diff options
Diffstat (limited to 'pkg/sentry/fs')
| -rw-r--r-- | pkg/sentry/fs/g3doc/fuse.md | 99 | ||||
| -rw-r--r-- | pkg/sentry/fs/proc/BUILD | 1 | ||||
| -rw-r--r-- | pkg/sentry/fs/proc/sys_net.go | 120 | ||||
| -rw-r--r-- | pkg/sentry/fs/proc/sys_net_state.go | 15 | ||||
| -rw-r--r-- | pkg/sentry/fs/proc/sys_net_test.go | 73 | ||||
| -rw-r--r-- | pkg/sentry/fs/proc/task.go | 44 |
6 files changed, 349 insertions, 3 deletions
diff --git a/pkg/sentry/fs/g3doc/fuse.md b/pkg/sentry/fs/g3doc/fuse.md index 2ca84dd74..05e043583 100644 --- a/pkg/sentry/fs/g3doc/fuse.md +++ b/pkg/sentry/fs/g3doc/fuse.md @@ -79,7 +79,7 @@ ops can be implemented in parallel. - Implement `/dev/fuse` - a character device used to establish an FD for communication between the sentry and the server daemon. -- Implement basic FUSE ops like `FUSE_INIT`, `FUSE_DESTROY`. +- Implement basic FUSE ops like `FUSE_INIT`. #### Read-only mount with basic file operations @@ -95,6 +95,103 @@ ops can be implemented in parallel. - Implement the remaining FUSE ops and decide if we can omit rarely used operations like ioctl. +### Design Details + +#### Lifecycle for a FUSE Request + +- User invokes a syscall +- Sentry prepares corresponding request + - If FUSE device is available + - Write the request in binary + - If FUSE device is full + - Kernel task blocked until available +- Sentry notifies the readers of fuse device that it's ready for read +- FUSE daemon reads the request and processes it +- Sentry waits until a reply is written to the FUSE device + - but returns directly for async requests +- FUSE daemon writes to the fuse device +- Sentry processes the reply + - For sync requests, unblock blocked kernel task + - For async requests, execute pre-specified callback if any +- Sentry returns the syscall to the user + +#### Channels and Queues for Requests in Different Stages + +`connection.initializedChan` + +- a channel that the requests issued before connection initialization blocks + on. + +`fd.queue` + +- a queue of requests that haven’t been read by the FUSE daemon yet. + +`fd.completions` + +- a map of the requests that have been prepared but not yet received a + response, including the ones on the `fd.queue`. + +`fd.waitQueue` + +- a queue of waiters that is waiting for the fuse device fd to be available, + such as the FUSE daemon. + +`fd.fullQueueCh` + +- a channel that the kernel task will be blocked on when the fd is not + available. + +#### Basic I/O Implementation + +Currently we have implemented basic functionalities of read and write for our +FUSE. We describe the design and ways to improve it here: + +##### Basic FUSE Read + +The vfs2 expects implementations of `vfs.FileDescriptionImpl.Read()` and +`vfs.FileDescriptionImpl.PRead()`. When a syscall is made, it will eventually +reach our implementation of those interface functions located at +`pkg/sentry/fsimpl/fuse/regular_file.go` for regular files. + +After validation checks of the input, sentry sends `FUSE_READ` requests to the +FUSE daemon. The FUSE daemon returns data after the `fuse_out_header` as the +responses. For the first version, we create a copy in kernel memory of those +data. They are represented as a byte slice in the marshalled struct. This +happens as a common process for all the FUSE responses at this moment at +`pkg/sentry/fsimpl/fuse/dev.go:writeLocked()`. We then directly copy from this +intermediate buffer to the input buffer provided by the read syscall. + +There is an extra requirement for FUSE: When mounting the FUSE fs, the mounter +or the FUSE daemon can specify a `max_read` or a `max_pages` parameter. They are +the upperbound of the bytes to read in each `FUSE_READ` request. We implemented +the code to handle the fragmented reads. + +To improve the performance: ideally we should have buffer cache to copy those +data from the responses of FUSE daemon into, as is also the design of several +other existing file system implementations for sentry, instead of a single-use +temporary buffer. Directly mapping the memory of one process to another could +also boost the performance, but to keep them isolated, we did not choose to do +so. + +##### Basic FUSE Write + +The vfs2 invokes implementations of `vfs.FileDescriptionImpl.Write()` and +`vfs.FileDescriptionImpl.PWrite()` on the regular file descriptor of FUSE when a +user makes write(2) and pwrite(2) syscall. + +For valid writes, sentry sends the bytes to write after a `FUSE_WRITE` header +(can be regarded as a request with 2 payloads) to the FUSE daemon. For the first +version, we allocate a buffer inside kernel memory to store the bytes from the +user, and copy directly from that buffer to the memory of FUSE daemon. This +happens at `pkg/sentry/fsimpl/fuse/dev.go:readLocked()` + +The parameters `max_write` and `max_pages` restrict the number of bytes in one +`FUSE_WRITE`. There are code handling fragmented writes in current +implementation. + +To have better performance: the extra copy created to store the bytes to write +can be replaced by the buffer cache as well. + # Appendix ## FUSE Protocol diff --git a/pkg/sentry/fs/proc/BUILD b/pkg/sentry/fs/proc/BUILD index 77c2c5c0e..b8b2281a8 100644 --- a/pkg/sentry/fs/proc/BUILD +++ b/pkg/sentry/fs/proc/BUILD @@ -50,6 +50,7 @@ go_library( "//pkg/sync", "//pkg/syserror", "//pkg/tcpip/header", + "//pkg/tcpip/network/ipv4", "//pkg/usermem", "//pkg/waiter", ], diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go index 8615b60f0..e555672ad 100644 --- a/pkg/sentry/fs/proc/sys_net.go +++ b/pkg/sentry/fs/proc/sys_net.go @@ -26,6 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -54,7 +55,7 @@ type tcpMemInode struct { // size stores the tcp buffer size during save, and sets the buffer // size in netstack in restore. We must save/restore this here, since - // netstack itself is stateless. + // a netstack instance is created on restore. size inet.TCPBufferSize // mu protects against concurrent reads/writes to files based on this @@ -258,6 +259,9 @@ func (f *tcpSackFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSeque if src.NumBytes() == 0 { return 0, nil } + + // Only consider size of one memory page for input for performance reasons. + // We are only reading if it's zero or not anyway. src = src.TakeFirst(usermem.PageSize - 1) var v int32 @@ -383,11 +387,125 @@ func (p *proc) newSysNetCore(ctx context.Context, msrc *fs.MountSource, s inet.S return newProcInode(ctx, d, msrc, fs.SpecialDirectory, nil) } +// ipForwarding implements fs.InodeOperations. +// +// ipForwarding is used to enable/disable packet forwarding of netstack. +// +// +stateify savable +type ipForwarding struct { + fsutil.SimpleFileInode + + stack inet.Stack `state:"wait"` + + // enabled stores the IPv4 forwarding state on save. + // We must save/restore this here, since a netstack instance + // is created on restore. + enabled *bool +} + +func newIPForwardingInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode { + ipf := &ipForwarding{ + SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0444), linux.PROC_SUPER_MAGIC), + stack: s, + } + sattr := fs.StableAttr{ + DeviceID: device.ProcDevice.DeviceID(), + InodeID: device.ProcDevice.NextIno(), + BlockSize: usermem.PageSize, + Type: fs.SpecialFile, + } + return fs.NewInode(ctx, ipf, msrc, sattr) +} + +// 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 (*ipForwarding) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + +// +stateify savable +type ipForwardingFile 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"` + + ipf *ipForwarding + + stack inet.Stack `state:"wait"` +} + +// GetFile implements fs.InodeOperations.GetFile. +func (ipf *ipForwarding) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + flags.Pread = true + flags.Pwrite = true + return fs.NewFile(ctx, dirent, flags, &ipForwardingFile{ + stack: ipf.stack, + ipf: ipf, + }), nil +} + +// Read implements fs.FileOperations.Read. +func (f *ipForwardingFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + if offset != 0 { + return 0, io.EOF + } + + if f.ipf.enabled == nil { + enabled := f.stack.Forwarding(ipv4.ProtocolNumber) + f.ipf.enabled = &enabled + } + + val := "0\n" + if *f.ipf.enabled { + // Technically, this is not quite compatible with Linux. Linux + // stores these as an integer, so if you write "2" into + // ip_forward, you should get 2 back. + val = "1\n" + } + n, err := dst.CopyOut(ctx, []byte(val)) + return int64(n), err +} + +// Write implements fs.FileOperations.Write. +// +// Offset is ignored, multiple writes are not supported. +func (f *ipForwardingFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Only consider size of one memory page for input for performance reasons. + // We are only reading if it's zero or not anyway. + 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 n, err + } + if f.ipf.enabled == nil { + f.ipf.enabled = new(bool) + } + *f.ipf.enabled = v != 0 + return n, f.stack.SetForwarding(ipv4.ProtocolNumber, *f.ipf.enabled) +} + func (p *proc) newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode { contents := map[string]*fs.Inode{ // Add tcp_sack. "tcp_sack": newTCPSackInode(ctx, msrc, s), + // Add ip_forward. + "ip_forward": newIPForwardingInode(ctx, msrc, s), + // The following files are simple stubs until they are // implemented in netstack, most of these files are // configuration related. We use the value closest to the diff --git a/pkg/sentry/fs/proc/sys_net_state.go b/pkg/sentry/fs/proc/sys_net_state.go index 6eba709c6..4cb4741af 100644 --- a/pkg/sentry/fs/proc/sys_net_state.go +++ b/pkg/sentry/fs/proc/sys_net_state.go @@ -14,7 +14,11 @@ package proc -import "fmt" +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" +) // beforeSave is invoked by stateify. func (t *tcpMemInode) beforeSave() { @@ -40,3 +44,12 @@ func (s *tcpSack) afterLoad() { } } } + +// afterLoad is invoked by stateify. +func (ipf *ipForwarding) afterLoad() { + if ipf.enabled != nil { + if err := ipf.stack.SetForwarding(ipv4.ProtocolNumber, *ipf.enabled); err != nil { + panic(fmt.Sprintf("failed to set IPv4 forwarding [%v]: %v", *ipf.enabled, err)) + } + } +} diff --git a/pkg/sentry/fs/proc/sys_net_test.go b/pkg/sentry/fs/proc/sys_net_test.go index 355e83d47..6ef5738e7 100644 --- a/pkg/sentry/fs/proc/sys_net_test.go +++ b/pkg/sentry/fs/proc/sys_net_test.go @@ -123,3 +123,76 @@ func TestConfigureRecvBufferSize(t *testing.T) { } } } + +// TestIPForwarding tests the implementation of +// /proc/sys/net/ipv4/ip_forwarding +func TestIPForwarding(t *testing.T) { + ctx := context.Background() + s := inet.NewTestStack() + + var cases = []struct { + comment string + initial bool + str string + final bool + }{ + { + comment: `Forwarding is disabled; write 1 and enable forwarding`, + initial: false, + str: "1", + final: true, + }, + { + comment: `Forwarding is disabled; write 0 and disable forwarding`, + initial: false, + str: "0", + final: false, + }, + { + comment: `Forwarding is enabled; write 1 and enable forwarding`, + initial: true, + str: "1", + final: true, + }, + { + comment: `Forwarding is enabled; write 0 and disable forwarding`, + initial: true, + str: "0", + final: false, + }, + { + comment: `Forwarding is disabled; write 2404 and enable forwarding`, + initial: false, + str: "2404", + final: true, + }, + { + comment: `Forwarding is enabled; write 2404 and enable forwarding`, + initial: true, + str: "2404", + final: true, + }, + } + for _, c := range cases { + t.Run(c.comment, func(t *testing.T) { + s.IPForwarding = c.initial + ipf := &ipForwarding{stack: s} + file := &ipForwardingFile{ + stack: s, + ipf: ipf, + } + + // Write the values. + src := usermem.BytesIOSequence([]byte(c.str)) + if n, err := file.Write(ctx, nil, src, 0); n != int64(len(c.str)) || err != nil { + t.Errorf("file.Write(ctx, nil, %q, 0) = (%d, %v); want (%d, nil)", c.str, n, err, len(c.str)) + } + + // Read the values from the stack and check them. + if got, want := s.IPForwarding, c.final; got != want { + t.Errorf("s.IPForwarding incorrect; got: %v, want: %v", got, want) + } + + }) + } +} diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index 103bfc600..22d658acf 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -84,6 +84,7 @@ func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bo "auxv": newAuxvec(t, msrc), "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), "comm": newComm(t, msrc), + "cwd": newCwd(t, msrc), "environ": newExecArgInode(t, msrc, environExecArg), "exe": newExe(t, msrc), "fd": newFdDir(t, msrc), @@ -300,6 +301,49 @@ func (e *exe) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { return exec.PathnameWithDeleted(ctx), nil } +// cwd is an fs.InodeOperations symlink for the /proc/PID/cwd file. +// +// +stateify savable +type cwd struct { + ramfs.Symlink + + t *kernel.Task +} + +func newCwd(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + cwdSymlink := &cwd{ + Symlink: *ramfs.NewSymlink(t, fs.RootOwner, ""), + t: t, + } + return newProcInode(t, cwdSymlink, msrc, fs.Symlink, t) +} + +// Readlink implements fs.InodeOperations. +func (e *cwd) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { + if !kernel.ContextCanTrace(ctx, e.t, false) { + return "", syserror.EACCES + } + if err := checkTaskState(e.t); err != nil { + return "", err + } + cwd := e.t.FSContext().WorkingDirectory() + if cwd == nil { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer cwd.DecRef(ctx) + + root := fs.RootFromContext(ctx) + if root == nil { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer root.DecRef(ctx) + + name, _ := cwd.FullName(root) + return name, nil +} + // namespaceSymlink represents a symlink in the namespacefs, such as the files // in /proc/<pid>/ns. // |