erofs: add checkpoint/restore support

Updates #8956

Signed-off-by: Tiwei Bie <tiwei.btw@antgroup.com>

pkg/erofs/erofs.go

@@ -223,6 +223,25 @@ func OpenImage(src *os.File) (*Image, error) {
 	return i, nil
 }
 
+// UpdateImage updates the underlying image file. This is typically used in checkpoint/restore.
+//
+// On success, the ownership of src is transferred to Image.
+//
+// Preconditions:
+//   - i.src == nil.
+//   - i.bytes == nil.
+func (i *Image) UpdateImage(src *os.File) error {
+	newImage, err := OpenImage(src)
+	if err != nil {
+		return err
+	}
+	if newImage.sb != i.sb {
+		return fmt.Errorf("superblock mismatch detected, got %+v, expected %+v", newImage.sb, i.sb)
+	}
+	*i = *newImage
+	return nil
+}
+
 // Close closes the image.
 func (i *Image) Close() {
 	unix.Munmap(i.bytes)

pkg/sentry/fsimpl/erofs/erofs.go

@@ -28,6 +28,7 @@ import (
 	"gvisor.dev/gvisor/pkg/erofs"
 	"gvisor.dev/gvisor/pkg/errors/linuxerr"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/memmap"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 )
 
@@ -51,6 +52,7 @@ type filesystem struct {
 
 	// Immutable options.
 	mopts string
+	iopts InternalFilesystemOptions
 
 	// devMinor is the filesystem's minor device number. devMinor is immutable.
 	devMinor uint32
@@ -70,6 +72,16 @@ type filesystem struct {
 	inodeBuckets []inodeBucket
 }
 
+// InternalFilesystemOptions may be passed as
+// vfs.GetFilesystemOptions.InternalData to FilesystemType.GetFilesystem.
+//
+// +stateify savable
+type InternalFilesystemOptions struct {
+	// If UniqueID is non-empty, it is an opaque string used to reassociate the
+	// filesystem with a new image FD during restoration from checkpoint.
+	UniqueID string
+}
+
 // Name implements vfs.FilesystemType.Name.
 func (FilesystemType) Name() string {
 	return Name
@@ -98,6 +110,12 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 	}
 	cu.Add(func() { image.Close() })
 
+	iopts, ok := opts.InternalData.(InternalFilesystemOptions)
+	if opts.InternalData != nil && !ok {
+		ctx.Warningf("erofs.FilesystemType.GetFilesystem: GetFilesystemOptions.InternalData has type %T, wanted erofs.InternalFilesystemOptions", opts.InternalData)
+		return nil, nil, linuxerr.EINVAL
+	}
+
 	devMinor, err := vfsObj.GetAnonBlockDevMinor()
 	if err != nil {
 		return nil, nil, err
@@ -105,6 +123,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 
 	fs := &filesystem{
 		mopts:    opts.Data,
+		iopts:    iopts,
 		image:    image,
 		devMinor: devMinor,
 		mf:       imageMemmapFile{image: image},
@@ -243,6 +262,14 @@ type inode struct {
 	// +checklocks:dirMu
 	dirents []vfs.Dirent `state:"nosave"`
 
+	// mapsMu protects mappings.
+	mapsMu sync.Mutex `state:"nosave"`
+
+	// mappings tracks the mappings of the file into memmap.MappingSpaces
+	// if this inode represents a regular file.
+	// +checklocks:mapsMu
+	mappings memmap.MappingSet
+
 	// locks supports POSIX and BSD style locks.
 	locks vfs.FileLocks
 

pkg/sentry/fsimpl/erofs/regular_file.go

@@ -131,15 +131,22 @@ func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpt
 
 // AddMapping implements memmap.Mappable.AddMapping.
 func (i *inode) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar hostarch.AddrRange, offset uint64, writable bool) error {
+	i.mapsMu.Lock()
+	i.mappings.AddMapping(ms, ar, offset, writable)
+	i.mapsMu.Unlock()
 	return nil
 }
 
 // RemoveMapping implements memmap.Mappable.RemoveMapping.
 func (i *inode) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar hostarch.AddrRange, offset uint64, writable bool) {
+	i.mapsMu.Lock()
+	i.mappings.RemoveMapping(ms, ar, offset, writable)
+	i.mapsMu.Unlock()
 }
 
 // CopyMapping implements memmap.Mappable.CopyMapping.
 func (i *inode) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR hostarch.AddrRange, offset uint64, writable bool) error {
+	i.AddMapping(ctx, ms, dstAR, offset, writable)
 	return nil
 }
 
@@ -175,6 +182,9 @@ func (i *inode) Translate(ctx context.Context, required, optional memmap.Mappabl
 
 // InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
 func (i *inode) InvalidateUnsavable(ctx context.Context) error {
+	i.mapsMu.Lock()
+	i.mappings.InvalidateAll(memmap.InvalidateOpts{})
+	i.mapsMu.Unlock()
 	return nil
 }
 

pkg/sentry/fsimpl/erofs/save_restore.go

@@ -14,7 +14,36 @@
 
 package erofs
 
-// TODO: support checkpoint/restore.
+import (
+	"fmt"
+	"os"
+
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+)
+
+// Compile-time assertion that filesystem implements vfs.FilesystemImplSaveRestoreExtension.
+var _ = vfs.FilesystemImplSaveRestoreExtension((*filesystem)(nil))
+
+// PreprareSave implements vfs.FilesystemImplSaveRestoreExtension.PrepareSave.
+func (fs *filesystem) PrepareSave(ctx context.Context) error {
+	return nil
+}
+
+// CompleteRestore implements
+// vfs.FilesystemImplSaveRestoreExtension.CompleteRestore.
+func (fs *filesystem) CompleteRestore(ctx context.Context, opts vfs.CompleteRestoreOptions) error {
+	fdmapv := ctx.Value(vfs.CtxRestoreFilesystemFDMap)
+	if fdmapv == nil {
+		return fmt.Errorf("no image FD map available")
+	}
+	fdmap := fdmapv.(map[string]int)
+	fd, ok := fdmap[fs.iopts.UniqueID]
+	if !ok {
+		return fmt.Errorf("no image FD available for filesystem with unique ID %q", fs.iopts.UniqueID)
+	}
+	return fs.image.UpdateImage(os.NewFile(uintptr(fd), "EROFS image file"))
+}
 
 // saveParent is called by stateify.
 func (d *dentry) saveParent() *dentry {