ func (s *session) SaveInodeMapping(inode *fs.Inode, path string) { // This is very unintuitive. We *CANNOT* trust the inode's StableAttrs, // because overlay copyUp may have changed them out from under us. // So much for "immutable". - sattr := inode.InodeOperations.(*inodeOperations).fileState.sattr - s.inodeMappings[sattr.InodeID] = path + switch iops := inode.InodeOperations.(type) { + case *inodeOperations: + s.inodeMappings[iops.fileState.sattr.InodeID] = path + case *fifo: + s.inodeMappings[iops.fileIops.fileState.sattr.InodeID] = path + default: + panic(fmt.Sprintf("Invalid type: %T", iops)) + } } -// newInodeOperations creates a new 9p fs.InodeOperations backed by a p9.File and attributes -// (p9.QID, p9.AttrMask, p9.Attr). +// newInodeOperations creates a new 9p fs.InodeOperations backed by a p9.File +// and attributes (p9.QID, p9.AttrMask, p9.Attr). // // Endpoints lock must not be held if socket == false. -func newInodeOperations(ctx context.Context, s *session, file contextFile, qid p9.QID, valid p9.AttrMask, attr p9.Attr, socket bool) (fs.StableAttr, *inodeOperations) { +func newInodeOperations(ctx context.Context, s *session, file contextFile, qid p9.QID, valid p9.AttrMask, attr p9.Attr) (fs.StableAttr, *inodeOperations) { deviceKey := device.MultiDeviceKey{ Device: attr.RDev, SecondaryDevice: s.connID, @@ -201,17 +254,11 @@ func newInodeOperations(ctx context.Context, s *session, file contextFile, qid p BlockSize: bsize(attr), } - if s.endpoints != nil { - if socket { - sattr.Type = fs.Socket - } else { - // If unix sockets are allowed on this filesystem, check if this file is - // supposed to be a socket file. - unlock := s.endpoints.lock() - if s.endpoints.get(deviceKey) != nil { - sattr.Type = fs.Socket - } - unlock() + if s.overrides != nil && sattr.Type == fs.RegularFile { + // If overrides are allowed on this filesystem, check if this file is + // supposed to be of a different type, e.g. socket. + if t, ok := s.overrides.getType(deviceKey); ok { + sattr.Type = t } } @@ -267,7 +314,7 @@ func Root(ctx context.Context, dev string, filesystem fs.Filesystem, superBlockF s.EnableLeakCheck("gofer.session") if o.privateunixsocket { - s.endpoints = newEndpointMaps() + s.overrides = newOverrideMaps() } // Construct the MountSource with the session and superBlockFlags. @@ -305,26 +352,24 @@ func Root(ctx context.Context, dev string, filesystem fs.Filesystem, superBlockF return nil, err } - sattr, iops := newInodeOperations(ctx, &s, s.attach, qid, valid, attr, false) + sattr, iops := newInodeOperations(ctx, &s, s.attach, qid, valid, attr) return fs.NewInode(ctx, iops, m, sattr), nil } -// newEndpointMaps creates a new endpointMaps. -func newEndpointMaps() *endpointMaps { - return &endpointMaps{ - direntMap: make(map[transport.BoundEndpoint]*fs.Dirent), - keyMap: make(map[device.MultiDeviceKey]transport.BoundEndpoint), - pathMap: make(map[transport.BoundEndpoint]string), +// newOverrideMaps creates a new overrideMaps. +func newOverrideMaps() *overrideMaps { + return &overrideMaps{ + keyMap: make(map[device.MultiDeviceKey]*overrideInfo), + pathMap: make(map[*overrideInfo]string), } } -// fillKeyMap populates key and dirent maps upon restore from saved -// pathmap. +// fillKeyMap populates key and dirent maps upon restore from saved pathmap. func (s *session) fillKeyMap(ctx context.Context) error { - unlock := s.endpoints.lock() + unlock := s.overrides.lock() defer unlock() - for ep, dirPath := range s.endpoints.pathMap { + for ep, dirPath := range s.overrides.pathMap { _, file, err := s.attach.walk(ctx, splitAbsolutePath(dirPath)) if err != nil { return fmt.Errorf("error filling endpointmaps, failed to walk to %q: %v", dirPath, err) @@ -341,25 +386,25 @@ func (s *session) fillKeyMap(ctx context.Context) error { Inode: qid.Path, } - s.endpoints.keyMap[key] = ep + s.overrides.keyMap[key] = ep } return nil } -// fillPathMap populates paths for endpoints from dirents in direntMap +// fillPathMap populates paths for overrides from dirents in direntMap // before save. func (s *session) fillPathMap() error { - unlock := s.endpoints.lock() + unlock := s.overrides.lock() defer unlock() - for ep, dir := range s.endpoints.direntMap { - mountRoot := dir.MountRoot() + for _, endpoint := range s.overrides.keyMap { + mountRoot := endpoint.dirent.MountRoot() defer mountRoot.DecRef() - dirPath, _ := dir.FullName(mountRoot) + dirPath, _ := endpoint.dirent.FullName(mountRoot) if dirPath == "" { return fmt.Errorf("error getting path from dirent") } - s.endpoints.pathMap[ep] = dirPath + s.overrides.pathMap[endpoint] = dirPath } return nil } @@ -368,7 +413,7 @@ func (s *session) fillPathMap() error { func (s *session) restoreEndpointMaps(ctx context.Context) error { // When restoring, only need to create the keyMap because the dirent and path // maps got stored through the save. - s.endpoints.keyMap = make(map[device.MultiDeviceKey]transport.BoundEndpoint) + s.overrides.keyMap = make(map[device.MultiDeviceKey]*overrideInfo) if err := s.fillKeyMap(ctx); err != nil { return fmt.Errorf("failed to insert sockets into endpoint map: %v", err) } @@ -376,6 +421,6 @@ func (s *session) restoreEndpointMaps(ctx context.Context) error { // Re-create pathMap because it can no longer be trusted as socket paths can // change while process continues to run. Empty pathMap will be re-filled upon // next save. - s.endpoints.pathMap = make(map[transport.BoundEndpoint]string) + s.overrides.pathMap = make(map[*overrideInfo]string) return nil } diff --git a/pkg/sentry/fs/gofer/session_state.go b/pkg/sentry/fs/gofer/session_state.go index 0285c5361..111da59f9 100644 --- a/pkg/sentry/fs/gofer/session_state.go +++ b/pkg/sentry/fs/gofer/session_state.go @@ -25,9 +25,9 @@ import ( // beforeSave is invoked by stateify. func (s *session) beforeSave() { - if s.endpoints != nil { + if s.overrides != nil { if err := s.fillPathMap(); err != nil { - panic("failed to save paths to endpoint map before saving" + err.Error()) + panic("failed to save paths to override map before saving" + err.Error()) } } } @@ -74,10 +74,10 @@ func (s *session) afterLoad() { panic(fmt.Sprintf("new attach name %v, want %v", opts.aname, s.aname)) } - // Check if endpointMaps exist when uds sockets are enabled - // (only pathmap will actualy have been saved). - if opts.privateunixsocket != (s.endpoints != nil) { - panic(fmt.Sprintf("new privateunixsocket option %v, want %v", opts.privateunixsocket, s.endpoints != nil)) + // Check if overrideMaps exist when uds sockets are enabled (only pathmaps + // will actually have been saved). + if opts.privateunixsocket != (s.overrides != nil) { + panic(fmt.Sprintf("new privateunixsocket option %v, want %v", opts.privateunixsocket, s.overrides != nil)) } if args.Flags != s.superBlockFlags { panic(fmt.Sprintf("new mount flags %v, want %v", args.Flags, s.superBlockFlags)) diff --git a/pkg/sentry/fs/gofer/socket.go b/pkg/sentry/fs/gofer/socket.go index 376cfce2c..10ba2f5f0 100644 --- a/pkg/sentry/fs/gofer/socket.go +++ b/pkg/sentry/fs/gofer/socket.go @@ -32,15 +32,15 @@ func (i *inodeOperations) BoundEndpoint(inode *fs.Inode, path string) transport. return nil } - if i.session().endpoints != nil { - unlock := i.session().endpoints.lock() + if i.session().overrides != nil { + unlock := i.session().overrides.lock() defer unlock() - ep := i.session().endpoints.get(i.fileState.key) + ep := i.session().overrides.getBoundEndpoint(i.fileState.key) if ep != nil { return ep } - // Not found in endpoints map, it may be a gofer backed unix socket... + // Not found in overrides map, it may be a gofer backed unix socket... } inode.IncRef() diff --git a/test/syscalls/BUILD b/test/syscalls/BUILD index 40e974314..8f2b75a1c 100644 --- a/test/syscalls/BUILD +++ b/test/syscalls/BUILD @@ -225,7 +225,6 @@ syscall_test( syscall_test( add_overlay = True, test = "//test/syscalls/linux:mknod_test", - use_tmpfs = True, # mknod is not supported over gofer. ) syscall_test( -- cgit v1.2.3 From 492229d0176c1af2ab4ea4cf91bf211e940b5b12 Mon Sep 17 00:00:00 2001 From: Jamie Liu Date: Tue, 4 Feb 2020 11:28:36 -0800 Subject: VFS2 gofer client Updates #1198 Opening host pipes (by spinning in fdpipe) and host sockets is not yet complete, and will be done in a future CL. Major differences from VFS1 gofer client (sentry/fs/gofer), with varying levels of backportability: - "Cache policies" are replaced by InteropMode, which control the behavior of timestamps in addition to caching. Under InteropModeExclusive (analogous to cacheAll) and InteropModeWritethrough (analogous to cacheAllWritethrough), client timestamps are *not* written back to the server (it is not possible in 9P or Linux for clients to set ctime, so writing back client-authoritative timestamps results in incoherence between atime/mtime and ctime). Under InteropModeShared (analogous to cacheRemoteRevalidating), client timestamps are not used at all (remote filesystem clocks are authoritative). cacheNone is translated to InteropModeShared + new option filesystemOptions.specialRegularFiles. - Under InteropModeShared, "unstable attribute" reloading for permission checks, lookup, and revalidation are fused, which is feasible in VFS2 since gofer.filesystem controls path resolution. This results in a ~33% reduction in RPCs for filesystem operations compared to cacheRemoteRevalidating. For example, consider stat("/foo/bar/baz") where "/foo/bar/baz" fails revalidation, resulting in the instantiation of a new dentry: VFS1 RPCs: getattr("/") // fs.MountNamespace.FindLink() => fs.Inode.CheckPermission() => gofer.inodeOperations.check() => gofer.inodeOperations.UnstableAttr() walkgetattr("/", "foo") = fid1 // fs.Dirent.walk() => gofer.session.Revalidate() => gofer.cachePolicy.Revalidate() clunk(fid1) getattr("/foo") // CheckPermission walkgetattr("/foo", "bar") = fid2 // Revalidate clunk(fid2) getattr("/foo/bar") // CheckPermission walkgetattr("/foo/bar", "baz") = fid3 // Revalidate clunk(fid3) walkgetattr("/foo/bar", "baz") = fid4 // fs.Dirent.walk() => gofer.inodeOperations.Lookup getattr("/foo/bar/baz") // linux.stat() => gofer.inodeOperations.UnstableAttr() VFS2 RPCs: getattr("/") // gofer.filesystem.walkExistingLocked() walkgetattr("/", "foo") = fid1 // gofer.filesystem.stepExistingLocked() clunk(fid1) // No getattr: walkgetattr already updated metadata for permission check walkgetattr("/foo", "bar") = fid2 clunk(fid2) walkgetattr("/foo/bar", "baz") = fid3 // No clunk: fid3 used for new gofer.dentry // No getattr: walkgetattr already updated metadata for stat() - gofer.filesystem.unlinkAt() does not require instantiation of a dentry that represents the file to be deleted. Updates #898. - gofer.regularFileFD.OnClose() skips Tflushf for regular files under InteropModeExclusive, as it's nonsensical to request a remote file flush without flushing locally-buffered writes to that remote file first. - Symlink targets are cached when InteropModeShared is not in effect. - p9.QID.Path (which is already required to be unique for each file within a server, and is accordingly already synthesized from device/inode numbers in all known gofers) is used as-is for inode numbers, rather than being mapped along with attr.RDev in the client to yet another synthetic inode number. - Relevant parts of fsutil.CachingInodeOperations are inlined directly into gofer package code. This avoids having to duplicate part of its functionality in fsutil.HostMappable. PiperOrigin-RevId: 293190213 --- pkg/safemem/seq_unsafe.go | 17 + pkg/sentry/fs/fsutil/BUILD | 4 +- pkg/sentry/fs/fsutil/frame_ref_set.go | 13 +- pkg/sentry/fs/fsutil/inode_cached.go | 2 +- pkg/sentry/fsimpl/gofer/BUILD | 55 ++ pkg/sentry/fsimpl/gofer/directory.go | 190 +++++ pkg/sentry/fsimpl/gofer/filesystem.go | 1087 ++++++++++++++++++++++++++++ pkg/sentry/fsimpl/gofer/gofer.go | 1147 ++++++++++++++++++++++++++++++ pkg/sentry/fsimpl/gofer/handle.go | 135 ++++ pkg/sentry/fsimpl/gofer/handle_unsafe.go | 66 ++ pkg/sentry/fsimpl/gofer/p9file.go | 219 ++++++ pkg/sentry/fsimpl/gofer/pagemath.go | 31 + pkg/sentry/fsimpl/gofer/regular_file.go | 860 ++++++++++++++++++++++ pkg/sentry/fsimpl/gofer/special_file.go | 159 +++++ pkg/sentry/fsimpl/gofer/symlink.go | 47 ++ pkg/sentry/fsimpl/gofer/time.go | 75 ++ pkg/sentry/fsimpl/tmpfs/filesystem.go | 2 +- pkg/sentry/socket/hostinet/socket.go | 23 +- 18 files changed, 4103 insertions(+), 29 deletions(-) create mode 100644 pkg/sentry/fsimpl/gofer/BUILD create mode 100644 pkg/sentry/fsimpl/gofer/directory.go create mode 100644 pkg/sentry/fsimpl/gofer/filesystem.go create mode 100644 pkg/sentry/fsimpl/gofer/gofer.go create mode 100644 pkg/sentry/fsimpl/gofer/handle.go create mode 100644 pkg/sentry/fsimpl/gofer/handle_unsafe.go create mode 100644 pkg/sentry/fsimpl/gofer/p9file.go create mode 100644 pkg/sentry/fsimpl/gofer/pagemath.go create mode 100644 pkg/sentry/fsimpl/gofer/regular_file.go create mode 100644 pkg/sentry/fsimpl/gofer/special_file.go create mode 100644 pkg/sentry/fsimpl/gofer/symlink.go create mode 100644 pkg/sentry/fsimpl/gofer/time.go (limited to 'pkg/sentry/fs') diff --git a/pkg/safemem/seq_unsafe.go b/pkg/safemem/seq_unsafe.go index 354a95dde..dcdfc9600 100644 --- a/pkg/safemem/seq_unsafe.go +++ b/pkg/safemem/seq_unsafe.go @@ -18,6 +18,7 @@ import ( "bytes" "fmt" "reflect" + "syscall" "unsafe" ) @@ -297,3 +298,19 @@ func ZeroSeq(dsts BlockSeq) (uint64, error) { } return done, nil } + +// IovecsFromBlockSeq returns a []syscall.Iovec representing seq. +func IovecsFromBlockSeq(bs BlockSeq) []syscall.Iovec { + iovs := make([]syscall.Iovec, 0, bs.NumBlocks()) + for ; !bs.IsEmpty(); bs = bs.Tail() { + b := bs.Head() + iovs = append(iovs, syscall.Iovec{ + Base: &b.ToSlice()[0], + Len: uint64(b.Len()), + }) + // We don't need to care about b.NeedSafecopy(), because the host + // kernel will handle such address ranges just fine (by returning + // EFAULT). + } + return iovs +} diff --git a/pkg/sentry/fs/fsutil/BUILD b/pkg/sentry/fs/fsutil/BUILD index 4ab2a384f..789369220 100644 --- a/pkg/sentry/fs/fsutil/BUILD +++ b/pkg/sentry/fs/fsutil/BUILD @@ -28,13 +28,13 @@ go_template_instance( "platform": "gvisor.dev/gvisor/pkg/sentry/platform", }, package = "fsutil", - prefix = "frameRef", + prefix = "FrameRef", template = "//pkg/segment:generic_set", types = { "Key": "uint64", "Range": "platform.FileRange", "Value": "uint64", - "Functions": "frameRefSetFunctions", + "Functions": "FrameRefSetFunctions", }, ) diff --git a/pkg/sentry/fs/fsutil/frame_ref_set.go b/pkg/sentry/fs/fsutil/frame_ref_set.go index dd63db32b..6564fd0c6 100644 --- a/pkg/sentry/fs/fsutil/frame_ref_set.go +++ b/pkg/sentry/fs/fsutil/frame_ref_set.go @@ -20,24 +20,25 @@ import ( "gvisor.dev/gvisor/pkg/sentry/platform" ) -type frameRefSetFunctions struct{} +// FrameRefSetFunctions implements segment.Functions for FrameRefSet. +type FrameRefSetFunctions struct{} // MinKey implements segment.Functions.MinKey. -func (frameRefSetFunctions) MinKey() uint64 { +func (FrameRefSetFunctions) MinKey() uint64 { return 0 } // MaxKey implements segment.Functions.MaxKey. -func (frameRefSetFunctions) MaxKey() uint64 { +func (FrameRefSetFunctions) MaxKey() uint64 { return math.MaxUint64 } // ClearValue implements segment.Functions.ClearValue. -func (frameRefSetFunctions) ClearValue(val *uint64) { +func (FrameRefSetFunctions) ClearValue(val *uint64) { } // Merge implements segment.Functions.Merge. -func (frameRefSetFunctions) Merge(_ platform.FileRange, val1 uint64, _ platform.FileRange, val2 uint64) (uint64, bool) { +func (FrameRefSetFunctions) Merge(_ platform.FileRange, val1 uint64, _ platform.FileRange, val2 uint64) (uint64, bool) { if val1 != val2 { return 0, false } @@ -45,6 +46,6 @@ func (frameRefSetFunctions) Merge(_ platform.FileRange, val1 uint64, _ platform. } // Split implements segment.Functions.Split. -func (frameRefSetFunctions) Split(_ platform.FileRange, val uint64, _ uint64) (uint64, uint64) { +func (FrameRefSetFunctions) Split(_ platform.FileRange, val uint64, _ uint64) (uint64, uint64) { return val, val } diff --git a/pkg/sentry/fs/fsutil/inode_cached.go b/pkg/sentry/fs/fsutil/inode_cached.go index 573b8586e..800c8b4e1 100644 --- a/pkg/sentry/fs/fsutil/inode_cached.go +++ b/pkg/sentry/fs/fsutil/inode_cached.go @@ -111,7 +111,7 @@ type CachingInodeOperations struct { // refs tracks active references to data in the cache. // // refs is protected by dataMu. - refs frameRefSet + refs FrameRefSet } // CachingInodeOperationsOptions configures a CachingInodeOperations. diff --git a/pkg/sentry/fsimpl/gofer/BUILD b/pkg/sentry/fsimpl/gofer/BUILD new file mode 100644 index 000000000..4ba76a1e8 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/BUILD @@ -0,0 +1,55 @@ +load("//tools:defs.bzl", "go_library") +load("//tools/go_generics:defs.bzl", "go_template_instance") + +licenses(["notice"]) + +go_template_instance( + name = "dentry_list", + out = "dentry_list.go", + package = "gofer", + prefix = "dentry", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*dentry", + "Linker": "*dentry", + }, +) + +go_library( + name = "gofer", + srcs = [ + "dentry_list.go", + "directory.go", + "filesystem.go", + "gofer.go", + "handle.go", + "handle_unsafe.go", + "p9file.go", + "pagemath.go", + "regular_file.go", + "special_file.go", + "symlink.go", + "time.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fd", + "//pkg/fspath", + "//pkg/log", + "//pkg/p9", + "//pkg/safemem", + "//pkg/sentry/fs/fsutil", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/time", + "//pkg/sentry/memmap", + "//pkg/sentry/pgalloc", + "//pkg/sentry/platform", + "//pkg/sentry/usage", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/unet", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go new file mode 100644 index 000000000..baa2cdd8e --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/directory.go @@ -0,0 +1,190 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +func (d *dentry) isDir() bool { + return d.fileType() == linux.S_IFDIR +} + +// Preconditions: d.dirMu must be locked. d.isDir(). fs.opts.interop != +// InteropModeShared. +func (d *dentry) cacheNegativeChildLocked(name string) { + if d.negativeChildren == nil { + d.negativeChildren = make(map[string]struct{}) + } + d.negativeChildren[name] = struct{}{} +} + +type directoryFD struct { + fileDescription + vfs.DirectoryFileDescriptionDefaultImpl + + mu sync.Mutex + off int64 + dirents []vfs.Dirent +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *directoryFD) Release() { +} + +// IterDirents implements vfs.FileDescriptionImpl.IterDirents. +func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { + fd.mu.Lock() + defer fd.mu.Unlock() + + if fd.dirents == nil { + ds, err := fd.dentry().getDirents(ctx) + if err != nil { + return err + } + fd.dirents = ds + } + + for fd.off < int64(len(fd.dirents)) { + if !cb.Handle(fd.dirents[fd.off]) { + return nil + } + fd.off++ + } + return nil +} + +// Preconditions: d.isDir(). There exists at least one directoryFD representing d. +func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { + // 9P2000.L's readdir does not specify behavior in the presence of + // concurrent mutation of an iterated directory, so implementations may + // duplicate or omit entries in this case, which violates POSIX semantics. + // Thus we read all directory entries while holding d.dirMu to exclude + // directory mutations. (Note that it is impossible for the client to + // exclude concurrent mutation from other remote filesystem users. Since + // there is no way to detect if the server has incorrectly omitted + // directory entries, we simply assume that the server is well-behaved + // under InteropModeShared.) This is inconsistent with Linux (which appears + // to assume that directory fids have the correct semantics, and translates + // struct file_operations::readdir calls directly to readdir RPCs), but is + // consistent with VFS1. + + d.fs.renameMu.RLock() + defer d.fs.renameMu.RUnlock() + d.dirMu.Lock() + defer d.dirMu.Unlock() + if d.dirents != nil { + return d.dirents, nil + } + + // It's not clear if 9P2000.L's readdir is expected to return "." and "..", + // so we generate them here. + parent := d.vfsd.ParentOrSelf().Impl().(*dentry) + dirents := []vfs.Dirent{ + { + Name: ".", + Type: linux.DT_DIR, + Ino: d.ino, + NextOff: 1, + }, + { + Name: "..", + Type: uint8(atomic.LoadUint32(&parent.mode) >> 12), + Ino: parent.ino, + NextOff: 2, + }, + } + off := uint64(0) + const count = 64 * 1024 // for consistency with the vfs1 client + d.handleMu.RLock() + defer d.handleMu.RUnlock() + if !d.handleReadable { + // This should not be possible because a readable handle should have + // been opened when the calling directoryFD was opened. + panic("gofer.dentry.getDirents called without a readable handle") + } + for { + p9ds, err := d.handle.file.readdir(ctx, off, count) + if err != nil { + return nil, err + } + if len(p9ds) == 0 { + // Cache dirents for future directoryFDs if permitted. + if d.fs.opts.interop != InteropModeShared { + d.dirents = dirents + } + return dirents, nil + } + for _, p9d := range p9ds { + if p9d.Name == "." || p9d.Name == ".." { + continue + } + dirent := vfs.Dirent{ + Name: p9d.Name, + Ino: p9d.QID.Path, + NextOff: int64(len(dirents) + 1), + } + // p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or + // DMSOCKET. + switch p9d.Type { + case p9.TypeSymlink: + dirent.Type = linux.DT_LNK + case p9.TypeDir: + dirent.Type = linux.DT_DIR + default: + dirent.Type = linux.DT_REG + } + dirents = append(dirents, dirent) + } + off = p9ds[len(p9ds)-1].Offset + } +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fd.mu.Lock() + defer fd.mu.Unlock() + + switch whence { + case linux.SEEK_SET: + if offset < 0 { + return 0, syserror.EINVAL + } + if offset == 0 { + // Ensure that the next call to fd.IterDirents() calls + // fd.dentry().getDirents(). + fd.dirents = nil + } + fd.off = offset + return fd.off, nil + case linux.SEEK_CUR: + offset += fd.off + if offset < 0 { + return 0, syserror.EINVAL + } + // Don't clear fd.dirents in this case, even if offset == 0. + fd.off = offset + return fd.off, nil + default: + return 0, syserror.EINVAL + } +} diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go new file mode 100644 index 000000000..8eb61debf --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -0,0 +1,1087 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Sync implements vfs.FilesystemImpl.Sync. +func (fs *filesystem) Sync(ctx context.Context) error { + // Snapshot current dentries and special files. + fs.syncMu.Lock() + ds := make([]*dentry, 0, len(fs.dentries)) + for d := range fs.dentries { + ds = append(ds, d) + } + sffds := make([]*specialFileFD, 0, len(fs.specialFileFDs)) + for sffd := range fs.specialFileFDs { + sffds = append(sffds, sffd) + } + fs.syncMu.Unlock() + + // Return the first error we encounter, but sync everything we can + // regardless. + var retErr error + + // Sync regular files. + for _, d := range ds { + if !d.TryIncRef() { + continue + } + err := d.syncSharedHandle(ctx) + d.DecRef() + if err != nil && retErr == nil { + retErr = err + } + } + + // Sync special files, which may be writable but do not use dentry shared + // handles (so they won't be synced by the above). + for _, sffd := range sffds { + if !sffd.vfsfd.TryIncRef() { + continue + } + err := sffd.Sync(ctx) + sffd.vfsfd.DecRef() + if err != nil && retErr == nil { + retErr = err + } + } + + return retErr +} + +// maxFilenameLen is the maximum length of a filename. This is dictated by 9P's +// encoding of strings, which uses 2 bytes for the length prefix. +const maxFilenameLen = (1 << 16) - 1 + +// dentrySlicePool is a pool of *[]*dentry used to store dentries for which +// dentry.checkCachingLocked() must be called. The pool holds pointers to +// slices because Go lacks generics, so sync.Pool operates on interface{}, so +// every call to (what should be) sync.Pool<[]*dentry>.Put() allocates a copy +// of the slice header on the heap. +var dentrySlicePool = sync.Pool{ + New: func() interface{} { + ds := make([]*dentry, 0, 4) // arbitrary non-zero initial capacity + return &ds + }, +} + +func appendDentry(ds *[]*dentry, d *dentry) *[]*dentry { + if ds == nil { + ds = dentrySlicePool.Get().(*[]*dentry) + } + *ds = append(*ds, d) + return ds +} + +// Preconditions: ds != nil. +func putDentrySlice(ds *[]*dentry) { + // Allow dentries to be GC'd. + for i := range *ds { + (*ds)[i] = nil + } + *ds = (*ds)[:0] + dentrySlicePool.Put(ds) +} + +// stepLocked resolves rp.Component() to an existing file, starting from the +// given directory. +// +// Dentries which may become cached as a result of the traversal are appended +// to *ds. +// +// Preconditions: fs.renameMu must be locked. d.dirMu must be locked. +// !rp.Done(). If fs.opts.interop == InteropModeShared, then d's cached +// metadata must be up to date. +func (fs *filesystem) stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { + if !d.isDir() { + return nil, syserror.ENOTDIR + } + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } +afterSymlink: + name := rp.Component() + if name == "." { + rp.Advance() + return d, nil + } + if name == ".." { + parentVFSD, err := rp.ResolveParent(&d.vfsd) + if err != nil { + return nil, err + } + parent := parentVFSD.Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // We must assume that parentVFSD is correct, because if d has been + // moved elsewhere in the remote filesystem so that its parent has + // changed, we have no way of determining its new parent's location + // in the filesystem. Get updated metadata for parentVFSD. + _, attrMask, attr, err := parent.file.getAttr(ctx, dentryAttrMask()) + if err != nil { + return nil, err + } + parent.updateFromP9Attrs(attrMask, &attr) + } + rp.Advance() + return parent, nil + } + childVFSD, err := rp.ResolveChild(&d.vfsd, name) + if err != nil { + return nil, err + } + // FIXME(jamieliu): Linux performs revalidation before mount lookup + // (fs/namei.c:lookup_fast() => __d_lookup_rcu(), d_revalidate(), + // __follow_mount_rcu()). + child, err := fs.revalidateChildLocked(ctx, rp.VirtualFilesystem(), d, name, childVFSD, ds) + if err != nil { + return nil, err + } + if child == nil { + return nil, syserror.ENOENT + } + if child.isSymlink() && rp.ShouldFollowSymlink() { + target, err := child.readlink(ctx, rp.Mount()) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + goto afterSymlink // don't check the current directory again + } + rp.Advance() + return child, nil +} + +// revalidateChildLocked must be called after a call to parent.vfsd.Child(name) +// or vfs.ResolvingPath.ResolveChild(name) returns childVFSD (which may be +// nil) to verify that the returned child (or lack thereof) is correct. If no file +// exists at name, revalidateChildLocked returns (nil, nil). +// +// Preconditions: fs.renameMu must be locked. parent.dirMu must be locked. +// parent.isDir(). name is not "." or "..". +// +// Postconditions: If revalidateChildLocked returns a non-nil dentry, its +// cached metadata is up to date. +func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, childVFSD *vfs.Dentry, ds **[]*dentry) (*dentry, error) { + if childVFSD != nil && fs.opts.interop != InteropModeShared { + // We have a cached dentry that is assumed to be correct. + return childVFSD.Impl().(*dentry), nil + } + // We either don't have a cached dentry or need to verify that it's still + // correct, either of which requires a remote lookup. Check if this name is + // valid before performing the lookup. + if len(name) > maxFilenameLen { + return nil, syserror.ENAMETOOLONG + } + // Check if we've already cached this lookup with a negative result. + if _, ok := parent.negativeChildren[name]; ok { + return nil, nil + } + // Perform the remote lookup. + qid, file, attrMask, attr, err := parent.file.walkGetAttrOne(ctx, name) + if err != nil && err != syserror.ENOENT { + return nil, err + } + if childVFSD != nil { + child := childVFSD.Impl().(*dentry) + if !file.isNil() && qid.Path == child.ino { + // The file at this path hasn't changed. Just update cached + // metadata. + file.close(ctx) + child.updateFromP9Attrs(attrMask, &attr) + return child, nil + } + // The file at this path has changed or no longer exists. Remove + // the stale dentry from the tree, and re-evaluate its caching + // status (i.e. if it has 0 references, drop it). + vfsObj.ForceDeleteDentry(childVFSD) + *ds = appendDentry(*ds, child) + childVFSD = nil + } + if file.isNil() { + // No file exists at this path now. Cache the negative lookup if + // allowed. + if fs.opts.interop != InteropModeShared { + parent.cacheNegativeChildLocked(name) + } + return nil, nil + } + // Create a new dentry representing the file. + child, err := fs.newDentry(ctx, file, qid, attrMask, &attr) + if err != nil { + file.close(ctx) + return nil, err + } + parent.IncRef() // reference held by child on its parent + parent.vfsd.InsertChild(&child.vfsd, name) + // For now, child has 0 references, so our caller should call + // child.checkCachingLocked(). + *ds = appendDentry(*ds, child) + return child, nil +} + +// walkParentDirLocked resolves all but the last path component of rp to an +// existing directory, starting from the given directory (which is usually +// rp.Start().Impl().(*dentry)). It does not check that the returned directory +// is searchable by the provider of rp. +// +// Preconditions: fs.renameMu must be locked. !rp.Done(). If fs.opts.interop == +// InteropModeShared, then d's cached metadata must be up to date. +func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { + for !rp.Final() { + d.dirMu.Lock() + next, err := fs.stepLocked(ctx, rp, d, ds) + d.dirMu.Unlock() + if err != nil { + return nil, err + } + d = next + } + if !d.isDir() { + return nil, syserror.ENOTDIR + } + return d, nil +} + +// resolveLocked resolves rp to an existing file. +// +// Preconditions: fs.renameMu must be locked. +func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, ds **[]*dentry) (*dentry, error) { + d := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for rp.Start() as required by fs.stepLocked(). + if err := d.updateFromGetattr(ctx); err != nil { + return nil, err + } + } + for !rp.Done() { + d.dirMu.Lock() + next, err := fs.stepLocked(ctx, rp, d, ds) + d.dirMu.Unlock() + if err != nil { + return nil, err + } + d = next + } + if rp.MustBeDir() && !d.isDir() { + return nil, syserror.ENOTDIR + } + return d, nil +} + +// doCreateAt checks that creating a file at rp is permitted, then invokes +// create to do so. +// +// Preconditions: !rp.Done(). For the final path component in rp, +// !rp.ShouldFollowSymlink(). +func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + start := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for start as required by + // fs.walkParentDirLocked(). + if err := start.updateFromGetattr(ctx); err != nil { + return err + } + } + parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return err + } + if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + return err + } + if parent.isDeleted() { + return syserror.ENOENT + } + name := rp.Component() + if name == "." || name == ".." { + return syserror.EEXIST + } + if len(name) > maxFilenameLen { + return syserror.ENAMETOOLONG + } + if !dir && rp.MustBeDir() { + return syserror.ENOENT + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + if fs.opts.interop == InteropModeShared { + // The existence of a dentry at name would be inconclusive because the + // file it represents may have been deleted from the remote filesystem, + // so we would need to make an RPC to revalidate the dentry. Just + // attempt the file creation RPC instead. If a file does exist, the RPC + // will fail with EEXIST like we would have. If the RPC succeeds, and a + // stale dentry exists, the dentry will fail revalidation next time + // it's used. + return create(parent, name) + } + if parent.vfsd.Child(name) != nil { + return syserror.EEXIST + } + // No cached dentry exists; however, there might still be an existing file + // at name. As above, we attempt the file creation RPC anyway. + if err := create(parent, name); err != nil { + return err + } + parent.touchCMtime(ctx) + delete(parent.negativeChildren, name) + parent.dirents = nil + return nil +} + +// Preconditions: !rp.Done(). +func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + start := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for start as required by + // fs.walkParentDirLocked(). + if err := start.updateFromGetattr(ctx); err != nil { + return err + } + } + parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return err + } + if err := parent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + return err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return err + } + defer rp.Mount().EndWrite() + + name := rp.Component() + if dir { + if name == "." { + return syserror.EINVAL + } + if name == ".." { + return syserror.ENOTEMPTY + } + } else { + if name == "." || name == ".." { + return syserror.EISDIR + } + } + vfsObj := rp.VirtualFilesystem() + mntns := vfs.MountNamespaceFromContext(ctx) + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + childVFSD := parent.vfsd.Child(name) + var child *dentry + // We only need a dentry representing the file at name if it can be a mount + // point. If childVFSD is nil, then it can't be a mount point. If childVFSD + // is non-nil but stale, the actual file can't be a mount point either; we + // detect this case by just speculatively calling PrepareDeleteDentry and + // only revalidating the dentry if that fails (indicating that the existing + // dentry is a mount point). + if childVFSD != nil { + child = childVFSD.Impl().(*dentry) + if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { + child, err = fs.revalidateChildLocked(ctx, vfsObj, parent, name, childVFSD, &ds) + if err != nil { + return err + } + if child != nil { + childVFSD = &child.vfsd + if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { + return err + } + } else { + childVFSD = nil + } + } + } else if _, ok := parent.negativeChildren[name]; ok { + return syserror.ENOENT + } + flags := uint32(0) + if dir { + if child != nil && !child.isDir() { + return syserror.ENOTDIR + } + flags = linux.AT_REMOVEDIR + } else { + if child != nil && child.isDir() { + return syserror.EISDIR + } + if rp.MustBeDir() { + return syserror.ENOTDIR + } + } + err = parent.file.unlinkAt(ctx, name, flags) + if err != nil { + if childVFSD != nil { + vfsObj.AbortDeleteDentry(childVFSD) + } + return err + } + if fs.opts.interop != InteropModeShared { + parent.touchCMtime(ctx) + parent.cacheNegativeChildLocked(name) + parent.dirents = nil + } + if child != nil { + child.setDeleted() + vfsObj.CommitDeleteDentry(childVFSD) + ds = appendDentry(ds, child) + } + return nil +} + +// renameMuRUnlockAndCheckCaching calls fs.renameMu.RUnlock(), then calls +// dentry.checkCachingLocked on all dentries in *ds with fs.renameMu locked for +// writing. +// +// ds is a pointer-to-pointer since defer evaluates its arguments immediately, +// but dentry slices are allocated lazily, and it's much easier to say "defer +// fs.renameMuRUnlockAndCheckCaching(&ds)" than "defer func() { +// fs.renameMuRUnlockAndCheckCaching(ds) }()" to work around this. +func (fs *filesystem) renameMuRUnlockAndCheckCaching(ds **[]*dentry) { + fs.renameMu.RUnlock() + if *ds == nil { + return + } + if len(**ds) != 0 { + fs.renameMu.Lock() + for _, d := range **ds { + d.checkCachingLocked() + } + fs.renameMu.Unlock() + } + putDentrySlice(*ds) +} + +func (fs *filesystem) renameMuUnlockAndCheckCaching(ds **[]*dentry) { + if *ds == nil { + fs.renameMu.Unlock() + return + } + for _, d := range **ds { + d.checkCachingLocked() + } + fs.renameMu.Unlock() + putDentrySlice(*ds) +} + +// GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. +func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return nil, err + } + if opts.CheckSearchable { + if !d.isDir() { + return nil, syserror.ENOTDIR + } + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } + } + d.IncRef() + return &d.vfsd, nil +} + +// GetParentDentryAt implements vfs.FilesystemImpl.GetParentDentryAt. +func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + start := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for start as required by + // fs.walkParentDirLocked(). + if err := start.updateFromGetattr(ctx); err != nil { + return nil, err + } + } + d, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return nil, err + } + d.IncRef() + return &d.vfsd, nil +} + +// LinkAt implements vfs.FilesystemImpl.LinkAt. +func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, childName string) error { + if rp.Mount() != vd.Mount() { + return syserror.EXDEV + } + // 9P2000.L supports hard links, but we don't. + return syserror.EPERM + }) +} + +// MkdirAt implements vfs.FilesystemImpl.MkdirAt. +func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { + return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string) error { + creds := rp.Credentials() + _, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + return err + }) +} + +// MknodAt implements vfs.FilesystemImpl.MknodAt. +func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error { + creds := rp.Credentials() + _, err := parent.file.mknod(ctx, name, (p9.FileMode)(opts.Mode), opts.DevMajor, opts.DevMinor, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + return err + }) +} + +// OpenAt implements vfs.FilesystemImpl.OpenAt. +func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + // Reject O_TMPFILE, which is not supported; supporting it correctly in the + // presence of other remote filesystem users requires remote filesystem + // support, and it isn't clear that there's any way to implement this in + // 9P. + if opts.Flags&linux.O_TMPFILE != 0 { + return nil, syserror.EOPNOTSUPP + } + mayCreate := opts.Flags&linux.O_CREAT != 0 + mustCreate := opts.Flags&(linux.O_CREAT|linux.O_EXCL) == (linux.O_CREAT | linux.O_EXCL) + + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + + start := rp.Start().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + // Get updated metadata for start as required by fs.stepLocked(). + if err := start.updateFromGetattr(ctx); err != nil { + return nil, err + } + } + if rp.Done() { + return start.openLocked(ctx, rp, opts.Flags) + } + +afterTrailingSymlink: + parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return nil, err + } + // Check for search permission in the parent directory. + if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } + // Determine whether or not we need to create a file. + parent.dirMu.Lock() + child, err := fs.stepLocked(ctx, rp, parent, &ds) + if err == syserror.ENOENT && mayCreate { + fd, err := parent.createAndOpenChildLocked(ctx, rp, &opts) + parent.dirMu.Unlock() + return fd, err + } + if err != nil { + parent.dirMu.Unlock() + return nil, err + } + // Open existing child or follow symlink. + parent.dirMu.Unlock() + if mustCreate { + return nil, syserror.EEXIST + } + if child.isSymlink() && rp.ShouldFollowSymlink() { + target, err := child.readlink(ctx, rp.Mount()) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + start = parent + goto afterTrailingSymlink + } + return child.openLocked(ctx, rp, opts.Flags) +} + +// Preconditions: fs.renameMu must be locked. +func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, flags uint32) (*vfs.FileDescription, error) { + ats := vfs.AccessTypesForOpenFlags(flags) + if err := d.checkPermissions(rp.Credentials(), ats, d.isDir()); err != nil { + return nil, err + } + mnt := rp.Mount() + filetype := d.fileType() + switch { + case filetype == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD: + if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, flags&linux.O_TRUNC != 0); err != nil { + return nil, err + } + fd := ®ularFileFD{} + if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{ + AllowDirectIO: true, + }); err != nil { + return nil, err + } + return &fd.vfsfd, nil + case filetype == linux.S_IFDIR: + // Can't open directories with O_CREAT. + if flags&linux.O_CREAT != 0 { + return nil, syserror.EISDIR + } + // Can't open directories writably. + if ats&vfs.MayWrite != 0 { + return nil, syserror.EISDIR + } + if flags&linux.O_DIRECT != 0 { + return nil, syserror.EINVAL + } + if err := d.ensureSharedHandle(ctx, ats&vfs.MayRead != 0, false /* write */, false /* trunc */); err != nil { + return nil, err + } + fd := &directoryFD{} + if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return &fd.vfsfd, nil + case filetype == linux.S_IFLNK: + // Can't open symlinks without O_PATH (which is unimplemented). + return nil, syserror.ELOOP + default: + if flags&linux.O_DIRECT != 0 { + return nil, syserror.EINVAL + } + h, err := openHandle(ctx, d.file, ats&vfs.MayRead != 0, ats&vfs.MayWrite != 0, flags&linux.O_TRUNC != 0) + if err != nil { + return nil, err + } + fd := &specialFileFD{ + handle: h, + } + if err := fd.vfsfd.Init(fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + h.close(ctx) + return nil, err + } + return &fd.vfsfd, nil + } +} + +// Preconditions: d.fs.renameMu must be locked. d.dirMu must be locked. +func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { + if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + return nil, err + } + if d.isDeleted() { + return nil, syserror.ENOENT + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return nil, err + } + defer mnt.EndWrite() + + // 9P2000.L's lcreate takes a fid representing the parent directory, and + // converts it into an open fid representing the created file, so we need + // to duplicate the directory fid first. + _, dirfile, err := d.file.walk(ctx, nil) + if err != nil { + return nil, err + } + creds := rp.Credentials() + name := rp.Component() + fdobj, openFile, createQID, _, err := dirfile.create(ctx, name, (p9.OpenFlags)(opts.Flags), (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + if err != nil { + dirfile.close(ctx) + return nil, err + } + // Then we need to walk to the file we just created to get a non-open fid + // representing it, and to get its metadata. This must use d.file since, as + // explained above, dirfile was invalidated by dirfile.Create(). + walkQID, nonOpenFile, attrMask, attr, err := d.file.walkGetAttrOne(ctx, name) + if err != nil { + openFile.close(ctx) + if fdobj != nil { + fdobj.Close() + } + return nil, err + } + // Sanity-check that we walked to the file we created. + if createQID.Path != walkQID.Path { + // Probably due to concurrent remote filesystem mutation? + ctx.Warningf("gofer.dentry.createAndOpenChildLocked: created file has QID %v before walk, QID %v after (interop=%v)", createQID, walkQID, d.fs.opts.interop) + nonOpenFile.close(ctx) + openFile.close(ctx) + if fdobj != nil { + fdobj.Close() + } + return nil, syserror.EAGAIN + } + + // Construct the new dentry. + child, err := d.fs.newDentry(ctx, nonOpenFile, createQID, attrMask, &attr) + if err != nil { + nonOpenFile.close(ctx) + openFile.close(ctx) + if fdobj != nil { + fdobj.Close() + } + return nil, err + } + // Incorporate the fid that was opened by lcreate. + useRegularFileFD := child.fileType() == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD + if useRegularFileFD { + child.handleMu.Lock() + child.handle.file = openFile + if fdobj != nil { + child.handle.fd = int32(fdobj.Release()) + } + child.handleReadable = vfs.MayReadFileWithOpenFlags(opts.Flags) + child.handleWritable = vfs.MayWriteFileWithOpenFlags(opts.Flags) + child.handleMu.Unlock() + } + // Take a reference on the new dentry to be held by the new file + // description. (This reference also means that the new dentry is not + // eligible for caching yet, so we don't need to append to a dentry slice.) + child.refs = 1 + // Insert the dentry into the tree. + d.IncRef() // reference held by child on its parent d + d.vfsd.InsertChild(&child.vfsd, name) + if d.fs.opts.interop != InteropModeShared { + d.touchCMtime(ctx) + delete(d.negativeChildren, name) + d.dirents = nil + } + + // Finally, construct a file description representing the created file. + var childVFSFD *vfs.FileDescription + mnt.IncRef() + if useRegularFileFD { + fd := ®ularFileFD{} + if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{ + AllowDirectIO: true, + }); err != nil { + return nil, err + } + childVFSFD = &fd.vfsfd + } else { + fd := &specialFileFD{ + handle: handle{ + file: openFile, + fd: -1, + }, + } + if fdobj != nil { + fd.handle.fd = int32(fdobj.Release()) + } + if err := fd.vfsfd.Init(fd, opts.Flags, mnt, &child.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + fd.handle.close(ctx) + return nil, err + } + childVFSFD = &fd.vfsfd + } + return childVFSFD, nil +} + +// ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. +func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return "", err + } + if !d.isSymlink() { + return "", syserror.EINVAL + } + return d.readlink(ctx, rp.Mount()) +} + +// RenameAt implements vfs.FilesystemImpl.RenameAt. +func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldParentVD vfs.VirtualDentry, oldName string, opts vfs.RenameOptions) error { + if opts.Flags != 0 { + // Requires 9P support. + return syserror.EINVAL + } + + var ds *[]*dentry + fs.renameMu.Lock() + defer fs.renameMuUnlockAndCheckCaching(&ds) + newParent, err := fs.walkParentDirLocked(ctx, rp, rp.Start().Impl().(*dentry), &ds) + if err != nil { + return err + } + newName := rp.Component() + if newName == "." || newName == ".." { + return syserror.EBUSY + } + mnt := rp.Mount() + if mnt != oldParentVD.Mount() { + return syserror.EXDEV + } + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + + oldParent := oldParentVD.Dentry().Impl().(*dentry) + if fs.opts.interop == InteropModeShared { + if err := oldParent.updateFromGetattr(ctx); err != nil { + return err + } + } + if err := oldParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + return err + } + vfsObj := rp.VirtualFilesystem() + // We need a dentry representing the renamed file since, if it's a + // directory, we need to check for write permission on it. + oldParent.dirMu.Lock() + defer oldParent.dirMu.Unlock() + renamed, err := fs.revalidateChildLocked(ctx, vfsObj, oldParent, oldName, oldParent.vfsd.Child(oldName), &ds) + if err != nil { + return err + } + if renamed == nil { + return syserror.ENOENT + } + if renamed.isDir() { + if renamed == newParent || renamed.vfsd.IsAncestorOf(&newParent.vfsd) { + return syserror.EINVAL + } + if oldParent != newParent { + if err := renamed.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + return err + } + } + } else { + if opts.MustBeDir || rp.MustBeDir() { + return syserror.ENOTDIR + } + } + + if oldParent != newParent { + if err := newParent.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true); err != nil { + return err + } + newParent.dirMu.Lock() + defer newParent.dirMu.Unlock() + } + if newParent.isDeleted() { + return syserror.ENOENT + } + replacedVFSD := newParent.vfsd.Child(newName) + var replaced *dentry + // This is similar to unlinkAt, except: + // + // - We revalidate the replaced dentry unconditionally for simplicity. + // + // - If rp.MustBeDir(), then we need a dentry representing the replaced + // file regardless to confirm that it's a directory. + if replacedVFSD != nil || rp.MustBeDir() { + replaced, err = fs.revalidateChildLocked(ctx, vfsObj, newParent, newName, replacedVFSD, &ds) + if err != nil { + return err + } + if replaced != nil { + if replaced.isDir() { + if !renamed.isDir() { + return syserror.EISDIR + } + } else { + if rp.MustBeDir() || renamed.isDir() { + return syserror.ENOTDIR + } + } + replacedVFSD = &replaced.vfsd + } else { + replacedVFSD = nil + } + } + + if oldParent == newParent && oldName == newName { + return nil + } + if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), &renamed.vfsd, replacedVFSD); err != nil { + return err + } + if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + if fs.opts.interop != InteropModeShared { + oldParent.cacheNegativeChildLocked(oldName) + oldParent.dirents = nil + delete(newParent.negativeChildren, newName) + newParent.dirents = nil + } + vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, &newParent.vfsd, newName, replacedVFSD) + return nil +} + +// RmdirAt implements vfs.FilesystemImpl.RmdirAt. +func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { + return fs.unlinkAt(ctx, rp, true /* dir */) +} + +// SetStatAt implements vfs.FilesystemImpl.SetStatAt. +func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return err + } + return d.setStat(ctx, rp.Credentials(), &opts.Stat, rp.Mount()) +} + +// StatAt implements vfs.FilesystemImpl.StatAt. +func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return linux.Statx{}, err + } + // Since walking updates metadata for all traversed dentries under + // InteropModeShared, including the returned one, we can return cached + // metadata here regardless of fs.opts.interop. + var stat linux.Statx + d.statTo(&stat) + return stat, nil +} + +// StatFSAt implements vfs.FilesystemImpl.StatFSAt. +func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return linux.Statfs{}, err + } + fsstat, err := d.file.statFS(ctx) + if err != nil { + return linux.Statfs{}, err + } + nameLen := uint64(fsstat.NameLength) + if nameLen > maxFilenameLen { + nameLen = maxFilenameLen + } + return linux.Statfs{ + // This is primarily for distinguishing a gofer file system in + // tests. Testing is important, so instead of defining + // something completely random, use a standard value. + Type: linux.V9FS_MAGIC, + BlockSize: int64(fsstat.BlockSize), + Blocks: fsstat.Blocks, + BlocksFree: fsstat.BlocksFree, + BlocksAvailable: fsstat.BlocksAvailable, + Files: fsstat.Files, + FilesFree: fsstat.FilesFree, + NameLength: nameLen, + }, nil +} + +// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. +func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error { + creds := rp.Credentials() + _, err := parent.file.symlink(ctx, target, name, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + return err + }) +} + +// UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. +func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { + return fs.unlinkAt(ctx, rp, false /* dir */) +} + +// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return nil, err + } + return d.listxattr(ctx) +} + +// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return "", err + } + return d.getxattr(ctx, name) +} + +// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. +func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return err + } + return d.setxattr(ctx, &opts) +} + +// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. +func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckCaching(&ds) + d, err := fs.resolveLocked(ctx, rp, &ds) + if err != nil { + return err + } + return d.removexattr(ctx, name) +} + +// PrependPath implements vfs.FilesystemImpl.PrependPath. +func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error { + fs.renameMu.RLock() + defer fs.renameMu.RUnlock() + return vfs.GenericPrependPath(vfsroot, vd, b) +} diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go new file mode 100644 index 000000000..d0552bd99 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -0,0 +1,1147 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package gofer provides a filesystem implementation that is backed by a 9p +// server, interchangably referred to as "gofers" throughout this package. +// +// Lock order: +// regularFileFD/directoryFD.mu +// filesystem.renameMu +// dentry.dirMu +// filesystem.syncMu +// dentry.metadataMu +// *** "memmap.Mappable locks" below this point +// dentry.mapsMu +// *** "memmap.Mappable locks taken by Translate" below this point +// dentry.handleMu +// dentry.dataMu +// +// Locking dentry.dirMu in multiple dentries requires holding +// filesystem.renameMu for writing. +package gofer + +import ( + "fmt" + "strconv" + "sync" + "sync/atomic" + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/unet" + "gvisor.dev/gvisor/pkg/usermem" +) + +// FilesystemType implements vfs.FilesystemType. +type FilesystemType struct{} + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + vfsfs vfs.Filesystem + + // mfp is used to allocate memory that caches regular file contents. mfp is + // immutable. + mfp pgalloc.MemoryFileProvider + + // Immutable options. + opts filesystemOptions + + // client is the client used by this filesystem. client is immutable. + client *p9.Client + + // uid and gid are the effective KUID and KGID of the filesystem's creator, + // and are used as the owner and group for files that don't specify one. + // uid and gid are immutable. + uid auth.KUID + gid auth.KGID + + // renameMu serves two purposes: + // + // - It synchronizes path resolution with renaming initiated by this + // client. + // + // - It is held by path resolution to ensure that reachable dentries remain + // valid. A dentry is reachable by path resolution if it has a non-zero + // reference count (such that it is usable as vfs.ResolvingPath.Start() or + // is reachable from its children), or if it is a child dentry (such that + // it is reachable from its parent). + renameMu sync.RWMutex + + // cachedDentries contains all dentries with 0 references. (Due to race + // conditions, it may also contain dentries with non-zero references.) + // cachedDentriesLen is the number of dentries in cachedDentries. These + // fields are protected by renameMu. + cachedDentries dentryList + cachedDentriesLen uint64 + + // dentries contains all dentries in this filesystem. specialFileFDs + // contains all open specialFileFDs. These fields are protected by syncMu. + syncMu sync.Mutex + dentries map[*dentry]struct{} + specialFileFDs map[*specialFileFD]struct{} +} + +type filesystemOptions struct { + // "Standard" 9P options. + fd int + aname string + interop InteropMode // derived from the "cache" mount option + msize uint32 + version string + + // maxCachedDentries is the maximum number of dentries with 0 references + // retained by the client. + maxCachedDentries uint64 + + // If forcePageCache is true, host FDs may not be used for application + // memory mappings even if available; instead, the client must perform its + // own caching of regular file pages. This is primarily useful for testing. + forcePageCache bool + + // If limitHostFDTranslation is true, apply maxFillRange() constraints to + // host FD mappings returned by dentry.(memmap.Mappable).Translate(). This + // makes memory accounting behavior more consistent between cases where + // host FDs are / are not available, but may increase the frequency of + // sentry-handled page faults on files for which a host FD is available. + limitHostFDTranslation bool + + // If overlayfsStaleRead is true, O_RDONLY host FDs provided by the remote + // filesystem may not be coherent with writable host FDs opened later, so + // mappings of the former must be replaced by mappings of the latter. This + // is usually only the case when the remote filesystem is an overlayfs + // mount on Linux < 4.19. + overlayfsStaleRead bool + + // If regularFilesUseSpecialFileFD is true, application FDs representing + // regular files will use distinct file handles for each FD, in the same + // way that application FDs representing "special files" such as sockets + // do. Note that this disables client caching and mmap for regular files. + regularFilesUseSpecialFileFD bool +} + +// InteropMode controls the client's interaction with other remote filesystem +// users. +type InteropMode uint32 + +const ( + // InteropModeExclusive is appropriate when the filesystem client is the + // only user of the remote filesystem. + // + // - The client may cache arbitrary filesystem state (file data, metadata, + // filesystem structure, etc.). + // + // - Client changes to filesystem state may be sent to the remote + // filesystem asynchronously, except when server permission checks are + // necessary. + // + // - File timestamps are based on client clocks. This ensures that users of + // the client observe timestamps that are coherent with their own clocks + // and consistent with Linux's semantics. However, since it is not always + // possible for clients to set arbitrary atimes and mtimes, and never + // possible for clients to set arbitrary ctimes, file timestamp changes are + // stored in the client only and never sent to the remote filesystem. + InteropModeExclusive InteropMode = iota + + // InteropModeWritethrough is appropriate when there are read-only users of + // the remote filesystem that expect to observe changes made by the + // filesystem client. + // + // - The client may cache arbitrary filesystem state. + // + // - Client changes to filesystem state must be sent to the remote + // filesystem synchronously. + // + // - File timestamps are based on client clocks. As a corollary, access + // timestamp changes from other remote filesystem users will not be visible + // to the client. + InteropModeWritethrough + + // InteropModeShared is appropriate when there are users of the remote + // filesystem that may mutate its state other than the client. + // + // - The client must verify cached filesystem state before using it. + // + // - Client changes to filesystem state must be sent to the remote + // filesystem synchronously. + // + // - File timestamps are based on server clocks. This is necessary to + // ensure that timestamp changes are synchronized between remote filesystem + // users. + // + // Note that the correctness of InteropModeShared depends on the server + // correctly implementing 9P fids (i.e. each fid immutably represents a + // single filesystem object), even in the presence of remote filesystem + // mutations from other users. If this is violated, the behavior of the + // client is undefined. + InteropModeShared +) + +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. +func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + mfp := pgalloc.MemoryFileProviderFromContext(ctx) + if mfp == nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: context does not provide a pgalloc.MemoryFileProvider") + return nil, nil, syserror.EINVAL + } + + mopts := vfs.GenericParseMountOptions(opts.Data) + var fsopts filesystemOptions + + // Check that the transport is "fd". + trans, ok := mopts["trans"] + if !ok { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: transport must be specified as 'trans=fd'") + return nil, nil, syserror.EINVAL + } + delete(mopts, "trans") + if trans != "fd" { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: unsupported transport: trans=%s", trans) + return nil, nil, syserror.EINVAL + } + + // Check that read and write FDs are provided and identical. + rfdstr, ok := mopts["rfdno"] + if !ok { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: read FD must be specified as 'rfdno=") + return nil, nil, syserror.EINVAL + } + delete(mopts, "rfdno") + rfd, err := strconv.Atoi(rfdstr) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid read FD: rfdno=%s", rfdstr) + return nil, nil, syserror.EINVAL + } + wfdstr, ok := mopts["wfdno"] + if !ok { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: write FD must be specified as 'wfdno=") + return nil, nil, syserror.EINVAL + } + delete(mopts, "wfdno") + wfd, err := strconv.Atoi(wfdstr) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid write FD: wfdno=%s", wfdstr) + return nil, nil, syserror.EINVAL + } + if rfd != wfd { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: read FD (%d) and write FD (%d) must be equal", rfd, wfd) + return nil, nil, syserror.EINVAL + } + fsopts.fd = rfd + + // Get the attach name. + fsopts.aname = "/" + if aname, ok := mopts["aname"]; ok { + delete(mopts, "aname") + fsopts.aname = aname + } + + // Parse the cache policy. For historical reasons, this defaults to the + // least generally-applicable option, InteropModeExclusive. + fsopts.interop = InteropModeExclusive + if cache, ok := mopts["cache"]; ok { + delete(mopts, "cache") + switch cache { + case "fscache": + fsopts.interop = InteropModeExclusive + case "fscache_writethrough": + fsopts.interop = InteropModeWritethrough + case "none": + fsopts.regularFilesUseSpecialFileFD = true + fallthrough + case "remote_revalidating": + fsopts.interop = InteropModeShared + default: + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid cache policy: cache=%s", cache) + return nil, nil, syserror.EINVAL + } + } + + // Parse the 9P message size. + fsopts.msize = 1024 * 1024 // 1M, tested to give good enough performance up to 64M + if msizestr, ok := mopts["msize"]; ok { + delete(mopts, "msize") + msize, err := strconv.ParseUint(msizestr, 10, 32) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid message size: msize=%s", msizestr) + return nil, nil, syserror.EINVAL + } + fsopts.msize = uint32(msize) + } + + // Parse the 9P protocol version. + fsopts.version = p9.HighestVersionString() + if version, ok := mopts["version"]; ok { + delete(mopts, "version") + fsopts.version = version + } + + // Parse the dentry cache limit. + fsopts.maxCachedDentries = 1000 + if str, ok := mopts["dentry_cache_limit"]; ok { + delete(mopts, "dentry_cache_limit") + maxCachedDentries, err := strconv.ParseUint(str, 10, 64) + if err != nil { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid dentry cache limit: dentry_cache_limit=%s", str) + return nil, nil, syserror.EINVAL + } + fsopts.maxCachedDentries = maxCachedDentries + } + + // Handle simple flags. + if _, ok := mopts["force_page_cache"]; ok { + delete(mopts, "force_page_cache") + fsopts.forcePageCache = true + } + if _, ok := mopts["limit_host_fd_translation"]; ok { + delete(mopts, "limit_host_fd_translation") + fsopts.limitHostFDTranslation = true + } + if _, ok := mopts["overlayfs_stale_read"]; ok { + delete(mopts, "overlayfs_stale_read") + fsopts.overlayfsStaleRead = true + } + // fsopts.regularFilesUseSpecialFileFD can only be enabled by specifying + // "cache=none". + + // Check for unparsed options. + if len(mopts) != 0 { + ctx.Warningf("gofer.FilesystemType.GetFilesystem: unknown options: %v", mopts) + return nil, nil, syserror.EINVAL + } + + // Establish a connection with the server. + conn, err := unet.NewSocket(fsopts.fd) + if err != nil { + return nil, nil, err + } + + // Perform version negotiation with the server. + ctx.UninterruptibleSleepStart(false) + client, err := p9.NewClient(conn, fsopts.msize, fsopts.version) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + conn.Close() + return nil, nil, err + } + // Ownership of conn has been transferred to client. + + // Perform attach to obtain the filesystem root. + ctx.UninterruptibleSleepStart(false) + attached, err := client.Attach(fsopts.aname) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + client.Close() + return nil, nil, err + } + attachFile := p9file{attached} + qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask()) + if err != nil { + attachFile.close(ctx) + client.Close() + return nil, nil, err + } + + // Construct the filesystem object. + fs := &filesystem{ + mfp: mfp, + opts: fsopts, + uid: creds.EffectiveKUID, + gid: creds.EffectiveKGID, + client: client, + dentries: make(map[*dentry]struct{}), + specialFileFDs: make(map[*specialFileFD]struct{}), + } + fs.vfsfs.Init(vfsObj, fs) + + // Construct the root dentry. + root, err := fs.newDentry(ctx, attachFile, qid, attrMask, &attr) + if err != nil { + attachFile.close(ctx) + fs.vfsfs.DecRef() + return nil, nil, err + } + // Set the root's reference count to 2. One reference is returned to the + // caller, and the other is deliberately leaked to prevent the root from + // being "cached" and subsequently evicted. Its resources will still be + // cleaned up by fs.Release(). + root.refs = 2 + + return &fs.vfsfs, &root.vfsd, nil +} + +// Release implements vfs.FilesystemImpl.Release. +func (fs *filesystem) Release() { + ctx := context.Background() + mf := fs.mfp.MemoryFile() + + fs.syncMu.Lock() + for d := range fs.dentries { + d.handleMu.Lock() + d.dataMu.Lock() + if d.handleWritable { + // Write dirty cached data to the remote file. + if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt); err != nil { + log.Warningf("gofer.filesystem.Release: failed to flush dentry: %v", err) + } + // TODO(jamieliu): Do we need to flushf/fsync d? + } + // Discard cached pages. + d.cache.DropAll(mf) + d.dirty.RemoveAll() + d.dataMu.Unlock() + // Close the host fd if one exists. + if d.handle.fd >= 0 { + syscall.Close(int(d.handle.fd)) + d.handle.fd = -1 + } + d.handleMu.Unlock() + } + // There can't be any specialFileFDs still using fs, since each such + // FileDescription would hold a reference on a Mount holding a reference on + // fs. + fs.syncMu.Unlock() + + // Close the connection to the server. This implicitly clunks all fids. + fs.client.Close() +} + +// dentry implements vfs.DentryImpl. +type dentry struct { + vfsd vfs.Dentry + + // refs is the reference count. Each dentry holds a reference on its + // parent, even if disowned. refs is accessed using atomic memory + // operations. + refs int64 + + // fs is the owning filesystem. fs is immutable. + fs *filesystem + + // We don't support hard links, so each dentry maps 1:1 to an inode. + + // file is the unopened p9.File that backs this dentry. file is immutable. + file p9file + + // If deleted is non-zero, the file represented by this dentry has been + // deleted. deleted is accessed using atomic memory operations. + deleted uint32 + + // If cached is true, dentryEntry links dentry into + // filesystem.cachedDentries. cached and dentryEntry are protected by + // filesystem.renameMu. + cached bool + dentryEntry + + dirMu sync.Mutex + + // If this dentry represents a directory, and InteropModeShared is not in + // effect, negativeChildren is a set of child names in this directory that + // are known not to exist. negativeChildren is protected by dirMu. + negativeChildren map[string]struct{} + + // If this dentry represents a directory, InteropModeShared is not in + // effect, and dirents is not nil, it is a cache of all entries in the + // directory, in the order they were returned by the server. dirents is + // protected by dirMu. + dirents []vfs.Dirent + + // Cached metadata; protected by metadataMu and accessed using atomic + // memory operations unless otherwise specified. + metadataMu sync.Mutex + ino uint64 // immutable + mode uint32 // type is immutable, perms are mutable + uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic + gid uint32 // auth.KGID, but ... + blockSize uint32 // 0 if unknown + // Timestamps, all nsecs from the Unix epoch. + atime int64 + mtime int64 + ctime int64 + btime int64 + // File size, protected by both metadataMu and dataMu (i.e. both must be + // locked to mutate it). + size uint64 + + mapsMu sync.Mutex + + // If this dentry represents a regular file, mappings tracks mappings of + // the file into memmap.MappingSpaces. mappings is protected by mapsMu. + mappings memmap.MappingSet + + // If this dentry represents a regular file or directory: + // + // - handle is the I/O handle used by all regularFileFDs/directoryFDs + // representing this dentry. + // + // - handleReadable is true if handle is readable. + // + // - handleWritable is true if handle is writable. + // + // Invariants: + // + // - If handleReadable == handleWritable == false, then handle.file == nil + // (i.e. there is no open handle). Conversely, if handleReadable || + // handleWritable == true, then handle.file != nil (i.e. there is an open + // handle). + // + // - handleReadable and handleWritable cannot transition from true to false + // (i.e. handles may not be downgraded). + // + // These fields are protected by handleMu. + handleMu sync.RWMutex + handle handle + handleReadable bool + handleWritable bool + + dataMu sync.RWMutex + + // If this dentry represents a regular file that is client-cached, cache + // maps offsets into the cached file to offsets into + // filesystem.mfp.MemoryFile() that store the file's data. cache is + // protected by dataMu. + cache fsutil.FileRangeSet + + // If this dentry represents a regular file that is client-cached, dirty + // tracks dirty segments in cache. dirty is protected by dataMu. + dirty fsutil.DirtySet + + // pf implements platform.File for mappings of handle.fd. + pf dentryPlatformFile + + // If this dentry represents a symbolic link, InteropModeShared is not in + // effect, and haveTarget is true, target is the symlink target. haveTarget + // and target are protected by dataMu. + haveTarget bool + target string +} + +// dentryAttrMask returns a p9.AttrMask enabling all attributes used by the +// gofer client. +func dentryAttrMask() p9.AttrMask { + return p9.AttrMask{ + Mode: true, + UID: true, + GID: true, + ATime: true, + MTime: true, + CTime: true, + Size: true, + BTime: true, + } +} + +// newDentry creates a new dentry representing the given file. The dentry +// initially has no references, but is not cached; it is the caller's +// responsibility to set the dentry's reference count and/or call +// dentry.checkCachingLocked() as appropriate. +func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, mask p9.AttrMask, attr *p9.Attr) (*dentry, error) { + if !mask.Mode { + ctx.Warningf("can't create gofer.dentry without file type") + return nil, syserror.EIO + } + if attr.Mode.FileType() == p9.ModeRegular && !mask.Size { + ctx.Warningf("can't create regular file gofer.dentry without file size") + return nil, syserror.EIO + } + + d := &dentry{ + fs: fs, + file: file, + ino: qid.Path, + mode: uint32(attr.Mode), + uid: uint32(fs.uid), + gid: uint32(fs.gid), + blockSize: usermem.PageSize, + handle: handle{ + fd: -1, + }, + } + d.pf.dentry = d + if mask.UID { + d.uid = uint32(attr.UID) + } + if mask.GID { + d.gid = uint32(attr.GID) + } + if mask.Size { + d.size = attr.Size + } + if attr.BlockSize != 0 { + d.blockSize = uint32(attr.BlockSize) + } + if mask.ATime { + d.atime = dentryTimestampFromP9(attr.ATimeSeconds, attr.ATimeNanoSeconds) + } + if mask.MTime { + d.mtime = dentryTimestampFromP9(attr.MTimeSeconds, attr.MTimeNanoSeconds) + } + if mask.CTime { + d.ctime = dentryTimestampFromP9(attr.CTimeSeconds, attr.CTimeNanoSeconds) + } + if mask.BTime { + d.btime = dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds) + } + d.vfsd.Init(d) + + fs.syncMu.Lock() + fs.dentries[d] = struct{}{} + fs.syncMu.Unlock() + return d, nil +} + +// updateFromP9Attrs is called to update d's metadata after an update from the +// remote filesystem. +func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) { + d.metadataMu.Lock() + if mask.Mode { + if got, want := uint32(attr.Mode.FileType()), d.fileType(); got != want { + d.metadataMu.Unlock() + panic(fmt.Sprintf("gofer.dentry file type changed from %#o to %#o", want, got)) + } + atomic.StoreUint32(&d.mode, uint32(attr.Mode)) + } + if mask.UID { + atomic.StoreUint32(&d.uid, uint32(attr.UID)) + } + if mask.GID { + atomic.StoreUint32(&d.gid, uint32(attr.GID)) + } + // There is no P9_GETATTR_* bit for I/O block size. + if attr.BlockSize != 0 { + atomic.StoreUint32(&d.blockSize, uint32(attr.BlockSize)) + } + if mask.ATime { + atomic.StoreInt64(&d.atime, dentryTimestampFromP9(attr.ATimeSeconds, attr.ATimeNanoSeconds)) + } + if mask.MTime { + atomic.StoreInt64(&d.mtime, dentryTimestampFromP9(attr.MTimeSeconds, attr.MTimeNanoSeconds)) + } + if mask.CTime { + atomic.StoreInt64(&d.ctime, dentryTimestampFromP9(attr.CTimeSeconds, attr.CTimeNanoSeconds)) + } + if mask.BTime { + atomic.StoreInt64(&d.btime, dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds)) + } + if mask.Size { + d.dataMu.Lock() + atomic.StoreUint64(&d.size, attr.Size) + d.dataMu.Unlock() + } + d.metadataMu.Unlock() +} + +func (d *dentry) updateFromGetattr(ctx context.Context) error { + // Use d.handle.file, which represents a 9P fid that has been opened, in + // preference to d.file, which represents a 9P fid that has not. This may + // be significantly more efficient in some implementations. + var ( + file p9file + handleMuRLocked bool + ) + d.handleMu.RLock() + if !d.handle.file.isNil() { + file = d.handle.file + handleMuRLocked = true + } else { + file = d.file + d.handleMu.RUnlock() + } + _, attrMask, attr, err := file.getAttr(ctx, dentryAttrMask()) + if handleMuRLocked { + d.handleMu.RUnlock() + } + if err != nil { + return err + } + d.updateFromP9Attrs(attrMask, &attr) + return nil +} + +func (d *dentry) fileType() uint32 { + return atomic.LoadUint32(&d.mode) & linux.S_IFMT +} + +func (d *dentry) statTo(stat *linux.Statx) { + stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS | linux.STATX_BTIME + stat.Blksize = atomic.LoadUint32(&d.blockSize) + stat.Nlink = 1 + if d.isDir() { + stat.Nlink = 2 + } + stat.UID = atomic.LoadUint32(&d.uid) + stat.GID = atomic.LoadUint32(&d.gid) + stat.Mode = uint16(atomic.LoadUint32(&d.mode)) + stat.Ino = d.ino + stat.Size = atomic.LoadUint64(&d.size) + // This is consistent with regularFileFD.Seek(), which treats regular files + // as having no holes. + stat.Blocks = (stat.Size + 511) / 512 + stat.Atime = statxTimestampFromDentry(atomic.LoadInt64(&d.atime)) + stat.Btime = statxTimestampFromDentry(atomic.LoadInt64(&d.btime)) + stat.Ctime = statxTimestampFromDentry(atomic.LoadInt64(&d.ctime)) + stat.Mtime = statxTimestampFromDentry(atomic.LoadInt64(&d.mtime)) + // TODO(jamieliu): device number +} + +func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx, mnt *vfs.Mount) error { + if stat.Mask == 0 { + return nil + } + if stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_SIZE) != 0 { + return syserror.EPERM + } + if err := vfs.CheckSetStat(creds, stat, uint16(atomic.LoadUint32(&d.mode))&^linux.S_IFMT, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { + return err + } + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + setLocalAtime := false + setLocalMtime := false + if d.fs.opts.interop != InteropModeShared { + // Timestamp updates will be handled locally. + setLocalAtime = stat.Mask&linux.STATX_ATIME != 0 + setLocalMtime = stat.Mask&linux.STATX_MTIME != 0 + stat.Mask &^= linux.STATX_ATIME | linux.STATX_MTIME + if !setLocalMtime && (stat.Mask&linux.STATX_SIZE != 0) { + // Truncate updates mtime. + setLocalMtime = true + stat.Mtime.Nsec = linux.UTIME_NOW + } + } + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + if stat.Mask != 0 { + if err := d.file.setAttr(ctx, p9.SetAttrMask{ + Permissions: stat.Mask&linux.STATX_MODE != 0, + UID: stat.Mask&linux.STATX_UID != 0, + GID: stat.Mask&linux.STATX_GID != 0, + Size: stat.Mask&linux.STATX_SIZE != 0, + ATime: stat.Mask&linux.STATX_ATIME != 0, + MTime: stat.Mask&linux.STATX_MTIME != 0, + ATimeNotSystemTime: stat.Atime.Nsec != linux.UTIME_NOW, + MTimeNotSystemTime: stat.Mtime.Nsec != linux.UTIME_NOW, + }, p9.SetAttr{ + Permissions: p9.FileMode(stat.Mode), + UID: p9.UID(stat.UID), + GID: p9.GID(stat.GID), + Size: stat.Size, + ATimeSeconds: uint64(stat.Atime.Sec), + ATimeNanoSeconds: uint64(stat.Atime.Nsec), + MTimeSeconds: uint64(stat.Mtime.Sec), + MTimeNanoSeconds: uint64(stat.Mtime.Nsec), + }); err != nil { + return err + } + } + if d.fs.opts.interop == InteropModeShared { + // There's no point to updating d's metadata in this case since it'll + // be overwritten by revalidation before the next time it's used + // anyway. (InteropModeShared inhibits client caching of regular file + // data, so there's no cache to truncate either.) + return nil + } + now, haveNow := nowFromContext(ctx) + if !haveNow { + ctx.Warningf("gofer.dentry.setStat: current time not available") + } + if stat.Mask&linux.STATX_MODE != 0 { + atomic.StoreUint32(&d.mode, d.fileType()|uint32(stat.Mode)) + } + if stat.Mask&linux.STATX_UID != 0 { + atomic.StoreUint32(&d.uid, stat.UID) + } + if stat.Mask&linux.STATX_GID != 0 { + atomic.StoreUint32(&d.gid, stat.GID) + } + if setLocalAtime { + if stat.Atime.Nsec == linux.UTIME_NOW { + if haveNow { + atomic.StoreInt64(&d.atime, now) + } + } else { + atomic.StoreInt64(&d.atime, dentryTimestampFromStatx(stat.Atime)) + } + } + if setLocalMtime { + if stat.Mtime.Nsec == linux.UTIME_NOW { + if haveNow { + atomic.StoreInt64(&d.mtime, now) + } + } else { + atomic.StoreInt64(&d.mtime, dentryTimestampFromStatx(stat.Mtime)) + } + } + if haveNow { + atomic.StoreInt64(&d.ctime, now) + } + if stat.Mask&linux.STATX_SIZE != 0 { + d.dataMu.Lock() + oldSize := d.size + d.size = stat.Size + // d.dataMu must be unlocked to lock d.mapsMu and invalidate mappings + // below. This allows concurrent calls to Read/Translate/etc. These + // functions synchronize with truncation by refusing to use cache + // contents beyond the new d.size. (We are still holding d.metadataMu, + // so we can't race with Write or another truncate.) + d.dataMu.Unlock() + if d.size < oldSize { + oldpgend := pageRoundUp(oldSize) + newpgend := pageRoundUp(d.size) + if oldpgend != newpgend { + d.mapsMu.Lock() + d.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{ + // Compare Linux's mm/truncate.c:truncate_setsize() => + // truncate_pagecache() => + // mm/memory.c:unmap_mapping_range(evencows=1). + InvalidatePrivate: true, + }) + d.mapsMu.Unlock() + } + // We are now guaranteed that there are no translations of + // truncated pages, and can remove them from the cache. Since + // truncated pages have been removed from the remote file, they + // should be dropped without being written back. + d.dataMu.Lock() + d.cache.Truncate(d.size, d.fs.mfp.MemoryFile()) + d.dirty.KeepClean(memmap.MappableRange{d.size, oldpgend}) + d.dataMu.Unlock() + } + } + return nil +} + +func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error { + return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&d.mode))&0777, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) +} + +// IncRef implements vfs.DentryImpl.IncRef. +func (d *dentry) IncRef() { + // d.refs may be 0 if d.fs.renameMu is locked, which serializes against + // d.checkCachingLocked(). + atomic.AddInt64(&d.refs, 1) +} + +// TryIncRef implements vfs.DentryImpl.TryIncRef. +func (d *dentry) TryIncRef() bool { + for { + refs := atomic.LoadInt64(&d.refs) + if refs == 0 { + return false + } + if atomic.CompareAndSwapInt64(&d.refs, refs, refs+1) { + return true + } + } +} + +// DecRef implements vfs.DentryImpl.DecRef. +func (d *dentry) DecRef() { + if refs := atomic.AddInt64(&d.refs, -1); refs == 0 { + d.fs.renameMu.Lock() + d.checkCachingLocked() + d.fs.renameMu.Unlock() + } else if refs < 0 { + panic("gofer.dentry.DecRef() called without holding a reference") + } +} + +// checkCachingLocked should be called after d's reference count becomes 0 or it +// becomes disowned. +// +// Preconditions: d.fs.renameMu must be locked for writing. +func (d *dentry) checkCachingLocked() { + // Dentries with a non-zero reference count must be retained. (The only way + // to obtain a reference on a dentry with zero references is via path + // resolution, which requires renameMu, so if d.refs is zero then it will + // remain zero while we hold renameMu for writing.) + if atomic.LoadInt64(&d.refs) != 0 { + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentriesLen-- + d.cached = false + } + return + } + // Non-child dentries with zero references are no longer reachable by path + // resolution and should be dropped immediately. + if d.vfsd.Parent() == nil || d.vfsd.IsDisowned() { + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentriesLen-- + d.cached = false + } + d.destroyLocked() + return + } + // If d is already cached, just move it to the front of the LRU. + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentries.PushFront(d) + return + } + // Cache the dentry, then evict the least recently used cached dentry if + // the cache becomes over-full. + d.fs.cachedDentries.PushFront(d) + d.fs.cachedDentriesLen++ + d.cached = true + if d.fs.cachedDentriesLen > d.fs.opts.maxCachedDentries { + victim := d.fs.cachedDentries.Back() + d.fs.cachedDentries.Remove(victim) + d.fs.cachedDentriesLen-- + victim.cached = false + // victim.refs may have become non-zero from an earlier path + // resolution since it was inserted into fs.cachedDentries; see + // dentry.incRefLocked(). Either way, we brought + // fs.cachedDentriesLen back down to fs.opts.maxCachedDentries, so + // we don't loop. + if atomic.LoadInt64(&victim.refs) == 0 { + if victimParentVFSD := victim.vfsd.Parent(); victimParentVFSD != nil { + victimParent := victimParentVFSD.Impl().(*dentry) + victimParent.dirMu.Lock() + if !victim.vfsd.IsDisowned() { + // victim can't be a mount point (in any mount + // namespace), since VFS holds references on mount + // points. + d.fs.vfsfs.VirtualFilesystem().ForceDeleteDentry(&victim.vfsd) + // We're only deleting the dentry, not the file it + // represents, so we don't need to update + // victimParent.dirents etc. + } + victimParent.dirMu.Unlock() + } + victim.destroyLocked() + } + } +} + +// Preconditions: d.fs.renameMu must be locked for writing. d.refs == 0. d is +// not a child dentry. +func (d *dentry) destroyLocked() { + ctx := context.Background() + d.handleMu.Lock() + if !d.handle.file.isNil() { + mf := d.fs.mfp.MemoryFile() + d.dataMu.Lock() + // Write dirty pages back to the remote filesystem. + if d.handleWritable { + if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil { + log.Warningf("gofer.dentry.DecRef: failed to write dirty data back: %v", err) + } + } + // Discard cached data. + d.cache.DropAll(mf) + d.dirty.RemoveAll() + d.dataMu.Unlock() + // Clunk open fids and close open host FDs. + d.handle.close(ctx) + } + d.handleMu.Unlock() + d.file.close(ctx) + // Remove d from the set of all dentries. + d.fs.syncMu.Lock() + delete(d.fs.dentries, d) + d.fs.syncMu.Unlock() + // Drop the reference held by d on its parent. + if parentVFSD := d.vfsd.Parent(); parentVFSD != nil { + parent := parentVFSD.Impl().(*dentry) + // This is parent.DecRef() without recursive locking of d.fs.renameMu. + if refs := atomic.AddInt64(&parent.refs, -1); refs == 0 { + parent.checkCachingLocked() + } else if refs < 0 { + panic("gofer.dentry.DecRef() called without holding a reference") + } + } +} + +func (d *dentry) isDeleted() bool { + return atomic.LoadUint32(&d.deleted) != 0 +} + +func (d *dentry) setDeleted() { + atomic.StoreUint32(&d.deleted, 1) +} + +func (d *dentry) listxattr(ctx context.Context) ([]string, error) { + return nil, syserror.ENOTSUP +} + +func (d *dentry) getxattr(ctx context.Context, name string) (string, error) { + // TODO(jamieliu): add vfs.GetxattrOptions.Size + return d.file.getXattr(ctx, name, linux.XATTR_SIZE_MAX) +} + +func (d *dentry) setxattr(ctx context.Context, opts *vfs.SetxattrOptions) error { + return d.file.setXattr(ctx, opts.Name, opts.Value, opts.Flags) +} + +func (d *dentry) removexattr(ctx context.Context, name string) error { + return syserror.ENOTSUP +} + +// Preconditions: d.isRegularFile() || d.isDirectory(). +func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool) error { + // O_TRUNC unconditionally requires us to obtain a new handle (opened with + // O_TRUNC). + if !trunc { + d.handleMu.RLock() + if (!read || d.handleReadable) && (!write || d.handleWritable) { + // The current handle is sufficient. + d.handleMu.RUnlock() + return nil + } + d.handleMu.RUnlock() + } + + haveOldFD := false + d.handleMu.Lock() + if (read && !d.handleReadable) || (write && !d.handleWritable) || trunc { + // Get a new handle. + wantReadable := d.handleReadable || read + wantWritable := d.handleWritable || write + h, err := openHandle(ctx, d.file, wantReadable, wantWritable, trunc) + if err != nil { + d.handleMu.Unlock() + return err + } + if !d.handle.file.isNil() { + // Check that old and new handles are compatible: If the old handle + // includes a host file descriptor but the new one does not, or + // vice versa, old and new memory mappings may be incoherent. + haveOldFD = d.handle.fd >= 0 + haveNewFD := h.fd >= 0 + if haveOldFD != haveNewFD { + d.handleMu.Unlock() + ctx.Warningf("gofer.dentry.ensureSharedHandle: can't change host FD availability from %v to %v across dentry handle upgrade", haveOldFD, haveNewFD) + h.close(ctx) + return syserror.EIO + } + if haveOldFD { + // We may have raced with callers of d.pf.FD() that are now + // using the old file descriptor, preventing us from safely + // closing it. We could handle this by invalidating existing + // memmap.Translations, but this is expensive. Instead, use + // dup2() to make the old file descriptor refer to the new file + // description, then close the new file descriptor (which is no + // longer needed). Racing callers may use the old or new file + // description, but this doesn't matter since they refer to the + // same file (unless d.fs.opts.overlayfsStaleRead is true, + // which we handle separately). + if err := syscall.Dup2(int(h.fd), int(d.handle.fd)); err != nil { + d.handleMu.Unlock() + ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to dup fd %d to fd %d: %v", h.fd, d.handle.fd, err) + h.close(ctx) + return err + } + syscall.Close(int(h.fd)) + h.fd = d.handle.fd + if d.fs.opts.overlayfsStaleRead { + // Replace sentry mappings of the old FD with mappings of + // the new FD, since the two are not necessarily coherent. + if err := d.pf.hostFileMapper.RegenerateMappings(int(h.fd)); err != nil { + d.handleMu.Unlock() + ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to replace sentry mappings of old FD with mappings of new FD: %v", err) + h.close(ctx) + return err + } + } + // Clunk the old fid before making the new handle visible (by + // unlocking d.handleMu). + d.handle.file.close(ctx) + } + } + // Switch to the new handle. + d.handle = h + d.handleReadable = wantReadable + d.handleWritable = wantWritable + } + d.handleMu.Unlock() + + if d.fs.opts.overlayfsStaleRead && haveOldFD { + // Invalidate application mappings that may be using the old FD; they + // will be replaced with mappings using the new FD after future calls + // to d.Translate(). This requires holding d.mapsMu, which precedes + // d.handleMu in the lock order. + d.mapsMu.Lock() + d.mappings.InvalidateAll(memmap.InvalidateOpts{}) + d.mapsMu.Unlock() + } + + return nil +} + +// fileDescription is embedded by gofer implementations of +// vfs.FileDescriptionImpl. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl +} + +func (fd *fileDescription) filesystem() *filesystem { + return fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) +} + +func (fd *fileDescription) dentry() *dentry { + return fd.vfsfd.Dentry().Impl().(*dentry) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + d := fd.dentry() + if d.fs.opts.interop == InteropModeShared && opts.Mask&(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE|linux.STATX_BLOCKS|linux.STATX_BTIME) != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC { + // TODO(jamieliu): Use specialFileFD.handle.file for the getattr if + // available? + if err := d.updateFromGetattr(ctx); err != nil { + return linux.Statx{}, err + } + } + var stat linux.Statx + d.statTo(&stat) + return stat, nil +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + return fd.dentry().setStat(ctx, auth.CredentialsFromContext(ctx), &opts.Stat, fd.vfsfd.Mount()) +} + +// Listxattr implements vfs.FileDescriptionImpl.Listxattr. +func (fd *fileDescription) Listxattr(ctx context.Context) ([]string, error) { + return fd.dentry().listxattr(ctx) +} + +// Getxattr implements vfs.FileDescriptionImpl.Getxattr. +func (fd *fileDescription) Getxattr(ctx context.Context, name string) (string, error) { + return fd.dentry().getxattr(ctx, name) +} + +// Setxattr implements vfs.FileDescriptionImpl.Setxattr. +func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { + return fd.dentry().setxattr(ctx, &opts) +} + +// Removexattr implements vfs.FileDescriptionImpl.Removexattr. +func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { + return fd.dentry().removexattr(ctx, name) +} diff --git a/pkg/sentry/fsimpl/gofer/handle.go b/pkg/sentry/fsimpl/gofer/handle.go new file mode 100644 index 000000000..cfe66f797 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/handle.go @@ -0,0 +1,135 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/safemem" +) + +// handle represents a remote "open file descriptor", consisting of an opened +// fid (p9.File) and optionally a host file descriptor. +type handle struct { + file p9file + fd int32 // -1 if unavailable +} + +// Preconditions: read || write. +func openHandle(ctx context.Context, file p9file, read, write, trunc bool) (handle, error) { + _, newfile, err := file.walk(ctx, nil) + if err != nil { + return handle{fd: -1}, err + } + var flags p9.OpenFlags + switch { + case read && !write: + flags = p9.ReadOnly + case !read && write: + flags = p9.WriteOnly + case read && write: + flags = p9.ReadWrite + } + if trunc { + flags |= p9.OpenTruncate + } + fdobj, _, _, err := newfile.open(ctx, flags) + if err != nil { + newfile.close(ctx) + return handle{fd: -1}, err + } + fd := int32(-1) + if fdobj != nil { + fd = int32(fdobj.Release()) + } + return handle{ + file: newfile, + fd: fd, + }, nil +} + +func (h *handle) close(ctx context.Context) { + h.file.close(ctx) + h.file = p9file{} + if h.fd >= 0 { + syscall.Close(int(h.fd)) + h.fd = -1 + } +} + +func (h *handle) readToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error) { + if dsts.IsEmpty() { + return 0, nil + } + if h.fd >= 0 { + ctx.UninterruptibleSleepStart(false) + n, err := hostPreadv(h.fd, dsts, int64(offset)) + ctx.UninterruptibleSleepFinish(false) + return n, err + } + if dsts.NumBlocks() == 1 && !dsts.Head().NeedSafecopy() { + n, err := h.file.readAt(ctx, dsts.Head().ToSlice(), offset) + return uint64(n), err + } + // Buffer the read since p9.File.ReadAt() takes []byte. + buf := make([]byte, dsts.NumBytes()) + n, err := h.file.readAt(ctx, buf, offset) + if n == 0 { + return 0, err + } + if cp, cperr := safemem.CopySeq(dsts, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:n]))); cperr != nil { + return cp, cperr + } + return uint64(n), err +} + +func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) { + if srcs.IsEmpty() { + return 0, nil + } + if h.fd >= 0 { + ctx.UninterruptibleSleepStart(false) + n, err := hostPwritev(h.fd, srcs, int64(offset)) + ctx.UninterruptibleSleepFinish(false) + return n, err + } + if srcs.NumBlocks() == 1 && !srcs.Head().NeedSafecopy() { + n, err := h.file.writeAt(ctx, srcs.Head().ToSlice(), offset) + return uint64(n), err + } + // Buffer the write since p9.File.WriteAt() takes []byte. + buf := make([]byte, srcs.NumBytes()) + cp, cperr := safemem.CopySeq(safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), srcs) + if cp == 0 { + return 0, cperr + } + n, err := h.file.writeAt(ctx, buf[:cp], offset) + if err != nil { + return uint64(n), err + } + return cp, cperr +} + +func (h *handle) sync(ctx context.Context) error { + if h.fd >= 0 { + ctx.UninterruptibleSleepStart(false) + err := syscall.Fsync(int(h.fd)) + ctx.UninterruptibleSleepFinish(false) + return err + } + return h.file.fsync(ctx) +} diff --git a/pkg/sentry/fsimpl/gofer/handle_unsafe.go b/pkg/sentry/fsimpl/gofer/handle_unsafe.go new file mode 100644 index 000000000..19560ab26 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/handle_unsafe.go @@ -0,0 +1,66 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "syscall" + "unsafe" + + "gvisor.dev/gvisor/pkg/safemem" +) + +// Preconditions: !dsts.IsEmpty(). +func hostPreadv(fd int32, dsts safemem.BlockSeq, off int64) (uint64, error) { + // No buffering is necessary regardless of safecopy; host syscalls will + // return EFAULT if appropriate, instead of raising SIGBUS. + if dsts.NumBlocks() == 1 { + // Use pread() instead of preadv() to avoid iovec allocation and + // copying. + dst := dsts.Head() + n, _, e := syscall.Syscall6(syscall.SYS_PREAD64, uintptr(fd), dst.Addr(), uintptr(dst.Len()), uintptr(off), 0, 0) + if e != 0 { + return 0, e + } + return uint64(n), nil + } + iovs := safemem.IovecsFromBlockSeq(dsts) + n, _, e := syscall.Syscall6(syscall.SYS_PREADV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0) + if e != 0 { + return 0, e + } + return uint64(n), nil +} + +// Preconditions: !srcs.IsEmpty(). +func hostPwritev(fd int32, srcs safemem.BlockSeq, off int64) (uint64, error) { + // No buffering is necessary regardless of safecopy; host syscalls will + // return EFAULT if appropriate, instead of raising SIGBUS. + if srcs.NumBlocks() == 1 { + // Use pwrite() instead of pwritev() to avoid iovec allocation and + // copying. + src := srcs.Head() + n, _, e := syscall.Syscall6(syscall.SYS_PWRITE64, uintptr(fd), src.Addr(), uintptr(src.Len()), uintptr(off), 0, 0) + if e != 0 { + return 0, e + } + return uint64(n), nil + } + iovs := safemem.IovecsFromBlockSeq(srcs) + n, _, e := syscall.Syscall6(syscall.SYS_PWRITEV, uintptr(fd), uintptr((unsafe.Pointer)(&iovs[0])), uintptr(len(iovs)), uintptr(off), 0, 0) + if e != 0 { + return 0, e + } + return uint64(n), nil +} diff --git a/pkg/sentry/fsimpl/gofer/p9file.go b/pkg/sentry/fsimpl/gofer/p9file.go new file mode 100644 index 000000000..755ac2985 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/p9file.go @@ -0,0 +1,219 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fd" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/syserror" +) + +// p9file is a wrapper around p9.File that provides methods that are +// Context-aware. +type p9file struct { + file p9.File +} + +func (f p9file) isNil() bool { + return f.file == nil +} + +func (f p9file) walk(ctx context.Context, names []string) ([]p9.QID, p9file, error) { + ctx.UninterruptibleSleepStart(false) + qids, newfile, err := f.file.Walk(names) + ctx.UninterruptibleSleepFinish(false) + return qids, p9file{newfile}, err +} + +func (f p9file) walkGetAttr(ctx context.Context, names []string) ([]p9.QID, p9file, p9.AttrMask, p9.Attr, error) { + ctx.UninterruptibleSleepStart(false) + qids, newfile, attrMask, attr, err := f.file.WalkGetAttr(names) + ctx.UninterruptibleSleepFinish(false) + return qids, p9file{newfile}, attrMask, attr, err +} + +// walkGetAttrOne is a wrapper around p9.File.WalkGetAttr that takes a single +// path component and returns a single qid. +func (f p9file) walkGetAttrOne(ctx context.Context, name string) (p9.QID, p9file, p9.AttrMask, p9.Attr, error) { + ctx.UninterruptibleSleepStart(false) + qids, newfile, attrMask, attr, err := f.file.WalkGetAttr([]string{name}) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + return p9.QID{}, p9file{}, p9.AttrMask{}, p9.Attr{}, err + } + if len(qids) != 1 { + ctx.Warningf("p9.File.WalkGetAttr returned %d qids (%v), wanted 1", len(qids), qids) + if newfile != nil { + p9file{newfile}.close(ctx) + } + return p9.QID{}, p9file{}, p9.AttrMask{}, p9.Attr{}, syserror.EIO + } + return qids[0], p9file{newfile}, attrMask, attr, nil +} + +func (f p9file) statFS(ctx context.Context) (p9.FSStat, error) { + ctx.UninterruptibleSleepStart(false) + fsstat, err := f.file.StatFS() + ctx.UninterruptibleSleepFinish(false) + return fsstat, err +} + +func (f p9file) getAttr(ctx context.Context, req p9.AttrMask) (p9.QID, p9.AttrMask, p9.Attr, error) { + ctx.UninterruptibleSleepStart(false) + qid, attrMask, attr, err := f.file.GetAttr(req) + ctx.UninterruptibleSleepFinish(false) + return qid, attrMask, attr, err +} + +func (f p9file) setAttr(ctx context.Context, valid p9.SetAttrMask, attr p9.SetAttr) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.SetAttr(valid, attr) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) getXattr(ctx context.Context, name string, size uint64) (string, error) { + ctx.UninterruptibleSleepStart(false) + val, err := f.file.GetXattr(name, size) + ctx.UninterruptibleSleepFinish(false) + return val, err +} + +func (f p9file) setXattr(ctx context.Context, name, value string, flags uint32) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.SetXattr(name, value, flags) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) allocate(ctx context.Context, mode p9.AllocateMode, offset, length uint64) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Allocate(mode, offset, length) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) close(ctx context.Context) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Close() + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) open(ctx context.Context, flags p9.OpenFlags) (*fd.FD, p9.QID, uint32, error) { + ctx.UninterruptibleSleepStart(false) + fdobj, qid, iounit, err := f.file.Open(flags) + ctx.UninterruptibleSleepFinish(false) + return fdobj, qid, iounit, err +} + +func (f p9file) readAt(ctx context.Context, p []byte, offset uint64) (int, error) { + ctx.UninterruptibleSleepStart(false) + n, err := f.file.ReadAt(p, offset) + ctx.UninterruptibleSleepFinish(false) + return n, err +} + +func (f p9file) writeAt(ctx context.Context, p []byte, offset uint64) (int, error) { + ctx.UninterruptibleSleepStart(false) + n, err := f.file.WriteAt(p, offset) + ctx.UninterruptibleSleepFinish(false) + return n, err +} + +func (f p9file) fsync(ctx context.Context) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.FSync() + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) create(ctx context.Context, name string, flags p9.OpenFlags, permissions p9.FileMode, uid p9.UID, gid p9.GID) (*fd.FD, p9file, p9.QID, uint32, error) { + ctx.UninterruptibleSleepStart(false) + fdobj, newfile, qid, iounit, err := f.file.Create(name, flags, permissions, uid, gid) + ctx.UninterruptibleSleepFinish(false) + return fdobj, p9file{newfile}, qid, iounit, err +} + +func (f p9file) mkdir(ctx context.Context, name string, permissions p9.FileMode, uid p9.UID, gid p9.GID) (p9.QID, error) { + ctx.UninterruptibleSleepStart(false) + qid, err := f.file.Mkdir(name, permissions, uid, gid) + ctx.UninterruptibleSleepFinish(false) + return qid, err +} + +func (f p9file) symlink(ctx context.Context, oldName string, newName string, uid p9.UID, gid p9.GID) (p9.QID, error) { + ctx.UninterruptibleSleepStart(false) + qid, err := f.file.Symlink(oldName, newName, uid, gid) + ctx.UninterruptibleSleepFinish(false) + return qid, err +} + +func (f p9file) link(ctx context.Context, target p9file, newName string) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Link(target.file, newName) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) mknod(ctx context.Context, name string, mode p9.FileMode, major uint32, minor uint32, uid p9.UID, gid p9.GID) (p9.QID, error) { + ctx.UninterruptibleSleepStart(false) + qid, err := f.file.Mknod(name, mode, major, minor, uid, gid) + ctx.UninterruptibleSleepFinish(false) + return qid, err +} + +func (f p9file) rename(ctx context.Context, newDir p9file, newName string) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Rename(newDir.file, newName) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) unlinkAt(ctx context.Context, name string, flags uint32) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.UnlinkAt(name, flags) + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) readdir(ctx context.Context, offset uint64, count uint32) ([]p9.Dirent, error) { + ctx.UninterruptibleSleepStart(false) + dirents, err := f.file.Readdir(offset, count) + ctx.UninterruptibleSleepFinish(false) + return dirents, err +} + +func (f p9file) readlink(ctx context.Context) (string, error) { + ctx.UninterruptibleSleepStart(false) + target, err := f.file.Readlink() + ctx.UninterruptibleSleepFinish(false) + return target, err +} + +func (f p9file) flush(ctx context.Context) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.Flush() + ctx.UninterruptibleSleepFinish(false) + return err +} + +func (f p9file) connect(ctx context.Context, flags p9.ConnectFlags) (*fd.FD, error) { + ctx.UninterruptibleSleepStart(false) + fdobj, err := f.file.Connect(flags) + ctx.UninterruptibleSleepFinish(false) + return fdobj, err +} diff --git a/pkg/sentry/fsimpl/gofer/pagemath.go b/pkg/sentry/fsimpl/gofer/pagemath.go new file mode 100644 index 000000000..847cb0784 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/pagemath.go @@ -0,0 +1,31 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "gvisor.dev/gvisor/pkg/usermem" +) + +// This are equivalent to usermem.Addr.RoundDown/Up, but without the +// potentially truncating conversion to usermem.Addr. This is necessary because +// there is no way to define generic "PageRoundDown/Up" functions in Go. + +func pageRoundDown(x uint64) uint64 { + return x &^ (usermem.PageSize - 1) +} + +func pageRoundUp(x uint64) uint64 { + return pageRoundDown(x + usermem.PageSize - 1) +} diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go new file mode 100644 index 000000000..8e11e06b3 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -0,0 +1,860 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "fmt" + "io" + "math" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +func (d *dentry) isRegularFile() bool { + return d.fileType() == linux.S_IFREG +} + +type regularFileFD struct { + fileDescription + + // off is the file offset. off is protected by mu. + mu sync.Mutex + off int64 +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *regularFileFD) Release() { +} + +// OnClose implements vfs.FileDescriptionImpl.OnClose. +func (fd *regularFileFD) OnClose(ctx context.Context) error { + if !fd.vfsfd.IsWritable() { + return nil + } + // Skip flushing if writes may be buffered by the client, since (as with + // the VFS1 client) we don't flush buffered writes on close anyway. + d := fd.dentry() + if d.fs.opts.interop == InteropModeExclusive { + return nil + } + d.handleMu.RLock() + defer d.handleMu.RUnlock() + return d.handle.file.flush(ctx) +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + // Check for reading at EOF before calling into MM (but not under + // InteropModeShared, which makes d.size unreliable). + d := fd.dentry() + if d.fs.opts.interop != InteropModeShared && uint64(offset) >= atomic.LoadUint64(&d.size) { + return 0, io.EOF + } + + if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + // Lock d.metadataMu for the rest of the read to prevent d.size from + // changing. + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + // Write dirty cached pages that will be touched by the read back to + // the remote file. + if err := d.writeback(ctx, offset, dst.NumBytes()); err != nil { + return 0, err + } + } + + rw := getDentryReadWriter(ctx, d, offset) + if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + // Require the read to go to the remote file. + rw.direct = true + } + n, err := dst.CopyOutFrom(ctx, rw) + putDentryReadWriter(rw) + if d.fs.opts.interop != InteropModeShared { + // Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed(). + d.touchAtime(ctx, fd.vfsfd.Mount()) + } + return n, err +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + fd.mu.Lock() + n, err := fd.PRead(ctx, dst, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + d := fd.dentry() + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + if d.fs.opts.interop != InteropModeShared { + // Compare Linux's mm/filemap.c:__generic_file_write_iter() => + // file_update_time(). This is d.touchCMtime(), but without locking + // d.metadataMu (recursively). + if now, ok := nowFromContext(ctx); ok { + atomic.StoreInt64(&d.mtime, now) + atomic.StoreInt64(&d.ctime, now) + } + } + if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + // Write dirty cached pages that will be touched by the write back to + // the remote file. + if err := d.writeback(ctx, offset, src.NumBytes()); err != nil { + return 0, err + } + // Remove touched pages from the cache. + pgstart := pageRoundDown(uint64(offset)) + pgend := pageRoundUp(uint64(offset + src.NumBytes())) + if pgend < pgstart { + return 0, syserror.EINVAL + } + mr := memmap.MappableRange{pgstart, pgend} + var freed []platform.FileRange + d.dataMu.Lock() + cseg := d.cache.LowerBoundSegment(mr.Start) + for cseg.Ok() && cseg.Start() < mr.End { + cseg = d.cache.Isolate(cseg, mr) + freed = append(freed, platform.FileRange{cseg.Value(), cseg.Value() + cseg.Range().Length()}) + cseg = d.cache.Remove(cseg).NextSegment() + } + d.dataMu.Unlock() + // Invalidate mappings of removed pages. + d.mapsMu.Lock() + d.mappings.Invalidate(mr, memmap.InvalidateOpts{}) + d.mapsMu.Unlock() + // Finally free pages removed from the cache. + mf := d.fs.mfp.MemoryFile() + for _, freedFR := range freed { + mf.DecRef(freedFR) + } + } + rw := getDentryReadWriter(ctx, d, offset) + if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + // Require the write to go to the remote file. + rw.direct = true + } + n, err := src.CopyInTo(ctx, rw) + putDentryReadWriter(rw) + if n != 0 && fd.vfsfd.StatusFlags()&(linux.O_DSYNC|linux.O_SYNC) != 0 { + // Write dirty cached pages touched by the write back to the remote + // file. + if err := d.writeback(ctx, offset, src.NumBytes()); err != nil { + return 0, err + } + // Request the remote filesystem to sync the remote file. + if err := d.handle.file.fsync(ctx); err != nil { + return 0, err + } + } + return n, err +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + fd.mu.Lock() + n, err := fd.PWrite(ctx, src, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +type dentryReadWriter struct { + ctx context.Context + d *dentry + off uint64 + direct bool +} + +var dentryReadWriterPool = sync.Pool{ + New: func() interface{} { + return &dentryReadWriter{} + }, +} + +func getDentryReadWriter(ctx context.Context, d *dentry, offset int64) *dentryReadWriter { + rw := dentryReadWriterPool.Get().(*dentryReadWriter) + rw.ctx = ctx + rw.d = d + rw.off = uint64(offset) + rw.direct = false + return rw +} + +func putDentryReadWriter(rw *dentryReadWriter) { + rw.ctx = nil + rw.d = nil + dentryReadWriterPool.Put(rw) +} + +// ReadToBlocks implements safemem.Reader.ReadToBlocks. +func (rw *dentryReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { + if dsts.IsEmpty() { + return 0, nil + } + + // If we have a mmappable host FD (which must be used here to ensure + // coherence with memory-mapped I/O), or if InteropModeShared is in effect + // (which prevents us from caching file contents and makes dentry.size + // unreliable), or if the file was opened O_DIRECT, read directly from + // dentry.handle without locking dentry.dataMu. + rw.d.handleMu.RLock() + if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct { + n, err := rw.d.handle.readToBlocksAt(rw.ctx, dsts, rw.off) + rw.d.handleMu.RUnlock() + rw.off += n + return n, err + } + + // Otherwise read from/through the cache. + mf := rw.d.fs.mfp.MemoryFile() + fillCache := mf.ShouldCacheEvictable() + var dataMuUnlock func() + if fillCache { + rw.d.dataMu.Lock() + dataMuUnlock = rw.d.dataMu.Unlock + } else { + rw.d.dataMu.RLock() + dataMuUnlock = rw.d.dataMu.RUnlock + } + + // Compute the range to read (limited by file size and overflow-checked). + if rw.off >= rw.d.size { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return 0, io.EOF + } + end := rw.d.size + if rend := rw.off + dsts.NumBytes(); rend > rw.off && rend < end { + end = rend + } + + var done uint64 + seg, gap := rw.d.cache.Find(rw.off) + for rw.off < end { + mr := memmap.MappableRange{rw.off, end} + switch { + case seg.Ok(): + // Get internal mappings from the cache. + ims, err := mf.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read) + if err != nil { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, err + } + + // Copy from internal mappings. + n, err := safemem.CopySeq(dsts, ims) + done += n + rw.off += n + dsts = dsts.DropFirst64(n) + if err != nil { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, err + } + + // Continue. + seg, gap = seg.NextNonEmpty() + + case gap.Ok(): + gapMR := gap.Range().Intersect(mr) + if fillCache { + // Read into the cache, then re-enter the loop to read from the + // cache. + reqMR := memmap.MappableRange{ + Start: pageRoundDown(gapMR.Start), + End: pageRoundUp(gapMR.End), + } + optMR := gap.Range() + err := rw.d.cache.Fill(rw.ctx, reqMR, maxFillRange(reqMR, optMR), mf, usage.PageCache, rw.d.handle.readToBlocksAt) + mf.MarkEvictable(rw.d, pgalloc.EvictableRange{optMR.Start, optMR.End}) + seg, gap = rw.d.cache.Find(rw.off) + if !seg.Ok() { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, err + } + // err might have occurred in part of gap.Range() outside + // gapMR. Forget about it for now; if the error matters and + // persists, we'll run into it again in a later iteration of + // this loop. + } else { + // Read directly from the file. + gapDsts := dsts.TakeFirst64(gapMR.Length()) + n, err := rw.d.handle.readToBlocksAt(rw.ctx, gapDsts, gapMR.Start) + done += n + rw.off += n + dsts = dsts.DropFirst64(n) + // Partial reads are fine. But we must stop reading. + if n != gapDsts.NumBytes() || err != nil { + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, err + } + + // Continue. + seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{} + } + } + } + dataMuUnlock() + rw.d.handleMu.RUnlock() + return done, nil +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +// +// Preconditions: rw.d.metadataMu must be locked. +func (rw *dentryReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + if srcs.IsEmpty() { + return 0, nil + } + + // If we have a mmappable host FD (which must be used here to ensure + // coherence with memory-mapped I/O), or if InteropModeShared is in effect + // (which prevents us from caching file contents), or if the file was + // opened with O_DIRECT, write directly to dentry.handle without locking + // dentry.dataMu. + rw.d.handleMu.RLock() + if (rw.d.handle.fd >= 0 && !rw.d.fs.opts.forcePageCache) || rw.d.fs.opts.interop == InteropModeShared || rw.direct { + n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, srcs, rw.off) + rw.d.handleMu.RUnlock() + rw.off += n + return n, err + } + + // Otherwise write to/through the cache. + mf := rw.d.fs.mfp.MemoryFile() + rw.d.dataMu.Lock() + + // Compute the range to write (overflow-checked). + start := rw.off + end := rw.off + srcs.NumBytes() + if end <= rw.off { + end = math.MaxInt64 + } + + var ( + done uint64 + retErr error + ) + seg, gap := rw.d.cache.Find(rw.off) + for rw.off < end { + mr := memmap.MappableRange{rw.off, end} + switch { + case seg.Ok(): + // Get internal mappings from the cache. + segMR := seg.Range().Intersect(mr) + ims, err := mf.MapInternal(seg.FileRangeOf(segMR), usermem.Write) + if err != nil { + retErr = err + goto exitLoop + } + + // Copy to internal mappings. + n, err := safemem.CopySeq(ims, srcs) + done += n + rw.off += n + srcs = srcs.DropFirst64(n) + rw.d.dirty.MarkDirty(segMR) + if err != nil { + retErr = err + goto exitLoop + } + + // Continue. + seg, gap = seg.NextNonEmpty() + + case gap.Ok(): + // Write directly to the file. At present, we never fill the cache + // when writing, since doing so can convert small writes into + // inefficient read-modify-write cycles, and we have no mechanism + // for detecting or avoiding this. + gapMR := gap.Range().Intersect(mr) + gapSrcs := srcs.TakeFirst64(gapMR.Length()) + n, err := rw.d.handle.writeFromBlocksAt(rw.ctx, gapSrcs, gapMR.Start) + done += n + rw.off += n + srcs = srcs.DropFirst64(n) + // Partial writes are fine. But we must stop writing. + if n != gapSrcs.NumBytes() || err != nil { + retErr = err + goto exitLoop + } + + // Continue. + seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{} + } + } +exitLoop: + if rw.off > rw.d.size { + atomic.StoreUint64(&rw.d.size, rw.off) + // The remote file's size will implicitly be extended to the correct + // value when we write back to it. + } + // If InteropModeWritethrough is in effect, flush written data back to the + // remote filesystem. + if rw.d.fs.opts.interop == InteropModeWritethrough && done != 0 { + if err := fsutil.SyncDirty(rw.ctx, memmap.MappableRange{ + Start: start, + End: rw.off, + }, &rw.d.cache, &rw.d.dirty, rw.d.size, mf, rw.d.handle.writeFromBlocksAt); err != nil { + // We have no idea how many bytes were actually flushed. + rw.off = start + done = 0 + retErr = err + } + } + rw.d.dataMu.Unlock() + rw.d.handleMu.RUnlock() + return done, retErr +} + +func (d *dentry) writeback(ctx context.Context, offset, size int64) error { + if size == 0 { + return nil + } + d.handleMu.RLock() + defer d.handleMu.RUnlock() + d.dataMu.Lock() + defer d.dataMu.Unlock() + // Compute the range of valid bytes (overflow-checked). + if uint64(offset) >= d.size { + return nil + } + end := int64(d.size) + if rend := offset + size; rend > offset && rend < end { + end = rend + } + return fsutil.SyncDirty(ctx, memmap.MappableRange{ + Start: uint64(offset), + End: uint64(end), + }, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt) +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *regularFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fd.mu.Lock() + defer fd.mu.Unlock() + switch whence { + case linux.SEEK_SET: + // Use offset as specified. + case linux.SEEK_CUR: + offset += fd.off + case linux.SEEK_END, linux.SEEK_DATA, linux.SEEK_HOLE: + // Ensure file size is up to date. + d := fd.dentry() + if fd.filesystem().opts.interop == InteropModeShared { + if err := d.updateFromGetattr(ctx); err != nil { + return 0, err + } + } + size := int64(atomic.LoadUint64(&d.size)) + // For SEEK_DATA and SEEK_HOLE, treat the file as a single contiguous + // block of data. + switch whence { + case linux.SEEK_END: + offset += size + case linux.SEEK_DATA: + if offset > size { + return 0, syserror.ENXIO + } + // Use offset as specified. + case linux.SEEK_HOLE: + if offset > size { + return 0, syserror.ENXIO + } + offset = size + } + default: + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + fd.off = offset + return offset, nil +} + +// Sync implements vfs.FileDescriptionImpl.Sync. +func (fd *regularFileFD) Sync(ctx context.Context) error { + return fd.dentry().syncSharedHandle(ctx) +} + +func (d *dentry) syncSharedHandle(ctx context.Context) error { + d.handleMu.RLock() + if !d.handleWritable { + d.handleMu.RUnlock() + return nil + } + d.dataMu.Lock() + // Write dirty cached data to the remote file. + err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), d.handle.writeFromBlocksAt) + d.dataMu.Unlock() + if err == nil { + // Sync the remote file. + err = d.handle.sync(ctx) + } + d.handleMu.RUnlock() + return err +} + +// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. +func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + d := fd.dentry() + switch d.fs.opts.interop { + case InteropModeExclusive: + // Any mapping is fine. + case InteropModeWritethrough: + // Shared writable mappings require a host FD, since otherwise we can't + // synchronously flush memory-mapped writes to the remote file. + if opts.Private || !opts.MaxPerms.Write { + break + } + fallthrough + case InteropModeShared: + // All mappings require a host FD to be coherent with other filesystem + // users. + if d.fs.opts.forcePageCache { + // Whether or not we have a host FD, we're not allowed to use it. + return syserror.ENODEV + } + d.handleMu.RLock() + haveFD := d.handle.fd >= 0 + d.handleMu.RUnlock() + if !haveFD { + return syserror.ENODEV + } + default: + panic(fmt.Sprintf("unknown InteropMode %v", d.fs.opts.interop)) + } + return vfs.GenericConfigureMMap(&fd.vfsfd, d, opts) +} + +func (d *dentry) mayCachePages() bool { + if d.fs.opts.interop == InteropModeShared { + return false + } + if d.fs.opts.forcePageCache { + return true + } + d.handleMu.RLock() + haveFD := d.handle.fd >= 0 + d.handleMu.RUnlock() + return haveFD +} + +// AddMapping implements memmap.Mappable.AddMapping. +func (d *dentry) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error { + d.mapsMu.Lock() + mapped := d.mappings.AddMapping(ms, ar, offset, writable) + // Do this unconditionally since whether we have a host FD can change + // across save/restore. + for _, r := range mapped { + d.pf.hostFileMapper.IncRefOn(r) + } + if d.mayCachePages() { + // d.Evict() will refuse to evict memory-mapped pages, so tell the + // MemoryFile to not bother trying. + mf := d.fs.mfp.MemoryFile() + for _, r := range mapped { + mf.MarkUnevictable(d, pgalloc.EvictableRange{r.Start, r.End}) + } + } + d.mapsMu.Unlock() + return nil +} + +// RemoveMapping implements memmap.Mappable.RemoveMapping. +func (d *dentry) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) { + d.mapsMu.Lock() + unmapped := d.mappings.RemoveMapping(ms, ar, offset, writable) + for _, r := range unmapped { + d.pf.hostFileMapper.DecRefOn(r) + } + if d.mayCachePages() { + // Pages that are no longer referenced by any application memory + // mappings are now considered unused; allow MemoryFile to evict them + // when necessary. + mf := d.fs.mfp.MemoryFile() + d.dataMu.Lock() + for _, r := range unmapped { + // Since these pages are no longer mapped, they are no longer + // concurrently dirtyable by a writable memory mapping. + d.dirty.AllowClean(r) + mf.MarkEvictable(d, pgalloc.EvictableRange{r.Start, r.End}) + } + d.dataMu.Unlock() + } + d.mapsMu.Unlock() +} + +// CopyMapping implements memmap.Mappable.CopyMapping. +func (d *dentry) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error { + return d.AddMapping(ctx, ms, dstAR, offset, writable) +} + +// Translate implements memmap.Mappable.Translate. +func (d *dentry) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) { + d.handleMu.RLock() + if d.handle.fd >= 0 && !d.fs.opts.forcePageCache { + d.handleMu.RUnlock() + mr := optional + if d.fs.opts.limitHostFDTranslation { + mr = maxFillRange(required, optional) + } + return []memmap.Translation{ + { + Source: mr, + File: &d.pf, + Offset: mr.Start, + Perms: usermem.AnyAccess, + }, + }, nil + } + + d.dataMu.Lock() + + // Constrain translations to d.size (rounded up) to prevent translation to + // pages that may be concurrently truncated. + pgend := pageRoundUp(d.size) + var beyondEOF bool + if required.End > pgend { + if required.Start >= pgend { + d.dataMu.Unlock() + d.handleMu.RUnlock() + return nil, &memmap.BusError{io.EOF} + } + beyondEOF = true + required.End = pgend + } + if optional.End > pgend { + optional.End = pgend + } + + mf := d.fs.mfp.MemoryFile() + cerr := d.cache.Fill(ctx, required, maxFillRange(required, optional), mf, usage.PageCache, d.handle.readToBlocksAt) + + var ts []memmap.Translation + var translatedEnd uint64 + for seg := d.cache.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() { + segMR := seg.Range().Intersect(optional) + // TODO(jamieliu): Make Translations writable even if writability is + // not required if already kept-dirty by another writable translation. + perms := usermem.AccessType{ + Read: true, + Execute: true, + } + if at.Write { + // From this point forward, this memory can be dirtied through the + // mapping at any time. + d.dirty.KeepDirty(segMR) + perms.Write = true + } + ts = append(ts, memmap.Translation{ + Source: segMR, + File: mf, + Offset: seg.FileRangeOf(segMR).Start, + Perms: perms, + }) + translatedEnd = segMR.End + } + + d.dataMu.Unlock() + d.handleMu.RUnlock() + + // Don't return the error returned by c.cache.Fill if it occurred outside + // of required. + if translatedEnd < required.End && cerr != nil { + return ts, &memmap.BusError{cerr} + } + if beyondEOF { + return ts, &memmap.BusError{io.EOF} + } + return ts, nil +} + +func maxFillRange(required, optional memmap.MappableRange) memmap.MappableRange { + const maxReadahead = 64 << 10 // 64 KB, chosen arbitrarily + if required.Length() >= maxReadahead { + return required + } + if optional.Length() <= maxReadahead { + return optional + } + optional.Start = required.Start + if optional.Length() <= maxReadahead { + return optional + } + optional.End = optional.Start + maxReadahead + return optional +} + +// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable. +func (d *dentry) InvalidateUnsavable(ctx context.Context) error { + // Whether we have a host fd (and consequently what platform.File is + // mapped) can change across save/restore, so invalidate all translations + // unconditionally. + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + d.mappings.InvalidateAll(memmap.InvalidateOpts{}) + + // Write the cache's contents back to the remote file so that if we have a + // host fd after restore, the remote file's contents are coherent. + mf := d.fs.mfp.MemoryFile() + d.dataMu.Lock() + defer d.dataMu.Unlock() + if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil { + return err + } + + // Discard the cache so that it's not stored in saved state. This is safe + // because per InvalidateUnsavable invariants, no new translations can have + // been returned after we invalidated all existing translations above. + d.cache.DropAll(mf) + d.dirty.RemoveAll() + + return nil +} + +// Evict implements pgalloc.EvictableMemoryUser.Evict. +func (d *dentry) Evict(ctx context.Context, er pgalloc.EvictableRange) { + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + d.dataMu.Lock() + defer d.dataMu.Unlock() + + mr := memmap.MappableRange{er.Start, er.End} + mf := d.fs.mfp.MemoryFile() + // Only allow pages that are no longer memory-mapped to be evicted. + for mgap := d.mappings.LowerBoundGap(mr.Start); mgap.Ok() && mgap.Start() < mr.End; mgap = mgap.NextGap() { + mgapMR := mgap.Range().Intersect(mr) + if mgapMR.Length() == 0 { + continue + } + if err := fsutil.SyncDirty(ctx, mgapMR, &d.cache, &d.dirty, d.size, mf, d.handle.writeFromBlocksAt); err != nil { + log.Warningf("Failed to writeback cached data %v: %v", mgapMR, err) + } + d.cache.Drop(mgapMR, mf) + d.dirty.KeepClean(mgapMR) + } +} + +// dentryPlatformFile implements platform.File. It exists solely because dentry +// cannot implement both vfs.DentryImpl.IncRef and platform.File.IncRef. +// +// dentryPlatformFile is only used when a host FD representing the remote file +// is available (i.e. dentry.handle.fd >= 0), and that FD is used for +// application memory mappings (i.e. !filesystem.opts.forcePageCache). +type dentryPlatformFile struct { + *dentry + + // fdRefs counts references on platform.File offsets. fdRefs is protected + // by dentry.dataMu. + fdRefs fsutil.FrameRefSet + + // If this dentry represents a regular file, and handle.fd >= 0, + // hostFileMapper caches mappings of handle.fd. + hostFileMapper fsutil.HostFileMapper +} + +// IncRef implements platform.File.IncRef. +func (d *dentryPlatformFile) IncRef(fr platform.FileRange) { + d.dataMu.Lock() + seg, gap := d.fdRefs.Find(fr.Start) + for { + switch { + case seg.Ok() && seg.Start() < fr.End: + seg = d.fdRefs.Isolate(seg, fr) + seg.SetValue(seg.Value() + 1) + seg, gap = seg.NextNonEmpty() + case gap.Ok() && gap.Start() < fr.End: + newRange := gap.Range().Intersect(fr) + usage.MemoryAccounting.Inc(newRange.Length(), usage.Mapped) + seg, gap = d.fdRefs.InsertWithoutMerging(gap, newRange, 1).NextNonEmpty() + default: + d.fdRefs.MergeAdjacent(fr) + d.dataMu.Unlock() + return + } + } +} + +// DecRef implements platform.File.DecRef. +func (d *dentryPlatformFile) DecRef(fr platform.FileRange) { + d.dataMu.Lock() + seg := d.fdRefs.FindSegment(fr.Start) + + for seg.Ok() && seg.Start() < fr.End { + seg = d.fdRefs.Isolate(seg, fr) + if old := seg.Value(); old == 1 { + usage.MemoryAccounting.Dec(seg.Range().Length(), usage.Mapped) + seg = d.fdRefs.Remove(seg).NextSegment() + } else { + seg.SetValue(old - 1) + seg = seg.NextSegment() + } + } + d.fdRefs.MergeAdjacent(fr) + d.dataMu.Unlock() + +} + +// MapInternal implements platform.File.MapInternal. +func (d *dentryPlatformFile) MapInternal(fr platform.FileRange, at usermem.AccessType) (safemem.BlockSeq, error) { + d.handleMu.RLock() + bs, err := d.hostFileMapper.MapInternal(fr, int(d.handle.fd), at.Write) + d.handleMu.RUnlock() + return bs, err +} + +// FD implements platform.File.FD. +func (d *dentryPlatformFile) FD() int { + d.handleMu.RLock() + fd := d.handle.fd + d.handleMu.RUnlock() + return int(fd) +} diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go new file mode 100644 index 000000000..08c691c47 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/special_file.go @@ -0,0 +1,159 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// specialFileFD implements vfs.FileDescriptionImpl for files other than +// regular files, directories, and symlinks: pipes, sockets, etc. It is also +// used for regular files when filesystemOptions.specialRegularFiles is in +// effect. specialFileFD differs from regularFileFD by using per-FD handles +// instead of shared per-dentry handles, and never buffering I/O. +type specialFileFD struct { + fileDescription + + // handle is immutable. + handle handle + + // off is the file offset. off is protected by mu. (POSIX 2.9.7 only + // requires operations using the file offset to be atomic for regular files + // and symlinks; however, since specialFileFD may be used for regular + // files, we apply this atomicity unconditionally.) + mu sync.Mutex + off int64 +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *specialFileFD) Release() { + fd.handle.close(context.Background()) + fs := fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) + fs.syncMu.Lock() + delete(fs.specialFileFDs, fd) + fs.syncMu.Unlock() +} + +// OnClose implements vfs.FileDescriptionImpl.OnClose. +func (fd *specialFileFD) OnClose(ctx context.Context) error { + if !fd.vfsfd.IsWritable() { + return nil + } + return fd.handle.file.flush(ctx) +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + // Going through dst.CopyOutFrom() holds MM locks around file operations of + // unknown duration. For regularFileFD, doing so is necessary to support + // mmap due to lock ordering; MM locks precede dentry.dataMu. That doesn't + // hold here since specialFileFD doesn't client-cache data. Just buffer the + // read instead. + if d := fd.dentry(); d.fs.opts.interop != InteropModeShared { + d.touchAtime(ctx, fd.vfsfd.Mount()) + } + buf := make([]byte, dst.NumBytes()) + n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) + if n == 0 { + return 0, err + } + if cp, cperr := dst.CopyOut(ctx, buf[:n]); cperr != nil { + return int64(cp), cperr + } + return int64(n), err +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *specialFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + fd.mu.Lock() + n, err := fd.PRead(ctx, dst, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *specialFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + if opts.Flags != 0 { + return 0, syserror.EOPNOTSUPP + } + + // Do a buffered write. See rationale in PRead. + if d := fd.dentry(); d.fs.opts.interop != InteropModeShared { + d.touchCMtime(ctx) + } + buf := make([]byte, src.NumBytes()) + // Don't do partial writes if we get a partial read from src. + if _, err := src.CopyIn(ctx, buf); err != nil { + return 0, err + } + n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) + return int64(n), err +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *specialFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + fd.mu.Lock() + n, err := fd.PWrite(ctx, src, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *specialFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fd.mu.Lock() + defer fd.mu.Unlock() + switch whence { + case linux.SEEK_SET: + // Use offset as given. + case linux.SEEK_CUR: + offset += fd.off + default: + // SEEK_END, SEEK_DATA, and SEEK_HOLE aren't supported since it's not + // clear that file size is even meaningful for these files. + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + fd.off = offset + return offset, nil +} + +// Sync implements vfs.FileDescriptionImpl.Sync. +func (fd *specialFileFD) Sync(ctx context.Context) error { + if !fd.vfsfd.IsWritable() { + return nil + } + return fd.handle.sync(ctx) +} diff --git a/pkg/sentry/fsimpl/gofer/symlink.go b/pkg/sentry/fsimpl/gofer/symlink.go new file mode 100644 index 000000000..adf43be60 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/symlink.go @@ -0,0 +1,47 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +func (d *dentry) isSymlink() bool { + return d.fileType() == linux.S_IFLNK +} + +// Precondition: d.isSymlink(). +func (d *dentry) readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { + if d.fs.opts.interop != InteropModeShared { + d.touchAtime(ctx, mnt) + d.dataMu.Lock() + if d.haveTarget { + target := d.target + d.dataMu.Unlock() + return target, nil + } + } + target, err := d.file.readlink(ctx) + if d.fs.opts.interop != InteropModeShared { + if err == nil { + d.haveTarget = true + d.target = target + } + d.dataMu.Unlock() + } + return target, err +} diff --git a/pkg/sentry/fsimpl/gofer/time.go b/pkg/sentry/fsimpl/gofer/time.go new file mode 100644 index 000000000..7598ec6a8 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/time.go @@ -0,0 +1,75 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package gofer + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +func dentryTimestampFromP9(s, ns uint64) int64 { + return int64(s*1e9 + ns) +} + +func dentryTimestampFromStatx(ts linux.StatxTimestamp) int64 { + return ts.Sec*1e9 + int64(ts.Nsec) +} + +func statxTimestampFromDentry(ns int64) linux.StatxTimestamp { + return linux.StatxTimestamp{ + Sec: ns / 1e9, + Nsec: uint32(ns % 1e9), + } +} + +func nowFromContext(ctx context.Context) (int64, bool) { + if clock := ktime.RealtimeClockFromContext(ctx); clock != nil { + return clock.Now().Nanoseconds(), true + } + return 0, false +} + +// Preconditions: fs.interop != InteropModeShared. +func (d *dentry) touchAtime(ctx context.Context, mnt *vfs.Mount) { + if err := mnt.CheckBeginWrite(); err != nil { + return + } + now, ok := nowFromContext(ctx) + if !ok { + mnt.EndWrite() + return + } + d.metadataMu.Lock() + atomic.StoreInt64(&d.atime, now) + d.metadataMu.Unlock() + mnt.EndWrite() +} + +// Preconditions: fs.interop != InteropModeShared. The caller has successfully +// called vfs.Mount.CheckBeginWrite(). +func (d *dentry) touchCMtime(ctx context.Context) { + now, ok := nowFromContext(ctx) + if !ok { + return + } + d.metadataMu.Lock() + atomic.StoreInt64(&d.mtime, now) + atomic.StoreInt64(&d.ctime, now) + d.metadataMu.Unlock() +} diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index 5ee9cf1e9..72bc15264 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -622,7 +622,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error if child.inode.isDir() { return syserror.EISDIR } - if !rp.MustBeDir() { + if rp.MustBeDir() { return syserror.ENOTDIR } mnt := rp.Mount() diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index bde4c7a1e..34f63986f 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -126,7 +126,7 @@ func (s *socketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS } return uint64(n), nil } - return readv(s.fd, iovecsFromBlockSeq(dsts)) + return readv(s.fd, safemem.IovecsFromBlockSeq(dsts)) })) return int64(n), err } @@ -149,7 +149,7 @@ func (s *socketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IO } return uint64(n), nil } - return writev(s.fd, iovecsFromBlockSeq(srcs)) + return writev(s.fd, safemem.IovecsFromBlockSeq(srcs)) })) return int64(n), err } @@ -402,7 +402,7 @@ func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags // We always do a non-blocking recv*(). sysflags := flags | syscall.MSG_DONTWAIT - iovs := iovecsFromBlockSeq(dsts) + iovs := safemem.IovecsFromBlockSeq(dsts) msg := syscall.Msghdr{ Iov: &iovs[0], Iovlen: uint64(len(iovs)), @@ -522,7 +522,7 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] return uint64(n), nil } - iovs := iovecsFromBlockSeq(srcs) + iovs := safemem.IovecsFromBlockSeq(srcs) msg := syscall.Msghdr{ Iov: &iovs[0], Iovlen: uint64(len(iovs)), @@ -567,21 +567,6 @@ func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to [] return int(n), syserr.FromError(err) } -func iovecsFromBlockSeq(bs safemem.BlockSeq) []syscall.Iovec { - iovs := make([]syscall.Iovec, 0, bs.NumBlocks()) - for ; !bs.IsEmpty(); bs = bs.Tail() { - b := bs.Head() - iovs = append(iovs, syscall.Iovec{ - Base: &b.ToSlice()[0], - Len: uint64(b.Len()), - }) - // We don't need to care about b.NeedSafecopy(), because the host - // kernel will handle such address ranges just fine (by returning - // EFAULT). - } - return iovs -} - func translateIOSyscallError(err error) error { if err == syscall.EAGAIN || err == syscall.EWOULDBLOCK { return syserror.ErrWouldBlock -- cgit v1.2.3 From 1b6a12a768216a99a5e0428c42ea4faf79cf3b50 Mon Sep 17 00:00:00 2001 From: Adin Scannell Date: Wed, 5 Feb 2020 22:45:44 -0800 Subject: Add notes to relevant tests. These were out-of-band notes that can help provide additional context and simplify automated imports. PiperOrigin-RevId: 293525915 --- pkg/metric/metric.go | 1 - pkg/sentry/arch/arch_x86.go | 4 ++ pkg/sentry/arch/signal_amd64.go | 2 +- pkg/sentry/fs/file_overlay_test.go | 1 + pkg/sentry/fs/proc/README.md | 4 ++ pkg/sentry/kernel/BUILD | 1 + pkg/sentry/kernel/kernel.go | 3 ++ pkg/sentry/kernel/kernel_opts.go | 20 +++++++ pkg/sentry/socket/hostinet/BUILD | 1 + pkg/sentry/socket/hostinet/socket.go | 5 +- pkg/sentry/socket/hostinet/sockopt_impl.go | 27 ++++++++++ pkg/tcpip/transport/tcp/endpoint.go | 3 ++ runsc/boot/filter/BUILD | 1 + runsc/boot/filter/config.go | 13 ----- runsc/boot/filter/config_profile.go | 34 ++++++++++++ runsc/container/console_test.go | 5 +- runsc/dockerutil/dockerutil.go | 11 ++-- runsc/testutil/BUILD | 5 +- runsc/testutil/testutil.go | 54 ------------------- runsc/testutil/testutil_runfiles.go | 75 +++++++++++++++++++++++++++ test/image/image_test.go | 8 +-- test/syscalls/build_defs.bzl | 35 +++++++++++-- test/syscalls/linux/chroot.cc | 2 +- test/syscalls/linux/concurrency.cc | 3 +- test/syscalls/linux/exec_proc_exe_workload.cc | 6 +++ test/syscalls/linux/fork.cc | 5 +- test/syscalls/linux/mmap.cc | 8 +-- test/syscalls/linux/open_create.cc | 1 + test/syscalls/linux/preadv.cc | 1 + test/syscalls/linux/proc.cc | 46 +++++++++++++--- test/syscalls/linux/readv.cc | 4 +- test/syscalls/linux/rseq.cc | 2 +- test/syscalls/linux/select.cc | 2 +- test/syscalls/linux/shm.cc | 2 +- test/syscalls/linux/sigprocmask.cc | 2 +- test/syscalls/linux/socket_unix_non_stream.cc | 4 +- test/syscalls/linux/symlink.cc | 2 +- test/syscalls/linux/tcp_socket.cc | 3 +- test/syscalls/linux/time.cc | 1 + test/syscalls/linux/tkill.cc | 2 +- test/util/temp_path.cc | 1 + tools/build/tags.bzl | 4 ++ tools/defs.bzl | 17 +++++- 43 files changed, 318 insertions(+), 113 deletions(-) create mode 100644 pkg/sentry/kernel/kernel_opts.go create mode 100644 pkg/sentry/socket/hostinet/sockopt_impl.go create mode 100644 runsc/boot/filter/config_profile.go create mode 100644 runsc/testutil/testutil_runfiles.go (limited to 'pkg/sentry/fs') diff --git a/pkg/metric/metric.go b/pkg/metric/metric.go index 93d4f2b8c..006fcd9ab 100644 --- a/pkg/metric/metric.go +++ b/pkg/metric/metric.go @@ -46,7 +46,6 @@ var ( // // TODO(b/67298402): Support non-cumulative metrics. // TODO(b/67298427): Support metric fields. -// type Uint64Metric struct { // value is the actual value of the metric. It must be accessed // atomically. diff --git a/pkg/sentry/arch/arch_x86.go b/pkg/sentry/arch/arch_x86.go index a18093155..3db8bd34b 100644 --- a/pkg/sentry/arch/arch_x86.go +++ b/pkg/sentry/arch/arch_x86.go @@ -114,6 +114,10 @@ func newX86FPStateSlice() []byte { size, align := cpuid.HostFeatureSet().ExtendedStateSize() capacity := size // Always use at least 4096 bytes. + // + // For the KVM platform, this state is a fixed 4096 bytes, so make sure + // that the underlying array is at _least_ that size otherwise we will + // corrupt random memory. This is not a pleasant thing to debug. if capacity < 4096 { capacity = 4096 } diff --git a/pkg/sentry/arch/signal_amd64.go b/pkg/sentry/arch/signal_amd64.go index 81b92bb43..6fb756f0e 100644 --- a/pkg/sentry/arch/signal_amd64.go +++ b/pkg/sentry/arch/signal_amd64.go @@ -55,7 +55,7 @@ type SignalContext64 struct { Trapno uint64 Oldmask linux.SignalSet Cr2 uint64 - // Pointer to a struct _fpstate. + // Pointer to a struct _fpstate. See b/33003106#comment8. Fpstate uint64 Reserved [8]uint64 } diff --git a/pkg/sentry/fs/file_overlay_test.go b/pkg/sentry/fs/file_overlay_test.go index 02538bb4f..a76d87e3a 100644 --- a/pkg/sentry/fs/file_overlay_test.go +++ b/pkg/sentry/fs/file_overlay_test.go @@ -177,6 +177,7 @@ func TestReaddirRevalidation(t *testing.T) { // TestReaddirOverlayFrozen tests that calling Readdir on an overlay file with // a frozen dirent tree does not make Readdir calls to the underlying files. +// This is a regression test for b/114808269. func TestReaddirOverlayFrozen(t *testing.T) { ctx := contexttest.Context(t) diff --git a/pkg/sentry/fs/proc/README.md b/pkg/sentry/fs/proc/README.md index 5d4ec6c7b..6667a0916 100644 --- a/pkg/sentry/fs/proc/README.md +++ b/pkg/sentry/fs/proc/README.md @@ -11,6 +11,8 @@ inconsistency, please file a bug. The following files are implemented: + + | File /proc/ | Content | | :------------------------ | :---------------------------------------------------- | | [cpuinfo](#cpuinfo) | Info about the CPU | @@ -22,6 +24,8 @@ The following files are implemented: | [uptime](#uptime) | Wall clock since boot, combined idle time of all cpus | | [version](#version) | Kernel version | + + ### cpuinfo ```bash diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index a27628c0a..2231d6973 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -91,6 +91,7 @@ go_library( "fs_context.go", "ipc_namespace.go", "kernel.go", + "kernel_opts.go", "kernel_state.go", "pending_signals.go", "pending_signals_list.go", diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index dcd6e91c4..3ee760ba2 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -235,6 +235,9 @@ type Kernel struct { // events. This is initialized lazily on the first unimplemented // syscall. unimplementedSyscallEmitter eventchannel.Emitter `state:"nosave"` + + // SpecialOpts contains special kernel options. + SpecialOpts } // InitKernelArgs holds arguments to Init. diff --git a/pkg/sentry/kernel/kernel_opts.go b/pkg/sentry/kernel/kernel_opts.go new file mode 100644 index 000000000..2e66ec587 --- /dev/null +++ b/pkg/sentry/kernel/kernel_opts.go @@ -0,0 +1,20 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kernel + +// SpecialOpts contains non-standard options for the kernel. +// +// +stateify savable +type SpecialOpts struct{} diff --git a/pkg/sentry/socket/hostinet/BUILD b/pkg/sentry/socket/hostinet/BUILD index 5a07d5d0e..023bad156 100644 --- a/pkg/sentry/socket/hostinet/BUILD +++ b/pkg/sentry/socket/hostinet/BUILD @@ -10,6 +10,7 @@ go_library( "save_restore.go", "socket.go", "socket_unsafe.go", + "sockopt_impl.go", "stack.go", ], visibility = ["//pkg/sentry:internal"], diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index 34f63986f..de76388ac 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -285,7 +285,7 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt } // Whitelist options and constrain option length. - var optlen int + optlen := getSockOptLen(t, level, name) switch level { case linux.SOL_IP: switch name { @@ -330,7 +330,7 @@ func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outPt // SetSockOpt implements socket.Socket.SetSockOpt. func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error { // Whitelist options and constrain option length. - var optlen int + optlen := setSockOptLen(t, level, name) switch level { case linux.SOL_IP: switch name { @@ -353,6 +353,7 @@ func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt [ optlen = sizeofInt32 } } + if optlen == 0 { // Pretend to accept socket options we don't understand. This seems // dangerous, but it's what netstack does... diff --git a/pkg/sentry/socket/hostinet/sockopt_impl.go b/pkg/sentry/socket/hostinet/sockopt_impl.go new file mode 100644 index 000000000..8a783712e --- /dev/null +++ b/pkg/sentry/socket/hostinet/sockopt_impl.go @@ -0,0 +1,27 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package hostinet + +import ( + "gvisor.dev/gvisor/pkg/sentry/kernel" +) + +func getSockOptLen(t *kernel.Task, level, name int) int { + return 0 // No custom options. +} + +func setSockOptLen(t *kernel.Task, level, name int) int { + return 0 // No custom options. +} diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index e4a6b1b8b..f2be0e651 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -2166,6 +2166,9 @@ func (e *endpoint) listen(backlog int) *tcpip.Error { e.isRegistered = true e.setEndpointState(StateListen) + // The channel may be non-nil when we're restoring the endpoint, and it + // may be pre-populated with some previously accepted (but not Accepted) + // endpoints. if e.acceptedChan == nil { e.acceptedChan = make(chan *endpoint, backlog) } diff --git a/runsc/boot/filter/BUILD b/runsc/boot/filter/BUILD index ce30f6c53..ed18f0047 100644 --- a/runsc/boot/filter/BUILD +++ b/runsc/boot/filter/BUILD @@ -8,6 +8,7 @@ go_library( "config.go", "config_amd64.go", "config_arm64.go", + "config_profile.go", "extra_filters.go", "extra_filters_msan.go", "extra_filters_race.go", diff --git a/runsc/boot/filter/config.go b/runsc/boot/filter/config.go index f8d351c7b..c69f4c602 100644 --- a/runsc/boot/filter/config.go +++ b/runsc/boot/filter/config.go @@ -536,16 +536,3 @@ func controlServerFilters(fd int) seccomp.SyscallRules { }, } } - -// profileFilters returns extra syscalls made by runtime/pprof package. -func profileFilters() seccomp.SyscallRules { - return seccomp.SyscallRules{ - syscall.SYS_OPENAT: []seccomp.Rule{ - { - seccomp.AllowAny{}, - seccomp.AllowAny{}, - seccomp.AllowValue(syscall.O_RDONLY | syscall.O_LARGEFILE | syscall.O_CLOEXEC), - }, - }, - } -} diff --git a/runsc/boot/filter/config_profile.go b/runsc/boot/filter/config_profile.go new file mode 100644 index 000000000..194952a7b --- /dev/null +++ b/runsc/boot/filter/config_profile.go @@ -0,0 +1,34 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package filter + +import ( + "syscall" + + "gvisor.dev/gvisor/pkg/seccomp" +) + +// profileFilters returns extra syscalls made by runtime/pprof package. +func profileFilters() seccomp.SyscallRules { + return seccomp.SyscallRules{ + syscall.SYS_OPENAT: []seccomp.Rule{ + { + seccomp.AllowAny{}, + seccomp.AllowAny{}, + seccomp.AllowValue(syscall.O_RDONLY | syscall.O_LARGEFILE | syscall.O_CLOEXEC), + }, + }, + } +} diff --git a/runsc/container/console_test.go b/runsc/container/console_test.go index 060b63bf3..c2518d52b 100644 --- a/runsc/container/console_test.go +++ b/runsc/container/console_test.go @@ -196,7 +196,10 @@ func TestJobControlSignalExec(t *testing.T) { defer ptyMaster.Close() defer ptySlave.Close() - // Exec bash and attach a terminal. + // Exec bash and attach a terminal. Note that occasionally /bin/sh + // may be a different shell or have a different configuration (such + // as disabling interactive mode and job control). Since we want to + // explicitly test interactive mode, use /bin/bash. See b/116981926. execArgs := &control.ExecArgs{ Filename: "/bin/bash", // Don't let bash execute from profile or rc files, otherwise diff --git a/runsc/dockerutil/dockerutil.go b/runsc/dockerutil/dockerutil.go index 9b6346ca2..1ff5e8cc3 100644 --- a/runsc/dockerutil/dockerutil.go +++ b/runsc/dockerutil/dockerutil.go @@ -143,8 +143,11 @@ func PrepareFiles(names ...string) (string, error) { return "", fmt.Errorf("os.Chmod(%q, 0777) failed: %v", dir, err) } for _, name := range names { - src := getLocalPath(name) - dst := path.Join(dir, name) + src, err := testutil.FindFile(name) + if err != nil { + return "", fmt.Errorf("testutil.Preparefiles(%q) failed: %v", name, err) + } + dst := path.Join(dir, path.Base(name)) if err := testutil.Copy(src, dst); err != nil { return "", fmt.Errorf("testutil.Copy(%q, %q) failed: %v", src, dst, err) } @@ -152,10 +155,6 @@ func PrepareFiles(names ...string) (string, error) { return dir, nil } -func getLocalPath(file string) string { - return path.Join(".", file) -} - // do executes docker command. func do(args ...string) (string, error) { log.Printf("Running: docker %s\n", args) diff --git a/runsc/testutil/BUILD b/runsc/testutil/BUILD index f845120b0..945405303 100644 --- a/runsc/testutil/BUILD +++ b/runsc/testutil/BUILD @@ -5,7 +5,10 @@ package(licenses = ["notice"]) go_library( name = "testutil", testonly = 1, - srcs = ["testutil.go"], + srcs = [ + "testutil.go", + "testutil_runfiles.go", + ], visibility = ["//:sandbox"], deps = [ "//pkg/log", diff --git a/runsc/testutil/testutil.go b/runsc/testutil/testutil.go index edf2e809a..80c2c9680 100644 --- a/runsc/testutil/testutil.go +++ b/runsc/testutil/testutil.go @@ -79,60 +79,6 @@ func ConfigureExePath() error { return nil } -// FindFile searchs for a file inside the test run environment. It returns the -// full path to the file. It fails if none or more than one file is found. -func FindFile(path string) (string, error) { - wd, err := os.Getwd() - if err != nil { - return "", err - } - - // The test root is demarcated by a path element called "__main__". Search for - // it backwards from the working directory. - root := wd - for { - dir, name := filepath.Split(root) - if name == "__main__" { - break - } - if len(dir) == 0 { - return "", fmt.Errorf("directory __main__ not found in %q", wd) - } - // Remove ending slash to loop around. - root = dir[:len(dir)-1] - } - - // Annoyingly, bazel adds the build type to the directory path for go - // binaries, but not for c++ binaries. We use two different patterns to - // to find our file. - patterns := []string{ - // Try the obvious path first. - filepath.Join(root, path), - // If it was a go binary, use a wildcard to match the build - // type. The pattern is: /test-path/__main__/directories/*/file. - filepath.Join(root, filepath.Dir(path), "*", filepath.Base(path)), - } - - for _, p := range patterns { - matches, err := filepath.Glob(p) - if err != nil { - // "The only possible returned error is ErrBadPattern, - // when pattern is malformed." -godoc - return "", fmt.Errorf("error globbing %q: %v", p, err) - } - switch len(matches) { - case 0: - // Try the next pattern. - case 1: - // We found it. - return matches[0], nil - default: - return "", fmt.Errorf("more than one match found for %q: %s", path, matches) - } - } - return "", fmt.Errorf("file %q not found", path) -} - // TestConfig returns the default configuration to use in tests. Note that // 'RootDir' must be set by caller if required. func TestConfig() *boot.Config { diff --git a/runsc/testutil/testutil_runfiles.go b/runsc/testutil/testutil_runfiles.go new file mode 100644 index 000000000..ece9ea9a1 --- /dev/null +++ b/runsc/testutil/testutil_runfiles.go @@ -0,0 +1,75 @@ +// Copyright 2018 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package testutil + +import ( + "fmt" + "os" + "path/filepath" +) + +// FindFile searchs for a file inside the test run environment. It returns the +// full path to the file. It fails if none or more than one file is found. +func FindFile(path string) (string, error) { + wd, err := os.Getwd() + if err != nil { + return "", err + } + + // The test root is demarcated by a path element called "__main__". Search for + // it backwards from the working directory. + root := wd + for { + dir, name := filepath.Split(root) + if name == "__main__" { + break + } + if len(dir) == 0 { + return "", fmt.Errorf("directory __main__ not found in %q", wd) + } + // Remove ending slash to loop around. + root = dir[:len(dir)-1] + } + + // Annoyingly, bazel adds the build type to the directory path for go + // binaries, but not for c++ binaries. We use two different patterns to + // to find our file. + patterns := []string{ + // Try the obvious path first. + filepath.Join(root, path), + // If it was a go binary, use a wildcard to match the build + // type. The pattern is: /test-path/__main__/directories/*/file. + filepath.Join(root, filepath.Dir(path), "*", filepath.Base(path)), + } + + for _, p := range patterns { + matches, err := filepath.Glob(p) + if err != nil { + // "The only possible returned error is ErrBadPattern, + // when pattern is malformed." -godoc + return "", fmt.Errorf("error globbing %q: %v", p, err) + } + switch len(matches) { + case 0: + // Try the next pattern. + case 1: + // We found it. + return matches[0], nil + default: + return "", fmt.Errorf("more than one match found for %q: %s", path, matches) + } + } + return "", fmt.Errorf("file %q not found", path) +} diff --git a/test/image/image_test.go b/test/image/image_test.go index d0dcb1861..0a1e19d6f 100644 --- a/test/image/image_test.go +++ b/test/image/image_test.go @@ -107,7 +107,7 @@ func TestHttpd(t *testing.T) { } d := dockerutil.MakeDocker("http-test") - dir, err := dockerutil.PrepareFiles("latin10k.txt") + dir, err := dockerutil.PrepareFiles("test/image/latin10k.txt") if err != nil { t.Fatalf("PrepareFiles() failed: %v", err) } @@ -139,7 +139,7 @@ func TestNginx(t *testing.T) { } d := dockerutil.MakeDocker("net-test") - dir, err := dockerutil.PrepareFiles("latin10k.txt") + dir, err := dockerutil.PrepareFiles("test/image/latin10k.txt") if err != nil { t.Fatalf("PrepareFiles() failed: %v", err) } @@ -183,7 +183,7 @@ func TestMysql(t *testing.T) { } client := dockerutil.MakeDocker("mysql-client-test") - dir, err := dockerutil.PrepareFiles("mysql.sql") + dir, err := dockerutil.PrepareFiles("test/image/mysql.sql") if err != nil { t.Fatalf("PrepareFiles() failed: %v", err) } @@ -283,7 +283,7 @@ func TestRuby(t *testing.T) { } d := dockerutil.MakeDocker("ruby-test") - dir, err := dockerutil.PrepareFiles("ruby.rb", "ruby.sh") + dir, err := dockerutil.PrepareFiles("test/image/ruby.rb", "test/image/ruby.sh") if err != nil { t.Fatalf("PrepareFiles() failed: %v", err) } diff --git a/test/syscalls/build_defs.bzl b/test/syscalls/build_defs.bzl index 1df761dd0..cbab85ef7 100644 --- a/test/syscalls/build_defs.bzl +++ b/test/syscalls/build_defs.bzl @@ -2,8 +2,6 @@ load("//tools:defs.bzl", "loopback") -# syscall_test is a macro that will create targets to run the given test target -# on the host (native) and runsc. def syscall_test( test, shard_count = 5, @@ -13,6 +11,19 @@ def syscall_test( add_uds_tree = False, add_hostinet = False, tags = None): + """syscall_test is a macro that will create targets for all platforms. + + Args: + test: the test target. + shard_count: shards for defined tests. + size: the defined test size. + use_tmpfs: use tmpfs in the defined tests. + add_overlay: add an overlay test. + add_uds_tree: add a UDS test. + add_hostinet: add a hostinet test. + tags: starting test tags. + """ + _syscall_test( test = test, shard_count = shard_count, @@ -111,6 +122,19 @@ def _syscall_test( # all the tests on a specific flavor. Use --test_tag_filters=ptrace,file_shared. tags += [full_platform, "file_" + file_access] + # Hash this target into one of 15 buckets. This can be used to + # randomly split targets between different workflows. + hash15 = hash(native.package_name() + name) % 15 + tags.append("hash15:" + str(hash15)) + + # TODO(b/139838000): Tests using hostinet must be disabled on Guitar until + # we figure out how to request ipv4 sockets on Guitar machines. + if network == "host": + tags.append("noguitar") + + # Disable off-host networking. + tags.append("requires-net:loopback") + # Add tag to prevent the tests from running in a Bazel sandbox. # TODO(b/120560048): Make the tests run without this tag. tags.append("no-sandbox") @@ -118,8 +142,11 @@ def _syscall_test( # TODO(b/112165693): KVM tests are tagged "manual" to until the platform is # more stable. if platform == "kvm": - tags += ["manual"] - tags += ["requires-kvm"] + tags.append("manual") + tags.append("requires-kvm") + + # TODO(b/112165693): Remove when tests pass reliably. + tags.append("notap") args = [ # Arguments are passed directly to syscall_test_runner binary. diff --git a/test/syscalls/linux/chroot.cc b/test/syscalls/linux/chroot.cc index 0a2d44a2c..85ec013d5 100644 --- a/test/syscalls/linux/chroot.cc +++ b/test/syscalls/linux/chroot.cc @@ -167,7 +167,7 @@ TEST(ChrootTest, DotDotFromOpenFD) { } // Test that link resolution in a chroot can escape the root by following an -// open proc fd. +// open proc fd. Regression test for b/32316719. TEST(ChrootTest, ProcFdLinkResolutionInChroot) { SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_SYS_CHROOT))); diff --git a/test/syscalls/linux/concurrency.cc b/test/syscalls/linux/concurrency.cc index f41f99900..7cd6a75bd 100644 --- a/test/syscalls/linux/concurrency.cc +++ b/test/syscalls/linux/concurrency.cc @@ -46,7 +46,8 @@ TEST(ConcurrencyTest, SingleProcessMultithreaded) { } // Test that multiple threads in this process continue to execute in parallel, -// even if an unrelated second process is spawned. +// even if an unrelated second process is spawned. Regression test for +// b/32119508. TEST(ConcurrencyTest, MultiProcessMultithreaded) { // In PID 1, start TIDs 1 and 2, and put both to sleep. // diff --git a/test/syscalls/linux/exec_proc_exe_workload.cc b/test/syscalls/linux/exec_proc_exe_workload.cc index b790fe5be..2989379b7 100644 --- a/test/syscalls/linux/exec_proc_exe_workload.cc +++ b/test/syscalls/linux/exec_proc_exe_workload.cc @@ -21,6 +21,12 @@ #include "test/util/posix_error.h" int main(int argc, char** argv, char** envp) { + // This is annoying. Because remote build systems may put these binaries + // in a content-addressable-store, you may wind up with /proc/self/exe + // pointing to some random path (but with a sensible argv[0]). + // + // Therefore, this test simply checks that the /proc/self/exe + // is absolute and *doesn't* match argv[1]. std::string exe = gvisor::testing::ProcessExePath(getpid()).ValueOrDie(); if (exe[0] != '/') { diff --git a/test/syscalls/linux/fork.cc b/test/syscalls/linux/fork.cc index 906f3358d..ff8bdfeb0 100644 --- a/test/syscalls/linux/fork.cc +++ b/test/syscalls/linux/fork.cc @@ -271,7 +271,7 @@ TEST_F(ForkTest, Alarm) { EXPECT_EQ(0, alarmed); } -// Child cannot affect parent private memory. +// Child cannot affect parent private memory. Regression test for b/24137240. TEST_F(ForkTest, PrivateMemory) { std::atomic local(0); @@ -298,6 +298,9 @@ TEST_F(ForkTest, PrivateMemory) { } // Kernel-accessed buffers should remain coherent across COW. +// +// The buffer must be >= usermem.ZeroCopyMinBytes, as UnsafeAccess operates +// differently. Regression test for b/33811887. TEST_F(ForkTest, COWSegment) { constexpr int kBufSize = 1024; char* read_buf = private_; diff --git a/test/syscalls/linux/mmap.cc b/test/syscalls/linux/mmap.cc index 1c4d9f1c7..11fb1b457 100644 --- a/test/syscalls/linux/mmap.cc +++ b/test/syscalls/linux/mmap.cc @@ -1418,7 +1418,7 @@ TEST_P(MMapFileParamTest, NoSigBusOnPageContainingEOF) { // // On most platforms this is trivial, but when the file is mapped via the sentry // page cache (which does not yet support writing to shared mappings), a bug -// caused reads to fail unnecessarily on such mappings. +// caused reads to fail unnecessarily on such mappings. See b/28913513. TEST_F(MMapFileTest, ReadingWritableSharedFilePageSucceeds) { uintptr_t addr; size_t len = strlen(kFileContents); @@ -1435,7 +1435,7 @@ TEST_F(MMapFileTest, ReadingWritableSharedFilePageSucceeds) { // Tests that EFAULT is returned when invoking a syscall that requires the OS to // read past end of file (resulting in a fault in sentry context in the gVisor -// case). +// case). See b/28913513. TEST_F(MMapFileTest, InternalSigBus) { uintptr_t addr; ASSERT_THAT(addr = Map(0, 2 * kPageSize, PROT_READ | PROT_WRITE, MAP_PRIVATE, @@ -1578,7 +1578,7 @@ TEST_F(MMapFileTest, Bug38498194) { } // Tests that reading from a file to a memory mapping of the same file does not -// deadlock. +// deadlock. See b/34813270. TEST_F(MMapFileTest, SelfRead) { uintptr_t addr; ASSERT_THAT(addr = Map(0, kPageSize, PROT_READ | PROT_WRITE, MAP_SHARED, @@ -1590,7 +1590,7 @@ TEST_F(MMapFileTest, SelfRead) { } // Tests that writing to a file from a memory mapping of the same file does not -// deadlock. +// deadlock. Regression test for b/34813270. TEST_F(MMapFileTest, SelfWrite) { uintptr_t addr; ASSERT_THAT(addr = Map(0, kPageSize, PROT_READ, MAP_SHARED, fd_.get(), 0), diff --git a/test/syscalls/linux/open_create.cc b/test/syscalls/linux/open_create.cc index 431733dbe..902d0a0dc 100644 --- a/test/syscalls/linux/open_create.cc +++ b/test/syscalls/linux/open_create.cc @@ -132,6 +132,7 @@ TEST(CreateTest, CreateFailsOnDirWithoutWritePerms) { } // A file originally created RW, but opened RO can later be opened RW. +// Regression test for b/65385065. TEST(CreateTest, OpenCreateROThenRW) { TempPath file(NewTempAbsPath()); diff --git a/test/syscalls/linux/preadv.cc b/test/syscalls/linux/preadv.cc index f7ea44054..5b0743fe9 100644 --- a/test/syscalls/linux/preadv.cc +++ b/test/syscalls/linux/preadv.cc @@ -37,6 +37,7 @@ namespace testing { namespace { +// Stress copy-on-write. Attempts to reproduce b/38430174. TEST(PreadvTest, MMConcurrencyStress) { // Fill a one-page file with zeroes (the contents don't really matter). const auto f = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFileWith( diff --git a/test/syscalls/linux/proc.cc b/test/syscalls/linux/proc.cc index 169b723eb..a23fdb58d 100644 --- a/test/syscalls/linux/proc.cc +++ b/test/syscalls/linux/proc.cc @@ -1352,13 +1352,19 @@ TEST(ProcPidSymlink, SubprocessZombied) { // FIXME(gvisor.dev/issue/164): Inconsistent behavior between gVisor and linux // on proc files. - // 4.17 & gVisor: Syscall succeeds and returns 1 + // + // ~4.3: Syscall fails with EACCES. + // 4.17 & gVisor: Syscall succeeds and returns 1. + // // EXPECT_THAT(ReadlinkWhileZombied("ns/pid", buf, sizeof(buf)), // SyscallFailsWithErrno(EACCES)); // FIXME(gvisor.dev/issue/164): Inconsistent behavior between gVisor and linux // on proc files. - // 4.17 & gVisor: Syscall succeeds and returns 1. + // + // ~4.3: Syscall fails with EACCES. + // 4.17 & gVisor: Syscall succeeds and returns 1. + // // EXPECT_THAT(ReadlinkWhileZombied("ns/user", buf, sizeof(buf)), // SyscallFailsWithErrno(EACCES)); } @@ -1431,8 +1437,12 @@ TEST(ProcPidFile, SubprocessRunning) { TEST(ProcPidFile, SubprocessZombie) { char buf[1]; - // 4.17: Succeeds and returns 1 - // gVisor: Succeeds and returns 0 + // FIXME(gvisor.dev/issue/164): Loosen requirement due to inconsistent + // behavior on different kernels. + // + // ~4.3: Succeds and returns 0. + // 4.17: Succeeds and returns 1. + // gVisor: Succeeds and returns 0. EXPECT_THAT(ReadWhileZombied("auxv", buf, sizeof(buf)), SyscallSucceeds()); EXPECT_THAT(ReadWhileZombied("cmdline", buf, sizeof(buf)), @@ -1458,7 +1468,10 @@ TEST(ProcPidFile, SubprocessZombie) { // FIXME(gvisor.dev/issue/164): Inconsistent behavior between gVisor and linux // on proc files. + // + // ~4.3: Fails and returns EACCES. // gVisor & 4.17: Succeeds and returns 1. + // // EXPECT_THAT(ReadWhileZombied("io", buf, sizeof(buf)), // SyscallFailsWithErrno(EACCES)); } @@ -1467,9 +1480,12 @@ TEST(ProcPidFile, SubprocessZombie) { TEST(ProcPidFile, SubprocessExited) { char buf[1]; - // FIXME(gvisor.dev/issue/164): Inconsistent behavior between kernels + // FIXME(gvisor.dev/issue/164): Inconsistent behavior between kernels. + // + // ~4.3: Fails and returns ESRCH. // gVisor: Fails with ESRCH. // 4.17: Succeeds and returns 1. + // // EXPECT_THAT(ReadWhileExited("auxv", buf, sizeof(buf)), // SyscallFailsWithErrno(ESRCH)); @@ -1641,7 +1657,7 @@ TEST(ProcTask, KilledThreadsDisappear) { EXPECT_NO_ERRNO(DirContainsExactly("/proc/self/task", TaskFiles(initial, {child1.Tid()}))); - // Stat child1's task file. + // Stat child1's task file. Regression test for b/32097707. struct stat statbuf; const std::string child1_task_file = absl::StrCat("/proc/self/task/", child1.Tid()); @@ -1669,7 +1685,7 @@ TEST(ProcTask, KilledThreadsDisappear) { EXPECT_NO_ERRNO(EventuallyDirContainsExactly( "/proc/self/task", TaskFiles(initial, {child3.Tid(), child5.Tid()}))); - // Stat child1's task file again. This time it should fail. + // Stat child1's task file again. This time it should fail. See b/32097707. EXPECT_THAT(stat(child1_task_file.c_str(), &statbuf), SyscallFailsWithErrno(ENOENT)); @@ -1824,7 +1840,7 @@ TEST(ProcSysVmOvercommitMemory, HasNumericValue) { } // Check that link for proc fd entries point the target node, not the -// symlink itself. +// symlink itself. Regression test for b/31155070. TEST(ProcTaskFd, FstatatFollowsSymlink) { const TempPath file = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateFile()); const FileDescriptor fd = @@ -1883,6 +1899,20 @@ TEST(ProcMounts, IsSymlink) { EXPECT_EQ(link, "self/mounts"); } +TEST(ProcSelfMountinfo, RequiredFieldsArePresent) { + auto mountinfo = + ASSERT_NO_ERRNO_AND_VALUE(GetContents("/proc/self/mountinfo")); + EXPECT_THAT( + mountinfo, + AllOf( + // Root mount. + ContainsRegex( + R"([0-9]+ [0-9]+ [0-9]+:[0-9]+ / / (rw|ro).*- \S+ \S+ (rw|ro)\S*)"), + // Proc mount - always rw. + ContainsRegex( + R"([0-9]+ [0-9]+ [0-9]+:[0-9]+ / /proc rw.*- \S+ \S+ rw\S*)"))); +} + // Check that /proc/self/mounts looks something like a real mounts file. TEST(ProcSelfMounts, RequiredFieldsArePresent) { auto mounts = ASSERT_NO_ERRNO_AND_VALUE(GetContents("/proc/self/mounts")); diff --git a/test/syscalls/linux/readv.cc b/test/syscalls/linux/readv.cc index 4069cbc7e..baaf9f757 100644 --- a/test/syscalls/linux/readv.cc +++ b/test/syscalls/linux/readv.cc @@ -254,7 +254,9 @@ TEST_F(ReadvTest, IovecOutsideTaskAddressRangeInNonemptyArray) { // This test depends on the maximum extent of a single readv() syscall, so // we can't tolerate interruption from saving. TEST(ReadvTestNoFixture, TruncatedAtMax_NoRandomSave) { - // Ensure that we won't be interrupted by ITIMER_PROF. + // Ensure that we won't be interrupted by ITIMER_PROF. This is particularly + // important in environments where automated profiling tools may start + // ITIMER_PROF automatically. struct itimerval itv = {}; auto const cleanup_itimer = ASSERT_NO_ERRNO_AND_VALUE(ScopedItimer(ITIMER_PROF, itv)); diff --git a/test/syscalls/linux/rseq.cc b/test/syscalls/linux/rseq.cc index 106c045e3..4bfb1ff56 100644 --- a/test/syscalls/linux/rseq.cc +++ b/test/syscalls/linux/rseq.cc @@ -36,7 +36,7 @@ namespace { // We must be very careful about how these tests are written. Each thread may // only have one struct rseq registration, which may be done automatically at // thread start (as of 2019-11-13, glibc does *not* support rseq and thus does -// not do so). +// not do so, but other libraries do). // // Testing of rseq is thus done primarily in a child process with no // registration. This means exec'ing a nostdlib binary, as rseq registration can diff --git a/test/syscalls/linux/select.cc b/test/syscalls/linux/select.cc index 424e2a67f..be2364fb8 100644 --- a/test/syscalls/linux/select.cc +++ b/test/syscalls/linux/select.cc @@ -146,7 +146,7 @@ TEST_F(SelectTest, IgnoreBitsAboveNfds) { // This test illustrates Linux's behavior of 'select' calls passing after // setrlimit RLIMIT_NOFILE is called. In particular, versions of sshd rely on -// this behavior. +// this behavior. See b/122318458. TEST_F(SelectTest, SetrlimitCallNOFILE) { fd_set read_set; FD_ZERO(&read_set); diff --git a/test/syscalls/linux/shm.cc b/test/syscalls/linux/shm.cc index 7ba752599..c7fdbb924 100644 --- a/test/syscalls/linux/shm.cc +++ b/test/syscalls/linux/shm.cc @@ -473,7 +473,7 @@ TEST(ShmTest, PartialUnmap) { } // Check that sentry does not panic when asked for a zero-length private shm -// segment. +// segment. Regression test for b/110694797. TEST(ShmTest, GracefullyFailOnZeroLenSegmentCreation) { EXPECT_THAT(Shmget(IPC_PRIVATE, 0, 0), PosixErrorIs(EINVAL, _)); } diff --git a/test/syscalls/linux/sigprocmask.cc b/test/syscalls/linux/sigprocmask.cc index 654c6a47f..a603fc1d1 100644 --- a/test/syscalls/linux/sigprocmask.cc +++ b/test/syscalls/linux/sigprocmask.cc @@ -237,7 +237,7 @@ TEST_F(SigProcMaskTest, SignalHandler) { } // Check that sigprocmask correctly handles aliasing of the set and oldset -// pointers. +// pointers. Regression test for b/30502311. TEST_F(SigProcMaskTest, AliasedSets) { sigset_t mask; diff --git a/test/syscalls/linux/socket_unix_non_stream.cc b/test/syscalls/linux/socket_unix_non_stream.cc index 276a94eb8..884319e1d 100644 --- a/test/syscalls/linux/socket_unix_non_stream.cc +++ b/test/syscalls/linux/socket_unix_non_stream.cc @@ -109,7 +109,7 @@ PosixErrorOr> CreateFragmentedRegion(const int size, } // A contiguous iov that is heavily fragmented in FileMem can still be sent -// successfully. +// successfully. See b/115833655. TEST_P(UnixNonStreamSocketPairTest, FragmentedSendMsg) { auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); @@ -165,7 +165,7 @@ TEST_P(UnixNonStreamSocketPairTest, FragmentedSendMsg) { } // A contiguous iov that is heavily fragmented in FileMem can still be received -// into successfully. +// into successfully. Regression test for b/115833655. TEST_P(UnixNonStreamSocketPairTest, FragmentedRecvMsg) { auto sockets = ASSERT_NO_ERRNO_AND_VALUE(NewSocketPair()); diff --git a/test/syscalls/linux/symlink.cc b/test/syscalls/linux/symlink.cc index b249ff91f..03ee1250d 100644 --- a/test/syscalls/linux/symlink.cc +++ b/test/syscalls/linux/symlink.cc @@ -38,7 +38,7 @@ mode_t FilePermission(const std::string& path) { } // Test that name collisions are checked on the new link path, not the source -// path. +// path. Regression test for b/31782115. TEST(SymlinkTest, CanCreateSymlinkWithCachedSourceDirent) { const std::string srcname = NewTempAbsPath(); const std::string newname = NewTempAbsPath(); diff --git a/test/syscalls/linux/tcp_socket.cc b/test/syscalls/linux/tcp_socket.cc index 8a8b68e75..c4591a3b9 100644 --- a/test/syscalls/linux/tcp_socket.cc +++ b/test/syscalls/linux/tcp_socket.cc @@ -244,7 +244,8 @@ TEST_P(TcpSocketTest, ZeroWriteAllowed) { } // Test that a non-blocking write with a buffer that is larger than the send -// buffer size will not actually write the whole thing at once. +// buffer size will not actually write the whole thing at once. Regression test +// for b/64438887. TEST_P(TcpSocketTest, NonblockingLargeWrite) { // Set the FD to O_NONBLOCK. int opts; diff --git a/test/syscalls/linux/time.cc b/test/syscalls/linux/time.cc index c7eead17e..1ccb95733 100644 --- a/test/syscalls/linux/time.cc +++ b/test/syscalls/linux/time.cc @@ -62,6 +62,7 @@ TEST(TimeTest, VsyscallTime_InvalidAddressSIGSEGV) { ::testing::KilledBySignal(SIGSEGV), ""); } +// Mimics the gettimeofday(2) wrapper from the Go runtime <= 1.2. int vsyscall_gettimeofday(struct timeval* tv, struct timezone* tz) { constexpr uint64_t kVsyscallGettimeofdayEntry = 0xffffffffff600000; return reinterpret_cast( diff --git a/test/syscalls/linux/tkill.cc b/test/syscalls/linux/tkill.cc index bae377c69..8d8ebbb24 100644 --- a/test/syscalls/linux/tkill.cc +++ b/test/syscalls/linux/tkill.cc @@ -54,7 +54,7 @@ void SigHandler(int sig, siginfo_t* info, void* context) { TEST_CHECK(info->si_code == SI_TKILL); } -// Test with a real signal. +// Test with a real signal. Regression test for b/24790092. TEST(TkillTest, ValidTIDAndRealSignal) { struct sigaction sa; sa.sa_sigaction = SigHandler; diff --git a/test/util/temp_path.cc b/test/util/temp_path.cc index 35aacb172..9c10b6674 100644 --- a/test/util/temp_path.cc +++ b/test/util/temp_path.cc @@ -77,6 +77,7 @@ std::string NewTempAbsPath() { std::string NewTempRelPath() { return NextTempBasename(); } std::string GetAbsoluteTestTmpdir() { + // Note that TEST_TMPDIR is guaranteed to be set. char* env_tmpdir = getenv("TEST_TMPDIR"); std::string tmp_dir = env_tmpdir != nullptr ? std::string(env_tmpdir) : "/tmp"; diff --git a/tools/build/tags.bzl b/tools/build/tags.bzl index e99c87f81..a6db44e47 100644 --- a/tools/build/tags.bzl +++ b/tools/build/tags.bzl @@ -33,4 +33,8 @@ go_suffixes = [ "_wasm_unsafe", "_linux", "_linux_unsafe", + "_opts", + "_opts_unsafe", + "_impl", + "_impl_unsafe", ] diff --git a/tools/defs.bzl b/tools/defs.bzl index 5d5fa134a..c03b557ae 100644 --- a/tools/defs.bzl +++ b/tools/defs.bzl @@ -73,6 +73,16 @@ def calculate_sets(srcs): result[target].append(file) return result +def go_imports(name, src, out): + """Simplify a single Go source file by eliminating unused imports.""" + native.genrule( + name = name, + srcs = [src], + outs = [out], + tools = ["@org_golang_x_tools//cmd/goimports:goimports"], + cmd = ("$(location @org_golang_x_tools//cmd/goimports:goimports) $(SRCS) > $@"), + ) + def go_library(name, srcs, deps = [], imports = [], stateify = True, marshal = False, **kwargs): """Wraps the standard go_library and does stateification and marshalling. @@ -107,10 +117,15 @@ def go_library(name, srcs, deps = [], imports = [], stateify = True, marshal = F state_sets = calculate_sets(srcs) for (suffix, srcs) in state_sets.items(): go_stateify( - name = name + suffix + "_state_autogen", + name = name + suffix + "_state_autogen_with_imports", srcs = srcs, imports = imports, package = name, + out = name + suffix + "_state_autogen_with_imports.go", + ) + go_imports( + name = name + suffix + "_state_autogen", + src = name + suffix + "_state_autogen_with_imports.go", out = name + suffix + "_state_autogen.go", ) all_srcs = all_srcs + [ -- cgit v1.2.3 From 17b9f5e66238bde1e4ed3bd9e5fb67342c8b58ec Mon Sep 17 00:00:00 2001 From: Dean Deng Date: Fri, 7 Feb 2020 14:46:24 -0800 Subject: Support listxattr and removexattr syscalls. Note that these are only implemented for tmpfs, and other impls will still return EOPNOTSUPP. PiperOrigin-RevId: 293899385 --- pkg/p9/client_file.go | 33 ++++ pkg/p9/file.go | 16 ++ pkg/p9/handlers.go | 33 ++++ pkg/p9/messages.go | 199 +++++++++++++++---- pkg/p9/p9.go | 4 + pkg/p9/version.go | 8 +- pkg/sentry/fs/copy_up.go | 2 +- pkg/sentry/fs/fsutil/inode.go | 20 +- pkg/sentry/fs/gofer/context_file.go | 14 ++ pkg/sentry/fs/gofer/inode.go | 13 +- pkg/sentry/fs/inode.go | 14 +- pkg/sentry/fs/inode_operations.go | 13 +- pkg/sentry/fs/inode_overlay.go | 18 +- pkg/sentry/fs/tmpfs/tmpfs.go | 9 +- pkg/sentry/syscalls/linux/linux64_amd64.go | 27 ++- pkg/sentry/syscalls/linux/linux64_arm64.go | 37 ++-- pkg/sentry/syscalls/linux/sys_xattr.go | 200 +++++++++++++++++++- runsc/fsgofer/fsgofer.go | 16 +- test/syscalls/linux/BUILD | 1 + test/syscalls/linux/xattr.cc | 294 ++++++++++++++++------------- 20 files changed, 733 insertions(+), 238 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/p9/client_file.go b/pkg/p9/client_file.go index 0254e4ccc..2ee07b664 100644 --- a/pkg/p9/client_file.go +++ b/pkg/p9/client_file.go @@ -194,6 +194,39 @@ func (c *clientFile) SetXattr(name, value string, flags uint32) error { return c.client.sendRecv(&Tsetxattr{FID: c.fid, Name: name, Value: value, Flags: flags}, &Rsetxattr{}) } +// ListXattr implements File.ListXattr. +func (c *clientFile) ListXattr(size uint64) (map[string]struct{}, error) { + if atomic.LoadUint32(&c.closed) != 0 { + return nil, syscall.EBADF + } + if !versionSupportsListRemoveXattr(c.client.version) { + return nil, syscall.EOPNOTSUPP + } + + rlistxattr := Rlistxattr{} + if err := c.client.sendRecv(&Tlistxattr{FID: c.fid, Size: size}, &rlistxattr); err != nil { + return nil, err + } + + xattrs := make(map[string]struct{}, len(rlistxattr.Xattrs)) + for _, x := range rlistxattr.Xattrs { + xattrs[x] = struct{}{} + } + return xattrs, nil +} + +// RemoveXattr implements File.RemoveXattr. +func (c *clientFile) RemoveXattr(name string) error { + if atomic.LoadUint32(&c.closed) != 0 { + return syscall.EBADF + } + if !versionSupportsListRemoveXattr(c.client.version) { + return syscall.EOPNOTSUPP + } + + return c.client.sendRecv(&Tremovexattr{FID: c.fid, Name: name}, &Rremovexattr{}) +} + // Allocate implements File.Allocate. func (c *clientFile) Allocate(mode AllocateMode, offset, length uint64) error { if atomic.LoadUint32(&c.closed) != 0 { diff --git a/pkg/p9/file.go b/pkg/p9/file.go index 4607cfcdf..d4ffbc8e3 100644 --- a/pkg/p9/file.go +++ b/pkg/p9/file.go @@ -105,6 +105,22 @@ type File interface { // TODO(b/127675828): Determine concurrency guarantees once implemented. SetXattr(name, value string, flags uint32) error + // ListXattr lists the names of the extended attributes on this node. + // + // Size indicates the size of the buffer that has been allocated to hold the + // attribute list. If the list would be larger than size, implementations may + // return ERANGE to indicate that the buffer is too small, but they are also + // free to ignore the hint entirely (i.e. the value returned may be larger + // than size). All size checking is done independently at the syscall layer. + // + // TODO(b/148303075): Determine concurrency guarantees once implemented. + ListXattr(size uint64) (map[string]struct{}, error) + + // RemoveXattr removes extended attributes on this node. + // + // TODO(b/148303075): Determine concurrency guarantees once implemented. + RemoveXattr(name string) error + // Allocate allows the caller to directly manipulate the allocated disk space // for the file. See fallocate(2) for more details. Allocate(mode AllocateMode, offset, length uint64) error diff --git a/pkg/p9/handlers.go b/pkg/p9/handlers.go index 7d6653a07..2ac45eb80 100644 --- a/pkg/p9/handlers.go +++ b/pkg/p9/handlers.go @@ -941,6 +941,39 @@ func (t *Tsetxattr) handle(cs *connState) message { return &Rsetxattr{} } +// handle implements handler.handle. +func (t *Tlistxattr) handle(cs *connState) message { + ref, ok := cs.LookupFID(t.FID) + if !ok { + return newErr(syscall.EBADF) + } + defer ref.DecRef() + + xattrs, err := ref.file.ListXattr(t.Size) + if err != nil { + return newErr(err) + } + xattrList := make([]string, 0, len(xattrs)) + for x := range xattrs { + xattrList = append(xattrList, x) + } + return &Rlistxattr{Xattrs: xattrList} +} + +// handle implements handler.handle. +func (t *Tremovexattr) handle(cs *connState) message { + ref, ok := cs.LookupFID(t.FID) + if !ok { + return newErr(syscall.EBADF) + } + defer ref.DecRef() + + if err := ref.file.RemoveXattr(t.Name); err != nil { + return newErr(err) + } + return &Rremovexattr{} +} + // handle implements handler.handle. func (t *Treaddir) handle(cs *connState) message { ref, ok := cs.LookupFID(t.Directory) diff --git a/pkg/p9/messages.go b/pkg/p9/messages.go index ceb723d86..b1cede5f5 100644 --- a/pkg/p9/messages.go +++ b/pkg/p9/messages.go @@ -174,11 +174,11 @@ type Rflush struct { } // Decode implements encoder.Decode. -func (*Rflush) Decode(b *buffer) { +func (*Rflush) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rflush) Encode(b *buffer) { +func (*Rflush) Encode(*buffer) { } // Type implements message.Type. @@ -188,7 +188,7 @@ func (*Rflush) Type() MsgType { // String implements fmt.Stringer. func (r *Rflush) String() string { - return fmt.Sprintf("RFlush{}") + return "RFlush{}" } // Twalk is a walk request. @@ -300,11 +300,11 @@ type Rclunk struct { } // Decode implements encoder.Decode. -func (*Rclunk) Decode(b *buffer) { +func (*Rclunk) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rclunk) Encode(b *buffer) { +func (*Rclunk) Encode(*buffer) { } // Type implements message.Type. @@ -314,7 +314,7 @@ func (*Rclunk) Type() MsgType { // String implements fmt.Stringer. func (r *Rclunk) String() string { - return fmt.Sprintf("Rclunk{}") + return "Rclunk{}" } // Tremove is a remove request. @@ -350,11 +350,11 @@ type Rremove struct { } // Decode implements encoder.Decode. -func (*Rremove) Decode(b *buffer) { +func (*Rremove) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rremove) Encode(b *buffer) { +func (*Rremove) Encode(*buffer) { } // Type implements message.Type. @@ -364,7 +364,7 @@ func (*Rremove) Type() MsgType { // String implements fmt.Stringer. func (r *Rremove) String() string { - return fmt.Sprintf("Rremove{}") + return "Rremove{}" } // Rlerror is an error response. @@ -745,16 +745,16 @@ func (*Rlink) Type() MsgType { } // Decode implements encoder.Decode. -func (*Rlink) Decode(b *buffer) { +func (*Rlink) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rlink) Encode(b *buffer) { +func (*Rlink) Encode(*buffer) { } // String implements fmt.Stringer. func (r *Rlink) String() string { - return fmt.Sprintf("Rlink{}") + return "Rlink{}" } // Trenameat is a rename request. @@ -803,11 +803,11 @@ type Rrenameat struct { } // Decode implements encoder.Decode. -func (*Rrenameat) Decode(b *buffer) { +func (*Rrenameat) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rrenameat) Encode(b *buffer) { +func (*Rrenameat) Encode(*buffer) { } // Type implements message.Type. @@ -817,7 +817,7 @@ func (*Rrenameat) Type() MsgType { // String implements fmt.Stringer. func (r *Rrenameat) String() string { - return fmt.Sprintf("Rrenameat{}") + return "Rrenameat{}" } // Tunlinkat is an unlink request. @@ -861,11 +861,11 @@ type Runlinkat struct { } // Decode implements encoder.Decode. -func (*Runlinkat) Decode(b *buffer) { +func (*Runlinkat) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Runlinkat) Encode(b *buffer) { +func (*Runlinkat) Encode(*buffer) { } // Type implements message.Type. @@ -875,7 +875,7 @@ func (*Runlinkat) Type() MsgType { // String implements fmt.Stringer. func (r *Runlinkat) String() string { - return fmt.Sprintf("Runlinkat{}") + return "Runlinkat{}" } // Trename is a rename request. @@ -922,11 +922,11 @@ type Rrename struct { } // Decode implements encoder.Decode. -func (*Rrename) Decode(b *buffer) { +func (*Rrename) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rrename) Encode(b *buffer) { +func (*Rrename) Encode(*buffer) { } // Type implements message.Type. @@ -936,7 +936,7 @@ func (*Rrename) Type() MsgType { // String implements fmt.Stringer. func (r *Rrename) String() string { - return fmt.Sprintf("Rrename{}") + return "Rrename{}" } // Treadlink is a readlink request. @@ -1409,11 +1409,11 @@ type Rsetattr struct { } // Decode implements encoder.Decode. -func (*Rsetattr) Decode(b *buffer) { +func (*Rsetattr) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rsetattr) Encode(b *buffer) { +func (*Rsetattr) Encode(*buffer) { } // Type implements message.Type. @@ -1423,7 +1423,7 @@ func (*Rsetattr) Type() MsgType { // String implements fmt.Stringer. func (r *Rsetattr) String() string { - return fmt.Sprintf("Rsetattr{}") + return "Rsetattr{}" } // Tallocate is an allocate request. This is an extension to 9P protocol, not @@ -1466,11 +1466,11 @@ type Rallocate struct { } // Decode implements encoder.Decode. -func (*Rallocate) Decode(b *buffer) { +func (*Rallocate) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rallocate) Encode(b *buffer) { +func (*Rallocate) Encode(*buffer) { } // Type implements message.Type. @@ -1480,7 +1480,71 @@ func (*Rallocate) Type() MsgType { // String implements fmt.Stringer. func (r *Rallocate) String() string { - return fmt.Sprintf("Rallocate{}") + return "Rallocate{}" +} + +// Tlistxattr is a listxattr request. +type Tlistxattr struct { + // FID refers to the file on which to list xattrs. + FID FID + + // Size is the buffer size for the xattr list. + Size uint64 +} + +// Decode implements encoder.Decode. +func (t *Tlistxattr) Decode(b *buffer) { + t.FID = b.ReadFID() + t.Size = b.Read64() +} + +// Encode implements encoder.Encode. +func (t *Tlistxattr) Encode(b *buffer) { + b.WriteFID(t.FID) + b.Write64(t.Size) +} + +// Type implements message.Type. +func (*Tlistxattr) Type() MsgType { + return MsgTlistxattr +} + +// String implements fmt.Stringer. +func (t *Tlistxattr) String() string { + return fmt.Sprintf("Tlistxattr{FID: %d, Size: %d}", t.FID, t.Size) +} + +// Rlistxattr is a listxattr response. +type Rlistxattr struct { + // Xattrs is a list of extended attribute names. + Xattrs []string +} + +// Decode implements encoder.Decode. +func (r *Rlistxattr) Decode(b *buffer) { + n := b.Read16() + r.Xattrs = r.Xattrs[:0] + for i := 0; i < int(n); i++ { + r.Xattrs = append(r.Xattrs, b.ReadString()) + } +} + +// Encode implements encoder.Encode. +func (r *Rlistxattr) Encode(b *buffer) { + b.Write16(uint16(len(r.Xattrs))) + for _, x := range r.Xattrs { + b.WriteString(x) + } +} + +// Type implements message.Type. +func (*Rlistxattr) Type() MsgType { + return MsgRlistxattr +} + +// String implements fmt.Stringer. +func (r *Rlistxattr) String() string { + return fmt.Sprintf("Rlistxattr{Xattrs: %v}", r.Xattrs) } // Txattrwalk walks extended attributes. @@ -1594,11 +1658,11 @@ type Rxattrcreate struct { } // Decode implements encoder.Decode. -func (r *Rxattrcreate) Decode(b *buffer) { +func (r *Rxattrcreate) Decode(*buffer) { } // Encode implements encoder.Encode. -func (r *Rxattrcreate) Encode(b *buffer) { +func (r *Rxattrcreate) Encode(*buffer) { } // Type implements message.Type. @@ -1608,7 +1672,7 @@ func (*Rxattrcreate) Type() MsgType { // String implements fmt.Stringer. func (r *Rxattrcreate) String() string { - return fmt.Sprintf("Rxattrcreate{}") + return "Rxattrcreate{}" } // Tgetxattr is a getxattr request. @@ -1719,11 +1783,11 @@ type Rsetxattr struct { } // Decode implements encoder.Decode. -func (r *Rsetxattr) Decode(b *buffer) { +func (r *Rsetxattr) Decode(*buffer) { } // Encode implements encoder.Encode. -func (r *Rsetxattr) Encode(b *buffer) { +func (r *Rsetxattr) Encode(*buffer) { } // Type implements message.Type. @@ -1733,7 +1797,60 @@ func (*Rsetxattr) Type() MsgType { // String implements fmt.Stringer. func (r *Rsetxattr) String() string { - return fmt.Sprintf("Rsetxattr{}") + return "Rsetxattr{}" +} + +// Tremovexattr is a removexattr request. +type Tremovexattr struct { + // FID refers to the file on which to set xattrs. + FID FID + + // Name is the attribute name. + Name string +} + +// Decode implements encoder.Decode. +func (t *Tremovexattr) Decode(b *buffer) { + t.FID = b.ReadFID() + t.Name = b.ReadString() +} + +// Encode implements encoder.Encode. +func (t *Tremovexattr) Encode(b *buffer) { + b.WriteFID(t.FID) + b.WriteString(t.Name) +} + +// Type implements message.Type. +func (*Tremovexattr) Type() MsgType { + return MsgTremovexattr +} + +// String implements fmt.Stringer. +func (t *Tremovexattr) String() string { + return fmt.Sprintf("Tremovexattr{FID: %d, Name: %s}", t.FID, t.Name) +} + +// Rremovexattr is a removexattr response. +type Rremovexattr struct { +} + +// Decode implements encoder.Decode. +func (r *Rremovexattr) Decode(*buffer) { +} + +// Encode implements encoder.Encode. +func (r *Rremovexattr) Encode(*buffer) { +} + +// Type implements message.Type. +func (*Rremovexattr) Type() MsgType { + return MsgRremovexattr +} + +// String implements fmt.Stringer. +func (r *Rremovexattr) String() string { + return "Rremovexattr{}" } // Treaddir is a readdir request. @@ -1880,11 +1997,11 @@ type Rfsync struct { } // Decode implements encoder.Decode. -func (*Rfsync) Decode(b *buffer) { +func (*Rfsync) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rfsync) Encode(b *buffer) { +func (*Rfsync) Encode(*buffer) { } // Type implements message.Type. @@ -1894,7 +2011,7 @@ func (*Rfsync) Type() MsgType { // String implements fmt.Stringer. func (r *Rfsync) String() string { - return fmt.Sprintf("Rfsync{}") + return "Rfsync{}" } // Tstatfs is a stat request. @@ -1980,11 +2097,11 @@ type Rflushf struct { } // Decode implements encoder.Decode. -func (*Rflushf) Decode(b *buffer) { +func (*Rflushf) Decode(*buffer) { } // Encode implements encoder.Encode. -func (*Rflushf) Encode(b *buffer) { +func (*Rflushf) Encode(*buffer) { } // Type implements message.Type. @@ -1994,7 +2111,7 @@ func (*Rflushf) Type() MsgType { // String implements fmt.Stringer. func (*Rflushf) String() string { - return fmt.Sprintf("Rflushf{}") + return "Rflushf{}" } // Twalkgetattr is a walk request. @@ -2484,6 +2601,8 @@ func init() { msgRegistry.register(MsgRgetattr, func() message { return &Rgetattr{} }) msgRegistry.register(MsgTsetattr, func() message { return &Tsetattr{} }) msgRegistry.register(MsgRsetattr, func() message { return &Rsetattr{} }) + msgRegistry.register(MsgTlistxattr, func() message { return &Tlistxattr{} }) + msgRegistry.register(MsgRlistxattr, func() message { return &Rlistxattr{} }) msgRegistry.register(MsgTxattrwalk, func() message { return &Txattrwalk{} }) msgRegistry.register(MsgRxattrwalk, func() message { return &Rxattrwalk{} }) msgRegistry.register(MsgTxattrcreate, func() message { return &Txattrcreate{} }) @@ -2492,6 +2611,8 @@ func init() { msgRegistry.register(MsgRgetxattr, func() message { return &Rgetxattr{} }) msgRegistry.register(MsgTsetxattr, func() message { return &Tsetxattr{} }) msgRegistry.register(MsgRsetxattr, func() message { return &Rsetxattr{} }) + msgRegistry.register(MsgTremovexattr, func() message { return &Tremovexattr{} }) + msgRegistry.register(MsgRremovexattr, func() message { return &Rremovexattr{} }) msgRegistry.register(MsgTreaddir, func() message { return &Treaddir{} }) msgRegistry.register(MsgRreaddir, func() message { return &Rreaddir{} }) msgRegistry.register(MsgTfsync, func() message { return &Tfsync{} }) diff --git a/pkg/p9/p9.go b/pkg/p9/p9.go index 5ab00d625..20ab31f7a 100644 --- a/pkg/p9/p9.go +++ b/pkg/p9/p9.go @@ -335,6 +335,8 @@ const ( MsgRgetattr = 25 MsgTsetattr = 26 MsgRsetattr = 27 + MsgTlistxattr = 28 + MsgRlistxattr = 29 MsgTxattrwalk = 30 MsgRxattrwalk = 31 MsgTxattrcreate = 32 @@ -343,6 +345,8 @@ const ( MsgRgetxattr = 35 MsgTsetxattr = 36 MsgRsetxattr = 37 + MsgTremovexattr = 38 + MsgRremovexattr = 39 MsgTreaddir = 40 MsgRreaddir = 41 MsgTfsync = 50 diff --git a/pkg/p9/version.go b/pkg/p9/version.go index 34a15eb55..09cde9f5a 100644 --- a/pkg/p9/version.go +++ b/pkg/p9/version.go @@ -26,7 +26,7 @@ const ( // // Clients are expected to start requesting this version number and // to continuously decrement it until a Tversion request succeeds. - highestSupportedVersion uint32 = 10 + highestSupportedVersion uint32 = 11 // lowestSupportedVersion is the lowest supported version X in a // version string of the format 9P2000.L.Google.X. @@ -167,3 +167,9 @@ func VersionSupportsOpenTruncateFlag(v uint32) bool { func versionSupportsGetSetXattr(v uint32) bool { return v >= 10 } + +// versionSupportsListRemoveXattr returns true if version v supports +// the Tlistxattr and Tremovexattr messages. +func versionSupportsListRemoveXattr(v uint32) bool { + return v >= 11 +} diff --git a/pkg/sentry/fs/copy_up.go b/pkg/sentry/fs/copy_up.go index f6c79e51b..b060a12ff 100644 --- a/pkg/sentry/fs/copy_up.go +++ b/pkg/sentry/fs/copy_up.go @@ -401,7 +401,7 @@ func copyAttributesLocked(ctx context.Context, upper *Inode, lower *Inode) error if err != nil { return err } - lowerXattr, err := lower.ListXattr(ctx) + lowerXattr, err := lower.ListXattr(ctx, linux.XATTR_SIZE_MAX) if err != nil && err != syserror.EOPNOTSUPP { return err } diff --git a/pkg/sentry/fs/fsutil/inode.go b/pkg/sentry/fs/fsutil/inode.go index 252830572..daecc4ffe 100644 --- a/pkg/sentry/fs/fsutil/inode.go +++ b/pkg/sentry/fs/fsutil/inode.go @@ -247,7 +247,7 @@ func (i *InodeSimpleExtendedAttributes) SetXattr(_ context.Context, _ *fs.Inode, } // ListXattr implements fs.InodeOperations.ListXattr. -func (i *InodeSimpleExtendedAttributes) ListXattr(context.Context, *fs.Inode) (map[string]struct{}, error) { +func (i *InodeSimpleExtendedAttributes) ListXattr(context.Context, *fs.Inode, uint64) (map[string]struct{}, error) { i.mu.RLock() names := make(map[string]struct{}, len(i.xattrs)) for name := range i.xattrs { @@ -257,6 +257,17 @@ func (i *InodeSimpleExtendedAttributes) ListXattr(context.Context, *fs.Inode) (m return names, nil } +// RemoveXattr implements fs.InodeOperations.RemoveXattr. +func (i *InodeSimpleExtendedAttributes) RemoveXattr(_ context.Context, _ *fs.Inode, name string) error { + i.mu.RLock() + defer i.mu.RUnlock() + if _, ok := i.xattrs[name]; ok { + delete(i.xattrs, name) + return nil + } + return syserror.ENOATTR +} + // staticFile is a file with static contents. It is returned by // InodeStaticFileGetter.GetFile. // @@ -460,10 +471,15 @@ func (InodeNoExtendedAttributes) SetXattr(context.Context, *fs.Inode, string, st } // ListXattr implements fs.InodeOperations.ListXattr. -func (InodeNoExtendedAttributes) ListXattr(context.Context, *fs.Inode) (map[string]struct{}, error) { +func (InodeNoExtendedAttributes) ListXattr(context.Context, *fs.Inode, uint64) (map[string]struct{}, error) { return nil, syserror.EOPNOTSUPP } +// RemoveXattr implements fs.InodeOperations.RemoveXattr. +func (InodeNoExtendedAttributes) RemoveXattr(context.Context, *fs.Inode, string) error { + return syserror.EOPNOTSUPP +} + // InodeNoopRelease implements fs.InodeOperations.Release as a noop. type InodeNoopRelease struct{} diff --git a/pkg/sentry/fs/gofer/context_file.go b/pkg/sentry/fs/gofer/context_file.go index 3da818aed..125907d70 100644 --- a/pkg/sentry/fs/gofer/context_file.go +++ b/pkg/sentry/fs/gofer/context_file.go @@ -73,6 +73,20 @@ func (c *contextFile) setXattr(ctx context.Context, name, value string, flags ui return err } +func (c *contextFile) listXattr(ctx context.Context, size uint64) (map[string]struct{}, error) { + ctx.UninterruptibleSleepStart(false) + xattrs, err := c.file.ListXattr(size) + ctx.UninterruptibleSleepFinish(false) + return xattrs, err +} + +func (c *contextFile) removeXattr(ctx context.Context, name string) error { + ctx.UninterruptibleSleepStart(false) + err := c.file.RemoveXattr(name) + ctx.UninterruptibleSleepFinish(false) + return err +} + func (c *contextFile) allocate(ctx context.Context, mode p9.AllocateMode, offset, length uint64) error { ctx.UninterruptibleSleepStart(false) err := c.file.Allocate(mode, offset, length) diff --git a/pkg/sentry/fs/gofer/inode.go b/pkg/sentry/fs/gofer/inode.go index ac28174d2..1c934981b 100644 --- a/pkg/sentry/fs/gofer/inode.go +++ b/pkg/sentry/fs/gofer/inode.go @@ -604,18 +604,23 @@ func (i *inodeOperations) Truncate(ctx context.Context, inode *fs.Inode, length } // GetXattr implements fs.InodeOperations.GetXattr. -func (i *inodeOperations) GetXattr(ctx context.Context, inode *fs.Inode, name string, size uint64) (string, error) { +func (i *inodeOperations) GetXattr(ctx context.Context, _ *fs.Inode, name string, size uint64) (string, error) { return i.fileState.file.getXattr(ctx, name, size) } // SetXattr implements fs.InodeOperations.SetXattr. -func (i *inodeOperations) SetXattr(ctx context.Context, inode *fs.Inode, name string, value string, flags uint32) error { +func (i *inodeOperations) SetXattr(ctx context.Context, _ *fs.Inode, name string, value string, flags uint32) error { return i.fileState.file.setXattr(ctx, name, value, flags) } // ListXattr implements fs.InodeOperations.ListXattr. -func (i *inodeOperations) ListXattr(context.Context, *fs.Inode) (map[string]struct{}, error) { - return nil, syscall.EOPNOTSUPP +func (i *inodeOperations) ListXattr(ctx context.Context, _ *fs.Inode, size uint64) (map[string]struct{}, error) { + return i.fileState.file.listXattr(ctx, size) +} + +// RemoveXattr implements fs.InodeOperations.RemoveXattr. +func (i *inodeOperations) RemoveXattr(ctx context.Context, _ *fs.Inode, name string) error { + return i.fileState.file.removeXattr(ctx, name) } // Allocate implements fs.InodeOperations.Allocate. diff --git a/pkg/sentry/fs/inode.go b/pkg/sentry/fs/inode.go index b66c091ab..55fb71c16 100644 --- a/pkg/sentry/fs/inode.go +++ b/pkg/sentry/fs/inode.go @@ -278,11 +278,19 @@ func (i *Inode) SetXattr(ctx context.Context, d *Dirent, name, value string, fla } // ListXattr calls i.InodeOperations.ListXattr with i as the Inode. -func (i *Inode) ListXattr(ctx context.Context) (map[string]struct{}, error) { +func (i *Inode) ListXattr(ctx context.Context, size uint64) (map[string]struct{}, error) { if i.overlay != nil { - return overlayListXattr(ctx, i.overlay) + return overlayListXattr(ctx, i.overlay, size) } - return i.InodeOperations.ListXattr(ctx, i) + return i.InodeOperations.ListXattr(ctx, i, size) +} + +// RemoveXattr calls i.InodeOperations.RemoveXattr with i as the Inode. +func (i *Inode) RemoveXattr(ctx context.Context, d *Dirent, name string) error { + if i.overlay != nil { + return overlayRemoveXattr(ctx, i.overlay, d, name) + } + return i.InodeOperations.RemoveXattr(ctx, i, name) } // CheckPermission will check if the caller may access this file in the diff --git a/pkg/sentry/fs/inode_operations.go b/pkg/sentry/fs/inode_operations.go index 70f2eae96..2bbfb72ef 100644 --- a/pkg/sentry/fs/inode_operations.go +++ b/pkg/sentry/fs/inode_operations.go @@ -190,7 +190,18 @@ type InodeOperations interface { // ListXattr returns the set of all extended attributes names that // have values. Inodes that do not support extended attributes return // EOPNOTSUPP. - ListXattr(ctx context.Context, inode *Inode) (map[string]struct{}, error) + // + // If this is called through the listxattr(2) syscall, size indicates the + // size of the buffer that the application has allocated to hold the + // attribute list. If the list would be larger than size, implementations may + // return ERANGE to indicate that the buffer is too small, but they are also + // free to ignore the hint entirely. All size checking is done independently + // at the syscall layer. + ListXattr(ctx context.Context, inode *Inode, size uint64) (map[string]struct{}, error) + + // RemoveXattr removes an extended attribute specified by name. Inodes that + // do not support extended attributes return EOPNOTSUPP. + RemoveXattr(ctx context.Context, inode *Inode, name string) error // Check determines whether an Inode can be accessed with the // requested permission mask using the context (which gives access diff --git a/pkg/sentry/fs/inode_overlay.go b/pkg/sentry/fs/inode_overlay.go index 4729b4aac..5ada33a32 100644 --- a/pkg/sentry/fs/inode_overlay.go +++ b/pkg/sentry/fs/inode_overlay.go @@ -564,15 +564,15 @@ func overlaySetxattr(ctx context.Context, o *overlayEntry, d *Dirent, name, valu return o.upper.SetXattr(ctx, d, name, value, flags) } -func overlayListXattr(ctx context.Context, o *overlayEntry) (map[string]struct{}, error) { +func overlayListXattr(ctx context.Context, o *overlayEntry, size uint64) (map[string]struct{}, error) { o.copyMu.RLock() defer o.copyMu.RUnlock() var names map[string]struct{} var err error if o.upper != nil { - names, err = o.upper.ListXattr(ctx) + names, err = o.upper.ListXattr(ctx, size) } else { - names, err = o.lower.ListXattr(ctx) + names, err = o.lower.ListXattr(ctx, size) } for name := range names { // Same as overlayGetXattr, we shouldn't forward along @@ -584,6 +584,18 @@ func overlayListXattr(ctx context.Context, o *overlayEntry) (map[string]struct{} return names, err } +func overlayRemoveXattr(ctx context.Context, o *overlayEntry, d *Dirent, name string) error { + // Don't allow changes to overlay xattrs through a removexattr syscall. + if strings.HasPrefix(XattrOverlayPrefix, name) { + return syserror.EPERM + } + + if err := copyUp(ctx, d); err != nil { + return err + } + return o.upper.RemoveXattr(ctx, d, name) +} + func overlayCheck(ctx context.Context, o *overlayEntry, p PermMask) error { o.copyMu.RLock() // Hot path. Avoid defers. diff --git a/pkg/sentry/fs/tmpfs/tmpfs.go b/pkg/sentry/fs/tmpfs/tmpfs.go index c00cef0a5..3c2b583ae 100644 --- a/pkg/sentry/fs/tmpfs/tmpfs.go +++ b/pkg/sentry/fs/tmpfs/tmpfs.go @@ -159,8 +159,13 @@ func (d *Dir) SetXattr(ctx context.Context, i *fs.Inode, name, value string, fla } // ListXattr implements fs.InodeOperations.ListXattr. -func (d *Dir) ListXattr(ctx context.Context, i *fs.Inode) (map[string]struct{}, error) { - return d.ramfsDir.ListXattr(ctx, i) +func (d *Dir) ListXattr(ctx context.Context, i *fs.Inode, size uint64) (map[string]struct{}, error) { + return d.ramfsDir.ListXattr(ctx, i, size) +} + +// RemoveXattr implements fs.InodeOperations.RemoveXattr. +func (d *Dir) RemoveXattr(ctx context.Context, i *fs.Inode, name string) error { + return d.ramfsDir.RemoveXattr(ctx, i, name) } // Lookup implements fs.InodeOperations.Lookup. diff --git a/pkg/sentry/syscalls/linux/linux64_amd64.go b/pkg/sentry/syscalls/linux/linux64_amd64.go index 588f8b087..79066ad2a 100644 --- a/pkg/sentry/syscalls/linux/linux64_amd64.go +++ b/pkg/sentry/syscalls/linux/linux64_amd64.go @@ -228,21 +228,18 @@ var AMD64 = &kernel.SyscallTable{ 185: syscalls.Error("security", syserror.ENOSYS, "Not implemented in Linux.", nil), 186: syscalls.Supported("gettid", Gettid), 187: syscalls.Supported("readahead", Readahead), - // TODO(b/148303075): Enable set/getxattr (in their various - // forms) once we also have list and removexattr. The JVM - // assumes that if get/set exist, then list and remove do too. - 188: syscalls.ErrorWithEvent("setxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 189: syscalls.ErrorWithEvent("lsetxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 190: syscalls.ErrorWithEvent("fsetxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 191: syscalls.ErrorWithEvent("getxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 192: syscalls.ErrorWithEvent("lgetxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 193: syscalls.ErrorWithEvent("fgetxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 194: syscalls.ErrorWithEvent("listxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 195: syscalls.ErrorWithEvent("llistxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 196: syscalls.ErrorWithEvent("flistxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 197: syscalls.ErrorWithEvent("removexattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 198: syscalls.ErrorWithEvent("lremovexattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 199: syscalls.ErrorWithEvent("fremovexattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), + 188: syscalls.PartiallySupported("setxattr", SetXattr, "Only supported for tmpfs.", nil), + 189: syscalls.PartiallySupported("lsetxattr", LSetXattr, "Only supported for tmpfs.", nil), + 190: syscalls.PartiallySupported("fsetxattr", FSetXattr, "Only supported for tmpfs.", nil), + 191: syscalls.PartiallySupported("getxattr", GetXattr, "Only supported for tmpfs.", nil), + 192: syscalls.PartiallySupported("lgetxattr", LGetXattr, "Only supported for tmpfs.", nil), + 193: syscalls.PartiallySupported("fgetxattr", FGetXattr, "Only supported for tmpfs.", nil), + 194: syscalls.PartiallySupported("listxattr", ListXattr, "Only supported for tmpfs", nil), + 195: syscalls.PartiallySupported("llistxattr", LListXattr, "Only supported for tmpfs", nil), + 196: syscalls.PartiallySupported("flistxattr", FListXattr, "Only supported for tmpfs", nil), + 197: syscalls.PartiallySupported("removexattr", RemoveXattr, "Only supported for tmpfs", nil), + 198: syscalls.PartiallySupported("lremovexattr", LRemoveXattr, "Only supported for tmpfs", nil), + 199: syscalls.PartiallySupported("fremovexattr", FRemoveXattr, "Only supported for tmpfs", nil), 200: syscalls.Supported("tkill", Tkill), 201: syscalls.Supported("time", Time), 202: syscalls.PartiallySupported("futex", Futex, "Robust futexes not supported.", nil), diff --git a/pkg/sentry/syscalls/linux/linux64_arm64.go b/pkg/sentry/syscalls/linux/linux64_arm64.go index 06e5ee401..7421619de 100644 --- a/pkg/sentry/syscalls/linux/linux64_arm64.go +++ b/pkg/sentry/syscalls/linux/linux64_arm64.go @@ -36,26 +36,23 @@ var ARM64 = &kernel.SyscallTable{ }, AuditNumber: linux.AUDIT_ARCH_AARCH64, Table: map[uintptr]kernel.Syscall{ - 0: syscalls.PartiallySupported("io_setup", IoSetup, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 1: syscalls.PartiallySupported("io_destroy", IoDestroy, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 2: syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 3: syscalls.PartiallySupported("io_cancel", IoCancel, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - 4: syscalls.PartiallySupported("io_getevents", IoGetevents, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), - // TODO(b/148303075): Enable set/getxattr (in their various - // forms) once we also have list and removexattr. The JVM - // assumes that if get/set exist, then list and remove do too. - 5: syscalls.ErrorWithEvent("setxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 6: syscalls.ErrorWithEvent("lsetxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 7: syscalls.ErrorWithEvent("fsetxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 8: syscalls.ErrorWithEvent("getxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 9: syscalls.ErrorWithEvent("lgetxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 10: syscalls.ErrorWithEvent("fgetxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 11: syscalls.ErrorWithEvent("listxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 13: syscalls.ErrorWithEvent("llistxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 13: syscalls.ErrorWithEvent("flistxattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 14: syscalls.ErrorWithEvent("removexattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 15: syscalls.ErrorWithEvent("lremovexattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), - 16: syscalls.ErrorWithEvent("fremovexattr", syserror.ENOTSUP, "Requires filesystem support.", []string{"gvisor.dev/issue/1636"}), + 0: syscalls.PartiallySupported("io_setup", IoSetup, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 1: syscalls.PartiallySupported("io_destroy", IoDestroy, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 2: syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 3: syscalls.PartiallySupported("io_cancel", IoCancel, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 4: syscalls.PartiallySupported("io_getevents", IoGetevents, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}), + 5: syscalls.PartiallySupported("setxattr", SetXattr, "Only supported for tmpfs.", nil), + 6: syscalls.PartiallySupported("lsetxattr", LSetXattr, "Only supported for tmpfs.", nil), + 7: syscalls.PartiallySupported("fsetxattr", FSetXattr, "Only supported for tmpfs.", nil), + 8: syscalls.PartiallySupported("getxattr", GetXattr, "Only supported for tmpfs.", nil), + 9: syscalls.PartiallySupported("lgetxattr", LGetXattr, "Only supported for tmpfs.", nil), + 10: syscalls.PartiallySupported("fgetxattr", FGetXattr, "Only supported for tmpfs.", nil), + 11: syscalls.PartiallySupported("listxattr", ListXattr, "Only supported for tmpfs", nil), + 12: syscalls.PartiallySupported("llistxattr", LListXattr, "Only supported for tmpfs", nil), + 13: syscalls.PartiallySupported("flistxattr", FListXattr, "Only supported for tmpfs", nil), + 14: syscalls.PartiallySupported("removexattr", RemoveXattr, "Only supported for tmpfs", nil), + 15: syscalls.PartiallySupported("lremovexattr", LRemoveXattr, "Only supported for tmpfs", nil), + 16: syscalls.PartiallySupported("fremovexattr", FRemoveXattr, "Only supported for tmpfs", nil), 17: syscalls.Supported("getcwd", Getcwd), 18: syscalls.CapError("lookup_dcookie", linux.CAP_SYS_ADMIN, "", nil), 19: syscalls.Supported("eventfd2", Eventfd2), diff --git a/pkg/sentry/syscalls/linux/sys_xattr.go b/pkg/sentry/syscalls/linux/sys_xattr.go index efb95555c..342337726 100644 --- a/pkg/sentry/syscalls/linux/sys_xattr.go +++ b/pkg/sentry/syscalls/linux/sys_xattr.go @@ -72,7 +72,7 @@ func getXattrFromPath(t *kernel.Task, args arch.SyscallArguments, resolveSymlink } valueLen := 0 - err = fileOpOn(t, linux.AT_FDCWD, path, resolveSymlink, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { + err = fileOpOn(t, linux.AT_FDCWD, path, resolveSymlink, func(_ *fs.Dirent, d *fs.Dirent, _ uint) error { if dirPath && !fs.IsDir(d.Inode.StableAttr) { return syserror.ENOTDIR } @@ -172,7 +172,7 @@ func setXattrFromPath(t *kernel.Task, args arch.SyscallArguments, resolveSymlink return 0, nil, err } - return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, resolveSymlink, func(root *fs.Dirent, d *fs.Dirent, _ uint) error { + return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, resolveSymlink, func(_ *fs.Dirent, d *fs.Dirent, _ uint) error { if dirPath && !fs.IsDir(d.Inode.StableAttr) { return syserror.ENOTDIR } @@ -187,12 +187,12 @@ func setXattr(t *kernel.Task, d *fs.Dirent, nameAddr, valueAddr usermem.Addr, si return syserror.EINVAL } - if err := checkXattrPermissions(t, d.Inode, fs.PermMask{Write: true}); err != nil { + name, err := copyInXattrName(t, nameAddr) + if err != nil { return err } - name, err := copyInXattrName(t, nameAddr) - if err != nil { + if err := checkXattrPermissions(t, d.Inode, fs.PermMask{Write: true}); err != nil { return err } @@ -226,12 +226,18 @@ func copyInXattrName(t *kernel.Task, nameAddr usermem.Addr) (string, error) { return name, nil } +// Restrict xattrs to regular files and directories. +// +// TODO(b/148380782): In Linux, this restriction technically only applies to +// xattrs in the "user.*" namespace. Make file type checks specific to the +// namespace once we allow other xattr prefixes. +func xattrFileTypeOk(i *fs.Inode) bool { + return fs.IsRegular(i.StableAttr) || fs.IsDir(i.StableAttr) +} + func checkXattrPermissions(t *kernel.Task, i *fs.Inode, perms fs.PermMask) error { // Restrict xattrs to regular files and directories. - // - // In Linux, this restriction technically only applies to xattrs in the - // "user.*" namespace, but we don't allow any other xattr prefixes anyway. - if !fs.IsRegular(i.StableAttr) && !fs.IsDir(i.StableAttr) { + if !xattrFileTypeOk(i) { if perms.Write { return syserror.EPERM } @@ -240,3 +246,179 @@ func checkXattrPermissions(t *kernel.Task, i *fs.Inode, perms fs.PermMask) error return i.CheckPermission(t, perms) } + +// ListXattr implements linux syscall listxattr(2). +func ListXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return listXattrFromPath(t, args, true) +} + +// LListXattr implements linux syscall llistxattr(2). +func LListXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return listXattrFromPath(t, args, false) +} + +// FListXattr implements linux syscall flistxattr(2). +func FListXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + listAddr := args[1].Pointer() + size := uint64(args[2].SizeT()) + + // TODO(b/113957122): Return EBADF if the fd was opened with O_PATH. + f := t.GetFile(fd) + if f == nil { + return 0, nil, syserror.EBADF + } + defer f.DecRef() + + n, err := listXattr(t, f.Dirent, listAddr, size) + if err != nil { + return 0, nil, err + } + + return uintptr(n), nil, nil +} + +func listXattrFromPath(t *kernel.Task, args arch.SyscallArguments, resolveSymlink bool) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + listAddr := args[1].Pointer() + size := uint64(args[2].SizeT()) + + path, dirPath, err := copyInPath(t, pathAddr, false /* allowEmpty */) + if err != nil { + return 0, nil, err + } + + n := 0 + err = fileOpOn(t, linux.AT_FDCWD, path, resolveSymlink, func(_ *fs.Dirent, d *fs.Dirent, _ uint) error { + if dirPath && !fs.IsDir(d.Inode.StableAttr) { + return syserror.ENOTDIR + } + + n, err = listXattr(t, d, listAddr, size) + return err + }) + if err != nil { + return 0, nil, err + } + + return uintptr(n), nil, nil +} + +func listXattr(t *kernel.Task, d *fs.Dirent, addr usermem.Addr, size uint64) (int, error) { + if !xattrFileTypeOk(d.Inode) { + return 0, nil + } + + // If listxattr(2) is called with size 0, the buffer size needed to contain + // the xattr list will be returned successfully even if it is nonzero. In + // that case, we need to retrieve the entire list so we can compute and + // return the correct size. + requestedSize := size + if size == 0 || size > linux.XATTR_SIZE_MAX { + requestedSize = linux.XATTR_SIZE_MAX + } + xattrs, err := d.Inode.ListXattr(t, requestedSize) + if err != nil { + return 0, err + } + + // TODO(b/148380782): support namespaces other than "user". + for x := range xattrs { + if !strings.HasPrefix(x, linux.XATTR_USER_PREFIX) { + delete(xattrs, x) + } + } + + listSize := xattrListSize(xattrs) + if listSize > linux.XATTR_SIZE_MAX { + return 0, syserror.E2BIG + } + if uint64(listSize) > requestedSize { + return 0, syserror.ERANGE + } + + // Don't copy out the attributes if size is 0. + if size == 0 { + return listSize, nil + } + + buf := make([]byte, 0, listSize) + for x := range xattrs { + buf = append(buf, []byte(x)...) + buf = append(buf, 0) + } + if _, err := t.CopyOutBytes(addr, buf); err != nil { + return 0, err + } + + return len(buf), nil +} + +func xattrListSize(xattrs map[string]struct{}) int { + size := 0 + for x := range xattrs { + size += len(x) + 1 + } + return size +} + +// RemoveXattr implements linux syscall removexattr(2). +func RemoveXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return removeXattrFromPath(t, args, true) +} + +// LRemoveXattr implements linux syscall lremovexattr(2). +func LRemoveXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + return removeXattrFromPath(t, args, false) +} + +// FRemoveXattr implements linux syscall fremovexattr(2). +func FRemoveXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + nameAddr := args[1].Pointer() + + // TODO(b/113957122): Return EBADF if the fd was opened with O_PATH. + f := t.GetFile(fd) + if f == nil { + return 0, nil, syserror.EBADF + } + defer f.DecRef() + + return 0, nil, removeXattr(t, f.Dirent, nameAddr) +} + +func removeXattrFromPath(t *kernel.Task, args arch.SyscallArguments, resolveSymlink bool) (uintptr, *kernel.SyscallControl, error) { + pathAddr := args[0].Pointer() + nameAddr := args[1].Pointer() + + path, dirPath, err := copyInPath(t, pathAddr, false /* allowEmpty */) + if err != nil { + return 0, nil, err + } + + return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, resolveSymlink, func(_ *fs.Dirent, d *fs.Dirent, _ uint) error { + if dirPath && !fs.IsDir(d.Inode.StableAttr) { + return syserror.ENOTDIR + } + + return removeXattr(t, d, nameAddr) + }) +} + +// removeXattr implements removexattr(2) from the given *fs.Dirent. +func removeXattr(t *kernel.Task, d *fs.Dirent, nameAddr usermem.Addr) error { + name, err := copyInXattrName(t, nameAddr) + if err != nil { + return err + } + + if err := checkXattrPermissions(t, d.Inode, fs.PermMask{Write: true}); err != nil { + return err + } + + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + + return d.Inode.RemoveXattr(t, d, name) +} diff --git a/runsc/fsgofer/fsgofer.go b/runsc/fsgofer/fsgofer.go index 4d84ad999..cadd83273 100644 --- a/runsc/fsgofer/fsgofer.go +++ b/runsc/fsgofer/fsgofer.go @@ -768,12 +768,22 @@ func (l *localFile) SetAttr(valid p9.SetAttrMask, attr p9.SetAttr) error { } // TODO(b/127675828): support getxattr. -func (l *localFile) GetXattr(name string, size uint64) (string, error) { +func (*localFile) GetXattr(string, uint64) (string, error) { return "", syscall.EOPNOTSUPP } // TODO(b/127675828): support setxattr. -func (l *localFile) SetXattr(name, value string, flags uint32) error { +func (*localFile) SetXattr(string, string, uint32) error { + return syscall.EOPNOTSUPP +} + +// TODO(b/148303075): support listxattr. +func (*localFile) ListXattr(uint64) (map[string]struct{}, error) { + return nil, syscall.EOPNOTSUPP +} + +// TODO(b/148303075): support removexattr. +func (*localFile) RemoveXattr(string) error { return syscall.EOPNOTSUPP } @@ -790,7 +800,7 @@ func (l *localFile) Allocate(mode p9.AllocateMode, offset, length uint64) error } // Rename implements p9.File; this should never be called. -func (l *localFile) Rename(p9.File, string) error { +func (*localFile) Rename(p9.File, string) error { panic("rename called directly") } diff --git a/test/syscalls/linux/BUILD b/test/syscalls/linux/BUILD index 12d389c3e..ca1af209a 100644 --- a/test/syscalls/linux/BUILD +++ b/test/syscalls/linux/BUILD @@ -3782,6 +3782,7 @@ cc_binary( "//test/util:capability_util", "//test/util:file_descriptor", "//test/util:fs_util", + "@com_google_absl//absl/container:flat_hash_set", "@com_google_absl//absl/strings", gtest, "//test/util:posix_error", diff --git a/test/syscalls/linux/xattr.cc b/test/syscalls/linux/xattr.cc index 85eb31847..8b00ef44c 100644 --- a/test/syscalls/linux/xattr.cc +++ b/test/syscalls/linux/xattr.cc @@ -24,6 +24,7 @@ #include "gmock/gmock.h" #include "gtest/gtest.h" +#include "absl/container/flat_hash_set.h" #include "test/syscalls/linux/file_base.h" #include "test/util/capability_util.h" #include "test/util/file_descriptor.h" @@ -38,36 +39,36 @@ namespace { class XattrTest : public FileTest {}; -TEST_F(XattrTest, XattrNullName) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); +TEST_F(XattrTest, XattrNonexistentFile) { + const char* path = "/does/not/exist"; + EXPECT_THAT(setxattr(path, nullptr, nullptr, 0, /*flags=*/0), + SyscallFailsWithErrno(ENOENT)); + EXPECT_THAT(getxattr(path, nullptr, nullptr, 0), + SyscallFailsWithErrno(ENOENT)); + EXPECT_THAT(listxattr(path, nullptr, 0), SyscallFailsWithErrno(ENOENT)); + EXPECT_THAT(removexattr(path, nullptr), SyscallFailsWithErrno(ENOENT)); +} +TEST_F(XattrTest, XattrNullName) { const char* path = test_file_name_.c_str(); EXPECT_THAT(setxattr(path, nullptr, nullptr, 0, /*flags=*/0), SyscallFailsWithErrno(EFAULT)); EXPECT_THAT(getxattr(path, nullptr, nullptr, 0), SyscallFailsWithErrno(EFAULT)); + EXPECT_THAT(removexattr(path, nullptr), SyscallFailsWithErrno(EFAULT)); } TEST_F(XattrTest, XattrEmptyName) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); EXPECT_THAT(setxattr(path, "", nullptr, 0, /*flags=*/0), SyscallFailsWithErrno(ERANGE)); EXPECT_THAT(getxattr(path, "", nullptr, 0), SyscallFailsWithErrno(ERANGE)); + EXPECT_THAT(removexattr(path, ""), SyscallFailsWithErrno(ERANGE)); } TEST_F(XattrTest, XattrLargeName) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); std::string name = "user."; name += std::string(XATTR_NAME_MAX - name.length(), 'a'); @@ -86,28 +87,23 @@ TEST_F(XattrTest, XattrLargeName) { SyscallFailsWithErrno(ERANGE)); EXPECT_THAT(getxattr(path, name.c_str(), nullptr, 0), SyscallFailsWithErrno(ERANGE)); + EXPECT_THAT(removexattr(path, name.c_str()), SyscallFailsWithErrno(ERANGE)); } TEST_F(XattrTest, XattrInvalidPrefix) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); std::string name(XATTR_NAME_MAX, 'a'); EXPECT_THAT(setxattr(path, name.c_str(), nullptr, 0, /*flags=*/0), SyscallFailsWithErrno(EOPNOTSUPP)); EXPECT_THAT(getxattr(path, name.c_str(), nullptr, 0), SyscallFailsWithErrno(EOPNOTSUPP)); + EXPECT_THAT(removexattr(path, name.c_str()), + SyscallFailsWithErrno(EOPNOTSUPP)); } // Do not allow save/restore cycles after making the test file read-only, as // the restore will fail to open it with r/w permissions. TEST_F(XattrTest, XattrReadOnly_NoRandomSave) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - // Drop capabilities that allow us to override file and directory permissions. ASSERT_NO_ERRNO(SetCapability(CAP_DAC_OVERRIDE, false)); ASSERT_NO_ERRNO(SetCapability(CAP_DAC_READ_SEARCH, false)); @@ -124,19 +120,21 @@ TEST_F(XattrTest, XattrReadOnly_NoRandomSave) { EXPECT_THAT(setxattr(path, name, &val, size, /*flags=*/0), SyscallFailsWithErrno(EACCES)); + EXPECT_THAT(removexattr(path, name), SyscallFailsWithErrno(EACCES)); char buf = '-'; EXPECT_THAT(getxattr(path, name, &buf, size), SyscallSucceedsWithValue(size)); EXPECT_EQ(buf, val); + + char list[sizeof(name)]; + EXPECT_THAT(listxattr(path, list, sizeof(list)), + SyscallSucceedsWithValue(sizeof(name))); + EXPECT_STREQ(list, name); } // Do not allow save/restore cycles after making the test file write-only, as // the restore will fail to open it with r/w permissions. TEST_F(XattrTest, XattrWriteOnly_NoRandomSave) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - // Drop capabilities that allow us to override file and directory permissions. ASSERT_NO_ERRNO(SetCapability(CAP_DAC_OVERRIDE, false)); ASSERT_NO_ERRNO(SetCapability(CAP_DAC_READ_SEARCH, false)); @@ -152,6 +150,14 @@ TEST_F(XattrTest, XattrWriteOnly_NoRandomSave) { EXPECT_THAT(setxattr(path, name, &val, size, /*flags=*/0), SyscallSucceeds()); EXPECT_THAT(getxattr(path, name, nullptr, 0), SyscallFailsWithErrno(EACCES)); + + // listxattr will succeed even without read permissions. + char list[sizeof(name)]; + EXPECT_THAT(listxattr(path, list, sizeof(list)), + SyscallSucceedsWithValue(sizeof(name))); + EXPECT_STREQ(list, name); + + EXPECT_THAT(removexattr(path, name), SyscallSucceeds()); } TEST_F(XattrTest, XattrTrustedWithNonadmin) { @@ -163,64 +169,66 @@ TEST_F(XattrTest, XattrTrustedWithNonadmin) { const char name[] = "trusted.abc"; EXPECT_THAT(setxattr(path, name, nullptr, 0, /*flags=*/0), SyscallFailsWithErrno(EPERM)); + EXPECT_THAT(removexattr(path, name), SyscallFailsWithErrno(EPERM)); EXPECT_THAT(getxattr(path, name, nullptr, 0), SyscallFailsWithErrno(ENODATA)); } TEST_F(XattrTest, XattrOnDirectory) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - TempPath dir = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateDir()); const char name[] = "user.test"; - EXPECT_THAT(setxattr(dir.path().c_str(), name, NULL, 0, /*flags=*/0), + EXPECT_THAT(setxattr(dir.path().c_str(), name, nullptr, 0, /*flags=*/0), SyscallSucceeds()); - EXPECT_THAT(getxattr(dir.path().c_str(), name, NULL, 0), + EXPECT_THAT(getxattr(dir.path().c_str(), name, nullptr, 0), SyscallSucceedsWithValue(0)); + + char list[sizeof(name)]; + EXPECT_THAT(listxattr(dir.path().c_str(), list, sizeof(list)), + SyscallSucceedsWithValue(sizeof(name))); + EXPECT_STREQ(list, name); + + EXPECT_THAT(removexattr(dir.path().c_str(), name), SyscallSucceeds()); } TEST_F(XattrTest, XattrOnSymlink) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - TempPath dir = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateDir()); TempPath link = ASSERT_NO_ERRNO_AND_VALUE( TempPath::CreateSymlinkTo(dir.path(), test_file_name_)); const char name[] = "user.test"; - EXPECT_THAT(setxattr(link.path().c_str(), name, NULL, 0, /*flags=*/0), + EXPECT_THAT(setxattr(link.path().c_str(), name, nullptr, 0, /*flags=*/0), SyscallSucceeds()); - EXPECT_THAT(getxattr(link.path().c_str(), name, NULL, 0), + EXPECT_THAT(getxattr(link.path().c_str(), name, nullptr, 0), SyscallSucceedsWithValue(0)); + + char list[sizeof(name)]; + EXPECT_THAT(listxattr(link.path().c_str(), list, sizeof(list)), + SyscallSucceedsWithValue(sizeof(name))); + EXPECT_STREQ(list, name); + + EXPECT_THAT(removexattr(link.path().c_str(), name), SyscallSucceeds()); } TEST_F(XattrTest, XattrOnInvalidFileTypes) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char name[] = "user.test"; char char_device[] = "/dev/zero"; - EXPECT_THAT(setxattr(char_device, name, NULL, 0, /*flags=*/0), + EXPECT_THAT(setxattr(char_device, name, nullptr, 0, /*flags=*/0), SyscallFailsWithErrno(EPERM)); - EXPECT_THAT(getxattr(char_device, name, NULL, 0), + EXPECT_THAT(getxattr(char_device, name, nullptr, 0), SyscallFailsWithErrno(ENODATA)); + EXPECT_THAT(listxattr(char_device, nullptr, 0), SyscallSucceedsWithValue(0)); // Use tmpfs, where creation of named pipes is supported. const std::string fifo = NewTempAbsPathInDir("/dev/shm"); const char* path = fifo.c_str(); EXPECT_THAT(mknod(path, S_IFIFO | S_IRUSR | S_IWUSR, 0), SyscallSucceeds()); - EXPECT_THAT(setxattr(path, name, NULL, 0, /*flags=*/0), + EXPECT_THAT(setxattr(path, name, nullptr, 0, /*flags=*/0), SyscallFailsWithErrno(EPERM)); - EXPECT_THAT(getxattr(path, name, NULL, 0), SyscallFailsWithErrno(ENODATA)); + EXPECT_THAT(getxattr(path, name, nullptr, 0), SyscallFailsWithErrno(ENODATA)); + EXPECT_THAT(listxattr(path, nullptr, 0), SyscallSucceedsWithValue(0)); + EXPECT_THAT(removexattr(path, name), SyscallFailsWithErrno(EPERM)); } TEST_F(XattrTest, SetxattrSizeSmallerThanValue) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; std::vector val = {'a', 'a'}; @@ -236,10 +244,6 @@ TEST_F(XattrTest, SetxattrSizeSmallerThanValue) { } TEST_F(XattrTest, SetxattrZeroSize) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; char val = 'a'; @@ -252,10 +256,6 @@ TEST_F(XattrTest, SetxattrZeroSize) { } TEST_F(XattrTest, SetxattrSizeTooLarge) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; @@ -271,10 +271,6 @@ TEST_F(XattrTest, SetxattrSizeTooLarge) { } TEST_F(XattrTest, SetxattrNullValueAndNonzeroSize) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; EXPECT_THAT(setxattr(path, name, nullptr, 1, /*flags=*/0), @@ -284,10 +280,6 @@ TEST_F(XattrTest, SetxattrNullValueAndNonzeroSize) { } TEST_F(XattrTest, SetxattrNullValueAndZeroSize) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; EXPECT_THAT(setxattr(path, name, nullptr, 0, /*flags=*/0), SyscallSucceeds()); @@ -296,10 +288,6 @@ TEST_F(XattrTest, SetxattrNullValueAndZeroSize) { } TEST_F(XattrTest, SetxattrValueTooLargeButOKSize) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; std::vector val(XATTR_SIZE_MAX + 1); @@ -316,10 +304,6 @@ TEST_F(XattrTest, SetxattrValueTooLargeButOKSize) { } TEST_F(XattrTest, SetxattrReplaceWithSmaller) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; std::vector val = {'a', 'a'}; @@ -335,10 +319,6 @@ TEST_F(XattrTest, SetxattrReplaceWithSmaller) { } TEST_F(XattrTest, SetxattrReplaceWithLarger) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; std::vector val = {'a', 'a'}; @@ -353,10 +333,6 @@ TEST_F(XattrTest, SetxattrReplaceWithLarger) { } TEST_F(XattrTest, SetxattrCreateFlag) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; EXPECT_THAT(setxattr(path, name, nullptr, 0, XATTR_CREATE), @@ -368,10 +344,6 @@ TEST_F(XattrTest, SetxattrCreateFlag) { } TEST_F(XattrTest, SetxattrReplaceFlag) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; EXPECT_THAT(setxattr(path, name, nullptr, 0, XATTR_REPLACE), @@ -384,10 +356,6 @@ TEST_F(XattrTest, SetxattrReplaceFlag) { } TEST_F(XattrTest, SetxattrInvalidFlags) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); int invalid_flags = 0xff; EXPECT_THAT(setxattr(path, nullptr, nullptr, 0, invalid_flags), @@ -395,10 +363,6 @@ TEST_F(XattrTest, SetxattrInvalidFlags) { } TEST_F(XattrTest, Getxattr) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; int val = 1234; @@ -411,10 +375,6 @@ TEST_F(XattrTest, Getxattr) { } TEST_F(XattrTest, GetxattrSizeSmallerThanValue) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; std::vector val = {'a', 'a'}; @@ -427,10 +387,6 @@ TEST_F(XattrTest, GetxattrSizeSmallerThanValue) { } TEST_F(XattrTest, GetxattrSizeLargerThanValue) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; char val = 'a'; @@ -446,10 +402,6 @@ TEST_F(XattrTest, GetxattrSizeLargerThanValue) { } TEST_F(XattrTest, GetxattrZeroSize) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; char val = 'a'; @@ -463,10 +415,6 @@ TEST_F(XattrTest, GetxattrZeroSize) { } TEST_F(XattrTest, GetxattrSizeTooLarge) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; char val = 'a'; @@ -483,10 +431,6 @@ TEST_F(XattrTest, GetxattrSizeTooLarge) { } TEST_F(XattrTest, GetxattrNullValue) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; char val = 'a'; @@ -498,10 +442,6 @@ TEST_F(XattrTest, GetxattrNullValue) { } TEST_F(XattrTest, GetxattrNullValueAndZeroSize) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - const char* path = test_file_name_.c_str(); const char name[] = "user.test"; char val = 'a'; @@ -518,35 +458,109 @@ TEST_F(XattrTest, GetxattrNullValueAndZeroSize) { } TEST_F(XattrTest, GetxattrNonexistentName) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); + const char* path = test_file_name_.c_str(); + const char name[] = "user.test"; + EXPECT_THAT(getxattr(path, name, nullptr, 0), SyscallFailsWithErrno(ENODATA)); +} + +TEST_F(XattrTest, Listxattr) { + const char* path = test_file_name_.c_str(); + const std::string name = "user.test"; + const std::string name2 = "user.test2"; + const std::string name3 = "user.test3"; + EXPECT_THAT(setxattr(path, name.c_str(), nullptr, 0, /*flags=*/0), + SyscallSucceeds()); + EXPECT_THAT(setxattr(path, name2.c_str(), nullptr, 0, /*flags=*/0), + SyscallSucceeds()); + EXPECT_THAT(setxattr(path, name3.c_str(), nullptr, 0, /*flags=*/0), + SyscallSucceeds()); + std::vector list(name.size() + 1 + name2.size() + 1 + name3.size() + 1); + char* buf = list.data(); + EXPECT_THAT(listxattr(path, buf, XATTR_SIZE_MAX), + SyscallSucceedsWithValue(list.size())); + + absl::flat_hash_set got = {}; + for (char* p = buf; p < buf + list.size(); p += strlen(p) + 1) { + got.insert(std::string{p}); + } + + absl::flat_hash_set expected = {name, name2, name3}; + EXPECT_EQ(got, expected); +} + +TEST_F(XattrTest, ListxattrNoXattrs) { + const char* path = test_file_name_.c_str(); + + std::vector list, expected; + EXPECT_THAT(listxattr(path, list.data(), sizeof(list)), + SyscallSucceedsWithValue(0)); + EXPECT_EQ(list, expected); + + // Listxattr should succeed if there are no attributes, even if the buffer + // passed in is a nullptr. + EXPECT_THAT(listxattr(path, nullptr, sizeof(list)), + SyscallSucceedsWithValue(0)); +} + +TEST_F(XattrTest, ListxattrNullBuffer) { const char* path = test_file_name_.c_str(); const char name[] = "user.test"; + EXPECT_THAT(setxattr(path, name, nullptr, 0, /*flags=*/0), SyscallSucceeds()); + + EXPECT_THAT(listxattr(path, nullptr, sizeof(name)), + SyscallFailsWithErrno(EFAULT)); +} + +TEST_F(XattrTest, ListxattrSizeTooSmall) { + const char* path = test_file_name_.c_str(); + const char name[] = "user.test"; + EXPECT_THAT(setxattr(path, name, nullptr, 0, /*flags=*/0), SyscallSucceeds()); + + char list[sizeof(name) - 1]; + EXPECT_THAT(listxattr(path, list, sizeof(list)), + SyscallFailsWithErrno(ERANGE)); +} + +TEST_F(XattrTest, ListxattrZeroSize) { + const char* path = test_file_name_.c_str(); + const char name[] = "user.test"; + EXPECT_THAT(setxattr(path, name, nullptr, 0, /*flags=*/0), SyscallSucceeds()); + EXPECT_THAT(listxattr(path, nullptr, 0), + SyscallSucceedsWithValue(sizeof(name))); +} + +TEST_F(XattrTest, RemoveXattr) { + const char* path = test_file_name_.c_str(); + const char name[] = "user.test"; + EXPECT_THAT(setxattr(path, name, nullptr, 0, /*flags=*/0), SyscallSucceeds()); + EXPECT_THAT(removexattr(path, name), SyscallSucceeds()); EXPECT_THAT(getxattr(path, name, nullptr, 0), SyscallFailsWithErrno(ENODATA)); } -TEST_F(XattrTest, LGetSetxattrOnSymlink) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); +TEST_F(XattrTest, RemoveXattrNonexistentName) { + const char* path = test_file_name_.c_str(); + const char name[] = "user.test"; + EXPECT_THAT(removexattr(path, name), SyscallFailsWithErrno(ENODATA)); +} +TEST_F(XattrTest, LXattrOnSymlink) { + const char name[] = "user.test"; TempPath dir = ASSERT_NO_ERRNO_AND_VALUE(TempPath::CreateDir()); TempPath link = ASSERT_NO_ERRNO_AND_VALUE( TempPath::CreateSymlinkTo(dir.path(), test_file_name_)); - EXPECT_THAT(lsetxattr(link.path().c_str(), nullptr, nullptr, 0, 0), + EXPECT_THAT(lsetxattr(link.path().c_str(), name, nullptr, 0, 0), SyscallFailsWithErrno(EPERM)); - EXPECT_THAT(lgetxattr(link.path().c_str(), nullptr, nullptr, 0), + EXPECT_THAT(lgetxattr(link.path().c_str(), name, nullptr, 0), SyscallFailsWithErrno(ENODATA)); + EXPECT_THAT(llistxattr(link.path().c_str(), nullptr, 0), + SyscallSucceedsWithValue(0)); + EXPECT_THAT(lremovexattr(link.path().c_str(), name), + SyscallFailsWithErrno(EPERM)); } -TEST_F(XattrTest, LGetSetxattrOnNonsymlink) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); - +TEST_F(XattrTest, LXattrOnNonsymlink) { const char* path = test_file_name_.c_str(); const char name[] = "user.test"; int val = 1234; @@ -558,13 +572,16 @@ TEST_F(XattrTest, LGetSetxattrOnNonsymlink) { EXPECT_THAT(lgetxattr(path, name, &buf, size), SyscallSucceedsWithValue(size)); EXPECT_EQ(buf, val); -} -TEST_F(XattrTest, FGetSetxattr) { - // TODO(gvisor.dev/issue/1636): Re-enable once list/remove xattr are - // supported, and get/set have been added pack to the syscall table. - SKIP_IF(IsRunningOnGvisor()); + char list[sizeof(name)]; + EXPECT_THAT(llistxattr(path, list, sizeof(list)), + SyscallSucceedsWithValue(sizeof(name))); + EXPECT_STREQ(list, name); + EXPECT_THAT(lremovexattr(path, name), SyscallSucceeds()); +} + +TEST_F(XattrTest, XattrWithFD) { const FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(test_file_name_.c_str(), 0)); const char name[] = "user.test"; @@ -577,6 +594,13 @@ TEST_F(XattrTest, FGetSetxattr) { EXPECT_THAT(fgetxattr(fd.get(), name, &buf, size), SyscallSucceedsWithValue(size)); EXPECT_EQ(buf, val); + + char list[sizeof(name)]; + EXPECT_THAT(flistxattr(fd.get(), list, sizeof(list)), + SyscallSucceedsWithValue(sizeof(name))); + EXPECT_STREQ(list, name); + + EXPECT_THAT(fremovexattr(fd.get(), name), SyscallSucceeds()); } } // namespace -- cgit v1.2.3 From a6f9361c2f7c5b46a200de1dc891a0ce059ad90e Mon Sep 17 00:00:00 2001 From: Adin Scannell Date: Mon, 10 Feb 2020 13:51:26 -0800 Subject: Add context to comments. PiperOrigin-RevId: 294295852 --- pkg/sentry/fs/gofer/attr.go | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/gofer/attr.go b/pkg/sentry/fs/gofer/attr.go index 71cccdc34..6db4b762d 100644 --- a/pkg/sentry/fs/gofer/attr.go +++ b/pkg/sentry/fs/gofer/attr.go @@ -88,8 +88,9 @@ func bsize(pattr p9.Attr) int64 { if pattr.BlockSize > 0 { return int64(pattr.BlockSize) } - // Some files may have no clue of their block size. Better not to report - // something misleading or buggy and have a safe default. + // Some files, particularly those that are not on a local file system, + // may have no clue of their block size. Better not to report something + // misleading or buggy and have a safe default. return usermem.PageSize } @@ -149,6 +150,7 @@ func links(valid p9.AttrMask, pattr p9.Attr) uint64 { } // This node is likely backed by a file system that doesn't support links. + // // We could readdir() and count children directories to provide an accurate // link count. However this may be expensive since the gofer may be backed by remote // storage. Instead, simply return 2 links for directories and 1 for everything else -- cgit v1.2.3 From 115898e368e4afe5418a7290d9545fafc7f6f25e Mon Sep 17 00:00:00 2001 From: Adin Scannell Date: Tue, 11 Feb 2020 11:37:12 -0800 Subject: Prevent DATA RACE in UnstableAttr. The slaveInodeOperations is currently copying the object when truncate is called (which is a no-op). This may result in a (unconsequential) data race when being modified concurrently. PiperOrigin-RevId: 294484276 --- pkg/sentry/fs/tty/slave.go | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/tty/slave.go b/pkg/sentry/fs/tty/slave.go index db55cdc48..6a2dbc576 100644 --- a/pkg/sentry/fs/tty/slave.go +++ b/pkg/sentry/fs/tty/slave.go @@ -73,7 +73,7 @@ func (si *slaveInodeOperations) Release(ctx context.Context) { } // Truncate implements fs.InodeOperations.Truncate. -func (slaveInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { +func (*slaveInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { return nil } -- cgit v1.2.3 From 6dced977eab69401a114257e386addb9cb03a39d Mon Sep 17 00:00:00 2001 From: gVisor bot Date: Tue, 11 Feb 2020 17:38:04 -0800 Subject: Ensure fsimpl/gofer.dentryPlatformFile.hostFileMapper is initialized. Fixes #1812. (The more direct cause of the deadlock is panic unsafety because the historically high cost of defer means that we avoid it in hot paths, including much of MM; defer is much cheaper as of Go 1.14, but still a measurable overhead.) PiperOrigin-RevId: 294560316 --- pkg/sentry/fs/fsutil/host_file_mapper.go | 17 +++++++++++------ pkg/sentry/fsimpl/gofer/regular_file.go | 5 +++++ 2 files changed, 16 insertions(+), 6 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/fsutil/host_file_mapper.go b/pkg/sentry/fs/fsutil/host_file_mapper.go index 67278aa86..e82afd112 100644 --- a/pkg/sentry/fs/fsutil/host_file_mapper.go +++ b/pkg/sentry/fs/fsutil/host_file_mapper.go @@ -65,13 +65,18 @@ type mapping struct { writable bool } -// NewHostFileMapper returns a HostFileMapper with no references or cached -// mappings. +// Init must be called on zero-value HostFileMappers before first use. +func (f *HostFileMapper) Init() { + f.refs = make(map[uint64]int32) + f.mappings = make(map[uint64]mapping) +} + +// NewHostFileMapper returns an initialized HostFileMapper allocated on the +// heap with no references or cached mappings. func NewHostFileMapper() *HostFileMapper { - return &HostFileMapper{ - refs: make(map[uint64]int32), - mappings: make(map[uint64]mapping), - } + f := &HostFileMapper{} + f.Init() + return f } // IncRefOn increments the reference count on all offsets in mr. diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go index 8e11e06b3..54c1031a7 100644 --- a/pkg/sentry/fsimpl/gofer/regular_file.go +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -571,6 +571,8 @@ func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpt default: panic(fmt.Sprintf("unknown InteropMode %v", d.fs.opts.interop)) } + // After this point, d may be used as a memmap.Mappable. + d.pf.hostFileMapperInitOnce.Do(d.pf.hostFileMapper.Init) return vfs.GenericConfigureMMap(&fd.vfsfd, d, opts) } @@ -799,6 +801,9 @@ type dentryPlatformFile struct { // If this dentry represents a regular file, and handle.fd >= 0, // hostFileMapper caches mappings of handle.fd. hostFileMapper fsutil.HostFileMapper + + // hostFileMapperInitOnce is used to lazily initialize hostFileMapper. + hostFileMapperInitOnce sync.Once } // IncRef implements platform.File.IncRef. -- cgit v1.2.3 From 4075de11be44372c454aae7f9650cdc814c52229 Mon Sep 17 00:00:00 2001 From: gVisor bot Date: Fri, 14 Feb 2020 11:11:55 -0800 Subject: Plumb VFS2 inside the Sentry - Added fsbridge package with interface that can be used to open and read from VFS1 and VFS2 files. - Converted ELF loader to use fsbridge - Added VFS2 types to FSContext - Added vfs.MountNamespace to ThreadGroup Updates #1623 PiperOrigin-RevId: 295183950 --- pkg/sentry/control/BUILD | 5 + pkg/sentry/control/proc.go | 127 +++++++++++++-- pkg/sentry/fs/proc/BUILD | 1 + pkg/sentry/fs/proc/task.go | 17 +- pkg/sentry/fsbridge/BUILD | 24 +++ pkg/sentry/fsbridge/bridge.go | 54 ++++++ pkg/sentry/fsbridge/fs.go | 181 +++++++++++++++++++++ pkg/sentry/fsbridge/vfs.go | 134 +++++++++++++++ pkg/sentry/fsimpl/devtmpfs/devtmpfs.go | 4 + pkg/sentry/fsimpl/gofer/filesystem.go | 5 +- pkg/sentry/fsimpl/gofer/gofer.go | 3 + pkg/sentry/fsimpl/kernfs/filesystem.go | 10 +- pkg/sentry/fsimpl/proc/BUILD | 1 + pkg/sentry/fsimpl/proc/filesystem.go | 18 +- pkg/sentry/fsimpl/proc/tasks_test.go | 17 +- pkg/sentry/fsimpl/sys/BUILD | 1 + pkg/sentry/fsimpl/sys/sys.go | 3 + pkg/sentry/fsimpl/sys/sys_test.go | 7 +- pkg/sentry/fsimpl/testutil/BUILD | 2 +- pkg/sentry/fsimpl/testutil/kernel.go | 24 +-- pkg/sentry/fsimpl/testutil/testutil.go | 12 +- pkg/sentry/fsimpl/tmpfs/filesystem.go | 12 +- pkg/sentry/fsimpl/tmpfs/tmpfs.go | 3 + pkg/sentry/kernel/BUILD | 2 + pkg/sentry/kernel/fs_context.go | 98 +++++++++-- pkg/sentry/kernel/kernel.go | 145 +++++++++++++---- pkg/sentry/kernel/task.go | 27 +++ pkg/sentry/kernel/task_clone.go | 11 +- pkg/sentry/kernel/task_context.go | 2 +- pkg/sentry/kernel/task_exit.go | 7 + pkg/sentry/kernel/task_log.go | 15 +- pkg/sentry/kernel/task_start.go | 49 +++--- pkg/sentry/kernel/thread_group.go | 6 +- pkg/sentry/loader/BUILD | 2 + pkg/sentry/loader/elf.go | 28 ++-- pkg/sentry/loader/interpreter.go | 6 +- pkg/sentry/loader/loader.go | 179 ++++++-------------- pkg/sentry/loader/vdso.go | 7 +- pkg/sentry/mm/BUILD | 2 +- pkg/sentry/mm/metadata.go | 10 +- pkg/sentry/mm/mm.go | 4 +- pkg/sentry/strace/strace.go | 28 ++++ pkg/sentry/syscalls/linux/BUILD | 1 + pkg/sentry/syscalls/linux/sys_prctl.go | 3 +- pkg/sentry/syscalls/linux/sys_thread.go | 17 +- .../syscalls/linux/vfs2/linux64_override_amd64.go | 106 ++++++++++++ pkg/sentry/vfs/BUILD | 1 + pkg/sentry/vfs/context.go | 7 +- pkg/sentry/vfs/mount.go | 10 +- pkg/sentry/vfs/options.go | 2 +- pkg/sentry/vfs/vfs.go | 5 +- runsc/boot/loader.go | 11 +- 52 files changed, 1134 insertions(+), 322 deletions(-) create mode 100644 pkg/sentry/fsbridge/BUILD create mode 100644 pkg/sentry/fsbridge/bridge.go create mode 100644 pkg/sentry/fsbridge/fs.go create mode 100644 pkg/sentry/fsbridge/vfs.go (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/control/BUILD b/pkg/sentry/control/BUILD index e69496477..d16d78aa5 100644 --- a/pkg/sentry/control/BUILD +++ b/pkg/sentry/control/BUILD @@ -16,10 +16,13 @@ go_library( ], deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/fd", + "//pkg/fspath", "//pkg/log", "//pkg/sentry/fs", "//pkg/sentry/fs/host", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", @@ -27,8 +30,10 @@ go_library( "//pkg/sentry/state", "//pkg/sentry/strace", "//pkg/sentry/usage", + "//pkg/sentry/vfs", "//pkg/sentry/watchdog", "//pkg/sync", + "//pkg/syserror", "//pkg/tcpip/link/sniffer", "//pkg/urpc", ], diff --git a/pkg/sentry/control/proc.go b/pkg/sentry/control/proc.go index ced51c66c..8973754c8 100644 --- a/pkg/sentry/control/proc.go +++ b/pkg/sentry/control/proc.go @@ -18,19 +18,26 @@ import ( "bytes" "encoding/json" "fmt" + "path" "sort" "strings" "text/tabwriter" "time" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/host" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/urpc" ) @@ -60,6 +67,12 @@ type ExecArgs struct { // process's MountNamespace. MountNamespace *fs.MountNamespace + // MountNamespaceVFS2 is the mount namespace to execute the new process in. + // A reference on MountNamespace must be held for the lifetime of the + // ExecArgs. If MountNamespace is nil, it will default to the init + // process's MountNamespace. + MountNamespaceVFS2 *vfs.MountNamespace + // WorkingDirectory defines the working directory for the new process. WorkingDirectory string `json:"wd"` @@ -150,6 +163,7 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI Envv: args.Envv, WorkingDirectory: args.WorkingDirectory, MountNamespace: args.MountNamespace, + MountNamespaceVFS2: args.MountNamespaceVFS2, Credentials: creds, FDTable: fdTable, Umask: 0022, @@ -166,24 +180,53 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI // be donated to the new process in CreateProcess. initArgs.MountNamespace.IncRef() } + if initArgs.MountNamespaceVFS2 != nil { + // initArgs must hold a reference on MountNamespaceVFS2, which will + // be donated to the new process in CreateProcess. + initArgs.MountNamespaceVFS2.IncRef() + } ctx := initArgs.NewContext(proc.Kernel) if initArgs.Filename == "" { - // Get the full path to the filename from the PATH env variable. - paths := fs.GetPath(initArgs.Envv) - mns := initArgs.MountNamespace - if mns == nil { - mns = proc.Kernel.GlobalInit().Leader().MountNamespace() - } - f, err := mns.ResolveExecutablePath(ctx, initArgs.WorkingDirectory, initArgs.Argv[0], paths) - if err != nil { - return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err) + if kernel.VFS2Enabled { + // Get the full path to the filename from the PATH env variable. + if initArgs.MountNamespaceVFS2 == nil { + // Set initArgs so that 'ctx' returns the namespace. + // + // MountNamespaceVFS2 adds a reference to the namespace, which is + // transferred to the new process. + initArgs.MountNamespaceVFS2 = proc.Kernel.GlobalInit().Leader().MountNamespaceVFS2() + } + + paths := fs.GetPath(initArgs.Envv) + vfsObj := proc.Kernel.VFS + file, err := ResolveExecutablePath(ctx, vfsObj, initArgs.WorkingDirectory, initArgs.Argv[0], paths) + if err != nil { + return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err) + } + initArgs.File = fsbridge.NewVFSFile(file) + } else { + // Get the full path to the filename from the PATH env variable. + paths := fs.GetPath(initArgs.Envv) + if initArgs.MountNamespace == nil { + // Set initArgs so that 'ctx' returns the namespace. + initArgs.MountNamespace = proc.Kernel.GlobalInit().Leader().MountNamespace() + + // initArgs must hold a reference on MountNamespace, which will + // be donated to the new process in CreateProcess. + initArgs.MountNamespaceVFS2.IncRef() + } + f, err := initArgs.MountNamespace.ResolveExecutablePath(ctx, initArgs.WorkingDirectory, initArgs.Argv[0], paths) + if err != nil { + return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err) + } + initArgs.Filename = f } - initArgs.Filename = f } mounter := fs.FileOwnerFromContext(ctx) + // TODO(gvisor.dev/issue/1623): Use host FD when supported in VFS2. var ttyFile *fs.File for appFD, hostFile := range args.FilePayload.Files { var appFile *fs.File @@ -411,3 +454,67 @@ func ttyName(tty *kernel.TTY) string { } return fmt.Sprintf("pts/%d", tty.Index) } + +// ResolveExecutablePath resolves the given executable name given a set of +// paths that might contain it. +func ResolveExecutablePath(ctx context.Context, vfsObj *vfs.VirtualFilesystem, wd, name string, paths []string) (*vfs.FileDescription, error) { + root := vfs.RootFromContext(ctx) + defer root.DecRef() + creds := auth.CredentialsFromContext(ctx) + + // Absolute paths can be used directly. + if path.IsAbs(name) { + return openExecutable(ctx, vfsObj, creds, root, name) + } + + // Paths with '/' in them should be joined to the working directory, or + // to the root if working directory is not set. + if strings.IndexByte(name, '/') > 0 { + if len(wd) == 0 { + wd = "/" + } + if !path.IsAbs(wd) { + return nil, fmt.Errorf("working directory %q must be absolute", wd) + } + return openExecutable(ctx, vfsObj, creds, root, path.Join(wd, name)) + } + + // Otherwise, we must lookup the name in the paths, starting from the + // calling context's root directory. + for _, p := range paths { + if !path.IsAbs(p) { + // Relative paths aren't safe, no one should be using them. + log.Warningf("Skipping relative path %q in $PATH", p) + continue + } + + binPath := path.Join(p, name) + f, err := openExecutable(ctx, vfsObj, creds, root, binPath) + if err != nil { + return nil, err + } + if f == nil { + continue // Not found/no access. + } + return f, nil + } + return nil, syserror.ENOENT +} + +func openExecutable(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, root vfs.VirtualDentry, path string) (*vfs.FileDescription, error) { + pop := vfs.PathOperation{ + Root: root, + Start: root, // binPath is absolute, Start can be anything. + Path: fspath.Parse(path), + FollowFinalSymlink: true, + } + opts := &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, + } + f, err := vfsObj.OpenAt(ctx, creds, &pop, opts) + if err == syserror.ENOENT || err == syserror.EACCES { + return nil, nil + } + return f, err +} diff --git a/pkg/sentry/fs/proc/BUILD b/pkg/sentry/fs/proc/BUILD index 280093c5e..77c2c5c0e 100644 --- a/pkg/sentry/fs/proc/BUILD +++ b/pkg/sentry/fs/proc/BUILD @@ -36,6 +36,7 @@ go_library( "//pkg/sentry/fs/proc/device", "//pkg/sentry/fs/proc/seqfile", "//pkg/sentry/fs/ramfs", + "//pkg/sentry/fsbridge", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index ca020e11e..8ab8d8a02 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -28,6 +28,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs/proc/device" "gvisor.dev/gvisor/pkg/sentry/fs/proc/seqfile" "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -249,7 +250,7 @@ func newExe(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { return newProcInode(t, exeSymlink, msrc, fs.Symlink, t) } -func (e *exe) executable() (d *fs.Dirent, err error) { +func (e *exe) executable() (file fsbridge.File, err error) { e.t.WithMuLocked(func(t *kernel.Task) { mm := t.MemoryManager() if mm == nil { @@ -262,8 +263,8 @@ func (e *exe) executable() (d *fs.Dirent, err error) { // The MemoryManager may be destroyed, in which case // MemoryManager.destroy will simply set the executable to nil // (with locks held). - d = mm.Executable() - if d == nil { + file = mm.Executable() + if file == nil { err = syserror.ENOENT } }) @@ -283,15 +284,7 @@ func (e *exe) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { } defer exec.DecRef() - root := fs.RootFromContext(ctx) - if root == nil { - // This doesn't correspond to anything in Linux because the vfs is - // global there. - return "", syserror.EINVAL - } - defer root.DecRef() - n, _ := exec.FullName(root) - return n, nil + return exec.PathnameWithDeleted(ctx), nil } // namespaceSymlink represents a symlink in the namespacefs, such as the files diff --git a/pkg/sentry/fsbridge/BUILD b/pkg/sentry/fsbridge/BUILD new file mode 100644 index 000000000..6c798f0bd --- /dev/null +++ b/pkg/sentry/fsbridge/BUILD @@ -0,0 +1,24 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "fsbridge", + srcs = [ + "bridge.go", + "fs.go", + "vfs.go", + ], + visibility = ["//pkg/sentry:internal"], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/sentry/fs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", + ], +) diff --git a/pkg/sentry/fsbridge/bridge.go b/pkg/sentry/fsbridge/bridge.go new file mode 100644 index 000000000..8e7590721 --- /dev/null +++ b/pkg/sentry/fsbridge/bridge.go @@ -0,0 +1,54 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package fsbridge provides common interfaces to bridge between VFS1 and VFS2 +// files. +package fsbridge + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// File provides a common interface to bridge between VFS1 and VFS2 files. +type File interface { + // PathnameWithDeleted returns an absolute pathname to vd, consistent with + // Linux's d_path(). In particular, if vd.Dentry() has been disowned, + // PathnameWithDeleted appends " (deleted)" to the returned pathname. + PathnameWithDeleted(ctx context.Context) string + + // ReadFull read all contents from the file. + ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) + + // ConfigureMMap mutates opts to implement mmap(2) for the file. + ConfigureMMap(context.Context, *memmap.MMapOpts) error + + // Type returns the file type, e.g. linux.S_IFREG. + Type(context.Context) (linux.FileMode, error) + + // IncRef increments reference. + IncRef() + + // DecRef decrements reference. + DecRef() +} + +// Lookup provides a common interface to open files. +type Lookup interface { + // OpenPath opens a file. + OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, remainingTraversals *uint, resolveFinal bool) (File, error) +} diff --git a/pkg/sentry/fsbridge/fs.go b/pkg/sentry/fsbridge/fs.go new file mode 100644 index 000000000..093ce1fb3 --- /dev/null +++ b/pkg/sentry/fsbridge/fs.go @@ -0,0 +1,181 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fsbridge + +import ( + "io" + "strings" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// fsFile implements File interface over fs.File. +// +// +stateify savable +type fsFile struct { + file *fs.File +} + +var _ File = (*fsFile)(nil) + +// NewFSFile creates a new File over fs.File. +func NewFSFile(file *fs.File) File { + return &fsFile{file: file} +} + +// PathnameWithDeleted implements File. +func (f *fsFile) PathnameWithDeleted(ctx context.Context) string { + root := fs.RootFromContext(ctx) + if root == nil { + // This doesn't correspond to anything in Linux because the vfs is + // global there. + return "" + } + defer root.DecRef() + + name, _ := f.file.Dirent.FullName(root) + return name +} + +// ReadFull implements File. +func (f *fsFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) { + var total int64 + for dst.NumBytes() > 0 { + n, err := f.file.Preadv(ctx, dst, offset+total) + total += n + if err == io.EOF && total != 0 { + return total, io.ErrUnexpectedEOF + } else if err != nil { + return total, err + } + dst = dst.DropFirst64(n) + } + return total, nil +} + +// ConfigureMMap implements File. +func (f *fsFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return f.file.ConfigureMMap(ctx, opts) +} + +// Type implements File. +func (f *fsFile) Type(context.Context) (linux.FileMode, error) { + return linux.FileMode(f.file.Dirent.Inode.StableAttr.Type.LinuxType()), nil +} + +// IncRef implements File. +func (f *fsFile) IncRef() { + f.file.IncRef() +} + +// DecRef implements File. +func (f *fsFile) DecRef() { + f.file.DecRef() +} + +// fsLookup implements Lookup interface using fs.File. +// +// +stateify savable +type fsLookup struct { + mntns *fs.MountNamespace + + root *fs.Dirent + workingDir *fs.Dirent +} + +var _ Lookup = (*fsLookup)(nil) + +// NewFSLookup creates a new Lookup using VFS1. +func NewFSLookup(mntns *fs.MountNamespace, root, workingDir *fs.Dirent) Lookup { + return &fsLookup{ + mntns: mntns, + root: root, + workingDir: workingDir, + } +} + +// OpenPath implements Lookup. +func (l *fsLookup) OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, remainingTraversals *uint, resolveFinal bool) (File, error) { + var d *fs.Dirent + var err error + if resolveFinal { + d, err = l.mntns.FindInode(ctx, l.root, l.workingDir, path, remainingTraversals) + } else { + d, err = l.mntns.FindLink(ctx, l.root, l.workingDir, path, remainingTraversals) + } + if err != nil { + return nil, err + } + defer d.DecRef() + + if !resolveFinal && fs.IsSymlink(d.Inode.StableAttr) { + return nil, syserror.ELOOP + } + + fsPerm := openOptionsToPermMask(&opts) + if err := d.Inode.CheckPermission(ctx, fsPerm); err != nil { + return nil, err + } + + // If they claim it's a directory, then make sure. + if strings.HasSuffix(path, "/") { + if d.Inode.StableAttr.Type != fs.Directory { + return nil, syserror.ENOTDIR + } + } + + if opts.FileExec && d.Inode.StableAttr.Type != fs.RegularFile { + ctx.Infof("%q is not a regular file: %v", path, d.Inode.StableAttr.Type) + return nil, syserror.EACCES + } + + f, err := d.Inode.GetFile(ctx, d, flagsToFileFlags(opts.Flags)) + if err != nil { + return nil, err + } + + return &fsFile{file: f}, nil +} + +func openOptionsToPermMask(opts *vfs.OpenOptions) fs.PermMask { + mode := opts.Flags & linux.O_ACCMODE + return fs.PermMask{ + Read: mode == linux.O_RDONLY || mode == linux.O_RDWR, + Write: mode == linux.O_WRONLY || mode == linux.O_RDWR, + Execute: opts.FileExec, + } +} + +func flagsToFileFlags(flags uint32) fs.FileFlags { + return fs.FileFlags{ + Direct: flags&linux.O_DIRECT != 0, + DSync: flags&(linux.O_DSYNC|linux.O_SYNC) != 0, + Sync: flags&linux.O_SYNC != 0, + NonBlocking: flags&linux.O_NONBLOCK != 0, + Read: (flags & linux.O_ACCMODE) != linux.O_WRONLY, + Write: (flags & linux.O_ACCMODE) != linux.O_RDONLY, + Append: flags&linux.O_APPEND != 0, + Directory: flags&linux.O_DIRECTORY != 0, + Async: flags&linux.O_ASYNC != 0, + LargeFile: flags&linux.O_LARGEFILE != 0, + Truncate: flags&linux.O_TRUNC != 0, + } +} diff --git a/pkg/sentry/fsbridge/vfs.go b/pkg/sentry/fsbridge/vfs.go new file mode 100644 index 000000000..e657c39bc --- /dev/null +++ b/pkg/sentry/fsbridge/vfs.go @@ -0,0 +1,134 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fsbridge + +import ( + "io" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// fsFile implements File interface over vfs.FileDescription. +// +// +stateify savable +type vfsFile struct { + file *vfs.FileDescription +} + +var _ File = (*vfsFile)(nil) + +// NewVFSFile creates a new File over fs.File. +func NewVFSFile(file *vfs.FileDescription) File { + return &vfsFile{file: file} +} + +// PathnameWithDeleted implements File. +func (f *vfsFile) PathnameWithDeleted(ctx context.Context) string { + root := vfs.RootFromContext(ctx) + defer root.DecRef() + + vfsObj := f.file.VirtualDentry().Mount().Filesystem().VirtualFilesystem() + name, _ := vfsObj.PathnameWithDeleted(ctx, root, f.file.VirtualDentry()) + return name +} + +// ReadFull implements File. +func (f *vfsFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) { + var total int64 + for dst.NumBytes() > 0 { + n, err := f.file.PRead(ctx, dst, offset+total, vfs.ReadOptions{}) + total += n + if err == io.EOF && total != 0 { + return total, io.ErrUnexpectedEOF + } else if err != nil { + return total, err + } + dst = dst.DropFirst64(n) + } + return total, nil +} + +// ConfigureMMap implements File. +func (f *vfsFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return f.file.ConfigureMMap(ctx, opts) +} + +// Type implements File. +func (f *vfsFile) Type(ctx context.Context) (linux.FileMode, error) { + stat, err := f.file.Stat(ctx, vfs.StatOptions{}) + if err != nil { + return 0, err + } + return linux.FileMode(stat.Mode).FileType(), nil +} + +// IncRef implements File. +func (f *vfsFile) IncRef() { + f.file.IncRef() +} + +// DecRef implements File. +func (f *vfsFile) DecRef() { + f.file.DecRef() +} + +// fsLookup implements Lookup interface using fs.File. +// +// +stateify savable +type vfsLookup struct { + mntns *vfs.MountNamespace + + root vfs.VirtualDentry + workingDir vfs.VirtualDentry +} + +var _ Lookup = (*vfsLookup)(nil) + +// NewVFSLookup creates a new Lookup using VFS2. +func NewVFSLookup(mntns *vfs.MountNamespace, root, workingDir vfs.VirtualDentry) Lookup { + return &vfsLookup{ + mntns: mntns, + root: root, + workingDir: workingDir, + } +} + +// OpenPath implements Lookup. +// +// remainingTraversals is not configurable in VFS2, all callers are using the +// default anyways. +// +// TODO(gvisor.dev/issue/1623): Check mount has read and exec permission. +func (l *vfsLookup) OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, _ *uint, resolveFinal bool) (File, error) { + vfsObj := l.mntns.Root().Mount().Filesystem().VirtualFilesystem() + creds := auth.CredentialsFromContext(ctx) + pop := &vfs.PathOperation{ + Root: l.root, + Start: l.root, + Path: fspath.Parse(path), + FollowFinalSymlink: resolveFinal, + } + fd, err := vfsObj.OpenAt(ctx, creds, pop, &opts) + if err != nil { + return nil, err + } + return &vfsFile{file: fd}, nil +} diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go index e03a0c665..abd4f24e7 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go @@ -28,6 +28,9 @@ import ( "gvisor.dev/gvisor/pkg/sync" ) +// Name is the default filesystem name. +const Name = "devtmpfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct { initOnce sync.Once @@ -107,6 +110,7 @@ func (a *Accessor) wrapContext(ctx context.Context) *accessorContext { func (ac *accessorContext) Value(key interface{}) interface{} { switch key { case vfs.CtxMountNamespace: + ac.a.mntns.IncRef() return ac.a.mntns case vfs.CtxRoot: ac.a.root.IncRef() diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go index 138adb9f7..5cfb0dc4c 100644 --- a/pkg/sentry/fsimpl/gofer/filesystem.go +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -400,6 +400,7 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b } vfsObj := rp.VirtualFilesystem() mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() parent.dirMu.Lock() defer parent.dirMu.Unlock() childVFSD := parent.vfsd.Child(name) @@ -934,7 +935,9 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if oldParent == newParent && oldName == newName { return nil } - if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), &renamed.vfsd, replacedVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil { return err } if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index d0552bd99..d00850e25 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -52,6 +52,9 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// Name is the default filesystem name. +const Name = "9p" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index ee98eb66a..292f58afd 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -544,6 +544,7 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() virtfs := rp.VirtualFilesystem() srcDirDentry := srcDirVFSD.Impl().(*Dentry) @@ -595,7 +596,10 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error parentDentry := vfsd.Parent().Impl().(*Dentry) parentDentry.dirMu.Lock() defer parentDentry.dirMu.Unlock() - if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { + + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil { @@ -697,7 +701,9 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error parentDentry := vfsd.Parent().Impl().(*Dentry) parentDentry.dirMu.Lock() defer parentDentry.dirMu.Unlock() - if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil { diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index 12aac2e6a..a83245866 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -14,6 +14,7 @@ go_library( "tasks_net.go", "tasks_sys.go", ], + visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", diff --git a/pkg/sentry/fsimpl/proc/filesystem.go b/pkg/sentry/fsimpl/proc/filesystem.go index 11477b6a9..5c19d5522 100644 --- a/pkg/sentry/fsimpl/proc/filesystem.go +++ b/pkg/sentry/fsimpl/proc/filesystem.go @@ -26,15 +26,18 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" ) -// procFSType is the factory class for procfs. +// Name is the default filesystem name. +const Name = "proc" + +// FilesystemType is the factory class for procfs. // // +stateify savable -type procFSType struct{} +type FilesystemType struct{} -var _ vfs.FilesystemType = (*procFSType)(nil) +var _ vfs.FilesystemType = (*FilesystemType)(nil) // GetFilesystem implements vfs.FilesystemType. -func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +func (ft *FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { k := kernel.KernelFromContext(ctx) if k == nil { return nil, nil, fmt.Errorf("procfs requires a kernel") @@ -47,12 +50,13 @@ func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFile procfs := &kernfs.Filesystem{} procfs.VFSFilesystem().Init(vfsObj, procfs) - var data *InternalData + var cgroups map[string]string if opts.InternalData != nil { - data = opts.InternalData.(*InternalData) + data := opts.InternalData.(*InternalData) + cgroups = data.Cgroups } - _, dentry := newTasksInode(procfs, k, pidns, data.Cgroups) + _, dentry := newTasksInode(procfs, k, pidns, cgroups) return procfs.VFSFilesystem(), dentry.VFSDentry(), nil } diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index 6fc3524db..96c72cbc9 100644 --- a/pkg/sentry/fsimpl/proc/tasks_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_test.go @@ -90,8 +90,7 @@ func setup(t *testing.T) *testutil.System { ctx := k.SupervisorContext() creds := auth.CredentialsFromContext(ctx) - vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("procfs", &procFSType{}, &vfs.RegisterFilesystemTypeOptions{ + k.VFS.MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) fsOpts := vfs.GetFilesystemOptions{ @@ -102,11 +101,11 @@ func setup(t *testing.T) *testutil.System { }, }, } - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "procfs", &fsOpts) + mntns, err := k.VFS.NewMountNamespace(ctx, creds, "", Name, &fsOpts) if err != nil { t.Fatalf("NewMountNamespace(): %v", err) } - return testutil.NewSystem(ctx, t, vfsObj, mntns) + return testutil.NewSystem(ctx, t, k.VFS, mntns) } func TestTasksEmpty(t *testing.T) { @@ -131,7 +130,7 @@ func TestTasks(t *testing.T) { var tasks []*kernel.Task for i := 0; i < 5; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc) + task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } @@ -213,7 +212,7 @@ func TestTasksOffset(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) for i := 0; i < 3; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - if _, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc); err != nil { + if _, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root); err != nil { t.Fatalf("CreateTask(): %v", err) } } @@ -337,7 +336,7 @@ func TestTask(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - _, err := testutil.CreateTask(s.Ctx, "name", tc) + _, err := testutil.CreateTask(s.Ctx, "name", tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } @@ -352,7 +351,7 @@ func TestProcSelf(t *testing.T) { k := kernel.KernelFromContext(s.Ctx) tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, "name", tc) + task, err := testutil.CreateTask(s.Ctx, "name", tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } @@ -433,7 +432,7 @@ func TestTree(t *testing.T) { var tasks []*kernel.Task for i := 0; i < 5; i++ { tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) - task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc) + task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root) if err != nil { t.Fatalf("CreateTask(): %v", err) } diff --git a/pkg/sentry/fsimpl/sys/BUILD b/pkg/sentry/fsimpl/sys/BUILD index 66c0d8bc8..a741e2bb6 100644 --- a/pkg/sentry/fsimpl/sys/BUILD +++ b/pkg/sentry/fsimpl/sys/BUILD @@ -7,6 +7,7 @@ go_library( srcs = [ "sys.go", ], + visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", diff --git a/pkg/sentry/fsimpl/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go index d693fceae..c36c4fa11 100644 --- a/pkg/sentry/fsimpl/sys/sys.go +++ b/pkg/sentry/fsimpl/sys/sys.go @@ -28,6 +28,9 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// Name is the default filesystem name. +const Name = "sysfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} diff --git a/pkg/sentry/fsimpl/sys/sys_test.go b/pkg/sentry/fsimpl/sys/sys_test.go index 8b1cf0bd0..5d1ba5867 100644 --- a/pkg/sentry/fsimpl/sys/sys_test.go +++ b/pkg/sentry/fsimpl/sys/sys_test.go @@ -34,16 +34,15 @@ func newTestSystem(t *testing.T) *testutil.System { } ctx := k.SupervisorContext() creds := auth.CredentialsFromContext(ctx) - v := vfs.New() - v.MustRegisterFilesystemType("sysfs", sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + k.VFS.MustRegisterFilesystemType(sys.Name, sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mns, err := v.NewMountNamespace(ctx, creds, "", "sysfs", &vfs.GetFilesystemOptions{}) + mns, err := k.VFS.NewMountNamespace(ctx, creds, "", sys.Name, &vfs.GetFilesystemOptions{}) if err != nil { t.Fatalf("Failed to create new mount namespace: %v", err) } - return testutil.NewSystem(ctx, t, v, mns) + return testutil.NewSystem(ctx, t, k.VFS, mns) } func TestReadCPUFile(t *testing.T) { diff --git a/pkg/sentry/fsimpl/testutil/BUILD b/pkg/sentry/fsimpl/testutil/BUILD index efd5974c4..e4f36f4ae 100644 --- a/pkg/sentry/fsimpl/testutil/BUILD +++ b/pkg/sentry/fsimpl/testutil/BUILD @@ -16,7 +16,7 @@ go_library( "//pkg/cpuid", "//pkg/fspath", "//pkg/memutil", - "//pkg/sentry/fs", + "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/sched", diff --git a/pkg/sentry/fsimpl/testutil/kernel.go b/pkg/sentry/fsimpl/testutil/kernel.go index 89f8c4915..a91b3ec4d 100644 --- a/pkg/sentry/fsimpl/testutil/kernel.go +++ b/pkg/sentry/fsimpl/testutil/kernel.go @@ -24,7 +24,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/memutil" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" @@ -33,6 +33,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/sentry/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" // Platforms are plugable. _ "gvisor.dev/gvisor/pkg/sentry/platform/kvm" @@ -99,26 +100,27 @@ func Boot() (*kernel.Kernel, error) { return nil, fmt.Errorf("initializing kernel: %v", err) } - ctx := k.SupervisorContext() + kernel.VFS2Enabled = true + + vfsObj := vfs.New() + vfsObj.MustRegisterFilesystemType(tmpfs.Name, &tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + AllowUserList: true, + }) + k.VFS = vfsObj - // Create mount namespace without root as it's the minimum required to create - // the global thread group. - mntns, err := fs.NewMountNamespace(ctx, nil) - if err != nil { - return nil, err - } ls, err := limits.NewLinuxLimitSet() if err != nil { return nil, err } - tg := k.NewThreadGroup(mntns, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls) + tg := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls) k.TestOnly_SetGlobalInit(tg) return k, nil } // CreateTask creates a new bare bones task for tests. -func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup) (*kernel.Task, error) { +func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns *vfs.MountNamespace, root, cwd vfs.VirtualDentry) (*kernel.Task, error) { k := kernel.KernelFromContext(ctx) config := &kernel.TaskConfig{ Kernel: k, @@ -129,6 +131,8 @@ func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup) (*kern UTSNamespace: kernel.UTSNamespaceFromContext(ctx), IPCNamespace: kernel.IPCNamespaceFromContext(ctx), AbstractSocketNamespace: kernel.NewAbstractSocketNamespace(), + MountNamespaceVFS2: mntns, + FSContext: kernel.NewFSContextVFS2(root, cwd, 0022), } return k.TaskSet().NewTask(config) } diff --git a/pkg/sentry/fsimpl/testutil/testutil.go b/pkg/sentry/fsimpl/testutil/testutil.go index 69fd84ddd..b97e3534a 100644 --- a/pkg/sentry/fsimpl/testutil/testutil.go +++ b/pkg/sentry/fsimpl/testutil/testutil.go @@ -41,12 +41,12 @@ type System struct { Creds *auth.Credentials VFS *vfs.VirtualFilesystem Root vfs.VirtualDentry - mns *vfs.MountNamespace + MntNs *vfs.MountNamespace } // NewSystem constructs a System. // -// Precondition: Caller must hold a reference on mns, whose ownership +// Precondition: Caller must hold a reference on MntNs, whose ownership // is transferred to the new System. func NewSystem(ctx context.Context, t *testing.T, v *vfs.VirtualFilesystem, mns *vfs.MountNamespace) *System { s := &System{ @@ -54,7 +54,7 @@ func NewSystem(ctx context.Context, t *testing.T, v *vfs.VirtualFilesystem, mns Ctx: ctx, Creds: auth.CredentialsFromContext(ctx), VFS: v, - mns: mns, + MntNs: mns, Root: mns.Root(), } return s @@ -75,7 +75,7 @@ func (s *System) WithSubtest(t *testing.T) *System { Ctx: s.Ctx, Creds: s.Creds, VFS: s.VFS, - mns: s.mns, + MntNs: s.MntNs, Root: s.Root, } } @@ -90,7 +90,7 @@ func (s *System) WithTemporaryContext(ctx context.Context) *System { Ctx: ctx, Creds: s.Creds, VFS: s.VFS, - mns: s.mns, + MntNs: s.MntNs, Root: s.Root, } } @@ -98,7 +98,7 @@ func (s *System) WithTemporaryContext(ctx context.Context) *System { // Destroy release resources associated with a test system. func (s *System) Destroy() { s.Root.DecRef() - s.mns.DecRef() // Reference on mns passed to NewSystem. + s.MntNs.DecRef() // Reference on MntNs passed to NewSystem. } // ReadToEnd reads the contents of fd until EOF to a string. diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index 8785452b6..7f7b791c4 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -486,7 +486,9 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa vfsObj := rp.VirtualFilesystem() oldParentDir := oldParent.inode.impl.(*directory) newParentDir := newParent.inode.impl.(*directory) - if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), renamedVFSD, replacedVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareRenameDentry(mntns, renamedVFSD, replacedVFSD); err != nil { return err } if replaced != nil { @@ -543,7 +545,9 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error } defer mnt.EndWrite() vfsObj := rp.VirtualFilesystem() - if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { return err } parent.inode.impl.(*directory).childList.Remove(child) @@ -631,7 +635,9 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error } defer mnt.EndWrite() vfsObj := rp.VirtualFilesystem() - if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef() + if err := vfsObj.PrepareDeleteDentry(mntns, childVFSD); err != nil { return err } parent.inode.impl.(*directory).childList.Remove(child) diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go index 2108d0f4d..c5bb17562 100644 --- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -40,6 +40,9 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// Name is the default filesystem name. +const Name = "tmpfs" + // FilesystemType implements vfs.FilesystemType. type FilesystemType struct{} diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index 2231d6973..46306945f 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -157,6 +157,7 @@ go_library( "//pkg/context", "//pkg/cpuid", "//pkg/eventchannel", + "//pkg/fspath", "//pkg/log", "//pkg/metric", "//pkg/refs", @@ -167,6 +168,7 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/fs/timerfd", + "//pkg/sentry/fsbridge", "//pkg/sentry/hostcpu", "//pkg/sentry/inet", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go index 2448c1d99..7218aa24e 100644 --- a/pkg/sentry/kernel/fs_context.go +++ b/pkg/sentry/kernel/fs_context.go @@ -19,6 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" ) @@ -37,10 +38,16 @@ type FSContext struct { // destroyed. root *fs.Dirent + // rootVFS2 is the filesystem root. + rootVFS2 vfs.VirtualDentry + // cwd is the current working directory. Will be nil iff the FSContext // has been destroyed. cwd *fs.Dirent + // cwdVFS2 is the current working directory. + cwdVFS2 vfs.VirtualDentry + // umask is the current file mode creation mask. When a thread using this // context invokes a syscall that creates a file, bits set in umask are // removed from the permissions that the file is created with. @@ -60,6 +67,19 @@ func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext { return &f } +// NewFSContextVFS2 returns a new filesystem context. +func NewFSContextVFS2(root, cwd vfs.VirtualDentry, umask uint) *FSContext { + root.IncRef() + cwd.IncRef() + f := FSContext{ + rootVFS2: root, + cwdVFS2: cwd, + umask: umask, + } + f.EnableLeakCheck("kernel.FSContext") + return &f +} + // destroy is the destructor for an FSContext. // // This will call DecRef on both root and cwd Dirents. If either call to @@ -75,11 +95,17 @@ func (f *FSContext) destroy() { f.mu.Lock() defer f.mu.Unlock() - f.root.DecRef() - f.root = nil - - f.cwd.DecRef() - f.cwd = nil + if VFS2Enabled { + f.rootVFS2.DecRef() + f.rootVFS2 = vfs.VirtualDentry{} + f.cwdVFS2.DecRef() + f.cwdVFS2 = vfs.VirtualDentry{} + } else { + f.root.DecRef() + f.root = nil + f.cwd.DecRef() + f.cwd = nil + } } // DecRef implements RefCounter.DecRef with destructor f.destroy. @@ -93,12 +119,21 @@ func (f *FSContext) DecRef() { func (f *FSContext) Fork() *FSContext { f.mu.Lock() defer f.mu.Unlock() - f.cwd.IncRef() - f.root.IncRef() + + if VFS2Enabled { + f.cwdVFS2.IncRef() + f.rootVFS2.IncRef() + } else { + f.cwd.IncRef() + f.root.IncRef() + } + return &FSContext{ - cwd: f.cwd, - root: f.root, - umask: f.umask, + cwd: f.cwd, + root: f.root, + cwdVFS2: f.cwdVFS2, + rootVFS2: f.rootVFS2, + umask: f.umask, } } @@ -109,12 +144,23 @@ func (f *FSContext) Fork() *FSContext { func (f *FSContext) WorkingDirectory() *fs.Dirent { f.mu.Lock() defer f.mu.Unlock() - if f.cwd != nil { - f.cwd.IncRef() - } + + f.cwd.IncRef() return f.cwd } +// WorkingDirectoryVFS2 returns the current working directory. +// +// This will return nil if called after destroy(), otherwise it will return a +// Dirent with a reference taken. +func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry { + f.mu.Lock() + defer f.mu.Unlock() + + f.cwdVFS2.IncRef() + return f.cwdVFS2 +} + // SetWorkingDirectory sets the current working directory. // This will take an extra reference on the Dirent. // @@ -137,6 +183,20 @@ func (f *FSContext) SetWorkingDirectory(d *fs.Dirent) { old.DecRef() } +// SetWorkingDirectoryVFS2 sets the current working directory. +// This will take an extra reference on the VirtualDentry. +// +// This is not a valid call after destroy. +func (f *FSContext) SetWorkingDirectoryVFS2(d vfs.VirtualDentry) { + f.mu.Lock() + defer f.mu.Unlock() + + old := f.cwdVFS2 + f.cwdVFS2 = d + d.IncRef() + old.DecRef() +} + // RootDirectory returns the current filesystem root. // // This will return nil if called after destroy(), otherwise it will return a @@ -150,6 +210,18 @@ func (f *FSContext) RootDirectory() *fs.Dirent { return f.root } +// RootDirectoryVFS2 returns the current filesystem root. +// +// This will return nil if called after destroy(), otherwise it will return a +// Dirent with a reference taken. +func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry { + f.mu.Lock() + defer f.mu.Unlock() + + f.rootVFS2.IncRef() + return f.rootVFS2 +} + // SetRootDirectory sets the root directory. // This will take an extra reference on the Dirent. // diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 3ee760ba2..2665f057c 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -43,11 +43,13 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/eventchannel" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/hostcpu" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -71,6 +73,10 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" ) +// VFS2Enabled is set to true when VFS2 is enabled. Added as a global for allow +// easy access everywhere. To be removed once VFS2 becomes the default. +var VFS2Enabled = false + // Kernel represents an emulated Linux kernel. It must be initialized by calling // Init() or LoadFrom(). // @@ -238,6 +244,9 @@ type Kernel struct { // SpecialOpts contains special kernel options. SpecialOpts + + // VFS keeps the filesystem state used across the kernel. + VFS *vfs.VirtualFilesystem } // InitKernelArgs holds arguments to Init. @@ -624,7 +633,7 @@ type CreateProcessArgs struct { // File is a passed host FD pointing to a file to load as the init binary. // // This is checked if and only if Filename is "". - File *fs.File + File fsbridge.File // Argvv is a list of arguments. Argv []string @@ -673,6 +682,13 @@ type CreateProcessArgs struct { // increment it). MountNamespace *fs.MountNamespace + // MountNamespaceVFS2 optionally contains the mount namespace for this + // process. If nil, the init process's mount namespace is used. + // + // Anyone setting MountNamespaceVFS2 must donate a reference (i.e. + // increment it). + MountNamespaceVFS2 *vfs.MountNamespace + // ContainerID is the container that the process belongs to. ContainerID string } @@ -711,11 +727,22 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { return ctx.args.Credentials case fs.CtxRoot: if ctx.args.MountNamespace != nil { - // MountNamespace.Root() will take a reference on the root - // dirent for us. + // MountNamespace.Root() will take a reference on the root dirent for us. return ctx.args.MountNamespace.Root() } return nil + case vfs.CtxRoot: + if ctx.args.MountNamespaceVFS2 == nil { + return nil + } + // MountNamespaceVFS2.Root() takes a reference on the root dirent for us. + return ctx.args.MountNamespaceVFS2.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2 takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter case ktime.CtxRealtimeClock: @@ -757,34 +784,77 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, defer k.extMu.Unlock() log.Infof("EXEC: %v", args.Argv) - // Grab the mount namespace. - mounts := args.MountNamespace - if mounts == nil { - mounts = k.GlobalInit().Leader().MountNamespace() - mounts.IncRef() - } - - tg := k.NewThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) ctx := args.NewContext(k) - // Get the root directory from the MountNamespace. - root := mounts.Root() - // The call to newFSContext below will take a reference on root, so we - // don't need to hold this one. - defer root.DecRef() - - // Grab the working directory. - remainingTraversals := uint(args.MaxSymlinkTraversals) - wd := root // Default. - if args.WorkingDirectory != "" { - var err error - wd, err = mounts.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) - if err != nil { - return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + var ( + opener fsbridge.Lookup + fsContext *FSContext + mntns *fs.MountNamespace + ) + + if VFS2Enabled { + mntnsVFS2 := args.MountNamespaceVFS2 + if mntnsVFS2 == nil { + // MountNamespaceVFS2 adds a reference to the namespace, which is + // transferred to the new process. + mntnsVFS2 = k.GlobalInit().Leader().MountNamespaceVFS2() + } + // Get the root directory from the MountNamespace. + root := args.MountNamespaceVFS2.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + wd := root // Default. + if args.WorkingDirectory != "" { + pop := vfs.PathOperation{ + Root: root, + Start: wd, + Path: fspath.Parse(args.WorkingDirectory), + FollowFinalSymlink: true, + } + var err error + wd, err = k.VFS.GetDentryAt(ctx, args.Credentials, &pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewVFSLookup(mntnsVFS2, root, wd) + fsContext = NewFSContextVFS2(root, wd, args.Umask) + + } else { + mntns = args.MountNamespace + if mntns == nil { + mntns = k.GlobalInit().Leader().MountNamespace() + mntns.IncRef() } - defer wd.DecRef() + // Get the root directory from the MountNamespace. + root := mntns.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + remainingTraversals := args.MaxSymlinkTraversals + wd := root // Default. + if args.WorkingDirectory != "" { + var err error + wd, err = mntns.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewFSLookup(mntns, root, wd) + fsContext = newFSContext(root, wd, args.Umask) } + tg := k.NewThreadGroup(mntns, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) + // Check which file to start from. switch { case args.Filename != "": @@ -805,11 +875,9 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, } // Create a fresh task context. - remainingTraversals = uint(args.MaxSymlinkTraversals) + remainingTraversals := args.MaxSymlinkTraversals loadArgs := loader.LoadArgs{ - Mounts: mounts, - Root: root, - WorkingDirectory: wd, + Opener: opener, RemainingTraversals: &remainingTraversals, ResolveFinal: true, Filename: args.Filename, @@ -834,13 +902,14 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, Kernel: k, ThreadGroup: tg, TaskContext: tc, - FSContext: newFSContext(root, wd, args.Umask), + FSContext: fsContext, FDTable: args.FDTable, Credentials: args.Credentials, AllowedCPUMask: sched.NewFullCPUSet(k.applicationCores), UTSNamespace: args.UTSNamespace, IPCNamespace: args.IPCNamespace, AbstractSocketNamespace: args.AbstractSocketNamespace, + MountNamespaceVFS2: args.MountNamespaceVFS2, ContainerID: args.ContainerID, } t, err := k.tasks.NewTask(config) @@ -1378,6 +1447,20 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { return ctx.k.globalInit.mounts.Root() } return nil + case vfs.CtxRoot: + if ctx.k.globalInit == nil { + return vfs.VirtualDentry{} + } + mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + defer mntns.DecRef() + // Root() takes a reference on the root dirent for us. + return mntns.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2() takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter case ktime.CtxRealtimeClock: diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index 981e8c7fe..a3443ff21 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -424,6 +424,11 @@ type Task struct { // abstractSockets is protected by mu. abstractSockets *AbstractSocketNamespace + // mountNamespaceVFS2 is the task's mount namespace. + // + // It is protected by mu. It is owned by the task goroutine. + mountNamespaceVFS2 *vfs.MountNamespace + // parentDeathSignal is sent to this task's thread group when its parent exits. // // parentDeathSignal is protected by mu. @@ -638,6 +643,11 @@ func (t *Task) Value(key interface{}) interface{} { return int32(t.ThreadGroup().ID()) case fs.CtxRoot: return t.fsContext.RootDirectory() + case vfs.CtxRoot: + return t.fsContext.RootDirectoryVFS2() + case vfs.CtxMountNamespace: + t.mountNamespaceVFS2.IncRef() + return t.mountNamespaceVFS2 case fs.CtxDirentCacheLimiter: return t.k.DirentCacheLimiter case inet.CtxStack: @@ -701,6 +711,14 @@ func (t *Task) SyscallRestartBlock() SyscallRestartBlock { // Preconditions: The caller must be running on the task goroutine, or t.mu // must be locked. func (t *Task) IsChrooted() bool { + if VFS2Enabled { + realRoot := t.mountNamespaceVFS2.Root() + defer realRoot.DecRef() + root := t.fsContext.RootDirectoryVFS2() + defer root.DecRef() + return root != realRoot + } + realRoot := t.tg.mounts.Root() defer realRoot.DecRef() root := t.fsContext.RootDirectory() @@ -796,6 +814,15 @@ func (t *Task) MountNamespace() *fs.MountNamespace { return t.tg.mounts } +// MountNamespaceVFS2 returns t's MountNamespace. A reference is taken on the +// returned mount namespace. +func (t *Task) MountNamespaceVFS2() *vfs.MountNamespace { + t.mu.Lock() + defer t.mu.Unlock() + t.mountNamespaceVFS2.IncRef() + return t.mountNamespaceVFS2 +} + // AbstractSockets returns t's AbstractSocketNamespace. func (t *Task) AbstractSockets() *AbstractSocketNamespace { return t.abstractSockets diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 53d4d211b..ba74b4c1c 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -199,6 +199,12 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { ipcns = NewIPCNamespace(userns) } + // TODO(b/63601033): Implement CLONE_NEWNS. + mntnsVFS2 := t.mountNamespaceVFS2 + if mntnsVFS2 != nil { + mntnsVFS2.IncRef() + } + tc, err := t.tc.Fork(t, t.k, !opts.NewAddressSpace) if err != nil { return 0, nil, err @@ -241,7 +247,9 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { rseqAddr := usermem.Addr(0) rseqSignature := uint32(0) if opts.NewThreadGroup { - tg.mounts.IncRef() + if tg.mounts != nil { + tg.mounts.IncRef() + } sh := t.tg.signalHandlers if opts.NewSignalHandlers { sh = sh.Fork() @@ -265,6 +273,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { UTSNamespace: utsns, IPCNamespace: ipcns, AbstractSocketNamespace: t.abstractSockets, + MountNamespaceVFS2: mntnsVFS2, RSeqAddr: rseqAddr, RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index 2d6e7733c..2be982684 100644 --- a/pkg/sentry/kernel/task_context.go +++ b/pkg/sentry/kernel/task_context.go @@ -136,7 +136,7 @@ func (t *Task) Stack() *arch.Stack { func (k *Kernel) LoadTaskImage(ctx context.Context, args loader.LoadArgs) (*TaskContext, *syserr.Error) { // If File is not nil, we should load that instead of resolving Filename. if args.File != nil { - args.Filename = args.File.MappedName(ctx) + args.Filename = args.File.PathnameWithDeleted(ctx) } // Prepare a new user address space to load into. diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go index 435761e5a..c4ade6e8e 100644 --- a/pkg/sentry/kernel/task_exit.go +++ b/pkg/sentry/kernel/task_exit.go @@ -269,6 +269,13 @@ func (*runExitMain) execute(t *Task) taskRunState { t.fsContext.DecRef() t.fdTable.DecRef() + t.mu.Lock() + if t.mountNamespaceVFS2 != nil { + t.mountNamespaceVFS2.DecRef() + t.mountNamespaceVFS2 = nil + } + t.mu.Unlock() + // If this is the last task to exit from the thread group, release the // thread group's resources. if lastExiter { diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go index 41259210c..6d737d3e5 100644 --- a/pkg/sentry/kernel/task_log.go +++ b/pkg/sentry/kernel/task_log.go @@ -198,18 +198,11 @@ func (t *Task) traceExecEvent(tc *TaskContext) { if !trace.IsEnabled() { return } - d := tc.MemoryManager.Executable() - if d == nil { + file := tc.MemoryManager.Executable() + if file == nil { trace.Logf(t.traceContext, traceCategory, "exec: << unknown >>") return } - defer d.DecRef() - root := t.fsContext.RootDirectory() - if root == nil { - trace.Logf(t.traceContext, traceCategory, "exec: << no root directory >>") - return - } - defer root.DecRef() - n, _ := d.FullName(root) - trace.Logf(t.traceContext, traceCategory, "exec: %s", n) + defer file.DecRef() + trace.Logf(t.traceContext, traceCategory, "exec: %s", file.PathnameWithDeleted(t)) } diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index de838beef..f9236a842 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -80,6 +81,9 @@ type TaskConfig struct { // AbstractSocketNamespace is the AbstractSocketNamespace of the new task. AbstractSocketNamespace *AbstractSocketNamespace + // MountNamespaceVFS2 is the MountNamespace of the new task. + MountNamespaceVFS2 *vfs.MountNamespace + // RSeqAddr is a pointer to the the userspace linux.RSeq structure. RSeqAddr usermem.Addr @@ -116,28 +120,29 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { parent: cfg.Parent, children: make(map[*Task]struct{}), }, - runState: (*runApp)(nil), - interruptChan: make(chan struct{}, 1), - signalMask: cfg.SignalMask, - signalStack: arch.SignalStack{Flags: arch.SignalStackFlagDisable}, - tc: *tc, - fsContext: cfg.FSContext, - fdTable: cfg.FDTable, - p: cfg.Kernel.Platform.NewContext(), - k: cfg.Kernel, - ptraceTracees: make(map[*Task]struct{}), - allowedCPUMask: cfg.AllowedCPUMask.Copy(), - ioUsage: &usage.IO{}, - niceness: cfg.Niceness, - netns: cfg.NetworkNamespaced, - utsns: cfg.UTSNamespace, - ipcns: cfg.IPCNamespace, - abstractSockets: cfg.AbstractSocketNamespace, - rseqCPU: -1, - rseqAddr: cfg.RSeqAddr, - rseqSignature: cfg.RSeqSignature, - futexWaiter: futex.NewWaiter(), - containerID: cfg.ContainerID, + runState: (*runApp)(nil), + interruptChan: make(chan struct{}, 1), + signalMask: cfg.SignalMask, + signalStack: arch.SignalStack{Flags: arch.SignalStackFlagDisable}, + tc: *tc, + fsContext: cfg.FSContext, + fdTable: cfg.FDTable, + p: cfg.Kernel.Platform.NewContext(), + k: cfg.Kernel, + ptraceTracees: make(map[*Task]struct{}), + allowedCPUMask: cfg.AllowedCPUMask.Copy(), + ioUsage: &usage.IO{}, + niceness: cfg.Niceness, + netns: cfg.NetworkNamespaced, + utsns: cfg.UTSNamespace, + ipcns: cfg.IPCNamespace, + abstractSockets: cfg.AbstractSocketNamespace, + mountNamespaceVFS2: cfg.MountNamespaceVFS2, + rseqCPU: -1, + rseqAddr: cfg.RSeqAddr, + rseqSignature: cfg.RSeqSignature, + futexWaiter: futex.NewWaiter(), + containerID: cfg.ContainerID, } t.creds.Store(cfg.Credentials) t.endStopCond.L = &t.tg.signalHandlers.mu diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 768e958d2..268f62e9d 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -256,7 +256,7 @@ type ThreadGroup struct { tty *TTY } -// NewThreadGroup returns a new, empty thread group in PID namespace ns. The +// NewThreadGroup returns a new, empty thread group in PID namespace pidns. The // thread group leader will send its parent terminationSignal when it exits. // The new thread group isn't visible to the system until a task has been // created inside of it by a successful call to TaskSet.NewTask. @@ -317,7 +317,9 @@ func (tg *ThreadGroup) release() { for _, it := range its { it.DestroyTimer() } - tg.mounts.DecRef() + if tg.mounts != nil { + tg.mounts.DecRef() + } } // forEachChildThreadGroupLocked indicates over all child ThreadGroups. diff --git a/pkg/sentry/loader/BUILD b/pkg/sentry/loader/BUILD index 23790378a..c6aa65f28 100644 --- a/pkg/sentry/loader/BUILD +++ b/pkg/sentry/loader/BUILD @@ -33,6 +33,7 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/fs/anon", "//pkg/sentry/fs/fsutil", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel/auth", "//pkg/sentry/limits", "//pkg/sentry/memmap", @@ -40,6 +41,7 @@ go_library( "//pkg/sentry/pgalloc", "//pkg/sentry/uniqueid", "//pkg/sentry/usage", + "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/syserror", "//pkg/usermem", diff --git a/pkg/sentry/loader/elf.go b/pkg/sentry/loader/elf.go index 122ed05c2..616fafa2c 100644 --- a/pkg/sentry/loader/elf.go +++ b/pkg/sentry/loader/elf.go @@ -27,7 +27,7 @@ import ( "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -97,11 +97,11 @@ type elfInfo struct { // accepts from the ELF, and it doesn't parse unnecessary parts of the file. // // ctx may be nil if f does not need it. -func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { +func parseHeader(ctx context.Context, f fsbridge.File) (elfInfo, error) { // Check ident first; it will tell us the endianness of the rest of the // structs. var ident [elf.EI_NIDENT]byte - _, err := readFull(ctx, f, usermem.BytesIOSequence(ident[:]), 0) + _, err := f.ReadFull(ctx, usermem.BytesIOSequence(ident[:]), 0) if err != nil { log.Infof("Error reading ELF ident: %v", err) // The entire ident array always exists. @@ -137,7 +137,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { var hdr elf.Header64 hdrBuf := make([]byte, header64Size) - _, err = readFull(ctx, f, usermem.BytesIOSequence(hdrBuf), 0) + _, err = f.ReadFull(ctx, usermem.BytesIOSequence(hdrBuf), 0) if err != nil { log.Infof("Error reading ELF header: %v", err) // The entire header always exists. @@ -187,7 +187,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { } phdrBuf := make([]byte, totalPhdrSize) - _, err = readFull(ctx, f, usermem.BytesIOSequence(phdrBuf), int64(hdr.Phoff)) + _, err = f.ReadFull(ctx, usermem.BytesIOSequence(phdrBuf), int64(hdr.Phoff)) if err != nil { log.Infof("Error reading ELF phdrs: %v", err) // If phdrs were specified, they should all exist. @@ -227,7 +227,7 @@ func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) { // mapSegment maps a phdr into the Task. offset is the offset to apply to // phdr.Vaddr. -func mapSegment(ctx context.Context, m *mm.MemoryManager, f *fs.File, phdr *elf.ProgHeader, offset usermem.Addr) error { +func mapSegment(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, phdr *elf.ProgHeader, offset usermem.Addr) error { // We must make a page-aligned mapping. adjust := usermem.Addr(phdr.Vaddr).PageOffset() @@ -395,7 +395,7 @@ type loadedELF struct { // // Preconditions: // * f is an ELF file -func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info elfInfo, sharedLoadOffset usermem.Addr) (loadedELF, error) { +func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, info elfInfo, sharedLoadOffset usermem.Addr) (loadedELF, error) { first := true var start, end usermem.Addr var interpreter string @@ -431,7 +431,7 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info el } path := make([]byte, phdr.Filesz) - _, err := readFull(ctx, f, usermem.BytesIOSequence(path), int64(phdr.Off)) + _, err := f.ReadFull(ctx, usermem.BytesIOSequence(path), int64(phdr.Off)) if err != nil { // If an interpreter was specified, it should exist. ctx.Infof("Error reading PT_INTERP path: %v", err) @@ -564,7 +564,7 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info el // Preconditions: // * f is an ELF file // * f is the first ELF loaded into m -func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureSet, f *fs.File) (loadedELF, arch.Context, error) { +func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureSet, f fsbridge.File) (loadedELF, arch.Context, error) { info, err := parseHeader(ctx, f) if err != nil { ctx.Infof("Failed to parse initial ELF: %v", err) @@ -602,7 +602,7 @@ func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureS // // Preconditions: // * f is an ELF file -func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, initial loadedELF) (loadedELF, error) { +func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, initial loadedELF) (loadedELF, error) { info, err := parseHeader(ctx, f) if err != nil { if err == syserror.ENOEXEC { @@ -649,16 +649,14 @@ func loadELF(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, error // Refresh the traversal limit. *args.RemainingTraversals = linux.MaxSymlinkTraversals args.Filename = bin.interpreter - d, i, err := openPath(ctx, args) + intFile, err := openPath(ctx, args) if err != nil { ctx.Infof("Error opening interpreter %s: %v", bin.interpreter, err) return loadedELF{}, nil, err } - defer i.DecRef() - // We don't need the Dirent. - d.DecRef() + defer intFile.DecRef() - interp, err = loadInterpreterELF(ctx, args.MemoryManager, i, bin) + interp, err = loadInterpreterELF(ctx, args.MemoryManager, intFile, bin) if err != nil { ctx.Infof("Error loading interpreter: %v", err) return loadedELF{}, nil, err diff --git a/pkg/sentry/loader/interpreter.go b/pkg/sentry/loader/interpreter.go index 098a45d36..3886b4d33 100644 --- a/pkg/sentry/loader/interpreter.go +++ b/pkg/sentry/loader/interpreter.go @@ -19,7 +19,7 @@ import ( "io" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -37,9 +37,9 @@ const ( ) // parseInterpreterScript returns the interpreter path and argv. -func parseInterpreterScript(ctx context.Context, filename string, f *fs.File, argv []string) (newpath string, newargv []string, err error) { +func parseInterpreterScript(ctx context.Context, filename string, f fsbridge.File, argv []string) (newpath string, newargv []string, err error) { line := make([]byte, interpMaxLineLength) - n, err := readFull(ctx, f, usermem.BytesIOSequence(line), 0) + n, err := f.ReadFull(ctx, usermem.BytesIOSequence(line), 0) // Short read is OK. if err != nil && err != io.ErrUnexpectedEOF { if err == io.EOF { diff --git a/pkg/sentry/loader/loader.go b/pkg/sentry/loader/loader.go index 9a613d6b7..d6675b8f0 100644 --- a/pkg/sentry/loader/loader.go +++ b/pkg/sentry/loader/loader.go @@ -20,7 +20,6 @@ import ( "fmt" "io" "path" - "strings" "gvisor.dev/gvisor/pkg/abi" "gvisor.dev/gvisor/pkg/abi/linux" @@ -29,8 +28,10 @@ import ( "gvisor.dev/gvisor/pkg/rand" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/mm" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -41,16 +42,6 @@ type LoadArgs struct { // MemoryManager is the memory manager to load the executable into. MemoryManager *mm.MemoryManager - // Mounts is the mount namespace in which to look up Filename. - Mounts *fs.MountNamespace - - // Root is the root directory under which to look up Filename. - Root *fs.Dirent - - // WorkingDirectory is the working directory under which to look up - // Filename. - WorkingDirectory *fs.Dirent - // RemainingTraversals is the maximum number of symlinks to follow to // resolve Filename. This counter is passed by reference to keep it // updated throughout the call stack. @@ -65,7 +56,12 @@ type LoadArgs struct { // File is an open fs.File object of the executable. If File is not // nil, then File will be loaded and Filename will be ignored. - File *fs.File + // + // The caller is responsible for checking that the user can execute this file. + File fsbridge.File + + // Opener is used to open the executable file when 'File' is nil. + Opener fsbridge.Lookup // CloseOnExec indicates that the executable (or one of its parent // directories) was opened with O_CLOEXEC. If the executable is an @@ -106,103 +102,32 @@ func readFull(ctx context.Context, f *fs.File, dst usermem.IOSequence, offset in // installed in the Task FDTable. The caller takes ownership of both. // // args.Filename must be a readable, executable, regular file. -func openPath(ctx context.Context, args LoadArgs) (*fs.Dirent, *fs.File, error) { +func openPath(ctx context.Context, args LoadArgs) (fsbridge.File, error) { if args.Filename == "" { ctx.Infof("cannot open empty name") - return nil, nil, syserror.ENOENT - } - - var d *fs.Dirent - var err error - if args.ResolveFinal { - d, err = args.Mounts.FindInode(ctx, args.Root, args.WorkingDirectory, args.Filename, args.RemainingTraversals) - } else { - d, err = args.Mounts.FindLink(ctx, args.Root, args.WorkingDirectory, args.Filename, args.RemainingTraversals) - } - if err != nil { - return nil, nil, err - } - // Defer a DecRef for the sake of failure cases. - defer d.DecRef() - - if !args.ResolveFinal && fs.IsSymlink(d.Inode.StableAttr) { - return nil, nil, syserror.ELOOP - } - - if err := checkPermission(ctx, d); err != nil { - return nil, nil, err - } - - // If they claim it's a directory, then make sure. - // - // N.B. we reject directories below, but we must first reject - // non-directories passed as directories. - if strings.HasSuffix(args.Filename, "/") && !fs.IsDir(d.Inode.StableAttr) { - return nil, nil, syserror.ENOTDIR - } - - if err := checkIsRegularFile(ctx, d, args.Filename); err != nil { - return nil, nil, err - } - - f, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true}) - if err != nil { - return nil, nil, err - } - // Defer a DecRef for the sake of failure cases. - defer f.DecRef() - - if err := checkPread(ctx, f, args.Filename); err != nil { - return nil, nil, err - } - - d.IncRef() - f.IncRef() - return d, f, err -} - -// checkFile performs checks on a file to be executed. -func checkFile(ctx context.Context, f *fs.File, filename string) error { - if err := checkPermission(ctx, f.Dirent); err != nil { - return err - } - - if err := checkIsRegularFile(ctx, f.Dirent, filename); err != nil { - return err + return nil, syserror.ENOENT } - return checkPread(ctx, f, filename) -} - -// checkPermission checks whether the file is readable and executable. -func checkPermission(ctx context.Context, d *fs.Dirent) error { - perms := fs.PermMask{ - // TODO(gvisor.dev/issue/160): Linux requires only execute - // permission, not read. However, our backing filesystems may - // prevent us from reading the file without read permission. - // - // Additionally, a task with a non-readable executable has - // additional constraints on access via ptrace and procfs. - Read: true, - Execute: true, + // TODO(gvisor.dev/issue/160): Linux requires only execute permission, + // not read. However, our backing filesystems may prevent us from reading + // the file without read permission. Additionally, a task with a + // non-readable executable has additional constraints on access via + // ptrace and procfs. + opts := vfs.OpenOptions{ + Flags: linux.O_RDONLY, + FileExec: true, } - return d.Inode.CheckPermission(ctx, perms) + return args.Opener.OpenPath(ctx, args.Filename, opts, args.RemainingTraversals, args.ResolveFinal) } // checkIsRegularFile prevents us from trying to execute a directory, pipe, etc. -func checkIsRegularFile(ctx context.Context, d *fs.Dirent, filename string) error { - attr := d.Inode.StableAttr - if !fs.IsRegular(attr) { - ctx.Infof("%s is not regular: %v", filename, attr) - return syserror.EACCES +func checkIsRegularFile(ctx context.Context, file fsbridge.File, filename string) error { + t, err := file.Type(ctx) + if err != nil { + return err } - return nil -} - -// checkPread checks whether we can read the file at arbitrary offsets. -func checkPread(ctx context.Context, f *fs.File, filename string) error { - if !f.Flags().Pread { - ctx.Infof("%s cannot be read at an offset: %+v", filename, f.Flags()) + if t != linux.ModeRegular { + ctx.Infof("%q is not a regular file: %v", filename, t) return syserror.EACCES } return nil @@ -224,8 +149,10 @@ const ( maxLoaderAttempts = 6 ) -// loadExecutable loads an executable that is pointed to by args.File. If nil, -// the path args.Filename is resolved and loaded. If the executable is an +// loadExecutable loads an executable that is pointed to by args.File. The +// caller is responsible for checking that the user can execute this file. +// If nil, the path args.Filename is resolved and loaded (check that the user +// can execute this file is done here in this case). If the executable is an // interpreter script rather than an ELF, the binary of the corresponding // interpreter will be loaded. // @@ -234,37 +161,27 @@ const ( // * arch.Context matching the binary arch // * fs.Dirent of the binary file // * Possibly updated args.Argv -func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, *fs.Dirent, []string, error) { +func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, fsbridge.File, []string, error) { for i := 0; i < maxLoaderAttempts; i++ { - var ( - d *fs.Dirent - err error - ) if args.File == nil { - d, args.File, err = openPath(ctx, args) - // We will return d in the successful case, but defer a DecRef for the - // sake of intermediate loops and failure cases. - if d != nil { - defer d.DecRef() - } - if args.File != nil { - defer args.File.DecRef() + var err error + args.File, err = openPath(ctx, args) + if err != nil { + ctx.Infof("Error opening %s: %v", args.Filename, err) + return loadedELF{}, nil, nil, nil, err } + // Ensure file is release in case the code loops or errors out. + defer args.File.DecRef() } else { - d = args.File.Dirent - d.IncRef() - defer d.DecRef() - err = checkFile(ctx, args.File, args.Filename) - } - if err != nil { - ctx.Infof("Error opening %s: %v", args.Filename, err) - return loadedELF{}, nil, nil, nil, err + if err := checkIsRegularFile(ctx, args.File, args.Filename); err != nil { + return loadedELF{}, nil, nil, nil, err + } } // Check the header. Is this an ELF or interpreter script? var hdr [4]uint8 // N.B. We assume that reading from a regular file cannot block. - _, err = readFull(ctx, args.File, usermem.BytesIOSequence(hdr[:]), 0) + _, err := args.File.ReadFull(ctx, usermem.BytesIOSequence(hdr[:]), 0) // Allow unexpected EOF, as a valid executable could be only three bytes // (e.g., #!a). if err != nil && err != io.ErrUnexpectedEOF { @@ -281,9 +198,10 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context ctx.Infof("Error loading ELF: %v", err) return loadedELF{}, nil, nil, nil, err } - // An ELF is always terminal. Hold on to d. - d.IncRef() - return loaded, ac, d, args.Argv, err + // An ELF is always terminal. Hold on to file. + args.File.IncRef() + return loaded, ac, args.File, args.Argv, err + case bytes.Equal(hdr[:2], []byte(interpreterScriptMagic)): if args.CloseOnExec { return loadedELF{}, nil, nil, nil, syserror.ENOENT @@ -295,6 +213,7 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context } // Refresh the traversal limit for the interpreter. *args.RemainingTraversals = linux.MaxSymlinkTraversals + default: ctx.Infof("Unknown magic: %v", hdr) return loadedELF{}, nil, nil, nil, syserror.ENOEXEC @@ -317,11 +236,11 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context // * Load is called on the Task goroutine. func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *VDSO) (abi.OS, arch.Context, string, *syserr.Error) { // Load the executable itself. - loaded, ac, d, newArgv, err := loadExecutable(ctx, args) + loaded, ac, file, newArgv, err := loadExecutable(ctx, args) if err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to load %s: %v", args.Filename, err), syserr.FromError(err).ToLinux()) } - defer d.DecRef() + defer file.DecRef() // Load the VDSO. vdsoAddr, err := loadVDSO(ctx, args.MemoryManager, vdso, loaded) @@ -390,7 +309,7 @@ func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *V m.SetEnvvStart(sl.EnvvStart) m.SetEnvvEnd(sl.EnvvEnd) m.SetAuxv(auxv) - m.SetExecutable(d) + m.SetExecutable(file) ac.SetIP(uintptr(loaded.entry)) ac.SetStack(uintptr(stack.Bottom)) diff --git a/pkg/sentry/loader/vdso.go b/pkg/sentry/loader/vdso.go index 52f446ed7..161b28c2c 100644 --- a/pkg/sentry/loader/vdso.go +++ b/pkg/sentry/loader/vdso.go @@ -27,6 +27,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/anon" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/mm" "gvisor.dev/gvisor/pkg/sentry/pgalloc" @@ -69,6 +70,8 @@ type byteReader struct { var _ fs.FileOperations = (*byteReader)(nil) // newByteReaderFile creates a fake file to read data from. +// +// TODO(gvisor.dev/issue/1623): Convert to VFS2. func newByteReaderFile(ctx context.Context, data []byte) *fs.File { // Create a fake inode. inode := fs.NewInode( @@ -123,7 +126,7 @@ func (b *byteReader) Write(ctx context.Context, file *fs.File, src usermem.IOSeq // * PT_LOAD segments don't extend beyond the end of the file. // // ctx may be nil if f does not need it. -func validateVDSO(ctx context.Context, f *fs.File, size uint64) (elfInfo, error) { +func validateVDSO(ctx context.Context, f fsbridge.File, size uint64) (elfInfo, error) { info, err := parseHeader(ctx, f) if err != nil { log.Infof("Unable to parse VDSO header: %v", err) @@ -221,7 +224,7 @@ type VDSO struct { // PrepareVDSO validates the system VDSO and returns a VDSO, containing the // param page for updating by the kernel. func PrepareVDSO(ctx context.Context, mfp pgalloc.MemoryFileProvider) (*VDSO, error) { - vdsoFile := newByteReaderFile(ctx, vdsoBin) + vdsoFile := fsbridge.NewFSFile(newByteReaderFile(ctx, vdsoBin)) // First make sure the VDSO is valid. vdsoFile does not use ctx, so a // nil context can be passed. diff --git a/pkg/sentry/mm/BUILD b/pkg/sentry/mm/BUILD index e5729ced5..73591dab7 100644 --- a/pkg/sentry/mm/BUILD +++ b/pkg/sentry/mm/BUILD @@ -105,8 +105,8 @@ go_library( "//pkg/safecopy", "//pkg/safemem", "//pkg/sentry/arch", - "//pkg/sentry/fs", "//pkg/sentry/fs/proc/seqfile", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/futex", "//pkg/sentry/kernel/shm", diff --git a/pkg/sentry/mm/metadata.go b/pkg/sentry/mm/metadata.go index f550acae0..6a49334f4 100644 --- a/pkg/sentry/mm/metadata.go +++ b/pkg/sentry/mm/metadata.go @@ -16,7 +16,7 @@ package mm import ( "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/usermem" ) @@ -132,7 +132,7 @@ func (mm *MemoryManager) SetAuxv(auxv arch.Auxv) { // // An additional reference will be taken in the case of a non-nil executable, // which must be released by the caller. -func (mm *MemoryManager) Executable() *fs.Dirent { +func (mm *MemoryManager) Executable() fsbridge.File { mm.metadataMu.Lock() defer mm.metadataMu.Unlock() @@ -147,15 +147,15 @@ func (mm *MemoryManager) Executable() *fs.Dirent { // SetExecutable sets the executable. // // This takes a reference on d. -func (mm *MemoryManager) SetExecutable(d *fs.Dirent) { +func (mm *MemoryManager) SetExecutable(file fsbridge.File) { mm.metadataMu.Lock() // Grab a new reference. - d.IncRef() + file.IncRef() // Set the executable. orig := mm.executable - mm.executable = d + mm.executable = file mm.metadataMu.Unlock() diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go index 09e582dd3..637383c7a 100644 --- a/pkg/sentry/mm/mm.go +++ b/pkg/sentry/mm/mm.go @@ -37,7 +37,7 @@ package mm import ( "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" @@ -215,7 +215,7 @@ type MemoryManager struct { // is not nil, it holds a reference on the Dirent. // // executable is protected by metadataMu. - executable *fs.Dirent + executable fsbridge.File // dumpability describes if and how this MemoryManager may be dumped to // userspace. diff --git a/pkg/sentry/strace/strace.go b/pkg/sentry/strace/strace.go index a796b2396..46cb2a1cc 100644 --- a/pkg/sentry/strace/strace.go +++ b/pkg/sentry/strace/strace.go @@ -141,6 +141,10 @@ func path(t *kernel.Task, addr usermem.Addr) string { } func fd(t *kernel.Task, fd int32) string { + if kernel.VFS2Enabled { + return fdVFS2(t, fd) + } + root := t.FSContext().RootDirectory() if root != nil { defer root.DecRef() @@ -169,6 +173,30 @@ func fd(t *kernel.Task, fd int32) string { return fmt.Sprintf("%#x %s", fd, name) } +func fdVFS2(t *kernel.Task, fd int32) string { + root := t.FSContext().RootDirectoryVFS2() + defer root.DecRef() + + vfsObj := root.Mount().Filesystem().VirtualFilesystem() + if fd == linux.AT_FDCWD { + wd := t.FSContext().WorkingDirectoryVFS2() + defer wd.DecRef() + + name, _ := vfsObj.PathnameWithDeleted(t, root, wd) + return fmt.Sprintf("AT_FDCWD %s", name) + } + + file := t.GetFileVFS2(fd) + if file == nil { + // Cast FD to uint64 to avoid printing negative hex. + return fmt.Sprintf("%#x (bad FD)", uint64(fd)) + } + defer file.DecRef() + + name, _ := vfsObj.PathnameWithDeleted(t, root, file.VirtualDentry()) + return fmt.Sprintf("%#x %s", fd, name) +} + func fdpair(t *kernel.Task, addr usermem.Addr) string { var fds [2]int32 _, err := t.CopyIn(addr, &fds) diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD index be16ee686..0d24fd3c4 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -74,6 +74,7 @@ go_library( "//pkg/sentry/fs/lock", "//pkg/sentry/fs/timerfd", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/fsbridge", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/epoll", diff --git a/pkg/sentry/syscalls/linux/sys_prctl.go b/pkg/sentry/syscalls/linux/sys_prctl.go index 98db32d77..9c6728530 100644 --- a/pkg/sentry/syscalls/linux/sys_prctl.go +++ b/pkg/sentry/syscalls/linux/sys_prctl.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/mm" @@ -135,7 +136,7 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } // Set the underlying executable. - t.MemoryManager().SetExecutable(file.Dirent) + t.MemoryManager().SetExecutable(fsbridge.NewFSFile(file)) case linux.PR_SET_MM_AUXV, linux.PR_SET_MM_START_CODE, diff --git a/pkg/sentry/syscalls/linux/sys_thread.go b/pkg/sentry/syscalls/linux/sys_thread.go index 0c9e2255d..00915fdde 100644 --- a/pkg/sentry/syscalls/linux/sys_thread.go +++ b/pkg/sentry/syscalls/linux/sys_thread.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/loader" @@ -119,7 +120,7 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user defer root.DecRef() var wd *fs.Dirent - var executable *fs.File + var executable fsbridge.File var closeOnExec bool if dirFD == linux.AT_FDCWD || path.IsAbs(pathname) { // Even if the pathname is absolute, we may still need the wd @@ -136,7 +137,15 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user closeOnExec = fdFlags.CloseOnExec if atEmptyPath && len(pathname) == 0 { - executable = f + // TODO(gvisor.dev/issue/160): Linux requires only execute permission, + // not read. However, our backing filesystems may prevent us from reading + // the file without read permission. Additionally, a task with a + // non-readable executable has additional constraints on access via + // ptrace and procfs. + if err := f.Dirent.Inode.CheckPermission(t, fs.PermMask{Read: true, Execute: true}); err != nil { + return 0, nil, err + } + executable = fsbridge.NewFSFile(f) } else { wd = f.Dirent wd.IncRef() @@ -152,9 +161,7 @@ func execveat(t *kernel.Task, dirFD int32, pathnameAddr, argvAddr, envvAddr user // Load the new TaskContext. remainingTraversals := uint(linux.MaxSymlinkTraversals) loadArgs := loader.LoadArgs{ - Mounts: t.MountNamespace(), - Root: root, - WorkingDirectory: wd, + Opener: fsbridge.NewFSLookup(t.MountNamespace(), root, wd), RemainingTraversals: &remainingTraversals, ResolveFinal: resolveFinal, Filename: pathname, diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go index c134714ee..e0ac32b33 100644 --- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go +++ b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go @@ -22,4 +22,110 @@ import ( // Override syscall table to add syscalls implementations from this package. func Override(table map[uintptr]kernel.Syscall) { table[0] = syscalls.Supported("read", Read) + + // Remove syscalls that haven't been converted yet. It's better to get ENOSYS + // rather than a SIGSEGV deep in the stack. + delete(table, 1) // write + delete(table, 2) // open + delete(table, 3) // close + delete(table, 4) // stat + delete(table, 5) // fstat + delete(table, 6) // lstat + delete(table, 7) // poll + delete(table, 8) // lseek + delete(table, 9) // mmap + delete(table, 16) // ioctl + delete(table, 17) // pread64 + delete(table, 18) // pwrite64 + delete(table, 19) // readv + delete(table, 20) // writev + delete(table, 21) // access + delete(table, 22) // pipe + delete(table, 32) // dup + delete(table, 33) // dup2 + delete(table, 40) // sendfile + delete(table, 59) // execve + delete(table, 72) // fcntl + delete(table, 73) // flock + delete(table, 74) // fsync + delete(table, 75) // fdatasync + delete(table, 76) // truncate + delete(table, 77) // ftruncate + delete(table, 78) // getdents + delete(table, 79) // getcwd + delete(table, 80) // chdir + delete(table, 81) // fchdir + delete(table, 82) // rename + delete(table, 83) // mkdir + delete(table, 84) // rmdir + delete(table, 85) // creat + delete(table, 86) // link + delete(table, 87) // unlink + delete(table, 88) // symlink + delete(table, 89) // readlink + delete(table, 90) // chmod + delete(table, 91) // fchmod + delete(table, 92) // chown + delete(table, 93) // fchown + delete(table, 94) // lchown + delete(table, 133) // mknod + delete(table, 137) // statfs + delete(table, 138) // fstatfs + delete(table, 161) // chroot + delete(table, 162) // sync + delete(table, 165) // mount + delete(table, 166) // umount2 + delete(table, 172) // iopl + delete(table, 173) // ioperm + delete(table, 187) // readahead + delete(table, 188) // setxattr + delete(table, 189) // lsetxattr + delete(table, 190) // fsetxattr + delete(table, 191) // getxattr + delete(table, 192) // lgetxattr + delete(table, 193) // fgetxattr + delete(table, 206) // io_setup + delete(table, 207) // io_destroy + delete(table, 208) // io_getevents + delete(table, 209) // io_submit + delete(table, 210) // io_cancel + delete(table, 213) // epoll_create + delete(table, 214) // epoll_ctl_old + delete(table, 215) // epoll_wait_old + delete(table, 216) // remap_file_pages + delete(table, 217) // getdents64 + delete(table, 232) // epoll_wait + delete(table, 233) // epoll_ctl + delete(table, 253) // inotify_init + delete(table, 254) // inotify_add_watch + delete(table, 255) // inotify_rm_watch + delete(table, 257) // openat + delete(table, 258) // mkdirat + delete(table, 259) // mknodat + delete(table, 260) // fchownat + delete(table, 261) // futimesat + delete(table, 262) // fstatat + delete(table, 263) // unlinkat + delete(table, 264) // renameat + delete(table, 265) // linkat + delete(table, 266) // symlinkat + delete(table, 267) // readlinkat + delete(table, 268) // fchmodat + delete(table, 269) // faccessat + delete(table, 270) // pselect + delete(table, 271) // ppoll + delete(table, 285) // fallocate + delete(table, 291) // epoll_create1 + delete(table, 292) // dup3 + delete(table, 293) // pipe2 + delete(table, 294) // inotify_init1 + delete(table, 295) // preadv + delete(table, 296) // pwritev + delete(table, 306) // syncfs + delete(table, 316) // renameat2 + delete(table, 319) // memfd_create + delete(table, 322) // execveat + delete(table, 327) // preadv2 + delete(table, 328) // pwritev2 + delete(table, 332) // statx } diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 14b39eb9d..0b4f18ab5 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -43,6 +43,7 @@ go_library( "//pkg/abi/linux", "//pkg/context", "//pkg/fspath", + "//pkg/log", "//pkg/sentry/arch", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/vfs/context.go b/pkg/sentry/vfs/context.go index d97362b9a..82781e6d3 100644 --- a/pkg/sentry/vfs/context.go +++ b/pkg/sentry/vfs/context.go @@ -29,9 +29,10 @@ const ( CtxRoot ) -// MountNamespaceFromContext returns the MountNamespace used by ctx. It does -// not take a reference on the returned MountNamespace. If ctx is not -// associated with a MountNamespace, MountNamespaceFromContext returns nil. +// MountNamespaceFromContext returns the MountNamespace used by ctx. If ctx is +// not associated with a MountNamespace, MountNamespaceFromContext returns nil. +// +// A reference is taken on the returned MountNamespace. func MountNamespaceFromContext(ctx context.Context) *MountNamespace { if v := ctx.Value(CtxMountNamespace); v != nil { return v.(*MountNamespace) diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 1fbb420f9..ad2c9fcf4 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -114,6 +114,7 @@ type MountNamespace struct { func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth.Credentials, source, fsTypeName string, opts *GetFilesystemOptions) (*MountNamespace, error) { rft := vfs.getFilesystemType(fsTypeName) if rft == nil { + ctx.Warningf("Unknown filesystem: %s", fsTypeName) return nil, syserror.ENODEV } fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, *opts) @@ -231,9 +232,12 @@ func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credenti return syserror.EINVAL } vfs.mountMu.Lock() - if mntns := MountNamespaceFromContext(ctx); mntns != nil && mntns != vd.mount.ns { - vfs.mountMu.Unlock() - return syserror.EINVAL + if mntns := MountNamespaceFromContext(ctx); mntns != nil { + defer mntns.DecRef() + if mntns != vd.mount.ns { + vfs.mountMu.Unlock() + return syserror.EINVAL + } } // TODO(jamieliu): Linux special-cases umount of the caller's root, which diff --git a/pkg/sentry/vfs/options.go b/pkg/sentry/vfs/options.go index fdf8be157..6af7fdac1 100644 --- a/pkg/sentry/vfs/options.go +++ b/pkg/sentry/vfs/options.go @@ -61,7 +61,7 @@ type MountOptions struct { type OpenOptions struct { // Flags contains access mode and flags as specified for open(2). // - // FilesystemImpls is reponsible for implementing the following flags: + // FilesystemImpls are responsible for implementing the following flags: // O_RDONLY, O_WRONLY, O_RDWR, O_APPEND, O_CREAT, O_DIRECT, O_DSYNC, // O_EXCL, O_NOATIME, O_NOCTTY, O_NONBLOCK, O_PATH, O_SYNC, O_TMPFILE, and // O_TRUNC. VFS is responsible for handling O_DIRECTORY, O_LARGEFILE, and diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index 9629afee9..51deae313 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -393,7 +393,8 @@ func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credential // be executed. return nil, syserror.EACCES } - if linux.FileMode(stat.Mode).FileType() != linux.ModeRegular { + if t := linux.FileMode(stat.Mode).FileType(); t != linux.ModeRegular { + ctx.Infof("%q is not a regular file: %v", pop.Path, t) return nil, syserror.EACCES } } @@ -743,6 +744,8 @@ func (vfs *VirtualFilesystem) SyncAllFilesystems(ctx context.Context) error { // VirtualDentry methods require that a reference is held on the VirtualDentry. // // VirtualDentry is analogous to Linux's struct path. +// +// +stateify savable type VirtualDentry struct { mount *Mount dentry *Dentry diff --git a/runsc/boot/loader.go b/runsc/boot/loader.go index 9f0d5d7af..239ca5302 100644 --- a/runsc/boot/loader.go +++ b/runsc/boot/loader.go @@ -795,16 +795,19 @@ func (l *Loader) executeAsync(args *control.ExecArgs) (kernel.ThreadID, error) { return 0, fmt.Errorf("container %q not started", args.ContainerID) } + // TODO(gvisor.dev/issue/1623): Add VFS2 support + // Get the container MountNamespace from the Task. tg.Leader().WithMuLocked(func(t *kernel.Task) { - // task.MountNamespace() does not take a ref, so we must do so - // ourselves. + // task.MountNamespace() does not take a ref, so we must do so ourselves. args.MountNamespace = t.MountNamespace() args.MountNamespace.IncRef() }) - defer args.MountNamespace.DecRef() + if args.MountNamespace != nil { + defer args.MountNamespace.DecRef() + } - // Add the HOME enviroment varible if it is not already set. + // Add the HOME environment variable if it is not already set. root := args.MountNamespace.Root() defer root.DecRef() ctx := fs.WithRoot(l.k.SupervisorContext(), root) -- cgit v1.2.3 From 4a73bae269ae9f52a962ae3b08a17ccaacf7ba80 Mon Sep 17 00:00:00 2001 From: gVisor bot Date: Thu, 20 Feb 2020 15:19:40 -0800 Subject: Initial network namespace support. TCP/IP will work with netstack networking. hostinet doesn't work, and sockets will have the same behavior as it is now. Before the userspace is able to create device, the default loopback device can be used to test. /proc/net and /sys/net will still be connected to the root network stack; this is the same behavior now. Issue #1833 PiperOrigin-RevId: 296309389 --- pkg/sentry/fs/proc/net.go | 5 +- pkg/sentry/fs/proc/sys_net.go | 4 +- pkg/sentry/fsimpl/proc/tasks_net.go | 5 +- pkg/sentry/fsimpl/proc/tasks_sys.go | 4 +- pkg/sentry/fsimpl/testutil/kernel.go | 1 + pkg/sentry/inet/BUILD | 1 + pkg/sentry/inet/namespace.go | 99 +++++++++++++++++++++++++ pkg/sentry/kernel/kernel.go | 26 ++++--- pkg/sentry/kernel/task.go | 9 +-- pkg/sentry/kernel/task_clone.go | 16 ++-- pkg/sentry/kernel/task_net.go | 19 +++-- pkg/sentry/kernel/task_start.go | 8 +- pkg/tcpip/time_unsafe.go | 2 + runsc/boot/BUILD | 2 +- runsc/boot/controller.go | 11 +-- runsc/boot/loader.go | 121 +++++++++++++++++++++---------- runsc/boot/network.go | 27 +++++++ runsc/boot/pprof.go | 18 ----- runsc/boot/pprof/BUILD | 11 +++ runsc/boot/pprof/pprof.go | 20 +++++ runsc/sandbox/network.go | 25 +------ test/syscalls/BUILD | 2 + test/syscalls/linux/BUILD | 17 +++++ test/syscalls/linux/network_namespace.cc | 121 +++++++++++++++++++++++++++++++ 24 files changed, 451 insertions(+), 123 deletions(-) create mode 100644 pkg/sentry/inet/namespace.go delete mode 100644 runsc/boot/pprof.go create mode 100644 runsc/boot/pprof/BUILD create mode 100644 runsc/boot/pprof/pprof.go create mode 100644 test/syscalls/linux/network_namespace.cc (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/proc/net.go b/pkg/sentry/fs/proc/net.go index 6f2775344..95d5817ff 100644 --- a/pkg/sentry/fs/proc/net.go +++ b/pkg/sentry/fs/proc/net.go @@ -43,7 +43,10 @@ import ( // newNet creates a new proc net entry. func (p *proc) newNetDir(ctx context.Context, k *kernel.Kernel, msrc *fs.MountSource) *fs.Inode { var contents map[string]*fs.Inode - if s := p.k.NetworkStack(); s != nil { + // TODO(gvisor.dev/issue/1833): Support for using the network stack in the + // network namespace of the calling process. We should make this per-process, + // a.k.a. /proc/PID/net, and make /proc/net a symlink to /proc/self/net. + if s := p.k.RootNetworkNamespace().Stack(); s != nil { contents = map[string]*fs.Inode{ "dev": seqfile.NewSeqFileInode(ctx, &netDev{s: s}, msrc), "snmp": seqfile.NewSeqFileInode(ctx, &netSnmp{s: s}, msrc), diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go index 0772d4ae4..d4c4b533d 100644 --- a/pkg/sentry/fs/proc/sys_net.go +++ b/pkg/sentry/fs/proc/sys_net.go @@ -357,7 +357,9 @@ func (p *proc) newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s ine func (p *proc) newSysNetDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode { var contents map[string]*fs.Inode - if s := p.k.NetworkStack(); s != nil { + // TODO(gvisor.dev/issue/1833): Support for using the network stack in the + // network namespace of the calling process. + if s := p.k.RootNetworkNamespace().Stack(); s != nil { contents = map[string]*fs.Inode{ "ipv4": p.newSysNetIPv4Dir(ctx, msrc, s), "core": p.newSysNetCore(ctx, msrc, s), diff --git a/pkg/sentry/fsimpl/proc/tasks_net.go b/pkg/sentry/fsimpl/proc/tasks_net.go index 608fec017..d4e1812d8 100644 --- a/pkg/sentry/fsimpl/proc/tasks_net.go +++ b/pkg/sentry/fsimpl/proc/tasks_net.go @@ -39,7 +39,10 @@ import ( func newNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { var contents map[string]*kernfs.Dentry - if stack := k.NetworkStack(); stack != nil { + // TODO(gvisor.dev/issue/1833): Support for using the network stack in the + // network namespace of the calling process. We should make this per-process, + // a.k.a. /proc/PID/net, and make /proc/net a symlink to /proc/self/net. + if stack := k.RootNetworkNamespace().Stack(); stack != nil { const ( arp = "IP address HW type Flags HW address Mask Device\n" netlink = "sk Eth Pid Groups Rmem Wmem Dump Locks Drops Inode\n" diff --git a/pkg/sentry/fsimpl/proc/tasks_sys.go b/pkg/sentry/fsimpl/proc/tasks_sys.go index c7ce74883..3d5dc463c 100644 --- a/pkg/sentry/fsimpl/proc/tasks_sys.go +++ b/pkg/sentry/fsimpl/proc/tasks_sys.go @@ -50,7 +50,9 @@ func newSysDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *k func newSysNetDir(root *auth.Credentials, inoGen InoGenerator, k *kernel.Kernel) *kernfs.Dentry { var contents map[string]*kernfs.Dentry - if stack := k.NetworkStack(); stack != nil { + // TODO(gvisor.dev/issue/1833): Support for using the network stack in the + // network namespace of the calling process. + if stack := k.RootNetworkNamespace().Stack(); stack != nil { contents = map[string]*kernfs.Dentry{ "ipv4": kernfs.NewStaticDir(root, inoGen.NextIno(), 0555, map[string]*kernfs.Dentry{ "tcp_sack": newDentry(root, inoGen.NextIno(), 0644, &tcpSackData{stack: stack}), diff --git a/pkg/sentry/fsimpl/testutil/kernel.go b/pkg/sentry/fsimpl/testutil/kernel.go index d0be32e72..488478e29 100644 --- a/pkg/sentry/fsimpl/testutil/kernel.go +++ b/pkg/sentry/fsimpl/testutil/kernel.go @@ -128,6 +128,7 @@ func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns ThreadGroup: tc, TaskContext: &kernel.TaskContext{Name: name}, Credentials: auth.CredentialsFromContext(ctx), + NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.ApplicationCores()), UTSNamespace: kernel.UTSNamespaceFromContext(ctx), IPCNamespace: kernel.IPCNamespaceFromContext(ctx), diff --git a/pkg/sentry/inet/BUILD b/pkg/sentry/inet/BUILD index 334432abf..07bf39fed 100644 --- a/pkg/sentry/inet/BUILD +++ b/pkg/sentry/inet/BUILD @@ -10,6 +10,7 @@ go_library( srcs = [ "context.go", "inet.go", + "namespace.go", "test_stack.go", ], deps = [ diff --git a/pkg/sentry/inet/namespace.go b/pkg/sentry/inet/namespace.go new file mode 100644 index 000000000..c16667e7f --- /dev/null +++ b/pkg/sentry/inet/namespace.go @@ -0,0 +1,99 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package inet + +// Namespace represents a network namespace. See network_namespaces(7). +// +// +stateify savable +type Namespace struct { + // stack is the network stack implementation of this network namespace. + stack Stack `state:"nosave"` + + // creator allows kernel to create new network stack for network namespaces. + // If nil, no networking will function if network is namespaced. + creator NetworkStackCreator + + // isRoot indicates whether this is the root network namespace. + isRoot bool +} + +// NewRootNamespace creates the root network namespace, with creator +// allowing new network namespaces to be created. If creator is nil, no +// networking will function if the network is namespaced. +func NewRootNamespace(stack Stack, creator NetworkStackCreator) *Namespace { + return &Namespace{ + stack: stack, + creator: creator, + isRoot: true, + } +} + +// NewNamespace creates a new network namespace from the root. +func NewNamespace(root *Namespace) *Namespace { + n := &Namespace{ + creator: root.creator, + } + n.init() + return n +} + +// Stack returns the network stack of n. Stack may return nil if no network +// stack is configured. +func (n *Namespace) Stack() Stack { + return n.stack +} + +// IsRoot returns whether n is the root network namespace. +func (n *Namespace) IsRoot() bool { + return n.isRoot +} + +// RestoreRootStack restores the root network namespace with stack. This should +// only be called when restoring kernel. +func (n *Namespace) RestoreRootStack(stack Stack) { + if !n.isRoot { + panic("RestoreRootStack can only be called on root network namespace") + } + if n.stack != nil { + panic("RestoreRootStack called after a stack has already been set") + } + n.stack = stack +} + +func (n *Namespace) init() { + // Root network namespace will have stack assigned later. + if n.isRoot { + return + } + if n.creator != nil { + var err error + n.stack, err = n.creator.CreateStack() + if err != nil { + panic(err) + } + } +} + +// afterLoad is invoked by stateify. +func (n *Namespace) afterLoad() { + n.init() +} + +// NetworkStackCreator allows new instances of a network stack to be created. It +// is used by the kernel to create new network namespaces when requested. +type NetworkStackCreator interface { + // CreateStack creates a new network stack for a network namespace. + CreateStack() (Stack, error) +} diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 7da0368f1..c62fd6eb1 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -111,7 +111,7 @@ type Kernel struct { timekeeper *Timekeeper tasks *TaskSet rootUserNamespace *auth.UserNamespace - networkStack inet.Stack `state:"nosave"` + rootNetworkNamespace *inet.Namespace applicationCores uint useHostCores bool extraAuxv []arch.AuxEntry @@ -260,8 +260,9 @@ type InitKernelArgs struct { // RootUserNamespace is the root user namespace. RootUserNamespace *auth.UserNamespace - // NetworkStack is the TCP/IP network stack. NetworkStack may be nil. - NetworkStack inet.Stack + // RootNetworkNamespace is the root network namespace. If nil, no networking + // will be available. + RootNetworkNamespace *inet.Namespace // ApplicationCores is the number of logical CPUs visible to sandboxed // applications. The set of logical CPU IDs is [0, ApplicationCores); thus @@ -320,7 +321,10 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.rootUTSNamespace = args.RootUTSNamespace k.rootIPCNamespace = args.RootIPCNamespace k.rootAbstractSocketNamespace = args.RootAbstractSocketNamespace - k.networkStack = args.NetworkStack + k.rootNetworkNamespace = args.RootNetworkNamespace + if k.rootNetworkNamespace == nil { + k.rootNetworkNamespace = inet.NewRootNamespace(nil, nil) + } k.applicationCores = args.ApplicationCores if args.UseHostCores { k.useHostCores = true @@ -543,8 +547,6 @@ func (ts *TaskSet) unregisterEpollWaiters() { func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) error { loadStart := time.Now() - k.networkStack = net - initAppCores := k.applicationCores // Load the pre-saved CPUID FeatureSet. @@ -575,6 +577,10 @@ func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) log.Infof("Kernel load stats: %s", &stats) log.Infof("Kernel load took [%s].", time.Since(kernelStart)) + // rootNetworkNamespace should be populated after loading the state file. + // Restore the root network stack. + k.rootNetworkNamespace.RestoreRootStack(net) + // Load the memory file's state. memoryStart := time.Now() if err := k.mf.LoadFrom(k.SupervisorContext(), r); err != nil { @@ -905,6 +911,7 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, FSContext: fsContext, FDTable: args.FDTable, Credentials: args.Credentials, + NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.applicationCores), UTSNamespace: args.UTSNamespace, IPCNamespace: args.IPCNamespace, @@ -1255,10 +1262,9 @@ func (k *Kernel) RootAbstractSocketNamespace() *AbstractSocketNamespace { return k.rootAbstractSocketNamespace } -// NetworkStack returns the network stack. NetworkStack may return nil if no -// network stack is available. -func (k *Kernel) NetworkStack() inet.Stack { - return k.networkStack +// RootNetworkNamespace returns the root network namespace, always non-nil. +func (k *Kernel) RootNetworkNamespace() *inet.Namespace { + return k.rootNetworkNamespace } // GlobalInit returns the thread group with ID 1 in the root PID namespace, or diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index a3443ff21..e37e23231 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -486,13 +486,10 @@ type Task struct { numaPolicy int32 numaNodeMask uint64 - // If netns is true, the task is in a non-root network namespace. Network - // namespaces aren't currently implemented in full; being in a network - // namespace simply prevents the task from observing any network devices - // (including loopback) or using abstract socket addresses (see unix(7)). + // netns is the task's network namespace. netns is never nil. // - // netns is protected by mu. netns is owned by the task goroutine. - netns bool + // netns is protected by mu. + netns *inet.Namespace // If rseqPreempted is true, before the next call to p.Switch(), // interrupt rseq critical regions as defined by rseqAddr and diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index ba74b4c1c..78866f280 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -17,6 +17,7 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -54,8 +55,7 @@ type SharingOptions struct { NewUserNamespace bool // If NewNetworkNamespace is true, the task should have an independent - // network namespace. (Note that network namespaces are not really - // implemented; see comment on Task.netns for details.) + // network namespace. NewNetworkNamespace bool // If NewFiles is true, the task should use an independent file descriptor @@ -199,6 +199,11 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { ipcns = NewIPCNamespace(userns) } + netns := t.NetworkNamespace() + if opts.NewNetworkNamespace { + netns = inet.NewNamespace(netns) + } + // TODO(b/63601033): Implement CLONE_NEWNS. mntnsVFS2 := t.mountNamespaceVFS2 if mntnsVFS2 != nil { @@ -268,7 +273,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { FDTable: fdTable, Credentials: creds, Niceness: t.Niceness(), - NetworkNamespaced: t.netns, + NetworkNamespace: netns, AllowedCPUMask: t.CPUMask(), UTSNamespace: utsns, IPCNamespace: ipcns, @@ -283,9 +288,6 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { } else { cfg.InheritParent = t } - if opts.NewNetworkNamespace { - cfg.NetworkNamespaced = true - } nt, err := t.tg.pidns.owner.NewTask(cfg) if err != nil { if opts.NewThreadGroup { @@ -482,7 +484,7 @@ func (t *Task) Unshare(opts *SharingOptions) error { t.mu.Unlock() return syserror.EPERM } - t.netns = true + t.netns = inet.NewNamespace(t.netns) } if opts.NewUTSNamespace { if !haveCapSysAdmin { diff --git a/pkg/sentry/kernel/task_net.go b/pkg/sentry/kernel/task_net.go index 172a31e1d..f7711232c 100644 --- a/pkg/sentry/kernel/task_net.go +++ b/pkg/sentry/kernel/task_net.go @@ -22,14 +22,23 @@ import ( func (t *Task) IsNetworkNamespaced() bool { t.mu.Lock() defer t.mu.Unlock() - return t.netns + return !t.netns.IsRoot() } // NetworkContext returns the network stack used by the task. NetworkContext // may return nil if no network stack is available. +// +// TODO(gvisor.dev/issue/1833): Migrate callers of this method to +// NetworkNamespace(). func (t *Task) NetworkContext() inet.Stack { - if t.IsNetworkNamespaced() { - return nil - } - return t.k.networkStack + t.mu.Lock() + defer t.mu.Unlock() + return t.netns.Stack() +} + +// NetworkNamespace returns the network namespace observed by the task. +func (t *Task) NetworkNamespace() *inet.Namespace { + t.mu.Lock() + defer t.mu.Unlock() + return t.netns } diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index f9236a842..a5035bb7f 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -17,6 +17,7 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" @@ -65,9 +66,8 @@ type TaskConfig struct { // Niceness is the niceness of the new task. Niceness int - // If NetworkNamespaced is true, the new task should observe a non-root - // network namespace. - NetworkNamespaced bool + // NetworkNamespace is the network namespace to be used for the new task. + NetworkNamespace *inet.Namespace // AllowedCPUMask contains the cpus that this task can run on. AllowedCPUMask sched.CPUSet @@ -133,7 +133,7 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { allowedCPUMask: cfg.AllowedCPUMask.Copy(), ioUsage: &usage.IO{}, niceness: cfg.Niceness, - netns: cfg.NetworkNamespaced, + netns: cfg.NetworkNamespace, utsns: cfg.UTSNamespace, ipcns: cfg.IPCNamespace, abstractSockets: cfg.AbstractSocketNamespace, diff --git a/pkg/tcpip/time_unsafe.go b/pkg/tcpip/time_unsafe.go index 48764b978..2f98a996f 100644 --- a/pkg/tcpip/time_unsafe.go +++ b/pkg/tcpip/time_unsafe.go @@ -25,6 +25,8 @@ import ( ) // StdClock implements Clock with the time package. +// +// +stateify savable type StdClock struct{} var _ Clock = (*StdClock)(nil) diff --git a/runsc/boot/BUILD b/runsc/boot/BUILD index ae4dd102a..26f68fe3d 100644 --- a/runsc/boot/BUILD +++ b/runsc/boot/BUILD @@ -19,7 +19,6 @@ go_library( "loader_amd64.go", "loader_arm64.go", "network.go", - "pprof.go", "strace.go", "user.go", ], @@ -91,6 +90,7 @@ go_library( "//pkg/usermem", "//runsc/boot/filter", "//runsc/boot/platforms", + "//runsc/boot/pprof", "//runsc/specutils", "@com_github_golang_protobuf//proto:go_default_library", "@com_github_opencontainers_runtime-spec//specs-go:go_default_library", diff --git a/runsc/boot/controller.go b/runsc/boot/controller.go index 9c9e94864..17e774e0c 100644 --- a/runsc/boot/controller.go +++ b/runsc/boot/controller.go @@ -32,6 +32,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/watchdog" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/urpc" + "gvisor.dev/gvisor/runsc/boot/pprof" "gvisor.dev/gvisor/runsc/specutils" ) @@ -142,7 +143,7 @@ func newController(fd int, l *Loader) (*controller, error) { } srv.Register(manager) - if eps, ok := l.k.NetworkStack().(*netstack.Stack); ok { + if eps, ok := l.k.RootNetworkNamespace().Stack().(*netstack.Stack); ok { net := &Network{ Stack: eps.Stack, } @@ -341,7 +342,7 @@ func (cm *containerManager) Restore(o *RestoreOpts, _ *struct{}) error { return fmt.Errorf("creating memory file: %v", err) } k.SetMemoryFile(mf) - networkStack := cm.l.k.NetworkStack() + networkStack := cm.l.k.RootNetworkNamespace().Stack() cm.l.k = k // Set up the restore environment. @@ -365,9 +366,9 @@ func (cm *containerManager) Restore(o *RestoreOpts, _ *struct{}) error { } if cm.l.conf.ProfileEnable { - // initializePProf opens /proc/self/maps, so has to be - // called before installing seccomp filters. - initializePProf() + // pprof.Initialize opens /proc/self/maps, so has to be called before + // installing seccomp filters. + pprof.Initialize() } // Seccomp filters have to be applied before parsing the state file. diff --git a/runsc/boot/loader.go b/runsc/boot/loader.go index eef43b9df..e7ca98134 100644 --- a/runsc/boot/loader.go +++ b/runsc/boot/loader.go @@ -49,6 +49,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/watchdog" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/link/loopback" "gvisor.dev/gvisor/pkg/tcpip/link/sniffer" "gvisor.dev/gvisor/pkg/tcpip/network/arp" "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" @@ -60,6 +61,7 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/transport/udp" "gvisor.dev/gvisor/runsc/boot/filter" _ "gvisor.dev/gvisor/runsc/boot/platforms" // register all platforms. + "gvisor.dev/gvisor/runsc/boot/pprof" "gvisor.dev/gvisor/runsc/specutils" // Include supported socket providers. @@ -230,11 +232,8 @@ func New(args Args) (*Loader, error) { return nil, fmt.Errorf("enabling strace: %v", err) } - // Create an empty network stack because the network namespace may be empty at - // this point. Netns is configured before Run() is called. Netstack is - // configured using a control uRPC message. Host network is configured inside - // Run(). - networkStack, err := newEmptyNetworkStack(args.Conf, k, k) + // Create root network namespace/stack. + netns, err := newRootNetworkNamespace(args.Conf, k, k) if err != nil { return nil, fmt.Errorf("creating network: %v", err) } @@ -277,7 +276,7 @@ func New(args Args) (*Loader, error) { FeatureSet: cpuid.HostFeatureSet(), Timekeeper: tk, RootUserNamespace: creds.UserNamespace, - NetworkStack: networkStack, + RootNetworkNamespace: netns, ApplicationCores: uint(args.NumCPU), Vdso: vdso, RootUTSNamespace: kernel.NewUTSNamespace(args.Spec.Hostname, args.Spec.Hostname, creds.UserNamespace), @@ -466,7 +465,7 @@ func (l *Loader) run() error { // Delay host network configuration to this point because network namespace // is configured after the loader is created and before Run() is called. log.Debugf("Configuring host network") - stack := l.k.NetworkStack().(*hostinet.Stack) + stack := l.k.RootNetworkNamespace().Stack().(*hostinet.Stack) if err := stack.Configure(); err != nil { return err } @@ -485,7 +484,7 @@ func (l *Loader) run() error { // l.restore is set by the container manager when a restore call is made. if !l.restore { if l.conf.ProfileEnable { - initializePProf() + pprof.Initialize() } // Finally done with all configuration. Setup filters before user code @@ -908,48 +907,92 @@ func (l *Loader) WaitExit() kernel.ExitStatus { return l.k.GlobalInit().ExitStatus() } -func newEmptyNetworkStack(conf *Config, clock tcpip.Clock, uniqueID stack.UniqueID) (inet.Stack, error) { +func newRootNetworkNamespace(conf *Config, clock tcpip.Clock, uniqueID stack.UniqueID) (*inet.Namespace, error) { + // Create an empty network stack because the network namespace may be empty at + // this point. Netns is configured before Run() is called. Netstack is + // configured using a control uRPC message. Host network is configured inside + // Run(). switch conf.Network { case NetworkHost: - return hostinet.NewStack(), nil + // No network namespacing support for hostinet yet, hence creator is nil. + return inet.NewRootNamespace(hostinet.NewStack(), nil), nil case NetworkNone, NetworkSandbox: - // NetworkNone sets up loopback using netstack. - netProtos := []stack.NetworkProtocol{ipv4.NewProtocol(), ipv6.NewProtocol(), arp.NewProtocol()} - transProtos := []stack.TransportProtocol{tcp.NewProtocol(), udp.NewProtocol(), icmp.NewProtocol4()} - s := netstack.Stack{stack.New(stack.Options{ - NetworkProtocols: netProtos, - TransportProtocols: transProtos, - Clock: clock, - Stats: netstack.Metrics, - HandleLocal: true, - // Enable raw sockets for users with sufficient - // privileges. - RawFactory: raw.EndpointFactory{}, - UniqueID: uniqueID, - })} - - // Enable SACK Recovery. - if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcp.SACKEnabled(true)); err != nil { - return nil, fmt.Errorf("failed to enable SACK: %v", err) + s, err := newEmptySandboxNetworkStack(clock, uniqueID) + if err != nil { + return nil, err } + creator := &sandboxNetstackCreator{ + clock: clock, + uniqueID: uniqueID, + } + return inet.NewRootNamespace(s, creator), nil - // Set default TTLs as required by socket/netstack. - s.Stack.SetNetworkProtocolOption(ipv4.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL)) - s.Stack.SetNetworkProtocolOption(ipv6.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL)) + default: + panic(fmt.Sprintf("invalid network configuration: %v", conf.Network)) + } - // Enable Receive Buffer Auto-Tuning. - if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.ModerateReceiveBufferOption(true)); err != nil { - return nil, fmt.Errorf("SetTransportProtocolOption failed: %v", err) - } +} - s.FillDefaultIPTables() +func newEmptySandboxNetworkStack(clock tcpip.Clock, uniqueID stack.UniqueID) (inet.Stack, error) { + netProtos := []stack.NetworkProtocol{ipv4.NewProtocol(), ipv6.NewProtocol(), arp.NewProtocol()} + transProtos := []stack.TransportProtocol{tcp.NewProtocol(), udp.NewProtocol(), icmp.NewProtocol4()} + s := netstack.Stack{stack.New(stack.Options{ + NetworkProtocols: netProtos, + TransportProtocols: transProtos, + Clock: clock, + Stats: netstack.Metrics, + HandleLocal: true, + // Enable raw sockets for users with sufficient + // privileges. + RawFactory: raw.EndpointFactory{}, + UniqueID: uniqueID, + })} - return &s, nil + // Enable SACK Recovery. + if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcp.SACKEnabled(true)); err != nil { + return nil, fmt.Errorf("failed to enable SACK: %v", err) + } - default: - panic(fmt.Sprintf("invalid network configuration: %v", conf.Network)) + // Set default TTLs as required by socket/netstack. + s.Stack.SetNetworkProtocolOption(ipv4.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL)) + s.Stack.SetNetworkProtocolOption(ipv6.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL)) + + // Enable Receive Buffer Auto-Tuning. + if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.ModerateReceiveBufferOption(true)); err != nil { + return nil, fmt.Errorf("SetTransportProtocolOption failed: %v", err) + } + + s.FillDefaultIPTables() + + return &s, nil +} + +// sandboxNetstackCreator implements kernel.NetworkStackCreator. +// +// +stateify savable +type sandboxNetstackCreator struct { + clock tcpip.Clock + uniqueID stack.UniqueID +} + +// CreateStack implements kernel.NetworkStackCreator.CreateStack. +func (f *sandboxNetstackCreator) CreateStack() (inet.Stack, error) { + s, err := newEmptySandboxNetworkStack(f.clock, f.uniqueID) + if err != nil { + return nil, err } + + // Setup loopback. + n := &Network{Stack: s.(*netstack.Stack).Stack} + nicID := tcpip.NICID(f.uniqueID.UniqueID()) + link := DefaultLoopbackLink + linkEP := loopback.New() + if err := n.createNICWithAddrs(nicID, link.Name, linkEP, link.Addresses); err != nil { + return nil, err + } + + return s, nil } // signal sends a signal to one or more processes in a container. If PID is 0, diff --git a/runsc/boot/network.go b/runsc/boot/network.go index 6a8765ec8..bee6ee336 100644 --- a/runsc/boot/network.go +++ b/runsc/boot/network.go @@ -17,6 +17,7 @@ package boot import ( "fmt" "net" + "strings" "syscall" "gvisor.dev/gvisor/pkg/log" @@ -31,6 +32,32 @@ import ( "gvisor.dev/gvisor/pkg/urpc" ) +var ( + // DefaultLoopbackLink contains IP addresses and routes of "127.0.0.1/8" and + // "::1/8" on "lo" interface. + DefaultLoopbackLink = LoopbackLink{ + Name: "lo", + Addresses: []net.IP{ + net.IP("\x7f\x00\x00\x01"), + net.IPv6loopback, + }, + Routes: []Route{ + { + Destination: net.IPNet{ + IP: net.IPv4(0x7f, 0, 0, 0), + Mask: net.IPv4Mask(0xff, 0, 0, 0), + }, + }, + { + Destination: net.IPNet{ + IP: net.IPv6loopback, + Mask: net.IPMask(strings.Repeat("\xff", net.IPv6len)), + }, + }, + }, + } +) + // Network exposes methods that can be used to configure a network stack. type Network struct { Stack *stack.Stack diff --git a/runsc/boot/pprof.go b/runsc/boot/pprof.go deleted file mode 100644 index 463362f02..000000000 --- a/runsc/boot/pprof.go +++ /dev/null @@ -1,18 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package boot - -func initializePProf() { -} diff --git a/runsc/boot/pprof/BUILD b/runsc/boot/pprof/BUILD new file mode 100644 index 000000000..29cb42b2f --- /dev/null +++ b/runsc/boot/pprof/BUILD @@ -0,0 +1,11 @@ +load("//tools:defs.bzl", "go_library") + +package(licenses = ["notice"]) + +go_library( + name = "pprof", + srcs = ["pprof.go"], + visibility = [ + "//runsc:__subpackages__", + ], +) diff --git a/runsc/boot/pprof/pprof.go b/runsc/boot/pprof/pprof.go new file mode 100644 index 000000000..1ded20dee --- /dev/null +++ b/runsc/boot/pprof/pprof.go @@ -0,0 +1,20 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package pprof provides a stub to initialize custom profilers. +package pprof + +// Initialize will be called at boot for initializing custom profilers. +func Initialize() { +} diff --git a/runsc/sandbox/network.go b/runsc/sandbox/network.go index 99e143696..bc093fba5 100644 --- a/runsc/sandbox/network.go +++ b/runsc/sandbox/network.go @@ -21,7 +21,6 @@ import ( "path/filepath" "runtime" "strconv" - "strings" "syscall" specs "github.com/opencontainers/runtime-spec/specs-go" @@ -75,30 +74,8 @@ func setupNetwork(conn *urpc.Client, pid int, spec *specs.Spec, conf *boot.Confi } func createDefaultLoopbackInterface(conn *urpc.Client) error { - link := boot.LoopbackLink{ - Name: "lo", - Addresses: []net.IP{ - net.IP("\x7f\x00\x00\x01"), - net.IPv6loopback, - }, - Routes: []boot.Route{ - { - Destination: net.IPNet{ - - IP: net.IPv4(0x7f, 0, 0, 0), - Mask: net.IPv4Mask(0xff, 0, 0, 0), - }, - }, - { - Destination: net.IPNet{ - IP: net.IPv6loopback, - Mask: net.IPMask(strings.Repeat("\xff", net.IPv6len)), - }, - }, - }, - } if err := conn.Call(boot.NetworkCreateLinksAndRoutes, &boot.CreateLinksAndRoutesArgs{ - LoopbackLinks: []boot.LoopbackLink{link}, + LoopbackLinks: []boot.LoopbackLink{boot.DefaultLoopbackLink}, }, nil); err != nil { return fmt.Errorf("creating loopback link and routes: %v", err) } diff --git a/test/syscalls/BUILD b/test/syscalls/BUILD index d69ac8356..d1977d4de 100644 --- a/test/syscalls/BUILD +++ b/test/syscalls/BUILD @@ -258,6 +258,8 @@ syscall_test( syscall_test(test = "//test/syscalls/linux:munmap_test") +syscall_test(test = "//test/syscalls/linux:network_namespace_test") + syscall_test( add_overlay = True, test = "//test/syscalls/linux:open_create_test", diff --git a/test/syscalls/linux/BUILD b/test/syscalls/linux/BUILD index 05a818795..aa303af84 100644 --- a/test/syscalls/linux/BUILD +++ b/test/syscalls/linux/BUILD @@ -3639,6 +3639,23 @@ cc_binary( ], ) +cc_binary( + name = "network_namespace_test", + testonly = 1, + srcs = ["network_namespace.cc"], + linkstatic = 1, + deps = [ + ":socket_test_util", + gtest, + "//test/util:capability_util", + "//test/util:memory_util", + "//test/util:test_main", + "//test/util:test_util", + "//test/util:thread_util", + "@com_google_absl//absl/synchronization", + ], +) + cc_binary( name = "semaphore_test", testonly = 1, diff --git a/test/syscalls/linux/network_namespace.cc b/test/syscalls/linux/network_namespace.cc new file mode 100644 index 000000000..6ea48c263 --- /dev/null +++ b/test/syscalls/linux/network_namespace.cc @@ -0,0 +1,121 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include +#include +#include + +#include "gmock/gmock.h" +#include "gtest/gtest.h" +#include "absl/synchronization/notification.h" +#include "test/syscalls/linux/socket_test_util.h" +#include "test/util/capability_util.h" +#include "test/util/memory_util.h" +#include "test/util/test_util.h" +#include "test/util/thread_util.h" + +namespace gvisor { +namespace testing { + +namespace { + +using TestFunc = std::function; +using RunFunc = std::function; + +struct NamespaceStrategy { + RunFunc run; + + static NamespaceStrategy Of(RunFunc run) { + NamespaceStrategy s; + s.run = run; + return s; + } +}; + +PosixError RunWithUnshare(TestFunc fn) { + PosixError err = PosixError(-1, "function did not return a value"); + ScopedThread t([&] { + if (unshare(CLONE_NEWNET) != 0) { + err = PosixError(errno); + return; + } + err = fn(); + }); + t.Join(); + return err; +} + +PosixError RunWithClone(TestFunc fn) { + struct Args { + absl::Notification n; + TestFunc fn; + PosixError err; + }; + Args args; + args.fn = fn; + args.err = PosixError(-1, "function did not return a value"); + + ASSIGN_OR_RETURN_ERRNO( + Mapping child_stack, + MmapAnon(kPageSize, PROT_READ | PROT_WRITE, MAP_PRIVATE)); + pid_t child = clone( + +[](void *arg) { + Args *args = reinterpret_cast(arg); + args->err = args->fn(); + args->n.Notify(); + syscall(SYS_exit, 0); // Exit manually. No return address on stack. + return 0; + }, + reinterpret_cast(child_stack.addr() + kPageSize), + CLONE_NEWNET | CLONE_THREAD | CLONE_SIGHAND | CLONE_VM, &args); + if (child < 0) { + return PosixError(errno, "clone() failed"); + } + args.n.WaitForNotification(); + return args.err; +} + +class NetworkNamespaceTest + : public ::testing::TestWithParam {}; + +TEST_P(NetworkNamespaceTest, LoopbackExists) { + SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))); + + EXPECT_NO_ERRNO(GetParam().run([]() { + // TODO(gvisor.dev/issue/1833): Update this to test that only "lo" exists. + // Check loopback device exists. + int sock = socket(AF_INET, SOCK_DGRAM, 0); + if (sock < 0) { + return PosixError(errno, "socket() failed"); + } + struct ifreq ifr; + snprintf(ifr.ifr_name, IFNAMSIZ, "lo"); + if (ioctl(sock, SIOCGIFINDEX, &ifr) < 0) { + return PosixError(errno, "ioctl() failed, lo cannot be found"); + } + return NoError(); + })); +} + +INSTANTIATE_TEST_SUITE_P( + AllNetworkNamespaceTest, NetworkNamespaceTest, + ::testing::Values(NamespaceStrategy::Of(RunWithUnshare), + NamespaceStrategy::Of(RunWithClone))); + +} // namespace + +} // namespace testing +} // namespace gvisor -- cgit v1.2.3 From b8f56c79be40d9c75f4e2f279c9d821d1c1c3569 Mon Sep 17 00:00:00 2001 From: Ting-Yu Wang Date: Fri, 21 Feb 2020 15:41:56 -0800 Subject: Implement tap/tun device in vfs. PiperOrigin-RevId: 296526279 --- pkg/abi/linux/BUILD | 1 + pkg/abi/linux/ioctl.go | 26 ++ pkg/abi/linux/ioctl_tun.go | 29 ++ pkg/sentry/fs/dev/BUILD | 5 + pkg/sentry/fs/dev/dev.go | 10 +- pkg/sentry/fs/dev/net_tun.go | 170 +++++++++++ pkg/syserror/syserror.go | 1 + pkg/tcpip/buffer/view.go | 6 + pkg/tcpip/link/channel/BUILD | 1 + pkg/tcpip/link/channel/channel.go | 180 +++++++++--- pkg/tcpip/link/tun/BUILD | 18 +- pkg/tcpip/link/tun/device.go | 352 +++++++++++++++++++++++ pkg/tcpip/link/tun/protocol.go | 56 ++++ pkg/tcpip/stack/nic.go | 32 +++ pkg/tcpip/stack/stack.go | 39 +++ test/syscalls/BUILD | 2 + test/syscalls/linux/BUILD | 30 ++ test/syscalls/linux/dev.cc | 7 + test/syscalls/linux/socket_netlink_route_util.cc | 163 +++++++++++ test/syscalls/linux/socket_netlink_route_util.h | 55 ++++ test/syscalls/linux/tuntap.cc | 346 ++++++++++++++++++++++ 21 files changed, 1490 insertions(+), 39 deletions(-) create mode 100644 pkg/abi/linux/ioctl_tun.go create mode 100644 pkg/sentry/fs/dev/net_tun.go create mode 100644 pkg/tcpip/link/tun/device.go create mode 100644 pkg/tcpip/link/tun/protocol.go create mode 100644 test/syscalls/linux/socket_netlink_route_util.cc create mode 100644 test/syscalls/linux/socket_netlink_route_util.h create mode 100644 test/syscalls/linux/tuntap.cc (limited to 'pkg/sentry/fs') diff --git a/pkg/abi/linux/BUILD b/pkg/abi/linux/BUILD index a89f34d4b..322d1ccc4 100644 --- a/pkg/abi/linux/BUILD +++ b/pkg/abi/linux/BUILD @@ -30,6 +30,7 @@ go_library( "futex.go", "inotify.go", "ioctl.go", + "ioctl_tun.go", "ip.go", "ipc.go", "limits.go", diff --git a/pkg/abi/linux/ioctl.go b/pkg/abi/linux/ioctl.go index 0e18db9ef..2062e6a4b 100644 --- a/pkg/abi/linux/ioctl.go +++ b/pkg/abi/linux/ioctl.go @@ -72,3 +72,29 @@ const ( SIOCGMIIPHY = 0x8947 SIOCGMIIREG = 0x8948 ) + +// ioctl(2) directions. Used to calculate requests number. +// Constants from asm-generic/ioctl.h. +const ( + _IOC_NONE = 0 + _IOC_WRITE = 1 + _IOC_READ = 2 +) + +// Constants from asm-generic/ioctl.h. +const ( + _IOC_NRBITS = 8 + _IOC_TYPEBITS = 8 + _IOC_SIZEBITS = 14 + _IOC_DIRBITS = 2 + + _IOC_NRSHIFT = 0 + _IOC_TYPESHIFT = _IOC_NRSHIFT + _IOC_NRBITS + _IOC_SIZESHIFT = _IOC_TYPESHIFT + _IOC_TYPEBITS + _IOC_DIRSHIFT = _IOC_SIZESHIFT + _IOC_SIZEBITS +) + +// IOC outputs the result of _IOC macro in asm-generic/ioctl.h. +func IOC(dir, typ, nr, size uint32) uint32 { + return uint32(dir)<<_IOC_DIRSHIFT | typ<<_IOC_TYPESHIFT | nr<<_IOC_NRSHIFT | size<<_IOC_SIZESHIFT +} diff --git a/pkg/abi/linux/ioctl_tun.go b/pkg/abi/linux/ioctl_tun.go new file mode 100644 index 000000000..c59c9c136 --- /dev/null +++ b/pkg/abi/linux/ioctl_tun.go @@ -0,0 +1,29 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package linux + +// ioctl(2) request numbers from linux/if_tun.h +var ( + TUNSETIFF = IOC(_IOC_WRITE, 'T', 202, 4) + TUNGETIFF = IOC(_IOC_READ, 'T', 210, 4) +) + +// Flags from net/if_tun.h +const ( + IFF_TUN = 0x0001 + IFF_TAP = 0x0002 + IFF_NO_PI = 0x1000 + IFF_NOFILTER = 0x1000 +) diff --git a/pkg/sentry/fs/dev/BUILD b/pkg/sentry/fs/dev/BUILD index 4c4b7d5cc..9b6bb26d0 100644 --- a/pkg/sentry/fs/dev/BUILD +++ b/pkg/sentry/fs/dev/BUILD @@ -9,6 +9,7 @@ go_library( "device.go", "fs.go", "full.go", + "net_tun.go", "null.go", "random.go", "tty.go", @@ -19,15 +20,19 @@ go_library( "//pkg/context", "//pkg/rand", "//pkg/safemem", + "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/ramfs", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/kernel", "//pkg/sentry/memmap", "//pkg/sentry/mm", "//pkg/sentry/pgalloc", + "//pkg/sentry/socket/netstack", "//pkg/syserror", + "//pkg/tcpip/link/tun", "//pkg/usermem", "//pkg/waiter", ], diff --git a/pkg/sentry/fs/dev/dev.go b/pkg/sentry/fs/dev/dev.go index 35bd23991..7e66c29b0 100644 --- a/pkg/sentry/fs/dev/dev.go +++ b/pkg/sentry/fs/dev/dev.go @@ -66,8 +66,8 @@ func newMemDevice(ctx context.Context, iops fs.InodeOperations, msrc *fs.MountSo }) } -func newDirectory(ctx context.Context, msrc *fs.MountSource) *fs.Inode { - iops := ramfs.NewDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555)) +func newDirectory(ctx context.Context, contents map[string]*fs.Inode, msrc *fs.MountSource) *fs.Inode { + iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) return fs.NewInode(ctx, iops, msrc, fs.StableAttr{ DeviceID: devDevice.DeviceID(), InodeID: devDevice.NextIno(), @@ -111,7 +111,7 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { // A devpts is typically mounted at /dev/pts to provide // pseudoterminal support. Place an empty directory there for // the devpts to be mounted over. - "pts": newDirectory(ctx, msrc), + "pts": newDirectory(ctx, nil, msrc), // Similarly, applications expect a ptmx device at /dev/ptmx // connected to the terminals provided by /dev/pts/. Rather // than creating a device directly (which requires a hairy @@ -124,6 +124,10 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { "ptmx": newSymlink(ctx, "pts/ptmx", msrc), "tty": newCharacterDevice(ctx, newTTYDevice(ctx, fs.RootOwner, 0666), msrc, ttyDevMajor, ttyDevMinor), + + "net": newDirectory(ctx, map[string]*fs.Inode{ + "tun": newCharacterDevice(ctx, newNetTunDevice(ctx, fs.RootOwner, 0666), msrc, netTunDevMajor, netTunDevMinor), + }, msrc), } iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) diff --git a/pkg/sentry/fs/dev/net_tun.go b/pkg/sentry/fs/dev/net_tun.go new file mode 100644 index 000000000..755644488 --- /dev/null +++ b/pkg/sentry/fs/dev/net_tun.go @@ -0,0 +1,170 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package dev + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/socket/netstack" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip/link/tun" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + netTunDevMajor = 10 + netTunDevMinor = 200 +) + +// +stateify savable +type netTunInodeOperations struct { + fsutil.InodeGenericChecker `state:"nosave"` + fsutil.InodeNoExtendedAttributes `state:"nosave"` + fsutil.InodeNoopAllocate `state:"nosave"` + fsutil.InodeNoopRelease `state:"nosave"` + fsutil.InodeNoopTruncate `state:"nosave"` + fsutil.InodeNoopWriteOut `state:"nosave"` + fsutil.InodeNotDirectory `state:"nosave"` + fsutil.InodeNotMappable `state:"nosave"` + fsutil.InodeNotSocket `state:"nosave"` + fsutil.InodeNotSymlink `state:"nosave"` + fsutil.InodeVirtual `state:"nosave"` + + fsutil.InodeSimpleAttributes +} + +var _ fs.InodeOperations = (*netTunInodeOperations)(nil) + +func newNetTunDevice(ctx context.Context, owner fs.FileOwner, mode linux.FileMode) *netTunInodeOperations { + return &netTunInodeOperations{ + InodeSimpleAttributes: fsutil.NewInodeSimpleAttributes(ctx, owner, fs.FilePermsFromMode(mode), linux.TMPFS_MAGIC), + } +} + +// GetFile implements fs.InodeOperations.GetFile. +func (iops *netTunInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + return fs.NewFile(ctx, d, flags, &netTunFileOperations{}), nil +} + +// +stateify savable +type netTunFileOperations struct { + fsutil.FileNoSeek `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + + device tun.Device +} + +var _ fs.FileOperations = (*netTunFileOperations)(nil) + +// Release implements fs.FileOperations.Release. +func (fops *netTunFileOperations) Release() { + fops.device.Release() +} + +// Ioctl implements fs.FileOperations.Ioctl. +func (fops *netTunFileOperations) Ioctl(ctx context.Context, file *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + request := args[1].Uint() + data := args[2].Pointer() + + switch request { + case linux.TUNSETIFF: + t := kernel.TaskFromContext(ctx) + if t == nil { + panic("Ioctl should be called from a task context") + } + if !t.HasCapability(linux.CAP_NET_ADMIN) { + return 0, syserror.EPERM + } + stack, ok := t.NetworkContext().(*netstack.Stack) + if !ok { + return 0, syserror.EINVAL + } + + var req linux.IFReq + if _, err := usermem.CopyObjectIn(ctx, io, data, &req, usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + return 0, err + } + flags := usermem.ByteOrder.Uint16(req.Data[:]) + return 0, fops.device.SetIff(stack.Stack, req.Name(), flags) + + case linux.TUNGETIFF: + var req linux.IFReq + + copy(req.IFName[:], fops.device.Name()) + + // Linux adds IFF_NOFILTER (the same value as IFF_NO_PI unfortunately) when + // there is no sk_filter. See __tun_chr_ioctl() in net/drivers/tun.c. + flags := fops.device.Flags() | linux.IFF_NOFILTER + usermem.ByteOrder.PutUint16(req.Data[:], flags) + + _, err := usermem.CopyObjectOut(ctx, io, data, &req, usermem.IOOpts{ + AddressSpaceActive: true, + }) + return 0, err + + default: + return 0, syserror.ENOTTY + } +} + +// Write implements fs.FileOperations.Write. +func (fops *netTunFileOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + data := make([]byte, src.NumBytes()) + if _, err := src.CopyIn(ctx, data); err != nil { + return 0, err + } + return fops.device.Write(data) +} + +// Read implements fs.FileOperations.Read. +func (fops *netTunFileOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + data, err := fops.device.Read() + if err != nil { + return 0, err + } + n, err := dst.CopyOut(ctx, data) + if n > 0 && n < len(data) { + // Not an error for partial copying. Packet truncated. + err = nil + } + return int64(n), err +} + +// Readiness implements watier.Waitable.Readiness. +func (fops *netTunFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask { + return fops.device.Readiness(mask) +} + +// EventRegister implements watier.Waitable.EventRegister. +func (fops *netTunFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + fops.device.EventRegister(e, mask) +} + +// EventUnregister implements watier.Waitable.EventUnregister. +func (fops *netTunFileOperations) EventUnregister(e *waiter.Entry) { + fops.device.EventUnregister(e) +} diff --git a/pkg/syserror/syserror.go b/pkg/syserror/syserror.go index 2269f6237..4b5a0fca6 100644 --- a/pkg/syserror/syserror.go +++ b/pkg/syserror/syserror.go @@ -29,6 +29,7 @@ var ( EACCES = error(syscall.EACCES) EAGAIN = error(syscall.EAGAIN) EBADF = error(syscall.EBADF) + EBADFD = error(syscall.EBADFD) EBUSY = error(syscall.EBUSY) ECHILD = error(syscall.ECHILD) ECONNREFUSED = error(syscall.ECONNREFUSED) diff --git a/pkg/tcpip/buffer/view.go b/pkg/tcpip/buffer/view.go index 150310c11..17e94c562 100644 --- a/pkg/tcpip/buffer/view.go +++ b/pkg/tcpip/buffer/view.go @@ -156,3 +156,9 @@ func (vv *VectorisedView) Append(vv2 VectorisedView) { vv.views = append(vv.views, vv2.views...) vv.size += vv2.size } + +// AppendView appends the given view into this vectorised view. +func (vv *VectorisedView) AppendView(v View) { + vv.views = append(vv.views, v) + vv.size += len(v) +} diff --git a/pkg/tcpip/link/channel/BUILD b/pkg/tcpip/link/channel/BUILD index 3974c464e..b8b93e78e 100644 --- a/pkg/tcpip/link/channel/BUILD +++ b/pkg/tcpip/link/channel/BUILD @@ -7,6 +7,7 @@ go_library( srcs = ["channel.go"], visibility = ["//visibility:public"], deps = [ + "//pkg/sync", "//pkg/tcpip", "//pkg/tcpip/buffer", "//pkg/tcpip/stack", diff --git a/pkg/tcpip/link/channel/channel.go b/pkg/tcpip/link/channel/channel.go index 78d447acd..5944ba190 100644 --- a/pkg/tcpip/link/channel/channel.go +++ b/pkg/tcpip/link/channel/channel.go @@ -20,6 +20,7 @@ package channel import ( "context" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/stack" @@ -33,6 +34,118 @@ type PacketInfo struct { Route stack.Route } +// Notification is the interface for receiving notification from the packet +// queue. +type Notification interface { + // WriteNotify will be called when a write happens to the queue. + WriteNotify() +} + +// NotificationHandle is an opaque handle to the registered notification target. +// It can be used to unregister the notification when no longer interested. +// +// +stateify savable +type NotificationHandle struct { + n Notification +} + +type queue struct { + // mu protects fields below. + mu sync.RWMutex + // c is the outbound packet channel. Sending to c should hold mu. + c chan PacketInfo + numWrite int + numRead int + notify []*NotificationHandle +} + +func (q *queue) Close() { + close(q.c) +} + +func (q *queue) Read() (PacketInfo, bool) { + q.mu.Lock() + defer q.mu.Unlock() + select { + case p := <-q.c: + q.numRead++ + return p, true + default: + return PacketInfo{}, false + } +} + +func (q *queue) ReadContext(ctx context.Context) (PacketInfo, bool) { + // We have to receive from channel without holding the lock, since it can + // block indefinitely. This will cause a window that numWrite - numRead + // produces a larger number, but won't go to negative. numWrite >= numRead + // still holds. + select { + case pkt := <-q.c: + q.mu.Lock() + defer q.mu.Unlock() + q.numRead++ + return pkt, true + case <-ctx.Done(): + return PacketInfo{}, false + } +} + +func (q *queue) Write(p PacketInfo) bool { + wrote := false + + // It's important to make sure nobody can see numWrite until we increment it, + // so numWrite >= numRead holds. + q.mu.Lock() + select { + case q.c <- p: + wrote = true + q.numWrite++ + default: + } + notify := q.notify + q.mu.Unlock() + + if wrote { + // Send notification outside of lock. + for _, h := range notify { + h.n.WriteNotify() + } + } + return wrote +} + +func (q *queue) Num() int { + q.mu.RLock() + defer q.mu.RUnlock() + n := q.numWrite - q.numRead + if n < 0 { + panic("numWrite < numRead") + } + return n +} + +func (q *queue) AddNotify(notify Notification) *NotificationHandle { + q.mu.Lock() + defer q.mu.Unlock() + h := &NotificationHandle{n: notify} + q.notify = append(q.notify, h) + return h +} + +func (q *queue) RemoveNotify(handle *NotificationHandle) { + q.mu.Lock() + defer q.mu.Unlock() + // Make a copy, since we reads the array outside of lock when notifying. + notify := make([]*NotificationHandle, 0, len(q.notify)) + for _, h := range q.notify { + if h != handle { + notify = append(notify, h) + } + } + q.notify = notify +} + // Endpoint is link layer endpoint that stores outbound packets in a channel // and allows injection of inbound packets. type Endpoint struct { @@ -41,14 +154,16 @@ type Endpoint struct { linkAddr tcpip.LinkAddress LinkEPCapabilities stack.LinkEndpointCapabilities - // c is where outbound packets are queued. - c chan PacketInfo + // Outbound packet queue. + q *queue } // New creates a new channel endpoint. func New(size int, mtu uint32, linkAddr tcpip.LinkAddress) *Endpoint { return &Endpoint{ - c: make(chan PacketInfo, size), + q: &queue{ + c: make(chan PacketInfo, size), + }, mtu: mtu, linkAddr: linkAddr, } @@ -57,43 +172,36 @@ func New(size int, mtu uint32, linkAddr tcpip.LinkAddress) *Endpoint { // Close closes e. Further packet injections will panic. Reads continue to // succeed until all packets are read. func (e *Endpoint) Close() { - close(e.c) + e.q.Close() } -// Read does non-blocking read for one packet from the outbound packet queue. +// Read does non-blocking read one packet from the outbound packet queue. func (e *Endpoint) Read() (PacketInfo, bool) { - select { - case pkt := <-e.c: - return pkt, true - default: - return PacketInfo{}, false - } + return e.q.Read() } // ReadContext does blocking read for one packet from the outbound packet queue. // It can be cancelled by ctx, and in this case, it returns false. func (e *Endpoint) ReadContext(ctx context.Context) (PacketInfo, bool) { - select { - case pkt := <-e.c: - return pkt, true - case <-ctx.Done(): - return PacketInfo{}, false - } + return e.q.ReadContext(ctx) } // Drain removes all outbound packets from the channel and counts them. func (e *Endpoint) Drain() int { c := 0 for { - select { - case <-e.c: - c++ - default: + if _, ok := e.Read(); !ok { return c } + c++ } } +// NumQueued returns the number of packet queued for outbound. +func (e *Endpoint) NumQueued() int { + return e.q.Num() +} + // InjectInbound injects an inbound packet. func (e *Endpoint) InjectInbound(protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) { e.InjectLinkAddr(protocol, "", pkt) @@ -155,10 +263,7 @@ func (e *Endpoint) WritePacket(r *stack.Route, gso *stack.GSO, protocol tcpip.Ne Route: route, } - select { - case e.c <- p: - default: - } + e.q.Write(p) return nil } @@ -171,7 +276,6 @@ func (e *Endpoint) WritePackets(r *stack.Route, gso *stack.GSO, pkts []tcpip.Pac route.Release() payloadView := pkts[0].Data.ToView() n := 0 -packetLoop: for _, pkt := range pkts { off := pkt.DataOffset size := pkt.DataSize @@ -185,12 +289,10 @@ packetLoop: Route: route, } - select { - case e.c <- p: - n++ - default: - break packetLoop + if !e.q.Write(p) { + break } + n++ } return n, nil @@ -204,13 +306,21 @@ func (e *Endpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Error { GSO: nil, } - select { - case e.c <- p: - default: - } + e.q.Write(p) return nil } // Wait implements stack.LinkEndpoint.Wait. func (*Endpoint) Wait() {} + +// AddNotify adds a notification target for receiving event about outgoing +// packets. +func (e *Endpoint) AddNotify(notify Notification) *NotificationHandle { + return e.q.AddNotify(notify) +} + +// RemoveNotify removes handle from the list of notification targets. +func (e *Endpoint) RemoveNotify(handle *NotificationHandle) { + e.q.RemoveNotify(handle) +} diff --git a/pkg/tcpip/link/tun/BUILD b/pkg/tcpip/link/tun/BUILD index e5096ea38..e0db6cf54 100644 --- a/pkg/tcpip/link/tun/BUILD +++ b/pkg/tcpip/link/tun/BUILD @@ -4,6 +4,22 @@ package(licenses = ["notice"]) go_library( name = "tun", - srcs = ["tun_unsafe.go"], + srcs = [ + "device.go", + "protocol.go", + "tun_unsafe.go", + ], visibility = ["//visibility:public"], + deps = [ + "//pkg/abi/linux", + "//pkg/refs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/tcpip", + "//pkg/tcpip/buffer", + "//pkg/tcpip/header", + "//pkg/tcpip/link/channel", + "//pkg/tcpip/stack", + "//pkg/waiter", + ], ) diff --git a/pkg/tcpip/link/tun/device.go b/pkg/tcpip/link/tun/device.go new file mode 100644 index 000000000..6ff47a742 --- /dev/null +++ b/pkg/tcpip/link/tun/device.go @@ -0,0 +1,352 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tun + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/buffer" + "gvisor.dev/gvisor/pkg/tcpip/header" + "gvisor.dev/gvisor/pkg/tcpip/link/channel" + "gvisor.dev/gvisor/pkg/tcpip/stack" + "gvisor.dev/gvisor/pkg/waiter" +) + +const ( + // drivers/net/tun.c:tun_net_init() + defaultDevMtu = 1500 + + // Queue length for outbound packet, arriving at fd side for read. Overflow + // causes packet drops. gVisor implementation-specific. + defaultDevOutQueueLen = 1024 +) + +var zeroMAC [6]byte + +// Device is an opened /dev/net/tun device. +// +// +stateify savable +type Device struct { + waiter.Queue + + mu sync.RWMutex `state:"nosave"` + endpoint *tunEndpoint + notifyHandle *channel.NotificationHandle + flags uint16 +} + +// beforeSave is invoked by stateify. +func (d *Device) beforeSave() { + d.mu.Lock() + defer d.mu.Unlock() + // TODO(b/110961832): Restore the device to stack. At this moment, the stack + // is not savable. + if d.endpoint != nil { + panic("/dev/net/tun does not support save/restore when a device is associated with it.") + } +} + +// Release implements fs.FileOperations.Release. +func (d *Device) Release() { + d.mu.Lock() + defer d.mu.Unlock() + + // Decrease refcount if there is an endpoint associated with this file. + if d.endpoint != nil { + d.endpoint.RemoveNotify(d.notifyHandle) + d.endpoint.DecRef() + d.endpoint = nil + } +} + +// SetIff services TUNSETIFF ioctl(2) request. +func (d *Device) SetIff(s *stack.Stack, name string, flags uint16) error { + d.mu.Lock() + defer d.mu.Unlock() + + if d.endpoint != nil { + return syserror.EINVAL + } + + // Input validations. + isTun := flags&linux.IFF_TUN != 0 + isTap := flags&linux.IFF_TAP != 0 + supportedFlags := uint16(linux.IFF_TUN | linux.IFF_TAP | linux.IFF_NO_PI) + if isTap && isTun || !isTap && !isTun || flags&^supportedFlags != 0 { + return syserror.EINVAL + } + + prefix := "tun" + if isTap { + prefix = "tap" + } + + endpoint, err := attachOrCreateNIC(s, name, prefix) + if err != nil { + return syserror.EINVAL + } + + d.endpoint = endpoint + d.notifyHandle = d.endpoint.AddNotify(d) + d.flags = flags + return nil +} + +func attachOrCreateNIC(s *stack.Stack, name, prefix string) (*tunEndpoint, error) { + for { + // 1. Try to attach to an existing NIC. + if name != "" { + if nic, found := s.GetNICByName(name); found { + endpoint, ok := nic.LinkEndpoint().(*tunEndpoint) + if !ok { + // Not a NIC created by tun device. + return nil, syserror.EOPNOTSUPP + } + if !endpoint.TryIncRef() { + // Race detected: NIC got deleted in between. + continue + } + return endpoint, nil + } + } + + // 2. Creating a new NIC. + id := tcpip.NICID(s.UniqueID()) + endpoint := &tunEndpoint{ + Endpoint: channel.New(defaultDevOutQueueLen, defaultDevMtu, ""), + stack: s, + nicID: id, + name: name, + } + if endpoint.name == "" { + endpoint.name = fmt.Sprintf("%s%d", prefix, id) + } + err := s.CreateNICWithOptions(endpoint.nicID, endpoint, stack.NICOptions{ + Name: endpoint.name, + }) + switch err { + case nil: + return endpoint, nil + case tcpip.ErrDuplicateNICID: + // Race detected: A NIC has been created in between. + continue + default: + return nil, syserror.EINVAL + } + } +} + +// Write inject one inbound packet to the network interface. +func (d *Device) Write(data []byte) (int64, error) { + d.mu.RLock() + endpoint := d.endpoint + d.mu.RUnlock() + if endpoint == nil { + return 0, syserror.EBADFD + } + if !endpoint.IsAttached() { + return 0, syserror.EIO + } + + dataLen := int64(len(data)) + + // Packet information. + var pktInfoHdr PacketInfoHeader + if !d.hasFlags(linux.IFF_NO_PI) { + if len(data) < PacketInfoHeaderSize { + // Ignore bad packet. + return dataLen, nil + } + pktInfoHdr = PacketInfoHeader(data[:PacketInfoHeaderSize]) + data = data[PacketInfoHeaderSize:] + } + + // Ethernet header (TAP only). + var ethHdr header.Ethernet + if d.hasFlags(linux.IFF_TAP) { + if len(data) < header.EthernetMinimumSize { + // Ignore bad packet. + return dataLen, nil + } + ethHdr = header.Ethernet(data[:header.EthernetMinimumSize]) + data = data[header.EthernetMinimumSize:] + } + + // Try to determine network protocol number, default zero. + var protocol tcpip.NetworkProtocolNumber + switch { + case pktInfoHdr != nil: + protocol = pktInfoHdr.Protocol() + case ethHdr != nil: + protocol = ethHdr.Type() + } + + // Try to determine remote link address, default zero. + var remote tcpip.LinkAddress + switch { + case ethHdr != nil: + remote = ethHdr.SourceAddress() + default: + remote = tcpip.LinkAddress(zeroMAC[:]) + } + + pkt := tcpip.PacketBuffer{ + Data: buffer.View(data).ToVectorisedView(), + } + if ethHdr != nil { + pkt.LinkHeader = buffer.View(ethHdr) + } + endpoint.InjectLinkAddr(protocol, remote, pkt) + return dataLen, nil +} + +// Read reads one outgoing packet from the network interface. +func (d *Device) Read() ([]byte, error) { + d.mu.RLock() + endpoint := d.endpoint + d.mu.RUnlock() + if endpoint == nil { + return nil, syserror.EBADFD + } + + for { + info, ok := endpoint.Read() + if !ok { + return nil, syserror.ErrWouldBlock + } + + v, ok := d.encodePkt(&info) + if !ok { + // Ignore unsupported packet. + continue + } + return v, nil + } +} + +// encodePkt encodes packet for fd side. +func (d *Device) encodePkt(info *channel.PacketInfo) (buffer.View, bool) { + var vv buffer.VectorisedView + + // Packet information. + if !d.hasFlags(linux.IFF_NO_PI) { + hdr := make(PacketInfoHeader, PacketInfoHeaderSize) + hdr.Encode(&PacketInfoFields{ + Protocol: info.Proto, + }) + vv.AppendView(buffer.View(hdr)) + } + + // If the packet does not already have link layer header, and the route + // does not exist, we can't compute it. This is possibly a raw packet, tun + // device doesn't support this at the moment. + if info.Pkt.LinkHeader == nil && info.Route.RemoteLinkAddress == "" { + return nil, false + } + + // Ethernet header (TAP only). + if d.hasFlags(linux.IFF_TAP) { + // Add ethernet header if not provided. + if info.Pkt.LinkHeader == nil { + hdr := &header.EthernetFields{ + SrcAddr: info.Route.LocalLinkAddress, + DstAddr: info.Route.RemoteLinkAddress, + Type: info.Proto, + } + if hdr.SrcAddr == "" { + hdr.SrcAddr = d.endpoint.LinkAddress() + } + + eth := make(header.Ethernet, header.EthernetMinimumSize) + eth.Encode(hdr) + vv.AppendView(buffer.View(eth)) + } else { + vv.AppendView(info.Pkt.LinkHeader) + } + } + + // Append upper headers. + vv.AppendView(buffer.View(info.Pkt.Header.View()[len(info.Pkt.LinkHeader):])) + // Append data payload. + vv.Append(info.Pkt.Data) + + return vv.ToView(), true +} + +// Name returns the name of the attached network interface. Empty string if +// unattached. +func (d *Device) Name() string { + d.mu.RLock() + defer d.mu.RUnlock() + if d.endpoint != nil { + return d.endpoint.name + } + return "" +} + +// Flags returns the flags set for d. Zero value if unset. +func (d *Device) Flags() uint16 { + d.mu.RLock() + defer d.mu.RUnlock() + return d.flags +} + +func (d *Device) hasFlags(flags uint16) bool { + return d.flags&flags == flags +} + +// Readiness implements watier.Waitable.Readiness. +func (d *Device) Readiness(mask waiter.EventMask) waiter.EventMask { + if mask&waiter.EventIn != 0 { + d.mu.RLock() + endpoint := d.endpoint + d.mu.RUnlock() + if endpoint != nil && endpoint.NumQueued() == 0 { + mask &= ^waiter.EventIn + } + } + return mask & (waiter.EventIn | waiter.EventOut) +} + +// WriteNotify implements channel.Notification.WriteNotify. +func (d *Device) WriteNotify() { + d.Notify(waiter.EventIn) +} + +// tunEndpoint is the link endpoint for the NIC created by the tun device. +// +// It is ref-counted as multiple opening files can attach to the same NIC. +// The last owner is responsible for deleting the NIC. +type tunEndpoint struct { + *channel.Endpoint + + refs.AtomicRefCount + + stack *stack.Stack + nicID tcpip.NICID + name string +} + +// DecRef decrements refcount of e, removes NIC if refcount goes to 0. +func (e *tunEndpoint) DecRef() { + e.DecRefWithDestructor(func() { + e.stack.RemoveNIC(e.nicID) + }) +} diff --git a/pkg/tcpip/link/tun/protocol.go b/pkg/tcpip/link/tun/protocol.go new file mode 100644 index 000000000..89d9d91a9 --- /dev/null +++ b/pkg/tcpip/link/tun/protocol.go @@ -0,0 +1,56 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package tun + +import ( + "encoding/binary" + + "gvisor.dev/gvisor/pkg/tcpip" +) + +const ( + // PacketInfoHeaderSize is the size of the packet information header. + PacketInfoHeaderSize = 4 + + offsetFlags = 0 + offsetProtocol = 2 +) + +// PacketInfoFields contains fields sent through the wire if IFF_NO_PI flag is +// not set. +type PacketInfoFields struct { + Flags uint16 + Protocol tcpip.NetworkProtocolNumber +} + +// PacketInfoHeader is the wire representation of the packet information sent if +// IFF_NO_PI flag is not set. +type PacketInfoHeader []byte + +// Encode encodes f into h. +func (h PacketInfoHeader) Encode(f *PacketInfoFields) { + binary.BigEndian.PutUint16(h[offsetFlags:][:2], f.Flags) + binary.BigEndian.PutUint16(h[offsetProtocol:][:2], uint16(f.Protocol)) +} + +// Flags returns the flag field in h. +func (h PacketInfoHeader) Flags() uint16 { + return binary.BigEndian.Uint16(h[offsetFlags:]) +} + +// Protocol returns the protocol field in h. +func (h PacketInfoHeader) Protocol() tcpip.NetworkProtocolNumber { + return tcpip.NetworkProtocolNumber(binary.BigEndian.Uint16(h[offsetProtocol:])) +} diff --git a/pkg/tcpip/stack/nic.go b/pkg/tcpip/stack/nic.go index 862954ab2..46d3a6646 100644 --- a/pkg/tcpip/stack/nic.go +++ b/pkg/tcpip/stack/nic.go @@ -298,6 +298,33 @@ func (n *NIC) enable() *tcpip.Error { return nil } +// remove detaches NIC from the link endpoint, and marks existing referenced +// network endpoints expired. This guarantees no packets between this NIC and +// the network stack. +func (n *NIC) remove() *tcpip.Error { + n.mu.Lock() + defer n.mu.Unlock() + + // Detach from link endpoint, so no packet comes in. + n.linkEP.Attach(nil) + + // Remove permanent and permanentTentative addresses, so no packet goes out. + var errs []*tcpip.Error + for nid, ref := range n.mu.endpoints { + switch ref.getKind() { + case permanentTentative, permanent: + if err := n.removePermanentAddressLocked(nid.LocalAddress); err != nil { + errs = append(errs, err) + } + } + } + if len(errs) > 0 { + return errs[0] + } + + return nil +} + // becomeIPv6Router transitions n into an IPv6 router. // // When transitioning into an IPv6 router, host-only state (NDP discovered @@ -1302,6 +1329,11 @@ func (n *NIC) Stack() *Stack { return n.stack } +// LinkEndpoint returns the link endpoint of n. +func (n *NIC) LinkEndpoint() LinkEndpoint { + return n.linkEP +} + // isAddrTentative returns true if addr is tentative on n. // // Note that if addr is not associated with n, then this function will return diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go index f0ed76fbe..900dd46c5 100644 --- a/pkg/tcpip/stack/stack.go +++ b/pkg/tcpip/stack/stack.go @@ -916,6 +916,18 @@ func (s *Stack) CreateNIC(id tcpip.NICID, ep LinkEndpoint) *tcpip.Error { return s.CreateNICWithOptions(id, ep, NICOptions{}) } +// GetNICByName gets the NIC specified by name. +func (s *Stack) GetNICByName(name string) (*NIC, bool) { + s.mu.RLock() + defer s.mu.RUnlock() + for _, nic := range s.nics { + if nic.Name() == name { + return nic, true + } + } + return nil, false +} + // EnableNIC enables the given NIC so that the link-layer endpoint can start // delivering packets to it. func (s *Stack) EnableNIC(id tcpip.NICID) *tcpip.Error { @@ -956,6 +968,33 @@ func (s *Stack) CheckNIC(id tcpip.NICID) bool { return nic.enabled() } +// RemoveNIC removes NIC and all related routes from the network stack. +func (s *Stack) RemoveNIC(id tcpip.NICID) *tcpip.Error { + s.mu.Lock() + defer s.mu.Unlock() + + nic, ok := s.nics[id] + if !ok { + return tcpip.ErrUnknownNICID + } + delete(s.nics, id) + + // Remove routes in-place. n tracks the number of routes written. + n := 0 + for i, r := range s.routeTable { + if r.NIC != id { + // Keep this route. + if i > n { + s.routeTable[n] = r + } + n++ + } + } + s.routeTable = s.routeTable[:n] + + return nic.remove() +} + // NICAddressRanges returns a map of NICIDs to their associated subnets. func (s *Stack) NICAddressRanges() map[tcpip.NICID][]tcpip.Subnet { s.mu.RLock() diff --git a/test/syscalls/BUILD b/test/syscalls/BUILD index d1977d4de..3518e862d 100644 --- a/test/syscalls/BUILD +++ b/test/syscalls/BUILD @@ -678,6 +678,8 @@ syscall_test( test = "//test/syscalls/linux:truncate_test", ) +syscall_test(test = "//test/syscalls/linux:tuntap_test") + syscall_test(test = "//test/syscalls/linux:udp_bind_test") syscall_test( diff --git a/test/syscalls/linux/BUILD b/test/syscalls/linux/BUILD index aa303af84..704bae17b 100644 --- a/test/syscalls/linux/BUILD +++ b/test/syscalls/linux/BUILD @@ -131,6 +131,17 @@ cc_library( ], ) +cc_library( + name = "socket_netlink_route_util", + testonly = 1, + srcs = ["socket_netlink_route_util.cc"], + hdrs = ["socket_netlink_route_util.h"], + deps = [ + ":socket_netlink_util", + "@com_google_absl//absl/types:optional", + ], +) + cc_library( name = "socket_test_util", testonly = 1, @@ -3430,6 +3441,25 @@ cc_binary( ], ) +cc_binary( + name = "tuntap_test", + testonly = 1, + srcs = ["tuntap.cc"], + linkstatic = 1, + deps = [ + ":socket_test_util", + gtest, + "//test/syscalls/linux:socket_netlink_route_util", + "//test/util:capability_util", + "//test/util:file_descriptor", + "//test/util:fs_util", + "//test/util:posix_error", + "//test/util:test_main", + "//test/util:test_util", + "@com_google_absl//absl/strings", + ], +) + cc_library( name = "udp_socket_test_cases", testonly = 1, diff --git a/test/syscalls/linux/dev.cc b/test/syscalls/linux/dev.cc index 4dd302eed..4e473268c 100644 --- a/test/syscalls/linux/dev.cc +++ b/test/syscalls/linux/dev.cc @@ -153,6 +153,13 @@ TEST(DevTest, TTYExists) { EXPECT_EQ(statbuf.st_mode, S_IFCHR | 0666); } +TEST(DevTest, NetTunExists) { + struct stat statbuf = {}; + ASSERT_THAT(stat("/dev/net/tun", &statbuf), SyscallSucceeds()); + // Check that it's a character device with rw-rw-rw- permissions. + EXPECT_EQ(statbuf.st_mode, S_IFCHR | 0666); +} + } // namespace } // namespace testing diff --git a/test/syscalls/linux/socket_netlink_route_util.cc b/test/syscalls/linux/socket_netlink_route_util.cc new file mode 100644 index 000000000..53eb3b6b2 --- /dev/null +++ b/test/syscalls/linux/socket_netlink_route_util.cc @@ -0,0 +1,163 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "test/syscalls/linux/socket_netlink_route_util.h" + +#include +#include +#include + +#include "absl/types/optional.h" +#include "test/syscalls/linux/socket_netlink_util.h" + +namespace gvisor { +namespace testing { +namespace { + +constexpr uint32_t kSeq = 12345; + +} // namespace + +PosixError DumpLinks( + const FileDescriptor& fd, uint32_t seq, + const std::function& fn) { + struct request { + struct nlmsghdr hdr; + struct ifinfomsg ifm; + }; + + struct request req = {}; + req.hdr.nlmsg_len = sizeof(req); + req.hdr.nlmsg_type = RTM_GETLINK; + req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_DUMP; + req.hdr.nlmsg_seq = seq; + req.ifm.ifi_family = AF_UNSPEC; + + return NetlinkRequestResponse(fd, &req, sizeof(req), fn, false); +} + +PosixErrorOr> DumpLinks() { + ASSIGN_OR_RETURN_ERRNO(FileDescriptor fd, NetlinkBoundSocket(NETLINK_ROUTE)); + + std::vector links; + RETURN_IF_ERRNO(DumpLinks(fd, kSeq, [&](const struct nlmsghdr* hdr) { + if (hdr->nlmsg_type != RTM_NEWLINK || + hdr->nlmsg_len < NLMSG_SPACE(sizeof(struct ifinfomsg))) { + return; + } + const struct ifinfomsg* msg = + reinterpret_cast(NLMSG_DATA(hdr)); + const auto* rta = FindRtAttr(hdr, msg, IFLA_IFNAME); + if (rta == nullptr) { + // Ignore links that do not have a name. + return; + } + + links.emplace_back(); + links.back().index = msg->ifi_index; + links.back().type = msg->ifi_type; + links.back().name = + std::string(reinterpret_cast(RTA_DATA(rta))); + })); + return links; +} + +PosixErrorOr> FindLoopbackLink() { + ASSIGN_OR_RETURN_ERRNO(auto links, DumpLinks()); + for (const auto& link : links) { + if (link.type == ARPHRD_LOOPBACK) { + return absl::optional(link); + } + } + return absl::optional(); +} + +PosixError LinkAddLocalAddr(int index, int family, int prefixlen, + const void* addr, int addrlen) { + ASSIGN_OR_RETURN_ERRNO(FileDescriptor fd, NetlinkBoundSocket(NETLINK_ROUTE)); + + struct request { + struct nlmsghdr hdr; + struct ifaddrmsg ifaddr; + char attrbuf[512]; + }; + + struct request req = {}; + req.hdr.nlmsg_len = NLMSG_LENGTH(sizeof(req.ifaddr)); + req.hdr.nlmsg_type = RTM_NEWADDR; + req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK; + req.hdr.nlmsg_seq = kSeq; + req.ifaddr.ifa_index = index; + req.ifaddr.ifa_family = family; + req.ifaddr.ifa_prefixlen = prefixlen; + + struct rtattr* rta = reinterpret_cast( + reinterpret_cast(&req) + NLMSG_ALIGN(req.hdr.nlmsg_len)); + rta->rta_type = IFA_LOCAL; + rta->rta_len = RTA_LENGTH(addrlen); + req.hdr.nlmsg_len = NLMSG_ALIGN(req.hdr.nlmsg_len) + RTA_LENGTH(addrlen); + memcpy(RTA_DATA(rta), addr, addrlen); + + return NetlinkRequestAckOrError(fd, kSeq, &req, req.hdr.nlmsg_len); +} + +PosixError LinkChangeFlags(int index, unsigned int flags, unsigned int change) { + ASSIGN_OR_RETURN_ERRNO(FileDescriptor fd, NetlinkBoundSocket(NETLINK_ROUTE)); + + struct request { + struct nlmsghdr hdr; + struct ifinfomsg ifinfo; + char pad[NLMSG_ALIGNTO]; + }; + + struct request req = {}; + req.hdr.nlmsg_len = NLMSG_LENGTH(sizeof(req.ifinfo)); + req.hdr.nlmsg_type = RTM_NEWLINK; + req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK; + req.hdr.nlmsg_seq = kSeq; + req.ifinfo.ifi_index = index; + req.ifinfo.ifi_flags = flags; + req.ifinfo.ifi_change = change; + + return NetlinkRequestAckOrError(fd, kSeq, &req, req.hdr.nlmsg_len); +} + +PosixError LinkSetMacAddr(int index, const void* addr, int addrlen) { + ASSIGN_OR_RETURN_ERRNO(FileDescriptor fd, NetlinkBoundSocket(NETLINK_ROUTE)); + + struct request { + struct nlmsghdr hdr; + struct ifinfomsg ifinfo; + char attrbuf[512]; + }; + + struct request req = {}; + req.hdr.nlmsg_len = NLMSG_LENGTH(sizeof(req.ifinfo)); + req.hdr.nlmsg_type = RTM_NEWLINK; + req.hdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_ACK; + req.hdr.nlmsg_seq = kSeq; + req.ifinfo.ifi_index = index; + + struct rtattr* rta = reinterpret_cast( + reinterpret_cast(&req) + NLMSG_ALIGN(req.hdr.nlmsg_len)); + rta->rta_type = IFLA_ADDRESS; + rta->rta_len = RTA_LENGTH(addrlen); + req.hdr.nlmsg_len = NLMSG_ALIGN(req.hdr.nlmsg_len) + RTA_LENGTH(addrlen); + memcpy(RTA_DATA(rta), addr, addrlen); + + return NetlinkRequestAckOrError(fd, kSeq, &req, req.hdr.nlmsg_len); +} + +} // namespace testing +} // namespace gvisor diff --git a/test/syscalls/linux/socket_netlink_route_util.h b/test/syscalls/linux/socket_netlink_route_util.h new file mode 100644 index 000000000..2c018e487 --- /dev/null +++ b/test/syscalls/linux/socket_netlink_route_util.h @@ -0,0 +1,55 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#ifndef GVISOR_TEST_SYSCALLS_LINUX_SOCKET_NETLINK_ROUTE_UTIL_H_ +#define GVISOR_TEST_SYSCALLS_LINUX_SOCKET_NETLINK_ROUTE_UTIL_H_ + +#include +#include + +#include + +#include "absl/types/optional.h" +#include "test/syscalls/linux/socket_netlink_util.h" + +namespace gvisor { +namespace testing { + +struct Link { + int index; + int16_t type; + std::string name; +}; + +PosixError DumpLinks(const FileDescriptor& fd, uint32_t seq, + const std::function& fn); + +PosixErrorOr> DumpLinks(); + +PosixErrorOr> FindLoopbackLink(); + +// LinkAddLocalAddr sets IFA_LOCAL attribute on the interface. +PosixError LinkAddLocalAddr(int index, int family, int prefixlen, + const void* addr, int addrlen); + +// LinkChangeFlags changes interface flags. E.g. IFF_UP. +PosixError LinkChangeFlags(int index, unsigned int flags, unsigned int change); + +// LinkSetMacAddr sets IFLA_ADDRESS attribute of the interface. +PosixError LinkSetMacAddr(int index, const void* addr, int addrlen); + +} // namespace testing +} // namespace gvisor + +#endif // GVISOR_TEST_SYSCALLS_LINUX_SOCKET_NETLINK_ROUTE_UTIL_H_ diff --git a/test/syscalls/linux/tuntap.cc b/test/syscalls/linux/tuntap.cc new file mode 100644 index 000000000..f6ac9d7b8 --- /dev/null +++ b/test/syscalls/linux/tuntap.cc @@ -0,0 +1,346 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "gmock/gmock.h" +#include "gtest/gtest.h" +#include "absl/strings/ascii.h" +#include "absl/strings/str_split.h" +#include "test/syscalls/linux/socket_netlink_route_util.h" +#include "test/syscalls/linux/socket_test_util.h" +#include "test/util/capability_util.h" +#include "test/util/file_descriptor.h" +#include "test/util/fs_util.h" +#include "test/util/posix_error.h" +#include "test/util/test_util.h" + +namespace gvisor { +namespace testing { +namespace { + +constexpr int kIPLen = 4; + +constexpr const char kDevNetTun[] = "/dev/net/tun"; +constexpr const char kTapName[] = "tap0"; + +constexpr const uint8_t kMacA[ETH_ALEN] = {0xAA, 0xAA, 0xAA, 0xAA, 0xAA, 0xAA}; +constexpr const uint8_t kMacB[ETH_ALEN] = {0xBB, 0xBB, 0xBB, 0xBB, 0xBB, 0xBB}; + +PosixErrorOr> DumpLinkNames() { + ASSIGN_OR_RETURN_ERRNO(auto links, DumpLinks()); + std::set names; + for (const auto& link : links) { + names.emplace(link.name); + } + return names; +} + +PosixErrorOr> GetLinkByName(const std::string& name) { + ASSIGN_OR_RETURN_ERRNO(auto links, DumpLinks()); + for (const auto& link : links) { + if (link.name == name) { + return absl::optional(link); + } + } + return absl::optional(); +} + +struct pihdr { + uint16_t pi_flags; + uint16_t pi_protocol; +} __attribute__((packed)); + +struct ping_pkt { + pihdr pi; + struct ethhdr eth; + struct iphdr ip; + struct icmphdr icmp; + char payload[64]; +} __attribute__((packed)); + +ping_pkt CreatePingPacket(const uint8_t srcmac[ETH_ALEN], const char* srcip, + const uint8_t dstmac[ETH_ALEN], const char* dstip) { + ping_pkt pkt = {}; + + pkt.pi.pi_protocol = htons(ETH_P_IP); + + memcpy(pkt.eth.h_dest, dstmac, sizeof(pkt.eth.h_dest)); + memcpy(pkt.eth.h_source, srcmac, sizeof(pkt.eth.h_source)); + pkt.eth.h_proto = htons(ETH_P_IP); + + pkt.ip.ihl = 5; + pkt.ip.version = 4; + pkt.ip.tos = 0; + pkt.ip.tot_len = htons(sizeof(struct iphdr) + sizeof(struct icmphdr) + + sizeof(pkt.payload)); + pkt.ip.id = 1; + pkt.ip.frag_off = 1 << 6; // Do not fragment + pkt.ip.ttl = 64; + pkt.ip.protocol = IPPROTO_ICMP; + inet_pton(AF_INET, dstip, &pkt.ip.daddr); + inet_pton(AF_INET, srcip, &pkt.ip.saddr); + pkt.ip.check = IPChecksum(pkt.ip); + + pkt.icmp.type = ICMP_ECHO; + pkt.icmp.code = 0; + pkt.icmp.checksum = 0; + pkt.icmp.un.echo.sequence = 1; + pkt.icmp.un.echo.id = 1; + + strncpy(pkt.payload, "abcd", sizeof(pkt.payload)); + pkt.icmp.checksum = ICMPChecksum(pkt.icmp, pkt.payload, sizeof(pkt.payload)); + + return pkt; +} + +struct arp_pkt { + pihdr pi; + struct ethhdr eth; + struct arphdr arp; + uint8_t arp_sha[ETH_ALEN]; + uint8_t arp_spa[kIPLen]; + uint8_t arp_tha[ETH_ALEN]; + uint8_t arp_tpa[kIPLen]; +} __attribute__((packed)); + +std::string CreateArpPacket(const uint8_t srcmac[ETH_ALEN], const char* srcip, + const uint8_t dstmac[ETH_ALEN], const char* dstip) { + std::string buffer; + buffer.resize(sizeof(arp_pkt)); + + arp_pkt* pkt = reinterpret_cast(&buffer[0]); + { + pkt->pi.pi_protocol = htons(ETH_P_ARP); + + memcpy(pkt->eth.h_dest, kMacA, sizeof(pkt->eth.h_dest)); + memcpy(pkt->eth.h_source, kMacB, sizeof(pkt->eth.h_source)); + pkt->eth.h_proto = htons(ETH_P_ARP); + + pkt->arp.ar_hrd = htons(ARPHRD_ETHER); + pkt->arp.ar_pro = htons(ETH_P_IP); + pkt->arp.ar_hln = ETH_ALEN; + pkt->arp.ar_pln = kIPLen; + pkt->arp.ar_op = htons(ARPOP_REPLY); + + memcpy(pkt->arp_sha, srcmac, sizeof(pkt->arp_sha)); + inet_pton(AF_INET, srcip, pkt->arp_spa); + memcpy(pkt->arp_tha, dstmac, sizeof(pkt->arp_tha)); + inet_pton(AF_INET, dstip, pkt->arp_tpa); + } + return buffer; +} + +} // namespace + +class TuntapTest : public ::testing::Test { + protected: + void TearDown() override { + if (ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))) { + // Bring back capability if we had dropped it in test case. + ASSERT_NO_ERRNO(SetCapability(CAP_NET_ADMIN, true)); + } + } +}; + +TEST_F(TuntapTest, CreateInterfaceNoCap) { + SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))); + + ASSERT_NO_ERRNO(SetCapability(CAP_NET_ADMIN, false)); + + FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(kDevNetTun, O_RDWR)); + + struct ifreq ifr = {}; + ifr.ifr_flags = IFF_TAP; + strncpy(ifr.ifr_name, kTapName, IFNAMSIZ); + + EXPECT_THAT(ioctl(fd.get(), TUNSETIFF, &ifr), SyscallFailsWithErrno(EPERM)); +} + +TEST_F(TuntapTest, CreateFixedNameInterface) { + SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))); + + FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(kDevNetTun, O_RDWR)); + + struct ifreq ifr_set = {}; + ifr_set.ifr_flags = IFF_TAP; + strncpy(ifr_set.ifr_name, kTapName, IFNAMSIZ); + EXPECT_THAT(ioctl(fd.get(), TUNSETIFF, &ifr_set), + SyscallSucceedsWithValue(0)); + + struct ifreq ifr_get = {}; + EXPECT_THAT(ioctl(fd.get(), TUNGETIFF, &ifr_get), + SyscallSucceedsWithValue(0)); + + struct ifreq ifr_expect = ifr_set; + // See __tun_chr_ioctl() in net/drivers/tun.c. + ifr_expect.ifr_flags |= IFF_NOFILTER; + + EXPECT_THAT(DumpLinkNames(), + IsPosixErrorOkAndHolds(::testing::Contains(kTapName))); + EXPECT_THAT(memcmp(&ifr_expect, &ifr_get, sizeof(ifr_get)), ::testing::Eq(0)); +} + +TEST_F(TuntapTest, CreateInterface) { + SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))); + + FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(kDevNetTun, O_RDWR)); + + struct ifreq ifr = {}; + ifr.ifr_flags = IFF_TAP; + // Empty ifr.ifr_name. Let kernel assign. + + EXPECT_THAT(ioctl(fd.get(), TUNSETIFF, &ifr), SyscallSucceedsWithValue(0)); + + struct ifreq ifr_get = {}; + EXPECT_THAT(ioctl(fd.get(), TUNGETIFF, &ifr_get), + SyscallSucceedsWithValue(0)); + + std::string ifname = ifr_get.ifr_name; + EXPECT_THAT(ifname, ::testing::StartsWith("tap")); + EXPECT_THAT(DumpLinkNames(), + IsPosixErrorOkAndHolds(::testing::Contains(ifname))); +} + +TEST_F(TuntapTest, InvalidReadWrite) { + SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))); + + FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(kDevNetTun, O_RDWR)); + + char buf[128] = {}; + EXPECT_THAT(read(fd.get(), buf, sizeof(buf)), SyscallFailsWithErrno(EBADFD)); + EXPECT_THAT(write(fd.get(), buf, sizeof(buf)), SyscallFailsWithErrno(EBADFD)); +} + +TEST_F(TuntapTest, WriteToDownDevice) { + SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))); + + // FIXME: gVisor always creates enabled/up'd interfaces. + SKIP_IF(IsRunningOnGvisor()); + + FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(kDevNetTun, O_RDWR)); + + // Device created should be down by default. + struct ifreq ifr = {}; + ifr.ifr_flags = IFF_TAP; + EXPECT_THAT(ioctl(fd.get(), TUNSETIFF, &ifr), SyscallSucceedsWithValue(0)); + + char buf[128] = {}; + EXPECT_THAT(write(fd.get(), buf, sizeof(buf)), SyscallFailsWithErrno(EIO)); +} + +// This test sets up a TAP device and pings kernel by sending ICMP echo request. +// +// It works as the following: +// * Open /dev/net/tun, and create kTapName interface. +// * Use rtnetlink to do initial setup of the interface: +// * Assign IP address 10.0.0.1/24 to kernel. +// * MAC address: kMacA +// * Bring up the interface. +// * Send an ICMP echo reqest (ping) packet from 10.0.0.2 (kMacB) to kernel. +// * Loop to receive packets from TAP device/fd: +// * If packet is an ICMP echo reply, it stops and passes the test. +// * If packet is an ARP request, it responds with canned reply and resends +// the +// ICMP request packet. +TEST_F(TuntapTest, PingKernel) { + SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_NET_ADMIN))); + + // Interface creation. + FileDescriptor fd = ASSERT_NO_ERRNO_AND_VALUE(Open(kDevNetTun, O_RDWR)); + + struct ifreq ifr_set = {}; + ifr_set.ifr_flags = IFF_TAP; + strncpy(ifr_set.ifr_name, kTapName, IFNAMSIZ); + EXPECT_THAT(ioctl(fd.get(), TUNSETIFF, &ifr_set), + SyscallSucceedsWithValue(0)); + + absl::optional link = + ASSERT_NO_ERRNO_AND_VALUE(GetLinkByName(kTapName)); + ASSERT_TRUE(link.has_value()); + + // Interface setup. + struct in_addr addr; + inet_pton(AF_INET, "10.0.0.1", &addr); + EXPECT_NO_ERRNO(LinkAddLocalAddr(link->index, AF_INET, /*prefixlen=*/24, + &addr, sizeof(addr))); + + if (!IsRunningOnGvisor()) { + // FIXME: gVisor doesn't support setting MAC address on interfaces yet. + EXPECT_NO_ERRNO(LinkSetMacAddr(link->index, kMacA, sizeof(kMacA))); + + // FIXME: gVisor always creates enabled/up'd interfaces. + EXPECT_NO_ERRNO(LinkChangeFlags(link->index, IFF_UP, IFF_UP)); + } + + ping_pkt ping_req = CreatePingPacket(kMacB, "10.0.0.2", kMacA, "10.0.0.1"); + std::string arp_rep = CreateArpPacket(kMacB, "10.0.0.2", kMacA, "10.0.0.1"); + + // Send ping, this would trigger an ARP request on Linux. + EXPECT_THAT(write(fd.get(), &ping_req, sizeof(ping_req)), + SyscallSucceedsWithValue(sizeof(ping_req))); + + // Receive loop to process inbound packets. + struct inpkt { + union { + pihdr pi; + ping_pkt ping; + arp_pkt arp; + }; + }; + while (1) { + inpkt r = {}; + int n = read(fd.get(), &r, sizeof(r)); + EXPECT_THAT(n, SyscallSucceeds()); + + if (n < sizeof(pihdr)) { + std::cerr << "Ignored packet, protocol: " << r.pi.pi_protocol + << " len: " << n << std::endl; + continue; + } + + // Process ARP packet. + if (n >= sizeof(arp_pkt) && r.pi.pi_protocol == htons(ETH_P_ARP)) { + // Respond with canned ARP reply. + EXPECT_THAT(write(fd.get(), arp_rep.data(), arp_rep.size()), + SyscallSucceedsWithValue(arp_rep.size())); + // First ping request might have been dropped due to mac address not in + // ARP cache. Send it again. + EXPECT_THAT(write(fd.get(), &ping_req, sizeof(ping_req)), + SyscallSucceedsWithValue(sizeof(ping_req))); + } + + // Process ping response packet. + if (n >= sizeof(ping_pkt) && r.pi.pi_protocol == ping_req.pi.pi_protocol && + r.ping.ip.protocol == ping_req.ip.protocol && + !memcmp(&r.ping.ip.saddr, &ping_req.ip.daddr, kIPLen) && + !memcmp(&r.ping.ip.daddr, &ping_req.ip.saddr, kIPLen) && + r.ping.icmp.type == 0 && r.ping.icmp.code == 0) { + // Ends and passes the test. + break; + } + } +} + +} // namespace testing +} // namespace gvisor -- cgit v1.2.3 From 53504e29ca27b8dc9e098fbb88983fdbce90cca3 Mon Sep 17 00:00:00 2001 From: Adin Scannell Date: Tue, 25 Feb 2020 12:16:43 -0800 Subject: Fix mount refcount issue. Each mount is holds a reference on a root Dirent, but the mount itself may live beyond it's own reference. This means that a call to Root() can come after the associated reference has been dropped. Instead of introducing a separate layer of references for mount objects, we simply change the Root() method to use TryIncRef() and allow it to return nil if the mount is already gone. This requires updating a small number of callers and minimizes the change (since VFSv2 will replace this code shortly). PiperOrigin-RevId: 297174230 --- pkg/sentry/fs/mount_test.go | 11 ++++++----- pkg/sentry/fs/mounts.go | 10 +++++++--- pkg/sentry/fs/proc/mounts.go | 16 ++++++++++++---- 3 files changed, 25 insertions(+), 12 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/mount_test.go b/pkg/sentry/fs/mount_test.go index e672a438c..a3d10770b 100644 --- a/pkg/sentry/fs/mount_test.go +++ b/pkg/sentry/fs/mount_test.go @@ -36,11 +36,12 @@ func mountPathsAre(root *Dirent, got []*Mount, want ...string) error { gotPaths := make(map[string]struct{}, len(got)) gotStr := make([]string, len(got)) for i, g := range got { - groot := g.Root() - name, _ := groot.FullName(root) - groot.DecRef() - gotStr[i] = name - gotPaths[name] = struct{}{} + if groot := g.Root(); groot != nil { + name, _ := groot.FullName(root) + groot.DecRef() + gotStr[i] = name + gotPaths[name] = struct{}{} + } } if len(got) != len(want) { return fmt.Errorf("mount paths are different, got: %q, want: %q", gotStr, want) diff --git a/pkg/sentry/fs/mounts.go b/pkg/sentry/fs/mounts.go index 574a2cc91..c7981f66e 100644 --- a/pkg/sentry/fs/mounts.go +++ b/pkg/sentry/fs/mounts.go @@ -100,10 +100,14 @@ func newUndoMount(d *Dirent) *Mount { } } -// Root returns the root dirent of this mount. Callers must call DecRef on the -// returned dirent. +// Root returns the root dirent of this mount. +// +// This may return nil if the mount has already been free. Callers must handle this +// case appropriately. If non-nil, callers must call DecRef on the returned *Dirent. func (m *Mount) Root() *Dirent { - m.root.IncRef() + if !m.root.TryIncRef() { + return nil + } return m.root } diff --git a/pkg/sentry/fs/proc/mounts.go b/pkg/sentry/fs/proc/mounts.go index c10888100..94deb553b 100644 --- a/pkg/sentry/fs/proc/mounts.go +++ b/pkg/sentry/fs/proc/mounts.go @@ -60,13 +60,15 @@ func forEachMount(t *kernel.Task, fn func(string, *fs.Mount)) { }) for _, m := range ms { mroot := m.Root() + if mroot == nil { + continue // No longer valid. + } mountPath, desc := mroot.FullName(rootDir) mroot.DecRef() if !desc { // MountSources that are not descendants of the chroot jail are ignored. continue } - fn(mountPath, m) } } @@ -91,6 +93,12 @@ func (mif *mountInfoFile) ReadSeqFileData(ctx context.Context, handle seqfile.Se var buf bytes.Buffer forEachMount(mif.t, func(mountPath string, m *fs.Mount) { + mroot := m.Root() + if mroot == nil { + return // No longer valid. + } + defer mroot.DecRef() + // Format: // 36 35 98:0 /mnt1 /mnt2 rw,noatime master:1 - ext3 /dev/root rw,errors=continue // (1)(2)(3) (4) (5) (6) (7) (8) (9) (10) (11) @@ -107,9 +115,6 @@ func (mif *mountInfoFile) ReadSeqFileData(ctx context.Context, handle seqfile.Se // (3) Major:Minor device ID. We don't have a superblock, so we // just use the root inode device number. - mroot := m.Root() - defer mroot.DecRef() - sa := mroot.Inode.StableAttr fmt.Fprintf(&buf, "%d:%d ", sa.DeviceFileMajor, sa.DeviceFileMinor) @@ -207,6 +212,9 @@ func (mf *mountsFile) ReadSeqFileData(ctx context.Context, handle seqfile.SeqHan // // The "needs dump"and fsck flags are always 0, which is allowed. root := m.Root() + if root == nil { + return // No longer valid. + } defer root.DecRef() flags := root.Inode.MountSource.Flags -- cgit v1.2.3 From fba479b3c78621cb122af20d1d677fe9193a971c Mon Sep 17 00:00:00 2001 From: Adin Scannell Date: Tue, 25 Feb 2020 19:03:23 -0800 Subject: Fix DATA RACE in fs.MayDelete. MayDelete must lock the directory also, otherwise concurrent renames may race. Note that this also changes the methods to be aligned with the actual Remove and RemoveDirectory methods to minimize confusion when reading the code. (It was hard to see that resolution was correct.) PiperOrigin-RevId: 297258304 --- pkg/sentry/fs/dirent.go | 25 ++++++++++++++----------- pkg/sentry/syscalls/linux/sys_file.go | 4 ++-- 2 files changed, 16 insertions(+), 13 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/dirent.go b/pkg/sentry/fs/dirent.go index acab0411a..e0b32e1c1 100644 --- a/pkg/sentry/fs/dirent.go +++ b/pkg/sentry/fs/dirent.go @@ -1438,8 +1438,8 @@ func lockForRename(oldParent *Dirent, oldName string, newParent *Dirent, newName }, nil } -func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error { - uattr, err := dir.Inode.UnstableAttr(ctx) +func (d *Dirent) checkSticky(ctx context.Context, victim *Dirent) error { + uattr, err := d.Inode.UnstableAttr(ctx) if err != nil { return syserror.EPERM } @@ -1465,30 +1465,33 @@ func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error { return syserror.EPERM } -// MayDelete determines whether `name`, a child of `dir`, can be deleted or +// MayDelete determines whether `name`, a child of `d`, can be deleted or // renamed by `ctx`. // // Compare Linux kernel fs/namei.c:may_delete. -func MayDelete(ctx context.Context, root, dir *Dirent, name string) error { - if err := dir.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { +func (d *Dirent) MayDelete(ctx context.Context, root *Dirent, name string) error { + if err := d.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil { return err } - victim, err := dir.Walk(ctx, root, name) + unlock := d.lockDirectory() + defer unlock() + + victim, err := d.walk(ctx, root, name, true /* may unlock */) if err != nil { return err } defer victim.DecRef() - return mayDelete(ctx, dir, victim) + return d.mayDelete(ctx, victim) } // mayDelete determines whether `victim`, a child of `dir`, can be deleted or // renamed by `ctx`. // // Preconditions: `dir` is writable and executable by `ctx`. -func mayDelete(ctx context.Context, dir, victim *Dirent) error { - if err := checkSticky(ctx, dir, victim); err != nil { +func (d *Dirent) mayDelete(ctx context.Context, victim *Dirent) error { + if err := d.checkSticky(ctx, victim); err != nil { return err } @@ -1542,7 +1545,7 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string defer renamed.DecRef() // Check that the renamed dirent is deletable. - if err := mayDelete(ctx, oldParent, renamed); err != nil { + if err := oldParent.mayDelete(ctx, renamed); err != nil { return err } @@ -1580,7 +1583,7 @@ func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string // across the Rename, so must call DecRef manually (no defer). // Check that we can delete replaced. - if err := mayDelete(ctx, newParent, replaced); err != nil { + if err := newParent.mayDelete(ctx, replaced); err != nil { replaced.DecRef() return err } diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go index c21f14dc0..d10a9bed8 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -1236,7 +1236,7 @@ func rmdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTEMPTY } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } @@ -1517,7 +1517,7 @@ func unlinkAt(t *kernel.Task, dirFD int32, addr usermem.Addr) error { return syserror.ENOTDIR } - if err := fs.MayDelete(t, root, d, name); err != nil { + if err := d.MayDelete(t, root, name); err != nil { return err } -- cgit v1.2.3 From 6b4d36e3253238dd72d0861ac1220d147e1de8dd Mon Sep 17 00:00:00 2001 From: Ting-Yu Wang Date: Fri, 28 Feb 2020 10:37:52 -0800 Subject: Hide /dev/net/tun when using hostinet. /dev/net/tun does not currently work with hostinet. This has caused some program starts failing because it thinks the feature exists. PiperOrigin-RevId: 297876196 --- pkg/sentry/fs/dev/BUILD | 1 + pkg/sentry/fs/dev/dev.go | 7 +++++-- pkg/sentry/fs/dev/net_tun.go | 7 +++++++ pkg/sentry/kernel/kernel.go | 4 ++++ test/syscalls/BUILD | 5 +++++ test/syscalls/linux/BUILD | 12 +++++++++++ test/syscalls/linux/dev.cc | 7 ------- test/syscalls/linux/tuntap.cc | 7 +++++++ test/syscalls/linux/tuntap_hostinet.cc | 37 ++++++++++++++++++++++++++++++++++ 9 files changed, 78 insertions(+), 9 deletions(-) create mode 100644 test/syscalls/linux/tuntap_hostinet.cc (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/dev/BUILD b/pkg/sentry/fs/dev/BUILD index 9b6bb26d0..9379a4d7b 100644 --- a/pkg/sentry/fs/dev/BUILD +++ b/pkg/sentry/fs/dev/BUILD @@ -26,6 +26,7 @@ go_library( "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/ramfs", "//pkg/sentry/fs/tmpfs", + "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/memmap", "//pkg/sentry/mm", diff --git a/pkg/sentry/fs/dev/dev.go b/pkg/sentry/fs/dev/dev.go index 7e66c29b0..acbd401a0 100644 --- a/pkg/sentry/fs/dev/dev.go +++ b/pkg/sentry/fs/dev/dev.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/usermem" ) @@ -124,10 +125,12 @@ func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode { "ptmx": newSymlink(ctx, "pts/ptmx", msrc), "tty": newCharacterDevice(ctx, newTTYDevice(ctx, fs.RootOwner, 0666), msrc, ttyDevMajor, ttyDevMinor), + } - "net": newDirectory(ctx, map[string]*fs.Inode{ + if isNetTunSupported(inet.StackFromContext(ctx)) { + contents["net"] = newDirectory(ctx, map[string]*fs.Inode{ "tun": newCharacterDevice(ctx, newNetTunDevice(ctx, fs.RootOwner, 0666), msrc, netTunDevMajor, netTunDevMinor), - }, msrc), + }, msrc) } iops := ramfs.NewDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555)) diff --git a/pkg/sentry/fs/dev/net_tun.go b/pkg/sentry/fs/dev/net_tun.go index 755644488..dc7ad075a 100644 --- a/pkg/sentry/fs/dev/net_tun.go +++ b/pkg/sentry/fs/dev/net_tun.go @@ -20,6 +20,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/socket/netstack" "gvisor.dev/gvisor/pkg/syserror" @@ -168,3 +169,9 @@ func (fops *netTunFileOperations) EventRegister(e *waiter.Entry, mask waiter.Eve func (fops *netTunFileOperations) EventUnregister(e *waiter.Entry) { fops.device.EventUnregister(e) } + +// isNetTunSupported returns whether /dev/net/tun device is supported for s. +func isNetTunSupported(s inet.Stack) bool { + _, ok := s.(*netstack.Stack) + return ok +} diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 8b76750e9..1d627564f 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -755,6 +755,8 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -1481,6 +1483,8 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: diff --git a/test/syscalls/BUILD b/test/syscalls/BUILD index 3518e862d..a69b0ce13 100644 --- a/test/syscalls/BUILD +++ b/test/syscalls/BUILD @@ -680,6 +680,11 @@ syscall_test( syscall_test(test = "//test/syscalls/linux:tuntap_test") +syscall_test( + add_hostinet = True, + test = "//test/syscalls/linux:tuntap_hostinet_test", +) + syscall_test(test = "//test/syscalls/linux:udp_bind_test") syscall_test( diff --git a/test/syscalls/linux/BUILD b/test/syscalls/linux/BUILD index 704bae17b..70c120e42 100644 --- a/test/syscalls/linux/BUILD +++ b/test/syscalls/linux/BUILD @@ -3460,6 +3460,18 @@ cc_binary( ], ) +cc_binary( + name = "tuntap_hostinet_test", + testonly = 1, + srcs = ["tuntap_hostinet.cc"], + linkstatic = 1, + deps = [ + gtest, + "//test/util:test_main", + "//test/util:test_util", + ], +) + cc_library( name = "udp_socket_test_cases", testonly = 1, diff --git a/test/syscalls/linux/dev.cc b/test/syscalls/linux/dev.cc index 4e473268c..4dd302eed 100644 --- a/test/syscalls/linux/dev.cc +++ b/test/syscalls/linux/dev.cc @@ -153,13 +153,6 @@ TEST(DevTest, TTYExists) { EXPECT_EQ(statbuf.st_mode, S_IFCHR | 0666); } -TEST(DevTest, NetTunExists) { - struct stat statbuf = {}; - ASSERT_THAT(stat("/dev/net/tun", &statbuf), SyscallSucceeds()); - // Check that it's a character device with rw-rw-rw- permissions. - EXPECT_EQ(statbuf.st_mode, S_IFCHR | 0666); -} - } // namespace } // namespace testing diff --git a/test/syscalls/linux/tuntap.cc b/test/syscalls/linux/tuntap.cc index f6ac9d7b8..f734511d6 100644 --- a/test/syscalls/linux/tuntap.cc +++ b/test/syscalls/linux/tuntap.cc @@ -153,6 +153,13 @@ std::string CreateArpPacket(const uint8_t srcmac[ETH_ALEN], const char* srcip, } // namespace +TEST(TuntapStaticTest, NetTunExists) { + struct stat statbuf; + ASSERT_THAT(stat(kDevNetTun, &statbuf), SyscallSucceeds()); + // Check that it's a character device with rw-rw-rw- permissions. + EXPECT_EQ(statbuf.st_mode, S_IFCHR | 0666); +} + class TuntapTest : public ::testing::Test { protected: void TearDown() override { diff --git a/test/syscalls/linux/tuntap_hostinet.cc b/test/syscalls/linux/tuntap_hostinet.cc new file mode 100644 index 000000000..0c527419e --- /dev/null +++ b/test/syscalls/linux/tuntap_hostinet.cc @@ -0,0 +1,37 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include +#include + +#include "gtest/gtest.h" +#include "test/util/test_util.h" + +namespace gvisor { +namespace testing { + +namespace { + +TEST(TuntapHostInetTest, NoNetTun) { + SKIP_IF(!IsRunningOnGvisor()); + + struct stat statbuf; + ASSERT_THAT(stat("/dev/net/tun", &statbuf), SyscallFailsWithErrno(ENOENT)); +} + +} // namespace +} // namespace testing + +} // namespace gvisor -- cgit v1.2.3 From 62bd3ca8a37543c060292ca86669e537b9ebf36d Mon Sep 17 00:00:00 2001 From: Michael Pratt Date: Mon, 2 Mar 2020 10:05:52 -0800 Subject: Take write lock when removing xattr PiperOrigin-RevId: 298380654 --- pkg/sentry/fs/fsutil/inode.go | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/fsutil/inode.go b/pkg/sentry/fs/fsutil/inode.go index daecc4ffe..1922ff08c 100644 --- a/pkg/sentry/fs/fsutil/inode.go +++ b/pkg/sentry/fs/fsutil/inode.go @@ -259,8 +259,8 @@ func (i *InodeSimpleExtendedAttributes) ListXattr(context.Context, *fs.Inode, ui // RemoveXattr implements fs.InodeOperations.RemoveXattr. func (i *InodeSimpleExtendedAttributes) RemoveXattr(_ context.Context, _ *fs.Inode, name string) error { - i.mu.RLock() - defer i.mu.RUnlock() + i.mu.Lock() + defer i.mu.Unlock() if _, ok := i.xattrs[name]; ok { delete(i.xattrs, name) return nil -- cgit v1.2.3 From da48fc6cca23a38faef51c5b5f8ae609940773a0 Mon Sep 17 00:00:00 2001 From: Ian Lewis Date: Thu, 5 Mar 2020 18:21:39 -0800 Subject: Stub oom_score_adj and oom_score. Adds an oom_score_adj and oom_score proc file stub. oom_score_adj accepts writes of values -1000 to 1000 and persists the value with the task. New tasks inherit the parent's oom_score_adj. oom_score is a read-only stub that always returns the value '0'. Issue #202 PiperOrigin-RevId: 299245355 --- pkg/sentry/fs/proc/task.go | 126 ++++++++++++++++++++++++++----- pkg/sentry/fsimpl/proc/task.go | 12 +-- pkg/sentry/fsimpl/proc/task_files.go | 43 +++++++++++ pkg/sentry/fsimpl/proc/tasks_test.go | 32 ++++---- pkg/sentry/kernel/task.go | 33 ++++++++ pkg/sentry/kernel/task_clone.go | 6 ++ pkg/sentry/kernel/task_start.go | 4 + test/syscalls/BUILD | 8 +- test/syscalls/linux/BUILD | 13 ++++ test/syscalls/linux/proc.cc | 21 ++++++ test/syscalls/linux/proc_pid_oomscore.cc | 72 ++++++++++++++++++ 11 files changed, 330 insertions(+), 40 deletions(-) create mode 100644 test/syscalls/linux/proc_pid_oomscore.cc (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index 8ab8d8a02..4e9b0fc00 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -72,24 +72,26 @@ var _ fs.InodeOperations = (*taskDir)(nil) // newTaskDir creates a new proc task entry. func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bool) *fs.Inode { contents := map[string]*fs.Inode{ - "auxv": newAuxvec(t, msrc), - "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), - "comm": newComm(t, msrc), - "environ": newExecArgInode(t, msrc, environExecArg), - "exe": newExe(t, msrc), - "fd": newFdDir(t, msrc), - "fdinfo": newFdInfoDir(t, msrc), - "gid_map": newGIDMap(t, msrc), - "io": newIO(t, msrc, isThreadGroup), - "maps": newMaps(t, msrc), - "mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), - "mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), - "ns": newNamespaceDir(t, msrc), - "smaps": newSmaps(t, msrc), - "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), - "statm": newStatm(t, msrc), - "status": newStatus(t, msrc, p.pidns), - "uid_map": newUIDMap(t, msrc), + "auxv": newAuxvec(t, msrc), + "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), + "comm": newComm(t, msrc), + "environ": newExecArgInode(t, msrc, environExecArg), + "exe": newExe(t, msrc), + "fd": newFdDir(t, msrc), + "fdinfo": newFdInfoDir(t, msrc), + "gid_map": newGIDMap(t, msrc), + "io": newIO(t, msrc, isThreadGroup), + "maps": newMaps(t, msrc), + "mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc), + "mounts": seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc), + "ns": newNamespaceDir(t, msrc), + "oom_score": newOOMScore(t, msrc), + "oom_score_adj": newOOMScoreAdj(t, msrc), + "smaps": newSmaps(t, msrc), + "stat": newTaskStat(t, msrc, isThreadGroup, p.pidns), + "statm": newStatm(t, msrc), + "status": newStatus(t, msrc, p.pidns), + "uid_map": newUIDMap(t, msrc), } if isThreadGroup { contents["task"] = p.newSubtasks(t, msrc) @@ -796,4 +798,92 @@ func (f *auxvecFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequenc return int64(n), err } +// newOOMScore returns a oom_score file. It is a stub that always returns 0. +// TODO(gvisor.dev/issue/1967) +func newOOMScore(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + return newStaticProcInode(t, msrc, []byte("0\n")) +} + +// oomScoreAdj is a file containing the oom_score adjustment for a task. +// +// +stateify savable +type oomScoreAdj struct { + fsutil.SimpleFileInode + + t *kernel.Task +} + +// +stateify savable +type oomScoreAdjFile struct { + fsutil.FileGenericSeek `state:"nosave"` + fsutil.FileNoIoctl `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNoopRelease `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + waiter.AlwaysReady `state:"nosave"` + + t *kernel.Task +} + +// newOOMScoreAdj returns a oom_score_adj file. +func newOOMScoreAdj(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + i := &oomScoreAdj{ + SimpleFileInode: *fsutil.NewSimpleFileInode(t, fs.RootOwner, fs.FilePermsFromMode(0644), linux.PROC_SUPER_MAGIC), + t: t, + } + return newProcInode(t, i, msrc, fs.SpecialFile, t) +} + +// Truncate implements fs.InodeOperations.Truncate. Truncate is called when +// O_TRUNC is specified for any kind of existing Dirent but is not called via +// (f)truncate for proc files. +func (*oomScoreAdj) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + +// GetFile implements fs.InodeOperations.GetFile. +func (o *oomScoreAdj) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + return fs.NewFile(ctx, dirent, flags, &oomScoreAdjFile{t: o.t}), nil +} + +// Read implements fs.FileOperations.Read. +func (f *oomScoreAdjFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + if offset != 0 { + return 0, io.EOF + } + adj, err := f.t.OOMScoreAdj() + if err != nil { + return 0, err + } + adjBytes := []byte(strconv.FormatInt(int64(adj), 10) + "\n") + n, err := dst.CopyOut(ctx, adjBytes) + return int64(n), err +} + +// Write implements fs.FileOperations.Write. +func (f *oomScoreAdjFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if err := f.t.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} + // LINT.ThenChange(../../fsimpl/proc/task.go|../../fsimpl/proc/task_files.go) diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index 2d814668a..18e5cd6f6 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -62,11 +62,13 @@ func newTaskInode(inoGen InoGenerator, task *kernel.Task, pidns *kernel.PIDNames "pid": newNamespaceSymlink(task, inoGen.NextIno(), "pid"), "user": newNamespaceSymlink(task, inoGen.NextIno(), "user"), }), - "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), - "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), - "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), - "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), - "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), + "oom_score": newTaskOwnedFile(task, inoGen.NextIno(), 0444, newStaticFile("0\n")), + "oom_score_adj": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &oomScoreAdj{task: task}), + "smaps": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &smapsData{task: task}), + "stat": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), + "statm": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statmData{task: task}), + "status": newTaskOwnedFile(task, inoGen.NextIno(), 0444, &statusData{task: task, pidns: pidns}), + "uid_map": newTaskOwnedFile(task, inoGen.NextIno(), 0644, &idMapData{task: task, gids: false}), } if isThreadGroup { contents["task"] = newSubtasks(task, pidns, inoGen, cgroupControllers) diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go index efd3b3453..5a231ac86 100644 --- a/pkg/sentry/fsimpl/proc/task_files.go +++ b/pkg/sentry/fsimpl/proc/task_files.go @@ -525,3 +525,46 @@ func (i *ioData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "cancelled_write_bytes: %d\n", io.BytesWriteCancelled) return nil } + +// oomScoreAdj is a stub of the /proc//oom_score_adj file. +// +// +stateify savable +type oomScoreAdj struct { + kernfs.DynamicBytesFile + + task *kernel.Task +} + +var _ vfs.WritableDynamicBytesSource = (*oomScoreAdj)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (o *oomScoreAdj) Generate(ctx context.Context, buf *bytes.Buffer) error { + adj, err := o.task.OOMScoreAdj() + if err != nil { + return err + } + fmt.Fprintf(buf, "%d\n", adj) + return nil +} + +// Write implements vfs.WritableDynamicBytesSource.Write. +func (o *oomScoreAdj) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + src = src.TakeFirst(usermem.PageSize - 1) + + var v int32 + n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts) + if err != nil { + return 0, err + } + + if err := o.task.SetOOMScoreAdj(v); err != nil { + return 0, err + } + + return n, nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index c5d531fe0..0eb401619 100644 --- a/pkg/sentry/fsimpl/proc/tasks_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_test.go @@ -63,21 +63,23 @@ var ( "thread-self": threadSelfLink.NextOff, } taskStaticFiles = map[string]testutil.DirentType{ - "auxv": linux.DT_REG, - "cgroup": linux.DT_REG, - "cmdline": linux.DT_REG, - "comm": linux.DT_REG, - "environ": linux.DT_REG, - "gid_map": linux.DT_REG, - "io": linux.DT_REG, - "maps": linux.DT_REG, - "ns": linux.DT_DIR, - "smaps": linux.DT_REG, - "stat": linux.DT_REG, - "statm": linux.DT_REG, - "status": linux.DT_REG, - "task": linux.DT_DIR, - "uid_map": linux.DT_REG, + "auxv": linux.DT_REG, + "cgroup": linux.DT_REG, + "cmdline": linux.DT_REG, + "comm": linux.DT_REG, + "environ": linux.DT_REG, + "gid_map": linux.DT_REG, + "io": linux.DT_REG, + "maps": linux.DT_REG, + "ns": linux.DT_DIR, + "oom_score": linux.DT_REG, + "oom_score_adj": linux.DT_REG, + "smaps": linux.DT_REG, + "stat": linux.DT_REG, + "statm": linux.DT_REG, + "status": linux.DT_REG, + "task": linux.DT_DIR, + "uid_map": linux.DT_REG, } ) diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index 2cee2e6ed..c0dbbe890 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -37,6 +37,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/usage" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -554,6 +555,13 @@ type Task struct { // // startTime is protected by mu. startTime ktime.Time + + // oomScoreAdj is the task's OOM score adjustment. This is currently not + // used but is maintained for consistency. + // TODO(gvisor.dev/issue/1967) + // + // oomScoreAdj is protected by mu, and is owned by the task goroutine. + oomScoreAdj int32 } func (t *Task) savePtraceTracer() *Task { @@ -847,3 +855,28 @@ func (t *Task) AbstractSockets() *AbstractSocketNamespace { func (t *Task) ContainerID() string { return t.containerID } + +// OOMScoreAdj gets the task's OOM score adjustment. +func (t *Task) OOMScoreAdj() (int32, error) { + t.mu.Lock() + defer t.mu.Unlock() + if t.ExitState() == TaskExitDead { + return 0, syserror.ESRCH + } + return t.oomScoreAdj, nil +} + +// SetOOMScoreAdj sets the task's OOM score adjustment. The value should be +// between -1000 and 1000 inclusive. +func (t *Task) SetOOMScoreAdj(adj int32) error { + t.mu.Lock() + defer t.mu.Unlock() + if t.ExitState() == TaskExitDead { + return syserror.ESRCH + } + if adj > 1000 || adj < -1000 { + return syserror.EINVAL + } + t.oomScoreAdj = adj + return nil +} diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 78866f280..dda502bb8 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -264,6 +264,11 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { rseqSignature = t.rseqSignature } + adj, err := t.OOMScoreAdj() + if err != nil { + return 0, nil, err + } + cfg := &TaskConfig{ Kernel: t.k, ThreadGroup: tg, @@ -282,6 +287,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { RSeqAddr: rseqAddr, RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), + OOMScoreAdj: adj, } if opts.NewThreadGroup { cfg.Parent = t diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index a5035bb7f..2bbf48bb8 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -93,6 +93,9 @@ type TaskConfig struct { // ContainerID is the container the new task belongs to. ContainerID string + + // oomScoreAdj is the task's OOM score adjustment. + OOMScoreAdj int32 } // NewTask creates a new task defined by cfg. @@ -143,6 +146,7 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { rseqSignature: cfg.RSeqSignature, futexWaiter: futex.NewWaiter(), containerID: cfg.ContainerID, + oomScoreAdj: cfg.OOMScoreAdj, } t.creds.Store(cfg.Credentials) t.endStopCond.L = &t.tg.signalHandlers.mu diff --git a/test/syscalls/BUILD b/test/syscalls/BUILD index a69b0ce13..9800a0cdf 100644 --- a/test/syscalls/BUILD +++ b/test/syscalls/BUILD @@ -318,10 +318,14 @@ syscall_test( test = "//test/syscalls/linux:proc_test", ) -syscall_test(test = "//test/syscalls/linux:proc_pid_uid_gid_map_test") - syscall_test(test = "//test/syscalls/linux:proc_net_test") +syscall_test(test = "//test/syscalls/linux:proc_pid_oomscore_test") + +syscall_test(test = "//test/syscalls/linux:proc_pid_smaps_test") + +syscall_test(test = "//test/syscalls/linux:proc_pid_uid_gid_map_test") + syscall_test( size = "medium", test = "//test/syscalls/linux:pselect_test", diff --git a/test/syscalls/linux/BUILD b/test/syscalls/linux/BUILD index 0fbd556de..43455f1a3 100644 --- a/test/syscalls/linux/BUILD +++ b/test/syscalls/linux/BUILD @@ -1631,6 +1631,19 @@ cc_binary( ], ) +cc_binary( + name = "proc_pid_oomscore_test", + testonly = 1, + srcs = ["proc_pid_oomscore.cc"], + linkstatic = 1, + deps = [ + "//test/util:fs_util", + "//test/util:test_main", + "//test/util:test_util", + "@com_google_absl//absl/strings", + ], +) + cc_binary( name = "proc_pid_smaps_test", testonly = 1, diff --git a/test/syscalls/linux/proc.cc b/test/syscalls/linux/proc.cc index f91187e75..5a70f6c3b 100644 --- a/test/syscalls/linux/proc.cc +++ b/test/syscalls/linux/proc.cc @@ -1431,6 +1431,12 @@ TEST(ProcPidFile, SubprocessRunning) { EXPECT_THAT(ReadWhileRunning("uid_map", buf, sizeof(buf)), SyscallSucceedsWithValue(sizeof(buf))); + + EXPECT_THAT(ReadWhileRunning("oom_score", buf, sizeof(buf)), + SyscallSucceedsWithValue(sizeof(buf))); + + EXPECT_THAT(ReadWhileRunning("oom_score_adj", buf, sizeof(buf)), + SyscallSucceedsWithValue(sizeof(buf))); } // Test whether /proc/PID/ files can be read for a zombie process. @@ -1466,6 +1472,12 @@ TEST(ProcPidFile, SubprocessZombie) { EXPECT_THAT(ReadWhileZombied("uid_map", buf, sizeof(buf)), SyscallSucceedsWithValue(sizeof(buf))); + EXPECT_THAT(ReadWhileZombied("oom_score", buf, sizeof(buf)), + SyscallSucceedsWithValue(sizeof(buf))); + + EXPECT_THAT(ReadWhileZombied("oom_score_adj", buf, sizeof(buf)), + SyscallSucceedsWithValue(sizeof(buf))); + // FIXME(gvisor.dev/issue/164): Inconsistent behavior between gVisor and linux // on proc files. // @@ -1527,6 +1539,15 @@ TEST(ProcPidFile, SubprocessExited) { EXPECT_THAT(ReadWhileExited("uid_map", buf, sizeof(buf)), SyscallSucceedsWithValue(sizeof(buf))); + + if (!IsRunningOnGvisor()) { + // FIXME(gvisor.dev/issue/164): Succeeds on gVisor. + EXPECT_THAT(ReadWhileExited("oom_score", buf, sizeof(buf)), + SyscallFailsWithErrno(ESRCH)); + } + + EXPECT_THAT(ReadWhileExited("oom_score_adj", buf, sizeof(buf)), + SyscallFailsWithErrno(ESRCH)); } PosixError DirContainsImpl(absl::string_view path, diff --git a/test/syscalls/linux/proc_pid_oomscore.cc b/test/syscalls/linux/proc_pid_oomscore.cc new file mode 100644 index 000000000..707821a3f --- /dev/null +++ b/test/syscalls/linux/proc_pid_oomscore.cc @@ -0,0 +1,72 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include + +#include +#include +#include + +#include "test/util/fs_util.h" +#include "test/util/test_util.h" + +namespace gvisor { +namespace testing { + +namespace { + +PosixErrorOr ReadProcNumber(std::string path) { + ASSIGN_OR_RETURN_ERRNO(std::string contents, GetContents(path)); + EXPECT_EQ(contents[contents.length() - 1], '\n'); + + int num; + if (!absl::SimpleAtoi(contents, &num)) { + return PosixError(EINVAL, "invalid value: " + contents); + } + + return num; +} + +TEST(ProcPidOomscoreTest, BasicRead) { + auto const oom_score = + ASSERT_NO_ERRNO_AND_VALUE(ReadProcNumber("/proc/self/oom_score")); + EXPECT_LE(oom_score, 1000); + EXPECT_GE(oom_score, -1000); +} + +TEST(ProcPidOomscoreAdjTest, BasicRead) { + auto const oom_score = + ASSERT_NO_ERRNO_AND_VALUE(ReadProcNumber("/proc/self/oom_score_adj")); + + // oom_score_adj defaults to 0. + EXPECT_EQ(oom_score, 0); +} + +TEST(ProcPidOomscoreAdjTest, BasicWrite) { + constexpr int test_value = 7; + FileDescriptor fd = + ASSERT_NO_ERRNO_AND_VALUE(Open("/proc/self/oom_score_adj", O_WRONLY)); + ASSERT_THAT( + RetryEINTR(write)(fd.get(), std::to_string(test_value).c_str(), 1), + SyscallSucceeds()); + + auto const oom_score = + ASSERT_NO_ERRNO_AND_VALUE(ReadProcNumber("/proc/self/oom_score_adj")); + EXPECT_EQ(oom_score, test_value); +} + +} // namespace + +} // namespace testing +} // namespace gvisor -- cgit v1.2.3 From 6fa5cee82c0f515b001dee5f3840e1f875b2f477 Mon Sep 17 00:00:00 2001 From: Tamir Duberstein Date: Fri, 6 Mar 2020 12:30:37 -0800 Subject: Prevent memory leaks in ilist When list elements are removed from a list but not discarded, it becomes important to invalidate the references they hold to their former neighbors to prevent memory leaks. PiperOrigin-RevId: 299412421 --- pkg/ilist/list.go | 8 ++++++-- pkg/sentry/fs/dirent_cache.go | 2 -- pkg/sentry/fs/inotify.go | 5 ++++- pkg/sentry/kernel/epoll/epoll_state.go | 13 ++++++++----- pkg/tcpip/network/fragmentation/fragmentation.go | 8 +++++--- 5 files changed, 23 insertions(+), 13 deletions(-) (limited to 'pkg/sentry/fs') diff --git a/pkg/ilist/list.go b/pkg/ilist/list.go index f3a609b57..8f93e4d6d 100644 --- a/pkg/ilist/list.go +++ b/pkg/ilist/list.go @@ -169,8 +169,9 @@ func (l *List) InsertBefore(a, e Element) { // Remove removes e from l. func (l *List) Remove(e Element) { - prev := ElementMapper{}.linkerFor(e).Prev() - next := ElementMapper{}.linkerFor(e).Next() + linker := ElementMapper{}.linkerFor(e) + prev := linker.Prev() + next := linker.Next() if prev != nil { ElementMapper{}.linkerFor(prev).SetNext(next) @@ -183,6 +184,9 @@ func (l *List) Remove(e Element) { } else { l.tail = prev } + + linker.SetNext(nil) + linker.SetPrev(nil) } // Entry is a default implementation of Linker. Users can add anonymous fields diff --git a/pkg/sentry/fs/dirent_cache.go b/pkg/sentry/fs/dirent_cache.go index 25514ace4..33de32c69 100644 --- a/pkg/sentry/fs/dirent_cache.go +++ b/pkg/sentry/fs/dirent_cache.go @@ -101,8 +101,6 @@ func (c *DirentCache) remove(d *Dirent) { panic(fmt.Sprintf("trying to remove %v, which is not in the dirent cache", d)) } c.list.Remove(d) - d.SetPrev(nil) - d.SetNext(nil) d.DecRef() c.currentSize-- if c.limit != nil { diff --git a/pkg/sentry/fs/inotify.go b/pkg/sentry/fs/inotify.go index 928c90aa0..e3a715c1f 100644 --- a/pkg/sentry/fs/inotify.go +++ b/pkg/sentry/fs/inotify.go @@ -143,7 +143,10 @@ func (i *Inotify) Read(ctx context.Context, _ *File, dst usermem.IOSequence, _ i } var writeLen int64 - for event := i.events.Front(); event != nil; event = event.Next() { + for it := i.events.Front(); it != nil; { + event := it + it = it.Next() + // Does the buffer have enough remaining space to hold the event we're // about to write out? if dst.NumBytes() < int64(event.sizeOf()) { diff --git a/pkg/sentry/kernel/epoll/epoll_state.go b/pkg/sentry/kernel/epoll/epoll_state.go index a0d35d350..8e9f200d0 100644 --- a/pkg/sentry/kernel/epoll/epoll_state.go +++ b/pkg/sentry/kernel/epoll/epoll_state.go @@ -38,11 +38,14 @@ func (e *EventPoll) afterLoad() { } } - for it := e.waitingList.Front(); it != nil; it = it.Next() { - if it.id.File.Readiness(it.mask) != 0 { - e.waitingList.Remove(it) - e.readyList.PushBack(it) - it.curList = &e.readyList + for it := e.waitingList.Front(); it != nil; { + entry := it + it = it.Next() + + if entry.id.File.Readiness(entry.mask) != 0 { + e.waitingList.Remove(entry) + e.readyList.PushBack(entry) + entry.curList = &e.readyList e.Notify(waiter.EventIn) } } diff --git a/pkg/tcpip/network/fragmentation/fragmentation.go b/pkg/tcpip/network/fragmentation/fragmentation.go index 92f2aa13a..f42abc4bb 100644 --- a/pkg/tcpip/network/fragmentation/fragmentation.go +++ b/pkg/tcpip/network/fragmentation/fragmentation.go @@ -115,10 +115,12 @@ func (f *Fragmentation) Process(id uint32, first, last uint16, more bool, vv buf // Evict reassemblers if we are consuming more memory than highLimit until // we reach lowLimit. if f.size > f.highLimit { - tail := f.rList.Back() - for f.size > f.lowLimit && tail != nil { + for f.size > f.lowLimit { + tail := f.rList.Back() + if tail == nil { + break + } f.release(tail) - tail = tail.Prev() } } f.mu.Unlock() -- cgit v1.2.3 From 960f6a975b7e44c0efe8fd38c66b02017c4fe137 Mon Sep 17 00:00:00 2001 From: Dean Deng Date: Fri, 6 Mar 2020 12:58:45 -0800 Subject: Add plumbing for importing fds in VFS2, along with non-socket, non-TTY impl. In VFS2, imported file descriptors are stored in a kernfs-based filesystem. Upon calling ImportFD, the host fd can be accessed in two ways: 1. a FileDescription that can be added to the FDTable, and 2. a Dentry in the host.filesystem mount, which we will want to access through magic symlinks in /proc/[pid]/fd/. An implementation of the kernfs.Inode interface stores a unique host fd. This inode can be inserted into file descriptions as well as dentries. This change also plumbs in three FileDescriptionImpls corresponding to fds for sockets, TTYs, and other files (only the latter is implemented here). These implementations will mostly make corresponding syscalls to the host. Where possible, the logic is ported over from pkg/sentry/fs/host. Updates #1672 PiperOrigin-RevId: 299417263 --- pkg/sentry/fs/host/util.go | 12 +- pkg/sentry/fsimpl/host/BUILD | 27 +++ pkg/sentry/fsimpl/host/default_file.go | 233 ++++++++++++++++++++ pkg/sentry/fsimpl/host/host.go | 286 +++++++++++++++++++++++++ pkg/sentry/fsimpl/host/util.go | 86 ++++++++ pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go | 2 +- pkg/sentry/fsimpl/kernfs/fd_impl_util.go | 18 +- pkg/sentry/fsimpl/kernfs/filesystem.go | 6 +- pkg/sentry/fsimpl/kernfs/inode_impl_util.go | 54 ++--- pkg/sentry/fsimpl/kernfs/kernfs.go | 7 +- pkg/sentry/fsimpl/proc/subtasks.go | 13 +- pkg/sentry/fsimpl/proc/task.go | 23 +- pkg/sentry/fsimpl/proc/tasks.go | 19 +- pkg/sentry/vfs/BUILD | 4 + pkg/sentry/vfs/file_description_impl_util.go | 4 +- 15 files changed, 731 insertions(+), 63 deletions(-) create mode 100644 pkg/sentry/fsimpl/host/BUILD create mode 100644 pkg/sentry/fsimpl/host/default_file.go create mode 100644 pkg/sentry/fsimpl/host/host.go create mode 100644 pkg/sentry/fsimpl/host/util.go (limited to 'pkg/sentry/fs') diff --git a/pkg/sentry/fs/host/util.go b/pkg/sentry/fs/host/util.go index e37e687c6..7c60dc1db 100644 --- a/pkg/sentry/fs/host/util.go +++ b/pkg/sentry/fs/host/util.go @@ -24,7 +24,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/syserror" ) @@ -152,9 +152,9 @@ func unstableAttr(mo *superOperations, s *syscall.Stat_t) fs.UnstableAttr { Usage: s.Blocks * 512, Perms: fs.FilePermsFromMode(linux.FileMode(s.Mode)), Owner: owner(mo, s), - AccessTime: ktime.FromUnix(s.Atim.Sec, s.Atim.Nsec), - ModificationTime: ktime.FromUnix(s.Mtim.Sec, s.Mtim.Nsec), - StatusChangeTime: ktime.FromUnix(s.Ctim.Sec, s.Ctim.Nsec), + AccessTime: time.FromUnix(s.Atim.Sec, s.Atim.Nsec), + ModificationTime: time.FromUnix(s.Mtim.Sec, s.Mtim.Nsec), + StatusChangeTime: time.FromUnix(s.Ctim.Sec, s.Ctim.Nsec), Links: uint64(s.Nlink), } } @@ -165,6 +165,8 @@ type dirInfo struct { bufp int // location of next record in buf. } +// LINT.IfChange + // isBlockError unwraps os errors and checks if they are caused by EAGAIN or // EWOULDBLOCK. This is so they can be transformed into syserror.ErrWouldBlock. func isBlockError(err error) bool { @@ -177,6 +179,8 @@ func isBlockError(err error) bool { return false } +// LINT.ThenChange(../../fsimpl/host/util.go) + func hostEffectiveKIDs() (uint32, []uint32, error) { gids, err := os.Getgroups() if err != nil { diff --git a/pkg/sentry/fsimpl/host/BUILD b/pkg/sentry/fsimpl/host/BUILD new file mode 100644 index 000000000..731f192b3 --- /dev/null +++ b/pkg/sentry/fsimpl/host/BUILD @@ -0,0 +1,27 @@ +load("//tools:defs.bzl", "go_library") + +licenses(["notice"]) + +go_library( + name = "host", + srcs = [ + "default_file.go", + "host.go", + "util.go", + ], + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/log", + "//pkg/refs", + "//pkg/safemem", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", + "//pkg/sentry/vfs", + "//pkg/sync", + "//pkg/syserror", + "//pkg/usermem", + "@org_golang_x_sys//unix:go_default_library", + ], +) diff --git a/pkg/sentry/fsimpl/host/default_file.go b/pkg/sentry/fsimpl/host/default_file.go new file mode 100644 index 000000000..172cdb161 --- /dev/null +++ b/pkg/sentry/fsimpl/host/default_file.go @@ -0,0 +1,233 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "math" + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// defaultFileFD implements FileDescriptionImpl for non-socket, non-TTY files. +type defaultFileFD struct { + fileDescription + + // canMap specifies whether we allow the file to be memory mapped. + canMap bool + + // mu protects the fields below. + mu sync.Mutex + + // offset specifies the current file offset. + offset int64 +} + +// TODO(gvisor.dev/issue/1672): Implement Waitable interface. + +// PRead implements FileDescriptionImpl. +func (f *defaultFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + return readFromHostFD(ctx, f.inode.hostFD, dst, offset, int(opts.Flags)) +} + +// Read implements FileDescriptionImpl. +func (f *defaultFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + // These files can't be memory mapped, assert this. + if f.canMap { + panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") + } + + f.mu.Lock() + n, err := readFromHostFD(ctx, f.inode.hostFD, dst, -1, int(opts.Flags)) + f.mu.Unlock() + if isBlockError(err) { + // If we got any data at all, return it as a "completed" partial read + // rather than retrying until complete. + if n != 0 { + err = nil + } else { + err = syserror.ErrWouldBlock + } + } + return n, err + } + // TODO(gvisor.dev/issue/1672): Cache pages, when forced to do so. + f.mu.Lock() + n, err := readFromHostFD(ctx, f.inode.hostFD, dst, f.offset, int(opts.Flags)) + f.offset += n + f.mu.Unlock() + return n, err +} + +func readFromHostFD(ctx context.Context, fd int, dst usermem.IOSequence, offset int64, flags int) (int64, error) { + if flags&^(linux.RWF_VALID) != 0 { + return 0, syserror.EOPNOTSUPP + } + + reader := safemem.FromVecReaderFunc{ + func(srcs [][]byte) (int64, error) { + n, err := unix.Preadv2(fd, srcs, offset, flags) + return int64(n), err + }, + } + n, err := dst.CopyOutFrom(ctx, reader) + return int64(n), err +} + +// PWrite implements FileDescriptionImpl. +func (f *defaultFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + return writeToHostFD(ctx, f.inode.hostFD, src, offset, int(opts.Flags)) +} + +// Write implements FileDescriptionImpl. +func (f *defaultFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // TODO(b/34716638): Some char devices do support offsets, e.g. /dev/null. + if f.inode.isStream { + // These files can't be memory mapped, assert this. + if f.canMap { + panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") + } + + f.mu.Lock() + n, err := writeToHostFD(ctx, f.inode.hostFD, src, -1, int(opts.Flags)) + f.mu.Unlock() + if isBlockError(err) { + err = syserror.ErrWouldBlock + } + return n, err + } + // TODO(gvisor.dev/issue/1672): Cache pages, when forced to do so. + // TODO(gvisor.dev/issue/1672): Write to end of file and update offset if O_APPEND is set on this file. + f.mu.Lock() + n, err := writeToHostFD(ctx, f.inode.hostFD, src, f.offset, int(opts.Flags)) + f.offset += n + f.mu.Unlock() + return n, err +} + +func writeToHostFD(ctx context.Context, fd int, src usermem.IOSequence, offset int64, flags int) (int64, error) { + if flags&^(linux.RWF_VALID) != 0 { + return 0, syserror.EOPNOTSUPP + } + + writer := safemem.FromVecWriterFunc{ + func(srcs [][]byte) (int64, error) { + n, err := unix.Pwritev2(fd, srcs, offset, flags) + return int64(n), err + }, + } + n, err := src.CopyInTo(ctx, writer) + return int64(n), err +} + +// Seek implements FileDescriptionImpl. +// +// Note that we do not support seeking on directories, since we do not even +// allow directory fds to be imported at all. +func (f *defaultFileFD) Seek(_ context.Context, offset int64, whence int32) (int64, error) { + // TODO(b/34716638): Some char devices do support seeking, e.g. /dev/null. + if f.inode.isStream { + return 0, syserror.ESPIPE + } + + f.mu.Lock() + defer f.mu.Unlock() + + switch whence { + case linux.SEEK_SET: + if offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset = offset + + case linux.SEEK_CUR: + // Check for overflow. Note that underflow cannot occur, since f.offset >= 0. + if offset > math.MaxInt64-f.offset { + return f.offset, syserror.EOVERFLOW + } + if f.offset+offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset += offset + + case linux.SEEK_END: + var s syscall.Stat_t + if err := syscall.Fstat(f.inode.hostFD, &s); err != nil { + return f.offset, err + } + size := s.Size + + // Check for overflow. Note that underflow cannot occur, since size >= 0. + if offset > math.MaxInt64-size { + return f.offset, syserror.EOVERFLOW + } + if size+offset < 0 { + return f.offset, syserror.EINVAL + } + f.offset = size + offset + + case linux.SEEK_DATA, linux.SEEK_HOLE: + // Modifying the offset in the host file table should not matter, since + // this is the only place where we use it. + // + // For reading and writing, we always rely on our internal offset. + n, err := unix.Seek(f.inode.hostFD, offset, int(whence)) + if err != nil { + return f.offset, err + } + f.offset = n + + default: + // Invalid whence. + return f.offset, syserror.EINVAL + } + + return f.offset, nil +} + +// Sync implements FileDescriptionImpl. +func (f *defaultFileFD) Sync(context.Context) error { + // TODO(gvisor.dev/issue/1672): Currently we do not support the SyncData optimization, so we always sync everything. + return unix.Fsync(f.inode.hostFD) +} + +// ConfigureMMap implements FileDescriptionImpl. +func (f *defaultFileFD) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error { + if !f.canMap { + return syserror.ENODEV + } + // TODO(gvisor.dev/issue/1672): Implement ConfigureMMap and Mappable interface. + return syserror.ENODEV +} diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go new file mode 100644 index 000000000..c205e6a0b --- /dev/null +++ b/pkg/sentry/fsimpl/host/host.go @@ -0,0 +1,286 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +// Package host provides a filesystem implementation for host files imported as +// file descriptors. +package host + +import ( + "errors" + "fmt" + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" +) + +// filesystem implements vfs.FilesystemImpl. +type filesystem struct { + kernfs.Filesystem +} + +// ImportFD sets up and returns a vfs.FileDescription from a donated fd. +func ImportFD(mnt *vfs.Mount, hostFD int, ownerUID auth.KUID, ownerGID auth.KGID, isTTY bool) (*vfs.FileDescription, error) { + // Must be importing to a mount of host.filesystem. + fs, ok := mnt.Filesystem().Impl().(*filesystem) + if !ok { + return nil, fmt.Errorf("can't import host FDs into filesystems of type %T", mnt.Filesystem().Impl()) + } + + // Retrieve metadata. + var s syscall.Stat_t + if err := syscall.Fstat(hostFD, &s); err != nil { + return nil, err + } + + fileMode := linux.FileMode(s.Mode) + fileType := fileMode.FileType() + // Pipes, character devices, and sockets can return EWOULDBLOCK for + // operations that would block. + isStream := fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK + + i := &inode{ + hostFD: hostFD, + isStream: isStream, + isTTY: isTTY, + ino: fs.NextIno(), + mode: fileMode, + uid: ownerUID, + gid: ownerGID, + } + + d := &kernfs.Dentry{} + d.Init(i) + // i.open will take a reference on d. + defer d.DecRef() + + return i.open(d.VFSDentry(), mnt) +} + +// inode implements kernfs.Inode. +type inode struct { + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + // When the reference count reaches zero, the host fd is closed. + refs.AtomicRefCount + + // hostFD contains the host fd that this file was originally created from, + // which must be available at time of restore. + // + // This field is initialized at creation time and is immutable. + hostFD int + + // isStream is true if the host fd points to a file representing a stream, + // e.g. a socket or a pipe. Such files are not seekable and can return + // EWOULDBLOCK for I/O operations. + // + // This field is initialized at creation time and is immutable. + isStream bool + + // isTTY is true if this file represents a TTY. + // + // This field is initialized at creation time and is immutable. + isTTY bool + + // ino is an inode number unique within this filesystem. + ino uint64 + + // mu protects the inode metadata below. + mu sync.Mutex + + // mode is the file mode of this inode. Note that this value may become out + // of date if the mode is changed on the host, e.g. with chmod. + mode linux.FileMode + + // uid and gid of the file owner. Note that these refer to the owner of the + // file created on import, not the fd on the host. + uid auth.KUID + gid auth.KGID +} + +// Note that these flags may become out of date, since they can be modified +// on the host, e.g. with fcntl. +func fileFlagsFromHostFD(fd int) (int, error) { + flags, err := unix.FcntlInt(uintptr(fd), syscall.F_GETFL, 0) + if err != nil { + log.Warningf("Failed to get file flags for donated FD %d: %v", fd, err) + return 0, err + } + // TODO(gvisor.dev/issue/1672): implement behavior corresponding to these allowed flags. + flags &= syscall.O_ACCMODE | syscall.O_DIRECT | syscall.O_NONBLOCK | syscall.O_DSYNC | syscall.O_SYNC | syscall.O_APPEND + return flags, nil +} + +// CheckPermissions implements kernfs.Inode. +func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, atx vfs.AccessTypes) error { + return vfs.GenericCheckPermissions(creds, atx, false /* isDir */, uint16(i.mode), i.uid, i.gid) +} + +// Mode implements kernfs.Inode. +func (i *inode) Mode() linux.FileMode { + return i.mode +} + +// Stat implements kernfs.Inode. +func (i *inode) Stat(_ *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + var s unix.Statx_t + if err := unix.Statx(i.hostFD, "", int(unix.AT_EMPTY_PATH|opts.Sync), int(opts.Mask), &s); err != nil { + return linux.Statx{}, err + } + ls := unixToLinuxStatx(s) + + // Use our own internal inode number and file owner. + // + // TODO(gvisor.dev/issue/1672): Use a kernfs-specific device number as well. + // If we use the device number from the host, it may collide with another + // sentry-internal device number. We handle device/inode numbers without + // relying on the host to prevent collisions. + ls.Ino = i.ino + ls.UID = uint32(i.uid) + ls.GID = uint32(i.gid) + + // Update file mode from the host. + i.mode = linux.FileMode(ls.Mode) + + return ls, nil +} + +// SetStat implements kernfs.Inode. +func (i *inode) SetStat(_ *vfs.Filesystem, opts vfs.SetStatOptions) error { + s := opts.Stat + + m := s.Mask + if m == 0 { + return nil + } + if m&(linux.STATX_UID|linux.STATX_GID) != 0 { + return syserror.EPERM + } + if m&linux.STATX_MODE != 0 { + if err := syscall.Fchmod(i.hostFD, uint32(s.Mode)); err != nil { + return err + } + i.mode = linux.FileMode(s.Mode) + } + if m&linux.STATX_SIZE != 0 { + if err := syscall.Ftruncate(i.hostFD, int64(s.Size)); err != nil { + return err + } + } + if m&(linux.STATX_ATIME|linux.STATX_MTIME) != 0 { + timestamps := []unix.Timespec{ + toTimespec(s.Atime, m&linux.STATX_ATIME == 0), + toTimespec(s.Mtime, m&linux.STATX_MTIME == 0), + } + if err := unix.UtimesNanoAt(i.hostFD, "", timestamps, unix.AT_EMPTY_PATH); err != nil { + return err + } + } + return nil +} + +// DecRef implements kernfs.Inode. +func (i *inode) DecRef() { + i.AtomicRefCount.DecRefWithDestructor(i.Destroy) +} + +// Destroy implements kernfs.Inode. +func (i *inode) Destroy() { + if err := unix.Close(i.hostFD); err != nil { + log.Warningf("failed to close host fd %d: %v", i.hostFD, err) + } +} + +// Open implements kernfs.Inode. +func (i *inode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.open(vfsd, rp.Mount()) +} + +func (i *inode) open(d *vfs.Dentry, mnt *vfs.Mount) (*vfs.FileDescription, error) { + + fileType := i.mode.FileType() + if fileType == syscall.S_IFSOCK { + if i.isTTY { + return nil, errors.New("cannot use host socket as TTY") + } + // TODO(gvisor.dev/issue/1672): support importing sockets. + return nil, errors.New("importing host sockets not supported") + } + + if i.isTTY { + // TODO(gvisor.dev/issue/1672): support importing host fd as TTY. + return nil, errors.New("importing host fd as TTY not supported") + } + + // For simplicity, set offset to 0. Technically, we should + // only set to 0 on files that are not seekable (sockets, pipes, etc.), + // and use the offset from the host fd otherwise. + fd := &defaultFileFD{ + fileDescription: fileDescription{ + inode: i, + }, + canMap: canMap(uint32(fileType)), + mu: sync.Mutex{}, + offset: 0, + } + + vfsfd := &fd.vfsfd + flags, err := fileFlagsFromHostFD(i.hostFD) + if err != nil { + return nil, err + } + + if err := vfsfd.Init(fd, uint32(flags), mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return vfsfd, nil +} + +// fileDescription is embedded by host fd implementations of FileDescriptionImpl. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + + // inode is vfsfd.Dentry().Impl().(*kernfs.Dentry).Inode().(*inode), but + // cached to reduce indirections and casting. fileDescription does not hold + // a reference on the inode through the inode field (since one is already + // held via the Dentry). + // + // inode is immutable after fileDescription creation. + inode *inode +} + +// SetStat implements vfs.FileDescriptionImpl. +func (f *fileDescription) SetStat(_ context.Context, opts vfs.SetStatOptions) error { + return f.inode.SetStat(nil, opts) +} + +// Stat implements vfs.FileDescriptionImpl. +func (f *fileDescription) Stat(_ context.Context, opts vfs.StatOptions) (linux.Statx, error) { + return f.inode.Stat(nil, opts) +} + +// Release implements vfs.FileDescriptionImpl. +func (f *fileDescription) Release() { + // noop +} diff --git a/pkg/sentry/fsimpl/host/util.go b/pkg/sentry/fsimpl/host/util.go new file mode 100644 index 000000000..e1ccacb4d --- /dev/null +++ b/pkg/sentry/fsimpl/host/util.go @@ -0,0 +1,86 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package host + +import ( + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/syserror" +) + +func toTimespec(ts linux.StatxTimestamp, omit bool) unix.Timespec { + if omit { + return unix.Timespec{ + Sec: 0, + Nsec: unix.UTIME_OMIT, + } + } + return unix.Timespec{ + Sec: int64(ts.Sec), + Nsec: int64(ts.Nsec), + } +} + +func unixToLinuxStatx(s unix.Statx_t) linux.Statx { + return linux.Statx{ + Mask: s.Mask, + Blksize: s.Blksize, + Attributes: s.Attributes, + Nlink: s.Nlink, + UID: s.Uid, + GID: s.Gid, + Mode: s.Mode, + Ino: s.Ino, + Size: s.Size, + Blocks: s.Blocks, + AttributesMask: s.Attributes_mask, + Atime: unixToLinuxStatxTimestamp(s.Atime), + Btime: unixToLinuxStatxTimestamp(s.Btime), + Ctime: unixToLinuxStatxTimestamp(s.Ctime), + Mtime: unixToLinuxStatxTimestamp(s.Mtime), + RdevMajor: s.Rdev_major, + RdevMinor: s.Rdev_minor, + DevMajor: s.Dev_major, + DevMinor: s.Dev_minor, + } +} + +func unixToLinuxStatxTimestamp(ts unix.StatxTimestamp) linux.StatxTimestamp { + return linux.StatxTimestamp{Sec: ts.Sec, Nsec: ts.Nsec} +} + +// wouldBlock returns true for file types that can return EWOULDBLOCK +// for blocking operations, e.g. pipes, character devices, and sockets. +func wouldBlock(fileType uint32) bool { + return fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK +} + +// canMap returns true if a file with fileType is allowed to be memory mapped. +// This is ported over from VFS1, but it's probably not the best way for us +// to check if a file can be memory mapped. +func canMap(fileType uint32) bool { + // TODO(gvisor.dev/issue/1672): Also allow "special files" to be mapped (see fs/host:canMap()). + // + // TODO(b/38213152): Some obscure character devices can be mapped. + return fileType == syscall.S_IFREG +} + +// isBlockError checks if an error is EAGAIN or EWOULDBLOCK. +// If so, they can be transformed into syserror.ErrWouldBlock. +func isBlockError(err error) bool { + return err == syserror.EAGAIN || err == syserror.EWOULDBLOCK +} diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go index d092ccb2a..1c026f4d8 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -122,7 +122,7 @@ func (fd *DynamicBytesFD) Release() {} // Stat implements vfs.FileDescriptionImpl.Stat. func (fd *DynamicBytesFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { fs := fd.vfsfd.VirtualDentry().Mount().Filesystem() - return fd.inode.Stat(fs), nil + return fd.inode.Stat(fs, opts) } // SetStat implements vfs.FileDescriptionImpl.SetStat. diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go index 5650512e0..da821d524 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -107,9 +107,13 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent fs.mu.Lock() defer fs.mu.Unlock() + opts := vfs.StatOptions{Mask: linux.STATX_INO} // Handle ".". if fd.off == 0 { - stat := fd.inode().Stat(vfsFS) + stat, err := fd.inode().Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: ".", Type: linux.DT_DIR, @@ -125,7 +129,10 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // Handle "..". if fd.off == 1 { parentInode := vfsd.ParentOrSelf().Impl().(*Dentry).inode - stat := parentInode.Stat(vfsFS) + stat, err := parentInode.Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: "..", Type: linux.FileMode(stat.Mode).DirentType(), @@ -146,7 +153,10 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent childIdx := fd.off - 2 for it := fd.children.nthLocked(childIdx); it != nil; it = it.Next() { inode := it.Dentry.Impl().(*Dentry).inode - stat := inode.Stat(vfsFS) + stat, err := inode.Stat(vfsFS, opts) + if err != nil { + return err + } dirent := vfs.Dirent{ Name: it.Name, Type: linux.FileMode(stat.Mode).DirentType(), @@ -190,7 +200,7 @@ func (fd *GenericDirectoryFD) Seek(ctx context.Context, offset int64, whence int func (fd *GenericDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { fs := fd.filesystem() inode := fd.inode() - return inode.Stat(fs), nil + return inode.Stat(fs, opts) } // SetStat implements vfs.FileDescriptionImpl.SetStat. diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index 292f58afd..1d7e04ad4 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -12,10 +12,10 @@ // See the License for the specific language governing permissions and // limitations under the License. -// This file implements vfs.FilesystemImpl for kernfs. - package kernfs +// This file implements vfs.FilesystemImpl for kernfs. + import ( "fmt" @@ -634,7 +634,7 @@ func (fs *Filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err != nil { return linux.Statx{}, err } - return inode.Stat(fs.VFSFilesystem()), nil + return inode.Stat(fs.VFSFilesystem(), opts) } // StatFSAt implements vfs.FilesystemImpl.StatFSAt. diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index 099d70a16..d50018b18 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -36,20 +36,20 @@ type InodeNoopRefCount struct { } // IncRef implements Inode.IncRef. -func (n *InodeNoopRefCount) IncRef() { +func (InodeNoopRefCount) IncRef() { } // DecRef implements Inode.DecRef. -func (n *InodeNoopRefCount) DecRef() { +func (InodeNoopRefCount) DecRef() { } // TryIncRef implements Inode.TryIncRef. -func (n *InodeNoopRefCount) TryIncRef() bool { +func (InodeNoopRefCount) TryIncRef() bool { return true } // Destroy implements Inode.Destroy. -func (n *InodeNoopRefCount) Destroy() { +func (InodeNoopRefCount) Destroy() { } // InodeDirectoryNoNewChildren partially implements the Inode interface. @@ -58,27 +58,27 @@ func (n *InodeNoopRefCount) Destroy() { type InodeDirectoryNoNewChildren struct{} // NewFile implements Inode.NewFile. -func (*InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (*InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewLink implements Inode.NewLink. -func (*InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (*InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (*InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { return nil, syserror.EPERM } @@ -90,62 +90,62 @@ type InodeNotDirectory struct { } // HasChildren implements Inode.HasChildren. -func (*InodeNotDirectory) HasChildren() bool { +func (InodeNotDirectory) HasChildren() bool { return false } // NewFile implements Inode.NewFile. -func (*InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { panic("NewFile called on non-directory inode") } // NewDir implements Inode.NewDir. -func (*InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { panic("NewDir called on non-directory inode") } // NewLink implements Inode.NewLinkink. -func (*InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { panic("NewLink called on non-directory inode") } // NewSymlink implements Inode.NewSymlink. -func (*InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { panic("NewSymlink called on non-directory inode") } // NewNode implements Inode.NewNode. -func (*InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { panic("NewNode called on non-directory inode") } // Unlink implements Inode.Unlink. -func (*InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { panic("Unlink called on non-directory inode") } // RmDir implements Inode.RmDir. -func (*InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { panic("RmDir called on non-directory inode") } // Rename implements Inode.Rename. -func (*InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { +func (InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { panic("Rename called on non-directory inode") } // Lookup implements Inode.Lookup. -func (*InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { panic("Lookup called on non-directory inode") } // IterDirents implements Inode.IterDirents. -func (*InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { +func (InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { panic("IterDirents called on non-directory inode") } // Valid implements Inode.Valid. -func (*InodeNotDirectory) Valid(context.Context) bool { +func (InodeNotDirectory) Valid(context.Context) bool { return true } @@ -157,17 +157,17 @@ func (*InodeNotDirectory) Valid(context.Context) bool { type InodeNoDynamicLookup struct{} // Lookup implements Inode.Lookup. -func (*InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { return nil, syserror.ENOENT } // IterDirents implements Inode.IterDirents. -func (*InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { +func (InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { return offset, nil } // Valid implements Inode.Valid. -func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { +func (InodeNoDynamicLookup) Valid(ctx context.Context) bool { return true } @@ -177,7 +177,7 @@ func (*InodeNoDynamicLookup) Valid(ctx context.Context) bool { type InodeNotSymlink struct{} // Readlink implements Inode.Readlink. -func (*InodeNotSymlink) Readlink(context.Context) (string, error) { +func (InodeNotSymlink) Readlink(context.Context) (string, error) { return "", syserror.EINVAL } @@ -219,7 +219,7 @@ func (a *InodeAttrs) Mode() linux.FileMode { // Stat partially implements Inode.Stat. Note that this function doesn't provide // all the stat fields, and the embedder should consider extending the result // with filesystem-specific fields. -func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { +func (a *InodeAttrs) Stat(*vfs.Filesystem, vfs.StatOptions) (linux.Statx, error) { var stat linux.Statx stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_NLINK stat.Ino = atomic.LoadUint64(&a.ino) @@ -230,7 +230,7 @@ func (a *InodeAttrs) Stat(*vfs.Filesystem) linux.Statx { // TODO: Implement other stat fields like timestamps. - return stat + return stat, nil } // SetStat implements Inode.SetStat. diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index c74fa999b..a8ab2a2ba 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -176,8 +176,6 @@ type Dentry struct { vfsd vfs.Dentry inode Inode - refs uint64 - // flags caches useful information about the dentry from the inode. See the // dflags* consts above. Must be accessed by atomic ops. flags uint32 @@ -302,7 +300,8 @@ type Inode interface { // this inode. The returned file description should hold a reference on the // inode for its lifetime. // - // Precondition: !rp.Done(). vfsd.Impl() must be a kernfs Dentry. + // Precondition: rp.Done(). vfsd.Impl() must be the kernfs Dentry containing + // the inode on which Open() is being called. Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) } @@ -328,7 +327,7 @@ type inodeMetadata interface { // Stat returns the metadata for this inode. This corresponds to // vfs.FilesystemImpl.StatAt. - Stat(fs *vfs.Filesystem) linux.Statx + Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) // SetStat updates the metadata for this inode. This corresponds to // vfs.FilesystemImpl.SetStatAt. diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index f3f4e49b4..611645f3f 100644 --- a/pkg/sentry/fsimpl/proc/subtasks.go +++ b/pkg/sentry/fsimpl/proc/subtasks.go @@ -121,8 +121,13 @@ func (i *subtasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.O } // Stat implements kernfs.Inode. -func (i *subtasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { - stat := i.InodeAttrs.Stat(vsfs) - stat.Nlink += uint32(i.task.ThreadGroup().Count()) - return stat +func (i *subtasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.InodeAttrs.Stat(vsfs, opts) + if err != nil { + return linux.Statx{}, err + } + if opts.Mask&linux.STATX_NLINK != 0 { + stat.Nlink += uint32(i.task.ThreadGroup().Count()) + } + return stat, nil } diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index 18e5cd6f6..c0d643f51 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -154,12 +154,21 @@ func newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.FileMode, childre } // Stat implements kernfs.Inode. -func (i *taskOwnedInode) Stat(fs *vfs.Filesystem) linux.Statx { - stat := i.Inode.Stat(fs) - uid, gid := i.getOwner(linux.FileMode(stat.Mode)) - stat.UID = uint32(uid) - stat.GID = uint32(gid) - return stat +func (i *taskOwnedInode) Stat(fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.Inode.Stat(fs, opts) + if err != nil { + return linux.Statx{}, err + } + if opts.Mask&(linux.STATX_UID|linux.STATX_GID) != 0 { + uid, gid := i.getOwner(linux.FileMode(stat.Mode)) + if opts.Mask&linux.STATX_UID != 0 { + stat.UID = uint32(uid) + } + if opts.Mask&linux.STATX_GID != 0 { + stat.GID = uint32(gid) + } + } + return stat, nil } // CheckPermissions implements kernfs.Inode. @@ -236,7 +245,7 @@ func newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentr // member, there is one entry containing three colon-separated fields: // hierarchy-ID:controller-list:cgroup-path" func newCgroupData(controllers map[string]string) dynamicInode { - buf := bytes.Buffer{} + var buf bytes.Buffer // The hierarchy ids must be positive integers (for cgroup v1), but the // exact number does not matter, so long as they are unique. We can diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go index 10c08fa90..b1e39c82f 100644 --- a/pkg/sentry/fsimpl/proc/tasks.go +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -211,17 +211,22 @@ func (i *tasksInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.Open return fd.VFSFileDescription(), nil } -func (i *tasksInode) Stat(vsfs *vfs.Filesystem) linux.Statx { - stat := i.InodeAttrs.Stat(vsfs) +func (i *tasksInode) Stat(vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + stat, err := i.InodeAttrs.Stat(vsfs, opts) + if err != nil { + return linux.Statx{}, err + } - // Add dynamic children to link count. - for _, tg := range i.pidns.ThreadGroups() { - if leader := tg.Leader(); leader != nil { - stat.Nlink++ + if opts.Mask&linux.STATX_NLINK != 0 { + // Add dynamic children to link count. + for _, tg := range i.pidns.ThreadGroups() { + if leader := tg.Leader(); leader != nil { + stat.Nlink++ + } } } - return stat + return stat, nil } func cpuInfoData(k *kernel.Kernel) string { diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 07c8383e6..cb4deb068 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -42,10 +42,13 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/fd", "//pkg/fspath", "//pkg/gohacks", "//pkg/log", + "//pkg/safemem", "//pkg/sentry/arch", + "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", @@ -53,6 +56,7 @@ go_library( "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index c2a52ec1b..45191d1c3 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -33,8 +33,8 @@ import ( // implementations to adapt: // - Have a local fileDescription struct (containing FileDescription) which // embeds FileDescriptionDefaultImpl and overrides the default methods -// which are common to all fd implementations for that for that filesystem -// like StatusFlags, SetStatusFlags, Stat, SetStat, StatFS, etc. +// which are common to all fd implementations for that filesystem like +// StatusFlags, SetStatusFlags, Stat, SetStat, StatFS, etc. // - This should be embedded in all file description implementations as the // first field by value. // - Directory FDs would also embed DirectoryFileDescriptionDefaultImpl. -- cgit v1.2.3