Create separate MemoryFile metadata file in compression=none mode.

This helps in parallelizing MemoryFile restore with kernel restore.

PiperOrigin-RevId: 629378902

pkg/sentry/control/state.go

@@ -44,12 +44,14 @@ type SaveOpts struct {
 	// Metadata is the set of metadata to prepend to the state file.
 	Metadata map[string]string `json:"metadata"`
 
-	// HavePagesFile indicates whether the checkpoint pages file is provided.
+	// HavePagesFile indicates whether the pages file and its corresponding
+	// metadata file is provided.
 	HavePagesFile bool `json:"have_pages_file"`
 
 	// FilePayload contains the following:
 	// 1. checkpoint state file.
-	// 2. optional checkpoint pages file.
+	// 2. optional checkpoint pages metadata file.
+	// 3. optional checkpoint pages file.
 	urpc.FilePayload
 
 	// Resume indicates if the sandbox process should continue running
@@ -61,15 +63,20 @@ type SaveOpts struct {
 func (s *State) Save(o *SaveOpts, _ *struct{}) error {
 	wantFiles := 1
 	if o.HavePagesFile {
-		wantFiles++
+		wantFiles += 2
 	}
 	if gotFiles := len(o.FilePayload.Files); gotFiles != wantFiles {
 		return fmt.Errorf("got %d files, wanted %d", gotFiles, wantFiles)
 	}
 
 	// Save to the first provided stream.
+	stateFile, err := o.ReleaseFD(0)
+	if err != nil {
+		return err
+	}
+	defer stateFile.Close()
 	saveOpts := state.SaveOpts{
-		Destination: o.FilePayload.Files[0],
+		Destination: stateFile,
 		Key:         o.Key,
 		Metadata:    o.Metadata,
 		Callback: func(err error) {
@@ -85,9 +92,17 @@ func (s *State) Save(o *SaveOpts, _ *struct{}) error {
 			}
 		},
 	}
-	defer o.FilePayload.Files[0].Close()
 	if o.HavePagesFile {
-		saveOpts.PagesFile = o.FilePayload.Files[1]
+		saveOpts.PagesMetadata, err = o.ReleaseFD(1)
+		if err != nil {
+			return err
+		}
+		defer saveOpts.PagesMetadata.Close()
+
+		saveOpts.PagesFile, err = o.ReleaseFD(2)
+		if err != nil {
+			return err
+		}
 		defer saveOpts.PagesFile.Close()
 	}
 	return saveOpts.Save(s.Kernel.SupervisorContext(), s.Kernel, s.Watchdog)

pkg/sentry/kernel/kernel.go

@@ -35,7 +35,6 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"os"
 	"path/filepath"
 	"time"
 
@@ -581,7 +580,7 @@ func loadPrivateMFs(ctx context.Context, r io.Reader, pr *statefile.AsyncReader)
 // SaveTo saves the state of k to w.
 //
 // Preconditions: The kernel must be paused throughout the call to SaveTo.
-func (k *Kernel) SaveTo(ctx context.Context, w io.Writer, pagesFile *os.File) error {
+func (k *Kernel) SaveTo(ctx context.Context, w io.Writer, pagesMetadata, pagesFile *fd.FD) error {
 	saveStart := time.Now()
 
 	// Do not allow other Kernel methods to affect it while it's being saved.
@@ -649,14 +648,18 @@ func (k *Kernel) SaveTo(ctx context.Context, w io.Writer, pagesFile *os.File) er
 
 	// Save the memory files' state.
 	memoryStart := time.Now()
-	pw := io.Writer(w)
+	pmw := w
+	if pagesMetadata != nil {
+		pmw = pagesMetadata
+	}
+	pw := w
 	if pagesFile != nil {
 		pw = pagesFile
 	}
-	if err := k.mf.SaveTo(ctx, w, pw); err != nil {
+	if err := k.mf.SaveTo(ctx, pmw, pw); err != nil {
 		return err
 	}
-	if err := savePrivateMFs(ctx, w, pw, mfsToSave); err != nil {
+	if err := savePrivateMFs(ctx, pmw, pw, mfsToSave); err != nil {
 		return err
 	}
 	log.Infof("Memory files save took [%s].", time.Since(memoryStart))
@@ -692,7 +695,7 @@ func (k *Kernel) invalidateUnsavableMappings(ctx context.Context) error {
 }
 
 // LoadFrom returns a new Kernel loaded from args.
-func (k *Kernel) LoadFrom(ctx context.Context, r io.Reader, pagesFile *fd.FD, timeReady chan struct{}, net inet.Stack, clocks sentrytime.Clocks, vfsOpts *vfs.CompleteRestoreOptions) error {
+func (k *Kernel) LoadFrom(ctx context.Context, r io.Reader, pagesMetadata, pagesFile *fd.FD, timeReady chan struct{}, net inet.Stack, clocks sentrytime.Clocks, vfsOpts *vfs.CompleteRestoreOptions) error {
 	loadStart := time.Now()
 
 	k.runningTasksCond.L = &k.runningTasksMu
@@ -734,14 +737,18 @@ func (k *Kernel) LoadFrom(ctx context.Context, r io.Reader, pagesFile *fd.FD, ti
 
 	// Load the memory files' state.
 	memoryStart := time.Now()
+	pmr := r
+	if pagesMetadata != nil {
+		pmr = pagesMetadata
+	}
 	var pr *statefile.AsyncReader
 	if pagesFile != nil {
 		pr = statefile.NewAsyncReader(pagesFile, 0 /* off */)
 	}
-	if err := k.mf.LoadFrom(ctx, r, pr); err != nil {
+	if err := k.mf.LoadFrom(ctx, pmr, pr); err != nil {
 		return err
 	}
-	if err := loadPrivateMFs(ctx, r, pr); err != nil {
+	if err := loadPrivateMFs(ctx, pmr, pr); err != nil {
 		return err
 	}
 	if pr != nil {

pkg/sentry/state/state.go

@@ -18,7 +18,6 @@ package state
 import (
 	"fmt"
 	"io"
-	"os"
 
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/errors/linuxerr"
@@ -50,9 +49,13 @@ type SaveOpts struct {
 	// Destination is the save target.
 	Destination io.Writer
 
+	// PagesMetadata is the file into which MemoryFile metadata is stored if
+	// PagesMetadata is non-nil. Otherwise this content is stored in Destination.
+	PagesMetadata *fd.FD
+
 	// PagesFile is the file in which all MemoryFile pages are stored if
-	// PagesFile is non-nil.
-	PagesFile *os.File
+	// PagesFile is non-nil. Otherwise this content is stored in Destination.
+	PagesFile *fd.FD
 
 	// Key is used for state integrity check.
 	Key []byte
@@ -92,7 +95,7 @@ func (opts SaveOpts) Save(ctx context.Context, k *kernel.Kernel, w *watchdog.Wat
 		err = ErrStateFile{err}
 	} else {
 		// Save the kernel.
-		err = k.SaveTo(ctx, wc, opts.PagesFile)
+		err = k.SaveTo(ctx, wc, opts.PagesMetadata, opts.PagesFile)
 
 		// ENOSPC is a state file error. This error can only come from
 		// writing the state file, and not from fs.FileOperations.Fsync
@@ -111,11 +114,15 @@ func (opts SaveOpts) Save(ctx context.Context, k *kernel.Kernel, w *watchdog.Wat
 
 // LoadOpts contains load-related options.
 type LoadOpts struct {
-	// Destination is the load source.
+	// Source is the load source.
 	Source io.Reader
 
+	// PagesMetadata is the file into which MemoryFile metadata is stored if
+	// PagesMetadata is non-nil. Otherwise this content is stored in Source.
+	PagesMetadata *fd.FD
+
 	// PagesFile is the file in which all MemoryFile pages are stored if
-	// PagesFile is non-nil.
+	// PagesFile is non-nil. Otherwise this content is stored in Source.
 	PagesFile *fd.FD
 
 	// Key is used for state integrity check.
@@ -133,5 +140,5 @@ func (opts LoadOpts) Load(ctx context.Context, k *kernel.Kernel, timeReady chan
 	previousMetadata = m
 
 	// Restore the Kernel object graph.
-	return k.LoadFrom(ctx, r, opts.PagesFile, timeReady, n, clocks, vfsOpts)
+	return k.LoadFrom(ctx, r, opts.PagesMetadata, opts.PagesFile, timeReady, n, clocks, vfsOpts)
 }

runsc/boot/controller.go

@@ -453,8 +453,9 @@ func (cm *containerManager) PortForward(opts *PortForwardOpts, _ *struct{}) erro
 type RestoreOpts struct {
 	// FilePayload contains the state file to be restored, followed in order by:
 	// 1. checkpoint state file.
-	// 2. optional checkpoint pages file.
-	// 3. optional platform device file.
+	// 2. optional checkpoint pages metadata file.
+	// 3. optional checkpoint pages file.
+	// 4. optional platform device file.
 	urpc.FilePayload
 	HavePagesFile  bool
 	HaveDeviceFile bool
@@ -496,12 +497,19 @@ func (cm *containerManager) Restore(o *RestoreOpts, _ *struct{}) error {
 
 	fileIdx := 1
 	if o.HavePagesFile {
-		pagesFile, err := o.ReleaseFD(fileIdx)
+		cm.restorer.pagesMetadata, err = o.ReleaseFD(fileIdx)
 		if err != nil {
 			return err
 		}
+		defer cm.restorer.pagesMetadata.Close()
+		fileIdx++
+
+		cm.restorer.pagesFile, err = o.ReleaseFD(fileIdx)
+		if err != nil {
+			return err
+		}
+		defer cm.restorer.pagesFile.Close()
 		fileIdx++
-		cm.restorer.pagesFile = pagesFile
 	}
 
 	if o.HaveDeviceFile {

runsc/boot/restore.go

@@ -48,6 +48,9 @@ const (
 	// CheckpointStateFileName is the file within the given image-path's
 	// directory which contains the container's saved state.
 	CheckpointStateFileName = "checkpoint.img"
+	// CheckpointPagesMetadataFileName is the file within the given image-path's
+	// directory containing the container's MemoryFile metadata.
+	CheckpointPagesMetadataFileName = "pages_meta.img"
 	// CheckpointPagesFileName is the file within the given image-path's
 	// directory containing the container's MemoryFile pages.
 	CheckpointPagesFileName = "pages.img"
@@ -68,8 +71,9 @@ type restorer struct {
 	containers []*containerInfo
 
 	// Files used by restore to rehydrate the state.
-	stateFile io.ReadCloser
-	pagesFile *fd.FD
+	stateFile     io.ReadCloser
+	pagesMetadata *fd.FD
+	pagesFile     *fd.FD
 
 	// deviceFile is the required to start the platform.
 	deviceFile *fd.FD
@@ -224,7 +228,7 @@ func (r *restorer) restore(l *Loader) error {
 	ctx = context.WithValue(ctx, pgalloc.CtxMemoryFileMap, mfmap)
 
 	// Load the state.
-	loadOpts := state.LoadOpts{Source: r.stateFile, PagesFile: r.pagesFile}
+	loadOpts := state.LoadOpts{Source: r.stateFile, PagesMetadata: r.pagesMetadata, PagesFile: r.pagesFile}
 	if err := loadOpts.Load(ctx, l.k, nil, netns.Stack(), time.NewCalibratedClocks(), &vfs.CompleteRestoreOptions{}); err != nil {
 		return err
 	}

runsc/sandbox/sandbox.go

@@ -468,8 +468,14 @@ func (s *Sandbox) Restore(conf *config.Config, cid string, imagePath string, dir
 	}
 	if pf, err := os.OpenFile(pagesFileName, pagesReadFlags, 0); err == nil {
 		defer pf.Close()
+		pagesMetadataFileName := path.Join(imagePath, boot.CheckpointPagesMetadataFileName)
+		pmf, err := os.Open(pagesMetadataFileName)
+		if err != nil {
+			return fmt.Errorf("opening restore image file %q failed: %v", pagesMetadataFileName, err)
+		}
+		defer pmf.Close()
 		opt.HavePagesFile = true
-		opt.FilePayload.Files = append(opt.FilePayload.Files, pf)
+		opt.FilePayload.Files = append(opt.FilePayload.Files, pmf, pf)
 	} else if !os.IsNotExist(err) {
 		return fmt.Errorf("opening restore image file %q failed: %v", pagesFileName, err)
 	}
@@ -1341,7 +1347,13 @@ func (s *Sandbox) Checkpoint(cid string, imagePath string, options statefile.Opt
 			return fmt.Errorf("creating checkpoint pages file %q: %w", pagesFilePath, err)
 		}
 		defer pf.Close()
-		opt.FilePayload.Files = append(opt.FilePayload.Files, pf)
+		pagesMetadataFilePath := filepath.Join(imagePath, boot.CheckpointPagesMetadataFileName)
+		pmf, err := os.OpenFile(pagesMetadataFilePath, os.O_CREATE|os.O_EXCL|os.O_RDWR, 0644)
+		if err != nil {
+			return fmt.Errorf("creating checkpoint pages metadata file %q: %w", pagesMetadataFilePath, err)
+		}
+		defer pmf.Close()
+		opt.FilePayload.Files = append(opt.FilePayload.Files, pmf, pf)
 		opt.HavePagesFile = true
 	}