diff options
Diffstat (limited to 'pkg/sentry/fsimpl/kernfs')
| -rw-r--r-- | pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go | 10 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/kernfs/fd_impl_util.go | 15 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/kernfs/filesystem.go | 257 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/kernfs/inode_impl_util.go | 98 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/kernfs/kernfs.go | 98 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/kernfs/kernfs_test.go | 28 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/kernfs/symlink.go | 6 | ||||
| -rw-r--r-- | pkg/sentry/fsimpl/kernfs/synthetic_directory.go | 18 |
8 files changed, 308 insertions, 222 deletions
diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go index 1ee089620..b929118b1 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -56,9 +56,9 @@ func (f *DynamicBytesFile) Init(creds *auth.Credentials, devMajor, devMinor uint } // Open implements Inode.Open. -func (f *DynamicBytesFile) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (f *DynamicBytesFile) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &DynamicBytesFD{} - if err := fd.Init(rp.Mount(), vfsd, f.data, &f.locks, opts.Flags); err != nil { + if err := fd.Init(rp.Mount(), d, f.data, &f.locks, opts.Flags); err != nil { return nil, err } return &fd.vfsfd, nil @@ -87,12 +87,12 @@ type DynamicBytesFD struct { } // Init initializes a DynamicBytesFD. -func (fd *DynamicBytesFD) Init(m *vfs.Mount, d *vfs.Dentry, data vfs.DynamicBytesSource, locks *vfs.FileLocks, flags uint32) error { +func (fd *DynamicBytesFD) Init(m *vfs.Mount, d *Dentry, data vfs.DynamicBytesSource, locks *vfs.FileLocks, flags uint32) error { fd.LockFD.Init(locks) - if err := fd.vfsfd.Init(fd, flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, flags, m, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return err } - fd.inode = d.Impl().(*Dentry).inode + fd.inode = d.inode fd.SetDataSource(data) return nil } diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go index 6518ff5cd..0a4cd4057 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -29,6 +29,8 @@ import ( ) // SeekEndConfig describes the SEEK_END behaviour for FDs. +// +// +stateify savable type SeekEndConfig int // Constants related to SEEK_END behaviour for FDs. @@ -41,6 +43,8 @@ const ( ) // GenericDirectoryFDOptions contains configuration for a GenericDirectoryFD. +// +// +stateify savable type GenericDirectoryFDOptions struct { SeekEnd SeekEndConfig } @@ -56,6 +60,8 @@ type GenericDirectoryFDOptions struct { // Must be initialize with Init before first use. // // Lock ordering: mu => children.mu. +// +// +stateify savable type GenericDirectoryFD struct { vfs.FileDescriptionDefaultImpl vfs.DirectoryFileDescriptionDefaultImpl @@ -68,7 +74,7 @@ type GenericDirectoryFD struct { children *OrderedChildren // mu protects the fields below. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // off is the current directory offset. Protected by "mu". off int64 @@ -76,12 +82,12 @@ type GenericDirectoryFD struct { // NewGenericDirectoryFD creates a new GenericDirectoryFD and returns its // dentry. -func NewGenericDirectoryFD(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, locks *vfs.FileLocks, opts *vfs.OpenOptions, fdOpts GenericDirectoryFDOptions) (*GenericDirectoryFD, error) { +func NewGenericDirectoryFD(m *vfs.Mount, d *Dentry, children *OrderedChildren, locks *vfs.FileLocks, opts *vfs.OpenOptions, fdOpts GenericDirectoryFDOptions) (*GenericDirectoryFD, error) { fd := &GenericDirectoryFD{} if err := fd.Init(children, locks, opts, fdOpts); err != nil { return nil, err } - if err := fd.vfsfd.Init(fd, opts.Flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, opts.Flags, m, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return fd, nil @@ -195,8 +201,7 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // these. childIdx := fd.off - 2 for it := fd.children.nthLocked(childIdx); it != nil; it = it.Next() { - inode := it.Dentry.Impl().(*Dentry).inode - stat, err := inode.Stat(ctx, fd.filesystem(), opts) + stat, err := it.Dentry.inode.Stat(ctx, fd.filesystem(), opts) if err != nil { return err } diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index d7d3e8f48..5cc1c4281 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -37,8 +37,7 @@ import ( // * !rp.Done(). // // Postcondition: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) stepExistingLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, mayFollowSymlinks bool) (*vfs.Dentry, error) { - d := vfsd.Impl().(*Dentry) +func (fs *Filesystem) stepExistingLocked(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, mayFollowSymlinks bool) (*Dentry, error) { if !d.isDir() { return nil, syserror.ENOTDIR } @@ -55,20 +54,20 @@ afterSymlink: // calls d_revalidate(), but walk_component() => handle_dots() does not. if name == "." { rp.Advance() - return vfsd, nil + return d, nil } if name == ".." { - if isRoot, err := rp.CheckRoot(ctx, vfsd); err != nil { + if isRoot, err := rp.CheckRoot(ctx, d.VFSDentry()); err != nil { return nil, err } else if isRoot || d.parent == nil { rp.Advance() - return vfsd, nil + return d, nil } - if err := rp.CheckMount(ctx, &d.parent.vfsd); err != nil { + if err := rp.CheckMount(ctx, d.parent.VFSDentry()); err != nil { return nil, err } rp.Advance() - return &d.parent.vfsd, nil + return d.parent, nil } if len(name) > linux.NAME_MAX { return nil, syserror.ENAMETOOLONG @@ -79,7 +78,7 @@ afterSymlink: if err != nil { return nil, err } - if err := rp.CheckMount(ctx, &next.vfsd); err != nil { + if err := rp.CheckMount(ctx, next.VFSDentry()); err != nil { return nil, err } // Resolve any symlink at current path component. @@ -102,7 +101,7 @@ afterSymlink: goto afterSymlink } rp.Advance() - return &next.vfsd, nil + return next, nil } // revalidateChildLocked must be called after a call to parent.vfsd.Child(name) @@ -122,25 +121,21 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir if !child.inode.Valid(ctx) { delete(parent.children, name) vfsObj.InvalidateDentry(ctx, &child.vfsd) - fs.deferDecRef(&child.vfsd) // Reference from Lookup. + fs.deferDecRef(child) // Reference from Lookup. child = nil } } if child == nil { - // Dentry isn't cached; it either doesn't exist or failed - // revalidation. Attempt to resolve it via Lookup. - // - // FIXME(gvisor.dev/issue/1193): Inode.Lookup() should return - // *(kernfs.)Dentry, not *vfs.Dentry, since (kernfs.)Filesystem assumes - // that all dentries in the filesystem are (kernfs.)Dentry and performs - // vfs.DentryImpl casts accordingly. - childVFSD, err := parent.inode.Lookup(ctx, name) + // Dentry isn't cached; it either doesn't exist or failed revalidation. + // Attempt to resolve it via Lookup. + c, err := parent.inode.Lookup(ctx, name) if err != nil { return nil, err } - // Reference on childVFSD dropped by a corresponding Valid. - child = childVFSD.Impl().(*Dentry) - parent.insertChildLocked(name, child) + // Reference on c (provided by Lookup) will be dropped when the dentry + // fails validation. + parent.InsertChildLocked(name, c) + child = c } return child, nil } @@ -153,20 +148,19 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // Preconditions: Filesystem.mu must be locked for at least reading. // // Postconditions: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, Inode, error) { - vfsd := rp.Start() +func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingPath) (*Dentry, error) { + d := rp.Start().Impl().(*Dentry) for !rp.Done() { var err error - vfsd, err = fs.stepExistingLocked(ctx, rp, vfsd, true /* mayFollowSymlinks */) + d, err = fs.stepExistingLocked(ctx, rp, d, true /* mayFollowSymlinks */) if err != nil { - return nil, nil, err + return nil, err } } - d := vfsd.Impl().(*Dentry) if rp.MustBeDir() && !d.isDir() { - return nil, nil, syserror.ENOTDIR + return nil, syserror.ENOTDIR } - return vfsd, d.inode, nil + return d, nil } // walkParentDirLocked resolves all but the last path component of rp to an @@ -181,20 +175,19 @@ func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingP // * !rp.Done(). // // Postconditions: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, Inode, error) { - vfsd := rp.Start() +func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath) (*Dentry, error) { + d := rp.Start().Impl().(*Dentry) for !rp.Final() { var err error - vfsd, err = fs.stepExistingLocked(ctx, rp, vfsd, true /* mayFollowSymlinks */) + d, err = fs.stepExistingLocked(ctx, rp, d, true /* mayFollowSymlinks */) if err != nil { - return nil, nil, err + return nil, err } } - d := vfsd.Impl().(*Dentry) if !d.isDir() { - return nil, nil, syserror.ENOTDIR + return nil, syserror.ENOTDIR } - return vfsd, d.inode, nil + return d, nil } // checkCreateLocked checks that a file named rp.Component() may be created in @@ -202,10 +195,9 @@ func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.Resolving // // Preconditions: // * Filesystem.mu must be locked for at least reading. -// * parentInode == parentVFSD.Impl().(*Dentry).Inode. // * isDir(parentInode) == true. -func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *vfs.Dentry, parentInode Inode) (string, error) { - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { +func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parent *Dentry) (string, error) { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return "", err } pc := rp.Component() @@ -215,11 +207,10 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v if len(pc) > linux.NAME_MAX { return "", syserror.ENAMETOOLONG } - // FIXME(gvisor.dev/issue/1193): Data race due to not holding dirMu. - if _, ok := parentVFSD.Impl().(*Dentry).children[pc]; ok { + if _, ok := parent.children[pc]; ok { return "", syserror.EEXIST } - if parentVFSD.IsDead() { + if parent.VFSDentry().IsDead() { return "", syserror.ENOENT } return pc, nil @@ -228,8 +219,8 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v // checkDeleteLocked checks that the file represented by vfsd may be deleted. // // Preconditions: Filesystem.mu must be locked for at least reading. -func checkDeleteLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry) error { - parent := vfsd.Impl().(*Dentry).parent +func checkDeleteLocked(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry) error { + parent := d.parent if parent == nil { return syserror.EBUSY } @@ -258,11 +249,11 @@ func (fs *Filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return err } - return inode.CheckPermissions(ctx, creds, ats) + return d.inode.CheckPermissions(ctx, creds, ats) } // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. @@ -270,20 +261,20 @@ func (fs *Filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op fs.mu.RLock() defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return nil, err } if opts.CheckSearchable { - d := vfsd.Impl().(*Dentry) if !d.isDir() { return nil, syserror.ENOTDIR } - if err := inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { return nil, err } } + vfsd := d.VFSDentry() vfsd.IncRef() // Ownership transferred to caller. return vfsd, nil } @@ -293,12 +284,12 @@ func (fs *Filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPa fs.mu.RLock() defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - vfsd, _, err := fs.walkParentDirLocked(ctx, rp) + d, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return nil, err } - vfsd.IncRef() // Ownership transferred to caller. - return vfsd, nil + d.IncRef() // Ownership transferred to caller. + return d.VFSDentry(), nil } // LinkAt implements vfs.FilesystemImpl.LinkAt. @@ -308,12 +299,15 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. } fs.mu.Lock() defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -330,11 +324,11 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EPERM } - childVFSD, err := parentInode.NewLink(ctx, pc, d.inode) + child, err := parent.inode.NewLink(ctx, pc, d.inode) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + parent.InsertChildLocked(pc, child) return nil } @@ -345,12 +339,15 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v } fs.mu.Lock() defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -358,14 +355,14 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - childVFSD, err := parentInode.NewDir(ctx, pc, opts) + child, err := parent.inode.NewDir(ctx, pc, opts) if err != nil { if !opts.ForSyntheticMountpoint || err == syserror.EEXIST { return err } - childVFSD = newSyntheticDirectory(rp.Credentials(), opts.Mode) + child = newSyntheticDirectory(rp.Credentials(), opts.Mode) } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + parent.InsertChildLocked(pc, child) return nil } @@ -376,12 +373,15 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v } fs.mu.Lock() defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -389,11 +389,11 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - newVFSD, err := parentInode.NewNode(ctx, pc, opts) + newD, err := parent.inode.NewNode(ctx, pc, opts) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, newVFSD.Impl().(*Dentry)) + parent.InsertChildLocked(pc, newD) return nil } @@ -409,28 +409,27 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf // Do not create new file. if opts.Flags&linux.O_CREAT == 0 { fs.mu.RLock() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) return nil, err } - if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) return nil, err } - inode.IncRef() - defer inode.DecRef(ctx) + d.inode.IncRef() + defer d.inode.DecRef(ctx) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) - return inode.Open(ctx, rp, vfsd, opts) + return d.inode.Open(ctx, rp, d, opts) } // May create new file. mustCreate := opts.Flags&linux.O_EXCL != 0 - vfsd := rp.Start() - inode := vfsd.Impl().(*Dentry).inode + d := rp.Start().Impl().(*Dentry) fs.mu.Lock() unlocked := false unlock := func() { @@ -447,22 +446,22 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if mustCreate { return nil, syserror.EEXIST } - if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - inode.IncRef() - defer inode.DecRef(ctx) + d.inode.IncRef() + defer d.inode.DecRef(ctx) unlock() - return inode.Open(ctx, rp, vfsd, opts) + return d.inode.Open(ctx, rp, d, opts) } afterTrailingSymlink: - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return nil, err } // Check for search permission in the parent directory. - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { return nil, err } // Reject attempts to open directories with O_CREAT. @@ -477,10 +476,10 @@ afterTrailingSymlink: return nil, syserror.ENAMETOOLONG } // Determine whether or not we need to create a file. - childVFSD, err := fs.stepExistingLocked(ctx, rp, parentVFSD, false /* mayFollowSymlinks */) + child, err := fs.stepExistingLocked(ctx, rp, parent, false /* mayFollowSymlinks */) if err == syserror.ENOENT { // Already checked for searchability above; now check for writability. - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } if err := rp.Mount().CheckBeginWrite(); err != nil { @@ -488,16 +487,18 @@ afterTrailingSymlink: } defer rp.Mount().EndWrite() // Create and open the child. - childVFSD, err = parentInode.NewFile(ctx, pc, opts) + child, err := parent.inode.NewFile(ctx, pc, opts) if err != nil { return nil, err } - child := childVFSD.Impl().(*Dentry) - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) + // FIXME(gvisor.dev/issue/1193): Race between checking existence with + // fs.stepExistingLocked and parent.InsertChild. If possible, we should hold + // dirMu from one to the other. + parent.InsertChild(pc, child) child.inode.IncRef() defer child.inode.DecRef(ctx) unlock() - return child.inode.Open(ctx, rp, childVFSD, opts) + return child.inode.Open(ctx, rp, child, opts) } if err != nil { return nil, err @@ -506,7 +507,6 @@ afterTrailingSymlink: if mustCreate { return nil, syserror.EEXIST } - child := childVFSD.Impl().(*Dentry) if rp.ShouldFollowSymlink() && child.isSymlink() { targetVD, targetPathname, err := child.inode.Getlink(ctx, rp.Mount()) if err != nil { @@ -533,22 +533,22 @@ afterTrailingSymlink: child.inode.IncRef() defer child.inode.DecRef(ctx) unlock() - return child.inode.Open(ctx, rp, &child.vfsd, opts) + return child.inode.Open(ctx, rp, child, opts) } // ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. func (fs *Filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) { fs.mu.RLock() - d, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { return "", err } - if !d.Impl().(*Dentry).isSymlink() { + if !d.isSymlink() { return "", syserror.EINVAL } - return inode.Readlink(ctx) + return d.inode.Readlink(ctx, rp.Mount()) } // RenameAt implements vfs.FilesystemImpl.RenameAt. @@ -565,11 +565,10 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // Resolve the destination directory first to verify that it's on this // Mount. - dstDirVFSD, dstDirInode, err := fs.walkParentDirLocked(ctx, rp) + dstDir, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return err } - dstDir := dstDirVFSD.Impl().(*Dentry) mnt := rp.Mount() if mnt != oldParentVD.Mount() { return syserror.EXDEV @@ -587,16 +586,15 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if err != nil { return err } - srcVFSD := &src.vfsd // Can we remove the src dentry? - if err := checkDeleteLocked(ctx, rp, srcVFSD); err != nil { + if err := checkDeleteLocked(ctx, rp, src); err != nil { return err } // Can we create the dst dentry? var dst *Dentry - pc, err := checkCreateLocked(ctx, rp, dstDirVFSD, dstDirInode) + pc, err := checkCreateLocked(ctx, rp, dstDir) switch err { case nil: // Ok, continue with rename as replacement. @@ -607,14 +605,14 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } dst = dstDir.children[pc] if dst == nil { - panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDirVFSD)) + panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDir)) } default: return err } var dstVFSD *vfs.Dentry if dst != nil { - dstVFSD = &dst.vfsd + dstVFSD = dst.VFSDentry() } mntns := vfs.MountNamespaceFromContext(ctx) @@ -630,17 +628,18 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa defer dstDir.dirMu.Unlock() } + srcVFSD := src.VFSDentry() if err := virtfs.PrepareRenameDentry(mntns, srcVFSD, dstVFSD); err != nil { return err } - replaced, err := srcDir.inode.Rename(ctx, src.name, pc, srcVFSD, dstDirVFSD) + replaced, err := srcDir.inode.Rename(ctx, src.name, pc, src, dstDir) if err != nil { virtfs.AbortRenameDentry(srcVFSD, dstVFSD) return err } delete(srcDir.children, src.name) if srcDir != dstDir { - fs.deferDecRef(srcDirVFSD) + fs.deferDecRef(srcDir) dstDir.IncRef() } src.parent = dstDir @@ -649,7 +648,11 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa dstDir.children = make(map[string]*Dentry) } dstDir.children[pc] = src - virtfs.CommitRenameReplaceDentry(ctx, srcVFSD, replaced) + var replaceVFSD *vfs.Dentry + if replaced != nil { + replaceVFSD = replaced.VFSDentry() + } + virtfs.CommitRenameReplaceDentry(ctx, srcVFSD, replaceVFSD) return nil } @@ -657,7 +660,8 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { fs.mu.Lock() defer fs.mu.Unlock() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + + d, err := fs.walkExistingLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err @@ -666,14 +670,13 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error return err } defer rp.Mount().EndWrite() - if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { + if err := checkDeleteLocked(ctx, rp, d); err != nil { return err } - d := vfsd.Impl().(*Dentry) if !d.isDir() { return syserror.ENOTDIR } - if inode.HasChildren() { + if d.inode.HasChildren() { return syserror.ENOTEMPTY } virtfs := rp.VirtualFilesystem() @@ -683,10 +686,12 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef(ctx) + vfsd := d.VFSDentry() if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } - if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil { + + if err := parentDentry.inode.RmDir(ctx, d.name, d); err != nil { virtfs.AbortDeleteDentry(vfsd) return err } @@ -697,7 +702,7 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error // SetStatAt implements vfs.FilesystemImpl.SetStatAt. func (fs *Filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { @@ -706,31 +711,31 @@ func (fs *Filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts if opts.Stat.Mask == 0 { return nil } - return inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts) + return d.inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts) } // StatAt implements vfs.FilesystemImpl.StatAt. func (fs *Filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { return linux.Statx{}, err } - return inode.Stat(ctx, fs.VFSFilesystem(), opts) + return d.inode.Stat(ctx, fs.VFSFilesystem(), opts) } // StatFSAt implements vfs.FilesystemImpl.StatFSAt. func (fs *Filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { return linux.Statfs{}, err } - return inode.StatFS(ctx, fs.VFSFilesystem()) + return d.inode.StatFS(ctx, fs.VFSFilesystem()) } // SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. @@ -740,12 +745,15 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ } fs.mu.Lock() defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) + parent, err := fs.walkParentDirLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -753,11 +761,11 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ return err } defer rp.Mount().EndWrite() - childVFSD, err := parentInode.NewSymlink(ctx, pc, target) + child, err := parent.inode.NewSymlink(ctx, pc, target) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + parent.InsertChildLocked(pc, child) return nil } @@ -765,7 +773,8 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { fs.mu.Lock() defer fs.mu.Unlock() - vfsd, _, err := fs.walkExistingLocked(ctx, rp) + + d, err := fs.walkExistingLocked(ctx, rp) fs.processDeferredDecRefsLocked(ctx) if err != nil { return err @@ -774,10 +783,9 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return err } defer rp.Mount().EndWrite() - if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { + if err := checkDeleteLocked(ctx, rp, d); err != nil { return err } - d := vfsd.Impl().(*Dentry) if d.isDir() { return syserror.EISDIR } @@ -787,10 +795,11 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error defer parentDentry.dirMu.Unlock() mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef(ctx) + vfsd := d.VFSDentry() if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } - if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil { + if err := parentDentry.inode.Unlink(ctx, d.name, d); err != nil { virtfs.AbortDeleteDentry(vfsd) return err } @@ -798,16 +807,16 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return nil } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *Filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { return nil, err } - if err := inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } return nil, syserror.ECONNREFUSED @@ -816,7 +825,7 @@ func (fs *Filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath // ListXattrAt implements vfs.FilesystemImpl.ListXattrAt. func (fs *Filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) + _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { @@ -829,7 +838,7 @@ func (fs *Filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, si // GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. func (fs *Filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetXattrOptions) (string, error) { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) + _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { @@ -842,7 +851,7 @@ func (fs *Filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt // SetXattrAt implements vfs.FilesystemImpl.SetXattrAt. func (fs *Filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) + _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { @@ -855,7 +864,7 @@ func (fs *Filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt // RemoveXattrAt implements vfs.FilesystemImpl.RemoveXattrAt. func (fs *Filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) + _, err := fs.walkExistingLocked(ctx, rp) fs.mu.RUnlock() fs.processDeferredDecRefs(ctx) if err != nil { diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index 74408e322..49210e748 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -31,6 +31,8 @@ import ( // count for inodes, performing no extra actions when references are obtained or // released. This is suitable for simple file inodes that don't reference any // resources. +// +// +stateify savable type InodeNoopRefCount struct { } @@ -50,30 +52,32 @@ func (InodeNoopRefCount) TryIncRef() bool { // InodeDirectoryNoNewChildren partially implements the Inode interface. // InodeDirectoryNoNewChildren represents a directory inode which does not // support creation of new children. +// +// +stateify savable type InodeDirectoryNoNewChildren struct{} // NewFile implements Inode.NewFile. -func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*Dentry, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*Dentry, error) { return nil, syserror.EPERM } // NewLink implements Inode.NewLink. -func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*Dentry, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*Dentry, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*Dentry, error) { return nil, syserror.EPERM } @@ -81,6 +85,8 @@ func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOpt // inodeDirectory and inodeDynamicDirectory sub interfaces. Inodes that do not // represent directories can embed this to provide no-op implementations for // directory-related functions. +// +// +stateify savable type InodeNotDirectory struct { } @@ -90,47 +96,47 @@ func (InodeNotDirectory) HasChildren() bool { } // NewFile implements Inode.NewFile. -func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*Dentry, error) { panic("NewFile called on non-directory inode") } // NewDir implements Inode.NewDir. -func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*Dentry, error) { panic("NewDir called on non-directory inode") } // NewLink implements Inode.NewLinkink. -func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*Dentry, error) { panic("NewLink called on non-directory inode") } // NewSymlink implements Inode.NewSymlink. -func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*Dentry, error) { panic("NewSymlink called on non-directory inode") } // NewNode implements Inode.NewNode. -func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*Dentry, error) { panic("NewNode called on non-directory inode") } // Unlink implements Inode.Unlink. -func (InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) Unlink(context.Context, string, *Dentry) error { panic("Unlink called on non-directory inode") } // RmDir implements Inode.RmDir. -func (InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) RmDir(context.Context, string, *Dentry) error { panic("RmDir called on non-directory inode") } // Rename implements Inode.Rename. -func (InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { +func (InodeNotDirectory) Rename(context.Context, string, string, *Dentry, *Dentry) (*Dentry, error) { panic("Rename called on non-directory inode") } // Lookup implements Inode.Lookup. -func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*Dentry, error) { panic("Lookup called on non-directory inode") } @@ -149,10 +155,12 @@ func (InodeNotDirectory) Valid(context.Context) bool { // dymanic entries (i.e. entries that are not "hashed" into the // vfs.Dentry.children) can embed this to provide no-op implementations for // functions related to dynamic entries. +// +// +stateify savable type InodeNoDynamicLookup struct{} // Lookup implements Inode.Lookup. -func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*Dentry, error) { return nil, syserror.ENOENT } @@ -169,10 +177,12 @@ func (InodeNoDynamicLookup) Valid(ctx context.Context) bool { // InodeNotSymlink partially implements the Inode interface, specifically the // inodeSymlink sub interface. All inodes that are not symlinks may embed this // to return the appropriate errors from symlink-related functions. +// +// +stateify savable type InodeNotSymlink struct{} // Readlink implements Inode.Readlink. -func (InodeNotSymlink) Readlink(context.Context) (string, error) { +func (InodeNotSymlink) Readlink(context.Context, *vfs.Mount) (string, error) { return "", syserror.EINVAL } @@ -186,6 +196,8 @@ func (InodeNotSymlink) Getlink(context.Context, *vfs.Mount) (vfs.VirtualDentry, // inode attributes. // // Must be initialized by Init prior to first use. +// +// +stateify savable type InodeAttrs struct { devMajor uint32 devMinor uint32 @@ -256,6 +268,13 @@ func (a *InodeAttrs) Stat(context.Context, *vfs.Filesystem, vfs.StatOptions) (li // SetStat implements Inode.SetStat. func (a *InodeAttrs) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + return a.SetInodeStat(ctx, fs, creds, opts) +} + +// SetInodeStat sets the corresponding attributes from opts to InodeAttrs. +// This function can be used by other kernfs-based filesystem implementation to +// sets the unexported attributes into kernfs.InodeAttrs. +func (a *InodeAttrs) SetInodeStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { if opts.Stat.Mask == 0 { return nil } @@ -323,13 +342,16 @@ func (a *InodeAttrs) DecLinks() { } } +// +stateify savable type slot struct { Name string - Dentry *vfs.Dentry + Dentry *Dentry slotEntry } // OrderedChildrenOptions contains initialization options for OrderedChildren. +// +// +stateify savable type OrderedChildrenOptions struct { // Writable indicates whether vfs.FilesystemImpl methods implemented by // OrderedChildren may modify the tracked children. This applies to @@ -345,12 +367,14 @@ type OrderedChildrenOptions struct { // directories. // // Must be initialize with Init before first use. +// +// +stateify savable type OrderedChildren struct { // Can children be modified by user syscalls? It set to false, interface // methods that would modify the children return EPERM. Immutable. writable bool - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` order slotList set map[string]*slot } @@ -383,7 +407,7 @@ func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint3 if child.isDir() { links++ } - if err := o.Insert(name, child.VFSDentry()); err != nil { + if err := o.Insert(name, child); err != nil { panic(fmt.Sprintf("Collision when attempting to insert child %q (%+v) into %+v", name, child, d)) } d.InsertChild(name, child) @@ -400,7 +424,7 @@ func (o *OrderedChildren) HasChildren() bool { // Insert inserts child into o. This ignores the writability of o, as this is // not part of the vfs.FilesystemImpl interface, and is a lower-level operation. -func (o *OrderedChildren) Insert(name string, child *vfs.Dentry) error { +func (o *OrderedChildren) Insert(name string, child *Dentry) error { o.mu.Lock() defer o.mu.Unlock() if _, ok := o.set[name]; ok { @@ -424,10 +448,10 @@ func (o *OrderedChildren) removeLocked(name string) { } // Precondition: caller must hold o.mu for writing. -func (o *OrderedChildren) replaceChildLocked(name string, new *vfs.Dentry) *vfs.Dentry { +func (o *OrderedChildren) replaceChildLocked(name string, new *Dentry) *Dentry { if s, ok := o.set[name]; ok { // Existing slot with given name, simply replace the dentry. - var old *vfs.Dentry + var old *Dentry old, s.Dentry = s.Dentry, new return old } @@ -443,7 +467,7 @@ func (o *OrderedChildren) replaceChildLocked(name string, new *vfs.Dentry) *vfs. } // Precondition: caller must hold o.mu for reading or writing. -func (o *OrderedChildren) checkExistingLocked(name string, child *vfs.Dentry) error { +func (o *OrderedChildren) checkExistingLocked(name string, child *Dentry) error { s, ok := o.set[name] if !ok { return syserror.ENOENT @@ -455,7 +479,7 @@ func (o *OrderedChildren) checkExistingLocked(name string, child *vfs.Dentry) er } // Unlink implements Inode.Unlink. -func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *vfs.Dentry) error { +func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *Dentry) error { if !o.writable { return syserror.EPERM } @@ -471,12 +495,13 @@ func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *vfs.De } // Rmdir implements Inode.Rmdir. -func (o *OrderedChildren) RmDir(ctx context.Context, name string, child *vfs.Dentry) error { +func (o *OrderedChildren) RmDir(ctx context.Context, name string, child *Dentry) error { // We're not responsible for checking that child is a directory, that it's // empty, or updating any link counts; so this is the same as unlink. return o.Unlink(ctx, name, child) } +// +stateify savable type renameAcrossDifferentImplementationsError struct{} func (renameAcrossDifferentImplementationsError) Error() string { @@ -492,8 +517,8 @@ func (renameAcrossDifferentImplementationsError) Error() string { // that will support Rename. // // Postcondition: reference on any replaced dentry transferred to caller. -func (o *OrderedChildren) Rename(ctx context.Context, oldname, newname string, child, dstDir *vfs.Dentry) (*vfs.Dentry, error) { - dst, ok := dstDir.Impl().(*Dentry).inode.(interface{}).(*OrderedChildren) +func (o *OrderedChildren) Rename(ctx context.Context, oldname, newname string, child, dstDir *Dentry) (*Dentry, error) { + dst, ok := dstDir.inode.(interface{}).(*OrderedChildren) if !ok { return nil, renameAcrossDifferentImplementationsError{} } @@ -535,12 +560,14 @@ func (o *OrderedChildren) nthLocked(i int64) *slot { } // InodeSymlink partially implements Inode interface for symlinks. +// +// +stateify savable type InodeSymlink struct { InodeNotDirectory } // Open implements Inode.Open. -func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { return nil, syserror.ELOOP } @@ -567,6 +594,7 @@ var _ Inode = (*StaticDirectory)(nil) func NewStaticDir(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode, children map[string]*Dentry, fdOpts GenericDirectoryFDOptions) *Dentry { inode := &StaticDirectory{} inode.Init(creds, devMajor, devMinor, ino, perm, fdOpts) + inode.EnableLeakCheck() dentry := &Dentry{} dentry.Init(inode) @@ -587,9 +615,9 @@ func (s *StaticDirectory) Init(creds *auth.Credentials, devMajor, devMinor uint3 s.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeDirectory|perm) } -// Open implements kernfs.Inode. -func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := NewGenericDirectoryFD(rp.Mount(), vfsd, &s.OrderedChildren, &s.locks, &opts, s.fdOpts) +// Open implements kernfs.Inode.Open. +func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := NewGenericDirectoryFD(rp.Mount(), d, &s.OrderedChildren, &s.locks, &opts, s.fdOpts) if err != nil { return nil, err } @@ -601,21 +629,25 @@ func (*StaticDirectory) SetStat(context.Context, *vfs.Filesystem, *auth.Credenti return syserror.EPERM } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (s *StaticDirectory) DecRef(context.Context) { s.StaticDirectoryRefs.DecRef(s.Destroy) } // AlwaysValid partially implements kernfs.inodeDynamicLookup. +// +// +stateify savable type AlwaysValid struct{} -// Valid implements kernfs.inodeDynamicLookup. +// Valid implements kernfs.inodeDynamicLookup.Valid. func (*AlwaysValid) Valid(context.Context) bool { return true } // InodeNoStatFS partially implements the Inode interface, where the client // filesystem doesn't support statfs(2). +// +// +stateify savable type InodeNoStatFS struct{} // StatFS implements Inode.StatFS. diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index 88fcd54aa..6d3d79333 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -29,7 +29,7 @@ // // Reference Model: // -// Kernfs dentries represents named pointers to inodes. Dentries and inode have +// Kernfs dentries represents named pointers to inodes. Dentries and inodes have // independent lifetimes and reference counts. A child dentry unconditionally // holds a reference on its parent directory's dentry. A dentry also holds a // reference on the inode it points to. Multiple dentries can point to the same @@ -60,20 +60,23 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" ) // Filesystem mostly implements vfs.FilesystemImpl for a generic in-memory // filesystem. Concrete implementations are expected to embed this in their own // Filesystem type. +// +// +stateify savable type Filesystem struct { vfsfs vfs.Filesystem - droppedDentriesMu sync.Mutex + droppedDentriesMu sync.Mutex `state:"nosave"` // droppedDentries is a list of dentries waiting to be DecRef()ed. This is // used to defer dentry destruction until mu can be acquired for // writing. Protected by droppedDentriesMu. - droppedDentries []*vfs.Dentry + droppedDentries []*Dentry // mu synchronizes the lifetime of Dentries on this filesystem. Holding it // for reading guarantees continued existence of any resolved dentries, but @@ -96,7 +99,7 @@ type Filesystem struct { // defer fs.mu.RUnlock() // ... // fs.deferDecRef(dentry) - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` // nextInoMinusOne is used to to allocate inode numbers on this // filesystem. Must be accessed by atomic operations. @@ -107,7 +110,7 @@ type Filesystem struct { // processDeferredDecRefs{,Locked}. See comment on Filesystem.mu. // // Precondition: d must not already be pending destruction. -func (fs *Filesystem) deferDecRef(d *vfs.Dentry) { +func (fs *Filesystem) deferDecRef(d *Dentry) { fs.droppedDentriesMu.Lock() fs.droppedDentries = append(fs.droppedDentries, d) fs.droppedDentriesMu.Unlock() @@ -159,6 +162,8 @@ const ( // to, and child dentries hold a reference on their parent. // // Must be initialized by Init prior to first use. +// +// +stateify savable type Dentry struct { DentryRefs @@ -172,7 +177,11 @@ type Dentry struct { name string // dirMu protects children and the names of child Dentries. - dirMu sync.Mutex + // + // Note that holding fs.mu for writing is not sufficient; + // revalidateChildLocked(), which is a very hot path, may modify children with + // fs.mu acquired for reading only. + dirMu sync.Mutex `state:"nosave"` children map[string]*Dentry inode Inode @@ -239,24 +248,25 @@ func (d *Dentry) Watches() *vfs.Watches { func (d *Dentry) OnZeroWatches(context.Context) {} // InsertChild inserts child into the vfs dentry cache with the given name under -// this dentry. This does not update the directory inode, so calling this on -// its own isn't sufficient to insert a child into a directory. InsertChild -// updates the link count on d if required. +// this dentry. This does not update the directory inode, so calling this on its +// own isn't sufficient to insert a child into a directory. // // Precondition: d must represent a directory inode. func (d *Dentry) InsertChild(name string, child *Dentry) { d.dirMu.Lock() - d.insertChildLocked(name, child) + d.InsertChildLocked(name, child) d.dirMu.Unlock() } -// insertChildLocked is equivalent to InsertChild, with additional +// InsertChildLocked is equivalent to InsertChild, with additional // preconditions. // -// Precondition: d.dirMu must be locked. -func (d *Dentry) insertChildLocked(name string, child *Dentry) { +// Preconditions: +// * d must represent a directory inode. +// * d.dirMu must be locked. +func (d *Dentry) InsertChildLocked(name string, child *Dentry) { if !d.isDir() { - panic(fmt.Sprintf("InsertChild called on non-directory Dentry: %+v.", d)) + panic(fmt.Sprintf("InsertChildLocked called on non-directory Dentry: %+v.", d)) } d.IncRef() // DecRef in child's Dentry.destroy. child.parent = d @@ -267,6 +277,36 @@ func (d *Dentry) insertChildLocked(name string, child *Dentry) { d.children[name] = child } +// RemoveChild removes child from the vfs dentry cache. This does not update the +// directory inode or modify the inode to be unlinked. So calling this on its own +// isn't sufficient to remove a child from a directory. +// +// Precondition: d must represent a directory inode. +func (d *Dentry) RemoveChild(name string, child *Dentry) error { + d.dirMu.Lock() + defer d.dirMu.Unlock() + return d.RemoveChildLocked(name, child) +} + +// RemoveChildLocked is equivalent to RemoveChild, with additional +// preconditions. +// +// Precondition: d.dirMu must be locked. +func (d *Dentry) RemoveChildLocked(name string, child *Dentry) error { + if !d.isDir() { + panic(fmt.Sprintf("RemoveChild called on non-directory Dentry: %+v.", d)) + } + c, ok := d.children[name] + if !ok { + return syserror.ENOENT + } + if c != child { + panic(fmt.Sprintf("Dentry hashed into inode doesn't match what vfs thinks! Child: %+v, vfs: %+v", c, child)) + } + delete(d.children, name) + return nil +} + // Inode returns the dentry's inode. func (d *Dentry) Inode() Inode { return d.inode @@ -287,7 +327,6 @@ func (d *Dentry) Inode() Inode { // // - Checking that dentries passed to methods are of the appropriate file type. // - Checking permissions. -// - Updating link and reference counts. // // Specific responsibilities of implementations are documented below. type Inode interface { @@ -297,7 +336,8 @@ type Inode interface { inodeRefs // Methods related to node metadata. A generic implementation is provided by - // InodeAttrs. + // InodeAttrs. Note that a concrete filesystem using kernfs is responsible for + // managing link counts. inodeMetadata // Method for inodes that represent symlink. InodeNotSymlink provides a @@ -315,11 +355,11 @@ type Inode interface { // Open creates a file description for the filesystem object represented by // this inode. The returned file description should hold a reference on the - // inode for its lifetime. + // dentry for its lifetime. // // Precondition: rp.Done(). vfsd.Impl() must be the kernfs Dentry containing // the inode on which Open() is being called. - Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) + Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) // StatFS returns filesystem statistics for the client filesystem. This // corresponds to vfs.FilesystemImpl.StatFSAt. If the client filesystem @@ -369,30 +409,30 @@ type inodeDirectory interface { HasChildren() bool // NewFile creates a new regular file inode. - NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) + NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*Dentry, error) // NewDir creates a new directory inode. - NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) + NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*Dentry, error) // NewLink creates a new hardlink to a specified inode in this // directory. Implementations should create a new kernfs Dentry pointing to // target, and update target's link count. - NewLink(ctx context.Context, name string, target Inode) (*vfs.Dentry, error) + NewLink(ctx context.Context, name string, target Inode) (*Dentry, error) // NewSymlink creates a new symbolic link inode. - NewSymlink(ctx context.Context, name, target string) (*vfs.Dentry, error) + NewSymlink(ctx context.Context, name, target string) (*Dentry, error) // NewNode creates a new filesystem node for a mknod syscall. - NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*vfs.Dentry, error) + NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*Dentry, error) // Unlink removes a child dentry from this directory inode. - Unlink(ctx context.Context, name string, child *vfs.Dentry) error + Unlink(ctx context.Context, name string, child *Dentry) error // RmDir removes an empty child directory from this directory // inode. Implementations must update the parent directory's link count, // if required. Implementations are not responsible for checking that child // is a directory, checking for an empty directory. - RmDir(ctx context.Context, name string, child *vfs.Dentry) error + RmDir(ctx context.Context, name string, child *Dentry) error // Rename is called on the source directory containing an inode being // renamed. child should point to the resolved child in the source @@ -400,7 +440,7 @@ type inodeDirectory interface { // should return the replaced dentry or nil otherwise. // // Precondition: Caller must serialize concurrent calls to Rename. - Rename(ctx context.Context, oldname, newname string, child, dstDir *vfs.Dentry) (replaced *vfs.Dentry, err error) + Rename(ctx context.Context, oldname, newname string, child, dstDir *Dentry) (replaced *Dentry, err error) } type inodeDynamicLookup interface { @@ -418,14 +458,14 @@ type inodeDynamicLookup interface { // // Lookup returns the child with an extra reference and the caller owns this // reference. - Lookup(ctx context.Context, name string) (*vfs.Dentry, error) + Lookup(ctx context.Context, name string) (*Dentry, error) // Valid should return true if this inode is still valid, or needs to // be resolved again by a call to Lookup. Valid(ctx context.Context) bool // IterDirents is used to iterate over dynamically created entries. It invokes - // cb on each entry in the directory represented by the FileDescription. + // cb on each entry in the directory represented by the Inode. // 'offset' is the offset for the entire IterDirents call, which may include // results from the caller (e.g. "." and ".."). 'relOffset' is the offset // inside the entries returned by this IterDirents invocation. In other words, @@ -437,7 +477,7 @@ type inodeDynamicLookup interface { type inodeSymlink interface { // Readlink returns the target of a symbolic link. If an inode is not a // symlink, the implementation should return EINVAL. - Readlink(ctx context.Context) (string, error) + Readlink(ctx context.Context, mnt *vfs.Mount) (string, error) // Getlink returns the target of a symbolic link, as used by path // resolution: diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go index 09806a3f2..e413242dc 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -121,8 +121,8 @@ func (fs *filesystem) newReadonlyDir(creds *auth.Credentials, mode linux.FileMod return &dir.dentry } -func (d *readonlyDir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ +func (d *readonlyDir) Open(ctx context.Context, rp *vfs.ResolvingPath, kd *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), kd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndStaticEntries, }) if err != nil { @@ -162,8 +162,8 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte return &dir.dentry } -func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, kd *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), kd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndStaticEntries, }) if err != nil { @@ -176,38 +176,36 @@ func (d *dir) DecRef(context.Context) { d.dirRefs.DecRef(d.Destroy) } -func (d *dir) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) { +func (d *dir) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*kernfs.Dentry, error) { creds := auth.CredentialsFromContext(ctx) dir := d.fs.newDir(creds, opts.Mode, nil) - dirVFSD := dir.VFSDentry() - if err := d.OrderedChildren.Insert(name, dirVFSD); err != nil { + if err := d.OrderedChildren.Insert(name, dir); err != nil { dir.DecRef(ctx) return nil, err } d.IncLinks(1) - return dirVFSD, nil + return dir, nil } -func (d *dir) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) { +func (d *dir) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*kernfs.Dentry, error) { creds := auth.CredentialsFromContext(ctx) f := d.fs.newFile(creds, "") - fVFSD := f.VFSDentry() - if err := d.OrderedChildren.Insert(name, fVFSD); err != nil { + if err := d.OrderedChildren.Insert(name, f); err != nil { f.DecRef(ctx) return nil, err } - return fVFSD, nil + return f, nil } -func (*dir) NewLink(context.Context, string, kernfs.Inode) (*vfs.Dentry, error) { +func (*dir) NewLink(context.Context, string, kernfs.Inode) (*kernfs.Dentry, error) { return nil, syserror.EPERM } -func (*dir) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (*dir) NewSymlink(context.Context, string, string) (*kernfs.Dentry, error) { return nil, syserror.EPERM } -func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (*kernfs.Dentry, error) { return nil, syserror.EPERM } diff --git a/pkg/sentry/fsimpl/kernfs/symlink.go b/pkg/sentry/fsimpl/kernfs/symlink.go index 64731a3e4..58a93eaac 100644 --- a/pkg/sentry/fsimpl/kernfs/symlink.go +++ b/pkg/sentry/fsimpl/kernfs/symlink.go @@ -24,6 +24,8 @@ import ( // StaticSymlink provides an Inode implementation for symlinks that point to // a immutable target. +// +// +stateify savable type StaticSymlink struct { InodeAttrs InodeNoopRefCount @@ -51,8 +53,8 @@ func (s *StaticSymlink) Init(creds *auth.Credentials, devMajor uint32, devMinor s.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeSymlink|0777) } -// Readlink implements Inode. -func (s *StaticSymlink) Readlink(_ context.Context) (string, error) { +// Readlink implements Inode.Readlink. +func (s *StaticSymlink) Readlink(_ context.Context, _ *vfs.Mount) (string, error) { return s.target, nil } diff --git a/pkg/sentry/fsimpl/kernfs/synthetic_directory.go b/pkg/sentry/fsimpl/kernfs/synthetic_directory.go index 01ba72fa8..ea7f073eb 100644 --- a/pkg/sentry/fsimpl/kernfs/synthetic_directory.go +++ b/pkg/sentry/fsimpl/kernfs/synthetic_directory.go @@ -41,12 +41,12 @@ type syntheticDirectory struct { var _ Inode = (*syntheticDirectory)(nil) -func newSyntheticDirectory(creds *auth.Credentials, perm linux.FileMode) *vfs.Dentry { +func newSyntheticDirectory(creds *auth.Credentials, perm linux.FileMode) *Dentry { inode := &syntheticDirectory{} inode.Init(creds, 0 /* devMajor */, 0 /* devMinor */, 0 /* ino */, perm) d := &Dentry{} d.Init(inode) - return &d.vfsd + return d } func (dir *syntheticDirectory) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode) { @@ -60,8 +60,8 @@ func (dir *syntheticDirectory) Init(creds *auth.Credentials, devMajor, devMinor } // Open implements Inode.Open. -func (dir *syntheticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := NewGenericDirectoryFD(rp.Mount(), vfsd, &dir.OrderedChildren, &dir.locks, &opts, GenericDirectoryFDOptions{}) +func (dir *syntheticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := NewGenericDirectoryFD(rp.Mount(), d, &dir.OrderedChildren, &dir.locks, &opts, GenericDirectoryFDOptions{}) if err != nil { return nil, err } @@ -69,12 +69,12 @@ func (dir *syntheticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, } // NewFile implements Inode.NewFile. -func (dir *syntheticDirectory) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*Dentry, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (dir *syntheticDirectory) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*Dentry, error) { if !opts.ForSyntheticMountpoint { return nil, syserror.EPERM } @@ -87,16 +87,16 @@ func (dir *syntheticDirectory) NewDir(ctx context.Context, name string, opts vfs } // NewLink implements Inode.NewLink. -func (dir *syntheticDirectory) NewLink(ctx context.Context, name string, target Inode) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewLink(ctx context.Context, name string, target Inode) (*Dentry, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (dir *syntheticDirectory) NewSymlink(ctx context.Context, name, target string) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewSymlink(ctx context.Context, name, target string) (*Dentry, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (dir *syntheticDirectory) NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*vfs.Dentry, error) { +func (dir *syntheticDirectory) NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*Dentry, error) { return nil, syserror.EPERM } |