Plumb VFS2 inside the Sentry

- Added fsbridge package with interface that can be used to open
  and read from VFS1 and VFS2 files.
- Converted ELF loader to use fsbridge
- Added VFS2 types to FSContext
- Added vfs.MountNamespace to ThreadGroup

Updates #1623

PiperOrigin-RevId: 295183950

pkg/sentry/control/BUILD

@@ -16,10 +16,13 @@ go_library(
     ],
     deps = [
         "//pkg/abi/linux",
+        "//pkg/context",
         "//pkg/fd",
+        "//pkg/fspath",
         "//pkg/log",
         "//pkg/sentry/fs",
         "//pkg/sentry/fs/host",
+        "//pkg/sentry/fsbridge",
         "//pkg/sentry/kernel",
         "//pkg/sentry/kernel/auth",
         "//pkg/sentry/kernel/time",
@@ -27,8 +30,10 @@ go_library(
         "//pkg/sentry/state",
         "//pkg/sentry/strace",
         "//pkg/sentry/usage",
+        "//pkg/sentry/vfs",
         "//pkg/sentry/watchdog",
         "//pkg/sync",
+        "//pkg/syserror",
         "//pkg/tcpip/link/sniffer",
         "//pkg/urpc",
     ],

pkg/sentry/control/proc.go

@@ -18,19 +18,26 @@ import (
 	"bytes"
 	"encoding/json"
 	"fmt"
+	"path"
 	"sort"
 	"strings"
 	"text/tabwriter"
 	"time"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/fspath"
+	"gvisor.dev/gvisor/pkg/log"
 	"gvisor.dev/gvisor/pkg/sentry/fs"
 	"gvisor.dev/gvisor/pkg/sentry/fs/host"
+	"gvisor.dev/gvisor/pkg/sentry/fsbridge"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time"
 	"gvisor.dev/gvisor/pkg/sentry/limits"
 	"gvisor.dev/gvisor/pkg/sentry/usage"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/urpc"
 )
 
@@ -60,6 +67,12 @@ type ExecArgs struct {
 	// process's MountNamespace.
 	MountNamespace *fs.MountNamespace
 
+	// MountNamespaceVFS2 is the mount namespace to execute the new process in.
+	// A reference on MountNamespace must be held for the lifetime of the
+	// ExecArgs. If MountNamespace is nil, it will default to the init
+	// process's MountNamespace.
+	MountNamespaceVFS2 *vfs.MountNamespace
+
 	// WorkingDirectory defines the working directory for the new process.
 	WorkingDirectory string `json:"wd"`
 
@@ -150,6 +163,7 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI
 		Envv:                    args.Envv,
 		WorkingDirectory:        args.WorkingDirectory,
 		MountNamespace:          args.MountNamespace,
+		MountNamespaceVFS2:      args.MountNamespaceVFS2,
 		Credentials:             creds,
 		FDTable:                 fdTable,
 		Umask:                   0022,
@@ -166,24 +180,53 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI
 		// be donated to the new process in CreateProcess.
 		initArgs.MountNamespace.IncRef()
 	}
+	if initArgs.MountNamespaceVFS2 != nil {
+		// initArgs must hold a reference on MountNamespaceVFS2, which will
+		// be donated to the new process in CreateProcess.
+		initArgs.MountNamespaceVFS2.IncRef()
+	}
 	ctx := initArgs.NewContext(proc.Kernel)
 
 	if initArgs.Filename == "" {
-		// Get the full path to the filename from the PATH env variable.
-		paths := fs.GetPath(initArgs.Envv)
-		mns := initArgs.MountNamespace
-		if mns == nil {
-			mns = proc.Kernel.GlobalInit().Leader().MountNamespace()
+		if kernel.VFS2Enabled {
+			// Get the full path to the filename from the PATH env variable.
+			if initArgs.MountNamespaceVFS2 == nil {
+				// Set initArgs so that 'ctx' returns the namespace.
+				//
+				// MountNamespaceVFS2 adds a reference to the namespace, which is
+				// transferred to the new process.
+				initArgs.MountNamespaceVFS2 = proc.Kernel.GlobalInit().Leader().MountNamespaceVFS2()
+			}
+
+			paths := fs.GetPath(initArgs.Envv)
+			vfsObj := proc.Kernel.VFS
+			file, err := ResolveExecutablePath(ctx, vfsObj, initArgs.WorkingDirectory, initArgs.Argv[0], paths)
+			if err != nil {
+				return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err)
+			}
+			initArgs.File = fsbridge.NewVFSFile(file)
+		} else {
+			// Get the full path to the filename from the PATH env variable.
+			paths := fs.GetPath(initArgs.Envv)
+			if initArgs.MountNamespace == nil {
+				// Set initArgs so that 'ctx' returns the namespace.
+				initArgs.MountNamespace = proc.Kernel.GlobalInit().Leader().MountNamespace()
+
+				// initArgs must hold a reference on MountNamespace, which will
+				// be donated to the new process in CreateProcess.
+				initArgs.MountNamespaceVFS2.IncRef()
+			}
+			f, err := initArgs.MountNamespace.ResolveExecutablePath(ctx, initArgs.WorkingDirectory, initArgs.Argv[0], paths)
+			if err != nil {
+				return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err)
+			}
+			initArgs.Filename = f
 		}
-		f, err := mns.ResolveExecutablePath(ctx, initArgs.WorkingDirectory, initArgs.Argv[0], paths)
-		if err != nil {
-			return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err)
-		}
-		initArgs.Filename = f
 	}
 
 	mounter := fs.FileOwnerFromContext(ctx)
 
+	// TODO(gvisor.dev/issue/1623): Use host FD when supported in VFS2.
 	var ttyFile *fs.File
 	for appFD, hostFile := range args.FilePayload.Files {
 		var appFile *fs.File
@@ -411,3 +454,67 @@ func ttyName(tty *kernel.TTY) string {
 	}
 	return fmt.Sprintf("pts/%d", tty.Index)
 }
+
+// ResolveExecutablePath resolves the given executable name given a set of
+// paths that might contain it.
+func ResolveExecutablePath(ctx context.Context, vfsObj *vfs.VirtualFilesystem, wd, name string, paths []string) (*vfs.FileDescription, error) {
+	root := vfs.RootFromContext(ctx)
+	defer root.DecRef()
+	creds := auth.CredentialsFromContext(ctx)
+
+	// Absolute paths can be used directly.
+	if path.IsAbs(name) {
+		return openExecutable(ctx, vfsObj, creds, root, name)
+	}
+
+	// Paths with '/' in them should be joined to the working directory, or
+	// to the root if working directory is not set.
+	if strings.IndexByte(name, '/') > 0 {
+		if len(wd) == 0 {
+			wd = "/"
+		}
+		if !path.IsAbs(wd) {
+			return nil, fmt.Errorf("working directory %q must be absolute", wd)
+		}
+		return openExecutable(ctx, vfsObj, creds, root, path.Join(wd, name))
+	}
+
+	// Otherwise, we must lookup the name in the paths, starting from the
+	// calling context's root directory.
+	for _, p := range paths {
+		if !path.IsAbs(p) {
+			// Relative paths aren't safe, no one should be using them.
+			log.Warningf("Skipping relative path %q in $PATH", p)
+			continue
+		}
+
+		binPath := path.Join(p, name)
+		f, err := openExecutable(ctx, vfsObj, creds, root, binPath)
+		if err != nil {
+			return nil, err
+		}
+		if f == nil {
+			continue // Not found/no access.
+		}
+		return f, nil
+	}
+	return nil, syserror.ENOENT
+}
+
+func openExecutable(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, root vfs.VirtualDentry, path string) (*vfs.FileDescription, error) {
+	pop := vfs.PathOperation{
+		Root:               root,
+		Start:              root, // binPath is absolute, Start can be anything.
+		Path:               fspath.Parse(path),
+		FollowFinalSymlink: true,
+	}
+	opts := &vfs.OpenOptions{
+		Flags:    linux.O_RDONLY,
+		FileExec: true,
+	}
+	f, err := vfsObj.OpenAt(ctx, creds, &pop, opts)
+	if err == syserror.ENOENT || err == syserror.EACCES {
+		return nil, nil
+	}
+	return f, err
+}

pkg/sentry/fs/proc/BUILD

@@ -36,6 +36,7 @@ go_library(
         "//pkg/sentry/fs/proc/device",
         "//pkg/sentry/fs/proc/seqfile",
         "//pkg/sentry/fs/ramfs",
+        "//pkg/sentry/fsbridge",
         "//pkg/sentry/inet",
         "//pkg/sentry/kernel",
         "//pkg/sentry/kernel/auth",

pkg/sentry/fs/proc/task.go

@@ -28,6 +28,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/fs/proc/device"
 	"gvisor.dev/gvisor/pkg/sentry/fs/proc/seqfile"
 	"gvisor.dev/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.dev/gvisor/pkg/sentry/fsbridge"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/limits"
 	"gvisor.dev/gvisor/pkg/sentry/mm"
@@ -249,7 +250,7 @@ func newExe(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
 	return newProcInode(t, exeSymlink, msrc, fs.Symlink, t)
 }
 
-func (e *exe) executable() (d *fs.Dirent, err error) {
+func (e *exe) executable() (file fsbridge.File, err error) {
 	e.t.WithMuLocked(func(t *kernel.Task) {
 		mm := t.MemoryManager()
 		if mm == nil {
@@ -262,8 +263,8 @@ func (e *exe) executable() (d *fs.Dirent, err error) {
 		// The MemoryManager may be destroyed, in which case
 		// MemoryManager.destroy will simply set the executable to nil
 		// (with locks held).
-		d = mm.Executable()
-		if d == nil {
+		file = mm.Executable()
+		if file == nil {
 			err = syserror.ENOENT
 		}
 	})
@@ -283,15 +284,7 @@ func (e *exe) Readlink(ctx context.Context, inode *fs.Inode) (string, error) {
 	}
 	defer exec.DecRef()
 
-	root := fs.RootFromContext(ctx)
-	if root == nil {
-		// This doesn't correspond to anything in Linux because the vfs is
-		// global there.
-		return "", syserror.EINVAL
-	}
-	defer root.DecRef()
-	n, _ := exec.FullName(root)
-	return n, nil
+	return exec.PathnameWithDeleted(ctx), nil
 }
 
 // namespaceSymlink represents a symlink in the namespacefs, such as the files

pkg/sentry/fsbridge/BUILD

@@ -0,0 +1,24 @@
+load("//tools:defs.bzl", "go_library")
+
+licenses(["notice"])
+
+go_library(
+    name = "fsbridge",
+    srcs = [
+        "bridge.go",
+        "fs.go",
+        "vfs.go",
+    ],
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/context",
+        "//pkg/fspath",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/vfs",
+        "//pkg/syserror",
+        "//pkg/usermem",
+    ],
+)

pkg/sentry/fsbridge/bridge.go

@@ -0,0 +1,54 @@
+// Copyright 2020 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 fsbridge provides common interfaces to bridge between VFS1 and VFS2
+// files.
+package fsbridge
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/memmap"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// File provides a common interface to bridge between VFS1 and VFS2 files.
+type File interface {
+	// PathnameWithDeleted returns an absolute pathname to vd, consistent with
+	// Linux's d_path(). In particular, if vd.Dentry() has been disowned,
+	// PathnameWithDeleted appends " (deleted)" to the returned pathname.
+	PathnameWithDeleted(ctx context.Context) string
+
+	// ReadFull read all contents from the file.
+	ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error)
+
+	// ConfigureMMap mutates opts to implement mmap(2) for the file.
+	ConfigureMMap(context.Context, *memmap.MMapOpts) error
+
+	// Type returns the file type, e.g. linux.S_IFREG.
+	Type(context.Context) (linux.FileMode, error)
+
+	// IncRef increments reference.
+	IncRef()
+
+	// DecRef decrements reference.
+	DecRef()
+}
+
+// Lookup provides a common interface to open files.
+type Lookup interface {
+	// OpenPath opens a file.
+	OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, remainingTraversals *uint, resolveFinal bool) (File, error)
+}

pkg/sentry/fsbridge/fs.go

@@ -0,0 +1,181 @@
+// Copyright 2020 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 fsbridge
+
+import (
+	"io"
+	"strings"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/fs"
+	"gvisor.dev/gvisor/pkg/sentry/memmap"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// fsFile implements File interface over fs.File.
+//
+// +stateify savable
+type fsFile struct {
+	file *fs.File
+}
+
+var _ File = (*fsFile)(nil)
+
+// NewFSFile creates a new File over fs.File.
+func NewFSFile(file *fs.File) File {
+	return &fsFile{file: file}
+}
+
+// PathnameWithDeleted implements File.
+func (f *fsFile) PathnameWithDeleted(ctx context.Context) string {
+	root := fs.RootFromContext(ctx)
+	if root == nil {
+		// This doesn't correspond to anything in Linux because the vfs is
+		// global there.
+		return ""
+	}
+	defer root.DecRef()
+
+	name, _ := f.file.Dirent.FullName(root)
+	return name
+}
+
+// ReadFull implements File.
+func (f *fsFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	var total int64
+	for dst.NumBytes() > 0 {
+		n, err := f.file.Preadv(ctx, dst, offset+total)
+		total += n
+		if err == io.EOF && total != 0 {
+			return total, io.ErrUnexpectedEOF
+		} else if err != nil {
+			return total, err
+		}
+		dst = dst.DropFirst64(n)
+	}
+	return total, nil
+}
+
+// ConfigureMMap implements File.
+func (f *fsFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error {
+	return f.file.ConfigureMMap(ctx, opts)
+}
+
+// Type implements File.
+func (f *fsFile) Type(context.Context) (linux.FileMode, error) {
+	return linux.FileMode(f.file.Dirent.Inode.StableAttr.Type.LinuxType()), nil
+}
+
+// IncRef implements File.
+func (f *fsFile) IncRef() {
+	f.file.IncRef()
+}
+
+// DecRef implements File.
+func (f *fsFile) DecRef() {
+	f.file.DecRef()
+}
+
+// fsLookup implements Lookup interface using fs.File.
+//
+// +stateify savable
+type fsLookup struct {
+	mntns *fs.MountNamespace
+
+	root       *fs.Dirent
+	workingDir *fs.Dirent
+}
+
+var _ Lookup = (*fsLookup)(nil)
+
+// NewFSLookup creates a new Lookup using VFS1.
+func NewFSLookup(mntns *fs.MountNamespace, root, workingDir *fs.Dirent) Lookup {
+	return &fsLookup{
+		mntns:      mntns,
+		root:       root,
+		workingDir: workingDir,
+	}
+}
+
+// OpenPath implements Lookup.
+func (l *fsLookup) OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, remainingTraversals *uint, resolveFinal bool) (File, error) {
+	var d *fs.Dirent
+	var err error
+	if resolveFinal {
+		d, err = l.mntns.FindInode(ctx, l.root, l.workingDir, path, remainingTraversals)
+	} else {
+		d, err = l.mntns.FindLink(ctx, l.root, l.workingDir, path, remainingTraversals)
+	}
+	if err != nil {
+		return nil, err
+	}
+	defer d.DecRef()
+
+	if !resolveFinal && fs.IsSymlink(d.Inode.StableAttr) {
+		return nil, syserror.ELOOP
+	}
+
+	fsPerm := openOptionsToPermMask(&opts)
+	if err := d.Inode.CheckPermission(ctx, fsPerm); err != nil {
+		return nil, err
+	}
+
+	// If they claim it's a directory, then make sure.
+	if strings.HasSuffix(path, "/") {
+		if d.Inode.StableAttr.Type != fs.Directory {
+			return nil, syserror.ENOTDIR
+		}
+	}
+
+	if opts.FileExec && d.Inode.StableAttr.Type != fs.RegularFile {
+		ctx.Infof("%q is not a regular file: %v", path, d.Inode.StableAttr.Type)
+		return nil, syserror.EACCES
+	}
+
+	f, err := d.Inode.GetFile(ctx, d, flagsToFileFlags(opts.Flags))
+	if err != nil {
+		return nil, err
+	}
+
+	return &fsFile{file: f}, nil
+}
+
+func openOptionsToPermMask(opts *vfs.OpenOptions) fs.PermMask {
+	mode := opts.Flags & linux.O_ACCMODE
+	return fs.PermMask{
+		Read:    mode == linux.O_RDONLY || mode == linux.O_RDWR,
+		Write:   mode == linux.O_WRONLY || mode == linux.O_RDWR,
+		Execute: opts.FileExec,
+	}
+}
+
+func flagsToFileFlags(flags uint32) fs.FileFlags {
+	return fs.FileFlags{
+		Direct:      flags&linux.O_DIRECT != 0,
+		DSync:       flags&(linux.O_DSYNC|linux.O_SYNC) != 0,
+		Sync:        flags&linux.O_SYNC != 0,
+		NonBlocking: flags&linux.O_NONBLOCK != 0,
+		Read:        (flags & linux.O_ACCMODE) != linux.O_WRONLY,
+		Write:       (flags & linux.O_ACCMODE) != linux.O_RDONLY,
+		Append:      flags&linux.O_APPEND != 0,
+		Directory:   flags&linux.O_DIRECTORY != 0,
+		Async:       flags&linux.O_ASYNC != 0,
+		LargeFile:   flags&linux.O_LARGEFILE != 0,
+		Truncate:    flags&linux.O_TRUNC != 0,
+	}
+}

pkg/sentry/fsbridge/vfs.go

@@ -0,0 +1,134 @@
+// Copyright 2020 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 fsbridge
+
+import (
+	"io"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/fspath"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/memmap"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// fsFile implements File interface over vfs.FileDescription.
+//
+// +stateify savable
+type vfsFile struct {
+	file *vfs.FileDescription
+}
+
+var _ File = (*vfsFile)(nil)
+
+// NewVFSFile creates a new File over fs.File.
+func NewVFSFile(file *vfs.FileDescription) File {
+	return &vfsFile{file: file}
+}
+
+// PathnameWithDeleted implements File.
+func (f *vfsFile) PathnameWithDeleted(ctx context.Context) string {
+	root := vfs.RootFromContext(ctx)
+	defer root.DecRef()
+
+	vfsObj := f.file.VirtualDentry().Mount().Filesystem().VirtualFilesystem()
+	name, _ := vfsObj.PathnameWithDeleted(ctx, root, f.file.VirtualDentry())
+	return name
+}
+
+// ReadFull implements File.
+func (f *vfsFile) ReadFull(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	var total int64
+	for dst.NumBytes() > 0 {
+		n, err := f.file.PRead(ctx, dst, offset+total, vfs.ReadOptions{})
+		total += n
+		if err == io.EOF && total != 0 {
+			return total, io.ErrUnexpectedEOF
+		} else if err != nil {
+			return total, err
+		}
+		dst = dst.DropFirst64(n)
+	}
+	return total, nil
+}
+
+// ConfigureMMap implements File.
+func (f *vfsFile) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error {
+	return f.file.ConfigureMMap(ctx, opts)
+}
+
+// Type implements File.
+func (f *vfsFile) Type(ctx context.Context) (linux.FileMode, error) {
+	stat, err := f.file.Stat(ctx, vfs.StatOptions{})
+	if err != nil {
+		return 0, err
+	}
+	return linux.FileMode(stat.Mode).FileType(), nil
+}
+
+// IncRef implements File.
+func (f *vfsFile) IncRef() {
+	f.file.IncRef()
+}
+
+// DecRef implements File.
+func (f *vfsFile) DecRef() {
+	f.file.DecRef()
+}
+
+// fsLookup implements Lookup interface using fs.File.
+//
+// +stateify savable
+type vfsLookup struct {
+	mntns *vfs.MountNamespace
+
+	root       vfs.VirtualDentry
+	workingDir vfs.VirtualDentry
+}
+
+var _ Lookup = (*vfsLookup)(nil)
+
+// NewVFSLookup creates a new Lookup using VFS2.
+func NewVFSLookup(mntns *vfs.MountNamespace, root, workingDir vfs.VirtualDentry) Lookup {
+	return &vfsLookup{
+		mntns:      mntns,
+		root:       root,
+		workingDir: workingDir,
+	}
+}
+
+// OpenPath implements Lookup.
+//
+// remainingTraversals is not configurable in VFS2, all callers are using the
+// default anyways.
+//
+// TODO(gvisor.dev/issue/1623): Check mount has read and exec permission.
+func (l *vfsLookup) OpenPath(ctx context.Context, path string, opts vfs.OpenOptions, _ *uint, resolveFinal bool) (File, error) {
+	vfsObj := l.mntns.Root().Mount().Filesystem().VirtualFilesystem()
+	creds := auth.CredentialsFromContext(ctx)
+	pop := &vfs.PathOperation{
+		Root:               l.root,
+		Start:              l.root,
+		Path:               fspath.Parse(path),
+		FollowFinalSymlink: resolveFinal,
+	}
+	fd, err := vfsObj.OpenAt(ctx, creds, pop, &opts)
+	if err != nil {
+		return nil, err
+	}
+	return &vfsFile{file: fd}, nil
+}

pkg/sentry/fsimpl/devtmpfs/devtmpfs.go

@@ -28,6 +28,9 @@ import (
 	"gvisor.dev/gvisor/pkg/sync"
 )
 
+// Name is the default filesystem name.
+const Name = "devtmpfs"
+
 // FilesystemType implements vfs.FilesystemType.
 type FilesystemType struct {
 	initOnce sync.Once
@@ -107,6 +110,7 @@ func (a *Accessor) wrapContext(ctx context.Context) *accessorContext {
 func (ac *accessorContext) Value(key interface{}) interface{} {
 	switch key {
 	case vfs.CtxMountNamespace:
+		ac.a.mntns.IncRef()
 		return ac.a.mntns
 	case vfs.CtxRoot:
 		ac.a.root.IncRef()

pkg/sentry/fsimpl/gofer/filesystem.go

@@ -400,6 +400,7 @@ func (fs *filesystem) unlinkAt(ctx context.Context, rp *vfs.ResolvingPath, dir b
 	}
 	vfsObj := rp.VirtualFilesystem()
 	mntns := vfs.MountNamespaceFromContext(ctx)
+	defer mntns.DecRef()
 	parent.dirMu.Lock()
 	defer parent.dirMu.Unlock()
 	childVFSD := parent.vfsd.Child(name)
@@ -934,7 +935,9 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa
 	if oldParent == newParent && oldName == newName {
 		return nil
 	}
-	if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), &renamed.vfsd, replacedVFSD); err != nil {
+	mntns := vfs.MountNamespaceFromContext(ctx)
+	defer mntns.DecRef()
+	if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil {
 		return err
 	}
 	if err := renamed.file.rename(ctx, newParent.file, newName); err != nil {

pkg/sentry/fsimpl/gofer/gofer.go

@@ -52,6 +52,9 @@ import (
 	"gvisor.dev/gvisor/pkg/usermem"
 )
 
+// Name is the default filesystem name.
+const Name = "9p"
+
 // FilesystemType implements vfs.FilesystemType.
 type FilesystemType struct{}
 

pkg/sentry/fsimpl/kernfs/filesystem.go

@@ -544,6 +544,7 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa
 	}
 
 	mntns := vfs.MountNamespaceFromContext(ctx)
+	defer mntns.DecRef()
 	virtfs := rp.VirtualFilesystem()
 
 	srcDirDentry := srcDirVFSD.Impl().(*Dentry)
@@ -595,7 +596,10 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error
 	parentDentry := vfsd.Parent().Impl().(*Dentry)
 	parentDentry.dirMu.Lock()
 	defer parentDentry.dirMu.Unlock()
-	if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil {
+
+	mntns := vfs.MountNamespaceFromContext(ctx)
+	defer mntns.DecRef()
+	if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil {
 		return err
 	}
 	if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil {
@@ -697,7 +701,9 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error
 	parentDentry := vfsd.Parent().Impl().(*Dentry)
 	parentDentry.dirMu.Lock()
 	defer parentDentry.dirMu.Unlock()
-	if err := virtfs.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), vfsd); err != nil {
+	mntns := vfs.MountNamespaceFromContext(ctx)
+	defer mntns.DecRef()
+	if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil {
 		return err
 	}
 	if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil {

pkg/sentry/fsimpl/proc/BUILD

@@ -14,6 +14,7 @@ go_library(
         "tasks_net.go",
         "tasks_sys.go",
     ],
+    visibility = ["//pkg/sentry:internal"],
     deps = [
         "//pkg/abi/linux",
         "//pkg/context",

pkg/sentry/fsimpl/proc/filesystem.go

@@ -26,15 +26,18 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 )
 
-// procFSType is the factory class for procfs.
+// Name is the default filesystem name.
+const Name = "proc"
+
+// FilesystemType is the factory class for procfs.
 //
 // +stateify savable
-type procFSType struct{}
+type FilesystemType struct{}
 
-var _ vfs.FilesystemType = (*procFSType)(nil)
+var _ vfs.FilesystemType = (*FilesystemType)(nil)
 
 // GetFilesystem implements vfs.FilesystemType.
-func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) {
+func (ft *FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) {
 	k := kernel.KernelFromContext(ctx)
 	if k == nil {
 		return nil, nil, fmt.Errorf("procfs requires a kernel")
@@ -47,12 +50,13 @@ func (ft *procFSType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFile
 	procfs := &kernfs.Filesystem{}
 	procfs.VFSFilesystem().Init(vfsObj, procfs)
 
-	var data *InternalData
+	var cgroups map[string]string
 	if opts.InternalData != nil {
-		data = opts.InternalData.(*InternalData)
+		data := opts.InternalData.(*InternalData)
+		cgroups = data.Cgroups
 	}
 
-	_, dentry := newTasksInode(procfs, k, pidns, data.Cgroups)
+	_, dentry := newTasksInode(procfs, k, pidns, cgroups)
 	return procfs.VFSFilesystem(), dentry.VFSDentry(), nil
 }
 

pkg/sentry/fsimpl/proc/tasks_test.go

@@ -90,8 +90,7 @@ func setup(t *testing.T) *testutil.System {
 	ctx := k.SupervisorContext()
 	creds := auth.CredentialsFromContext(ctx)
 
-	vfsObj := vfs.New()
-	vfsObj.MustRegisterFilesystemType("procfs", &procFSType{}, &vfs.RegisterFilesystemTypeOptions{
+	k.VFS.MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
 	fsOpts := vfs.GetFilesystemOptions{
@@ -102,11 +101,11 @@ func setup(t *testing.T) *testutil.System {
 			},
 		},
 	}
-	mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "procfs", &fsOpts)
+	mntns, err := k.VFS.NewMountNamespace(ctx, creds, "", Name, &fsOpts)
 	if err != nil {
 		t.Fatalf("NewMountNamespace(): %v", err)
 	}
-	return testutil.NewSystem(ctx, t, vfsObj, mntns)
+	return testutil.NewSystem(ctx, t, k.VFS, mntns)
 }
 
 func TestTasksEmpty(t *testing.T) {
@@ -131,7 +130,7 @@ func TestTasks(t *testing.T) {
 	var tasks []*kernel.Task
 	for i := 0; i < 5; i++ {
 		tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits())
-		task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc)
+		task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root)
 		if err != nil {
 			t.Fatalf("CreateTask(): %v", err)
 		}
@@ -213,7 +212,7 @@ func TestTasksOffset(t *testing.T) {
 	k := kernel.KernelFromContext(s.Ctx)
 	for i := 0; i < 3; i++ {
 		tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits())
-		if _, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc); err != nil {
+		if _, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root); err != nil {
 			t.Fatalf("CreateTask(): %v", err)
 		}
 	}
@@ -337,7 +336,7 @@ func TestTask(t *testing.T) {
 
 	k := kernel.KernelFromContext(s.Ctx)
 	tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits())
-	_, err := testutil.CreateTask(s.Ctx, "name", tc)
+	_, err := testutil.CreateTask(s.Ctx, "name", tc, s.MntNs, s.Root, s.Root)
 	if err != nil {
 		t.Fatalf("CreateTask(): %v", err)
 	}
@@ -352,7 +351,7 @@ func TestProcSelf(t *testing.T) {
 
 	k := kernel.KernelFromContext(s.Ctx)
 	tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits())
-	task, err := testutil.CreateTask(s.Ctx, "name", tc)
+	task, err := testutil.CreateTask(s.Ctx, "name", tc, s.MntNs, s.Root, s.Root)
 	if err != nil {
 		t.Fatalf("CreateTask(): %v", err)
 	}
@@ -433,7 +432,7 @@ func TestTree(t *testing.T) {
 	var tasks []*kernel.Task
 	for i := 0; i < 5; i++ {
 		tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits())
-		task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc)
+		task, err := testutil.CreateTask(s.Ctx, fmt.Sprintf("name-%d", i), tc, s.MntNs, s.Root, s.Root)
 		if err != nil {
 			t.Fatalf("CreateTask(): %v", err)
 		}

pkg/sentry/fsimpl/sys/BUILD

@@ -7,6 +7,7 @@ go_library(
     srcs = [
         "sys.go",
     ],
+    visibility = ["//pkg/sentry:internal"],
     deps = [
         "//pkg/abi/linux",
         "//pkg/context",

pkg/sentry/fsimpl/sys/sys.go

@@ -28,6 +28,9 @@ import (
 	"gvisor.dev/gvisor/pkg/syserror"
 )
 
+// Name is the default filesystem name.
+const Name = "sysfs"
+
 // FilesystemType implements vfs.FilesystemType.
 type FilesystemType struct{}
 

pkg/sentry/fsimpl/sys/sys_test.go

@@ -34,16 +34,15 @@ func newTestSystem(t *testing.T) *testutil.System {
 	}
 	ctx := k.SupervisorContext()
 	creds := auth.CredentialsFromContext(ctx)
-	v := vfs.New()
-	v.MustRegisterFilesystemType("sysfs", sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+	k.VFS.MustRegisterFilesystemType(sys.Name, sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
 
-	mns, err := v.NewMountNamespace(ctx, creds, "", "sysfs", &vfs.GetFilesystemOptions{})
+	mns, err := k.VFS.NewMountNamespace(ctx, creds, "", sys.Name, &vfs.GetFilesystemOptions{})
 	if err != nil {
 		t.Fatalf("Failed to create new mount namespace: %v", err)
 	}
-	return testutil.NewSystem(ctx, t, v, mns)
+	return testutil.NewSystem(ctx, t, k.VFS, mns)
 }
 
 func TestReadCPUFile(t *testing.T) {

pkg/sentry/fsimpl/testutil/BUILD

@@ -16,7 +16,7 @@ go_library(
         "//pkg/cpuid",
         "//pkg/fspath",
         "//pkg/memutil",
-        "//pkg/sentry/fs",
+        "//pkg/sentry/fsimpl/tmpfs",
         "//pkg/sentry/kernel",
         "//pkg/sentry/kernel/auth",
         "//pkg/sentry/kernel/sched",

pkg/sentry/fsimpl/testutil/kernel.go

@@ -24,7 +24,7 @@ import (
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/cpuid"
 	"gvisor.dev/gvisor/pkg/memutil"
-	"gvisor.dev/gvisor/pkg/sentry/fs"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/sched"
@@ -33,6 +33,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/pgalloc"
 	"gvisor.dev/gvisor/pkg/sentry/platform"
 	"gvisor.dev/gvisor/pkg/sentry/time"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
 
 	// Platforms are plugable.
 	_ "gvisor.dev/gvisor/pkg/sentry/platform/kvm"
@@ -99,26 +100,27 @@ func Boot() (*kernel.Kernel, error) {
 		return nil, fmt.Errorf("initializing kernel: %v", err)
 	}
 
-	ctx := k.SupervisorContext()
+	kernel.VFS2Enabled = true
+
+	vfsObj := vfs.New()
+	vfsObj.MustRegisterFilesystemType(tmpfs.Name, &tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+		AllowUserMount: true,
+		AllowUserList:  true,
+	})
+	k.VFS = vfsObj
 
-	// Create mount namespace without root as it's the minimum required to create
-	// the global thread group.
-	mntns, err := fs.NewMountNamespace(ctx, nil)
-	if err != nil {
-		return nil, err
-	}
 	ls, err := limits.NewLinuxLimitSet()
 	if err != nil {
 		return nil, err
 	}
-	tg := k.NewThreadGroup(mntns, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls)
+	tg := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, ls)
 	k.TestOnly_SetGlobalInit(tg)
 
 	return k, nil
 }
 
 // CreateTask creates a new bare bones task for tests.
-func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup) (*kernel.Task, error) {
+func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns *vfs.MountNamespace, root, cwd vfs.VirtualDentry) (*kernel.Task, error) {
 	k := kernel.KernelFromContext(ctx)
 	config := &kernel.TaskConfig{
 		Kernel:                  k,
@@ -129,6 +131,8 @@ func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup) (*kern
 		UTSNamespace:            kernel.UTSNamespaceFromContext(ctx),
 		IPCNamespace:            kernel.IPCNamespaceFromContext(ctx),
 		AbstractSocketNamespace: kernel.NewAbstractSocketNamespace(),
+		MountNamespaceVFS2:      mntns,
+		FSContext:               kernel.NewFSContextVFS2(root, cwd, 0022),
 	}
 	return k.TaskSet().NewTask(config)
 }