Complete vfs2 implementation of fallocate.

This change includes overlay, special regular gofer files, and hostfs.

Fixes #3589.

PiperOrigin-RevId: 332330860

pkg/sentry/fsimpl/gofer/filesystem.go

@@ -1026,7 +1026,7 @@ func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.Open
 		// step is required even if !d.cachedMetadataAuthoritative() because
 		// d.mappings has to be updated.
 		// d.metadataMu has already been acquired if trunc == true.
-		d.updateFileSizeLocked(0)
+		d.updateSizeLocked(0)
 
 		if d.cachedMetadataAuthoritative() {
 			d.touchCMtimeLocked()

pkg/sentry/fsimpl/gofer/gofer.go

@@ -833,7 +833,7 @@ func (d *dentry) updateFromP9AttrsLocked(mask p9.AttrMask, attr *p9.Attr) {
 		atomic.StoreUint32(&d.nlink, uint32(attr.NLink))
 	}
 	if mask.Size {
-		d.updateFileSizeLocked(attr.Size)
+		d.updateSizeLocked(attr.Size)
 	}
 }
 
@@ -987,7 +987,7 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs
 				// 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.updateFileSizeLocked(stat.Size)
+				d.updateSizeLocked(stat.Size)
 			}
 		}
 		if d.fs.opts.interop == InteropModeShared {
@@ -1024,8 +1024,31 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs
 	return nil
 }
 
+// doAllocate performs an allocate operation on d. Note that d.metadataMu will
+// be held when allocate is called.
+func (d *dentry) doAllocate(ctx context.Context, offset, length uint64, allocate func() error) error {
+	d.metadataMu.Lock()
+	defer d.metadataMu.Unlock()
+
+	// Allocating a smaller size is a noop.
+	size := offset + length
+	if d.cachedMetadataAuthoritative() && size <= d.size {
+		return nil
+	}
+
+	err := allocate()
+	if err != nil {
+		return err
+	}
+	d.updateSizeLocked(size)
+	if d.cachedMetadataAuthoritative() {
+		d.touchCMtimeLocked()
+	}
+	return nil
+}
+
 // Preconditions: d.metadataMu must be locked.
-func (d *dentry) updateFileSizeLocked(newSize uint64) {
+func (d *dentry) updateSizeLocked(newSize uint64) {
 	d.dataMu.Lock()
 	oldSize := d.size
 	atomic.StoreUint64(&d.size, newSize)

pkg/sentry/fsimpl/gofer/regular_file.go

@@ -79,28 +79,11 @@ func (fd *regularFileFD) OnClose(ctx context.Context) error {
 // Allocate implements vfs.FileDescriptionImpl.Allocate.
 func (fd *regularFileFD) Allocate(ctx context.Context, mode, offset, length uint64) error {
 	d := fd.dentry()
-	d.metadataMu.Lock()
-	defer d.metadataMu.Unlock()
-
-	// Allocating a smaller size is a noop.
-	size := offset + length
-	if d.cachedMetadataAuthoritative() && size <= d.size {
-		return nil
-	}
-
-	d.handleMu.RLock()
-	err := d.writeFile.allocate(ctx, p9.ToAllocateMode(mode), offset, length)
-	d.handleMu.RUnlock()
-	if err != nil {
-		return err
-	}
-	d.dataMu.Lock()
-	atomic.StoreUint64(&d.size, size)
-	d.dataMu.Unlock()
-	if d.cachedMetadataAuthoritative() {
-		d.touchCMtimeLocked()
-	}
-	return nil
+	return d.doAllocate(ctx, offset, length, func() error {
+		d.handleMu.RLock()
+		defer d.handleMu.RUnlock()
+		return d.writeFile.allocate(ctx, p9.ToAllocateMode(mode), offset, length)
+	})
 }
 
 // PRead implements vfs.FileDescriptionImpl.PRead.

pkg/sentry/fsimpl/gofer/special_file.go

@@ -22,6 +22,7 @@ import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/fdnotifier"
+	"gvisor.dev/gvisor/pkg/p9"
 	"gvisor.dev/gvisor/pkg/safemem"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/syserror"
@@ -135,6 +136,16 @@ func (fd *specialFileFD) EventUnregister(e *waiter.Entry) {
 	fd.fileDescription.EventUnregister(e)
 }
 
+func (fd *specialFileFD) Allocate(ctx context.Context, mode, offset, length uint64) error {
+	if fd.isRegularFile {
+		d := fd.dentry()
+		return d.doAllocate(ctx, offset, length, func() error {
+			return fd.handle.file.allocate(ctx, p9.ToAllocateMode(mode), offset, length)
+		})
+	}
+	return fd.FileDescriptionDefaultImpl.Allocate(ctx, mode, offset, length)
+}
+
 // PRead implements vfs.FileDescriptionImpl.PRead.
 func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
 	if fd.seekable && offset < 0 {

pkg/sentry/fsimpl/host/host.go

@@ -560,12 +560,7 @@ func (f *fileDescription) Release(context.Context) {
 
 // Allocate implements vfs.FileDescriptionImpl.
 func (f *fileDescription) Allocate(ctx context.Context, mode, offset, length uint64) error {
-	if !f.inode.seekable {
-		return syserror.ESPIPE
-	}
-
-	// TODO(gvisor.dev/issue/3589): Implement Allocate for non-pipe hostfds.
-	return syserror.EOPNOTSUPP
+	return unix.Fallocate(f.inode.hostFD, uint32(mode), int64(offset), int64(length))
 }
 
 // PRead implements FileDescriptionImpl.

pkg/sentry/fsimpl/overlay/non_directory.go

@@ -147,6 +147,16 @@ func (fd *nonDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux
 	return stat, nil
 }
 
+// Allocate implements vfs.FileDescriptionImpl.Allocate.
+func (fd *nonDirectoryFD) Allocate(ctx context.Context, mode, offset, length uint64) error {
+	wrappedFD, err := fd.getCurrentFD(ctx)
+	if err != nil {
+		return err
+	}
+	defer wrappedFD.DecRef(ctx)
+	return wrappedFD.Allocate(ctx, mode, offset, length)
+}
+
 // SetStat implements vfs.FileDescriptionImpl.SetStat.
 func (fd *nonDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error {
 	d := fd.dentry()

pkg/sentry/kernel/pipe/vfs.go

@@ -67,6 +67,11 @@ func (vp *VFSPipe) ReaderWriterPair(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlag
 	return vp.newFD(mnt, vfsd, linux.O_RDONLY|statusFlags, locks), vp.newFD(mnt, vfsd, linux.O_WRONLY|statusFlags, locks)
 }
 
+// Allocate implements vfs.FileDescriptionImpl.Allocate.
+func (*VFSPipe) Allocate(context.Context, uint64, uint64, uint64) error {
+	return syserror.ESPIPE
+}
+
 // Open opens the pipe represented by vp.
 func (vp *VFSPipe) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32, locks *vfs.FileLocks) (*vfs.FileDescription, error) {
 	vp.mu.Lock()

pkg/sentry/socket/hostinet/socket_vfs2.go

@@ -97,11 +97,6 @@ func (s *socketVFS2) Ioctl(ctx context.Context, uio usermem.IO, args arch.Syscal
 	return ioctl(ctx, s.fd, uio, args)
 }
 
-// Allocate implements vfs.FileDescriptionImpl.Allocate.
-func (s *socketVFS2) Allocate(ctx context.Context, mode, offset, length uint64) error {
-	return syserror.ENODEV
-}
-
 // PRead implements vfs.FileDescriptionImpl.PRead.
 func (s *socketVFS2) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
 	return 0, syserror.ESPIPE

pkg/sentry/vfs/file_description.go

@@ -326,6 +326,9 @@ type FileDescriptionImpl interface {
 	// Allocate grows the file to offset + length bytes.
 	// Only mode == 0 is supported currently.
 	//
+	// Allocate should return EISDIR on directories, ESPIPE on pipes, and ENODEV on
+	// other files where it is not supported.
+	//
 	// Preconditions: The FileDescription was opened for writing.
 	Allocate(ctx context.Context, mode, offset, length uint64) error
 

pkg/sentry/vfs/file_description_impl_util.go

@@ -57,7 +57,11 @@ func (FileDescriptionDefaultImpl) StatFS(ctx context.Context) (linux.Statfs, err
 }
 
 // Allocate implements FileDescriptionImpl.Allocate analogously to
-// fallocate called on regular file, directory or FIFO in Linux.
+// fallocate called on an invalid type of file in Linux.
+//
+// Note that directories can rely on this implementation even though they
+// should technically return EISDIR. Allocate should never be called for a
+// directory, because it requires a writable fd.
 func (FileDescriptionDefaultImpl) Allocate(ctx context.Context, mode, offset, length uint64) error {
 	return syserror.ENODEV
 }

test/syscalls/linux/fallocate.cc

@@ -179,6 +179,12 @@ TEST_F(AllocateTest, FallocateOtherFDs) {
   auto sock0 = FileDescriptor(socks[0]);
   auto sock1 = FileDescriptor(socks[1]);
   EXPECT_THAT(fallocate(sock0.get(), 0, 0, 10), SyscallFailsWithErrno(ENODEV));
+
+  int pipefds[2];
+  ASSERT_THAT(pipe(pipefds), SyscallSucceeds());
+  EXPECT_THAT(fallocate(pipefds[1], 0, 0, 10), SyscallFailsWithErrno(ESPIPE));
+  close(pipefds[0]);
+  close(pipefds[1]);
 }
 
 }  // namespace