Misc fixes to make stat_test pass (almost)

The only test failing now requires socket which is not
available in VFS2 yet.

Updates #1198

PiperOrigin-RevId: 302976572

7 files changed, 73 insertions, 14 deletions

diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go
index 38e4cdbc5..26b492185 100644
--- a/pkg/sentry/fsimpl/gofer/filesystem.go
+++ b/pkg/sentry/fsimpl/gofer/filesystem.go
@@ -454,6 +454,9 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b
 	}
 	if fs.opts.interop != InteropModeShared {
 		parent.touchCMtime(ctx)
+		if dir {
+			parent.decLinks()
+		}
 		parent.cacheNegativeChildLocked(name)
 		parent.dirents = nil
 	}
@@ -569,8 +572,13 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.
 func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error {
 	return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string) error {
 		creds := rp.Credentials()
-		_, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID))
-		return err
+		if _, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)); err != nil {
+			return err
+		}
+		if fs.opts.interop != InteropModeShared {
+			parent.incLinks()
+		}
+		return nil
 	})
 }
 
@@ -962,6 +970,10 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa
 		oldParent.dirents = nil
 		delete(newParent.negativeChildren, newName)
 		newParent.dirents = nil
+		if renamed.isDir() {
+			oldParent.decLinks()
+			newParent.incLinks()
+		}
 	}
 	vfsObj.CommitRenameReplaceDentry(&renamed.vfsd, &newParent.vfsd, newName, replacedVFSD)
 	return nil
diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go
index 999485492..13928ce36 100644
--- a/pkg/sentry/fsimpl/gofer/gofer.go
+++ b/pkg/sentry/fsimpl/gofer/gofer.go
@@ -485,6 +485,11 @@ type dentry struct {
 	// locked to mutate it).
 	size uint64
 
+	// nlink counts the number of hard links to this dentry. It's updated and
+	// accessed using atomic operations. It's not protected by metadataMu like the
+	// other metadata fields.
+	nlink uint32
+
 	mapsMu sync.Mutex
 
 	// If this dentry represents a regular file, mappings tracks mappings of
@@ -604,6 +609,9 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma
 	if mask.BTime {
 		d.btime = dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds)
 	}
+	if mask.NLink {
+		d.nlink = uint32(attr.NLink)
+	}
 	d.vfsd.Init(d)
 
 	fs.syncMu.Lock()
@@ -645,6 +653,9 @@ func (d *dentry) updateFromP9Attrs(mask p9.AttrMask, attr *p9.Attr) {
 	if mask.BTime {
 		atomic.StoreInt64(&d.btime, dentryTimestampFromP9(attr.BTimeSeconds, attr.BTimeNanoSeconds))
 	}
+	if mask.NLink {
+		atomic.StoreUint32(&d.nlink, uint32(attr.NLink))
+	}
 	if mask.Size {
 		d.dataMu.Lock()
 		atomic.StoreUint64(&d.size, attr.Size)
@@ -687,10 +698,7 @@ func (d *dentry) fileType() uint32 {
 func (d *dentry) statTo(stat *linux.Statx) {
 	stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK | linux.STATX_UID | linux.STATX_GID | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME | linux.STATX_INO | linux.STATX_SIZE | linux.STATX_BLOCKS | linux.STATX_BTIME
 	stat.Blksize = atomic.LoadUint32(&d.blockSize)
-	stat.Nlink = 1
-	if d.isDir() {
-		stat.Nlink = 2
-	}
+	stat.Nlink = atomic.LoadUint32(&d.nlink)
 	stat.UID = atomic.LoadUint32(&d.uid)
 	stat.GID = atomic.LoadUint32(&d.gid)
 	stat.Mode = uint16(atomic.LoadUint32(&d.mode))
@@ -703,7 +711,7 @@ func (d *dentry) statTo(stat *linux.Statx) {
 	stat.Btime = statxTimestampFromDentry(atomic.LoadInt64(&d.btime))
 	stat.Ctime = statxTimestampFromDentry(atomic.LoadInt64(&d.ctime))
 	stat.Mtime = statxTimestampFromDentry(atomic.LoadInt64(&d.mtime))
-	// TODO(jamieliu): device number
+	// TODO(gvisor.dev/issue/1198): device number
 }
 
 func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, stat *linux.Statx, mnt *vfs.Mount) error {
@@ -1094,6 +1102,26 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool
 	return nil
 }
 
+// incLinks increments link count.
+//
+// Preconditions: d.nlink != 0 && d.nlink < math.MaxUint32.
+func (d *dentry) incLinks() {
+	v := atomic.AddUint32(&d.nlink, 1)
+	if v < 2 {
+		panic(fmt.Sprintf("dentry.nlink is invalid (was 0 or overflowed): %d", v))
+	}
+}
+
+// decLinks decrements link count.
+//
+// Preconditions: d.nlink > 1.
+func (d *dentry) decLinks() {
+	v := atomic.AddUint32(&d.nlink, ^uint32(0))
+	if v == 0 {
+		panic(fmt.Sprintf("dentry.nlink must be greater than 0: %d", v))
+	}
+}
+
 // fileDescription is embedded by gofer implementations of
 // vfs.FileDescriptionImpl.
 type fileDescription struct {
@@ -1112,7 +1140,8 @@ func (fd *fileDescription) dentry() *dentry {
 // Stat implements vfs.FileDescriptionImpl.Stat.
 func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) {
 	d := fd.dentry()
-	if d.fs.opts.interop == InteropModeShared && opts.Mask&(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE|linux.STATX_BLOCKS|linux.STATX_BTIME) != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC {
+	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.fs.opts.interop == InteropModeShared && opts.Mask&(validMask) != 0 && opts.Sync != linux.AT_STATX_DONT_SYNC {
 		// TODO(jamieliu): Use specialFileFD.handle.file for the getattr if
 		// available?
 		if err := d.updateFromGetattr(ctx); err != nil {
diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD
index e7695e995..2eb210014 100644
--- a/pkg/sentry/syscalls/linux/vfs2/BUILD
+++ b/pkg/sentry/syscalls/linux/vfs2/BUILD
@@ -31,6 +31,7 @@ go_library(
     visibility = ["//:sandbox"],
     deps = [
         "//pkg/abi/linux",
+        "//pkg/bits",
         "//pkg/fspath",
         "//pkg/gohacks",
         "//pkg/sentry/arch",
diff --git a/pkg/sentry/syscalls/linux/vfs2/filesystem.go b/pkg/sentry/syscalls/linux/vfs2/filesystem.go
index fc5ceea4c..a859095e2 100644
--- a/pkg/sentry/syscalls/linux/vfs2/filesystem.go
+++ b/pkg/sentry/syscalls/linux/vfs2/filesystem.go
@@ -250,7 +250,7 @@ func rmdirat(t *kernel.Task, dirfd int32, pathAddr usermem.Addr) error {
 	if err != nil {
 		return err
 	}
-	tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, followFinalSymlink)
+	tpop, err := getTaskPathOperation(t, dirfd, path, disallowEmptyPath, nofollowFinalSymlink)
 	if err != nil {
 		return err
 	}
diff --git a/pkg/sentry/syscalls/linux/vfs2/getdents.go b/pkg/sentry/syscalls/linux/vfs2/getdents.go
index ddc140b65..a61cc5059 100644
--- a/pkg/sentry/syscalls/linux/vfs2/getdents.go
+++ b/pkg/sentry/syscalls/linux/vfs2/getdents.go
@@ -97,7 +97,7 @@ func (cb *getdentsCallback) Handle(dirent vfs.Dirent) error {
 		//     char           d_name[]; /* Filename (null-terminated) */
 		// };
 		size := 8 + 8 + 2 + 1 + 1 + len(dirent.Name)
-		if size < cb.remaining {
+		if size > cb.remaining {
 			return syserror.EINVAL
 		}
 		buf = cb.t.CopyScratchBuffer(size)
@@ -125,7 +125,7 @@ func (cb *getdentsCallback) Handle(dirent vfs.Dirent) error {
 			panic(fmt.Sprintf("unsupported sizeof(unsigned long): %d", cb.t.Arch().Width()))
 		}
 		size := 8 + 8 + 2 + 1 + 1 + 1 + len(dirent.Name)
-		if size < cb.remaining {
+		if size > cb.remaining {
 			return syserror.EINVAL
 		}
 		buf = cb.t.CopyScratchBuffer(size)
diff --git a/pkg/sentry/syscalls/linux/vfs2/stat.go b/pkg/sentry/syscalls/linux/vfs2/stat.go
index 068243132..fdfe49243 100644
--- a/pkg/sentry/syscalls/linux/vfs2/stat.go
+++ b/pkg/sentry/syscalls/linux/vfs2/stat.go
@@ -16,6 +16,7 @@ package vfs2
 
 import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/bits"
 	"gvisor.dev/gvisor/pkg/fspath"
 	"gvisor.dev/gvisor/pkg/gohacks"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
@@ -153,7 +154,11 @@ func Statx(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 	if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW|linux.AT_STATX_SYNC_TYPE) != 0 {
 		return 0, nil, syserror.EINVAL
 	}
-
+	// Make sure that only one sync type option is set.
+	syncType := uint32(flags & linux.AT_STATX_SYNC_TYPE)
+	if syncType != 0 && !bits.IsPowerOfTwo32(syncType) {
+		return 0, nil, syserror.EINVAL
+	}
 	if mask&linux.STATX__RESERVED != 0 {
 		return 0, nil, syserror.EINVAL
 	}
diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go
index eb4ebb511..8f31495da 100644
--- a/pkg/sentry/vfs/resolving_path.go
+++ b/pkg/sentry/vfs/resolving_path.go
@@ -329,10 +329,22 @@ func (rp *ResolvingPath) ResolveComponent(d *Dentry) (*Dentry, error) {
 // component in pcs represents a symbolic link, the symbolic link should be
 // followed.
 //
+// If path is terminated with '/', the '/' is considered the last element and
+// any symlink before that is followed:
+//   - For most non-creating walks, the last path component is handled by
+//     fs/namei.c:lookup_last(), which sets LOOKUP_FOLLOW if the first byte
+//     after the path component is non-NULL (which is only possible if it's '/')
+//     and the path component is of type LAST_NORM.
+//
+//   - For open/openat/openat2 without O_CREAT, the last path component is
+//     handled by fs/namei.c:do_last(), which does the same, though without the
+//     LAST_NORM check.
+//
 // Preconditions: !rp.Done().
 func (rp *ResolvingPath) ShouldFollowSymlink() bool {
-	// Non-final symlinks are always followed.
-	return rp.flags&rpflagsFollowFinalSymlink != 0 || !rp.Final()
+	// Non-final symlinks are always followed. Paths terminated with '/' are also
+	// always followed.
+	return rp.flags&rpflagsFollowFinalSymlink != 0 || !rp.Final() || rp.MustBeDir()
 }
 
 // HandleSymlink is called when the current path component is a symbolic link