Implement PRead for verity fs

PRead is implemented by read from the underlying file in blocks, and
verify each block. The verified contents are saved into the output
buffer.

PiperOrigin-RevId: 332092267

pkg/merkletree/merkletree.go

@@ -225,9 +225,9 @@ func Generate(data io.ReadSeeker, dataSize int64, treeReader io.ReadSeeker, tree
 // Verify will modify the cursor for data, but always restores it to its
 // original position upon exit. The cursor for tree is modified and not
 // restored.
-func Verify(w io.Writer, data, tree io.ReadSeeker, dataSize int64, readOffset int64, readSize int64, expectedRoot []byte, dataAndTreeInSameFile bool) error {
+func Verify(w io.Writer, data, tree io.ReadSeeker, dataSize int64, readOffset int64, readSize int64, expectedRoot []byte, dataAndTreeInSameFile bool) (int64, error) {
 	if readSize <= 0 {
-		return fmt.Errorf("Unexpected read size: %d", readSize)
+		return 0, fmt.Errorf("Unexpected read size: %d", readSize)
 	}
 	layout := InitLayout(int64(dataSize), dataAndTreeInSameFile)
 
@@ -240,29 +240,30 @@ func Verify(w io.Writer, data, tree io.ReadSeeker, dataSize int64, readOffset in
 	// finishes.
 	origOffset, err := data.Seek(0, io.SeekCurrent)
 	if err != nil {
-		return fmt.Errorf("Find current data offset failed: %v", err)
+		return 0, fmt.Errorf("Find current data offset failed: %v", err)
 	}
 	defer data.Seek(origOffset, io.SeekStart)
 
 	// Move to the first block that contains target data.
 	if _, err := data.Seek(firstDataBlock*layout.blockSize, io.SeekStart); err != nil {
-		return fmt.Errorf("Seek to datablock start failed: %v", err)
+		return 0, fmt.Errorf("Seek to datablock start failed: %v", err)
 	}
 
 	buf := make([]byte, layout.blockSize)
 	var readErr error
-	bytesRead := 0
+	total := int64(0)
 	for i := firstDataBlock; i <= lastDataBlock; i++ {
 		// Read a block that includes all or part of target range in
 		// input data.
-		bytesRead, readErr = data.Read(buf)
+		bytesRead, err := data.Read(buf)
+		readErr = err
 		// If at the end of input data and all previous blocks are
 		// verified, return the verified input data and EOF.
 		if readErr == io.EOF && bytesRead == 0 {
 			break
 		}
 		if readErr != nil && readErr != io.EOF {
-			return fmt.Errorf("Read from data failed: %v", err)
+			return 0, fmt.Errorf("Read from data failed: %v", err)
 		}
 		// If this is the end of file, zero the remaining bytes in buf,
 		// otherwise they are still from the previous block.
@@ -274,7 +275,7 @@ func Verify(w io.Writer, data, tree io.ReadSeeker, dataSize int64, readOffset in
 			}
 		}
 		if err := verifyBlock(tree, layout, buf, i, expectedRoot); err != nil {
-			return err
+			return 0, err
 		}
 		// startOff is the beginning of the read range within the
 		// current data block. Note that for all blocks other than the
@@ -298,10 +299,14 @@ func Verify(w io.Writer, data, tree io.ReadSeeker, dataSize int64, readOffset in
 		if endOff > int64(bytesRead) {
 			endOff = int64(bytesRead)
 		}
-		w.Write(buf[startOff:endOff])
+		n, err := w.Write(buf[startOff:endOff])
+		if err != nil {
+			return total, err
+		}
+		total += int64(n)
 
 	}
-	return readErr
+	return total, readErr
 }
 
 // verifyBlock verifies a block against tree. index is the number of block in

pkg/merkletree/merkletree_test.go

@@ -67,17 +67,17 @@ func TestLayout(t *testing.T) {
 		t.Run(fmt.Sprintf("%d", tc.dataSize), func(t *testing.T) {
 			l := InitLayout(tc.dataSize, tc.dataAndTreeInSameFile)
 			if l.blockSize != int64(usermem.PageSize) {
-				t.Errorf("got blockSize %d, want %d", l.blockSize, usermem.PageSize)
+				t.Errorf("Got blockSize %d, want %d", l.blockSize, usermem.PageSize)
 			}
 			if l.digestSize != sha256DigestSize {
-				t.Errorf("got digestSize %d, want %d", l.digestSize, sha256DigestSize)
+				t.Errorf("Got digestSize %d, want %d", l.digestSize, sha256DigestSize)
 			}
 			if l.numLevels() != len(tc.expectedLevelOffset) {
-				t.Errorf("got levels %d, want %d", l.numLevels(), len(tc.expectedLevelOffset))
+				t.Errorf("Got levels %d, want %d", l.numLevels(), len(tc.expectedLevelOffset))
 			}
 			for i := 0; i < l.numLevels() && i < len(tc.expectedLevelOffset); i++ {
 				if l.levelOffset[i] != tc.expectedLevelOffset[i] {
-					t.Errorf("got levelStart[%d] %d, want %d", i, l.levelOffset[i], tc.expectedLevelOffset[i])
+					t.Errorf("Got levelStart[%d] %d, want %d", i, l.levelOffset[i], tc.expectedLevelOffset[i])
 				}
 			}
 		})
@@ -169,11 +169,11 @@ func TestGenerate(t *testing.T) {
 					}, int64(len(tc.data)), &tree, &tree, dataAndTreeInSameFile)
 				}
 				if err != nil {
-					t.Fatalf("got err: %v, want nil", err)
+					t.Fatalf("Got err: %v, want nil", err)
 				}
 
 				if !bytes.Equal(root, tc.expectedRoot) {
-					t.Errorf("got root: %v, want %v", root, tc.expectedRoot)
+					t.Errorf("Got root: %v, want %v", root, tc.expectedRoot)
 				}
 			}
 		})
@@ -334,14 +334,21 @@ func TestVerify(t *testing.T) {
 				var buf bytes.Buffer
 				data[tc.modifyByte] ^= 1
 				if tc.shouldSucceed {
-					if err := Verify(&buf, bytes.NewReader(data), &tree, tc.dataSize, tc.verifyStart, tc.verifySize, root, dataAndTreeInSameFile); err != nil && err != io.EOF {
+					n, err := Verify(&buf, bytes.NewReader(data), &tree, tc.dataSize, tc.verifyStart, tc.verifySize, root, dataAndTreeInSameFile)
+					if err != nil && err != io.EOF {
 						t.Errorf("Verification failed when expected to succeed: %v", err)
 					}
-					if int64(buf.Len()) != tc.verifySize || !bytes.Equal(data[tc.verifyStart:tc.verifyStart+tc.verifySize], buf.Bytes()) {
-						t.Errorf("Incorrect output from Verify")
+					if n != tc.verifySize {
+						t.Errorf("Got Verify output size %d, want %d", n, tc.verifySize)
+					}
+					if int64(buf.Len()) != tc.verifySize {
+						t.Errorf("Got Verify output buf size %d, want %d,", buf.Len(), tc.verifySize)
+					}
+					if !bytes.Equal(data[tc.verifyStart:tc.verifyStart+tc.verifySize], buf.Bytes()) {
+						t.Errorf("Incorrect output buf from Verify")
 					}
 				} else {
-					if err := Verify(&buf, bytes.NewReader(data), &tree, tc.dataSize, tc.verifyStart, tc.verifySize, root, dataAndTreeInSameFile); err == nil {
+					if _, err := Verify(&buf, bytes.NewReader(data), &tree, tc.dataSize, tc.verifyStart, tc.verifySize, root, dataAndTreeInSameFile); err == nil {
 						t.Errorf("Verification succeeded when expected to fail")
 					}
 				}
@@ -382,14 +389,21 @@ func TestVerifyRandom(t *testing.T) {
 		var buf bytes.Buffer
 		// Checks that the random portion of data from the original data is
 		// verified successfully.
-		if err := Verify(&buf, bytes.NewReader(data), &tree, dataSize, start, size, root, dataAndTreeInSameFile); err != nil && err != io.EOF {
+		n, err := Verify(&buf, bytes.NewReader(data), &tree, dataSize, start, size, root, dataAndTreeInSameFile)
+		if err != nil && err != io.EOF {
 			t.Errorf("Verification failed for correct data: %v", err)
 		}
 		if size > dataSize-start {
 			size = dataSize - start
 		}
-		if int64(buf.Len()) != size || !bytes.Equal(data[start:start+size], buf.Bytes()) {
-			t.Errorf("Incorrect output from Verify")
+		if n != size {
+			t.Errorf("Got Verify output size %d, want %d", n, size)
+		}
+		if int64(buf.Len()) != size {
+			t.Errorf("Got Verify output buf size %d, want %d", buf.Len(), size)
+		}
+		if !bytes.Equal(data[start:start+size], buf.Bytes()) {
+			t.Errorf("Incorrect output buf from Verify")
 		}
 
 		buf.Reset()
@@ -397,7 +411,7 @@ func TestVerifyRandom(t *testing.T) {
 		randBytePos := rand.Int63n(size)
 		data[start+randBytePos] ^= 1
 
-		if err := Verify(&buf, bytes.NewReader(data), &tree, dataSize, start, size, root, dataAndTreeInSameFile); err == nil {
+		if _, err := Verify(&buf, bytes.NewReader(data), &tree, dataSize, start, size, root, dataAndTreeInSameFile); err == nil {
 			t.Errorf("Verification succeeded for modified data")
 		}
 	}

pkg/sentry/fsimpl/verity/filesystem.go

@@ -192,10 +192,7 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de
 	// contains the expected xattrs. If the file or the xattr does not
 	// exist, it indicates unexpected modifications to the file system.
 	if err == syserror.ENOENT || err == syserror.ENODATA {
-		if noCrashOnVerificationFailure {
-			return nil, err
-		}
-		panic(fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleOffsetInParentXattr, childPath, err))
+		return nil, alertIntegrityViolation(err, fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleOffsetInParentXattr, childPath, err))
 	}
 	if err != nil {
 		return nil, err
@@ -204,10 +201,7 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de
 	// unexpected modifications to the file system.
 	offset, err := strconv.Atoi(off)
 	if err != nil {
-		if noCrashOnVerificationFailure {
-			return nil, syserror.EINVAL
-		}
-		panic(fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleOffsetInParentXattr, childPath, err))
+		return nil, alertIntegrityViolation(err, fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleOffsetInParentXattr, childPath, err))
 	}
 
 	// Open parent Merkle tree file to read and verify child's root hash.
@@ -221,10 +215,7 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de
 	// The parent Merkle tree file should have been created. If it's
 	// missing, it indicates an unexpected modification to the file system.
 	if err == syserror.ENOENT {
-		if noCrashOnVerificationFailure {
-			return nil, err
-		}
-		panic(fmt.Sprintf("Failed to open parent Merkle file for %s: %v", childPath, err))
+		return nil, alertIntegrityViolation(err, fmt.Sprintf("Failed to open parent Merkle file for %s: %v", childPath, err))
 	}
 	if err != nil {
 		return nil, err
@@ -242,10 +233,7 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de
 	// contains the expected xattrs. If the file or the xattr does not
 	// exist, it indicates unexpected modifications to the file system.
 	if err == syserror.ENOENT || err == syserror.ENODATA {
-		if noCrashOnVerificationFailure {
-			return nil, err
-		}
-		panic(fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleSizeXattr, childPath, err))
+		return nil, alertIntegrityViolation(err, fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleSizeXattr, childPath, err))
 	}
 	if err != nil {
 		return nil, err
@@ -255,10 +243,7 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de
 	// unexpected modifications to the file system.
 	parentSize, err := strconv.Atoi(dataSize)
 	if err != nil {
-		if noCrashOnVerificationFailure {
-			return nil, syserror.EINVAL
-		}
-		panic(fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleSizeXattr, childPath, err))
+		return nil, alertIntegrityViolation(syserror.EINVAL, fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleSizeXattr, childPath, err))
 	}
 
 	fdReader := vfs.FileReadWriteSeeker{
@@ -270,11 +255,8 @@ func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *de
 	// contain the root hash of the children in the parent Merkle tree when
 	// Verify returns with success.
 	var buf bytes.Buffer
-	if err := merkletree.Verify(&buf, &fdReader, &fdReader, int64(parentSize), int64(offset), int64(merkletree.DigestSize()), parent.rootHash, true /* dataAndTreeInSameFile */); err != nil && err != io.EOF {
-		if noCrashOnVerificationFailure {
-			return nil, syserror.EIO
-		}
-		panic(fmt.Sprintf("Verification for %s failed: %v", childPath, err))
+	if _, err := merkletree.Verify(&buf, &fdReader, &fdReader, int64(parentSize), int64(offset), int64(merkletree.DigestSize()), parent.rootHash, true /* dataAndTreeInSameFile */); err != nil && err != io.EOF {
+		return nil, alertIntegrityViolation(syserror.EIO, fmt.Sprintf("Verification for %s failed: %v", childPath, err))
 	}
 
 	// Cache child root hash when it's verified the first time.
@@ -370,10 +352,7 @@ func (fs *filesystem) lookupAndVerifyLocked(ctx context.Context, parent *dentry,
 		// corresponding Merkle tree is found. This indicates an
 		// unexpected modification to the file system that
 		// removed/renamed the child.
-		if noCrashOnVerificationFailure {
-			return nil, childErr
-		}
-		panic(fmt.Sprintf("Target file %s is expected but missing", parentPath+"/"+name))
+		return nil, alertIntegrityViolation(childErr, fmt.Sprintf("Target file %s is expected but missing", parentPath+"/"+name))
 	} else if childErr == nil && childMerkleErr == syserror.ENOENT {
 		// If in allowRuntimeEnable mode, and the Merkle tree file is
 		// not created yet, we create an empty Merkle tree file, so that
@@ -409,10 +388,7 @@ func (fs *filesystem) lookupAndVerifyLocked(ctx context.Context, parent *dentry,
 			// If runtime enable is not allowed. This indicates an
 			// unexpected modification to the file system that
 			// removed/renamed the Merkle tree file.
-			if noCrashOnVerificationFailure {
-				return nil, childMerkleErr
-			}
-			panic(fmt.Sprintf("Expected Merkle file for target %s but none found", parentPath+"/"+name))
+			return nil, alertIntegrityViolation(childMerkleErr, fmt.Sprintf("Expected Merkle file for target %s but none found", parentPath+"/"+name))
 		}
 	} else if childErr == syserror.ENOENT && childMerkleErr == syserror.ENOENT {
 		// Both the child and the corresponding Merkle tree are missing.
@@ -421,7 +397,7 @@ func (fs *filesystem) lookupAndVerifyLocked(ctx context.Context, parent *dentry,
 		// TODO(b/167752508): Investigate possible ways to differentiate
 		// cases that both files are deleted from cases that they never
 		// exist in the file system.
-		panic(fmt.Sprintf("Failed to find file %s", parentPath+"/"+name))
+		return nil, alertIntegrityViolation(childErr, fmt.Sprintf("Failed to find file %s", parentPath+"/"+name))
 	}
 
 	mask := uint32(linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID)

pkg/sentry/fsimpl/verity/verity.go

@@ -22,6 +22,7 @@
 package verity
 
 import (
+	"fmt"
 	"strconv"
 	"sync/atomic"
 
@@ -29,7 +30,6 @@ import (
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/fspath"
 	"gvisor.dev/gvisor/pkg/marshal/primitive"
-
 	"gvisor.dev/gvisor/pkg/merkletree"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
 	fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock"
@@ -135,6 +135,16 @@ func (FilesystemType) Name() string {
 	return Name
 }
 
+// alertIntegrityViolation alerts a violation of integrity, which usually means
+// unexpected modification to the file system is detected. In
+// noCrashOnVerificationFailure mode, it returns an error, otherwise it panic.
+func alertIntegrityViolation(err error, msg string) error {
+	if noCrashOnVerificationFailure {
+		return err
+	}
+	panic(msg)
+}
+
 // GetFilesystem implements vfs.FilesystemType.GetFilesystem.
 func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) {
 	iopts, ok := opts.InternalData.(InternalFilesystemOptions)
@@ -204,15 +214,12 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 			return nil, nil, err
 		}
 	} else if err != nil {
-		// Failed to get dentry for the root Merkle file. This indicates
-		// an attack that removed/renamed the root Merkle file, or it's
-		// never generated.
-		if noCrashOnVerificationFailure {
-			fs.vfsfs.DecRef(ctx)
-			d.DecRef(ctx)
-			return nil, nil, err
-		}
-		panic("Failed to find root Merkle file")
+		// Failed to get dentry for the root Merkle file. This
+		// indicates an unexpected modification that removed/renamed
+		// the root Merkle file, or it's never generated.
+		fs.vfsfs.DecRef(ctx)
+		d.DecRef(ctx)
+		return nil, nil, alertIntegrityViolation(err, "Failed to find root Merkle file")
 	}
 	d.lowerMerkleVD = lowerMerkleVD
 
@@ -618,6 +625,54 @@ func (fd *fileDescription) Ioctl(ctx context.Context, uio usermem.IO, args arch.
 	}
 }
 
+// PRead implements vfs.FileDescriptionImpl.PRead.
+func (fd *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+	// No need to verify if the file is not enabled yet in
+	// allowRuntimeEnable mode.
+	if fd.d.fs.allowRuntimeEnable && len(fd.d.rootHash) == 0 {
+		return fd.lowerFD.PRead(ctx, dst, offset, opts)
+	}
+
+	// dataSize is the size of the whole file.
+	dataSize, err := fd.merkleReader.GetXattr(ctx, &vfs.GetXattrOptions{
+		Name: merkleSizeXattr,
+		Size: sizeOfInt32,
+	})
+
+	// The Merkle tree file for the child should have been created and
+	// contains the expected xattrs. If the xattr does not exist, it
+	// indicates unexpected modifications to the file system.
+	if err == syserror.ENODATA {
+		return 0, alertIntegrityViolation(err, fmt.Sprintf("Failed to get xattr %s: %v", merkleSizeXattr, err))
+	}
+	if err != nil {
+		return 0, err
+	}
+
+	// The dataSize xattr should be an integer. If it's not, it indicates
+	// unexpected modifications to the file system.
+	size, err := strconv.Atoi(dataSize)
+	if err != nil {
+		return 0, alertIntegrityViolation(err, fmt.Sprintf("Failed to convert xattr %s to int: %v", merkleSizeXattr, err))
+	}
+
+	dataReader := vfs.FileReadWriteSeeker{
+		FD:  fd.lowerFD,
+		Ctx: ctx,
+	}
+
+	merkleReader := vfs.FileReadWriteSeeker{
+		FD:  fd.merkleReader,
+		Ctx: ctx,
+	}
+
+	n, err := merkletree.Verify(dst.Writer(ctx), &dataReader, &merkleReader, int64(size), offset, dst.NumBytes(), fd.d.rootHash, false /* dataAndTreeInSameFile */)
+	if err != nil {
+		return 0, alertIntegrityViolation(syserror.EINVAL, fmt.Sprintf("Verification failed: %v", err))
+	}
+	return n, err
+}
+
 // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX.
 func (fd *fileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error {
 	return fd.Locks().LockPOSIX(ctx, &fd.vfsfd, uid, t, start, length, whence, block)