summaryrefslogtreecommitdiffhomepage
path: root/pkg/sentry/fsimpl/tmpfs
diff options
context:
space:
mode:
Diffstat (limited to 'pkg/sentry/fsimpl/tmpfs')
-rw-r--r--pkg/sentry/fsimpl/tmpfs/BUILD125
-rw-r--r--pkg/sentry/fsimpl/tmpfs/benchmark_test.go486
-rw-r--r--pkg/sentry/fsimpl/tmpfs/device_file.go49
-rw-r--r--pkg/sentry/fsimpl/tmpfs/directory.go232
-rw-r--r--pkg/sentry/fsimpl/tmpfs/filesystem.go860
-rw-r--r--pkg/sentry/fsimpl/tmpfs/named_pipe.go38
-rw-r--r--pkg/sentry/fsimpl/tmpfs/pipe_test.go238
-rw-r--r--pkg/sentry/fsimpl/tmpfs/regular_file.go637
-rw-r--r--pkg/sentry/fsimpl/tmpfs/regular_file_test.go349
-rw-r--r--pkg/sentry/fsimpl/tmpfs/socket_file.go34
-rw-r--r--pkg/sentry/fsimpl/tmpfs/stat_test.go236
-rw-r--r--pkg/sentry/fsimpl/tmpfs/symlink.go37
-rw-r--r--pkg/sentry/fsimpl/tmpfs/tmpfs.go775
-rw-r--r--pkg/sentry/fsimpl/tmpfs/tmpfs_test.go156
14 files changed, 4252 insertions, 0 deletions
diff --git a/pkg/sentry/fsimpl/tmpfs/BUILD b/pkg/sentry/fsimpl/tmpfs/BUILD
new file mode 100644
index 000000000..5cd428d64
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/BUILD
@@ -0,0 +1,125 @@
+load("//tools:defs.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+
+licenses(["notice"])
+
+go_template_instance(
+ name = "dentry_list",
+ out = "dentry_list.go",
+ package = "tmpfs",
+ prefix = "dentry",
+ template = "//pkg/ilist:generic_list",
+ types = {
+ "Element": "*dentry",
+ "Linker": "*dentry",
+ },
+)
+
+go_template_instance(
+ name = "fstree",
+ out = "fstree.go",
+ package = "tmpfs",
+ prefix = "generic",
+ template = "//pkg/sentry/vfs/genericfstree:generic_fstree",
+ types = {
+ "Dentry": "dentry",
+ },
+)
+
+go_template_instance(
+ name = "inode_refs",
+ out = "inode_refs.go",
+ package = "tmpfs",
+ prefix = "inode",
+ template = "//pkg/refs_vfs2:refs_template",
+ types = {
+ "T": "inode",
+ },
+)
+
+go_library(
+ name = "tmpfs",
+ srcs = [
+ "dentry_list.go",
+ "device_file.go",
+ "directory.go",
+ "filesystem.go",
+ "fstree.go",
+ "inode_refs.go",
+ "named_pipe.go",
+ "regular_file.go",
+ "socket_file.go",
+ "symlink.go",
+ "tmpfs.go",
+ ],
+ visibility = ["//pkg/sentry:internal"],
+ deps = [
+ "//pkg/abi/linux",
+ "//pkg/amutex",
+ "//pkg/context",
+ "//pkg/fspath",
+ "//pkg/log",
+ "//pkg/refs",
+ "//pkg/safemem",
+ "//pkg/sentry/arch",
+ "//pkg/sentry/fs",
+ "//pkg/sentry/fs/fsutil",
+ "//pkg/sentry/fs/lock",
+ "//pkg/sentry/kernel/auth",
+ "//pkg/sentry/kernel/pipe",
+ "//pkg/sentry/kernel/time",
+ "//pkg/sentry/memmap",
+ "//pkg/sentry/pgalloc",
+ "//pkg/sentry/platform",
+ "//pkg/sentry/socket/unix/transport",
+ "//pkg/sentry/uniqueid",
+ "//pkg/sentry/usage",
+ "//pkg/sentry/vfs",
+ "//pkg/sentry/vfs/memxattr",
+ "//pkg/sync",
+ "//pkg/syserror",
+ "//pkg/usermem",
+ ],
+)
+
+go_test(
+ name = "benchmark_test",
+ size = "small",
+ srcs = ["benchmark_test.go"],
+ deps = [
+ ":tmpfs",
+ "//pkg/abi/linux",
+ "//pkg/context",
+ "//pkg/fspath",
+ "//pkg/refs",
+ "//pkg/sentry/contexttest",
+ "//pkg/sentry/fs",
+ "//pkg/sentry/fs/tmpfs",
+ "//pkg/sentry/kernel/auth",
+ "//pkg/sentry/vfs",
+ "//pkg/syserror",
+ ],
+)
+
+go_test(
+ name = "tmpfs_test",
+ size = "small",
+ srcs = [
+ "pipe_test.go",
+ "regular_file_test.go",
+ "stat_test.go",
+ "tmpfs_test.go",
+ ],
+ library = ":tmpfs",
+ deps = [
+ "//pkg/abi/linux",
+ "//pkg/context",
+ "//pkg/fspath",
+ "//pkg/sentry/contexttest",
+ "//pkg/sentry/fs/lock",
+ "//pkg/sentry/kernel/auth",
+ "//pkg/sentry/vfs",
+ "//pkg/syserror",
+ "//pkg/usermem",
+ ],
+)
diff --git a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go
new file mode 100644
index 000000000..d263147c2
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go
@@ -0,0 +1,486 @@
+// 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 benchmark_test
+
+import (
+ "fmt"
+ "runtime"
+ "strings"
+ "testing"
+
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
+ "gvisor.dev/gvisor/pkg/fspath"
+ "gvisor.dev/gvisor/pkg/refs"
+ "gvisor.dev/gvisor/pkg/sentry/contexttest"
+ "gvisor.dev/gvisor/pkg/sentry/fs"
+ _ "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs"
+ "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/syserror"
+)
+
+// Differences from stat_benchmark:
+//
+// - Syscall interception, CopyInPath, copyOutStat, and overlayfs overheads are
+// not included.
+//
+// - *MountStat benchmarks use a tmpfs root mount and a tmpfs submount at /tmp.
+// Non-MountStat benchmarks use a tmpfs root mount and no submounts.
+// stat_benchmark uses a varying root mount, a tmpfs submount at /tmp, and a
+// subdirectory /tmp/<top_dir> (assuming TEST_TMPDIR == "/tmp"). Thus
+// stat_benchmark at depth 1 does a comparable amount of work to *MountStat
+// benchmarks at depth 2, and non-MountStat benchmarks at depth 3.
+var depths = []int{1, 2, 3, 8, 64, 100}
+
+const (
+ mountPointName = "tmp"
+ filename = "gvisor_test_temp_0_1557494568"
+)
+
+// This is copied from syscalls/linux/sys_file.go, with the dependency on
+// kernel.Task stripped out.
+func fileOpOn(ctx context.Context, mntns *fs.MountNamespace, root, wd *fs.Dirent, dirFD int32, path string, resolve bool, fn func(root *fs.Dirent, d *fs.Dirent) error) error {
+ var (
+ d *fs.Dirent // The file.
+ rel *fs.Dirent // The relative directory for search (if required.)
+ err error
+ )
+
+ // Extract the working directory (maybe).
+ if len(path) > 0 && path[0] == '/' {
+ // Absolute path; rel can be nil.
+ } else if dirFD == linux.AT_FDCWD {
+ // Need to reference the working directory.
+ rel = wd
+ } else {
+ // Need to extract the given FD.
+ return syserror.EBADF
+ }
+
+ // Lookup the node.
+ remainingTraversals := uint(linux.MaxSymlinkTraversals)
+ if resolve {
+ d, err = mntns.FindInode(ctx, root, rel, path, &remainingTraversals)
+ } else {
+ d, err = mntns.FindLink(ctx, root, rel, path, &remainingTraversals)
+ }
+ if err != nil {
+ return err
+ }
+
+ err = fn(root, d)
+ d.DecRef(ctx)
+ return err
+}
+
+func BenchmarkVFS1TmpfsStat(b *testing.B) {
+ for _, depth := range depths {
+ b.Run(fmt.Sprintf("%d", depth), func(b *testing.B) {
+ ctx := contexttest.Context(b)
+
+ // Create VFS.
+ tmpfsFS, ok := fs.FindFilesystem("tmpfs")
+ if !ok {
+ b.Fatalf("failed to find tmpfs filesystem type")
+ }
+ rootInode, err := tmpfsFS.Mount(ctx, "tmpfs", fs.MountSourceFlags{}, "", nil)
+ if err != nil {
+ b.Fatalf("failed to create tmpfs root mount: %v", err)
+ }
+ mntns, err := fs.NewMountNamespace(ctx, rootInode)
+ if err != nil {
+ b.Fatalf("failed to create mount namespace: %v", err)
+ }
+ defer mntns.DecRef(ctx)
+
+ var filePathBuilder strings.Builder
+ filePathBuilder.WriteByte('/')
+
+ // Create nested directories with given depth.
+ root := mntns.Root()
+ defer root.DecRef(ctx)
+ d := root
+ d.IncRef()
+ defer d.DecRef(ctx)
+ for i := depth; i > 0; i-- {
+ name := fmt.Sprintf("%d", i)
+ if err := d.Inode.CreateDirectory(ctx, d, name, fs.FilePermsFromMode(0755)); err != nil {
+ b.Fatalf("failed to create directory %q: %v", name, err)
+ }
+ next, err := d.Walk(ctx, root, name)
+ if err != nil {
+ b.Fatalf("failed to walk to directory %q: %v", name, err)
+ }
+ d.DecRef(ctx)
+ d = next
+ filePathBuilder.WriteString(name)
+ filePathBuilder.WriteByte('/')
+ }
+
+ // Create the file that will be stat'd.
+ file, err := d.Inode.Create(ctx, d, filename, fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0644))
+ if err != nil {
+ b.Fatalf("failed to create file %q: %v", filename, err)
+ }
+ file.DecRef(ctx)
+ filePathBuilder.WriteString(filename)
+ filePath := filePathBuilder.String()
+
+ dirPath := false
+ runtime.GC()
+ b.ResetTimer()
+ for i := 0; i < b.N; i++ {
+ err := fileOpOn(ctx, mntns, root, root, linux.AT_FDCWD, filePath, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+ if dirPath && !fs.IsDir(d.Inode.StableAttr) {
+ return syserror.ENOTDIR
+ }
+ uattr, err := d.Inode.UnstableAttr(ctx)
+ if err != nil {
+ return err
+ }
+ // Sanity check.
+ if uattr.Perms.User.Execute {
+ b.Fatalf("got wrong permissions (%0o)", uattr.Perms.LinuxMode())
+ }
+ return nil
+ })
+ if err != nil {
+ b.Fatalf("stat(%q) failed: %v", filePath, err)
+ }
+ }
+ // Don't include deferred cleanup in benchmark time.
+ b.StopTimer()
+ })
+ }
+}
+
+func BenchmarkVFS2TmpfsStat(b *testing.B) {
+ for _, depth := range depths {
+ b.Run(fmt.Sprintf("%d", depth), func(b *testing.B) {
+ ctx := contexttest.Context(b)
+ creds := auth.CredentialsFromContext(ctx)
+
+ // Create VFS.
+ vfsObj := vfs.VirtualFilesystem{}
+ if err := vfsObj.Init(ctx); err != nil {
+ b.Fatalf("VFS init: %v", err)
+ }
+ vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserMount: true,
+ })
+ mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{})
+ if err != nil {
+ b.Fatalf("failed to create tmpfs root mount: %v", err)
+ }
+ defer mntns.DecRef(ctx)
+
+ var filePathBuilder strings.Builder
+ filePathBuilder.WriteByte('/')
+
+ // Create nested directories with given depth.
+ root := mntns.Root()
+ defer root.DecRef(ctx)
+ vd := root
+ vd.IncRef()
+ for i := depth; i > 0; i-- {
+ name := fmt.Sprintf("%d", i)
+ pop := vfs.PathOperation{
+ Root: root,
+ Start: vd,
+ Path: fspath.Parse(name),
+ }
+ if err := vfsObj.MkdirAt(ctx, creds, &pop, &vfs.MkdirOptions{
+ Mode: 0755,
+ }); err != nil {
+ b.Fatalf("failed to create directory %q: %v", name, err)
+ }
+ nextVD, err := vfsObj.GetDentryAt(ctx, creds, &pop, &vfs.GetDentryOptions{})
+ if err != nil {
+ b.Fatalf("failed to walk to directory %q: %v", name, err)
+ }
+ vd.DecRef(ctx)
+ vd = nextVD
+ filePathBuilder.WriteString(name)
+ filePathBuilder.WriteByte('/')
+ }
+
+ // Create the file that will be stat'd.
+ fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: vd,
+ Path: fspath.Parse(filename),
+ FollowFinalSymlink: true,
+ }, &vfs.OpenOptions{
+ Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL,
+ Mode: 0644,
+ })
+ vd.DecRef(ctx)
+ vd = vfs.VirtualDentry{}
+ if err != nil {
+ b.Fatalf("failed to create file %q: %v", filename, err)
+ }
+ defer fd.DecRef(ctx)
+ filePathBuilder.WriteString(filename)
+ filePath := filePathBuilder.String()
+
+ runtime.GC()
+ b.ResetTimer()
+ for i := 0; i < b.N; i++ {
+ stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(filePath),
+ FollowFinalSymlink: true,
+ }, &vfs.StatOptions{})
+ if err != nil {
+ b.Fatalf("stat(%q) failed: %v", filePath, err)
+ }
+ // Sanity check.
+ if stat.Mode&^linux.S_IFMT != 0644 {
+ b.Fatalf("got wrong permissions (%0o)", stat.Mode)
+ }
+ }
+ // Don't include deferred cleanup in benchmark time.
+ b.StopTimer()
+ })
+ }
+}
+
+func BenchmarkVFS1TmpfsMountStat(b *testing.B) {
+ for _, depth := range depths {
+ b.Run(fmt.Sprintf("%d", depth), func(b *testing.B) {
+ ctx := contexttest.Context(b)
+
+ // Create VFS.
+ tmpfsFS, ok := fs.FindFilesystem("tmpfs")
+ if !ok {
+ b.Fatalf("failed to find tmpfs filesystem type")
+ }
+ rootInode, err := tmpfsFS.Mount(ctx, "tmpfs", fs.MountSourceFlags{}, "", nil)
+ if err != nil {
+ b.Fatalf("failed to create tmpfs root mount: %v", err)
+ }
+ mntns, err := fs.NewMountNamespace(ctx, rootInode)
+ if err != nil {
+ b.Fatalf("failed to create mount namespace: %v", err)
+ }
+ defer mntns.DecRef(ctx)
+
+ var filePathBuilder strings.Builder
+ filePathBuilder.WriteByte('/')
+
+ // Create and mount the submount.
+ root := mntns.Root()
+ defer root.DecRef(ctx)
+ if err := root.Inode.CreateDirectory(ctx, root, mountPointName, fs.FilePermsFromMode(0755)); err != nil {
+ b.Fatalf("failed to create mount point: %v", err)
+ }
+ mountPoint, err := root.Walk(ctx, root, mountPointName)
+ if err != nil {
+ b.Fatalf("failed to walk to mount point: %v", err)
+ }
+ defer mountPoint.DecRef(ctx)
+ submountInode, err := tmpfsFS.Mount(ctx, "tmpfs", fs.MountSourceFlags{}, "", nil)
+ if err != nil {
+ b.Fatalf("failed to create tmpfs submount: %v", err)
+ }
+ if err := mntns.Mount(ctx, mountPoint, submountInode); err != nil {
+ b.Fatalf("failed to mount tmpfs submount: %v", err)
+ }
+ filePathBuilder.WriteString(mountPointName)
+ filePathBuilder.WriteByte('/')
+
+ // Create nested directories with given depth.
+ d, err := root.Walk(ctx, root, mountPointName)
+ if err != nil {
+ b.Fatalf("failed to walk to mount root: %v", err)
+ }
+ defer d.DecRef(ctx)
+ for i := depth; i > 0; i-- {
+ name := fmt.Sprintf("%d", i)
+ if err := d.Inode.CreateDirectory(ctx, d, name, fs.FilePermsFromMode(0755)); err != nil {
+ b.Fatalf("failed to create directory %q: %v", name, err)
+ }
+ next, err := d.Walk(ctx, root, name)
+ if err != nil {
+ b.Fatalf("failed to walk to directory %q: %v", name, err)
+ }
+ d.DecRef(ctx)
+ d = next
+ filePathBuilder.WriteString(name)
+ filePathBuilder.WriteByte('/')
+ }
+
+ // Create the file that will be stat'd.
+ file, err := d.Inode.Create(ctx, d, filename, fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0644))
+ if err != nil {
+ b.Fatalf("failed to create file %q: %v", filename, err)
+ }
+ file.DecRef(ctx)
+ filePathBuilder.WriteString(filename)
+ filePath := filePathBuilder.String()
+
+ dirPath := false
+ runtime.GC()
+ b.ResetTimer()
+ for i := 0; i < b.N; i++ {
+ err := fileOpOn(ctx, mntns, root, root, linux.AT_FDCWD, filePath, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+ if dirPath && !fs.IsDir(d.Inode.StableAttr) {
+ return syserror.ENOTDIR
+ }
+ uattr, err := d.Inode.UnstableAttr(ctx)
+ if err != nil {
+ return err
+ }
+ // Sanity check.
+ if uattr.Perms.User.Execute {
+ b.Fatalf("got wrong permissions (%0o)", uattr.Perms.LinuxMode())
+ }
+ return nil
+ })
+ if err != nil {
+ b.Fatalf("stat(%q) failed: %v", filePath, err)
+ }
+ }
+ // Don't include deferred cleanup in benchmark time.
+ b.StopTimer()
+ })
+ }
+}
+
+func BenchmarkVFS2TmpfsMountStat(b *testing.B) {
+ for _, depth := range depths {
+ b.Run(fmt.Sprintf("%d", depth), func(b *testing.B) {
+ ctx := contexttest.Context(b)
+ creds := auth.CredentialsFromContext(ctx)
+
+ // Create VFS.
+ vfsObj := vfs.VirtualFilesystem{}
+ if err := vfsObj.Init(ctx); err != nil {
+ b.Fatalf("VFS init: %v", err)
+ }
+ vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserMount: true,
+ })
+ mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{})
+ if err != nil {
+ b.Fatalf("failed to create tmpfs root mount: %v", err)
+ }
+ defer mntns.DecRef(ctx)
+
+ var filePathBuilder strings.Builder
+ filePathBuilder.WriteByte('/')
+
+ // Create the mount point.
+ root := mntns.Root()
+ defer root.DecRef(ctx)
+ pop := vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(mountPointName),
+ }
+ if err := vfsObj.MkdirAt(ctx, creds, &pop, &vfs.MkdirOptions{
+ Mode: 0755,
+ }); err != nil {
+ b.Fatalf("failed to create mount point: %v", err)
+ }
+ // Save the mount point for later use.
+ mountPoint, err := vfsObj.GetDentryAt(ctx, creds, &pop, &vfs.GetDentryOptions{})
+ if err != nil {
+ b.Fatalf("failed to walk to mount point: %v", err)
+ }
+ defer mountPoint.DecRef(ctx)
+ // Create and mount the submount.
+ if err := vfsObj.MountAt(ctx, creds, "", &pop, "tmpfs", &vfs.MountOptions{}); err != nil {
+ b.Fatalf("failed to mount tmpfs submount: %v", err)
+ }
+ filePathBuilder.WriteString(mountPointName)
+ filePathBuilder.WriteByte('/')
+
+ // Create nested directories with given depth.
+ vd, err := vfsObj.GetDentryAt(ctx, creds, &pop, &vfs.GetDentryOptions{})
+ if err != nil {
+ b.Fatalf("failed to walk to mount root: %v", err)
+ }
+ for i := depth; i > 0; i-- {
+ name := fmt.Sprintf("%d", i)
+ pop := vfs.PathOperation{
+ Root: root,
+ Start: vd,
+ Path: fspath.Parse(name),
+ }
+ if err := vfsObj.MkdirAt(ctx, creds, &pop, &vfs.MkdirOptions{
+ Mode: 0755,
+ }); err != nil {
+ b.Fatalf("failed to create directory %q: %v", name, err)
+ }
+ nextVD, err := vfsObj.GetDentryAt(ctx, creds, &pop, &vfs.GetDentryOptions{})
+ if err != nil {
+ b.Fatalf("failed to walk to directory %q: %v", name, err)
+ }
+ vd.DecRef(ctx)
+ vd = nextVD
+ filePathBuilder.WriteString(name)
+ filePathBuilder.WriteByte('/')
+ }
+
+ // Create the file that will be stat'd.
+ fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: vd,
+ Path: fspath.Parse(filename),
+ FollowFinalSymlink: true,
+ }, &vfs.OpenOptions{
+ Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL,
+ Mode: 0644,
+ })
+ vd.DecRef(ctx)
+ if err != nil {
+ b.Fatalf("failed to create file %q: %v", filename, err)
+ }
+ fd.DecRef(ctx)
+ filePathBuilder.WriteString(filename)
+ filePath := filePathBuilder.String()
+
+ runtime.GC()
+ b.ResetTimer()
+ for i := 0; i < b.N; i++ {
+ stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(filePath),
+ FollowFinalSymlink: true,
+ }, &vfs.StatOptions{})
+ if err != nil {
+ b.Fatalf("stat(%q) failed: %v", filePath, err)
+ }
+ // Sanity check.
+ if stat.Mode&^linux.S_IFMT != 0644 {
+ b.Fatalf("got wrong permissions (%0o)", stat.Mode)
+ }
+ }
+ // Don't include deferred cleanup in benchmark time.
+ b.StopTimer()
+ })
+ }
+}
+
+func init() {
+ // Turn off reference leak checking for a fair comparison between vfs1 and
+ // vfs2.
+ refs.SetLeakMode(refs.NoLeakChecking)
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/device_file.go b/pkg/sentry/fsimpl/tmpfs/device_file.go
new file mode 100644
index 000000000..ac54d420d
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/device_file.go
@@ -0,0 +1,49 @@
+// 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 tmpfs
+
+import (
+ "fmt"
+
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+)
+
+type deviceFile struct {
+ inode inode
+ kind vfs.DeviceKind
+ major uint32
+ minor uint32
+}
+
+func (fs *filesystem) newDeviceFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode, kind vfs.DeviceKind, major, minor uint32) *inode {
+ file := &deviceFile{
+ kind: kind,
+ major: major,
+ minor: minor,
+ }
+ switch kind {
+ case vfs.BlockDevice:
+ mode |= linux.S_IFBLK
+ case vfs.CharDevice:
+ mode |= linux.S_IFCHR
+ default:
+ panic(fmt.Sprintf("invalid DeviceKind: %v", kind))
+ }
+ file.inode.init(file, fs, kuid, kgid, mode)
+ file.inode.nlink = 1 // from parent directory
+ return &file.inode
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go
new file mode 100644
index 000000000..78b4fc5be
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/directory.go
@@ -0,0 +1,232 @@
+// 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 tmpfs
+
+import (
+ "sync/atomic"
+
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/sync"
+ "gvisor.dev/gvisor/pkg/syserror"
+)
+
+type directory struct {
+ // Since directories can't be hard-linked, each directory can only be
+ // associated with a single dentry, which we can store in the directory
+ // struct.
+ dentry dentry
+ inode inode
+
+ // childMap maps the names of the directory's children to their dentries.
+ // childMap is protected by filesystem.mu.
+ childMap map[string]*dentry
+
+ // numChildren is len(childMap), but accessed using atomic memory
+ // operations to avoid locking in inode.statTo().
+ numChildren int64
+
+ // childList is a list containing (1) child dentries and (2) fake dentries
+ // (with inode == nil) that represent the iteration position of
+ // directoryFDs. childList is used to support directoryFD.IterDirents()
+ // efficiently. childList is protected by iterMu.
+ iterMu sync.Mutex
+ childList dentryList
+}
+
+func (fs *filesystem) newDirectory(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *directory {
+ dir := &directory{}
+ dir.inode.init(dir, fs, kuid, kgid, linux.S_IFDIR|mode)
+ dir.inode.nlink = 2 // from "." and parent directory or ".." for root
+ dir.dentry.inode = &dir.inode
+ dir.dentry.vfsd.Init(&dir.dentry)
+ return dir
+}
+
+// Preconditions: filesystem.mu must be locked for writing. dir must not
+// already contain a child with the given name.
+func (dir *directory) insertChildLocked(child *dentry, name string) {
+ child.parent = &dir.dentry
+ child.name = name
+ if dir.childMap == nil {
+ dir.childMap = make(map[string]*dentry)
+ }
+ dir.childMap[name] = child
+ atomic.AddInt64(&dir.numChildren, 1)
+ dir.iterMu.Lock()
+ dir.childList.PushBack(child)
+ dir.iterMu.Unlock()
+}
+
+// Preconditions: filesystem.mu must be locked for writing.
+func (dir *directory) removeChildLocked(child *dentry) {
+ delete(dir.childMap, child.name)
+ atomic.AddInt64(&dir.numChildren, -1)
+ dir.iterMu.Lock()
+ dir.childList.Remove(child)
+ dir.iterMu.Unlock()
+}
+
+func (dir *directory) mayDelete(creds *auth.Credentials, child *dentry) error {
+ return vfs.CheckDeleteSticky(creds, linux.FileMode(atomic.LoadUint32(&dir.inode.mode)), auth.KUID(atomic.LoadUint32(&child.inode.uid)))
+}
+
+type directoryFD struct {
+ fileDescription
+ vfs.DirectoryFileDescriptionDefaultImpl
+
+ // Protected by directory.iterMu.
+ iter *dentry
+ off int64
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (fd *directoryFD) Release(ctx context.Context) {
+ if fd.iter != nil {
+ dir := fd.inode().impl.(*directory)
+ dir.iterMu.Lock()
+ dir.childList.Remove(fd.iter)
+ dir.iterMu.Unlock()
+ fd.iter = nil
+ }
+}
+
+// IterDirents implements vfs.FileDescriptionImpl.IterDirents.
+func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error {
+ fs := fd.filesystem()
+ dir := fd.inode().impl.(*directory)
+
+ defer fd.dentry().InotifyWithParent(ctx, linux.IN_ACCESS, 0, vfs.PathEvent)
+
+ // fs.mu is required to read d.parent and dentry.name.
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ dir.iterMu.Lock()
+ defer dir.iterMu.Unlock()
+
+ fd.inode().touchAtime(fd.vfsfd.Mount())
+
+ if fd.off == 0 {
+ if err := cb.Handle(vfs.Dirent{
+ Name: ".",
+ Type: linux.DT_DIR,
+ Ino: dir.inode.ino,
+ NextOff: 1,
+ }); err != nil {
+ return err
+ }
+ fd.off++
+ }
+
+ if fd.off == 1 {
+ parentInode := genericParentOrSelf(&dir.dentry).inode
+ if err := cb.Handle(vfs.Dirent{
+ Name: "..",
+ Type: parentInode.direntType(),
+ Ino: parentInode.ino,
+ NextOff: 2,
+ }); err != nil {
+ return err
+ }
+ fd.off++
+ }
+
+ var child *dentry
+ if fd.iter == nil {
+ // Start iteration at the beginning of dir.
+ child = dir.childList.Front()
+ fd.iter = &dentry{}
+ } else {
+ // Continue iteration from where we left off.
+ child = fd.iter.Next()
+ dir.childList.Remove(fd.iter)
+ }
+ for child != nil {
+ // Skip other directoryFD iterators.
+ if child.inode != nil {
+ if err := cb.Handle(vfs.Dirent{
+ Name: child.name,
+ Type: child.inode.direntType(),
+ Ino: child.inode.ino,
+ NextOff: fd.off + 1,
+ }); err != nil {
+ dir.childList.InsertBefore(child, fd.iter)
+ return err
+ }
+ fd.off++
+ }
+ child = child.Next()
+ }
+ dir.childList.PushBack(fd.iter)
+ return nil
+}
+
+// Seek implements vfs.FileDescriptionImpl.Seek.
+func (fd *directoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
+ dir := fd.inode().impl.(*directory)
+ dir.iterMu.Lock()
+ defer dir.iterMu.Unlock()
+
+ switch whence {
+ case linux.SEEK_SET:
+ // Use offset as given.
+ case linux.SEEK_CUR:
+ offset += fd.off
+ default:
+ return 0, syserror.EINVAL
+ }
+ if offset < 0 {
+ return 0, syserror.EINVAL
+ }
+
+ // If the offset isn't changing (e.g. due to lseek(0, SEEK_CUR)), don't
+ // seek even if doing so might reposition the iterator due to concurrent
+ // mutation of the directory. Compare fs/libfs.c:dcache_dir_lseek().
+ if fd.off == offset {
+ return offset, nil
+ }
+
+ fd.off = offset
+ // Compensate for "." and "..".
+ remChildren := int64(0)
+ if offset >= 2 {
+ remChildren = offset - 2
+ }
+
+ // Ensure that fd.iter exists and is not linked into dir.childList.
+ if fd.iter == nil {
+ fd.iter = &dentry{}
+ } else {
+ dir.childList.Remove(fd.iter)
+ }
+ // Insert fd.iter before the remChildren'th child, or at the end of the
+ // list if remChildren >= number of children.
+ child := dir.childList.Front()
+ for child != nil {
+ // Skip other directoryFD iterators.
+ if child.inode != nil {
+ if remChildren == 0 {
+ dir.childList.InsertBefore(child, fd.iter)
+ return offset, nil
+ }
+ remChildren--
+ }
+ child = child.Next()
+ }
+ dir.childList.PushBack(fd.iter)
+ return offset, nil
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go
new file mode 100644
index 000000000..cb8b2d944
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go
@@ -0,0 +1,860 @@
+// 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 tmpfs
+
+import (
+ "fmt"
+ "sync/atomic"
+
+ "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/socket/unix/transport"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/syserror"
+ "gvisor.dev/gvisor/pkg/usermem"
+)
+
+// Sync implements vfs.FilesystemImpl.Sync.
+func (fs *filesystem) Sync(ctx context.Context) error {
+ // All filesystem state is in-memory.
+ return nil
+}
+
+// stepLocked resolves rp.Component() to an existing file, starting from the
+// given directory.
+//
+// stepLocked is loosely analogous to fs/namei.c:walk_component().
+//
+// Preconditions: filesystem.mu must be locked. !rp.Done().
+func stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry) (*dentry, error) {
+ dir, ok := d.inode.impl.(*directory)
+ if !ok {
+ return nil, syserror.ENOTDIR
+ }
+ if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil {
+ return nil, err
+ }
+afterSymlink:
+ name := rp.Component()
+ if name == "." {
+ rp.Advance()
+ return d, nil
+ }
+ if name == ".." {
+ if isRoot, err := rp.CheckRoot(ctx, &d.vfsd); err != nil {
+ return nil, err
+ } else if isRoot || d.parent == nil {
+ rp.Advance()
+ return d, nil
+ }
+ if err := rp.CheckMount(ctx, &d.parent.vfsd); err != nil {
+ return nil, err
+ }
+ rp.Advance()
+ return d.parent, nil
+ }
+ if len(name) > linux.NAME_MAX {
+ return nil, syserror.ENAMETOOLONG
+ }
+ child, ok := dir.childMap[name]
+ if !ok {
+ return nil, syserror.ENOENT
+ }
+ if err := rp.CheckMount(ctx, &child.vfsd); err != nil {
+ return nil, err
+ }
+ if symlink, ok := child.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() {
+ // Symlink traversal updates access time.
+ child.inode.touchAtime(rp.Mount())
+ if err := rp.HandleSymlink(symlink.target); err != nil {
+ return nil, err
+ }
+ goto afterSymlink // don't check the current directory again
+ }
+ rp.Advance()
+ return child, nil
+}
+
+// walkParentDirLocked resolves all but the last path component of rp to an
+// existing directory, starting from the given directory (which is usually
+// rp.Start().Impl().(*dentry)). It does not check that the returned directory
+// is searchable by the provider of rp.
+//
+// walkParentDirLocked is loosely analogous to Linux's
+// fs/namei.c:path_parentat().
+//
+// Preconditions: filesystem.mu must be locked. !rp.Done().
+func walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry) (*directory, error) {
+ for !rp.Final() {
+ next, err := stepLocked(ctx, rp, d)
+ if err != nil {
+ return nil, err
+ }
+ d = next
+ }
+ dir, ok := d.inode.impl.(*directory)
+ if !ok {
+ return nil, syserror.ENOTDIR
+ }
+ return dir, nil
+}
+
+// resolveLocked resolves rp to an existing file.
+//
+// resolveLocked is loosely analogous to Linux's fs/namei.c:path_lookupat().
+//
+// Preconditions: filesystem.mu must be locked.
+func resolveLocked(ctx context.Context, rp *vfs.ResolvingPath) (*dentry, error) {
+ d := rp.Start().Impl().(*dentry)
+ for !rp.Done() {
+ next, err := stepLocked(ctx, rp, d)
+ if err != nil {
+ return nil, err
+ }
+ d = next
+ }
+ if rp.MustBeDir() && !d.inode.isDir() {
+ return nil, syserror.ENOTDIR
+ }
+ return d, nil
+}
+
+// doCreateAt checks that creating a file at rp is permitted, then invokes
+// create to do so.
+//
+// doCreateAt is loosely analogous to a conjunction of Linux's
+// fs/namei.c:filename_create() and done_path_create().
+//
+// Preconditions: !rp.Done(). For the final path component in rp,
+// !rp.ShouldFollowSymlink().
+func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, create func(parentDir *directory, name string) error) error {
+ fs.mu.Lock()
+ defer fs.mu.Unlock()
+ parentDir, err := walkParentDirLocked(ctx, rp, rp.Start().Impl().(*dentry))
+ if err != nil {
+ return err
+ }
+ if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+ return err
+ }
+ name := rp.Component()
+ if name == "." || name == ".." {
+ return syserror.EEXIST
+ }
+ if len(name) > linux.NAME_MAX {
+ return syserror.ENAMETOOLONG
+ }
+ if _, ok := parentDir.childMap[name]; ok {
+ return syserror.EEXIST
+ }
+ if !dir && rp.MustBeDir() {
+ return syserror.ENOENT
+ }
+ // tmpfs never calls VFS.InvalidateDentry(), so parentDir.dentry can only
+ // be dead if it was deleted.
+ if parentDir.dentry.vfsd.IsDead() {
+ return syserror.ENOENT
+ }
+ mnt := rp.Mount()
+ if err := mnt.CheckBeginWrite(); err != nil {
+ return err
+ }
+ defer mnt.EndWrite()
+ if err := create(parentDir, name); err != nil {
+ return err
+ }
+
+ ev := linux.IN_CREATE
+ if dir {
+ ev |= linux.IN_ISDIR
+ }
+ parentDir.inode.watches.Notify(ctx, name, uint32(ev), 0, vfs.InodeEvent, false /* unlinked */)
+ parentDir.inode.touchCMtime()
+ return nil
+}
+
+// AccessAt implements vfs.Filesystem.Impl.AccessAt.
+func (fs *filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds *auth.Credentials, ats vfs.AccessTypes) error {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ return err
+ }
+ return d.inode.checkPermissions(creds, ats)
+}
+
+// GetDentryAt implements vfs.FilesystemImpl.GetDentryAt.
+func (fs *filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetDentryOptions) (*vfs.Dentry, error) {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ return nil, err
+ }
+ if opts.CheckSearchable {
+ if !d.inode.isDir() {
+ return nil, syserror.ENOTDIR
+ }
+ if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil {
+ return nil, err
+ }
+ }
+ d.IncRef()
+ return &d.vfsd, nil
+}
+
+// GetParentDentryAt implements vfs.FilesystemImpl.GetParentDentryAt.
+func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, error) {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ dir, err := walkParentDirLocked(ctx, rp, rp.Start().Impl().(*dentry))
+ if err != nil {
+ return nil, err
+ }
+ dir.dentry.IncRef()
+ return &dir.dentry.vfsd, nil
+}
+
+// LinkAt implements vfs.FilesystemImpl.LinkAt.
+func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error {
+ return fs.doCreateAt(ctx, rp, false /* dir */, func(parentDir *directory, name string) error {
+ if rp.Mount() != vd.Mount() {
+ return syserror.EXDEV
+ }
+ d := vd.Dentry().Impl().(*dentry)
+ i := d.inode
+ if i.isDir() {
+ return syserror.EPERM
+ }
+ if err := vfs.MayLink(auth.CredentialsFromContext(ctx), linux.FileMode(atomic.LoadUint32(&i.mode)), auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))); err != nil {
+ return err
+ }
+ if i.nlink == 0 {
+ return syserror.ENOENT
+ }
+ if i.nlink == maxLinks {
+ return syserror.EMLINK
+ }
+ i.incLinksLocked()
+ i.watches.Notify(ctx, "", linux.IN_ATTRIB, 0, vfs.InodeEvent, false /* unlinked */)
+ parentDir.insertChildLocked(fs.newDentry(i), name)
+ return nil
+ })
+}
+
+// MkdirAt implements vfs.FilesystemImpl.MkdirAt.
+func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error {
+ return fs.doCreateAt(ctx, rp, true /* dir */, func(parentDir *directory, name string) error {
+ creds := rp.Credentials()
+ if parentDir.inode.nlink == maxLinks {
+ return syserror.EMLINK
+ }
+ parentDir.inode.incLinksLocked() // from child's ".."
+ childDir := fs.newDirectory(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode)
+ parentDir.insertChildLocked(&childDir.dentry, name)
+ return nil
+ })
+}
+
+// MknodAt implements vfs.FilesystemImpl.MknodAt.
+func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error {
+ return fs.doCreateAt(ctx, rp, false /* dir */, func(parentDir *directory, name string) error {
+ creds := rp.Credentials()
+ var childInode *inode
+ switch opts.Mode.FileType() {
+ case linux.S_IFREG:
+ childInode = fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode)
+ case linux.S_IFIFO:
+ childInode = fs.newNamedPipe(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode)
+ case linux.S_IFBLK:
+ childInode = fs.newDeviceFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode, vfs.BlockDevice, opts.DevMajor, opts.DevMinor)
+ case linux.S_IFCHR:
+ childInode = fs.newDeviceFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode, vfs.CharDevice, opts.DevMajor, opts.DevMinor)
+ case linux.S_IFSOCK:
+ childInode = fs.newSocketFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode, opts.Endpoint)
+ default:
+ return syserror.EINVAL
+ }
+ child := fs.newDentry(childInode)
+ parentDir.insertChildLocked(child, name)
+ return nil
+ })
+}
+
+// OpenAt implements vfs.FilesystemImpl.OpenAt.
+func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) {
+ if opts.Flags&linux.O_TMPFILE != 0 {
+ // Not yet supported.
+ return nil, syserror.EOPNOTSUPP
+ }
+
+ // Handle O_CREAT and !O_CREAT separately, since in the latter case we
+ // don't need fs.mu for writing.
+ if opts.Flags&linux.O_CREAT == 0 {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ return nil, err
+ }
+ return d.open(ctx, rp, &opts, false /* afterCreate */)
+ }
+
+ mustCreate := opts.Flags&linux.O_EXCL != 0
+ start := rp.Start().Impl().(*dentry)
+ fs.mu.Lock()
+ defer fs.mu.Unlock()
+ if rp.Done() {
+ // Reject attempts to open mount root directory with O_CREAT.
+ if rp.MustBeDir() {
+ return nil, syserror.EISDIR
+ }
+ if mustCreate {
+ return nil, syserror.EEXIST
+ }
+ return start.open(ctx, rp, &opts, false /* afterCreate */)
+ }
+afterTrailingSymlink:
+ parentDir, err := walkParentDirLocked(ctx, rp, start)
+ if err != nil {
+ return nil, err
+ }
+ // Check for search permission in the parent directory.
+ if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil {
+ return nil, err
+ }
+ // Reject attempts to open directories with O_CREAT.
+ if rp.MustBeDir() {
+ return nil, syserror.EISDIR
+ }
+ name := rp.Component()
+ if name == "." || name == ".." {
+ return nil, syserror.EISDIR
+ }
+ if len(name) > linux.NAME_MAX {
+ return nil, syserror.ENAMETOOLONG
+ }
+ // Determine whether or not we need to create a file.
+ child, ok := parentDir.childMap[name]
+ if !ok {
+ // Already checked for searchability above; now check for writability.
+ if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil {
+ return nil, err
+ }
+ if err := rp.Mount().CheckBeginWrite(); err != nil {
+ return nil, err
+ }
+ defer rp.Mount().EndWrite()
+ // Create and open the child.
+ creds := rp.Credentials()
+ child := fs.newDentry(fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, opts.Mode))
+ parentDir.insertChildLocked(child, name)
+ fd, err := child.open(ctx, rp, &opts, true)
+ if err != nil {
+ return nil, err
+ }
+ parentDir.inode.watches.Notify(ctx, name, linux.IN_CREATE, 0, vfs.PathEvent, false /* unlinked */)
+ parentDir.inode.touchCMtime()
+ return fd, nil
+ }
+ if mustCreate {
+ return nil, syserror.EEXIST
+ }
+ // Is the file mounted over?
+ if err := rp.CheckMount(ctx, &child.vfsd); err != nil {
+ return nil, err
+ }
+ // Do we need to resolve a trailing symlink?
+ if symlink, ok := child.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() {
+ // Symlink traversal updates access time.
+ child.inode.touchAtime(rp.Mount())
+ if err := rp.HandleSymlink(symlink.target); err != nil {
+ return nil, err
+ }
+ start = &parentDir.dentry
+ goto afterTrailingSymlink
+ }
+ if rp.MustBeDir() && !child.inode.isDir() {
+ return nil, syserror.ENOTDIR
+ }
+ return child.open(ctx, rp, &opts, false)
+}
+
+func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions, afterCreate bool) (*vfs.FileDescription, error) {
+ ats := vfs.AccessTypesForOpenFlags(opts)
+ if !afterCreate {
+ if err := d.inode.checkPermissions(rp.Credentials(), ats); err != nil {
+ return nil, err
+ }
+ }
+ switch impl := d.inode.impl.(type) {
+ case *regularFile:
+ var fd regularFileFD
+ fd.LockFD.Init(&d.inode.locks)
+ if err := fd.vfsfd.Init(&fd, opts.Flags, rp.Mount(), &d.vfsd, &vfs.FileDescriptionOptions{AllowDirectIO: true}); err != nil {
+ return nil, err
+ }
+ if !afterCreate && opts.Flags&linux.O_TRUNC != 0 {
+ if _, err := impl.truncate(0); err != nil {
+ return nil, err
+ }
+ }
+ return &fd.vfsfd, nil
+ case *directory:
+ // Can't open directories writably.
+ if ats&vfs.MayWrite != 0 {
+ return nil, syserror.EISDIR
+ }
+ var fd directoryFD
+ fd.LockFD.Init(&d.inode.locks)
+ if err := fd.vfsfd.Init(&fd, opts.Flags, rp.Mount(), &d.vfsd, &vfs.FileDescriptionOptions{AllowDirectIO: true}); err != nil {
+ return nil, err
+ }
+ return &fd.vfsfd, nil
+ case *symlink:
+ // TODO(gvisor.dev/issue/2782): Can't open symlinks without O_PATH.
+ return nil, syserror.ELOOP
+ case *namedPipe:
+ return impl.pipe.Open(ctx, rp.Mount(), &d.vfsd, opts.Flags, &d.inode.locks)
+ case *deviceFile:
+ return rp.VirtualFilesystem().OpenDeviceSpecialFile(ctx, rp.Mount(), &d.vfsd, impl.kind, impl.major, impl.minor, opts)
+ case *socketFile:
+ return nil, syserror.ENXIO
+ default:
+ panic(fmt.Sprintf("unknown inode type: %T", d.inode.impl))
+ }
+}
+
+// ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt.
+func (fs *filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ return "", err
+ }
+ symlink, ok := d.inode.impl.(*symlink)
+ if !ok {
+ return "", syserror.EINVAL
+ }
+ symlink.inode.touchAtime(rp.Mount())
+ return symlink.target, nil
+}
+
+// RenameAt implements vfs.FilesystemImpl.RenameAt.
+func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldParentVD vfs.VirtualDentry, oldName string, opts vfs.RenameOptions) error {
+ if opts.Flags != 0 {
+ // TODO(b/145974740): Support renameat2 flags.
+ return syserror.EINVAL
+ }
+
+ // Resolve newParent first to verify that it's on this Mount.
+ fs.mu.Lock()
+ defer fs.mu.Unlock()
+ newParentDir, err := walkParentDirLocked(ctx, rp, rp.Start().Impl().(*dentry))
+ if err != nil {
+ return err
+ }
+ newName := rp.Component()
+ if newName == "." || newName == ".." {
+ return syserror.EBUSY
+ }
+ mnt := rp.Mount()
+ if mnt != oldParentVD.Mount() {
+ return syserror.EXDEV
+ }
+ if err := mnt.CheckBeginWrite(); err != nil {
+ return err
+ }
+ defer mnt.EndWrite()
+
+ oldParentDir := oldParentVD.Dentry().Impl().(*dentry).inode.impl.(*directory)
+ if err := oldParentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+ return err
+ }
+ renamed, ok := oldParentDir.childMap[oldName]
+ if !ok {
+ return syserror.ENOENT
+ }
+ if err := oldParentDir.mayDelete(rp.Credentials(), renamed); err != nil {
+ return err
+ }
+ // Note that we don't need to call rp.CheckMount(), since if renamed is a
+ // mount point then we want to rename the mount point, not anything in the
+ // mounted filesystem.
+ if renamed.inode.isDir() {
+ if renamed == &newParentDir.dentry || genericIsAncestorDentry(renamed, &newParentDir.dentry) {
+ return syserror.EINVAL
+ }
+ if oldParentDir != newParentDir {
+ // Writability is needed to change renamed's "..".
+ if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil {
+ return err
+ }
+ }
+ } else {
+ if opts.MustBeDir || rp.MustBeDir() {
+ return syserror.ENOTDIR
+ }
+ }
+
+ if err := newParentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+ return err
+ }
+ replaced, ok := newParentDir.childMap[newName]
+ if ok {
+ replacedDir, ok := replaced.inode.impl.(*directory)
+ if ok {
+ if !renamed.inode.isDir() {
+ return syserror.EISDIR
+ }
+ if len(replacedDir.childMap) != 0 {
+ return syserror.ENOTEMPTY
+ }
+ } else {
+ if rp.MustBeDir() {
+ return syserror.ENOTDIR
+ }
+ if renamed.inode.isDir() {
+ return syserror.ENOTDIR
+ }
+ }
+ } else {
+ if renamed.inode.isDir() && newParentDir.inode.nlink == maxLinks {
+ return syserror.EMLINK
+ }
+ }
+ // tmpfs never calls VFS.InvalidateDentry(), so newParentDir.dentry can
+ // only be dead if it was deleted.
+ if newParentDir.dentry.vfsd.IsDead() {
+ return syserror.ENOENT
+ }
+
+ // Linux places this check before some of those above; we do it here for
+ // simplicity, under the assumption that applications are not intentionally
+ // doing noop renames expecting them to succeed where non-noop renames
+ // would fail.
+ if renamed == replaced {
+ return nil
+ }
+ vfsObj := rp.VirtualFilesystem()
+ mntns := vfs.MountNamespaceFromContext(ctx)
+ defer mntns.DecRef(ctx)
+ var replacedVFSD *vfs.Dentry
+ if replaced != nil {
+ replacedVFSD = &replaced.vfsd
+ }
+ if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil {
+ return err
+ }
+ if replaced != nil {
+ newParentDir.removeChildLocked(replaced)
+ if replaced.inode.isDir() {
+ // Remove links for replaced/. and replaced/..
+ replaced.inode.decLinksLocked(ctx)
+ newParentDir.inode.decLinksLocked(ctx)
+ }
+ replaced.inode.decLinksLocked(ctx)
+ }
+ oldParentDir.removeChildLocked(renamed)
+ newParentDir.insertChildLocked(renamed, newName)
+ vfsObj.CommitRenameReplaceDentry(ctx, &renamed.vfsd, replacedVFSD)
+ oldParentDir.inode.touchCMtime()
+ if oldParentDir != newParentDir {
+ if renamed.inode.isDir() {
+ oldParentDir.inode.decLinksLocked(ctx)
+ newParentDir.inode.incLinksLocked()
+ }
+ newParentDir.inode.touchCMtime()
+ }
+ renamed.inode.touchCtime()
+
+ vfs.InotifyRename(ctx, &renamed.inode.watches, &oldParentDir.inode.watches, &newParentDir.inode.watches, oldName, newName, renamed.inode.isDir())
+ return nil
+}
+
+// RmdirAt implements vfs.FilesystemImpl.RmdirAt.
+func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error {
+ fs.mu.Lock()
+ defer fs.mu.Unlock()
+ parentDir, err := walkParentDirLocked(ctx, rp, rp.Start().Impl().(*dentry))
+ if err != nil {
+ return err
+ }
+ if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+ return err
+ }
+ name := rp.Component()
+ if name == "." {
+ return syserror.EINVAL
+ }
+ if name == ".." {
+ return syserror.ENOTEMPTY
+ }
+ child, ok := parentDir.childMap[name]
+ if !ok {
+ return syserror.ENOENT
+ }
+ if err := parentDir.mayDelete(rp.Credentials(), child); err != nil {
+ return err
+ }
+ childDir, ok := child.inode.impl.(*directory)
+ if !ok {
+ return syserror.ENOTDIR
+ }
+ if len(childDir.childMap) != 0 {
+ return syserror.ENOTEMPTY
+ }
+ mnt := rp.Mount()
+ if err := mnt.CheckBeginWrite(); err != nil {
+ return err
+ }
+ defer mnt.EndWrite()
+ vfsObj := rp.VirtualFilesystem()
+ mntns := vfs.MountNamespaceFromContext(ctx)
+ defer mntns.DecRef(ctx)
+ if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil {
+ return err
+ }
+ parentDir.removeChildLocked(child)
+ parentDir.inode.watches.Notify(ctx, name, linux.IN_DELETE|linux.IN_ISDIR, 0, vfs.InodeEvent, true /* unlinked */)
+ // Remove links for child, child/., and child/..
+ child.inode.decLinksLocked(ctx)
+ child.inode.decLinksLocked(ctx)
+ parentDir.inode.decLinksLocked(ctx)
+ vfsObj.CommitDeleteDentry(ctx, &child.vfsd)
+ parentDir.inode.touchCMtime()
+ return nil
+}
+
+// SetStatAt implements vfs.FilesystemImpl.SetStatAt.
+func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error {
+ fs.mu.RLock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ fs.mu.RUnlock()
+ return err
+ }
+ if err := d.inode.setStat(ctx, rp.Credentials(), &opts); err != nil {
+ fs.mu.RUnlock()
+ return err
+ }
+ fs.mu.RUnlock()
+
+ if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 {
+ d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent)
+ }
+ return nil
+}
+
+// StatAt implements vfs.FilesystemImpl.StatAt.
+func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ return linux.Statx{}, err
+ }
+ var stat linux.Statx
+ d.inode.statTo(&stat)
+ return stat, nil
+}
+
+// StatFSAt implements vfs.FilesystemImpl.StatFSAt.
+func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ if _, err := resolveLocked(ctx, rp); err != nil {
+ return linux.Statfs{}, err
+ }
+ statfs := linux.Statfs{
+ Type: linux.TMPFS_MAGIC,
+ BlockSize: usermem.PageSize,
+ FragmentSize: usermem.PageSize,
+ NameLength: linux.NAME_MAX,
+ // TODO(b/29637826): Allow configuring a tmpfs size and enforce it.
+ Blocks: 0,
+ BlocksFree: 0,
+ }
+ return statfs, nil
+}
+
+// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt.
+func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error {
+ return fs.doCreateAt(ctx, rp, false /* dir */, func(parentDir *directory, name string) error {
+ creds := rp.Credentials()
+ child := fs.newDentry(fs.newSymlink(creds.EffectiveKUID, creds.EffectiveKGID, 0777, target))
+ parentDir.insertChildLocked(child, name)
+ return nil
+ })
+}
+
+// UnlinkAt implements vfs.FilesystemImpl.UnlinkAt.
+func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error {
+ fs.mu.Lock()
+ defer fs.mu.Unlock()
+ parentDir, err := walkParentDirLocked(ctx, rp, rp.Start().Impl().(*dentry))
+ if err != nil {
+ return err
+ }
+ if err := parentDir.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil {
+ return err
+ }
+ name := rp.Component()
+ if name == "." || name == ".." {
+ return syserror.EISDIR
+ }
+ child, ok := parentDir.childMap[name]
+ if !ok {
+ return syserror.ENOENT
+ }
+ if err := parentDir.mayDelete(rp.Credentials(), child); err != nil {
+ return err
+ }
+ if child.inode.isDir() {
+ return syserror.EISDIR
+ }
+ if rp.MustBeDir() {
+ return syserror.ENOTDIR
+ }
+ mnt := rp.Mount()
+ if err := mnt.CheckBeginWrite(); err != nil {
+ return err
+ }
+ defer mnt.EndWrite()
+ vfsObj := rp.VirtualFilesystem()
+ mntns := vfs.MountNamespaceFromContext(ctx)
+ defer mntns.DecRef(ctx)
+ if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil {
+ return err
+ }
+
+ // Generate inotify events. Note that this must take place before the link
+ // count of the child is decremented, or else the watches may be dropped
+ // before these events are added.
+ vfs.InotifyRemoveChild(ctx, &child.inode.watches, &parentDir.inode.watches, name)
+
+ parentDir.removeChildLocked(child)
+ child.inode.decLinksLocked(ctx)
+ vfsObj.CommitDeleteDentry(ctx, &child.vfsd)
+ parentDir.inode.touchCMtime()
+ return nil
+}
+
+// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt.
+func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ return nil, err
+ }
+ if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil {
+ return nil, err
+ }
+ switch impl := d.inode.impl.(type) {
+ case *socketFile:
+ return impl.ep, nil
+ default:
+ return nil, syserror.ECONNREFUSED
+ }
+}
+
+// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt.
+func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ return nil, err
+ }
+ return d.inode.listxattr(size)
+}
+
+// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt.
+func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ return "", err
+ }
+ return d.inode.getxattr(rp.Credentials(), &opts)
+}
+
+// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt.
+func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error {
+ fs.mu.RLock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ fs.mu.RUnlock()
+ return err
+ }
+ if err := d.inode.setxattr(rp.Credentials(), &opts); err != nil {
+ fs.mu.RUnlock()
+ return err
+ }
+ fs.mu.RUnlock()
+
+ d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
+ return nil
+}
+
+// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt.
+func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error {
+ fs.mu.RLock()
+ d, err := resolveLocked(ctx, rp)
+ if err != nil {
+ fs.mu.RUnlock()
+ return err
+ }
+ if err := d.inode.removexattr(rp.Credentials(), name); err != nil {
+ fs.mu.RUnlock()
+ return err
+ }
+ fs.mu.RUnlock()
+
+ d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
+ return nil
+}
+
+// PrependPath implements vfs.FilesystemImpl.PrependPath.
+func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDentry, b *fspath.Builder) error {
+ fs.mu.RLock()
+ defer fs.mu.RUnlock()
+ mnt := vd.Mount()
+ d := vd.Dentry().Impl().(*dentry)
+ for {
+ if mnt == vfsroot.Mount() && &d.vfsd == vfsroot.Dentry() {
+ return vfs.PrependPathAtVFSRootError{}
+ }
+ if &d.vfsd == mnt.Root() {
+ return nil
+ }
+ if d.parent == nil {
+ if d.name != "" {
+ // This must be an anonymous memfd file.
+ b.PrependComponent("/" + d.name)
+ return vfs.PrependPathSyntheticError{}
+ }
+ return vfs.PrependPathAtNonMountRootError{}
+ }
+ b.PrependComponent(d.name)
+ d = d.parent
+ }
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/named_pipe.go b/pkg/sentry/fsimpl/tmpfs/named_pipe.go
new file mode 100644
index 000000000..739350cf0
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/named_pipe.go
@@ -0,0 +1,38 @@
+// 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 tmpfs
+
+import (
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/pipe"
+ "gvisor.dev/gvisor/pkg/usermem"
+)
+
+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(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *inode {
+ file := &namedPipe{pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize)}
+ file.inode.init(file, fs, kuid, kgid, linux.S_IFIFO|mode)
+ file.inode.nlink = 1 // Only the parent has a link.
+ return &file.inode
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/pipe_test.go b/pkg/sentry/fsimpl/tmpfs/pipe_test.go
new file mode 100644
index 000000000..ec2701d8b
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/pipe_test.go
@@ -0,0 +1,238 @@
+// 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 tmpfs
+
+import (
+ "bytes"
+ "testing"
+
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
+ "gvisor.dev/gvisor/pkg/fspath"
+ "gvisor.dev/gvisor/pkg/sentry/contexttest"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/syserror"
+ "gvisor.dev/gvisor/pkg/usermem"
+)
+
+const fileName = "mypipe"
+
+func TestSeparateFDs(t *testing.T) {
+ ctx, creds, vfsObj, root := setup(t)
+ defer root.DecRef(ctx)
+
+ // 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,
+ Path: fspath.Parse(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(ctx)
+
+ rfd, ok := <-rfdchan
+ if !ok {
+ t.Fatalf("failed to open pipe for reading %q", fileName)
+ }
+ defer rfd.DecRef(ctx)
+
+ 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(ctx)
+
+ // Open the read side as nonblocking.
+ pop := vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(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(ctx)
+
+ // 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(ctx)
+
+ 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(ctx)
+
+ // Open the write side as nonblocking, which should return ENXIO.
+ pop := vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(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(ctx)
+
+ // Open the pipe as readable and writable.
+ pop := vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(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(ctx)
+
+ 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.VirtualFilesystem{}
+ if err := vfsObj.Init(ctx); err != nil {
+ t.Fatalf("VFS init: %v", err)
+ }
+ vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserMount: true,
+ })
+ mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{})
+ 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,
+ Path: fspath.Parse(fileName),
+ }
+ 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,
+ Path: fspath.Parse(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.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.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.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/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go
new file mode 100644
index 000000000..0710b65db
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go
@@ -0,0 +1,637 @@
+// 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 tmpfs
+
+import (
+ "fmt"
+ "io"
+ "math"
+ "sync/atomic"
+
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
+ "gvisor.dev/gvisor/pkg/safemem"
+ "gvisor.dev/gvisor/pkg/sentry/fs"
+ "gvisor.dev/gvisor/pkg/sentry/fs/fsutil"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/memmap"
+ "gvisor.dev/gvisor/pkg/sentry/pgalloc"
+ "gvisor.dev/gvisor/pkg/sentry/usage"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/sync"
+ "gvisor.dev/gvisor/pkg/syserror"
+ "gvisor.dev/gvisor/pkg/usermem"
+)
+
+// regularFile is a regular (=S_IFREG) tmpfs file.
+type regularFile struct {
+ inode inode
+
+ // memFile is a platform.File used to allocate pages to this regularFile.
+ memFile *pgalloc.MemoryFile
+
+ // mapsMu protects mappings.
+ mapsMu sync.Mutex `state:"nosave"`
+
+ // mappings tracks mappings of the file into memmap.MappingSpaces.
+ //
+ // Protected by mapsMu.
+ mappings memmap.MappingSet
+
+ // writableMappingPages tracks how many pages of virtual memory are mapped
+ // as potentially writable from this file. If a page has multiple mappings,
+ // each mapping is counted separately.
+ //
+ // This counter is susceptible to overflow as we can potentially count
+ // mappings from many VMAs. We count pages rather than bytes to slightly
+ // mitigate this.
+ //
+ // Protected by mapsMu.
+ writableMappingPages uint64
+
+ // dataMu protects the fields below.
+ dataMu sync.RWMutex
+
+ // data maps offsets into the file to offsets into memFile that store
+ // the file's data.
+ //
+ // Protected by dataMu.
+ data fsutil.FileRangeSet
+
+ // seals represents file seals on this inode.
+ //
+ // Protected by dataMu.
+ seals uint32
+
+ // size is the size of data.
+ //
+ // Protected by both dataMu and inode.mu; reading it requires holding
+ // either mutex, while writing requires holding both AND using atomics.
+ // Readers that do not require consistency (like Stat) may read the
+ // value atomically without holding either lock.
+ size uint64
+}
+
+func (fs *filesystem) newRegularFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *inode {
+ file := &regularFile{
+ memFile: fs.memFile,
+ seals: linux.F_SEAL_SEAL,
+ }
+ file.inode.init(file, fs, kuid, kgid, linux.S_IFREG|mode)
+ file.inode.nlink = 1 // from parent directory
+ return &file.inode
+}
+
+// truncate grows or shrinks the file to the given size. It returns true if the
+// file size was updated.
+func (rf *regularFile) truncate(newSize uint64) (bool, error) {
+ rf.inode.mu.Lock()
+ defer rf.inode.mu.Unlock()
+ return rf.truncateLocked(newSize)
+}
+
+// Preconditions: rf.inode.mu must be held.
+func (rf *regularFile) truncateLocked(newSize uint64) (bool, error) {
+ oldSize := rf.size
+ if newSize == oldSize {
+ // Nothing to do.
+ return false, nil
+ }
+
+ // Need to hold inode.mu and dataMu while modifying size.
+ rf.dataMu.Lock()
+ if newSize > oldSize {
+ // Can we grow the file?
+ if rf.seals&linux.F_SEAL_GROW != 0 {
+ rf.dataMu.Unlock()
+ return false, syserror.EPERM
+ }
+ // We only need to update the file size.
+ atomic.StoreUint64(&rf.size, newSize)
+ rf.dataMu.Unlock()
+ return true, nil
+ }
+
+ // We are shrinking the file. First check if this is allowed.
+ if rf.seals&linux.F_SEAL_SHRINK != 0 {
+ rf.dataMu.Unlock()
+ return false, syserror.EPERM
+ }
+
+ // Update the file size.
+ atomic.StoreUint64(&rf.size, newSize)
+ rf.dataMu.Unlock()
+
+ // Invalidate past translations of truncated pages.
+ oldpgend := fs.OffsetPageEnd(int64(oldSize))
+ newpgend := fs.OffsetPageEnd(int64(newSize))
+ if newpgend < oldpgend {
+ rf.mapsMu.Lock()
+ rf.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{
+ // Compare Linux's mm/shmem.c:shmem_setattr() =>
+ // mm/memory.c:unmap_mapping_range(evencows=1).
+ InvalidatePrivate: true,
+ })
+ rf.mapsMu.Unlock()
+ }
+
+ // We are now guaranteed that there are no translations of truncated pages,
+ // and can remove them.
+ rf.dataMu.Lock()
+ rf.data.Truncate(newSize, rf.memFile)
+ rf.dataMu.Unlock()
+ return true, nil
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (rf *regularFile) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error {
+ rf.mapsMu.Lock()
+ defer rf.mapsMu.Unlock()
+ rf.dataMu.RLock()
+ defer rf.dataMu.RUnlock()
+
+ // Reject writable mapping if F_SEAL_WRITE is set.
+ if rf.seals&linux.F_SEAL_WRITE != 0 && writable {
+ return syserror.EPERM
+ }
+
+ rf.mappings.AddMapping(ms, ar, offset, writable)
+ if writable {
+ pagesBefore := rf.writableMappingPages
+
+ // ar is guaranteed to be page aligned per memmap.Mappable.
+ rf.writableMappingPages += uint64(ar.Length() / usermem.PageSize)
+
+ if rf.writableMappingPages < pagesBefore {
+ panic(fmt.Sprintf("Overflow while mapping potentially writable pages pointing to a tmpfs file. Before %v, after %v", pagesBefore, rf.writableMappingPages))
+ }
+ }
+
+ return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (rf *regularFile) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) {
+ rf.mapsMu.Lock()
+ defer rf.mapsMu.Unlock()
+
+ rf.mappings.RemoveMapping(ms, ar, offset, writable)
+
+ if writable {
+ pagesBefore := rf.writableMappingPages
+
+ // ar is guaranteed to be page aligned per memmap.Mappable.
+ rf.writableMappingPages -= uint64(ar.Length() / usermem.PageSize)
+
+ if rf.writableMappingPages > pagesBefore {
+ panic(fmt.Sprintf("Underflow while unmapping potentially writable pages pointing to a tmpfs file. Before %v, after %v", pagesBefore, rf.writableMappingPages))
+ }
+ }
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (rf *regularFile) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error {
+ return rf.AddMapping(ctx, ms, dstAR, offset, writable)
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (rf *regularFile) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+ rf.dataMu.Lock()
+ defer rf.dataMu.Unlock()
+
+ // Constrain translations to f.attr.Size (rounded up) to prevent
+ // translation to pages that may be concurrently truncated.
+ pgend := fs.OffsetPageEnd(int64(rf.size))
+ var beyondEOF bool
+ if required.End > pgend {
+ if required.Start >= pgend {
+ return nil, &memmap.BusError{io.EOF}
+ }
+ beyondEOF = true
+ required.End = pgend
+ }
+ if optional.End > pgend {
+ optional.End = pgend
+ }
+
+ cerr := rf.data.Fill(ctx, required, optional, rf.memFile, usage.Tmpfs, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) {
+ // Newly-allocated pages are zeroed, so we don't need to do anything.
+ return dsts.NumBytes(), nil
+ })
+
+ var ts []memmap.Translation
+ var translatedEnd uint64
+ for seg := rf.data.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() {
+ segMR := seg.Range().Intersect(optional)
+ ts = append(ts, memmap.Translation{
+ Source: segMR,
+ File: rf.memFile,
+ Offset: seg.FileRangeOf(segMR).Start,
+ Perms: usermem.AnyAccess,
+ })
+ translatedEnd = segMR.End
+ }
+
+ // Don't return the error returned by f.data.Fill if it occurred outside of
+ // required.
+ if translatedEnd < required.End && cerr != nil {
+ return ts, &memmap.BusError{cerr}
+ }
+ if beyondEOF {
+ return ts, &memmap.BusError{io.EOF}
+ }
+ return ts, nil
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (*regularFile) InvalidateUnsavable(context.Context) error {
+ return nil
+}
+
+type regularFileFD struct {
+ fileDescription
+
+ // off is the file offset. off is accessed using atomic memory operations.
+ // offMu serializes operations that may mutate off.
+ off int64
+ offMu sync.Mutex
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (fd *regularFileFD) Release(context.Context) {
+ // noop
+}
+
+// Allocate implements vfs.FileDescriptionImpl.Allocate.
+func (fd *regularFileFD) Allocate(ctx context.Context, mode, offset, length uint64) error {
+ f := fd.inode().impl.(*regularFile)
+
+ f.inode.mu.Lock()
+ defer f.inode.mu.Unlock()
+ oldSize := f.size
+ size := offset + length
+ if oldSize >= size {
+ return nil
+ }
+ _, err := f.truncateLocked(size)
+ return err
+}
+
+// PRead implements vfs.FileDescriptionImpl.PRead.
+func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+ if offset < 0 {
+ return 0, syserror.EINVAL
+ }
+
+ // Check that flags are supported. RWF_DSYNC/RWF_SYNC can be ignored since
+ // all state is in-memory.
+ //
+ // TODO(gvisor.dev/issue/2601): Support select preadv2 flags.
+ if opts.Flags&^(linux.RWF_HIPRI|linux.RWF_DSYNC|linux.RWF_SYNC) != 0 {
+ return 0, syserror.EOPNOTSUPP
+ }
+
+ if dst.NumBytes() == 0 {
+ return 0, nil
+ }
+ f := fd.inode().impl.(*regularFile)
+ rw := getRegularFileReadWriter(f, offset)
+ n, err := dst.CopyOutFrom(ctx, rw)
+ putRegularFileReadWriter(rw)
+ fd.inode().touchAtime(fd.vfsfd.Mount())
+ return n, err
+}
+
+// Read implements vfs.FileDescriptionImpl.Read.
+func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
+ fd.offMu.Lock()
+ n, err := fd.PRead(ctx, dst, fd.off, opts)
+ fd.off += n
+ fd.offMu.Unlock()
+ return n, err
+}
+
+// PWrite implements vfs.FileDescriptionImpl.PWrite.
+func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) {
+ n, _, err := fd.pwrite(ctx, src, offset, opts)
+ return n, err
+}
+
+// pwrite returns the number of bytes written, final offset and error. The
+// final offset should be ignored by PWrite.
+func (fd *regularFileFD) pwrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (written, finalOff int64, err error) {
+ if offset < 0 {
+ return 0, offset, syserror.EINVAL
+ }
+
+ // Check that flags are supported. RWF_DSYNC/RWF_SYNC can be ignored since
+ // all state is in-memory.
+ //
+ // TODO(gvisor.dev/issue/2601): Support select preadv2 flags.
+ if opts.Flags&^(linux.RWF_HIPRI|linux.RWF_DSYNC|linux.RWF_SYNC) != 0 {
+ return 0, offset, syserror.EOPNOTSUPP
+ }
+
+ srclen := src.NumBytes()
+ if srclen == 0 {
+ return 0, offset, nil
+ }
+ f := fd.inode().impl.(*regularFile)
+ f.inode.mu.Lock()
+ defer f.inode.mu.Unlock()
+ // If the file is opened with O_APPEND, update offset to file size.
+ if fd.vfsfd.StatusFlags()&linux.O_APPEND != 0 {
+ // Locking f.inode.mu is sufficient for reading f.size.
+ offset = int64(f.size)
+ }
+ if end := offset + srclen; end < offset {
+ // Overflow.
+ return 0, offset, syserror.EINVAL
+ }
+
+ srclen, err = vfs.CheckLimit(ctx, offset, srclen)
+ if err != nil {
+ return 0, offset, err
+ }
+ src = src.TakeFirst64(srclen)
+
+ rw := getRegularFileReadWriter(f, offset)
+ n, err := src.CopyInTo(ctx, rw)
+ f.inode.touchCMtimeLocked()
+ putRegularFileReadWriter(rw)
+ return n, n + offset, err
+}
+
+// Write implements vfs.FileDescriptionImpl.Write.
+func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
+ fd.offMu.Lock()
+ n, off, err := fd.pwrite(ctx, src, fd.off, opts)
+ fd.off = off
+ fd.offMu.Unlock()
+ return n, err
+}
+
+// Seek implements vfs.FileDescriptionImpl.Seek.
+func (fd *regularFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
+ fd.offMu.Lock()
+ defer fd.offMu.Unlock()
+ switch whence {
+ case linux.SEEK_SET:
+ // use offset as specified
+ case linux.SEEK_CUR:
+ offset += fd.off
+ case linux.SEEK_END:
+ offset += int64(atomic.LoadUint64(&fd.inode().impl.(*regularFile).size))
+ default:
+ return 0, syserror.EINVAL
+ }
+ if offset < 0 {
+ return 0, syserror.EINVAL
+ }
+ fd.off = offset
+ return offset, nil
+}
+
+// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap.
+func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error {
+ file := fd.inode().impl.(*regularFile)
+ return vfs.GenericConfigureMMap(&fd.vfsfd, file, opts)
+}
+
+// regularFileReadWriter implements safemem.Reader and Safemem.Writer.
+type regularFileReadWriter struct {
+ file *regularFile
+
+ // Offset into the file to read/write at. Note that this may be
+ // different from the FD offset if PRead/PWrite is used.
+ off uint64
+}
+
+var regularFileReadWriterPool = sync.Pool{
+ New: func() interface{} {
+ return &regularFileReadWriter{}
+ },
+}
+
+func getRegularFileReadWriter(file *regularFile, offset int64) *regularFileReadWriter {
+ rw := regularFileReadWriterPool.Get().(*regularFileReadWriter)
+ rw.file = file
+ rw.off = uint64(offset)
+ return rw
+}
+
+func putRegularFileReadWriter(rw *regularFileReadWriter) {
+ rw.file = nil
+ regularFileReadWriterPool.Put(rw)
+}
+
+// ReadToBlocks implements safemem.Reader.ReadToBlocks.
+func (rw *regularFileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) {
+ rw.file.dataMu.RLock()
+ defer rw.file.dataMu.RUnlock()
+ size := rw.file.size
+
+ // Compute the range to read (limited by file size and overflow-checked).
+ if rw.off >= size {
+ return 0, io.EOF
+ }
+ end := size
+ if rend := rw.off + dsts.NumBytes(); rend > rw.off && rend < end {
+ end = rend
+ }
+
+ var done uint64
+ seg, gap := rw.file.data.Find(uint64(rw.off))
+ for rw.off < end {
+ mr := memmap.MappableRange{uint64(rw.off), uint64(end)}
+ switch {
+ case seg.Ok():
+ // Get internal mappings.
+ ims, err := rw.file.memFile.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read)
+ if err != nil {
+ return done, err
+ }
+
+ // Copy from internal mappings.
+ n, err := safemem.CopySeq(dsts, ims)
+ done += n
+ rw.off += uint64(n)
+ dsts = dsts.DropFirst64(n)
+ if err != nil {
+ return done, err
+ }
+
+ // Continue.
+ seg, gap = seg.NextNonEmpty()
+
+ case gap.Ok():
+ // Tmpfs holes are zero-filled.
+ gapmr := gap.Range().Intersect(mr)
+ dst := dsts.TakeFirst64(gapmr.Length())
+ n, err := safemem.ZeroSeq(dst)
+ done += n
+ rw.off += uint64(n)
+ dsts = dsts.DropFirst64(n)
+ if err != nil {
+ return done, err
+ }
+
+ // Continue.
+ seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{}
+ }
+ }
+ return done, nil
+}
+
+// WriteFromBlocks implements safemem.Writer.WriteFromBlocks.
+//
+// Preconditions: inode.mu must be held.
+func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) {
+ // Hold dataMu so we can modify size.
+ rw.file.dataMu.Lock()
+ defer rw.file.dataMu.Unlock()
+
+ // Compute the range to write (overflow-checked).
+ end := rw.off + srcs.NumBytes()
+ if end <= rw.off {
+ end = math.MaxInt64
+ }
+
+ // Check if seals prevent either file growth or all writes.
+ switch {
+ case rw.file.seals&linux.F_SEAL_WRITE != 0: // Write sealed
+ return 0, syserror.EPERM
+ case end > rw.file.size && rw.file.seals&linux.F_SEAL_GROW != 0: // Grow sealed
+ // When growth is sealed, Linux effectively allows writes which would
+ // normally grow the file to partially succeed up to the current EOF,
+ // rounded down to the page boundary before the EOF.
+ //
+ // This happens because writes (and thus the growth check) for tmpfs
+ // files proceed page-by-page on Linux, and the final write to the page
+ // containing EOF fails, resulting in a partial write up to the start of
+ // that page.
+ //
+ // To emulate this behaviour, artifically truncate the write to the
+ // start of the page containing the current EOF.
+ //
+ // See Linux, mm/filemap.c:generic_perform_write() and
+ // mm/shmem.c:shmem_write_begin().
+ if pgstart := uint64(usermem.Addr(rw.file.size).RoundDown()); end > pgstart {
+ end = pgstart
+ }
+ if end <= rw.off {
+ // Truncation would result in no data being written.
+ return 0, syserror.EPERM
+ }
+ }
+
+ // Page-aligned mr for when we need to allocate memory. RoundUp can't
+ // overflow since end is an int64.
+ pgstartaddr := usermem.Addr(rw.off).RoundDown()
+ pgendaddr, _ := usermem.Addr(end).RoundUp()
+ pgMR := memmap.MappableRange{uint64(pgstartaddr), uint64(pgendaddr)}
+
+ var (
+ done uint64
+ retErr error
+ )
+ seg, gap := rw.file.data.Find(uint64(rw.off))
+ for rw.off < end {
+ mr := memmap.MappableRange{uint64(rw.off), uint64(end)}
+ switch {
+ case seg.Ok():
+ // Get internal mappings.
+ ims, err := rw.file.memFile.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Write)
+ if err != nil {
+ retErr = err
+ goto exitLoop
+ }
+
+ // Copy to internal mappings.
+ n, err := safemem.CopySeq(ims, srcs)
+ done += n
+ rw.off += uint64(n)
+ srcs = srcs.DropFirst64(n)
+ if err != nil {
+ retErr = err
+ goto exitLoop
+ }
+
+ // Continue.
+ seg, gap = seg.NextNonEmpty()
+
+ case gap.Ok():
+ // Allocate memory for the write.
+ gapMR := gap.Range().Intersect(pgMR)
+ fr, err := rw.file.memFile.Allocate(gapMR.Length(), usage.Tmpfs)
+ if err != nil {
+ retErr = err
+ goto exitLoop
+ }
+
+ // Write to that memory as usual.
+ seg, gap = rw.file.data.Insert(gap, gapMR, fr.Start), fsutil.FileRangeGapIterator{}
+ }
+ }
+exitLoop:
+ // If the write ends beyond the file's previous size, it causes the
+ // file to grow.
+ if rw.off > rw.file.size {
+ rw.file.size = rw.off
+ }
+
+ return done, retErr
+}
+
+// GetSeals returns the current set of seals on a memfd inode.
+func GetSeals(fd *vfs.FileDescription) (uint32, error) {
+ f, ok := fd.Impl().(*regularFileFD)
+ if !ok {
+ return 0, syserror.EINVAL
+ }
+ rf := f.inode().impl.(*regularFile)
+ rf.dataMu.RLock()
+ defer rf.dataMu.RUnlock()
+ return rf.seals, nil
+}
+
+// AddSeals adds new file seals to a memfd inode.
+func AddSeals(fd *vfs.FileDescription, val uint32) error {
+ f, ok := fd.Impl().(*regularFileFD)
+ if !ok {
+ return syserror.EINVAL
+ }
+ rf := f.inode().impl.(*regularFile)
+ rf.mapsMu.Lock()
+ defer rf.mapsMu.Unlock()
+ rf.dataMu.RLock()
+ defer rf.dataMu.RUnlock()
+
+ if rf.seals&linux.F_SEAL_SEAL != 0 {
+ // Seal applied which prevents addition of any new seals.
+ return syserror.EPERM
+ }
+
+ // F_SEAL_WRITE can only be added if there are no active writable maps.
+ if rf.seals&linux.F_SEAL_WRITE == 0 && val&linux.F_SEAL_WRITE != 0 {
+ if rf.writableMappingPages > 0 {
+ return syserror.EBUSY
+ }
+ }
+
+ // Seals can only be added, never removed.
+ rf.seals |= val
+ return nil
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file_test.go b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go
new file mode 100644
index 000000000..146c7fdfe
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go
@@ -0,0 +1,349 @@
+// 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 tmpfs
+
+import (
+ "bytes"
+ "fmt"
+ "io"
+ "testing"
+
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/sentry/contexttest"
+ "gvisor.dev/gvisor/pkg/sentry/fs/lock"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/syserror"
+ "gvisor.dev/gvisor/pkg/usermem"
+)
+
+// Test that we can write some data to a file and read it back.`
+func TestSimpleWriteRead(t *testing.T) {
+ ctx := contexttest.Context(t)
+ fd, cleanup, err := newFileFD(ctx, 0644)
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer cleanup()
+
+ // Write.
+ data := []byte("foobarbaz")
+ n, err := fd.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{})
+ if err != nil {
+ t.Fatalf("fd.Write failed: %v", err)
+ }
+ if n != int64(len(data)) {
+ t.Errorf("fd.Write got short write length %d, want %d", n, len(data))
+ }
+ if got, want := fd.Impl().(*regularFileFD).off, int64(len(data)); got != want {
+ t.Errorf("fd.Write left offset at %d, want %d", got, want)
+ }
+
+ // Seek back to beginning.
+ if _, err := fd.Seek(ctx, 0, linux.SEEK_SET); err != nil {
+ t.Fatalf("fd.Seek failed: %v", err)
+ }
+ if got, want := fd.Impl().(*regularFileFD).off, int64(0); got != want {
+ t.Errorf("fd.Seek(0) left offset at %d, want %d", got, want)
+ }
+
+ // Read.
+ buf := make([]byte, len(data))
+ n, err = fd.Read(ctx, usermem.BytesIOSequence(buf), vfs.ReadOptions{})
+ if err != nil && err != io.EOF {
+ t.Fatalf("fd.Read failed: %v", err)
+ }
+ if n != int64(len(data)) {
+ t.Errorf("fd.Read got short read length %d, want %d", n, len(data))
+ }
+ if got, want := string(buf), string(data); got != want {
+ t.Errorf("Read got %q want %s", got, want)
+ }
+ if got, want := fd.Impl().(*regularFileFD).off, int64(len(data)); got != want {
+ t.Errorf("fd.Write left offset at %d, want %d", got, want)
+ }
+}
+
+func TestPWrite(t *testing.T) {
+ ctx := contexttest.Context(t)
+ fd, cleanup, err := newFileFD(ctx, 0644)
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer cleanup()
+
+ // Fill file with 1k 'a's.
+ data := bytes.Repeat([]byte{'a'}, 1000)
+ n, err := fd.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{})
+ if err != nil {
+ t.Fatalf("fd.Write failed: %v", err)
+ }
+ if n != int64(len(data)) {
+ t.Errorf("fd.Write got short write length %d, want %d", n, len(data))
+ }
+
+ // Write "gVisor is awesome" at various offsets.
+ buf := []byte("gVisor is awesome")
+ offsets := []int{0, 1, 2, 10, 20, 50, 100, len(data) - 100, len(data) - 1, len(data), len(data) + 1}
+ for _, offset := range offsets {
+ name := fmt.Sprintf("PWrite offset=%d", offset)
+ t.Run(name, func(t *testing.T) {
+ n, err := fd.PWrite(ctx, usermem.BytesIOSequence(buf), int64(offset), vfs.WriteOptions{})
+ if err != nil {
+ t.Errorf("fd.PWrite got err %v want nil", err)
+ }
+ if n != int64(len(buf)) {
+ t.Errorf("fd.PWrite got %d bytes want %d", n, len(buf))
+ }
+
+ // Update data to reflect expected file contents.
+ if len(data) < offset+len(buf) {
+ data = append(data, make([]byte, (offset+len(buf))-len(data))...)
+ }
+ copy(data[offset:], buf)
+
+ // Read the whole file and compare with data.
+ readBuf := make([]byte, len(data))
+ n, err = fd.PRead(ctx, usermem.BytesIOSequence(readBuf), 0, vfs.ReadOptions{})
+ if err != nil {
+ t.Fatalf("fd.PRead failed: %v", err)
+ }
+ if n != int64(len(data)) {
+ t.Errorf("fd.PRead got short read length %d, want %d", n, len(data))
+ }
+ if got, want := string(readBuf), string(data); got != want {
+ t.Errorf("PRead got %q want %s", got, want)
+ }
+
+ })
+ }
+}
+
+func TestLocks(t *testing.T) {
+ ctx := contexttest.Context(t)
+ fd, cleanup, err := newFileFD(ctx, 0644)
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer cleanup()
+
+ uid1 := 123
+ uid2 := 456
+ if err := fd.Impl().LockBSD(ctx, uid1, lock.ReadLock, nil); err != nil {
+ t.Fatalf("fd.Impl().LockBSD failed: err = %v", err)
+ }
+ if err := fd.Impl().LockBSD(ctx, uid2, lock.ReadLock, nil); err != nil {
+ t.Fatalf("fd.Impl().LockBSD failed: err = %v", err)
+ }
+ if got, want := fd.Impl().LockBSD(ctx, uid2, lock.WriteLock, nil), syserror.ErrWouldBlock; got != want {
+ t.Fatalf("fd.Impl().LockBSD failed: got = %v, want = %v", got, want)
+ }
+ if err := fd.Impl().UnlockBSD(ctx, uid1); err != nil {
+ t.Fatalf("fd.Impl().UnlockBSD failed: err = %v", err)
+ }
+ if err := fd.Impl().LockBSD(ctx, uid2, lock.WriteLock, nil); err != nil {
+ t.Fatalf("fd.Impl().LockBSD failed: err = %v", err)
+ }
+
+ if err := fd.Impl().LockPOSIX(ctx, uid1, lock.ReadLock, 0, 1, linux.SEEK_SET, nil); err != nil {
+ t.Fatalf("fd.Impl().LockPOSIX failed: err = %v", err)
+ }
+ if err := fd.Impl().LockPOSIX(ctx, uid2, lock.ReadLock, 1, 2, linux.SEEK_SET, nil); err != nil {
+ t.Fatalf("fd.Impl().LockPOSIX failed: err = %v", err)
+ }
+ if err := fd.Impl().LockPOSIX(ctx, uid1, lock.WriteLock, 0, 1, linux.SEEK_SET, nil); err != nil {
+ t.Fatalf("fd.Impl().LockPOSIX failed: err = %v", err)
+ }
+ if got, want := fd.Impl().LockPOSIX(ctx, uid2, lock.ReadLock, 0, 1, linux.SEEK_SET, nil), syserror.ErrWouldBlock; got != want {
+ t.Fatalf("fd.Impl().LockPOSIX failed: got = %v, want = %v", got, want)
+ }
+ if err := fd.Impl().UnlockPOSIX(ctx, uid1, 0, 1, linux.SEEK_SET); err != nil {
+ t.Fatalf("fd.Impl().UnlockPOSIX failed: err = %v", err)
+ }
+}
+
+func TestPRead(t *testing.T) {
+ ctx := contexttest.Context(t)
+ fd, cleanup, err := newFileFD(ctx, 0644)
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer cleanup()
+
+ // Write 100 sequences of 'gVisor is awesome'.
+ data := bytes.Repeat([]byte("gVisor is awsome"), 100)
+ n, err := fd.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{})
+ if err != nil {
+ t.Fatalf("fd.Write failed: %v", err)
+ }
+ if n != int64(len(data)) {
+ t.Errorf("fd.Write got short write length %d, want %d", n, len(data))
+ }
+
+ // Read various sizes from various offsets.
+ sizes := []int{0, 1, 2, 10, 20, 50, 100, 1000}
+ offsets := []int{0, 1, 2, 10, 20, 50, 100, 1000, len(data) - 100, len(data) - 1, len(data), len(data) + 1}
+
+ for _, size := range sizes {
+ for _, offset := range offsets {
+ name := fmt.Sprintf("PRead offset=%d size=%d", offset, size)
+ t.Run(name, func(t *testing.T) {
+ var (
+ wantRead []byte
+ wantErr error
+ )
+ if offset < len(data) {
+ wantRead = data[offset:]
+ } else if size > 0 {
+ wantErr = io.EOF
+ }
+ if offset+size < len(data) {
+ wantRead = wantRead[:size]
+ }
+ buf := make([]byte, size)
+ n, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), int64(offset), vfs.ReadOptions{})
+ if err != wantErr {
+ t.Errorf("fd.PRead got err %v want %v", err, wantErr)
+ }
+ if n != int64(len(wantRead)) {
+ t.Errorf("fd.PRead got %d bytes want %d", n, len(wantRead))
+ }
+ if got := string(buf[:n]); got != string(wantRead) {
+ t.Errorf("fd.PRead got %q want %q", got, string(wantRead))
+ }
+ })
+ }
+ }
+}
+
+func TestTruncate(t *testing.T) {
+ ctx := contexttest.Context(t)
+ fd, cleanup, err := newFileFD(ctx, 0644)
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer cleanup()
+
+ // Fill the file with some data.
+ data := bytes.Repeat([]byte("gVisor is awsome"), 100)
+ written, err := fd.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{})
+ if err != nil {
+ t.Fatalf("fd.Write failed: %v", err)
+ }
+
+ // Size should be same as written.
+ sizeStatOpts := vfs.StatOptions{Mask: linux.STATX_SIZE}
+ stat, err := fd.Stat(ctx, sizeStatOpts)
+ if err != nil {
+ t.Fatalf("fd.Stat failed: %v", err)
+ }
+ if got, want := int64(stat.Size), written; got != want {
+ t.Errorf("fd.Stat got size %d, want %d", got, want)
+ }
+
+ // Truncate down.
+ newSize := uint64(10)
+ if err := fd.SetStat(ctx, vfs.SetStatOptions{
+ Stat: linux.Statx{
+ Mask: linux.STATX_SIZE,
+ Size: newSize,
+ },
+ }); err != nil {
+ t.Errorf("fd.Truncate failed: %v", err)
+ }
+ // Size should be updated.
+ statAfterTruncateDown, err := fd.Stat(ctx, sizeStatOpts)
+ if err != nil {
+ t.Fatalf("fd.Stat failed: %v", err)
+ }
+ if got, want := statAfterTruncateDown.Size, newSize; got != want {
+ t.Errorf("fd.Stat got size %d, want %d", got, want)
+ }
+ // We should only read newSize worth of data.
+ buf := make([]byte, 1000)
+ if n, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), 0, vfs.ReadOptions{}); err != nil && err != io.EOF {
+ t.Fatalf("fd.PRead failed: %v", err)
+ } else if uint64(n) != newSize {
+ t.Errorf("fd.PRead got size %d, want %d", n, newSize)
+ }
+ // Mtime and Ctime should be bumped.
+ if got := statAfterTruncateDown.Mtime.ToNsec(); got <= stat.Mtime.ToNsec() {
+ t.Errorf("fd.Stat got Mtime %v, want > %v", got, stat.Mtime)
+ }
+ if got := statAfterTruncateDown.Ctime.ToNsec(); got <= stat.Ctime.ToNsec() {
+ t.Errorf("fd.Stat got Ctime %v, want > %v", got, stat.Ctime)
+ }
+
+ // Truncate up.
+ newSize = 100
+ if err := fd.SetStat(ctx, vfs.SetStatOptions{
+ Stat: linux.Statx{
+ Mask: linux.STATX_SIZE,
+ Size: newSize,
+ },
+ }); err != nil {
+ t.Errorf("fd.Truncate failed: %v", err)
+ }
+ // Size should be updated.
+ statAfterTruncateUp, err := fd.Stat(ctx, sizeStatOpts)
+ if err != nil {
+ t.Fatalf("fd.Stat failed: %v", err)
+ }
+ if got, want := statAfterTruncateUp.Size, newSize; got != want {
+ t.Errorf("fd.Stat got size %d, want %d", got, want)
+ }
+ // We should read newSize worth of data.
+ buf = make([]byte, 1000)
+ if n, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), 0, vfs.ReadOptions{}); err != nil && err != io.EOF {
+ t.Fatalf("fd.PRead failed: %v", err)
+ } else if uint64(n) != newSize {
+ t.Errorf("fd.PRead got size %d, want %d", n, newSize)
+ }
+ // Bytes should be null after 10, since we previously truncated to 10.
+ for i := uint64(10); i < newSize; i++ {
+ if buf[i] != 0 {
+ t.Errorf("fd.PRead got byte %d=%x, want 0", i, buf[i])
+ break
+ }
+ }
+ // Mtime and Ctime should be bumped.
+ if got := statAfterTruncateUp.Mtime.ToNsec(); got <= statAfterTruncateDown.Mtime.ToNsec() {
+ t.Errorf("fd.Stat got Mtime %v, want > %v", got, statAfterTruncateDown.Mtime)
+ }
+ if got := statAfterTruncateUp.Ctime.ToNsec(); got <= statAfterTruncateDown.Ctime.ToNsec() {
+ t.Errorf("fd.Stat got Ctime %v, want > %v", got, stat.Ctime)
+ }
+
+ // Truncate to the current size.
+ newSize = statAfterTruncateUp.Size
+ if err := fd.SetStat(ctx, vfs.SetStatOptions{
+ Stat: linux.Statx{
+ Mask: linux.STATX_SIZE,
+ Size: newSize,
+ },
+ }); err != nil {
+ t.Errorf("fd.Truncate failed: %v", err)
+ }
+ statAfterTruncateNoop, err := fd.Stat(ctx, sizeStatOpts)
+ if err != nil {
+ t.Fatalf("fd.Stat failed: %v", err)
+ }
+ // Mtime and Ctime should not be bumped, since operation is a noop.
+ if got := statAfterTruncateNoop.Mtime.ToNsec(); got != statAfterTruncateUp.Mtime.ToNsec() {
+ t.Errorf("fd.Stat got Mtime %v, want %v", got, statAfterTruncateUp.Mtime)
+ }
+ if got := statAfterTruncateNoop.Ctime.ToNsec(); got != statAfterTruncateUp.Ctime.ToNsec() {
+ t.Errorf("fd.Stat got Ctime %v, want %v", got, statAfterTruncateUp.Ctime)
+ }
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/socket_file.go b/pkg/sentry/fsimpl/tmpfs/socket_file.go
new file mode 100644
index 000000000..3ed650474
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/socket_file.go
@@ -0,0 +1,34 @@
+// 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 tmpfs
+
+import (
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport"
+)
+
+// socketFile is a socket (=S_IFSOCK) tmpfs file.
+type socketFile struct {
+ inode inode
+ ep transport.BoundEndpoint
+}
+
+func (fs *filesystem) newSocketFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode, ep transport.BoundEndpoint) *inode {
+ file := &socketFile{ep: ep}
+ file.inode.init(file, fs, kuid, kgid, mode)
+ file.inode.nlink = 1 // from parent directory
+ return &file.inode
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/stat_test.go b/pkg/sentry/fsimpl/tmpfs/stat_test.go
new file mode 100644
index 000000000..f7ee4aab2
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/stat_test.go
@@ -0,0 +1,236 @@
+// 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 tmpfs
+
+import (
+ "fmt"
+ "testing"
+
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/sentry/contexttest"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+)
+
+func TestStatAfterCreate(t *testing.T) {
+ ctx := contexttest.Context(t)
+ mode := linux.FileMode(0644)
+
+ // Run with different file types.
+ for _, typ := range []string{"file", "dir", "pipe"} {
+ t.Run(fmt.Sprintf("type=%q", typ), func(t *testing.T) {
+ var (
+ fd *vfs.FileDescription
+ cleanup func()
+ err error
+ )
+ switch typ {
+ case "file":
+ fd, cleanup, err = newFileFD(ctx, mode)
+ case "dir":
+ fd, cleanup, err = newDirFD(ctx, mode)
+ case "pipe":
+ fd, cleanup, err = newPipeFD(ctx, mode)
+ default:
+ panic(fmt.Sprintf("unknown typ %q", typ))
+ }
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer cleanup()
+
+ got, err := fd.Stat(ctx, vfs.StatOptions{})
+ if err != nil {
+ t.Fatalf("Stat failed: %v", err)
+ }
+
+ // Atime, Ctime, Mtime should all be current time (non-zero).
+ atime, ctime, mtime := got.Atime.ToNsec(), got.Ctime.ToNsec(), got.Mtime.ToNsec()
+ if atime != ctime || ctime != mtime {
+ t.Errorf("got atime=%d ctime=%d mtime=%d, wanted equal values", atime, ctime, mtime)
+ }
+ if atime == 0 {
+ t.Errorf("got atime=%d, want non-zero", atime)
+ }
+
+ // Btime should be 0, as it is not set by tmpfs.
+ if btime := got.Btime.ToNsec(); btime != 0 {
+ t.Errorf("got btime %d, want 0", got.Btime.ToNsec())
+ }
+
+ // Size should be 0 (except for directories, which make up a size
+ // of 20 per entry, including the "." and ".." entries present in
+ // otherwise-empty directories).
+ wantSize := uint64(0)
+ if typ == "dir" {
+ wantSize = 40
+ }
+ if got.Size != wantSize {
+ t.Errorf("got size %d, want %d", got.Size, wantSize)
+ }
+
+ // Nlink should be 1 for files, 2 for dirs.
+ wantNlink := uint32(1)
+ if typ == "dir" {
+ wantNlink = 2
+ }
+ if got.Nlink != wantNlink {
+ t.Errorf("got nlink %d, want %d", got.Nlink, wantNlink)
+ }
+
+ // UID and GID are set from context creds.
+ creds := auth.CredentialsFromContext(ctx)
+ if got.UID != uint32(creds.EffectiveKUID) {
+ t.Errorf("got uid %d, want %d", got.UID, uint32(creds.EffectiveKUID))
+ }
+ if got.GID != uint32(creds.EffectiveKGID) {
+ t.Errorf("got gid %d, want %d", got.GID, uint32(creds.EffectiveKGID))
+ }
+
+ // Mode.
+ wantMode := uint16(mode)
+ switch typ {
+ case "file":
+ wantMode |= linux.S_IFREG
+ case "dir":
+ wantMode |= linux.S_IFDIR
+ case "pipe":
+ wantMode |= linux.S_IFIFO
+ default:
+ panic(fmt.Sprintf("unknown typ %q", typ))
+ }
+
+ if got.Mode != wantMode {
+ t.Errorf("got mode %x, want %x", got.Mode, wantMode)
+ }
+
+ // Ino.
+ if got.Ino == 0 {
+ t.Errorf("got ino %d, want not 0", got.Ino)
+ }
+ })
+ }
+}
+
+func TestSetStatAtime(t *testing.T) {
+ ctx := contexttest.Context(t)
+ fd, cleanup, err := newFileFD(ctx, 0644)
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer cleanup()
+
+ allStatOptions := vfs.StatOptions{Mask: linux.STATX_ALL}
+
+ // Get initial stat.
+ initialStat, err := fd.Stat(ctx, allStatOptions)
+ if err != nil {
+ t.Fatalf("Stat failed: %v", err)
+ }
+
+ // Set atime, but without the mask.
+ if err := fd.SetStat(ctx, vfs.SetStatOptions{Stat: linux.Statx{
+ Mask: 0,
+ Atime: linux.NsecToStatxTimestamp(100),
+ }}); err != nil {
+ t.Errorf("SetStat atime without mask failed: %v", err)
+ }
+ // Atime should be unchanged.
+ if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil {
+ t.Errorf("Stat got error: %v", err)
+ } else if gotStat.Atime != initialStat.Atime {
+ t.Errorf("Stat got atime %d, want %d", gotStat.Atime, initialStat.Atime)
+ }
+
+ // Set atime, this time included in the mask.
+ setStat := linux.Statx{
+ Mask: linux.STATX_ATIME,
+ Atime: linux.NsecToStatxTimestamp(100),
+ }
+ if err := fd.SetStat(ctx, vfs.SetStatOptions{Stat: setStat}); err != nil {
+ t.Errorf("SetStat atime with mask failed: %v", err)
+ }
+ if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil {
+ t.Errorf("Stat got error: %v", err)
+ } else if gotStat.Atime != setStat.Atime {
+ t.Errorf("Stat got atime %d, want %d", gotStat.Atime, setStat.Atime)
+ }
+}
+
+func TestSetStat(t *testing.T) {
+ ctx := contexttest.Context(t)
+ mode := linux.FileMode(0644)
+
+ // Run with different file types.
+ for _, typ := range []string{"file", "dir", "pipe"} {
+ t.Run(fmt.Sprintf("type=%q", typ), func(t *testing.T) {
+ var (
+ fd *vfs.FileDescription
+ cleanup func()
+ err error
+ )
+ switch typ {
+ case "file":
+ fd, cleanup, err = newFileFD(ctx, mode)
+ case "dir":
+ fd, cleanup, err = newDirFD(ctx, mode)
+ case "pipe":
+ fd, cleanup, err = newPipeFD(ctx, mode)
+ default:
+ panic(fmt.Sprintf("unknown typ %q", typ))
+ }
+ if err != nil {
+ t.Fatal(err)
+ }
+ defer cleanup()
+
+ allStatOptions := vfs.StatOptions{Mask: linux.STATX_ALL}
+
+ // Get initial stat.
+ initialStat, err := fd.Stat(ctx, allStatOptions)
+ if err != nil {
+ t.Fatalf("Stat failed: %v", err)
+ }
+
+ // Set atime, but without the mask.
+ if err := fd.SetStat(ctx, vfs.SetStatOptions{Stat: linux.Statx{
+ Mask: 0,
+ Atime: linux.NsecToStatxTimestamp(100),
+ }}); err != nil {
+ t.Errorf("SetStat atime without mask failed: %v", err)
+ }
+ // Atime should be unchanged.
+ if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil {
+ t.Errorf("Stat got error: %v", err)
+ } else if gotStat.Atime != initialStat.Atime {
+ t.Errorf("Stat got atime %d, want %d", gotStat.Atime, initialStat.Atime)
+ }
+
+ // Set atime, this time included in the mask.
+ setStat := linux.Statx{
+ Mask: linux.STATX_ATIME,
+ Atime: linux.NsecToStatxTimestamp(100),
+ }
+ if err := fd.SetStat(ctx, vfs.SetStatOptions{Stat: setStat}); err != nil {
+ t.Errorf("SetStat atime with mask failed: %v", err)
+ }
+ if gotStat, err := fd.Stat(ctx, allStatOptions); err != nil {
+ t.Errorf("Stat got error: %v", err)
+ } else if gotStat.Atime != setStat.Atime {
+ t.Errorf("Stat got atime %d, want %d", gotStat.Atime, setStat.Atime)
+ }
+ })
+ }
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/symlink.go b/pkg/sentry/fsimpl/tmpfs/symlink.go
new file mode 100644
index 000000000..b0de5fabe
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/symlink.go
@@ -0,0 +1,37 @@
+// 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 tmpfs
+
+import (
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+)
+
+type symlink struct {
+ inode inode
+ target string // immutable
+}
+
+func (fs *filesystem) newSymlink(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode, target string) *inode {
+ link := &symlink{
+ target: target,
+ }
+ link.inode.init(link, fs, kuid, kgid, linux.S_IFLNK|mode)
+ link.inode.nlink = 1 // from parent directory
+ return &link.inode
+}
+
+// O_PATH is unimplemented, so there's no way to get a FileDescription
+// representing a symlink yet.
diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
new file mode 100644
index 000000000..de2af6d01
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go
@@ -0,0 +1,775 @@
+// 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 tmpfs provides an in-memory filesystem whose contents are
+// application-mutable, consistent with Linux's tmpfs.
+//
+// Lock order:
+//
+// filesystem.mu
+// inode.mu
+// regularFileFD.offMu
+// *** "memmap.Mappable locks" below this point
+// regularFile.mapsMu
+// *** "memmap.Mappable locks taken by Translate" below this point
+// regularFile.dataMu
+// directory.iterMu
+package tmpfs
+
+import (
+ "fmt"
+ "math"
+ "strconv"
+ "strings"
+ "sync/atomic"
+
+ "gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
+ fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+ "gvisor.dev/gvisor/pkg/sentry/kernel/time"
+ "gvisor.dev/gvisor/pkg/sentry/pgalloc"
+ "gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/sentry/vfs/memxattr"
+ "gvisor.dev/gvisor/pkg/sync"
+ "gvisor.dev/gvisor/pkg/syserror"
+ "gvisor.dev/gvisor/pkg/usermem"
+)
+
+// Name is the default filesystem name.
+const Name = "tmpfs"
+
+// FilesystemType implements vfs.FilesystemType.
+type FilesystemType struct{}
+
+// filesystem implements vfs.FilesystemImpl.
+type filesystem struct {
+ vfsfs vfs.Filesystem
+
+ // memFile is used to allocate pages to for regular files.
+ memFile *pgalloc.MemoryFile
+
+ // clock is a realtime clock used to set timestamps in file operations.
+ clock time.Clock
+
+ // devMinor is the filesystem's minor device number. devMinor is immutable.
+ devMinor uint32
+
+ // mu serializes changes to the Dentry tree.
+ mu sync.RWMutex
+
+ nextInoMinusOne uint64 // accessed using atomic memory operations
+}
+
+// Name implements vfs.FilesystemType.Name.
+func (FilesystemType) Name() string {
+ return Name
+}
+
+// FilesystemOpts is used to pass configuration data to tmpfs.
+type FilesystemOpts struct {
+ // RootFileType is the FileType of the filesystem root. Valid values
+ // are: S_IFDIR, S_IFREG, and S_IFLNK. Defaults to S_IFDIR.
+ RootFileType uint16
+
+ // RootSymlinkTarget is the target of the root symlink. Only valid if
+ // RootFileType == S_IFLNK.
+ RootSymlinkTarget string
+
+ // FilesystemType allows setting a different FilesystemType for this
+ // tmpfs filesystem. This allows tmpfs to "impersonate" other
+ // filesystems, like ramdiskfs and cgroupfs.
+ FilesystemType vfs.FilesystemType
+}
+
+// GetFilesystem implements vfs.FilesystemType.GetFilesystem.
+func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, _ string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) {
+ memFileProvider := pgalloc.MemoryFileProviderFromContext(ctx)
+ if memFileProvider == nil {
+ panic("MemoryFileProviderFromContext returned nil")
+ }
+
+ rootFileType := uint16(linux.S_IFDIR)
+ newFSType := vfs.FilesystemType(&fstype)
+ tmpfsOpts, ok := opts.InternalData.(FilesystemOpts)
+ if ok {
+ if tmpfsOpts.RootFileType != 0 {
+ rootFileType = tmpfsOpts.RootFileType
+ }
+ if tmpfsOpts.FilesystemType != nil {
+ newFSType = tmpfsOpts.FilesystemType
+ }
+ }
+
+ mopts := vfs.GenericParseMountOptions(opts.Data)
+ rootMode := linux.FileMode(0777)
+ if rootFileType == linux.S_IFDIR {
+ rootMode = 01777
+ }
+ modeStr, ok := mopts["mode"]
+ if ok {
+ delete(mopts, "mode")
+ mode, err := strconv.ParseUint(modeStr, 8, 32)
+ if err != nil {
+ ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: invalid mode: %q", modeStr)
+ return nil, nil, syserror.EINVAL
+ }
+ rootMode = linux.FileMode(mode & 07777)
+ }
+ rootKUID := creds.EffectiveKUID
+ uidStr, ok := mopts["uid"]
+ if ok {
+ delete(mopts, "uid")
+ uid, err := strconv.ParseUint(uidStr, 10, 32)
+ if err != nil {
+ ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: invalid uid: %q", uidStr)
+ return nil, nil, syserror.EINVAL
+ }
+ kuid := creds.UserNamespace.MapToKUID(auth.UID(uid))
+ if !kuid.Ok() {
+ ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: unmapped uid: %d", uid)
+ return nil, nil, syserror.EINVAL
+ }
+ rootKUID = kuid
+ }
+ rootKGID := creds.EffectiveKGID
+ gidStr, ok := mopts["gid"]
+ if ok {
+ delete(mopts, "gid")
+ gid, err := strconv.ParseUint(gidStr, 10, 32)
+ if err != nil {
+ ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: invalid gid: %q", gidStr)
+ return nil, nil, syserror.EINVAL
+ }
+ kgid := creds.UserNamespace.MapToKGID(auth.GID(gid))
+ if !kgid.Ok() {
+ ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: unmapped gid: %d", gid)
+ return nil, nil, syserror.EINVAL
+ }
+ rootKGID = kgid
+ }
+ if len(mopts) != 0 {
+ ctx.Warningf("tmpfs.FilesystemType.GetFilesystem: unknown options: %v", mopts)
+ return nil, nil, syserror.EINVAL
+ }
+
+ devMinor, err := vfsObj.GetAnonBlockDevMinor()
+ if err != nil {
+ return nil, nil, err
+ }
+ clock := time.RealtimeClockFromContext(ctx)
+ fs := filesystem{
+ memFile: memFileProvider.MemoryFile(),
+ clock: clock,
+ devMinor: devMinor,
+ }
+ fs.vfsfs.Init(vfsObj, newFSType, &fs)
+
+ var root *dentry
+ switch rootFileType {
+ case linux.S_IFREG:
+ root = fs.newDentry(fs.newRegularFile(rootKUID, rootKGID, rootMode))
+ case linux.S_IFLNK:
+ root = fs.newDentry(fs.newSymlink(rootKUID, rootKGID, rootMode, tmpfsOpts.RootSymlinkTarget))
+ case linux.S_IFDIR:
+ root = &fs.newDirectory(rootKUID, rootKGID, rootMode).dentry
+ default:
+ fs.vfsfs.DecRef(ctx)
+ return nil, nil, fmt.Errorf("invalid tmpfs root file type: %#o", rootFileType)
+ }
+ return &fs.vfsfs, &root.vfsd, nil
+}
+
+// NewFilesystem returns a new tmpfs filesystem.
+func NewFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) (*vfs.Filesystem, *vfs.Dentry, error) {
+ return FilesystemType{}.GetFilesystem(ctx, vfsObj, creds, "", vfs.GetFilesystemOptions{})
+}
+
+// Release implements vfs.FilesystemImpl.Release.
+func (fs *filesystem) Release(ctx context.Context) {
+ fs.vfsfs.VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor)
+}
+
+// dentry implements vfs.DentryImpl.
+type dentry struct {
+ vfsd vfs.Dentry
+
+ // parent is this dentry's parent directory. Each referenced dentry holds a
+ // reference on parent.dentry. If this dentry is a filesystem root, parent
+ // is nil. parent is protected by filesystem.mu.
+ parent *dentry
+
+ // name is the name of this dentry in its parent. If this dentry is a
+ // filesystem root, name is the empty string. name is protected by
+ // filesystem.mu.
+ name string
+
+ // dentryEntry (ugh) links dentries into their parent directory.childList.
+ dentryEntry
+
+ // inode is the inode represented by this dentry. Multiple Dentries may
+ // share a single non-directory inode (with hard links). inode is
+ // immutable.
+ //
+ // tmpfs doesn't count references on dentries; because the dentry tree is
+ // the sole source of truth, it is by definition always consistent with the
+ // state of the filesystem. However, it does count references on inodes,
+ // because inode resources are released when all references are dropped.
+ // dentry therefore forwards reference counting directly to inode.
+ inode *inode
+}
+
+func (fs *filesystem) newDentry(inode *inode) *dentry {
+ d := &dentry{
+ inode: inode,
+ }
+ d.vfsd.Init(d)
+ return d
+}
+
+// IncRef implements vfs.DentryImpl.IncRef.
+func (d *dentry) IncRef() {
+ d.inode.incRef()
+}
+
+// TryIncRef implements vfs.DentryImpl.TryIncRef.
+func (d *dentry) TryIncRef() bool {
+ return d.inode.tryIncRef()
+}
+
+// DecRef implements vfs.DentryImpl.DecRef.
+func (d *dentry) DecRef(ctx context.Context) {
+ d.inode.decRef(ctx)
+}
+
+// InotifyWithParent implements vfs.DentryImpl.InotifyWithParent.
+func (d *dentry) InotifyWithParent(ctx context.Context, events, cookie uint32, et vfs.EventType) {
+ if d.inode.isDir() {
+ events |= linux.IN_ISDIR
+ }
+
+ // tmpfs never calls VFS.InvalidateDentry(), so d.vfsd.IsDead() indicates
+ // that d was deleted.
+ deleted := d.vfsd.IsDead()
+
+ d.inode.fs.mu.RLock()
+ // The ordering below is important, Linux always notifies the parent first.
+ if d.parent != nil {
+ d.parent.inode.watches.Notify(ctx, d.name, events, cookie, et, deleted)
+ }
+ d.inode.watches.Notify(ctx, "", events, cookie, et, deleted)
+ d.inode.fs.mu.RUnlock()
+}
+
+// Watches implements vfs.DentryImpl.Watches.
+func (d *dentry) Watches() *vfs.Watches {
+ return &d.inode.watches
+}
+
+// OnZeroWatches implements vfs.Dentry.OnZeroWatches.
+func (d *dentry) OnZeroWatches(context.Context) {}
+
+// inode represents a filesystem object.
+type inode struct {
+ // fs is the owning filesystem. fs is immutable.
+ fs *filesystem
+
+ // A reference is held on all inodes as long as they are reachable in the
+ // filesystem tree, i.e. nlink is nonzero. This reference is dropped when
+ // nlink reaches 0.
+ refs inodeRefs
+
+ // xattrs implements extended attributes.
+ //
+ // TODO(b/148380782): Support xattrs other than user.*
+ xattrs memxattr.SimpleExtendedAttributes
+
+ // Inode metadata. Writing multiple fields atomically requires holding
+ // mu, othewise atomic operations can be used.
+ mu sync.Mutex
+ mode uint32 // file type and mode
+ nlink uint32 // protected by filesystem.mu instead of inode.mu
+ uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic
+ gid uint32 // auth.KGID, but ...
+ ino uint64 // immutable
+
+ // Linux's tmpfs has no concept of btime.
+ atime int64 // nanoseconds
+ ctime int64 // nanoseconds
+ mtime int64 // nanoseconds
+
+ locks vfs.FileLocks
+
+ // Inotify watches for this inode.
+ watches vfs.Watches
+
+ impl interface{} // immutable
+}
+
+const maxLinks = math.MaxUint32
+
+func (i *inode) init(impl interface{}, fs *filesystem, kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) {
+ if mode.FileType() == 0 {
+ panic("file type is required in FileMode")
+ }
+ i.fs = fs
+ i.mode = uint32(mode)
+ i.uid = uint32(kuid)
+ i.gid = uint32(kgid)
+ i.ino = atomic.AddUint64(&fs.nextInoMinusOne, 1)
+ // Tmpfs creation sets atime, ctime, and mtime to current time.
+ now := fs.clock.Now().Nanoseconds()
+ i.atime = now
+ i.ctime = now
+ i.mtime = now
+ // i.nlink initialized by caller
+ i.impl = impl
+ i.refs.EnableLeakCheck()
+}
+
+// incLinksLocked increments i's link count.
+//
+// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0.
+// i.nlink < maxLinks.
+func (i *inode) incLinksLocked() {
+ if i.nlink == 0 {
+ panic("tmpfs.inode.incLinksLocked() called with no existing links")
+ }
+ if i.nlink == maxLinks {
+ panic("tmpfs.inode.incLinksLocked() called with maximum link count")
+ }
+ atomic.AddUint32(&i.nlink, 1)
+}
+
+// decLinksLocked decrements i's link count. If the link count reaches 0, we
+// remove a reference on i as well.
+//
+// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0.
+func (i *inode) decLinksLocked(ctx context.Context) {
+ if i.nlink == 0 {
+ panic("tmpfs.inode.decLinksLocked() called with no existing links")
+ }
+ if atomic.AddUint32(&i.nlink, ^uint32(0)) == 0 {
+ i.decRef(ctx)
+ }
+}
+
+func (i *inode) incRef() {
+ i.refs.IncRef()
+}
+
+func (i *inode) tryIncRef() bool {
+ return i.refs.TryIncRef()
+}
+
+func (i *inode) decRef(ctx context.Context) {
+ i.refs.DecRef(func() {
+ i.watches.HandleDeletion(ctx)
+ if regFile, ok := i.impl.(*regularFile); ok {
+ // Release memory used by regFile to store data. Since regFile is
+ // no longer usable, we don't need to grab any locks or update any
+ // metadata.
+ regFile.data.DropAll(regFile.memFile)
+ }
+ })
+}
+
+func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) error {
+ mode := linux.FileMode(atomic.LoadUint32(&i.mode))
+ return vfs.GenericCheckPermissions(creds, ats, mode, auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid)))
+}
+
+// Go won't inline this function, and returning linux.Statx (which is quite
+// big) means spending a lot of time in runtime.duffcopy(), so instead it's an
+// output parameter.
+//
+// Note that Linux does not guarantee to return consistent data (in the case of
+// a concurrent modification), so we do not require holding inode.mu.
+func (i *inode) statTo(stat *linux.Statx) {
+ stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_NLINK |
+ linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_SIZE |
+ linux.STATX_BLOCKS | linux.STATX_ATIME | linux.STATX_CTIME |
+ linux.STATX_MTIME
+ stat.Blksize = usermem.PageSize
+ stat.Nlink = atomic.LoadUint32(&i.nlink)
+ stat.UID = atomic.LoadUint32(&i.uid)
+ stat.GID = atomic.LoadUint32(&i.gid)
+ stat.Mode = uint16(atomic.LoadUint32(&i.mode))
+ stat.Ino = i.ino
+ stat.Atime = linux.NsecToStatxTimestamp(i.atime)
+ stat.Ctime = linux.NsecToStatxTimestamp(i.ctime)
+ stat.Mtime = linux.NsecToStatxTimestamp(i.mtime)
+ stat.DevMajor = linux.UNNAMED_MAJOR
+ stat.DevMinor = i.fs.devMinor
+ switch impl := i.impl.(type) {
+ case *regularFile:
+ stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS
+ stat.Size = uint64(atomic.LoadUint64(&impl.size))
+ // TODO(jamieliu): This should be impl.data.Span() / 512, but this is
+ // too expensive to compute here. Cache it in regularFile.
+ stat.Blocks = allocatedBlocksForSize(stat.Size)
+ case *directory:
+ // "20" is mm/shmem.c:BOGO_DIRENT_SIZE.
+ stat.Size = 20 * (2 + uint64(atomic.LoadInt64(&impl.numChildren)))
+ // stat.Blocks is 0.
+ case *symlink:
+ stat.Size = uint64(len(impl.target))
+ // stat.Blocks is 0.
+ case *namedPipe, *socketFile:
+ // stat.Size and stat.Blocks are 0.
+ case *deviceFile:
+ // stat.Size and stat.Blocks are 0.
+ stat.RdevMajor = impl.major
+ stat.RdevMinor = impl.minor
+ default:
+ panic(fmt.Sprintf("unknown inode type: %T", i.impl))
+ }
+}
+
+func (i *inode) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs.SetStatOptions) error {
+ stat := &opts.Stat
+ if stat.Mask == 0 {
+ return nil
+ }
+ if stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_CTIME|linux.STATX_SIZE) != 0 {
+ return syserror.EPERM
+ }
+ mode := linux.FileMode(atomic.LoadUint32(&i.mode))
+ if err := vfs.CheckSetStat(ctx, creds, opts, mode, auth.KUID(atomic.LoadUint32(&i.uid)), auth.KGID(atomic.LoadUint32(&i.gid))); err != nil {
+ return err
+ }
+ i.mu.Lock()
+ defer i.mu.Unlock()
+ var (
+ needsMtimeBump bool
+ needsCtimeBump bool
+ )
+ mask := stat.Mask
+ if mask&linux.STATX_MODE != 0 {
+ ft := atomic.LoadUint32(&i.mode) & linux.S_IFMT
+ atomic.StoreUint32(&i.mode, ft|uint32(stat.Mode&^linux.S_IFMT))
+ needsCtimeBump = true
+ }
+ if mask&linux.STATX_UID != 0 {
+ atomic.StoreUint32(&i.uid, stat.UID)
+ needsCtimeBump = true
+ }
+ if mask&linux.STATX_GID != 0 {
+ atomic.StoreUint32(&i.gid, stat.GID)
+ needsCtimeBump = true
+ }
+ if mask&linux.STATX_SIZE != 0 {
+ switch impl := i.impl.(type) {
+ case *regularFile:
+ updated, err := impl.truncateLocked(stat.Size)
+ if err != nil {
+ return err
+ }
+ if updated {
+ needsMtimeBump = true
+ needsCtimeBump = true
+ }
+ case *directory:
+ return syserror.EISDIR
+ default:
+ return syserror.EINVAL
+ }
+ }
+ now := i.fs.clock.Now().Nanoseconds()
+ if mask&linux.STATX_ATIME != 0 {
+ if stat.Atime.Nsec == linux.UTIME_NOW {
+ atomic.StoreInt64(&i.atime, now)
+ } else {
+ atomic.StoreInt64(&i.atime, stat.Atime.ToNsecCapped())
+ }
+ needsCtimeBump = true
+ }
+ if mask&linux.STATX_MTIME != 0 {
+ if stat.Mtime.Nsec == linux.UTIME_NOW {
+ atomic.StoreInt64(&i.mtime, now)
+ } else {
+ atomic.StoreInt64(&i.mtime, stat.Mtime.ToNsecCapped())
+ }
+ needsCtimeBump = true
+ // Ignore the mtime bump, since we just set it ourselves.
+ needsMtimeBump = false
+ }
+ if mask&linux.STATX_CTIME != 0 {
+ if stat.Ctime.Nsec == linux.UTIME_NOW {
+ atomic.StoreInt64(&i.ctime, now)
+ } else {
+ atomic.StoreInt64(&i.ctime, stat.Ctime.ToNsecCapped())
+ }
+ // Ignore the ctime bump, since we just set it ourselves.
+ needsCtimeBump = false
+ }
+ if needsMtimeBump {
+ atomic.StoreInt64(&i.mtime, now)
+ }
+ if needsCtimeBump {
+ atomic.StoreInt64(&i.ctime, now)
+ }
+
+ return nil
+}
+
+// allocatedBlocksForSize returns the number of 512B blocks needed to
+// accommodate the given size in bytes, as appropriate for struct
+// stat::st_blocks and struct statx::stx_blocks. (Note that this 512B block
+// size is independent of the "preferred block size for I/O", struct
+// stat::st_blksize and struct statx::stx_blksize.)
+func allocatedBlocksForSize(size uint64) uint64 {
+ return (size + 511) / 512
+}
+
+func (i *inode) direntType() uint8 {
+ switch impl := i.impl.(type) {
+ case *regularFile:
+ return linux.DT_REG
+ case *directory:
+ return linux.DT_DIR
+ case *symlink:
+ return linux.DT_LNK
+ case *socketFile:
+ return linux.DT_SOCK
+ case *namedPipe:
+ return linux.DT_FIFO
+ case *deviceFile:
+ switch impl.kind {
+ case vfs.BlockDevice:
+ return linux.DT_BLK
+ case vfs.CharDevice:
+ return linux.DT_CHR
+ default:
+ panic(fmt.Sprintf("unknown vfs.DeviceKind: %v", impl.kind))
+ }
+ default:
+ panic(fmt.Sprintf("unknown inode type: %T", i.impl))
+ }
+}
+
+func (i *inode) isDir() bool {
+ return linux.FileMode(i.mode).FileType() == linux.S_IFDIR
+}
+
+func (i *inode) touchAtime(mnt *vfs.Mount) {
+ if mnt.Flags.NoATime {
+ return
+ }
+ if err := mnt.CheckBeginWrite(); err != nil {
+ return
+ }
+ now := i.fs.clock.Now().Nanoseconds()
+ i.mu.Lock()
+ atomic.StoreInt64(&i.atime, now)
+ i.mu.Unlock()
+ mnt.EndWrite()
+}
+
+// Preconditions: The caller has called vfs.Mount.CheckBeginWrite().
+func (i *inode) touchCtime() {
+ now := i.fs.clock.Now().Nanoseconds()
+ i.mu.Lock()
+ atomic.StoreInt64(&i.ctime, now)
+ i.mu.Unlock()
+}
+
+// Preconditions: The caller has called vfs.Mount.CheckBeginWrite().
+func (i *inode) touchCMtime() {
+ now := i.fs.clock.Now().Nanoseconds()
+ i.mu.Lock()
+ atomic.StoreInt64(&i.mtime, now)
+ atomic.StoreInt64(&i.ctime, now)
+ i.mu.Unlock()
+}
+
+// Preconditions: The caller has called vfs.Mount.CheckBeginWrite() and holds
+// inode.mu.
+func (i *inode) touchCMtimeLocked() {
+ now := i.fs.clock.Now().Nanoseconds()
+ atomic.StoreInt64(&i.mtime, now)
+ atomic.StoreInt64(&i.ctime, now)
+}
+
+func (i *inode) listxattr(size uint64) ([]string, error) {
+ return i.xattrs.Listxattr(size)
+}
+
+func (i *inode) getxattr(creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) {
+ if err := i.checkPermissions(creds, vfs.MayRead); err != nil {
+ return "", err
+ }
+ if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) {
+ return "", syserror.EOPNOTSUPP
+ }
+ if !i.userXattrSupported() {
+ return "", syserror.ENODATA
+ }
+ return i.xattrs.Getxattr(opts)
+}
+
+func (i *inode) setxattr(creds *auth.Credentials, opts *vfs.SetxattrOptions) error {
+ if err := i.checkPermissions(creds, vfs.MayWrite); err != nil {
+ return err
+ }
+ if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) {
+ return syserror.EOPNOTSUPP
+ }
+ if !i.userXattrSupported() {
+ return syserror.EPERM
+ }
+ return i.xattrs.Setxattr(opts)
+}
+
+func (i *inode) removexattr(creds *auth.Credentials, name string) error {
+ if err := i.checkPermissions(creds, vfs.MayWrite); err != nil {
+ return err
+ }
+ if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) {
+ return syserror.EOPNOTSUPP
+ }
+ if !i.userXattrSupported() {
+ return syserror.EPERM
+ }
+ return i.xattrs.Removexattr(name)
+}
+
+// Extended attributes in the user.* namespace are only supported for regular
+// files and directories.
+func (i *inode) userXattrSupported() bool {
+ filetype := linux.S_IFMT & atomic.LoadUint32(&i.mode)
+ return filetype == linux.S_IFREG || filetype == linux.S_IFDIR
+}
+
+// fileDescription is embedded by tmpfs implementations of
+// vfs.FileDescriptionImpl.
+type fileDescription struct {
+ vfsfd vfs.FileDescription
+ vfs.FileDescriptionDefaultImpl
+ vfs.LockFD
+}
+
+func (fd *fileDescription) filesystem() *filesystem {
+ return fd.vfsfd.Mount().Filesystem().Impl().(*filesystem)
+}
+
+func (fd *fileDescription) dentry() *dentry {
+ return fd.vfsfd.Dentry().Impl().(*dentry)
+}
+
+func (fd *fileDescription) inode() *inode {
+ return fd.dentry().inode
+}
+
+// Stat implements vfs.FileDescriptionImpl.Stat.
+func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) {
+ var stat linux.Statx
+ fd.inode().statTo(&stat)
+ return stat, nil
+}
+
+// SetStat implements vfs.FileDescriptionImpl.SetStat.
+func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error {
+ creds := auth.CredentialsFromContext(ctx)
+ d := fd.dentry()
+ if err := d.inode.setStat(ctx, creds, &opts); err != nil {
+ return err
+ }
+
+ if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 {
+ d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent)
+ }
+ return nil
+}
+
+// Listxattr implements vfs.FileDescriptionImpl.Listxattr.
+func (fd *fileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) {
+ return fd.inode().listxattr(size)
+}
+
+// Getxattr implements vfs.FileDescriptionImpl.Getxattr.
+func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOptions) (string, error) {
+ return fd.inode().getxattr(auth.CredentialsFromContext(ctx), &opts)
+}
+
+// Setxattr implements vfs.FileDescriptionImpl.Setxattr.
+func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error {
+ d := fd.dentry()
+ if err := d.inode.setxattr(auth.CredentialsFromContext(ctx), &opts); err != nil {
+ return err
+ }
+
+ // Generate inotify events.
+ d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
+ return nil
+}
+
+// Removexattr implements vfs.FileDescriptionImpl.Removexattr.
+func (fd *fileDescription) Removexattr(ctx context.Context, name string) error {
+ d := fd.dentry()
+ if err := d.inode.removexattr(auth.CredentialsFromContext(ctx), name); err != nil {
+ return err
+ }
+
+ // Generate inotify events.
+ d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent)
+ return nil
+}
+
+// NewMemfd creates a new tmpfs regular file and file description that can back
+// an anonymous fd created by memfd_create.
+func NewMemfd(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, allowSeals bool, name string) (*vfs.FileDescription, error) {
+ fs, ok := mount.Filesystem().Impl().(*filesystem)
+ if !ok {
+ panic("NewMemfd() called with non-tmpfs mount")
+ }
+
+ // Per Linux, mm/shmem.c:__shmem_file_setup(), memfd inodes are set up with
+ // S_IRWXUGO.
+ inode := fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, 0777)
+ rf := inode.impl.(*regularFile)
+ if allowSeals {
+ rf.seals = 0
+ }
+
+ d := fs.newDentry(inode)
+ defer d.DecRef(ctx)
+ d.name = name
+
+ // Per Linux, mm/shmem.c:__shmem_file_setup(), memfd files are set up with
+ // FMODE_READ | FMODE_WRITE.
+ var fd regularFileFD
+ fd.Init(&inode.locks)
+ flags := uint32(linux.O_RDWR)
+ if err := fd.vfsfd.Init(&fd, flags, mount, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil {
+ return nil, err
+ }
+ return &fd.vfsfd, nil
+}
+
+// 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)
+}
+
+// UnlockPOSIX implements vfs.FileDescriptionImpl.UnlockPOSIX.
+func (fd *fileDescription) UnlockPOSIX(ctx context.Context, uid fslock.UniqueID, start, length uint64, whence int16) error {
+ return fd.Locks().UnlockPOSIX(ctx, &fd.vfsfd, uid, start, length, whence)
+}
+
+// Sync implements vfs.FileDescriptionImpl.Sync. It does nothing because all
+// filesystem state is in-memory.
+func (*fileDescription) Sync(context.Context) error {
+ return nil
+}
diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go b/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go
new file mode 100644
index 000000000..6f3e3ae6f
--- /dev/null
+++ b/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go
@@ -0,0 +1,156 @@
+// 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 tmpfs
+
+import (
+ "fmt"
+ "sync/atomic"
+
+ "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/vfs"
+)
+
+// nextFileID is used to generate unique file names.
+var nextFileID int64
+
+// newTmpfsRoot creates a new tmpfs mount, and returns the root. If the error
+// is not nil, then cleanup should be called when the root is no longer needed.
+func newTmpfsRoot(ctx context.Context) (*vfs.VirtualFilesystem, vfs.VirtualDentry, func(), error) {
+ creds := auth.CredentialsFromContext(ctx)
+
+ vfsObj := &vfs.VirtualFilesystem{}
+ if err := vfsObj.Init(ctx); err != nil {
+ return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("VFS init: %v", err)
+ }
+
+ vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+ AllowUserMount: true,
+ })
+ mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{})
+ if err != nil {
+ return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("failed to create tmpfs root mount: %v", err)
+ }
+ root := mntns.Root()
+ return vfsObj, root, func() {
+ root.DecRef(ctx)
+ mntns.DecRef(ctx)
+ }, nil
+}
+
+// newFileFD creates a new file in a new tmpfs mount, and returns the FD. If
+// the returned err is not nil, then cleanup should be called when the FD is no
+// longer needed.
+func newFileFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) {
+ creds := auth.CredentialsFromContext(ctx)
+ vfsObj, root, cleanup, err := newTmpfsRoot(ctx)
+ if err != nil {
+ return nil, nil, err
+ }
+
+ filename := fmt.Sprintf("tmpfs-test-file-%d", atomic.AddInt64(&nextFileID, 1))
+
+ // Create the file that will be write/read.
+ fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(filename),
+ }, &vfs.OpenOptions{
+ Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL,
+ Mode: linux.ModeRegular | mode,
+ })
+ if err != nil {
+ cleanup()
+ return nil, nil, fmt.Errorf("failed to create file %q: %v", filename, err)
+ }
+
+ return fd, cleanup, nil
+}
+
+// newDirFD is like newFileFD, but for directories.
+func newDirFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) {
+ creds := auth.CredentialsFromContext(ctx)
+ vfsObj, root, cleanup, err := newTmpfsRoot(ctx)
+ if err != nil {
+ return nil, nil, err
+ }
+
+ dirname := fmt.Sprintf("tmpfs-test-dir-%d", atomic.AddInt64(&nextFileID, 1))
+
+ // Create the dir.
+ if err := vfsObj.MkdirAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(dirname),
+ }, &vfs.MkdirOptions{
+ Mode: linux.ModeDirectory | mode,
+ }); err != nil {
+ cleanup()
+ return nil, nil, fmt.Errorf("failed to create directory %q: %v", dirname, err)
+ }
+
+ // Open the dir and return it.
+ fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(dirname),
+ }, &vfs.OpenOptions{
+ Flags: linux.O_RDONLY | linux.O_DIRECTORY,
+ })
+ if err != nil {
+ cleanup()
+ return nil, nil, fmt.Errorf("failed to open directory %q: %v", dirname, err)
+ }
+
+ return fd, cleanup, nil
+}
+
+// newPipeFD is like newFileFD, but for pipes.
+func newPipeFD(ctx context.Context, mode linux.FileMode) (*vfs.FileDescription, func(), error) {
+ creds := auth.CredentialsFromContext(ctx)
+ vfsObj, root, cleanup, err := newTmpfsRoot(ctx)
+ if err != nil {
+ return nil, nil, err
+ }
+
+ name := fmt.Sprintf("tmpfs-test-%d", atomic.AddInt64(&nextFileID, 1))
+
+ if err := vfsObj.MknodAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(name),
+ }, &vfs.MknodOptions{
+ Mode: linux.ModeNamedPipe | mode,
+ }); err != nil {
+ cleanup()
+ return nil, nil, fmt.Errorf("failed to create pipe %q: %v", name, err)
+ }
+
+ fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse(name),
+ }, &vfs.OpenOptions{
+ Flags: linux.O_RDWR,
+ })
+ if err != nil {
+ cleanup()
+ return nil, nil, fmt.Errorf("failed to open pipe %q: %v", name, err)
+ }
+
+ return fd, cleanup, nil
+}
generated by cgit v1.2.3 (git 2.39.1) at 2024-12-28 15:01:57 +0000