diff options
Diffstat (limited to 'pkg/sentry')
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/BUILD | 3 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/directory.go | 101 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/filesystem.go | 487 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/gofer.go | 668 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/gofer_test.go | 1 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/handle.go | 80 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/p9file.go | 8 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/regular_file.go | 26 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/revalidate.go | 50 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/save_restore.go | 143 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/socket.go | 45 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/special_file.go | 18 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/symlink.go | 8 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/gofer/time.go | 5 | ||||
| -rw-r--r-- | pkg/sentry/vfs/resolving_path.go | 6 |
15 files changed, 1318 insertions, 331 deletions
diff --git a/pkg/sentry/fsimpl/gofer/BUILD b/pkg/sentry/fsimpl/gofer/BUILD index 4244f2cf5..509dd0e1a 100644 --- a/pkg/sentry/fsimpl/gofer/BUILD +++ b/pkg/sentry/fsimpl/gofer/BUILD @@ -54,7 +54,10 @@ go_library( "//pkg/fdnotifier", "//pkg/fspath", "//pkg/hostarch", + "//pkg/lisafs", "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/metric", "//pkg/p9", "//pkg/refs", diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go index 5c48a9fee..d99a6112c 100644 --- a/pkg/sentry/fsimpl/gofer/directory.go +++ b/pkg/sentry/fsimpl/gofer/directory.go @@ -222,47 +222,88 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { off := uint64(0) const count = 64 * 1024 // for consistency with the vfs1 client d.handleMu.RLock() - if d.readFile.isNil() { + if !d.isReadFileOk() { // This should not be possible because a readable handle should // have been opened when the calling directoryFD was opened. d.handleMu.RUnlock() panic("gofer.dentry.getDirents called without a readable handle") } + // shouldSeek0 indicates whether the server should SEEK to 0 before reading + // directory entries. + shouldSeek0 := true for { - p9ds, err := d.readFile.readdir(ctx, off, count) - if err != nil { - d.handleMu.RUnlock() - return nil, err - } - if len(p9ds) == 0 { - d.handleMu.RUnlock() - break - } - for _, p9d := range p9ds { - if p9d.Name == "." || p9d.Name == ".." { - continue + if d.fs.opts.lisaEnabled { + countLisa := int32(count) + if shouldSeek0 { + // See lisafs.Getdents64Req.Count. + countLisa = -countLisa + shouldSeek0 = false + } + lisafsDs, err := d.readFDLisa.Getdents64(ctx, countLisa) + if err != nil { + d.handleMu.RUnlock() + return nil, err + } + if len(lisafsDs) == 0 { + d.handleMu.RUnlock() + break + } + for i := range lisafsDs { + name := string(lisafsDs[i].Name) + if name == "." || name == ".." { + continue + } + dirent := vfs.Dirent{ + Name: name, + Ino: d.fs.inoFromKey(inoKey{ + ino: uint64(lisafsDs[i].Ino), + devMinor: uint32(lisafsDs[i].DevMinor), + devMajor: uint32(lisafsDs[i].DevMajor), + }), + NextOff: int64(len(dirents) + 1), + Type: uint8(lisafsDs[i].Type), + } + dirents = append(dirents, dirent) + if realChildren != nil { + realChildren[name] = struct{}{} + } } - dirent := vfs.Dirent{ - Name: p9d.Name, - Ino: d.fs.inoFromQIDPath(p9d.QID.Path), - NextOff: int64(len(dirents) + 1), + } else { + p9ds, err := d.readFile.readdir(ctx, off, count) + if err != nil { + d.handleMu.RUnlock() + return nil, err } - // 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 + if len(p9ds) == 0 { + d.handleMu.RUnlock() + break } - dirents = append(dirents, dirent) - if realChildren != nil { - realChildren[p9d.Name] = struct{}{} + for _, p9d := range p9ds { + if p9d.Name == "." || p9d.Name == ".." { + continue + } + dirent := vfs.Dirent{ + Name: p9d.Name, + Ino: d.fs.inoFromQIDPath(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) + if realChildren != nil { + realChildren[p9d.Name] = struct{}{} + } } + off = p9ds[len(p9ds)-1].Offset } - off = p9ds[len(p9ds)-1].Offset } } // Emit entries for synthetic children. diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go index 9d943bd4a..f7b3446d3 100644 --- a/pkg/sentry/fsimpl/gofer/filesystem.go +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -21,10 +21,12 @@ import ( "sync" "sync/atomic" + "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/errors/linuxerr" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/lisafs" "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/sentry/fsimpl/host" "gvisor.dev/gvisor/pkg/sentry/fsmetric" @@ -53,9 +55,47 @@ func (fs *filesystem) Sync(ctx context.Context) error { // regardless. var retErr error + if fs.opts.lisaEnabled { + // Try accumulating all FDIDs to fsync and fsync then via one RPC as + // opposed to making an RPC per FDID. Passing a non-nil accFsyncFDIDs to + // dentry.syncCachedFile() and specialFileFD.sync() will cause them to not + // make an RPC, instead accumulate syncable FDIDs in the passed slice. + accFsyncFDIDs := make([]lisafs.FDID, 0, len(ds)+len(sffds)) + + // Sync syncable dentries. + for _, d := range ds { + if err := d.syncCachedFile(ctx, true /* forFilesystemSync */, &accFsyncFDIDs); err != nil { + ctx.Infof("gofer.filesystem.Sync: dentry.syncCachedFile failed: %v", err) + if 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 err := sffd.sync(ctx, true /* forFilesystemSync */, &accFsyncFDIDs); err != nil { + ctx.Infof("gofer.filesystem.Sync: specialFileFD.sync failed: %v", err) + if retErr == nil { + retErr = err + } + } + } + + if err := fs.clientLisa.SyncFDs(ctx, accFsyncFDIDs); err != nil { + ctx.Infof("gofer.filesystem.Sync: fs.fsyncMultipleFDLisa failed: %v", err) + if retErr == nil { + retErr = err + } + } + + return retErr + } + // Sync syncable dentries. for _, d := range ds { - if err := d.syncCachedFile(ctx, true /* forFilesystemSync */); err != nil { + if err := d.syncCachedFile(ctx, true /* forFilesystemSync */, nil /* accFsyncFDIDsLisa */); err != nil { ctx.Infof("gofer.filesystem.Sync: dentry.syncCachedFile failed: %v", err) if retErr == nil { retErr = err @@ -66,7 +106,7 @@ func (fs *filesystem) Sync(ctx context.Context) error { // 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 err := sffd.sync(ctx, true /* forFilesystemSync */); err != nil { + if err := sffd.sync(ctx, true /* forFilesystemSync */, nil /* accFsyncFDIDsLisa */); err != nil { ctx.Infof("gofer.filesystem.Sync: specialFileFD.sync failed: %v", err) if retErr == nil { retErr = err @@ -197,7 +237,13 @@ afterSymlink: rp.Advance() return d.parent, followedSymlink, nil } - child, err := fs.getChildLocked(ctx, d, name, ds) + var child *dentry + var err error + if fs.opts.lisaEnabled { + child, err = fs.getChildAndWalkPathLocked(ctx, d, rp, ds) + } else { + child, err = fs.getChildLocked(ctx, d, name, ds) + } if err != nil { return nil, false, err } @@ -219,6 +265,99 @@ afterSymlink: return child, followedSymlink, nil } +// Preconditions: +// * fs.opts.lisaEnabled. +// * fs.renameMu must be locked. +// * parent.dirMu must be locked. +// * parent.isDir(). +// * parent and the dentry at name have been revalidated. +func (fs *filesystem) getChildAndWalkPathLocked(ctx context.Context, parent *dentry, rp *vfs.ResolvingPath, ds **[]*dentry) (*dentry, error) { + // Note that pit is a copy of the iterator that does not affect rp. + pit := rp.Pit() + first := pit.String() + if len(first) > maxFilenameLen { + return nil, linuxerr.ENAMETOOLONG + } + if child, ok := parent.children[first]; ok || parent.isSynthetic() { + if child == nil { + return nil, linuxerr.ENOENT + } + return child, nil + } + + // Walk as much of the path as possible in 1 RPC. + names := []string{first} + for pit = pit.Next(); pit.Ok(); pit = pit.Next() { + name := pit.String() + if name == "." { + continue + } + if name == ".." { + break + } + names = append(names, name) + } + status, inodes, err := parent.controlFDLisa.WalkMultiple(ctx, names) + if err != nil { + return nil, err + } + if len(inodes) == 0 { + parent.cacheNegativeLookupLocked(first) + return nil, linuxerr.ENOENT + } + + // Add the walked inodes into the dentry tree. + curParent := parent + curParentDirMuLock := func() { + if curParent != parent { + curParent.dirMu.Lock() + } + } + curParentDirMuUnlock := func() { + if curParent != parent { + curParent.dirMu.Unlock() // +checklocksforce: locked via curParentDirMuLock(). + } + } + var ret *dentry + var dentryCreationErr error + for i := range inodes { + if dentryCreationErr != nil { + fs.clientLisa.CloseFDBatched(ctx, inodes[i].ControlFD) + continue + } + + child, err := fs.newDentryLisa(ctx, &inodes[i]) + if err != nil { + fs.clientLisa.CloseFDBatched(ctx, inodes[i].ControlFD) + dentryCreationErr = err + continue + } + curParentDirMuLock() + curParent.cacheNewChildLocked(child, names[i]) + curParentDirMuUnlock() + // For now, child has 0 references, so our caller should call + // child.checkCachingLocked(). curParent gained a ref so we should also + // call curParent.checkCachingLocked() so it can be removed from the cache + // if needed. We only do that for the first iteration because all + // subsequent parents would have already been added to ds. + if i == 0 { + *ds = appendDentry(*ds, curParent) + } + *ds = appendDentry(*ds, child) + curParent = child + if i == 0 { + ret = child + } + } + + if status == lisafs.WalkComponentDoesNotExist && curParent.isDir() { + curParentDirMuLock() + curParent.cacheNegativeLookupLocked(names[len(inodes)]) + curParentDirMuUnlock() + } + return ret, dentryCreationErr +} + // getChildLocked returns a dentry representing the child of parent with the // given name. Returns ENOENT if the child doesn't exist. // @@ -227,7 +366,7 @@ afterSymlink: // * parent.dirMu must be locked. // * parent.isDir(). // * name is not "." or "..". -// * dentry at name has been revalidated +// * parent and the dentry at name have been revalidated. func (fs *filesystem) getChildLocked(ctx context.Context, parent *dentry, name string, ds **[]*dentry) (*dentry, error) { if len(name) > maxFilenameLen { return nil, linuxerr.ENAMETOOLONG @@ -239,20 +378,35 @@ func (fs *filesystem) getChildLocked(ctx context.Context, parent *dentry, name s return child, nil } - qid, file, attrMask, attr, err := parent.file.walkGetAttrOne(ctx, name) - if err != nil { - if linuxerr.Equals(linuxerr.ENOENT, err) { - parent.cacheNegativeLookupLocked(name) + var child *dentry + if fs.opts.lisaEnabled { + childInode, err := parent.controlFDLisa.Walk(ctx, name) + if err != nil { + if linuxerr.Equals(linuxerr.ENOENT, err) { + parent.cacheNegativeLookupLocked(name) + } + return nil, err + } + // Create a new dentry representing the file. + child, err = fs.newDentryLisa(ctx, childInode) + if err != nil { + fs.clientLisa.CloseFDBatched(ctx, childInode.ControlFD) + return nil, err + } + } else { + qid, file, attrMask, attr, err := parent.file.walkGetAttrOne(ctx, name) + if err != nil { + if linuxerr.Equals(linuxerr.ENOENT, err) { + parent.cacheNegativeLookupLocked(name) + } + return nil, err + } + // 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 } - return nil, err - } - - // Create a new dentry representing the file. - child, err := fs.newDentry(ctx, file, qid, attrMask, &attr) - if err != nil { - file.close(ctx) - delete(parent.children, name) - return nil, err } parent.cacheNewChildLocked(child, name) appendNewChildDentry(ds, parent, child) @@ -328,7 +482,7 @@ func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, // Preconditions: // * !rp.Done(). // * For the final path component in rp, !rp.ShouldFollowSymlink(). -func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, createInRemoteDir func(parent *dentry, name string, ds **[]*dentry) error, createInSyntheticDir func(parent *dentry, name string) error) error { +func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, createInRemoteDir func(parent *dentry, name string, ds **[]*dentry) (*lisafs.Inode, error), createInSyntheticDir func(parent *dentry, name string) error, updateChild func(child *dentry)) error { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(ctx, &ds) @@ -415,9 +569,26 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir // No cached dentry exists; however, in InteropModeShared there might still be // an existing file at name. Just attempt the file creation RPC anyways. If a // file does exist, the RPC will fail with EEXIST like we would have. - if err := createInRemoteDir(parent, name, &ds); err != nil { + lisaInode, err := createInRemoteDir(parent, name, &ds) + if err != nil { return err } + // lisafs may aggresively cache newly created inodes. This has helped reduce + // Walk RPCs in practice. + if lisaInode != nil { + child, err := fs.newDentryLisa(ctx, lisaInode) + if err != nil { + fs.clientLisa.CloseFDBatched(ctx, lisaInode.ControlFD) + return err + } + parent.cacheNewChildLocked(child, name) + appendNewChildDentry(&ds, parent, child) + + // lisafs may update dentry properties upon successful creation. + if updateChild != nil { + updateChild(child) + } + } if fs.opts.interop != InteropModeShared { if child, ok := parent.children[name]; ok && child == nil { // Delete the now-stale negative dentry. @@ -565,7 +736,11 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b return linuxerr.ENOENT } } else if child == nil || !child.isSynthetic() { - err = parent.file.unlinkAt(ctx, name, flags) + if fs.opts.lisaEnabled { + err = parent.controlFDLisa.UnlinkAt(ctx, name, flags) + } else { + err = parent.file.unlinkAt(ctx, name, flags) + } if err != nil { if child != nil { vfsObj.AbortDeleteDentry(&child.vfsd) // +checklocksforce: see above. @@ -658,40 +833,43 @@ func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPa // 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, _ **[]*dentry) error { + err := fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, childName string, ds **[]*dentry) (*lisafs.Inode, error) { if rp.Mount() != vd.Mount() { - return linuxerr.EXDEV + return nil, linuxerr.EXDEV } d := vd.Dentry().Impl().(*dentry) if d.isDir() { - return linuxerr.EPERM + return nil, linuxerr.EPERM } gid := auth.KGID(atomic.LoadUint32(&d.gid)) uid := auth.KUID(atomic.LoadUint32(&d.uid)) mode := linux.FileMode(atomic.LoadUint32(&d.mode)) if err := vfs.MayLink(rp.Credentials(), mode, uid, gid); err != nil { - return err + return nil, err } if d.nlink == 0 { - return linuxerr.ENOENT + return nil, linuxerr.ENOENT } if d.nlink == math.MaxUint32 { - return linuxerr.EMLINK + return nil, linuxerr.EMLINK } - if err := parent.file.link(ctx, d.file, childName); err != nil { - return err + if fs.opts.lisaEnabled { + return parent.controlFDLisa.LinkAt(ctx, d.controlFDLisa.ID(), childName) } + return nil, parent.file.link(ctx, d.file, childName) + }, nil, nil) + if err == nil { // Success! - atomic.AddUint32(&d.nlink, 1) - return nil - }, nil) + vd.Dentry().Impl().(*dentry).incLinks() + } + return err } // MkdirAt implements vfs.FilesystemImpl.MkdirAt. func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { creds := rp.Credentials() - return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string, ds **[]*dentry) error { + return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string, ds **[]*dentry) (*lisafs.Inode, error) { // If the parent is a setgid directory, use the parent's GID // rather than the caller's and enable setgid. kgid := creds.EffectiveKGID @@ -700,9 +878,18 @@ func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v kgid = auth.KGID(atomic.LoadUint32(&parent.gid)) mode |= linux.S_ISGID } - if _, err := parent.file.mkdir(ctx, name, p9.FileMode(mode), (p9.UID)(creds.EffectiveKUID), p9.GID(kgid)); err != nil { + var ( + childDirInode *lisafs.Inode + err error + ) + if fs.opts.lisaEnabled { + childDirInode, err = parent.controlFDLisa.MkdirAt(ctx, name, mode, lisafs.UID(creds.EffectiveKUID), lisafs.GID(kgid)) + } else { + _, err = parent.file.mkdir(ctx, name, p9.FileMode(mode), (p9.UID)(creds.EffectiveKUID), p9.GID(kgid)) + } + if err != nil { if !opts.ForSyntheticMountpoint || linuxerr.Equals(linuxerr.EEXIST, err) { - return err + return nil, err } ctx.Infof("Failed to create remote directory %q: %v; falling back to synthetic directory", name, err) parent.createSyntheticChildLocked(&createSyntheticOpts{ @@ -716,7 +903,7 @@ func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v if fs.opts.interop != InteropModeShared { parent.incLinks() } - return nil + return childDirInode, nil }, func(parent *dentry, name string) error { if !opts.ForSyntheticMountpoint { // Can't create non-synthetic files in synthetic directories. @@ -730,16 +917,26 @@ func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v }) parent.incLinks() return nil - }) + }, nil) } // 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, ds **[]*dentry) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string, ds **[]*dentry) (*lisafs.Inode, 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)) - if !linuxerr.Equals(linuxerr.EPERM, err) { - return err + var ( + childInode *lisafs.Inode + err error + ) + if fs.opts.lisaEnabled { + childInode, err = parent.controlFDLisa.MknodAt(ctx, name, opts.Mode, lisafs.UID(creds.EffectiveKUID), lisafs.GID(creds.EffectiveKGID), opts.DevMinor, opts.DevMajor) + } else { + _, err = parent.file.mknod(ctx, name, (p9.FileMode)(opts.Mode), opts.DevMajor, opts.DevMinor, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) + } + if err == nil { + return childInode, nil + } else if !linuxerr.Equals(linuxerr.EPERM, err) { + return nil, err } // EPERM means that gofer does not allow creating a socket or pipe. Fallback @@ -750,10 +947,10 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v switch { case err == nil: // Step succeeded, another file exists. - return linuxerr.EEXIST + return nil, linuxerr.EEXIST case !linuxerr.Equals(linuxerr.ENOENT, err): // Unexpected error. - return err + return nil, err } switch opts.Mode.FileType() { @@ -766,7 +963,7 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v endpoint: opts.Endpoint, }) *ds = appendDentry(*ds, parent) - return nil + return nil, nil case linux.S_IFIFO: parent.createSyntheticChildLocked(&createSyntheticOpts{ name: name, @@ -776,11 +973,11 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize), }) *ds = appendDentry(*ds, parent) - return nil + return nil, nil } // Retain error from gofer if synthetic file cannot be created internally. - return linuxerr.EPERM - }, nil) + return nil, linuxerr.EPERM + }, nil, nil) } // OpenAt implements vfs.FilesystemImpl.OpenAt. @@ -986,6 +1183,23 @@ func (d *dentry) openSocketByConnecting(ctx context.Context, opts *vfs.OpenOptio if opts.Flags&linux.O_DIRECT != 0 { return nil, linuxerr.EINVAL } + if d.fs.opts.lisaEnabled { + // Note that special value of linux.SockType = 0 is interpreted by lisafs + // as "do not care about the socket type". Analogous to p9.AnonymousSocket. + sockFD, err := d.controlFDLisa.Connect(ctx, 0 /* sockType */) + if err != nil { + return nil, err + } + fd, err := host.NewFD(ctx, kernel.KernelFromContext(ctx).HostMount(), sockFD, &host.NewFDOptions{ + HaveFlags: true, + Flags: opts.Flags, + }) + if err != nil { + unix.Close(sockFD) + return nil, err + } + return fd, nil + } fdObj, err := d.file.connect(ctx, p9.AnonymousSocket) if err != nil { return nil, err @@ -998,6 +1212,7 @@ func (d *dentry) openSocketByConnecting(ctx context.Context, opts *vfs.OpenOptio fdObj.Close() return nil, err } + // Ownership has been transferred to fd. fdObj.Release() return fd, nil } @@ -1017,7 +1232,13 @@ func (d *dentry) openSpecialFile(ctx context.Context, mnt *vfs.Mount, opts *vfs. // since closed its end. isBlockingOpenOfNamedPipe := d.fileType() == linux.S_IFIFO && opts.Flags&linux.O_NONBLOCK == 0 retry: - h, err := openHandle(ctx, d.file, ats.MayRead(), ats.MayWrite(), opts.Flags&linux.O_TRUNC != 0) + var h handle + var err error + if d.fs.opts.lisaEnabled { + h, err = openHandleLisa(ctx, d.controlFDLisa, ats.MayRead(), ats.MayWrite(), opts.Flags&linux.O_TRUNC != 0) + } else { + h, err = openHandle(ctx, d.file, ats.MayRead(), ats.MayWrite(), opts.Flags&linux.O_TRUNC != 0) + } if err != nil { if isBlockingOpenOfNamedPipe && ats == vfs.MayWrite && linuxerr.Equals(linuxerr.ENXIO, err) { // An attempt to open a named pipe with O_WRONLY|O_NONBLOCK fails @@ -1061,18 +1282,8 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving } 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() - // We only want the access mode for creating the file. - createFlags := p9.OpenFlags(opts.Flags) & p9.OpenFlagsModeMask - // If the parent is a setgid directory, use the parent's GID rather // than the caller's. kgid := creds.EffectiveKGID @@ -1080,51 +1291,87 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving kgid = auth.KGID(atomic.LoadUint32(&d.gid)) } - fdobj, openFile, createQID, _, err := dirfile.create(ctx, name, createFlags, p9.FileMode(opts.Mode), (p9.UID)(creds.EffectiveKUID), p9.GID(kgid)) - 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(). - _, nonOpenFile, attrMask, attr, err := d.file.walkGetAttrOne(ctx, name) - if err != nil { - openFile.close(ctx) - if fdobj != nil { - fdobj.Close() + var child *dentry + var openP9File p9file + openLisaFD := lisafs.InvalidFDID + openHostFD := int32(-1) + if d.fs.opts.lisaEnabled { + ino, openFD, hostFD, err := d.controlFDLisa.OpenCreateAt(ctx, name, opts.Flags&linux.O_ACCMODE, opts.Mode, lisafs.UID(creds.EffectiveKUID), lisafs.GID(kgid)) + if err != nil { + return nil, err + } + openHostFD = int32(hostFD) + openLisaFD = openFD + + child, err = d.fs.newDentryLisa(ctx, &ino) + if err != nil { + d.fs.clientLisa.CloseFDBatched(ctx, ino.ControlFD) + d.fs.clientLisa.CloseFDBatched(ctx, openFD) + if hostFD >= 0 { + unix.Close(hostFD) + } + return nil, err + } + } else { + // 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 + } + // We only want the access mode for creating the file. + createFlags := p9.OpenFlags(opts.Flags) & p9.OpenFlagsModeMask + + fdobj, openFile, createQID, _, err := dirfile.create(ctx, name, createFlags, p9.FileMode(opts.Mode), (p9.UID)(creds.EffectiveKUID), p9.GID(kgid)) + 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(). + _, nonOpenFile, attrMask, attr, err := d.file.walkGetAttrOne(ctx, name) + if err != nil { + openFile.close(ctx) + if fdobj != nil { + fdobj.Close() + } + return nil, err + } + + // 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 } - return nil, err - } - // 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() + openHostFD = int32(fdobj.Release()) } - return nil, err + openP9File = openFile } // Incorporate the fid that was opened by lcreate. useRegularFileFD := child.fileType() == linux.S_IFREG && !d.fs.opts.regularFilesUseSpecialFileFD if useRegularFileFD { - openFD := int32(-1) - if fdobj != nil { - openFD = int32(fdobj.Release()) - } child.handleMu.Lock() if vfs.MayReadFileWithOpenFlags(opts.Flags) { - child.readFile = openFile - if fdobj != nil { - child.readFD = openFD - child.mmapFD = openFD + child.readFile = openP9File + child.readFDLisa = d.fs.clientLisa.NewFD(openLisaFD) + if openHostFD != -1 { + child.readFD = openHostFD + child.mmapFD = openHostFD } } if vfs.MayWriteFileWithOpenFlags(opts.Flags) { - child.writeFile = openFile - child.writeFD = openFD + child.writeFile = openP9File + child.writeFDLisa = d.fs.clientLisa.NewFD(openLisaFD) + child.writeFD = openHostFD } child.handleMu.Unlock() } @@ -1146,11 +1393,9 @@ func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.Resolving childVFSFD = &fd.vfsfd } else { h := handle{ - file: openFile, - fd: -1, - } - if fdobj != nil { - h.fd = int32(fdobj.Release()) + file: openP9File, + fdLisa: d.fs.clientLisa.NewFD(openLisaFD), + fd: openHostFD, } fd, err := newSpecialFileFD(h, mnt, child, opts.Flags) if err != nil { @@ -1304,7 +1549,12 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // Update the remote filesystem. if !renamed.isSynthetic() { - if err := renamed.file.rename(ctx, newParent.file, newName); err != nil { + if fs.opts.lisaEnabled { + err = renamed.controlFDLisa.RenameTo(ctx, newParent.controlFDLisa.ID(), newName) + } else { + err = renamed.file.rename(ctx, newParent.file, newName) + } + if err != nil { vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) return err } @@ -1315,7 +1565,12 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if replaced.isDir() { flags = linux.AT_REMOVEDIR } - if err := newParent.file.unlinkAt(ctx, newName, flags); err != nil { + if fs.opts.lisaEnabled { + err = newParent.controlFDLisa.UnlinkAt(ctx, newName, flags) + } else { + err = newParent.file.unlinkAt(ctx, newName, flags) + } + if err != nil { vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) return err } @@ -1431,6 +1686,28 @@ func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu for d.isSynthetic() { d = d.parent } + if fs.opts.lisaEnabled { + var statFS lisafs.StatFS + if err := d.controlFDLisa.StatFSTo(ctx, &statFS); err != nil { + return linux.Statfs{}, err + } + if statFS.NameLength > maxFilenameLen { + statFS.NameLength = 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: statFS.BlockSize, + Blocks: statFS.Blocks, + BlocksFree: statFS.BlocksFree, + BlocksAvailable: statFS.BlocksAvailable, + Files: statFS.Files, + FilesFree: statFS.FilesFree, + NameLength: statFS.NameLength, + }, nil + } fsstat, err := d.file.statFS(ctx) if err != nil { return linux.Statfs{}, err @@ -1456,11 +1733,21 @@ func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu // 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, _ **[]*dentry) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string, ds **[]*dentry) (*lisafs.Inode, error) { creds := rp.Credentials() + if fs.opts.lisaEnabled { + return parent.controlFDLisa.SymlinkAt(ctx, name, target, lisafs.UID(creds.EffectiveKUID), lisafs.GID(creds.EffectiveKGID)) + } _, err := parent.file.symlink(ctx, target, name, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) - return err - }, nil) + return nil, err + }, nil, func(child *dentry) { + if fs.opts.interop != InteropModeShared { + // lisafs caches the symlink target on creation. In practice, this + // helps avoid a lot of ReadLink RPCs. + child.haveTarget = true + child.target = target + } + }) } // UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. @@ -1505,7 +1792,7 @@ func (fs *filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, si if err != nil { return nil, err } - return d.listXattr(ctx, rp.Credentials(), size) + return d.listXattr(ctx, size) } // GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index 13971d086..7bef8242f 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -48,6 +48,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/errors/linuxerr" "gvisor.dev/gvisor/pkg/hostarch" + "gvisor.dev/gvisor/pkg/lisafs" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/p9" refs_vfs1 "gvisor.dev/gvisor/pkg/refs" @@ -119,6 +120,10 @@ type filesystem struct { // client is the client used by this filesystem. client is immutable. client *p9.Client `state:"nosave"` + // clientLisa is the client used for communicating with the server when + // lisafs is enabled. lisafsCient is immutable. + clientLisa *lisafs.Client `state:"nosave"` + // clock is a realtime clock used to set timestamps in file operations. clock ktime.Clock @@ -162,6 +167,12 @@ type filesystem struct { inoMu sync.Mutex `state:"nosave"` inoByQIDPath map[uint64]uint64 `state:"nosave"` + // inoByKey is the same as inoByQIDPath but only used by lisafs. It helps + // identify inodes based on the device ID and host inode number provided + // by the gofer process. It is not preserved across checkpoint/restore for + // the same reason as above. inoByKey is protected by inoMu. + inoByKey map[inoKey]uint64 `state:"nosave"` + // lastIno is the last inode number assigned to a file. lastIno is accessed // using atomic memory operations. lastIno uint64 @@ -471,44 +482,83 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt syncableDentries: make(map[*dentry]struct{}), specialFileFDs: make(map[*specialFileFD]struct{}), inoByQIDPath: make(map[uint64]uint64), + inoByKey: make(map[inoKey]uint64), } fs.vfsfs.Init(vfsObj, &fstype, fs) + if err := fs.initClientAndRoot(ctx); err != nil { + fs.vfsfs.DecRef(ctx) + return nil, nil, err + } + + return &fs.vfsfs, &fs.root.vfsd, nil +} + +func (fs *filesystem) initClientAndRoot(ctx context.Context) error { + var err error + if fs.opts.lisaEnabled { + var rootInode *lisafs.Inode + rootInode, err = fs.initClientLisa(ctx) + if err != nil { + return err + } + fs.root, err = fs.newDentryLisa(ctx, rootInode) + if err != nil { + fs.clientLisa.CloseFDBatched(ctx, rootInode.ControlFD) + } + } else { + fs.root, err = fs.initClient(ctx) + } + + // Set the root's reference count to 2. One reference is returned to the + // caller, and the other is held by fs to prevent the root from being "cached" + // and subsequently evicted. + if err == nil { + fs.root.refs = 2 + } + return err +} + +func (fs *filesystem) initClientLisa(ctx context.Context) (*lisafs.Inode, error) { + sock, err := unet.NewSocket(fs.opts.fd) + if err != nil { + return nil, err + } + + var rootInode *lisafs.Inode + ctx.UninterruptibleSleepStart(false) + fs.clientLisa, rootInode, err = lisafs.NewClient(sock, fs.opts.aname) + ctx.UninterruptibleSleepFinish(false) + return rootInode, err +} + +func (fs *filesystem) initClient(ctx context.Context) (*dentry, error) { // Connect to the server. if err := fs.dial(ctx); err != nil { - return nil, nil, err + return nil, err } // Perform attach to obtain the filesystem root. ctx.UninterruptibleSleepStart(false) - attached, err := fs.client.Attach(fsopts.aname) + attached, err := fs.client.Attach(fs.opts.aname) ctx.UninterruptibleSleepFinish(false) if err != nil { - fs.vfsfs.DecRef(ctx) - return nil, nil, err + return nil, err } attachFile := p9file{attached} qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask()) if err != nil { attachFile.close(ctx) - fs.vfsfs.DecRef(ctx) - return nil, nil, err + return nil, err } // Construct the root dentry. root, err := fs.newDentry(ctx, attachFile, qid, attrMask, &attr) if err != nil { attachFile.close(ctx) - fs.vfsfs.DecRef(ctx) - return nil, nil, err + return nil, err } - // Set the root's reference count to 2. One reference is returned to the - // caller, and the other is held by fs to prevent the root from being "cached" - // and subsequently evicted. - root.refs = 2 - fs.root = root - - return &fs.vfsfs, &root.vfsd, nil + return root, nil } func getFDFromMountOptionsMap(ctx context.Context, mopts map[string]string) (int, error) { @@ -626,7 +676,11 @@ func (fs *filesystem) Release(ctx context.Context) { if !fs.iopts.LeakConnection { // Close the connection to the server. This implicitly clunks all fids. - fs.client.Close() + if fs.opts.lisaEnabled { + fs.clientLisa.Close() + } else { + fs.client.Close() + } } fs.vfsfs.VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) @@ -657,6 +711,23 @@ func (d *dentry) releaseSyntheticRecursiveLocked(ctx context.Context) { } } +// inoKey is the key used to identify the inode backed by this dentry. +// +// +stateify savable +type inoKey struct { + ino uint64 + devMinor uint32 + devMajor uint32 +} + +func inoKeyFromStat(stat *linux.Statx) inoKey { + return inoKey{ + ino: stat.Ino, + devMinor: stat.DevMinor, + devMajor: stat.DevMajor, + } +} + // dentry implements vfs.DentryImpl. // // +stateify savable @@ -687,6 +758,9 @@ type dentry struct { // qidPath is the p9.QID.Path for this file. qidPath is immutable. qidPath uint64 + // inoKey is used to identify this dentry's inode. + inoKey inoKey + // file is the unopened p9.File that backs this dentry. file is immutable. // // If file.isNil(), this dentry represents a synthetic file, i.e. a file @@ -694,6 +768,14 @@ type dentry struct { // only files that can be synthetic are sockets, pipes, and directories. file p9file `state:"nosave"` + // controlFDLisa is used by lisafs to perform path based operations on this + // dentry. + // + // if !controlFDLisa.Ok(), this dentry represents a synthetic file, i.e. a + // file that does not exist on the remote filesystem. As of this writing, the + // only files that can be synthetic are sockets, pipes, and directories. + controlFDLisa lisafs.ClientFD `state:"nosave"` + // If deleted is non-zero, the file represented by this dentry has been // deleted. deleted is accessed using atomic memory operations. deleted uint32 @@ -804,12 +886,14 @@ type dentry struct { // always either -1 or equal to readFD; if !writeFile.isNil() (the file has // been opened for writing), it is additionally either -1 or equal to // writeFD. - handleMu sync.RWMutex `state:"nosave"` - readFile p9file `state:"nosave"` - writeFile p9file `state:"nosave"` - readFD int32 `state:"nosave"` - writeFD int32 `state:"nosave"` - mmapFD int32 `state:"nosave"` + handleMu sync.RWMutex `state:"nosave"` + readFile p9file `state:"nosave"` + writeFile p9file `state:"nosave"` + readFDLisa lisafs.ClientFD `state:"nosave"` + writeFDLisa lisafs.ClientFD `state:"nosave"` + readFD int32 `state:"nosave"` + writeFD int32 `state:"nosave"` + mmapFD int32 `state:"nosave"` dataMu sync.RWMutex `state:"nosave"` @@ -933,6 +1017,79 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma return d, nil } +func (fs *filesystem) newDentryLisa(ctx context.Context, ino *lisafs.Inode) (*dentry, error) { + if ino.Stat.Mask&linux.STATX_TYPE == 0 { + ctx.Warningf("can't create gofer.dentry without file type") + return nil, linuxerr.EIO + } + if ino.Stat.Mode&linux.FileTypeMask == linux.ModeRegular && ino.Stat.Mask&linux.STATX_SIZE == 0 { + ctx.Warningf("can't create regular file gofer.dentry without file size") + return nil, linuxerr.EIO + } + + inoKey := inoKeyFromStat(&ino.Stat) + d := &dentry{ + fs: fs, + inoKey: inoKey, + ino: fs.inoFromKey(inoKey), + mode: uint32(ino.Stat.Mode), + uid: uint32(fs.opts.dfltuid), + gid: uint32(fs.opts.dfltgid), + blockSize: hostarch.PageSize, + readFD: -1, + writeFD: -1, + mmapFD: -1, + controlFDLisa: fs.clientLisa.NewFD(ino.ControlFD), + } + + d.pf.dentry = d + if ino.Stat.Mask&linux.STATX_UID != 0 { + d.uid = dentryUIDFromLisaUID(lisafs.UID(ino.Stat.UID)) + } + if ino.Stat.Mask&linux.STATX_GID != 0 { + d.gid = dentryGIDFromLisaGID(lisafs.GID(ino.Stat.GID)) + } + if ino.Stat.Mask&linux.STATX_SIZE != 0 { + d.size = ino.Stat.Size + } + if ino.Stat.Blksize != 0 { + d.blockSize = ino.Stat.Blksize + } + if ino.Stat.Mask&linux.STATX_ATIME != 0 { + d.atime = dentryTimestampFromLisa(ino.Stat.Atime) + } + if ino.Stat.Mask&linux.STATX_MTIME != 0 { + d.mtime = dentryTimestampFromLisa(ino.Stat.Mtime) + } + if ino.Stat.Mask&linux.STATX_CTIME != 0 { + d.ctime = dentryTimestampFromLisa(ino.Stat.Ctime) + } + if ino.Stat.Mask&linux.STATX_BTIME != 0 { + d.btime = dentryTimestampFromLisa(ino.Stat.Btime) + } + if ino.Stat.Mask&linux.STATX_NLINK != 0 { + d.nlink = ino.Stat.Nlink + } + d.vfsd.Init(d) + refsvfs2.Register(d) + fs.syncMu.Lock() + fs.syncableDentries[d] = struct{}{} + fs.syncMu.Unlock() + return d, nil +} + +func (fs *filesystem) inoFromKey(key inoKey) uint64 { + fs.inoMu.Lock() + defer fs.inoMu.Unlock() + + if ino, ok := fs.inoByKey[key]; ok { + return ino + } + ino := fs.nextIno() + fs.inoByKey[key] = ino + return ino +} + func (fs *filesystem) inoFromQIDPath(qidPath uint64) uint64 { fs.inoMu.Lock() defer fs.inoMu.Unlock() @@ -949,7 +1106,7 @@ func (fs *filesystem) nextIno() uint64 { } func (d *dentry) isSynthetic() bool { - return d.file.isNil() + return !d.isControlFileOk() } func (d *dentry) cachedMetadataAuthoritative() bool { @@ -999,6 +1156,50 @@ func (d *dentry) updateFromP9AttrsLocked(mask p9.AttrMask, attr *p9.Attr) { } } +// updateFromLisaStatLocked is called to update d's metadata after an update +// from the remote filesystem. +// Precondition: d.metadataMu must be locked. +// +checklocks:d.metadataMu +func (d *dentry) updateFromLisaStatLocked(stat *linux.Statx) { + if stat.Mask&linux.STATX_TYPE != 0 { + if got, want := stat.Mode&linux.FileTypeMask, d.fileType(); uint32(got) != want { + panic(fmt.Sprintf("gofer.dentry file type changed from %#o to %#o", want, got)) + } + } + if stat.Mask&linux.STATX_MODE != 0 { + atomic.StoreUint32(&d.mode, uint32(stat.Mode)) + } + if stat.Mask&linux.STATX_UID != 0 { + atomic.StoreUint32(&d.uid, dentryUIDFromLisaUID(lisafs.UID(stat.UID))) + } + if stat.Mask&linux.STATX_GID != 0 { + atomic.StoreUint32(&d.uid, dentryGIDFromLisaGID(lisafs.GID(stat.GID))) + } + if stat.Blksize != 0 { + atomic.StoreUint32(&d.blockSize, stat.Blksize) + } + // Don't override newer client-defined timestamps with old server-defined + // ones. + if stat.Mask&linux.STATX_ATIME != 0 && atomic.LoadUint32(&d.atimeDirty) == 0 { + atomic.StoreInt64(&d.atime, dentryTimestampFromLisa(stat.Atime)) + } + if stat.Mask&linux.STATX_MTIME != 0 && atomic.LoadUint32(&d.mtimeDirty) == 0 { + atomic.StoreInt64(&d.mtime, dentryTimestampFromLisa(stat.Mtime)) + } + if stat.Mask&linux.STATX_CTIME != 0 { + atomic.StoreInt64(&d.ctime, dentryTimestampFromLisa(stat.Ctime)) + } + if stat.Mask&linux.STATX_BTIME != 0 { + atomic.StoreInt64(&d.btime, dentryTimestampFromLisa(stat.Btime)) + } + if stat.Mask&linux.STATX_NLINK != 0 { + atomic.StoreUint32(&d.nlink, stat.Nlink) + } + if stat.Mask&linux.STATX_SIZE != 0 { + d.updateSizeLocked(stat.Size) + } +} + // Preconditions: !d.isSynthetic(). // Preconditions: d.metadataMu is locked. // +checklocks:d.metadataMu @@ -1008,6 +1209,9 @@ func (d *dentry) refreshSizeLocked(ctx context.Context) error { if d.writeFD < 0 { d.handleMu.RUnlock() // Ask the gofer if we don't have a host FD. + if d.fs.opts.lisaEnabled { + return d.updateFromStatLisaLocked(ctx, nil) + } return d.updateFromGetattrLocked(ctx, p9file{}) } @@ -1027,6 +1231,9 @@ func (d *dentry) updateFromGetattr(ctx context.Context) error { // updating stale attributes in d.updateFromP9AttrsLocked(). d.metadataMu.Lock() defer d.metadataMu.Unlock() + if d.fs.opts.lisaEnabled { + return d.updateFromStatLisaLocked(ctx, nil) + } return d.updateFromGetattrLocked(ctx, p9file{}) } @@ -1034,6 +1241,45 @@ func (d *dentry) updateFromGetattr(ctx context.Context) error { // * !d.isSynthetic(). // * d.metadataMu is locked. // +checklocks:d.metadataMu +func (d *dentry) updateFromStatLisaLocked(ctx context.Context, fdLisa *lisafs.ClientFD) error { + handleMuRLocked := false + if fdLisa == nil { + // Use open FDs in preferenece to the control FD. This may be significantly + // more efficient in some implementations. Prefer a writable FD over a + // readable one since some filesystem implementations may update a writable + // FD's metadata after writes, without making metadata updates immediately + // visible to read-only FDs representing the same file. + d.handleMu.RLock() + switch { + case d.writeFDLisa.Ok(): + fdLisa = &d.writeFDLisa + handleMuRLocked = true + case d.readFDLisa.Ok(): + fdLisa = &d.readFDLisa + handleMuRLocked = true + default: + fdLisa = &d.controlFDLisa + d.handleMu.RUnlock() + } + } + + var stat linux.Statx + err := fdLisa.StatTo(ctx, &stat) + if handleMuRLocked { + // handleMu must be released before updateFromLisaStatLocked(). + d.handleMu.RUnlock() // +checklocksforce: complex case. + } + if err != nil { + return err + } + d.updateFromLisaStatLocked(&stat) + return nil +} + +// Preconditions: +// * !d.isSynthetic(). +// * d.metadataMu is locked. +// +checklocks:d.metadataMu func (d *dentry) updateFromGetattrLocked(ctx context.Context, file p9file) error { handleMuRLocked := false if file.isNil() { @@ -1173,6 +1419,13 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs } } + // failureMask indicates which attributes could not be set on the remote + // filesystem. p9 returns an error if any of the attributes could not be set + // but that leads to inconsistency as the server could have set a few + // attributes successfully but a later failure will cause the successful ones + // to not be updated in the dentry cache. + var failureMask uint32 + var failureErr error if !d.isSynthetic() { if stat.Mask != 0 { if stat.Mask&linux.STATX_SIZE != 0 { @@ -1182,35 +1435,50 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs // the remote file has been truncated). d.dataMu.Lock() } - 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.Mask&linux.STATX_ATIME != 0 && stat.Atime.Nsec != linux.UTIME_NOW, - MTimeNotSystemTime: stat.Mask&linux.STATX_MTIME != 0 && 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 { - if stat.Mask&linux.STATX_SIZE != 0 { - d.dataMu.Unlock() // +checklocksforce: locked conditionally above + if d.fs.opts.lisaEnabled { + var err error + failureMask, failureErr, err = d.controlFDLisa.SetStat(ctx, stat) + if err != nil { + if stat.Mask&linux.STATX_SIZE != 0 { + d.dataMu.Unlock() // +checklocksforce: locked conditionally above + } + return err + } + } else { + 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.Mask&linux.STATX_ATIME != 0 && stat.Atime.Nsec != linux.UTIME_NOW, + MTimeNotSystemTime: stat.Mask&linux.STATX_MTIME != 0 && 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 { + if stat.Mask&linux.STATX_SIZE != 0 { + d.dataMu.Unlock() // +checklocksforce: locked conditionally above + } + return err } - return err } if stat.Mask&linux.STATX_SIZE != 0 { - // d.size should be kept up to date, and privatized - // copy-on-write mappings of truncated pages need to be - // invalidated, even if InteropModeShared is in effect. - d.updateSizeAndUnlockDataMuLocked(stat.Size) // +checklocksforce: locked conditionally above + if failureMask&linux.STATX_SIZE == 0 { + // d.size should be kept up to date, and privatized + // copy-on-write mappings of truncated pages need to be + // invalidated, even if InteropModeShared is in effect. + d.updateSizeAndUnlockDataMuLocked(stat.Size) // +checklocksforce: locked conditionally above + } else { + d.dataMu.Unlock() // +checklocksforce: locked conditionally above + } } } if d.fs.opts.interop == InteropModeShared { @@ -1221,13 +1489,13 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs return nil } } - if stat.Mask&linux.STATX_MODE != 0 { + if stat.Mask&linux.STATX_MODE != 0 && failureMask&linux.STATX_MODE == 0 { atomic.StoreUint32(&d.mode, d.fileType()|uint32(stat.Mode)) } - if stat.Mask&linux.STATX_UID != 0 { + if stat.Mask&linux.STATX_UID != 0 && failureMask&linux.STATX_UID == 0 { atomic.StoreUint32(&d.uid, stat.UID) } - if stat.Mask&linux.STATX_GID != 0 { + if stat.Mask&linux.STATX_GID != 0 && failureMask&linux.STATX_GID == 0 { atomic.StoreUint32(&d.gid, stat.GID) } // Note that stat.Atime.Nsec and stat.Mtime.Nsec can't be UTIME_NOW because @@ -1235,15 +1503,19 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs // stat.Mtime to client-local timestamps above, and if // !d.cachedMetadataAuthoritative() then we returned after calling // d.file.setAttr(). For the same reason, now must have been initialized. - if stat.Mask&linux.STATX_ATIME != 0 { + if stat.Mask&linux.STATX_ATIME != 0 && failureMask&linux.STATX_ATIME == 0 { atomic.StoreInt64(&d.atime, stat.Atime.ToNsec()) atomic.StoreUint32(&d.atimeDirty, 0) } - if stat.Mask&linux.STATX_MTIME != 0 { + if stat.Mask&linux.STATX_MTIME != 0 && failureMask&linux.STATX_MTIME == 0 { atomic.StoreInt64(&d.mtime, stat.Mtime.ToNsec()) atomic.StoreUint32(&d.mtimeDirty, 0) } atomic.StoreInt64(&d.ctime, now) + if failureMask != 0 { + // Setting some attribute failed on the remote filesystem. + return failureErr + } return nil } @@ -1359,6 +1631,20 @@ func dentryGIDFromP9GID(gid p9.GID) uint32 { return uint32(gid) } +func dentryUIDFromLisaUID(uid lisafs.UID) uint32 { + if !uid.Ok() { + return uint32(auth.OverflowUID) + } + return uint32(uid) +} + +func dentryGIDFromLisaGID(gid lisafs.GID) uint32 { + if !gid.Ok() { + return uint32(auth.OverflowGID) + } + return uint32(gid) +} + // IncRef implements vfs.DentryImpl.IncRef. func (d *dentry) IncRef() { // d.refs may be 0 if d.fs.renameMu is locked, which serializes against @@ -1667,15 +1953,24 @@ func (d *dentry) destroyLocked(ctx context.Context) { d.dirty.RemoveAll() } d.dataMu.Unlock() - // Clunk open fids and close open host FDs. - if !d.readFile.isNil() { - _ = d.readFile.close(ctx) - } - if !d.writeFile.isNil() && d.readFile != d.writeFile { - _ = d.writeFile.close(ctx) + if d.fs.opts.lisaEnabled { + if d.readFDLisa.Ok() && d.readFDLisa.ID() != d.writeFDLisa.ID() { + d.readFDLisa.CloseBatched(ctx) + } + if d.writeFDLisa.Ok() { + d.writeFDLisa.CloseBatched(ctx) + } + } else { + // Clunk open fids and close open host FDs. + if !d.readFile.isNil() { + _ = d.readFile.close(ctx) + } + if !d.writeFile.isNil() && d.readFile != d.writeFile { + _ = d.writeFile.close(ctx) + } + d.readFile = p9file{} + d.writeFile = p9file{} } - d.readFile = p9file{} - d.writeFile = p9file{} if d.readFD >= 0 { _ = unix.Close(int(d.readFD)) } @@ -1687,7 +1982,7 @@ func (d *dentry) destroyLocked(ctx context.Context) { d.mmapFD = -1 d.handleMu.Unlock() - if !d.file.isNil() { + if d.isControlFileOk() { // Note that it's possible that d.atimeDirty or d.mtimeDirty are true, // i.e. client and server timestamps may differ (because e.g. a client // write was serviced by the page cache, and only written back to the @@ -1696,10 +1991,16 @@ func (d *dentry) destroyLocked(ctx context.Context) { // instantiated for the same file would remain coherent. Unfortunately, // this turns out to be too expensive in many cases, so for now we // don't do this. - if err := d.file.close(ctx); err != nil { - log.Warningf("gofer.dentry.destroyLocked: failed to close file: %v", err) + + // Close the control FD. + if d.fs.opts.lisaEnabled { + d.controlFDLisa.CloseBatched(ctx) + } else { + if err := d.file.close(ctx); err != nil { + log.Warningf("gofer.dentry.destroyLocked: failed to close file: %v", err) + } + d.file = p9file{} } - d.file = p9file{} // Remove d from the set of syncable dentries. d.fs.syncMu.Lock() @@ -1725,10 +2026,38 @@ func (d *dentry) setDeleted() { atomic.StoreUint32(&d.deleted, 1) } -func (d *dentry) listXattr(ctx context.Context, creds *auth.Credentials, size uint64) ([]string, error) { - if d.file.isNil() { +func (d *dentry) isControlFileOk() bool { + if d.fs.opts.lisaEnabled { + return d.controlFDLisa.Ok() + } + return !d.file.isNil() +} + +func (d *dentry) isReadFileOk() bool { + if d.fs.opts.lisaEnabled { + return d.readFDLisa.Ok() + } + return !d.readFile.isNil() +} + +func (d *dentry) listXattr(ctx context.Context, size uint64) ([]string, error) { + if !d.isControlFileOk() { return nil, nil } + + if d.fs.opts.lisaEnabled { + xattrs, err := d.controlFDLisa.ListXattr(ctx, size) + if err != nil { + return nil, err + } + + res := make([]string, 0, len(xattrs)) + for _, xattr := range xattrs { + res = append(res, xattr) + } + return res, nil + } + xattrMap, err := d.file.listXattr(ctx, size) if err != nil { return nil, err @@ -1741,32 +2070,41 @@ func (d *dentry) listXattr(ctx context.Context, creds *auth.Credentials, size ui } func (d *dentry) getXattr(ctx context.Context, creds *auth.Credentials, opts *vfs.GetXattrOptions) (string, error) { - if d.file.isNil() { + if !d.isControlFileOk() { return "", linuxerr.ENODATA } if err := d.checkXattrPermissions(creds, opts.Name, vfs.MayRead); err != nil { return "", err } + if d.fs.opts.lisaEnabled { + return d.controlFDLisa.GetXattr(ctx, opts.Name, opts.Size) + } return d.file.getXattr(ctx, opts.Name, opts.Size) } func (d *dentry) setXattr(ctx context.Context, creds *auth.Credentials, opts *vfs.SetXattrOptions) error { - if d.file.isNil() { + if !d.isControlFileOk() { return linuxerr.EPERM } if err := d.checkXattrPermissions(creds, opts.Name, vfs.MayWrite); err != nil { return err } + if d.fs.opts.lisaEnabled { + return d.controlFDLisa.SetXattr(ctx, opts.Name, opts.Value, opts.Flags) + } return d.file.setXattr(ctx, opts.Name, opts.Value, opts.Flags) } func (d *dentry) removeXattr(ctx context.Context, creds *auth.Credentials, name string) error { - if d.file.isNil() { + if !d.isControlFileOk() { return linuxerr.EPERM } if err := d.checkXattrPermissions(creds, name, vfs.MayWrite); err != nil { return err } + if d.fs.opts.lisaEnabled { + return d.controlFDLisa.RemoveXattr(ctx, name) + } return d.file.removeXattr(ctx, name) } @@ -1778,19 +2116,30 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool // O_TRUNC). if !trunc { d.handleMu.RLock() - if (!read || !d.readFile.isNil()) && (!write || !d.writeFile.isNil()) { + var canReuseCurHandle bool + if d.fs.opts.lisaEnabled { + canReuseCurHandle = (!read || d.readFDLisa.Ok()) && (!write || d.writeFDLisa.Ok()) + } else { + canReuseCurHandle = (!read || !d.readFile.isNil()) && (!write || !d.writeFile.isNil()) + } + d.handleMu.RUnlock() + if canReuseCurHandle { // Current handles are sufficient. - d.handleMu.RUnlock() return nil } - d.handleMu.RUnlock() } var fdsToCloseArr [2]int32 fdsToClose := fdsToCloseArr[:0] invalidateTranslations := false d.handleMu.Lock() - if (read && d.readFile.isNil()) || (write && d.writeFile.isNil()) || trunc { + var needNewHandle bool + if d.fs.opts.lisaEnabled { + needNewHandle = (read && !d.readFDLisa.Ok()) || (write && !d.writeFDLisa.Ok()) || trunc + } else { + needNewHandle = (read && d.readFile.isNil()) || (write && d.writeFile.isNil()) || trunc + } + if needNewHandle { // Get a new handle. If this file has been opened for both reading and // writing, try to get a single handle that is usable for both: // @@ -1799,9 +2148,21 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool // // - NOTE(b/141991141): Some filesystems may not ensure coherence // between multiple handles for the same file. - openReadable := !d.readFile.isNil() || read - openWritable := !d.writeFile.isNil() || write - h, err := openHandle(ctx, d.file, openReadable, openWritable, trunc) + var ( + openReadable bool + openWritable bool + h handle + err error + ) + if d.fs.opts.lisaEnabled { + openReadable = d.readFDLisa.Ok() || read + openWritable = d.writeFDLisa.Ok() || write + h, err = openHandleLisa(ctx, d.controlFDLisa, openReadable, openWritable, trunc) + } else { + openReadable = !d.readFile.isNil() || read + openWritable = !d.writeFile.isNil() || write + h, err = openHandle(ctx, d.file, openReadable, openWritable, trunc) + } if linuxerr.Equals(linuxerr.EACCES, err) && (openReadable != read || openWritable != write) { // It may not be possible to use a single handle for both // reading and writing, since permissions on the file may have @@ -1811,7 +2172,11 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool ctx.Debugf("gofer.dentry.ensureSharedHandle: bifurcating read/write handles for dentry %p", d) openReadable = read openWritable = write - h, err = openHandle(ctx, d.file, openReadable, openWritable, trunc) + if d.fs.opts.lisaEnabled { + h, err = openHandleLisa(ctx, d.controlFDLisa, openReadable, openWritable, trunc) + } else { + h, err = openHandle(ctx, d.file, openReadable, openWritable, trunc) + } } if err != nil { d.handleMu.Unlock() @@ -1873,9 +2238,16 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool // previously opened for reading (without an FD), then existing // translations of the file may use the internal page cache; // invalidate those mappings. - if d.writeFile.isNil() { - invalidateTranslations = !d.readFile.isNil() - atomic.StoreInt32(&d.mmapFD, h.fd) + if d.fs.opts.lisaEnabled { + if !d.writeFDLisa.Ok() { + invalidateTranslations = d.readFDLisa.Ok() + atomic.StoreInt32(&d.mmapFD, h.fd) + } + } else { + if d.writeFile.isNil() { + invalidateTranslations = !d.readFile.isNil() + atomic.StoreInt32(&d.mmapFD, h.fd) + } } } else if openWritable && d.writeFD < 0 { atomic.StoreInt32(&d.writeFD, h.fd) @@ -1902,24 +2274,45 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool atomic.StoreInt32(&d.mmapFD, -1) } - // Switch to new fids. - var oldReadFile p9file - if openReadable { - oldReadFile = d.readFile - d.readFile = h.file - } - var oldWriteFile p9file - if openWritable { - oldWriteFile = d.writeFile - d.writeFile = h.file - } - // NOTE(b/141991141): Clunk old fids before making new fids visible (by - // unlocking d.handleMu). - if !oldReadFile.isNil() { - oldReadFile.close(ctx) - } - if !oldWriteFile.isNil() && oldReadFile != oldWriteFile { - oldWriteFile.close(ctx) + // Switch to new fids/FDs. + if d.fs.opts.lisaEnabled { + oldReadFD := lisafs.InvalidFDID + if openReadable { + oldReadFD = d.readFDLisa.ID() + d.readFDLisa = h.fdLisa + } + oldWriteFD := lisafs.InvalidFDID + if openWritable { + oldWriteFD = d.writeFDLisa.ID() + d.writeFDLisa = h.fdLisa + } + // NOTE(b/141991141): Close old FDs before making new fids visible (by + // unlocking d.handleMu). + if oldReadFD.Ok() { + d.fs.clientLisa.CloseFDBatched(ctx, oldReadFD) + } + if oldWriteFD.Ok() && oldReadFD != oldWriteFD { + d.fs.clientLisa.CloseFDBatched(ctx, oldWriteFD) + } + } else { + var oldReadFile p9file + if openReadable { + oldReadFile = d.readFile + d.readFile = h.file + } + var oldWriteFile p9file + if openWritable { + oldWriteFile = d.writeFile + d.writeFile = h.file + } + // NOTE(b/141991141): Clunk old fids before making new fids visible (by + // unlocking d.handleMu). + if !oldReadFile.isNil() { + oldReadFile.close(ctx) + } + if !oldWriteFile.isNil() && oldReadFile != oldWriteFile { + oldWriteFile.close(ctx) + } } } d.handleMu.Unlock() @@ -1943,27 +2336,29 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool // Preconditions: d.handleMu must be locked. func (d *dentry) readHandleLocked() handle { return handle{ - file: d.readFile, - fd: d.readFD, + fdLisa: d.readFDLisa, + file: d.readFile, + fd: d.readFD, } } // Preconditions: d.handleMu must be locked. func (d *dentry) writeHandleLocked() handle { return handle{ - file: d.writeFile, - fd: d.writeFD, + fdLisa: d.writeFDLisa, + file: d.writeFile, + fd: d.writeFD, } } func (d *dentry) syncRemoteFile(ctx context.Context) error { d.handleMu.RLock() defer d.handleMu.RUnlock() - return d.syncRemoteFileLocked(ctx) + return d.syncRemoteFileLocked(ctx, nil /* accFsyncFDIDsLisa */) } // Preconditions: d.handleMu must be locked. -func (d *dentry) syncRemoteFileLocked(ctx context.Context) error { +func (d *dentry) syncRemoteFileLocked(ctx context.Context, accFsyncFDIDsLisa *[]lisafs.FDID) error { // If we have a host FD, fsyncing it is likely to be faster than an fsync // RPC. Prefer syncing write handles over read handles, since some remote // filesystem implementations may not sync changes made through write @@ -1974,7 +2369,13 @@ func (d *dentry) syncRemoteFileLocked(ctx context.Context) error { ctx.UninterruptibleSleepFinish(false) return err } - if !d.writeFile.isNil() { + if d.fs.opts.lisaEnabled && d.writeFDLisa.Ok() { + if accFsyncFDIDsLisa != nil { + *accFsyncFDIDsLisa = append(*accFsyncFDIDsLisa, d.writeFDLisa.ID()) + return nil + } + return d.writeFDLisa.Sync(ctx) + } else if !d.fs.opts.lisaEnabled && !d.writeFile.isNil() { return d.writeFile.fsync(ctx) } if d.readFD >= 0 { @@ -1983,13 +2384,19 @@ func (d *dentry) syncRemoteFileLocked(ctx context.Context) error { ctx.UninterruptibleSleepFinish(false) return err } - if !d.readFile.isNil() { + if d.fs.opts.lisaEnabled && d.readFDLisa.Ok() { + if accFsyncFDIDsLisa != nil { + *accFsyncFDIDsLisa = append(*accFsyncFDIDsLisa, d.readFDLisa.ID()) + return nil + } + return d.readFDLisa.Sync(ctx) + } else if !d.fs.opts.lisaEnabled && !d.readFile.isNil() { return d.readFile.fsync(ctx) } return nil } -func (d *dentry) syncCachedFile(ctx context.Context, forFilesystemSync bool) error { +func (d *dentry) syncCachedFile(ctx context.Context, forFilesystemSync bool, accFsyncFDIDsLisa *[]lisafs.FDID) error { d.handleMu.RLock() defer d.handleMu.RUnlock() h := d.writeHandleLocked() @@ -2002,7 +2409,7 @@ func (d *dentry) syncCachedFile(ctx context.Context, forFilesystemSync bool) err return err } } - if err := d.syncRemoteFileLocked(ctx); err != nil { + if err := d.syncRemoteFileLocked(ctx, accFsyncFDIDsLisa); err != nil { if !forFilesystemSync { return err } @@ -2059,18 +2466,33 @@ func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linu d := fd.dentry() const validMask = uint32(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) if !d.cachedMetadataAuthoritative() && opts.Mask&validMask != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC { - // Use specialFileFD.handle.file for the getattr if available, for the - // same reason that we try to use open file handles in - // dentry.updateFromGetattrLocked(). - var file p9file - if sffd, ok := fd.vfsfd.Impl().(*specialFileFD); ok { - file = sffd.handle.file - } - d.metadataMu.Lock() - err := d.updateFromGetattrLocked(ctx, file) - d.metadataMu.Unlock() - if err != nil { - return linux.Statx{}, err + if d.fs.opts.lisaEnabled { + // Use specialFileFD.handle.fileLisa for the Stat if available, for the + // same reason that we try to use open FD in updateFromStatLisaLocked(). + var fdLisa *lisafs.ClientFD + if sffd, ok := fd.vfsfd.Impl().(*specialFileFD); ok { + fdLisa = &sffd.handle.fdLisa + } + d.metadataMu.Lock() + err := d.updateFromStatLisaLocked(ctx, fdLisa) + d.metadataMu.Unlock() + if err != nil { + return linux.Statx{}, err + } + } else { + // Use specialFileFD.handle.file for the getattr if available, for the + // same reason that we try to use open file handles in + // dentry.updateFromGetattrLocked(). + var file p9file + if sffd, ok := fd.vfsfd.Impl().(*specialFileFD); ok { + file = sffd.handle.file + } + d.metadataMu.Lock() + err := d.updateFromGetattrLocked(ctx, file) + d.metadataMu.Unlock() + if err != nil { + return linux.Statx{}, err + } } } var stat linux.Statx @@ -2091,7 +2513,7 @@ func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) // ListXattr implements vfs.FileDescriptionImpl.ListXattr. func (fd *fileDescription) ListXattr(ctx context.Context, size uint64) ([]string, error) { - return fd.dentry().listXattr(ctx, auth.CredentialsFromContext(ctx), size) + return fd.dentry().listXattr(ctx, size) } // GetXattr implements vfs.FileDescriptionImpl.GetXattr. diff --git a/pkg/sentry/fsimpl/gofer/gofer_test.go b/pkg/sentry/fsimpl/gofer/gofer_test.go index 806392d50..d5cc73f33 100644 --- a/pkg/sentry/fsimpl/gofer/gofer_test.go +++ b/pkg/sentry/fsimpl/gofer/gofer_test.go @@ -33,6 +33,7 @@ func TestDestroyIdempotent(t *testing.T) { }, syncableDentries: make(map[*dentry]struct{}), inoByQIDPath: make(map[uint64]uint64), + inoByKey: make(map[inoKey]uint64), } attr := &p9.Attr{ diff --git a/pkg/sentry/fsimpl/gofer/handle.go b/pkg/sentry/fsimpl/gofer/handle.go index 02540a754..394aecd62 100644 --- a/pkg/sentry/fsimpl/gofer/handle.go +++ b/pkg/sentry/fsimpl/gofer/handle.go @@ -17,6 +17,7 @@ package gofer import ( "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/lisafs" "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/hostfd" @@ -26,10 +27,13 @@ import ( // handle represents a remote "open file descriptor", consisting of an opened // fid (p9.File) and optionally a host file descriptor. // +// If lisafs is being used, fdLisa points to an open file on the server. +// // These are explicitly not savable. type handle struct { - file p9file - fd int32 // -1 if unavailable + fdLisa lisafs.ClientFD + file p9file + fd int32 // -1 if unavailable } // Preconditions: read || write. @@ -65,13 +69,47 @@ func openHandle(ctx context.Context, file p9file, read, write, trunc bool) (hand }, nil } +// Preconditions: read || write. +func openHandleLisa(ctx context.Context, fdLisa lisafs.ClientFD, read, write, trunc bool) (handle, error) { + var flags uint32 + switch { + case read && write: + flags = unix.O_RDWR + case read: + flags = unix.O_RDONLY + case write: + flags = unix.O_WRONLY + default: + panic("tried to open unreadable and unwritable handle") + } + if trunc { + flags |= unix.O_TRUNC + } + openFD, hostFD, err := fdLisa.OpenAt(ctx, flags) + if err != nil { + return handle{fd: -1}, err + } + h := handle{ + fdLisa: fdLisa.Client().NewFD(openFD), + fd: int32(hostFD), + } + return h, nil +} + func (h *handle) isOpen() bool { + if h.fdLisa.Client() != nil { + return h.fdLisa.Ok() + } return !h.file.isNil() } func (h *handle) close(ctx context.Context) { - h.file.close(ctx) - h.file = p9file{} + if h.fdLisa.Client() != nil { + h.fdLisa.CloseBatched(ctx) + } else { + h.file.close(ctx) + h.file = p9file{} + } if h.fd >= 0 { unix.Close(int(h.fd)) h.fd = -1 @@ -89,19 +127,27 @@ func (h *handle) readToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offs return n, err } if dsts.NumBlocks() == 1 && !dsts.Head().NeedSafecopy() { - n, err := h.file.readAt(ctx, dsts.Head().ToSlice(), offset) - return uint64(n), err + if h.fdLisa.Client() != nil { + return h.fdLisa.Read(ctx, dsts.Head().ToSlice(), offset) + } + return h.file.readAt(ctx, dsts.Head().ToSlice(), offset) } // Buffer the read since p9.File.ReadAt() takes []byte. buf := make([]byte, dsts.NumBytes()) - n, err := h.file.readAt(ctx, buf, offset) + var n uint64 + var err error + if h.fdLisa.Client() != nil { + n, err = h.fdLisa.Read(ctx, buf, offset) + } else { + 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 + return n, err } func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) { @@ -115,8 +161,10 @@ func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, o return n, err } if srcs.NumBlocks() == 1 && !srcs.Head().NeedSafecopy() { - n, err := h.file.writeAt(ctx, srcs.Head().ToSlice(), offset) - return uint64(n), err + if h.fdLisa.Client() != nil { + return h.fdLisa.Write(ctx, srcs.Head().ToSlice(), offset) + } + return h.file.writeAt(ctx, srcs.Head().ToSlice(), offset) } // Buffer the write since p9.File.WriteAt() takes []byte. buf := make([]byte, srcs.NumBytes()) @@ -124,12 +172,18 @@ func (h *handle) writeFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, o if cp == 0 { return 0, cperr } - n, err := h.file.writeAt(ctx, buf[:cp], offset) + var n uint64 + var err error + if h.fdLisa.Client() != nil { + n, err = h.fdLisa.Write(ctx, buf[:cp], offset) + } else { + n, err = h.file.writeAt(ctx, buf[:cp], offset) + } // err takes precedence over cperr. if err != nil { - return uint64(n), err + return n, err } - return uint64(n), cperr + return n, cperr } type handleReadWriter struct { diff --git a/pkg/sentry/fsimpl/gofer/p9file.go b/pkg/sentry/fsimpl/gofer/p9file.go index 5a3ddfc9d..0d97b60fd 100644 --- a/pkg/sentry/fsimpl/gofer/p9file.go +++ b/pkg/sentry/fsimpl/gofer/p9file.go @@ -141,18 +141,18 @@ func (f p9file) open(ctx context.Context, flags p9.OpenFlags) (*fd.FD, p9.QID, u return fdobj, qid, iounit, err } -func (f p9file) readAt(ctx context.Context, p []byte, offset uint64) (int, error) { +func (f p9file) readAt(ctx context.Context, p []byte, offset uint64) (uint64, error) { ctx.UninterruptibleSleepStart(false) n, err := f.file.ReadAt(p, offset) ctx.UninterruptibleSleepFinish(false) - return n, err + return uint64(n), err } -func (f p9file) writeAt(ctx context.Context, p []byte, offset uint64) (int, error) { +func (f p9file) writeAt(ctx context.Context, p []byte, offset uint64) (uint64, error) { ctx.UninterruptibleSleepStart(false) n, err := f.file.WriteAt(p, offset) ctx.UninterruptibleSleepFinish(false) - return n, err + return uint64(n), err } func (f p9file) fsync(ctx context.Context) error { diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go index 947dbe05f..874f9873d 100644 --- a/pkg/sentry/fsimpl/gofer/regular_file.go +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -98,6 +98,12 @@ func (fd *regularFileFD) OnClose(ctx context.Context) error { } d.handleMu.RLock() defer d.handleMu.RUnlock() + if d.fs.opts.lisaEnabled { + if !d.writeFDLisa.Ok() { + return nil + } + return d.writeFDLisa.Flush(ctx) + } if d.writeFile.isNil() { return nil } @@ -110,6 +116,9 @@ func (fd *regularFileFD) Allocate(ctx context.Context, mode, offset, length uint return d.doAllocate(ctx, offset, length, func() error { d.handleMu.RLock() defer d.handleMu.RUnlock() + if d.fs.opts.lisaEnabled { + return d.writeFDLisa.Allocate(ctx, mode, offset, length) + } return d.writeFile.allocate(ctx, p9.ToAllocateMode(mode), offset, length) }) } @@ -282,8 +291,19 @@ func (fd *regularFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off // changes to the host. if newMode := vfs.ClearSUIDAndSGID(oldMode); newMode != oldMode { atomic.StoreUint32(&d.mode, newMode) - if err := d.file.setAttr(ctx, p9.SetAttrMask{Permissions: true}, p9.SetAttr{Permissions: p9.FileMode(newMode)}); err != nil { - return 0, offset, err + if d.fs.opts.lisaEnabled { + stat := linux.Statx{Mask: linux.STATX_MODE, Mode: uint16(newMode)} + failureMask, failureErr, err := d.controlFDLisa.SetStat(ctx, &stat) + if err != nil { + return 0, offset, err + } + if failureMask != 0 { + return 0, offset, failureErr + } + } else { + if err := d.file.setAttr(ctx, p9.SetAttrMask{Permissions: true}, p9.SetAttr{Permissions: p9.FileMode(newMode)}); err != nil { + return 0, offset, err + } } } } @@ -677,7 +697,7 @@ func regularFileSeekLocked(ctx context.Context, d *dentry, fdOffset, offset int6 // Sync implements vfs.FileDescriptionImpl.Sync. func (fd *regularFileFD) Sync(ctx context.Context) error { - return fd.dentry().syncCachedFile(ctx, false /* lowSyncExpectations */) + return fd.dentry().syncCachedFile(ctx, false /* forFilesystemSync */, nil /* accFsyncFDIDsLisa */) } // ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. diff --git a/pkg/sentry/fsimpl/gofer/revalidate.go b/pkg/sentry/fsimpl/gofer/revalidate.go index 226790a11..5d4009832 100644 --- a/pkg/sentry/fsimpl/gofer/revalidate.go +++ b/pkg/sentry/fsimpl/gofer/revalidate.go @@ -15,7 +15,9 @@ package gofer import ( + "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/sync" ) @@ -234,28 +236,54 @@ func (fs *filesystem) revalidateHelper(ctx context.Context, vfsObj *vfs.VirtualF } // Lock metadata on all dentries *before* getting attributes for them. state.lockAllMetadata() - stats, err := state.start.file.multiGetAttr(ctx, state.names) - if err != nil { - return err + + var ( + stats []p9.FullStat + statsLisa []linux.Statx + numStats int + ) + if fs.opts.lisaEnabled { + var err error + statsLisa, err = state.start.controlFDLisa.WalkStat(ctx, state.names) + if err != nil { + return err + } + numStats = len(statsLisa) + } else { + var err error + stats, err = state.start.file.multiGetAttr(ctx, state.names) + if err != nil { + return err + } + numStats = len(stats) } i := -1 for d := state.popFront(); d != nil; d = state.popFront() { i++ - found := i < len(stats) + found := i < numStats if i == 0 && len(state.names[0]) == 0 { if found && !d.isSynthetic() { // First dentry is where the search is starting, just update attributes // since it cannot be replaced. - d.updateFromP9AttrsLocked(stats[i].Valid, &stats[i].Attr) // +checklocksforce: acquired by lockAllMetadata. + if fs.opts.lisaEnabled { + d.updateFromLisaStatLocked(&statsLisa[i]) // +checklocksforce: acquired by lockAllMetadata. + } else { + d.updateFromP9AttrsLocked(stats[i].Valid, &stats[i].Attr) // +checklocksforce: acquired by lockAllMetadata. + } } d.metadataMu.Unlock() // +checklocksforce: see above. continue } - // Note that synthetic dentries will always fails the comparison check - // below. - if !found || d.qidPath != stats[i].QID.Path { + // Note that synthetic dentries will always fail this comparison check. + var shouldInvalidate bool + if fs.opts.lisaEnabled { + shouldInvalidate = !found || d.inoKey != inoKeyFromStat(&statsLisa[i]) + } else { + shouldInvalidate = !found || d.qidPath != stats[i].QID.Path + } + if shouldInvalidate { d.metadataMu.Unlock() // +checklocksforce: see above. if !found && d.isSynthetic() { // We have a synthetic file, and no remote file has arisen to replace @@ -298,7 +326,11 @@ func (fs *filesystem) revalidateHelper(ctx context.Context, vfsObj *vfs.VirtualF } // The file at this path hasn't changed. Just update cached metadata. - d.updateFromP9AttrsLocked(stats[i].Valid, &stats[i].Attr) // +checklocksforce: see above. + if fs.opts.lisaEnabled { + d.updateFromLisaStatLocked(&statsLisa[i]) // +checklocksforce: see above. + } else { + d.updateFromP9AttrsLocked(stats[i].Valid, &stats[i].Attr) // +checklocksforce: see above. + } d.metadataMu.Unlock() } diff --git a/pkg/sentry/fsimpl/gofer/save_restore.go b/pkg/sentry/fsimpl/gofer/save_restore.go index 8dcbc61ed..475322527 100644 --- a/pkg/sentry/fsimpl/gofer/save_restore.go +++ b/pkg/sentry/fsimpl/gofer/save_restore.go @@ -24,6 +24,7 @@ import ( "gvisor.dev/gvisor/pkg/errors/linuxerr" "gvisor.dev/gvisor/pkg/fdnotifier" "gvisor.dev/gvisor/pkg/hostarch" + "gvisor.dev/gvisor/pkg/lisafs" "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/refsvfs2" "gvisor.dev/gvisor/pkg/safemem" @@ -112,10 +113,19 @@ func (d *dentry) prepareSaveRecursive(ctx context.Context) error { return err } } - if !d.readFile.isNil() || !d.writeFile.isNil() { - d.fs.savedDentryRW[d] = savedDentryRW{ - read: !d.readFile.isNil(), - write: !d.writeFile.isNil(), + if d.fs.opts.lisaEnabled { + if d.readFDLisa.Ok() || d.writeFDLisa.Ok() { + d.fs.savedDentryRW[d] = savedDentryRW{ + read: d.readFDLisa.Ok(), + write: d.writeFDLisa.Ok(), + } + } + } else { + if !d.readFile.isNil() || !d.writeFile.isNil() { + d.fs.savedDentryRW[d] = savedDentryRW{ + read: !d.readFile.isNil(), + write: !d.writeFile.isNil(), + } } } d.dirMu.Lock() @@ -177,25 +187,37 @@ func (fs *filesystem) CompleteRestore(ctx context.Context, opts vfs.CompleteRest return fmt.Errorf("no server FD available for filesystem with unique ID %q", fs.iopts.UniqueID) } fs.opts.fd = fd - if err := fs.dial(ctx); err != nil { - return err - } fs.inoByQIDPath = make(map[uint64]uint64) + fs.inoByKey = make(map[inoKey]uint64) - // Restore the filesystem root. - ctx.UninterruptibleSleepStart(false) - attached, err := fs.client.Attach(fs.opts.aname) - ctx.UninterruptibleSleepFinish(false) - if err != nil { - return err - } - attachFile := p9file{attached} - qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask()) - if err != nil { - return err - } - if err := fs.root.restoreFile(ctx, attachFile, qid, attrMask, &attr, &opts); err != nil { - return err + if fs.opts.lisaEnabled { + rootInode, err := fs.initClientLisa(ctx) + if err != nil { + return err + } + if err := fs.root.restoreFileLisa(ctx, rootInode, &opts); err != nil { + return err + } + } else { + if err := fs.dial(ctx); err != nil { + return err + } + + // Restore the filesystem root. + ctx.UninterruptibleSleepStart(false) + attached, err := fs.client.Attach(fs.opts.aname) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + return err + } + attachFile := p9file{attached} + qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask()) + if err != nil { + return err + } + if err := fs.root.restoreFile(ctx, attachFile, qid, attrMask, &attr, &opts); err != nil { + return err + } } // Restore remaining dentries. @@ -283,6 +305,55 @@ func (d *dentry) restoreFile(ctx context.Context, file p9file, qid p9.QID, attrM return nil } +func (d *dentry) restoreFileLisa(ctx context.Context, inode *lisafs.Inode, opts *vfs.CompleteRestoreOptions) error { + d.controlFDLisa = d.fs.clientLisa.NewFD(inode.ControlFD) + + // Gofers do not preserve inoKey across checkpoint/restore, so: + // + // - We must assume that the remote filesystem did not change in a way that + // would invalidate dentries, since we can't revalidate dentries by + // checking inoKey. + // + // - We need to associate the new inoKey with the existing d.ino. + d.inoKey = inoKeyFromStat(&inode.Stat) + d.fs.inoMu.Lock() + d.fs.inoByKey[d.inoKey] = d.ino + d.fs.inoMu.Unlock() + + // Check metadata stability before updating metadata. + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + if d.isRegularFile() { + if opts.ValidateFileSizes { + if inode.Stat.Mask&linux.STATX_SIZE != 0 { + return fmt.Errorf("gofer.dentry(%q).restoreFile: file size validation failed: file size not available", genericDebugPathname(d)) + } + if d.size != inode.Stat.Size { + return fmt.Errorf("gofer.dentry(%q).restoreFile: file size validation failed: size changed from %d to %d", genericDebugPathname(d), d.size, inode.Stat.Size) + } + } + if opts.ValidateFileModificationTimestamps { + if inode.Stat.Mask&linux.STATX_MTIME != 0 { + return fmt.Errorf("gofer.dentry(%q).restoreFile: mtime validation failed: mtime not available", genericDebugPathname(d)) + } + if want := dentryTimestampFromLisa(inode.Stat.Mtime); d.mtime != want { + return fmt.Errorf("gofer.dentry(%q).restoreFile: mtime validation failed: mtime changed from %+v to %+v", genericDebugPathname(d), linux.NsecToStatxTimestamp(d.mtime), linux.NsecToStatxTimestamp(want)) + } + } + } + if !d.cachedMetadataAuthoritative() { + d.updateFromLisaStatLocked(&inode.Stat) + } + + if rw, ok := d.fs.savedDentryRW[d]; ok { + if err := d.ensureSharedHandle(ctx, rw.read, rw.write, false /* trunc */); err != nil { + return err + } + } + + return nil +} + // Preconditions: d is not synthetic. func (d *dentry) restoreDescendantsRecursive(ctx context.Context, opts *vfs.CompleteRestoreOptions) error { for _, child := range d.children { @@ -305,19 +376,35 @@ func (d *dentry) restoreDescendantsRecursive(ctx context.Context, opts *vfs.Comp // only be detected by checking filesystem.syncableDentries). d.parent has been // restored. func (d *dentry) restoreRecursive(ctx context.Context, opts *vfs.CompleteRestoreOptions) error { - qid, file, attrMask, attr, err := d.parent.file.walkGetAttrOne(ctx, d.name) - if err != nil { - return err - } - if err := d.restoreFile(ctx, file, qid, attrMask, &attr, opts); err != nil { - return err + if d.fs.opts.lisaEnabled { + inode, err := d.parent.controlFDLisa.Walk(ctx, d.name) + if err != nil { + return err + } + if err := d.restoreFileLisa(ctx, inode, opts); err != nil { + return err + } + } else { + qid, file, attrMask, attr, err := d.parent.file.walkGetAttrOne(ctx, d.name) + if err != nil { + return err + } + if err := d.restoreFile(ctx, file, qid, attrMask, &attr, opts); err != nil { + return err + } } return d.restoreDescendantsRecursive(ctx, opts) } func (fd *specialFileFD) completeRestore(ctx context.Context) error { d := fd.dentry() - h, err := openHandle(ctx, d.file, fd.vfsfd.IsReadable(), fd.vfsfd.IsWritable(), false /* trunc */) + var h handle + var err error + if d.fs.opts.lisaEnabled { + h, err = openHandleLisa(ctx, d.controlFDLisa, fd.vfsfd.IsReadable(), fd.vfsfd.IsWritable(), false /* trunc */) + } else { + h, err = openHandle(ctx, d.file, fd.vfsfd.IsReadable(), fd.vfsfd.IsWritable(), false /* trunc */) + } if err != nil { return err } diff --git a/pkg/sentry/fsimpl/gofer/socket.go b/pkg/sentry/fsimpl/gofer/socket.go index fe15f8583..86ab70453 100644 --- a/pkg/sentry/fsimpl/gofer/socket.go +++ b/pkg/sentry/fsimpl/gofer/socket.go @@ -59,11 +59,6 @@ func sockTypeToP9(t linux.SockType) (p9.ConnectFlags, bool) { // BidirectionalConnect implements ConnectableEndpoint.BidirectionalConnect. func (e *endpoint) BidirectionalConnect(ctx context.Context, ce transport.ConnectingEndpoint, returnConnect func(transport.Receiver, transport.ConnectedEndpoint)) *syserr.Error { - cf, ok := sockTypeToP9(ce.Type()) - if !ok { - return syserr.ErrConnectionRefused - } - // No lock ordering required as only the ConnectingEndpoint has a mutex. ce.Lock() @@ -77,7 +72,7 @@ func (e *endpoint) BidirectionalConnect(ctx context.Context, ce transport.Connec return syserr.ErrInvalidEndpointState } - c, err := e.newConnectedEndpoint(ctx, cf, ce.WaiterQueue()) + c, err := e.newConnectedEndpoint(ctx, ce.Type(), ce.WaiterQueue()) if err != nil { ce.Unlock() return err @@ -95,7 +90,7 @@ func (e *endpoint) BidirectionalConnect(ctx context.Context, ce transport.Connec // UnidirectionalConnect implements // transport.BoundEndpoint.UnidirectionalConnect. func (e *endpoint) UnidirectionalConnect(ctx context.Context) (transport.ConnectedEndpoint, *syserr.Error) { - c, err := e.newConnectedEndpoint(ctx, p9.DgramSocket, &waiter.Queue{}) + c, err := e.newConnectedEndpoint(ctx, linux.SOCK_DGRAM, &waiter.Queue{}) if err != nil { return nil, err } @@ -111,25 +106,39 @@ func (e *endpoint) UnidirectionalConnect(ctx context.Context) (transport.Connect return c, nil } -func (e *endpoint) newConnectedEndpoint(ctx context.Context, flags p9.ConnectFlags, queue *waiter.Queue) (*host.SCMConnectedEndpoint, *syserr.Error) { - hostFile, err := e.dentry.file.connect(ctx, flags) - if err != nil { +func (e *endpoint) newConnectedEndpoint(ctx context.Context, sockType linux.SockType, queue *waiter.Queue) (*host.SCMConnectedEndpoint, *syserr.Error) { + if e.dentry.fs.opts.lisaEnabled { + hostSockFD, err := e.dentry.controlFDLisa.Connect(ctx, sockType) + if err != nil { + return nil, syserr.ErrConnectionRefused + } + + c, serr := host.NewSCMEndpoint(ctx, hostSockFD, queue, e.path) + if serr != nil { + unix.Close(hostSockFD) + log.Warningf("Gofer returned invalid host socket for BidirectionalConnect; file %+v sockType %d: %v", e.dentry.file, sockType, serr) + return nil, serr + } + return c, nil + } + + flags, ok := sockTypeToP9(sockType) + if !ok { return nil, syserr.ErrConnectionRefused } - // Dup the fd so that the new endpoint can manage its lifetime. - hostFD, err := unix.Dup(hostFile.FD()) + hostFile, err := e.dentry.file.connect(ctx, flags) if err != nil { - log.Warningf("Could not dup host socket fd %d: %v", hostFile.FD(), err) - return nil, syserr.FromError(err) + return nil, syserr.ErrConnectionRefused } - // After duplicating, we no longer need hostFile. - hostFile.Close() - c, serr := host.NewSCMEndpoint(ctx, hostFD, queue, e.path) + c, serr := host.NewSCMEndpoint(ctx, hostFile.FD(), queue, e.path) if serr != nil { - log.Warningf("Gofer returned invalid host socket for BidirectionalConnect; file %+v flags %+v: %v", e.dentry.file, flags, serr) + hostFile.Close() + log.Warningf("Gofer returned invalid host socket for BidirectionalConnect; file %+v sockType %d: %v", e.dentry.file, sockType, serr) return nil, serr } + // Ownership has been transferred to c. + hostFile.Release() return c, nil } diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go index a8d47b65b..c568bbfd2 100644 --- a/pkg/sentry/fsimpl/gofer/special_file.go +++ b/pkg/sentry/fsimpl/gofer/special_file.go @@ -23,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/errors/linuxerr" "gvisor.dev/gvisor/pkg/fdnotifier" "gvisor.dev/gvisor/pkg/hostarch" + "gvisor.dev/gvisor/pkg/lisafs" "gvisor.dev/gvisor/pkg/metric" "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/safemem" @@ -149,6 +150,9 @@ func (fd *specialFileFD) OnClose(ctx context.Context) error { if !fd.vfsfd.IsWritable() { return nil } + if fs := fd.filesystem(); fs.opts.lisaEnabled { + return fd.handle.fdLisa.Flush(ctx) + } return fd.handle.file.flush(ctx) } @@ -184,6 +188,9 @@ func (fd *specialFileFD) Allocate(ctx context.Context, mode, offset, length uint if fd.isRegularFile { d := fd.dentry() return d.doAllocate(ctx, offset, length, func() error { + if d.fs.opts.lisaEnabled { + return fd.handle.fdLisa.Allocate(ctx, mode, offset, length) + } return fd.handle.file.allocate(ctx, p9.ToAllocateMode(mode), offset, length) }) } @@ -371,10 +378,10 @@ func (fd *specialFileFD) Seek(ctx context.Context, offset int64, whence int32) ( // Sync implements vfs.FileDescriptionImpl.Sync. func (fd *specialFileFD) Sync(ctx context.Context) error { - return fd.sync(ctx, false /* forFilesystemSync */) + return fd.sync(ctx, false /* forFilesystemSync */, nil /* accFsyncFDIDsLisa */) } -func (fd *specialFileFD) sync(ctx context.Context, forFilesystemSync bool) error { +func (fd *specialFileFD) sync(ctx context.Context, forFilesystemSync bool, accFsyncFDIDsLisa *[]lisafs.FDID) error { // Locks to ensure it didn't race with fd.Release(). fd.releaseMu.RLock() defer fd.releaseMu.RUnlock() @@ -391,6 +398,13 @@ func (fd *specialFileFD) sync(ctx context.Context, forFilesystemSync bool) error ctx.UninterruptibleSleepFinish(false) return err } + if fs := fd.filesystem(); fs.opts.lisaEnabled { + if accFsyncFDIDsLisa != nil { + *accFsyncFDIDsLisa = append(*accFsyncFDIDsLisa, fd.handle.fdLisa.ID()) + return nil + } + return fd.handle.fdLisa.Sync(ctx) + } return fd.handle.file.fsync(ctx) }() if err != nil { diff --git a/pkg/sentry/fsimpl/gofer/symlink.go b/pkg/sentry/fsimpl/gofer/symlink.go index dbd834c67..27d9be5c4 100644 --- a/pkg/sentry/fsimpl/gofer/symlink.go +++ b/pkg/sentry/fsimpl/gofer/symlink.go @@ -35,7 +35,13 @@ func (d *dentry) readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { return target, nil } } - target, err := d.file.readlink(ctx) + var target string + var err error + if d.fs.opts.lisaEnabled { + target, err = d.controlFDLisa.ReadLinkAt(ctx) + } else { + target, err = d.file.readlink(ctx) + } if d.fs.opts.interop != InteropModeShared { if err == nil { d.haveTarget = true diff --git a/pkg/sentry/fsimpl/gofer/time.go b/pkg/sentry/fsimpl/gofer/time.go index 9cbe805b9..07940b225 100644 --- a/pkg/sentry/fsimpl/gofer/time.go +++ b/pkg/sentry/fsimpl/gofer/time.go @@ -17,6 +17,7 @@ package gofer import ( "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/vfs" ) @@ -24,6 +25,10 @@ func dentryTimestampFromP9(s, ns uint64) int64 { return int64(s*1e9 + ns) } +func dentryTimestampFromLisa(t linux.StatxTimestamp) int64 { + return t.Sec*1e9 + int64(t.Nsec) +} + // Preconditions: d.cachedMetadataAuthoritative() == true. func (d *dentry) touchAtime(mnt *vfs.Mount) { if mnt.Flags.NoATime || mnt.ReadOnly() { diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go index 7fd7f000d..40aff2927 100644 --- a/pkg/sentry/vfs/resolving_path.go +++ b/pkg/sentry/vfs/resolving_path.go @@ -223,6 +223,12 @@ func (rp *ResolvingPath) Final() bool { return rp.curPart == 0 && !rp.pit.NextOk() } +// Pit returns a copy of rp's current path iterator. Modifying the iterator +// does not change rp. +func (rp *ResolvingPath) Pit() fspath.Iterator { + return rp.pit +} + // Component returns the current path component in the stream represented by // rp. // |