summaryrefslogtreecommitdiffhomepage
path: root/pkg/sentry/fs/gofer/path.go
diff options
context:
space:
mode:
Diffstat (limited to 'pkg/sentry/fs/gofer/path.go')
-rw-r--r--pkg/sentry/fs/gofer/path.go213
1 files changed, 137 insertions, 76 deletions
diff --git a/pkg/sentry/fs/gofer/path.go b/pkg/sentry/fs/gofer/path.go
index 8c17603f8..3c66dc3c2 100644
--- a/pkg/sentry/fs/gofer/path.go
+++ b/pkg/sentry/fs/gofer/path.go
@@ -16,21 +16,30 @@ package gofer
import (
"fmt"
- "syscall"
+ "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/log"
"gvisor.dev/gvisor/pkg/p9"
- "gvisor.dev/gvisor/pkg/sentry/context"
"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) {
@@ -58,7 +67,7 @@ func (i *inodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string
// Get a p9.File for name.
qids, newFile, mask, p9attr, err := i.fileState.file.walkGetAttr(ctx, []string{name})
if err != nil {
- if err == syscall.ENOENT {
+ if err == syserror.ENOENT {
if cp.cacheNegativeDirents() {
// Return a negative Dirent. It will stay cached until something
// is created over it.
@@ -69,8 +78,25 @@ 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
@@ -138,11 +164,11 @@ 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)
- defer d.DecRef()
+ defer d.DecRef(ctx)
// Construct the new file, caching the handles if allowed.
h := handles{
@@ -180,7 +206,7 @@ func (i *inodeOperations) CreateHardLink(ctx context.Context, inode *fs.Inode, t
targetOpts, ok := target.InodeOperations.(*inodeOperations)
if !ok {
- return syscall.EXDEV
+ return syserror.EXDEV
}
if err := i.fileState.file.link(ctx, &targetOpts.fileState.file, newName); err != nil {
@@ -223,82 +249,115 @@ func (i *inodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string,
return nil, syserror.ENAMETOOLONG
}
- if i.session().endpoints == nil {
- return nil, syscall.EOPNOTSUPP
+ if i.session().overrides == nil {
+ return nil, syserror.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
}
- // 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 {
- newFile.close(ctx)
- 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 != syserror.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 {
- newFile.close(ctx)
- 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 +366,23 @@ 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()
+ defer d.DecRef(ctx)
+
+ 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 +391,8 @@ 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(ctx, *key)
}
i.touchModificationAndStatusChangeTime(ctx, dir)
@@ -364,17 +426,16 @@ func (i *inodeOperations) Rename(ctx context.Context, inode *fs.Inode, oldParent
return syserror.ENAMETOOLONG
}
- // Unwrap the new parent to a *inodeOperations.
- newParentInodeOperations, ok := newParent.InodeOperations.(*inodeOperations)
- if !ok {
- return syscall.EXDEV
+ // Don't allow renames across different mounts.
+ if newParent.MountSource != oldParent.MountSource {
+ return syserror.EXDEV
}
+ // Unwrap the new parent to a *inodeOperations.
+ newParentInodeOperations := newParent.InodeOperations.(*inodeOperations)
+
// Unwrap the old parent to a *inodeOperations.
- oldParentInodeOperations, ok := oldParent.InodeOperations.(*inodeOperations)
- if !ok {
- return syscall.EXDEV
- }
+ oldParentInodeOperations := oldParent.InodeOperations.(*inodeOperations)
// Do the rename.
if err := i.fileState.file.rename(ctx, newParentInodeOperations.fileState.file, newName); err != nil {
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-17 03:46:35 +0000