Add support for pipes in VFS2.

PiperOrigin-RevId: 275650307

7 files changed, 380 insertions, 13 deletions

diff --git a/pkg/sentry/fsimpl/memfs/BUILD b/pkg/sentry/fsimpl/memfs/BUILD
index 7e364c5fd..04d667273 100644
--- a/pkg/sentry/fsimpl/memfs/BUILD
+++ b/pkg/sentry/fsimpl/memfs/BUILD
@@ -24,14 +24,18 @@ go_library(
         "directory.go",
         "filesystem.go",
         "memfs.go",
+        "named_pipe.go",
         "regular_file.go",
         "symlink.go",
     ],
     importpath = "gvisor.dev/gvisor/pkg/sentry/fsimpl/memfs",
     deps = [
         "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/sentry/arch",
         "//pkg/sentry/context",
         "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/pipe",
         "//pkg/sentry/usermem",
         "//pkg/sentry/vfs",
         "//pkg/syserror",
@@ -54,3 +58,19 @@ go_test(
         "//pkg/syserror",
     ],
 )
+
+go_test(
+    name = "memfs_test",
+    size = "small",
+    srcs = ["pipe_test.go"],
+    embed = [":memfs"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/usermem",
+        "//pkg/sentry/vfs",
+        "//pkg/syserror",
+    ],
+)
diff --git a/pkg/sentry/fsimpl/memfs/filesystem.go b/pkg/sentry/fsimpl/memfs/filesystem.go
index f79e2d9c8..f006c40cd 100644
--- a/pkg/sentry/fsimpl/memfs/filesystem.go
+++ b/pkg/sentry/fsimpl/memfs/filesystem.go
@@ -233,7 +233,7 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v
 	if err != nil {
 		return err
 	}
-	_, err = checkCreateLocked(rp, parentVFSD, parentInode)
+	pc, err := checkCreateLocked(rp, parentVFSD, parentInode)
 	if err != nil {
 		return err
 	}
@@ -241,8 +241,40 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v
 		return err
 	}
 	defer rp.Mount().EndWrite()
-	// TODO: actually implement mknod
-	return syserror.EPERM
+
+	switch opts.Mode.FileType() {
+	case 0:
+		// "Zero file type is equivalent to type S_IFREG." - mknod(2)
+		fallthrough
+	case linux.ModeRegular:
+		// TODO(b/138862511): Implement.
+		return syserror.EINVAL
+
+	case linux.ModeNamedPipe:
+		child := fs.newDentry(fs.newNamedPipe(rp.Credentials(), opts.Mode))
+		parentVFSD.InsertChild(&child.vfsd, pc)
+		parentInode.impl.(*directory).childList.PushBack(child)
+		return nil
+
+	case linux.ModeSocket:
+		// TODO(b/138862511): Implement.
+		return syserror.EINVAL
+
+	case linux.ModeCharacterDevice:
+		fallthrough
+	case linux.ModeBlockDevice:
+		// TODO(b/72101894): We don't support creating block or character
+		// devices at the moment.
+		//
+		// When we start supporting block and character devices, we'll
+		// need to check for CAP_MKNOD here.
+		return syserror.EPERM
+
+	default:
+		// "EINVAL - mode requested creation of something other than a
+		// regular file, device special file, FIFO or socket." - mknod(2)
+		return syserror.EINVAL
+	}
 }
 
 // OpenAt implements vfs.FilesystemImpl.OpenAt.
@@ -250,8 +282,9 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf
 	// Filter out flags that are not supported by memfs. O_DIRECTORY and
 	// O_NOFOLLOW have no effect here (they're handled by VFS by setting
 	// appropriate bits in rp), but are returned by
-	// FileDescriptionImpl.StatusFlags().
-	opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW
+	// FileDescriptionImpl.StatusFlags(). O_NONBLOCK is supported only by
+	// pipes.
+	opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NONBLOCK
 
 	if opts.Flags&linux.O_CREAT == 0 {
 		fs.mu.RLock()
@@ -260,7 +293,7 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf
 		if err != nil {
 			return nil, err
 		}
-		return inode.open(rp, vfsd, opts.Flags, false)
+		return inode.open(ctx, rp, vfsd, opts.Flags, false)
 	}
 
 	mustCreate := opts.Flags&linux.O_EXCL != 0
@@ -275,7 +308,7 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf
 		if mustCreate {
 			return nil, syserror.EEXIST
 		}
-		return inode.open(rp, vfsd, opts.Flags, false)
+		return inode.open(ctx, rp, vfsd, opts.Flags, false)
 	}
 afterTrailingSymlink:
 	// Walk to the parent directory of the last path component.
@@ -320,7 +353,7 @@ afterTrailingSymlink:
 		child := fs.newDentry(childInode)
 		vfsd.InsertChild(&child.vfsd, pc)
 		inode.impl.(*directory).childList.PushBack(child)
-		return childInode.open(rp, &child.vfsd, opts.Flags, true)
+		return childInode.open(ctx, rp, &child.vfsd, opts.Flags, true)
 	}
 	// Open existing file or follow symlink.
 	if mustCreate {
@@ -336,10 +369,10 @@ afterTrailingSymlink:
 		// symlink target.
 		goto afterTrailingSymlink
 	}
-	return childInode.open(rp, childVFSD, opts.Flags, false)
+	return childInode.open(ctx, rp, childVFSD, opts.Flags, false)
 }
 
-func (i *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32, afterCreate bool) (*vfs.FileDescription, error) {
+func (i *inode) open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32, afterCreate bool) (*vfs.FileDescription, error) {
 	ats := vfs.AccessTypesForOpenFlags(flags)
 	if !afterCreate {
 		if err := i.checkPermissions(rp.Credentials(), ats, i.isDir()); err != nil {
@@ -378,6 +411,8 @@ func (i *inode) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32, afte
 	case *symlink:
 		// Can't open symlinks without O_PATH (which is unimplemented).
 		return nil, syserror.ELOOP
+	case *namedPipe:
+		return newNamedPipeFD(ctx, impl, rp, vfsd, flags)
 	default:
 		panic(fmt.Sprintf("unknown inode type: %T", i.impl))
 	}
diff --git a/pkg/sentry/fsimpl/memfs/memfs.go b/pkg/sentry/fsimpl/memfs/memfs.go
index b78471c0f..64c851c1a 100644
--- a/pkg/sentry/fsimpl/memfs/memfs.go
+++ b/pkg/sentry/fsimpl/memfs/memfs.go
@@ -227,6 +227,8 @@ func (i *inode) statTo(stat *linux.Statx) {
 		stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS
 		stat.Size = uint64(len(impl.target))
 		stat.Blocks = allocatedBlocksForSize(stat.Size)
+	case *namedPipe:
+		stat.Mode |= linux.S_IFIFO
 	default:
 		panic(fmt.Sprintf("unknown inode type: %T", i.impl))
 	}
diff --git a/pkg/sentry/fsimpl/memfs/named_pipe.go b/pkg/sentry/fsimpl/memfs/named_pipe.go
new file mode 100644
index 000000000..732ed7c58
--- /dev/null
+++ b/pkg/sentry/fsimpl/memfs/named_pipe.go
@@ -0,0 +1,59 @@
+// Copyright 2019 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package memfs
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/context"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/pipe"
+	"gvisor.dev/gvisor/pkg/sentry/usermem"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+)
+
+type namedPipe struct {
+	inode inode
+
+	pipe *pipe.VFSPipe
+}
+
+// Preconditions:
+//   * fs.mu must be locked.
+//   * rp.Mount().CheckBeginWrite() has been called successfully.
+func (fs *filesystem) newNamedPipe(creds *auth.Credentials, mode linux.FileMode) *inode {
+	file := &namedPipe{pipe: pipe.NewVFSPipe(pipe.DefaultPipeSize, usermem.PageSize)}
+	file.inode.init(file, fs, creds, mode)
+	file.inode.nlink = 1 // Only the parent has a link.
+	return &file.inode
+}
+
+// namedPipeFD implements vfs.FileDescriptionImpl. Methods are implemented
+// entirely via struct embedding.
+type namedPipeFD struct {
+	fileDescription
+
+	*pipe.VFSPipeFD
+}
+
+func newNamedPipeFD(ctx context.Context, np *namedPipe, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) {
+	var err error
+	var fd namedPipeFD
+	fd.VFSPipeFD, err = np.pipe.NewVFSPipeFD(ctx, rp, vfsd, &fd.vfsfd, flags)
+	if err != nil {
+		return nil, err
+	}
+	fd.vfsfd.Init(&fd, rp.Mount(), vfsd)
+	return &fd.vfsfd, nil
+}
diff --git a/pkg/sentry/fsimpl/memfs/pipe_test.go b/pkg/sentry/fsimpl/memfs/pipe_test.go
new file mode 100644
index 000000000..0674b81a3
--- /dev/null
+++ b/pkg/sentry/fsimpl/memfs/pipe_test.go
@@ -0,0 +1,233 @@
+// Copyright 2019 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package memfs
+
+import (
+	"bytes"
+	"testing"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/context"
+	"gvisor.dev/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/usermem"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+const fileName = "mypipe"
+
+func TestSeparateFDs(t *testing.T) {
+	ctx, creds, vfsObj, root := setup(t)
+	defer root.DecRef()
+
+	// Open the read side. This is done in a concurrently because opening
+	// One end the pipe blocks until the other end is opened.
+	pop := vfs.PathOperation{
+		Root:               root,
+		Start:              root,
+		Pathname:           fileName,
+		FollowFinalSymlink: true,
+	}
+	rfdchan := make(chan *vfs.FileDescription)
+	go func() {
+		openOpts := vfs.OpenOptions{Flags: linux.O_RDONLY}
+		rfd, _ := vfsObj.OpenAt(ctx, creds, &pop, &openOpts)
+		rfdchan <- rfd
+	}()
+
+	// Open the write side.
+	openOpts := vfs.OpenOptions{Flags: linux.O_WRONLY}
+	wfd, err := vfsObj.OpenAt(ctx, creds, &pop, &openOpts)
+	if err != nil {
+		t.Fatalf("failed to open pipe for writing %q: %v", fileName, err)
+	}
+	defer wfd.DecRef()
+
+	rfd, ok := <-rfdchan
+	if !ok {
+		t.Fatalf("failed to open pipe for reading %q", fileName)
+	}
+	defer rfd.DecRef()
+
+	const msg = "vamos azul"
+	checkEmpty(ctx, t, rfd)
+	checkWrite(ctx, t, wfd, msg)
+	checkRead(ctx, t, rfd, msg)
+}
+
+func TestNonblockingRead(t *testing.T) {
+	ctx, creds, vfsObj, root := setup(t)
+	defer root.DecRef()
+
+	// Open the read side as nonblocking.
+	pop := vfs.PathOperation{
+		Root:               root,
+		Start:              root,
+		Pathname:           fileName,
+		FollowFinalSymlink: true,
+	}
+	openOpts := vfs.OpenOptions{Flags: linux.O_RDONLY | linux.O_NONBLOCK}
+	rfd, err := vfsObj.OpenAt(ctx, creds, &pop, &openOpts)
+	if err != nil {
+		t.Fatalf("failed to open pipe for reading %q: %v", fileName, err)
+	}
+	defer rfd.DecRef()
+
+	// Open the write side.
+	openOpts = vfs.OpenOptions{Flags: linux.O_WRONLY}
+	wfd, err := vfsObj.OpenAt(ctx, creds, &pop, &openOpts)
+	if err != nil {
+		t.Fatalf("failed to open pipe for writing %q: %v", fileName, err)
+	}
+	defer wfd.DecRef()
+
+	const msg = "geh blau"
+	checkEmpty(ctx, t, rfd)
+	checkWrite(ctx, t, wfd, msg)
+	checkRead(ctx, t, rfd, msg)
+}
+
+func TestNonblockingWriteError(t *testing.T) {
+	ctx, creds, vfsObj, root := setup(t)
+	defer root.DecRef()
+
+	// Open the write side as nonblocking, which should return ENXIO.
+	pop := vfs.PathOperation{
+		Root:               root,
+		Start:              root,
+		Pathname:           fileName,
+		FollowFinalSymlink: true,
+	}
+	openOpts := vfs.OpenOptions{Flags: linux.O_WRONLY | linux.O_NONBLOCK}
+	_, err := vfsObj.OpenAt(ctx, creds, &pop, &openOpts)
+	if err != syserror.ENXIO {
+		t.Fatalf("expected ENXIO, but got error: %v", err)
+	}
+}
+
+func TestSingleFD(t *testing.T) {
+	ctx, creds, vfsObj, root := setup(t)
+	defer root.DecRef()
+
+	// Open the pipe as readable and writable.
+	pop := vfs.PathOperation{
+		Root:               root,
+		Start:              root,
+		Pathname:           fileName,
+		FollowFinalSymlink: true,
+	}
+	openOpts := vfs.OpenOptions{Flags: linux.O_RDWR}
+	fd, err := vfsObj.OpenAt(ctx, creds, &pop, &openOpts)
+	if err != nil {
+		t.Fatalf("failed to open pipe for writing %q: %v", fileName, err)
+	}
+	defer fd.DecRef()
+
+	const msg = "forza blu"
+	checkEmpty(ctx, t, fd)
+	checkWrite(ctx, t, fd, msg)
+	checkRead(ctx, t, fd, msg)
+}
+
+// setup creates a VFS with a pipe in the root directory at path fileName. The
+// returned VirtualDentry must be DecRef()'d be the caller. It calls t.Fatal
+// upon failure.
+func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesystem, vfs.VirtualDentry) {
+	ctx := contexttest.Context(t)
+	creds := auth.CredentialsFromContext(ctx)
+
+	// Create VFS.
+	vfsObj := vfs.New()
+	vfsObj.MustRegisterFilesystemType("memfs", FilesystemType{})
+	mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "memfs", &vfs.NewFilesystemOptions{})
+	if err != nil {
+		t.Fatalf("failed to create tmpfs root mount: %v", err)
+	}
+
+	// Create the pipe.
+	root := mntns.Root()
+	pop := vfs.PathOperation{
+		Root:               root,
+		Start:              root,
+		Pathname:           fileName,
+		FollowFinalSymlink: true,
+	}
+	mknodOpts := vfs.MknodOptions{Mode: linux.ModeNamedPipe | 0644}
+	if err := vfsObj.MknodAt(ctx, creds, &pop, &mknodOpts); err != nil {
+		t.Fatalf("failed to create file %q: %v", fileName, err)
+	}
+
+	// Sanity check: the file pipe exists and has the correct mode.
+	stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{
+		Root:               root,
+		Start:              root,
+		Pathname:           fileName,
+		FollowFinalSymlink: true,
+	}, &vfs.StatOptions{})
+	if err != nil {
+		t.Fatalf("stat(%q) failed: %v", fileName, err)
+	}
+	if stat.Mode&^linux.S_IFMT != 0644 {
+		t.Errorf("got wrong permissions (%0o)", stat.Mode)
+	}
+	if stat.Mode&linux.S_IFMT != linux.ModeNamedPipe {
+		t.Errorf("got wrong file type (%0o)", stat.Mode)
+	}
+
+	return ctx, creds, vfsObj, root
+}
+
+// checkEmpty calls t.Fatal if the pipe in fd is not empty.
+func checkEmpty(ctx context.Context, t *testing.T, fd *vfs.FileDescription) {
+	readData := make([]byte, 1)
+	dst := usermem.BytesIOSequence(readData)
+	bytesRead, err := fd.Impl().Read(ctx, dst, vfs.ReadOptions{})
+	if err != syserror.ErrWouldBlock {
+		t.Fatalf("expected ErrWouldBlock reading from empty pipe %q, but got: %v", fileName, err)
+	}
+	if bytesRead != 0 {
+		t.Fatalf("expected to read 0 bytes, but got %d", bytesRead)
+	}
+}
+
+// checkWrite calls t.Fatal if it fails to write all of msg to fd.
+func checkWrite(ctx context.Context, t *testing.T, fd *vfs.FileDescription, msg string) {
+	writeData := []byte(msg)
+	src := usermem.BytesIOSequence(writeData)
+	bytesWritten, err := fd.Impl().Write(ctx, src, vfs.WriteOptions{})
+	if err != nil {
+		t.Fatalf("error writing to pipe %q: %v", fileName, err)
+	}
+	if bytesWritten != int64(len(writeData)) {
+		t.Fatalf("expected to write %d bytes, but wrote %d", len(writeData), bytesWritten)
+	}
+}
+
+// checkRead calls t.Fatal if it fails to read msg from fd.
+func checkRead(ctx context.Context, t *testing.T, fd *vfs.FileDescription, msg string) {
+	readData := make([]byte, len(msg))
+	dst := usermem.BytesIOSequence(readData)
+	bytesRead, err := fd.Impl().Read(ctx, dst, vfs.ReadOptions{})
+	if err != nil {
+		t.Fatalf("error reading from pipe %q: %v", fileName, err)
+	}
+	if bytesRead != int64(len(msg)) {
+		t.Fatalf("expected to read %d bytes, but got %d", len(msg), bytesRead)
+	}
+	if !bytes.Equal(readData, []byte(msg)) {
+		t.Fatalf("expected to read %q from pipe, but got %q", msg, string(readData))
+	}
+}
diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go
index 02320b830..6416e0dd8 100644
--- a/pkg/sentry/kernel/pipe/vfs.go
+++ b/pkg/sentry/kernel/pipe/vfs.go
@@ -182,7 +182,7 @@ func (fd *VFSPipeFD) Release() {
 }
 
 // OnClose implements vfs.FileDescriptionImpl.OnClose.
-func (fd *VFSPipeFD) OnClose() error {
+func (fd *VFSPipeFD) OnClose(_ context.Context) error {
 	return nil
 }
 
diff --git a/pkg/sentry/vfs/syscalls.go b/pkg/sentry/vfs/syscalls.go
index 23f2b9e08..abde0feaa 100644
--- a/pkg/sentry/vfs/syscalls.go
+++ b/pkg/sentry/vfs/syscalls.go
@@ -96,6 +96,26 @@ func (vfs *VirtualFilesystem) MkdirAt(ctx context.Context, creds *auth.Credentia
 	}
 }
 
+// MknodAt creates a file of the given mode at the given path. It returns an
+// error from the syserror package.
+func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *MknodOptions) error {
+	rp, err := vfs.getResolvingPath(creds, pop)
+	if err != nil {
+		return nil
+	}
+	for {
+		if err = rp.mount.fs.impl.MknodAt(ctx, rp, *opts); err == nil {
+			vfs.putResolvingPath(rp)
+			return nil
+		}
+		// Handle mount traversals.
+		if !rp.handleError(err) {
+			vfs.putResolvingPath(rp)
+			return err
+		}
+	}
+}
+
 // OpenAt returns a FileDescription providing access to the file at the given
 // path. A reference is taken on the returned FileDescription.
 func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *OpenOptions) (*FileDescription, error) {
@@ -198,8 +218,6 @@ func (fd *FileDescription) SetStatusFlags(ctx context.Context, flags uint32) err
 //
 // - VFS.LinkAt()
 //
-// - VFS.MknodAt()
-//
 // - VFS.ReadlinkAt()
 //
 // - VFS.RenameAt()