From 51f3ab85e024fcd74c49d273ce5202a207577d31 Mon Sep 17 00:00:00 2001 From: Nicolas Lacasse Date: Mon, 6 Jan 2020 12:51:35 -0800 Subject: Convert memfs into proto-tmpfs. - Renamed memfs to tmpfs. - Copied fileRangeSet bits from fs/fsutil/ to fsimpl/tmpfs/ - Changed tmpfs to be backed by filemem instead of byte slice. - regularFileReadWriter uses a sync.Pool, similar to gofer client. PiperOrigin-RevId: 288356380 --- pkg/sentry/fs/fsutil/BUILD | 2 +- pkg/sentry/fs/fsutil/file_range_set.go | 14 +- pkg/sentry/fsimpl/memfs/BUILD | 80 --- pkg/sentry/fsimpl/memfs/benchmark_test.go | 487 ------------------- pkg/sentry/fsimpl/memfs/directory.go | 187 ------- pkg/sentry/fsimpl/memfs/filesystem.go | 698 --------------------------- pkg/sentry/fsimpl/memfs/memfs.go | 293 ----------- pkg/sentry/fsimpl/memfs/named_pipe.go | 60 --- pkg/sentry/fsimpl/memfs/pipe_test.go | 235 --------- pkg/sentry/fsimpl/memfs/regular_file.go | 154 ------ pkg/sentry/fsimpl/memfs/symlink.go | 36 -- pkg/sentry/fsimpl/tmpfs/BUILD | 92 ++++ pkg/sentry/fsimpl/tmpfs/benchmark_test.go | 487 +++++++++++++++++++ pkg/sentry/fsimpl/tmpfs/directory.go | 187 +++++++ pkg/sentry/fsimpl/tmpfs/filesystem.go | 698 +++++++++++++++++++++++++++ pkg/sentry/fsimpl/tmpfs/named_pipe.go | 60 +++ pkg/sentry/fsimpl/tmpfs/pipe_test.go | 235 +++++++++ pkg/sentry/fsimpl/tmpfs/regular_file.go | 357 ++++++++++++++ pkg/sentry/fsimpl/tmpfs/regular_file_test.go | 224 +++++++++ pkg/sentry/fsimpl/tmpfs/symlink.go | 36 ++ pkg/sentry/fsimpl/tmpfs/tmpfs.go | 299 ++++++++++++ 21 files changed, 2683 insertions(+), 2238 deletions(-) delete mode 100644 pkg/sentry/fsimpl/memfs/BUILD delete mode 100644 pkg/sentry/fsimpl/memfs/benchmark_test.go delete mode 100644 pkg/sentry/fsimpl/memfs/directory.go delete mode 100644 pkg/sentry/fsimpl/memfs/filesystem.go delete mode 100644 pkg/sentry/fsimpl/memfs/memfs.go delete mode 100644 pkg/sentry/fsimpl/memfs/named_pipe.go delete mode 100644 pkg/sentry/fsimpl/memfs/pipe_test.go delete mode 100644 pkg/sentry/fsimpl/memfs/regular_file.go delete mode 100644 pkg/sentry/fsimpl/memfs/symlink.go create mode 100644 pkg/sentry/fsimpl/tmpfs/BUILD create mode 100644 pkg/sentry/fsimpl/tmpfs/benchmark_test.go create mode 100644 pkg/sentry/fsimpl/tmpfs/directory.go create mode 100644 pkg/sentry/fsimpl/tmpfs/filesystem.go create mode 100644 pkg/sentry/fsimpl/tmpfs/named_pipe.go create mode 100644 pkg/sentry/fsimpl/tmpfs/pipe_test.go create mode 100644 pkg/sentry/fsimpl/tmpfs/regular_file.go create mode 100644 pkg/sentry/fsimpl/tmpfs/regular_file_test.go create mode 100644 pkg/sentry/fsimpl/tmpfs/symlink.go create mode 100644 pkg/sentry/fsimpl/tmpfs/tmpfs.go (limited to 'pkg/sentry') diff --git a/pkg/sentry/fs/fsutil/BUILD b/pkg/sentry/fs/fsutil/BUILD index b2e8d9c77..9ca695a95 100644 --- a/pkg/sentry/fs/fsutil/BUILD +++ b/pkg/sentry/fs/fsutil/BUILD @@ -53,7 +53,7 @@ go_template_instance( "Key": "uint64", "Range": "memmap.MappableRange", "Value": "uint64", - "Functions": "fileRangeSetFunctions", + "Functions": "FileRangeSetFunctions", }, ) diff --git a/pkg/sentry/fs/fsutil/file_range_set.go b/pkg/sentry/fs/fsutil/file_range_set.go index 0a5466b0a..f52d712e3 100644 --- a/pkg/sentry/fs/fsutil/file_range_set.go +++ b/pkg/sentry/fs/fsutil/file_range_set.go @@ -34,25 +34,25 @@ import ( // // type FileRangeSet -// fileRangeSetFunctions implements segment.Functions for FileRangeSet. -type fileRangeSetFunctions struct{} +// FileRangeSetFunctions implements segment.Functions for FileRangeSet. +type FileRangeSetFunctions struct{} // MinKey implements segment.Functions.MinKey. -func (fileRangeSetFunctions) MinKey() uint64 { +func (FileRangeSetFunctions) MinKey() uint64 { return 0 } // MaxKey implements segment.Functions.MaxKey. -func (fileRangeSetFunctions) MaxKey() uint64 { +func (FileRangeSetFunctions) MaxKey() uint64 { return math.MaxUint64 } // ClearValue implements segment.Functions.ClearValue. -func (fileRangeSetFunctions) ClearValue(_ *uint64) { +func (FileRangeSetFunctions) ClearValue(_ *uint64) { } // Merge implements segment.Functions.Merge. -func (fileRangeSetFunctions) Merge(mr1 memmap.MappableRange, frstart1 uint64, _ memmap.MappableRange, frstart2 uint64) (uint64, bool) { +func (FileRangeSetFunctions) Merge(mr1 memmap.MappableRange, frstart1 uint64, _ memmap.MappableRange, frstart2 uint64) (uint64, bool) { if frstart1+mr1.Length() != frstart2 { return 0, false } @@ -60,7 +60,7 @@ func (fileRangeSetFunctions) Merge(mr1 memmap.MappableRange, frstart1 uint64, _ } // Split implements segment.Functions.Split. -func (fileRangeSetFunctions) Split(mr memmap.MappableRange, frstart uint64, split uint64) (uint64, uint64) { +func (FileRangeSetFunctions) Split(mr memmap.MappableRange, frstart uint64, split uint64) (uint64, uint64) { return frstart, frstart + (split - mr.Start) } diff --git a/pkg/sentry/fsimpl/memfs/BUILD b/pkg/sentry/fsimpl/memfs/BUILD deleted file mode 100644 index 5689bed3b..000000000 --- a/pkg/sentry/fsimpl/memfs/BUILD +++ /dev/null @@ -1,80 +0,0 @@ -load("//tools/go_stateify:defs.bzl", "go_library") -load("@io_bazel_rules_go//go:def.bzl", "go_test") - -package(licenses = ["notice"]) - -load("//tools/go_generics:defs.bzl", "go_template_instance") - -go_template_instance( - name = "dentry_list", - out = "dentry_list.go", - package = "memfs", - prefix = "dentry", - template = "//pkg/ilist:generic_list", - types = { - "Element": "*dentry", - "Linker": "*dentry", - }, -) - -go_library( - name = "memfs", - srcs = [ - "dentry_list.go", - "directory.go", - "filesystem.go", - "memfs.go", - "named_pipe.go", - "regular_file.go", - "symlink.go", - ], - importpath = "gvisor.dev/gvisor/pkg/sentry/fsimpl/memfs", - deps = [ - "//pkg/abi/linux", - "//pkg/amutex", - "//pkg/fspath", - "//pkg/sentry/arch", - "//pkg/sentry/context", - "//pkg/sentry/kernel/auth", - "//pkg/sentry/kernel/pipe", - "//pkg/sentry/usermem", - "//pkg/sentry/vfs", - "//pkg/syserror", - ], -) - -go_test( - name = "benchmark_test", - size = "small", - srcs = ["benchmark_test.go"], - deps = [ - ":memfs", - "//pkg/abi/linux", - "//pkg/fspath", - "//pkg/refs", - "//pkg/sentry/context", - "//pkg/sentry/context/contexttest", - "//pkg/sentry/fs", - "//pkg/sentry/fs/tmpfs", - "//pkg/sentry/kernel/auth", - "//pkg/sentry/vfs", - "//pkg/syserror", - ], -) - -go_test( - name = "memfs_test", - size = "small", - srcs = ["pipe_test.go"], - embed = [":memfs"], - deps = [ - "//pkg/abi/linux", - "//pkg/fspath", - "//pkg/sentry/context", - "//pkg/sentry/context/contexttest", - "//pkg/sentry/kernel/auth", - "//pkg/sentry/usermem", - "//pkg/sentry/vfs", - "//pkg/syserror", - ], -) diff --git a/pkg/sentry/fsimpl/memfs/benchmark_test.go b/pkg/sentry/fsimpl/memfs/benchmark_test.go deleted file mode 100644 index a27876a4e..000000000 --- a/pkg/sentry/fsimpl/memfs/benchmark_test.go +++ /dev/null @@ -1,487 +0,0 @@ -// 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/fspath" - "gvisor.dev/gvisor/pkg/refs" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" - "gvisor.dev/gvisor/pkg/sentry/fs" - _ "gvisor.dev/gvisor/pkg/sentry/fs/tmpfs" - "gvisor.dev/gvisor/pkg/sentry/fsimpl/memfs" - "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/ (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() - 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() - - var filePathBuilder strings.Builder - filePathBuilder.WriteByte('/') - - // Create nested directories with given depth. - root := mntns.Root() - defer root.DecRef() - d := root - d.IncRef() - defer d.DecRef() - 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() - 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() - 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 BenchmarkVFS2MemfsStat(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.New() - vfsObj.MustRegisterFilesystemType("memfs", memfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ - AllowUserMount: true, - }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "memfs", &vfs.GetFilesystemOptions{}) - if err != nil { - b.Fatalf("failed to create tmpfs root mount: %v", err) - } - defer mntns.DecRef(vfsObj) - - var filePathBuilder strings.Builder - filePathBuilder.WriteByte('/') - - // Create nested directories with given depth. - root := mntns.Root() - defer root.DecRef() - 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() - 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() - vd = vfs.VirtualDentry{} - if err != nil { - b.Fatalf("failed to create file %q: %v", filename, err) - } - defer fd.DecRef() - 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() - - var filePathBuilder strings.Builder - filePathBuilder.WriteByte('/') - - // Create and mount the submount. - root := mntns.Root() - defer root.DecRef() - 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() - 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() - 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() - 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() - 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 BenchmarkVFS2MemfsMountStat(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.New() - vfsObj.MustRegisterFilesystemType("memfs", memfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ - AllowUserMount: true, - }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "memfs", &vfs.GetFilesystemOptions{}) - if err != nil { - b.Fatalf("failed to create tmpfs root mount: %v", err) - } - defer mntns.DecRef(vfsObj) - - var filePathBuilder strings.Builder - filePathBuilder.WriteByte('/') - - // Create the mount point. - root := mntns.Root() - defer root.DecRef() - 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() - // Create and mount the submount. - if err := vfsObj.MountAt(ctx, creds, "", &pop, "memfs", &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() - vd = nextVD - filePathBuilder.WriteString(name) - filePathBuilder.WriteByte('/') - } - - // Verify that we didn't create any directories under the mount - // point (i.e. they were all created on the submount). - firstDirName := fmt.Sprintf("%d", depth) - if child := mountPoint.Dentry().Child(firstDirName); child != nil { - b.Fatalf("created directory %q under root mount, not submount", firstDirName) - } - - // 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() - if err != nil { - b.Fatalf("failed to create file %q: %v", filename, err) - } - fd.DecRef() - 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/memfs/directory.go b/pkg/sentry/fsimpl/memfs/directory.go deleted file mode 100644 index 0bd82e480..000000000 --- a/pkg/sentry/fsimpl/memfs/directory.go +++ /dev/null @@ -1,187 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package memfs - -import ( - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" -) - -type directory struct { - inode inode - - // 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 filesystem.mu. - childList dentryList -} - -func (fs *filesystem) newDirectory(creds *auth.Credentials, mode linux.FileMode) *inode { - dir := &directory{} - dir.inode.init(dir, fs, creds, mode) - dir.inode.nlink = 2 // from "." and parent directory or ".." for root - return &dir.inode -} - -func (i *inode) isDir() bool { - _, ok := i.impl.(*directory) - return ok -} - -type directoryFD struct { - fileDescription - vfs.DirectoryFileDescriptionDefaultImpl - - // Protected by filesystem.mu. - iter *dentry - off int64 -} - -// Release implements vfs.FileDescriptionImpl.Release. -func (fd *directoryFD) Release() { - if fd.iter != nil { - fs := fd.filesystem() - dir := fd.inode().impl.(*directory) - fs.mu.Lock() - dir.childList.Remove(fd.iter) - fs.mu.Unlock() - fd.iter = nil - } -} - -// IterDirents implements vfs.FileDescriptionImpl.IterDirents. -func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { - fs := fd.filesystem() - vfsd := fd.vfsfd.VirtualDentry().Dentry() - - fs.mu.Lock() - defer fs.mu.Unlock() - - if fd.off == 0 { - if !cb.Handle(vfs.Dirent{ - Name: ".", - Type: linux.DT_DIR, - Ino: vfsd.Impl().(*dentry).inode.ino, - NextOff: 1, - }) { - return nil - } - fd.off++ - } - if fd.off == 1 { - parentInode := vfsd.ParentOrSelf().Impl().(*dentry).inode - if !cb.Handle(vfs.Dirent{ - Name: "..", - Type: parentInode.direntType(), - Ino: parentInode.ino, - NextOff: 2, - }) { - return nil - } - fd.off++ - } - - dir := vfsd.Impl().(*dentry).inode.impl.(*directory) - 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 !cb.Handle(vfs.Dirent{ - Name: child.vfsd.Name(), - Type: child.inode.direntType(), - Ino: child.inode.ino, - NextOff: fd.off + 1, - }) { - dir.childList.InsertBefore(child, fd.iter) - return nil - } - 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) { - fs := fd.filesystem() - fs.mu.Lock() - defer fs.mu.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 - } - - dir := fd.inode().impl.(*directory) - - // 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/memfs/filesystem.go b/pkg/sentry/fsimpl/memfs/filesystem.go deleted file mode 100644 index b063e09a3..000000000 --- a/pkg/sentry/fsimpl/memfs/filesystem.go +++ /dev/null @@ -1,698 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package memfs - -import ( - "fmt" - "sync/atomic" - - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/fspath" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" -) - -// 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(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { - if !d.inode.isDir() { - return nil, syserror.ENOTDIR - } - if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { - return nil, err - } -afterSymlink: - nextVFSD, err := rp.ResolveComponent(&d.vfsd) - if err != nil { - return nil, err - } - if nextVFSD == nil { - // Since the Dentry tree is the sole source of truth for memfs, if it's - // not in the Dentry tree, it doesn't exist. - return nil, syserror.ENOENT - } - next := nextVFSD.Impl().(*dentry) - if symlink, ok := next.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { - // TODO: symlink traversals update access time - if err := rp.HandleSymlink(symlink.target); err != nil { - return nil, err - } - goto afterSymlink // don't check the current directory again - } - rp.Advance() - return next, 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(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { - for !rp.Final() { - next, err := stepLocked(rp, d) - if err != nil { - return nil, err - } - d = next - } - if !d.inode.isDir() { - return nil, syserror.ENOTDIR - } - return d, 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(rp *vfs.ResolvingPath) (*dentry, error) { - d := rp.Start().Impl().(*dentry) - for !rp.Done() { - next, err := stepLocked(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(rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { - fs.mu.Lock() - defer fs.mu.Unlock() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) - if err != nil { - return err - } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { - return err - } - name := rp.Component() - if name == "." || name == ".." { - return syserror.EEXIST - } - // Call parent.vfsd.Child() instead of stepLocked() or rp.ResolveChild(), - // because if the child exists we want to return EEXIST immediately instead - // of attempting symlink/mount traversal. - if parent.vfsd.Child(name) != nil { - return syserror.EEXIST - } - if !dir && rp.MustBeDir() { - return syserror.ENOENT - } - // In memfs, the only way to cause a dentry to be disowned is by removing - // it from the filesystem, so this check is equivalent to checking if - // parent has been removed. - if parent.vfsd.IsDisowned() { - return syserror.ENOENT - } - mnt := rp.Mount() - if err := mnt.CheckBeginWrite(); err != nil { - return err - } - defer mnt.EndWrite() - return create(parent, name) -} - -// 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(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, true /* isDir */); 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() - d, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) - if err != nil { - return nil, err - } - d.IncRef() - return &d.vfsd, nil -} - -// LinkAt implements vfs.FilesystemImpl.LinkAt. -func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { - if rp.Mount() != vd.Mount() { - return syserror.EXDEV - } - d := vd.Dentry().Impl().(*dentry) - if d.inode.isDir() { - return syserror.EPERM - } - if d.inode.nlink == 0 { - return syserror.ENOENT - } - if d.inode.nlink == maxLinks { - return syserror.EMLINK - } - d.inode.incLinksLocked() - child := fs.newDentry(d.inode) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) - return nil - }) -} - -// MkdirAt implements vfs.FilesystemImpl.MkdirAt. -func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { - return fs.doCreateAt(rp, true /* dir */, func(parent *dentry, name string) error { - if parent.inode.nlink == maxLinks { - return syserror.EMLINK - } - parent.inode.incLinksLocked() // from child's ".." - child := fs.newDentry(fs.newDirectory(rp.Credentials(), opts.Mode)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) - return nil - }) -} - -// MknodAt implements vfs.FilesystemImpl.MknodAt. -func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { - switch opts.Mode.FileType() { - case 0, linux.S_IFREG: - child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) - return nil - case linux.S_IFIFO: - child := fs.newDentry(fs.newNamedPipe(rp.Credentials(), opts.Mode)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) - return nil - case linux.S_IFBLK, linux.S_IFCHR, linux.S_IFSOCK: - // Not yet supported. - return syserror.EPERM - default: - return syserror.EINVAL - } - }) -} - -// 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(rp) - if err != nil { - return nil, err - } - return d.open(ctx, rp, opts.Flags, 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 directories with O_CREAT. - if rp.MustBeDir() { - return nil, syserror.EISDIR - } - if mustCreate { - return nil, syserror.EEXIST - } - return start.open(ctx, rp, opts.Flags, false /* afterCreate */) - } -afterTrailingSymlink: - parent, err := walkParentDirLocked(rp, start) - if err != nil { - return nil, err - } - // Check for search permission in the parent directory. - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); 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 - } - // Determine whether or not we need to create a file. - child, err := stepLocked(rp, parent) - if err == syserror.ENOENT { - // Already checked for searchability above; now check for writability. - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { - return nil, err - } - if err := rp.Mount().CheckBeginWrite(); err != nil { - return nil, err - } - defer rp.Mount().EndWrite() - // Create and open the child. - child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) - return child.open(ctx, rp, opts.Flags, true) - } - if err != nil { - return nil, err - } - // Do we need to resolve a trailing symlink? - if !rp.Done() { - start = parent - goto afterTrailingSymlink - } - // Open existing file. - if mustCreate { - return nil, syserror.EEXIST - } - return child.open(ctx, rp, opts.Flags, false) -} - -func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, flags uint32, afterCreate bool) (*vfs.FileDescription, error) { - ats := vfs.AccessTypesForOpenFlags(flags) - if !afterCreate { - if err := d.inode.checkPermissions(rp.Credentials(), ats, d.inode.isDir()); err != nil { - return nil, err - } - } - mnt := rp.Mount() - switch impl := d.inode.impl.(type) { - case *regularFile: - var fd regularFileFD - fd.readable = vfs.MayReadFileWithOpenFlags(flags) - fd.writable = vfs.MayWriteFileWithOpenFlags(flags) - if fd.writable { - if err := mnt.CheckBeginWrite(); err != nil { - return nil, err - } - // mnt.EndWrite() is called by regularFileFD.Release(). - } - fd.vfsfd.Init(&fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}) - if flags&linux.O_TRUNC != 0 { - impl.mu.Lock() - impl.data = impl.data[:0] - atomic.StoreInt64(&impl.dataLen, 0) - impl.mu.Unlock() - } - return &fd.vfsfd, nil - case *directory: - // Can't open directories writably. - if ats&vfs.MayWrite != 0 { - return nil, syserror.EISDIR - } - var fd directoryFD - fd.vfsfd.Init(&fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}) - return &fd.vfsfd, nil - case *symlink: - // Can't open symlinks without O_PATH (which is unimplemented). - return nil, syserror.ELOOP - case *namedPipe: - return newNamedPipeFD(ctx, impl, rp, &d.vfsd, flags) - 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(rp) - if err != nil { - return "", err - } - symlink, ok := d.inode.impl.(*symlink) - if !ok { - return "", syserror.EINVAL - } - 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() - newParent, err := walkParentDirLocked(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() - - oldParent := oldParentVD.Dentry().Impl().(*dentry) - if err := oldParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { - return err - } - // Call vfs.Dentry.Child() instead of stepLocked() or rp.ResolveChild(), - // because if the existing child is a symlink or mount point then we want - // to rename over it rather than follow it. - renamedVFSD := oldParent.vfsd.Child(oldName) - if renamedVFSD == nil { - return syserror.ENOENT - } - renamed := renamedVFSD.Impl().(*dentry) - if renamed.inode.isDir() { - if renamed == newParent || renamedVFSD.IsAncestorOf(&newParent.vfsd) { - return syserror.EINVAL - } - if oldParent != newParent { - // Writability is needed to change renamed's "..". - if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true /* isDir */); err != nil { - return err - } - } - } else { - if opts.MustBeDir || rp.MustBeDir() { - return syserror.ENOTDIR - } - } - - if err := newParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { - return err - } - replacedVFSD := newParent.vfsd.Child(newName) - var replaced *dentry - if replacedVFSD != nil { - replaced = replacedVFSD.Impl().(*dentry) - if replaced.inode.isDir() { - if !renamed.inode.isDir() { - return syserror.EISDIR - } - if replaced.vfsd.HasChildren() { - return syserror.ENOTEMPTY - } - } else { - if rp.MustBeDir() { - return syserror.ENOTDIR - } - if renamed.inode.isDir() { - return syserror.ENOTDIR - } - } - } else { - if renamed.inode.isDir() && newParent.inode.nlink == maxLinks { - return syserror.EMLINK - } - } - if newParent.vfsd.IsDisowned() { - 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 renamedVFSD == replacedVFSD { - return nil - } - vfsObj := rp.VirtualFilesystem() - oldParentDir := oldParent.inode.impl.(*directory) - newParentDir := newParent.inode.impl.(*directory) - if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), renamedVFSD, replacedVFSD); err != nil { - return err - } - if replaced != nil { - newParentDir.childList.Remove(replaced) - if replaced.inode.isDir() { - newParent.inode.decLinksLocked() // from replaced's ".." - } - replaced.inode.decLinksLocked() - } - oldParentDir.childList.Remove(renamed) - newParentDir.childList.PushBack(renamed) - if renamed.inode.isDir() { - oldParent.inode.decLinksLocked() - newParent.inode.incLinksLocked() - } - // TODO: update timestamps and parent directory sizes - vfsObj.CommitRenameReplaceDentry(renamedVFSD, &newParent.vfsd, newName, replacedVFSD) - 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() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) - if err != nil { - return err - } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { - return err - } - name := rp.Component() - if name == "." { - return syserror.EINVAL - } - if name == ".." { - return syserror.ENOTEMPTY - } - childVFSD := parent.vfsd.Child(name) - if childVFSD == nil { - return syserror.ENOENT - } - child := childVFSD.Impl().(*dentry) - if !child.inode.isDir() { - return syserror.ENOTDIR - } - if childVFSD.HasChildren() { - return syserror.ENOTEMPTY - } - mnt := rp.Mount() - if err := mnt.CheckBeginWrite(); err != nil { - return err - } - defer mnt.EndWrite() - vfsObj := rp.VirtualFilesystem() - if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { - return err - } - parent.inode.impl.(*directory).childList.Remove(child) - parent.inode.decLinksLocked() // from child's ".." - child.inode.decLinksLocked() - vfsObj.CommitDeleteDentry(childVFSD) - return nil -} - -// SetStatAt implements vfs.FilesystemImpl.SetStatAt. -func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { - fs.mu.RLock() - defer fs.mu.RUnlock() - _, err := resolveLocked(rp) - if err != nil { - return err - } - if opts.Stat.Mask == 0 { - return nil - } - // TODO: implement inode.setStat - return syserror.EPERM -} - -// 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(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() - _, err := resolveLocked(rp) - if err != nil { - return linux.Statfs{}, err - } - // TODO: actually implement statfs - return linux.Statfs{}, syserror.ENOSYS -} - -// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. -func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { - return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { - child := fs.newDentry(fs.newSymlink(rp.Credentials(), target)) - parent.vfsd.InsertChild(&child.vfsd, name) - parent.inode.impl.(*directory).childList.PushBack(child) - 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() - parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) - if err != nil { - return err - } - if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { - return err - } - name := rp.Component() - if name == "." || name == ".." { - return syserror.EISDIR - } - childVFSD := parent.vfsd.Child(name) - if childVFSD == nil { - return syserror.ENOENT - } - child := childVFSD.Impl().(*dentry) - 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() - if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { - return err - } - parent.inode.impl.(*directory).childList.Remove(child) - child.inode.decLinksLocked() - vfsObj.CommitDeleteDentry(childVFSD) - return nil -} - -// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { - fs.mu.RLock() - defer fs.mu.RUnlock() - _, err := resolveLocked(rp) - if err != nil { - return nil, err - } - // TODO(b/127675828): support extended attributes - return nil, syserror.ENOTSUP -} - -// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { - fs.mu.RLock() - defer fs.mu.RUnlock() - _, err := resolveLocked(rp) - if err != nil { - return "", err - } - // TODO(b/127675828): support extended attributes - return "", syserror.ENOTSUP -} - -// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. -func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { - fs.mu.RLock() - defer fs.mu.RUnlock() - _, err := resolveLocked(rp) - if err != nil { - return err - } - // TODO(b/127675828): support extended attributes - return syserror.ENOTSUP -} - -// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. -func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { - fs.mu.RLock() - defer fs.mu.RUnlock() - _, err := resolveLocked(rp) - if err != nil { - return err - } - // TODO(b/127675828): support extended attributes - return syserror.ENOTSUP -} - -// 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() - return vfs.GenericPrependPath(vfsroot, vd, b) -} diff --git a/pkg/sentry/fsimpl/memfs/memfs.go b/pkg/sentry/fsimpl/memfs/memfs.go deleted file mode 100644 index 8d0167c93..000000000 --- a/pkg/sentry/fsimpl/memfs/memfs.go +++ /dev/null @@ -1,293 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -// Package memfs provides a filesystem implementation that behaves like tmpfs: -// the Dentry tree is the sole source of truth for the state of the filesystem. -// -// memfs is intended primarily to demonstrate filesystem implementation -// patterns. Real uses cases for an in-memory filesystem should use tmpfs -// instead. -// -// Lock order: -// -// filesystem.mu -// regularFileFD.offMu -// regularFile.mu -// inode.mu -package memfs - -import ( - "fmt" - "math" - "sync" - "sync/atomic" - - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" -) - -// FilesystemType implements vfs.FilesystemType. -type FilesystemType struct{} - -// filesystem implements vfs.FilesystemImpl. -type filesystem struct { - vfsfs vfs.Filesystem - - // mu serializes changes to the Dentry tree. - mu sync.RWMutex - - nextInoMinusOne uint64 // accessed using atomic memory operations -} - -// GetFilesystem implements vfs.FilesystemType.GetFilesystem. -func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { - var fs filesystem - fs.vfsfs.Init(vfsObj, &fs) - root := fs.newDentry(fs.newDirectory(creds, 01777)) - return &fs.vfsfs, &root.vfsd, nil -} - -// Release implements vfs.FilesystemImpl.Release. -func (fs *filesystem) Release() { -} - -// dentry implements vfs.DentryImpl. -type dentry struct { - vfsd vfs.Dentry - - // inode is the inode represented by this dentry. Multiple Dentries may - // share a single non-directory inode (with hard links). inode is - // immutable. - inode *inode - - // memfs 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. - // (memfs doesn't really have resources to release, but we implement - // reference counting because tmpfs regular files will.) - - // dentryEntry (ugh) links dentries into their parent directory.childList. - dentryEntry -} - -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() { - d.inode.decRef() -} - -// inode represents a filesystem object. -type inode struct { - // refs is a reference count. refs is accessed using atomic memory - // operations. - // - // A reference is held on all inodes that are reachable in the filesystem - // tree. For non-directories (which may have multiple hard links), this - // means that a reference is dropped when nlink reaches 0. For directories, - // nlink never reaches 0 due to the "." entry; instead, - // filesystem.RmdirAt() drops the reference. - refs int64 - - // Inode metadata; protected by mu and accessed using atomic memory - // operations unless otherwise specified. - mu sync.RWMutex - mode uint32 // excluding file type bits, which are based on impl - 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 - - impl interface{} // immutable -} - -const maxLinks = math.MaxUint32 - -func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, mode linux.FileMode) { - i.refs = 1 - i.mode = uint32(mode) - i.uid = uint32(creds.EffectiveKUID) - i.gid = uint32(creds.EffectiveKGID) - i.ino = atomic.AddUint64(&fs.nextInoMinusOne, 1) - // i.nlink initialized by caller - i.impl = impl -} - -// 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("memfs.inode.incLinksLocked() called with no existing links") - } - if i.nlink == maxLinks { - panic("memfs.inode.incLinksLocked() called with maximum link count") - } - atomic.AddUint32(&i.nlink, 1) -} - -// decLinksLocked decrements i's link count. -// -// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0. -func (i *inode) decLinksLocked() { - if i.nlink == 0 { - panic("memfs.inode.decLinksLocked() called with no existing links") - } - atomic.AddUint32(&i.nlink, ^uint32(0)) -} - -func (i *inode) incRef() { - if atomic.AddInt64(&i.refs, 1) <= 1 { - panic("memfs.inode.incRef() called without holding a reference") - } -} - -func (i *inode) tryIncRef() bool { - for { - refs := atomic.LoadInt64(&i.refs) - if refs == 0 { - return false - } - if atomic.CompareAndSwapInt64(&i.refs, refs, refs+1) { - return true - } - } -} - -func (i *inode) decRef() { - if refs := atomic.AddInt64(&i.refs, -1); refs == 0 { - // This is unnecessary; it's mostly to simulate what tmpfs would do. - if regfile, ok := i.impl.(*regularFile); ok { - regfile.mu.Lock() - regfile.data = nil - atomic.StoreInt64(®file.dataLen, 0) - regfile.mu.Unlock() - } - } else if refs < 0 { - panic("memfs.inode.decRef() called without holding a reference") - } -} - -func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error { - return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&i.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. -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 - stat.Blksize = 1 // usermem.PageSize in tmpfs - 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 - // TODO: device number - switch impl := i.impl.(type) { - case *regularFile: - stat.Mode |= linux.S_IFREG - stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS - stat.Size = uint64(atomic.LoadInt64(&impl.dataLen)) - // In tmpfs, this will be FileRangeSet.Span() / 512 (but also cached in - // a uint64 accessed using atomic memory operations to avoid taking - // locks). - stat.Blocks = allocatedBlocksForSize(stat.Size) - case *directory: - stat.Mode |= linux.S_IFDIR - case *symlink: - stat.Mode |= linux.S_IFLNK - stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS - stat.Size = uint64(len(impl.target)) - stat.Blocks = allocatedBlocksForSize(stat.Size) - case *namedPipe: - stat.Mode |= linux.S_IFIFO - default: - panic(fmt.Sprintf("unknown inode type: %T", i.impl)) - } -} - -// 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 i.impl.(type) { - case *regularFile: - return linux.DT_REG - case *directory: - return linux.DT_DIR - case *symlink: - return linux.DT_LNK - default: - panic(fmt.Sprintf("unknown inode type: %T", i.impl)) - } -} - -// fileDescription is embedded by memfs implementations of -// vfs.FileDescriptionImpl. -type fileDescription struct { - vfsfd vfs.FileDescription - vfs.FileDescriptionDefaultImpl -} - -func (fd *fileDescription) filesystem() *filesystem { - return fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) -} - -func (fd *fileDescription) inode() *inode { - return fd.vfsfd.Dentry().Impl().(*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 { - if opts.Stat.Mask == 0 { - return nil - } - // TODO: implement inode.setStat - return syserror.EPERM -} diff --git a/pkg/sentry/fsimpl/memfs/named_pipe.go b/pkg/sentry/fsimpl/memfs/named_pipe.go deleted file mode 100644 index b5a204438..000000000 --- a/pkg/sentry/fsimpl/memfs/named_pipe.go +++ /dev/null @@ -1,60 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package memfs - -import ( - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" - "gvisor.dev/gvisor/pkg/sentry/usermem" - "gvisor.dev/gvisor/pkg/sentry/vfs" -) - -type namedPipe struct { - inode inode - - pipe *pipe.VFSPipe -} - -// Preconditions: -// * fs.mu must be locked. -// * rp.Mount().CheckBeginWrite() has been called successfully. -func (fs *filesystem) newNamedPipe(creds *auth.Credentials, mode linux.FileMode) *inode { - file := &namedPipe{pipe: pipe.NewVFSPipe(pipe.DefaultPipeSize, usermem.PageSize)} - file.inode.init(file, fs, creds, mode) - file.inode.nlink = 1 // Only the parent has a link. - return &file.inode -} - -// namedPipeFD implements vfs.FileDescriptionImpl. Methods are implemented -// entirely via struct embedding. -type namedPipeFD struct { - fileDescription - - *pipe.VFSPipeFD -} - -func newNamedPipeFD(ctx context.Context, np *namedPipe, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { - var err error - var fd namedPipeFD - fd.VFSPipeFD, err = np.pipe.NewVFSPipeFD(ctx, rp, vfsd, &fd.vfsfd, flags) - if err != nil { - return nil, err - } - mnt := rp.Mount() - fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}) - return &fd.vfsfd, nil -} diff --git a/pkg/sentry/fsimpl/memfs/pipe_test.go b/pkg/sentry/fsimpl/memfs/pipe_test.go deleted file mode 100644 index 807c1af7a..000000000 --- a/pkg/sentry/fsimpl/memfs/pipe_test.go +++ /dev/null @@ -1,235 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package memfs - -import ( - "bytes" - "testing" - - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/fspath" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/usermem" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" -) - -const fileName = "mypipe" - -func TestSeparateFDs(t *testing.T) { - ctx, creds, vfsObj, root := setup(t) - defer root.DecRef() - - // Open the read side. This is done in a concurrently because opening - // One end the pipe blocks until the other end is opened. - pop := vfs.PathOperation{ - Root: root, - Start: root, - 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() - - rfd, ok := <-rfdchan - if !ok { - t.Fatalf("failed to open pipe for reading %q", fileName) - } - defer rfd.DecRef() - - const msg = "vamos azul" - checkEmpty(ctx, t, rfd) - checkWrite(ctx, t, wfd, msg) - checkRead(ctx, t, rfd, msg) -} - -func TestNonblockingRead(t *testing.T) { - ctx, creds, vfsObj, root := setup(t) - defer root.DecRef() - - // Open the read side as nonblocking. - pop := vfs.PathOperation{ - Root: root, - Start: root, - 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() - - // Open the write side. - openOpts = vfs.OpenOptions{Flags: linux.O_WRONLY} - wfd, err := vfsObj.OpenAt(ctx, creds, &pop, &openOpts) - if err != nil { - t.Fatalf("failed to open pipe for writing %q: %v", fileName, err) - } - defer wfd.DecRef() - - const msg = "geh blau" - checkEmpty(ctx, t, rfd) - checkWrite(ctx, t, wfd, msg) - checkRead(ctx, t, rfd, msg) -} - -func TestNonblockingWriteError(t *testing.T) { - ctx, creds, vfsObj, root := setup(t) - defer root.DecRef() - - // Open the write side as nonblocking, which should return ENXIO. - pop := vfs.PathOperation{ - Root: root, - Start: root, - 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() - - // 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() - - const msg = "forza blu" - checkEmpty(ctx, t, fd) - checkWrite(ctx, t, fd, msg) - checkRead(ctx, t, fd, msg) -} - -// setup creates a VFS with a pipe in the root directory at path fileName. The -// returned VirtualDentry must be DecRef()'d be the caller. It calls t.Fatal -// upon failure. -func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesystem, vfs.VirtualDentry) { - ctx := contexttest.Context(t) - creds := auth.CredentialsFromContext(ctx) - - // Create VFS. - vfsObj := vfs.New() - vfsObj.MustRegisterFilesystemType("memfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ - AllowUserMount: true, - }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "memfs", &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/memfs/regular_file.go b/pkg/sentry/fsimpl/memfs/regular_file.go deleted file mode 100644 index b7f4853b3..000000000 --- a/pkg/sentry/fsimpl/memfs/regular_file.go +++ /dev/null @@ -1,154 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package memfs - -import ( - "io" - "sync" - "sync/atomic" - - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/usermem" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" -) - -type regularFile struct { - inode inode - - mu sync.RWMutex - data []byte - // dataLen is len(data), but accessed using atomic memory operations to - // avoid locking in inode.stat(). - dataLen int64 -} - -func (fs *filesystem) newRegularFile(creds *auth.Credentials, mode linux.FileMode) *inode { - file := ®ularFile{} - file.inode.init(file, fs, creds, mode) - file.inode.nlink = 1 // from parent directory - return &file.inode -} - -type regularFileFD struct { - fileDescription - - // These are immutable. - readable bool - writable bool - - // 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() { - if fd.writable { - fd.vfsfd.VirtualDentry().Mount().EndWrite() - } -} - -// PRead implements vfs.FileDescriptionImpl.PRead. -func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { - if !fd.readable { - return 0, syserror.EINVAL - } - f := fd.inode().impl.(*regularFile) - f.mu.RLock() - if offset >= int64(len(f.data)) { - f.mu.RUnlock() - return 0, io.EOF - } - n, err := dst.CopyOut(ctx, f.data[offset:]) - f.mu.RUnlock() - return int64(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) { - if !fd.writable { - return 0, syserror.EINVAL - } - if offset < 0 { - return 0, syserror.EINVAL - } - srclen := src.NumBytes() - if srclen == 0 { - return 0, nil - } - f := fd.inode().impl.(*regularFile) - f.mu.Lock() - end := offset + srclen - if end < offset { - // Overflow. - f.mu.Unlock() - return 0, syserror.EFBIG - } - if end > f.dataLen { - f.data = append(f.data, make([]byte, end-f.dataLen)...) - atomic.StoreInt64(&f.dataLen, end) - } - n, err := src.CopyIn(ctx, f.data[offset:end]) - f.mu.Unlock() - return int64(n), 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, err := fd.PWrite(ctx, src, fd.off, opts) - fd.off += n - 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 += atomic.LoadInt64(&fd.inode().impl.(*regularFile).dataLen) - default: - return 0, syserror.EINVAL - } - if offset < 0 { - return 0, syserror.EINVAL - } - fd.off = offset - return offset, nil -} - -// Sync implements vfs.FileDescriptionImpl.Sync. -func (fd *regularFileFD) Sync(ctx context.Context) error { - return nil -} diff --git a/pkg/sentry/fsimpl/memfs/symlink.go b/pkg/sentry/fsimpl/memfs/symlink.go deleted file mode 100644 index b2ac2cbeb..000000000 --- a/pkg/sentry/fsimpl/memfs/symlink.go +++ /dev/null @@ -1,36 +0,0 @@ -// Copyright 2019 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package memfs - -import ( - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" -) - -type symlink struct { - inode inode - target string // immutable -} - -func (fs *filesystem) newSymlink(creds *auth.Credentials, target string) *inode { - link := &symlink{ - target: target, - } - link.inode.init(link, fs, creds, 0777) - 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/BUILD b/pkg/sentry/fsimpl/tmpfs/BUILD new file mode 100644 index 000000000..a5b285987 --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/BUILD @@ -0,0 +1,92 @@ +load("//tools/go_stateify:defs.bzl", "go_library") +load("@io_bazel_rules_go//go:def.bzl", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") + +package(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_library( + name = "tmpfs", + srcs = [ + "dentry_list.go", + "directory.go", + "filesystem.go", + "named_pipe.go", + "regular_file.go", + "symlink.go", + "tmpfs.go", + ], + importpath = "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs", + deps = [ + "//pkg/abi/linux", + "//pkg/amutex", + "//pkg/fspath", + "//pkg/log", + "//pkg/sentry/arch", + "//pkg/sentry/context", + "//pkg/sentry/fs", + "//pkg/sentry/fs/fsutil", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/pipe", + "//pkg/sentry/memmap", + "//pkg/sentry/pgalloc", + "//pkg/sentry/platform", + "//pkg/sentry/safemem", + "//pkg/sentry/usage", + "//pkg/sentry/usermem", + "//pkg/sentry/vfs", + "//pkg/syserror", + ], +) + +go_test( + name = "benchmark_test", + size = "small", + srcs = ["benchmark_test.go"], + deps = [ + ":tmpfs", + "//pkg/abi/linux", + "//pkg/fspath", + "//pkg/refs", + "//pkg/sentry/context", + "//pkg/sentry/context/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", + ], + embed = [":tmpfs"], + deps = [ + "//pkg/abi/linux", + "//pkg/fspath", + "//pkg/sentry/context", + "//pkg/sentry/context/contexttest", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/contexttest", + "//pkg/sentry/usermem", + "//pkg/sentry/vfs", + "//pkg/syserror", + ], +) diff --git a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go new file mode 100644 index 000000000..d88c83499 --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go @@ -0,0 +1,487 @@ +// 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/fspath" + "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/context/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/ (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() + 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() + + var filePathBuilder strings.Builder + filePathBuilder.WriteByte('/') + + // Create nested directories with given depth. + root := mntns.Root() + defer root.DecRef() + d := root + d.IncRef() + defer d.DecRef() + 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() + 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() + 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 BenchmarkVFS2MemfsStat(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.New() + 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(vfsObj) + + var filePathBuilder strings.Builder + filePathBuilder.WriteByte('/') + + // Create nested directories with given depth. + root := mntns.Root() + defer root.DecRef() + 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() + 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() + vd = vfs.VirtualDentry{} + if err != nil { + b.Fatalf("failed to create file %q: %v", filename, err) + } + defer fd.DecRef() + 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() + + var filePathBuilder strings.Builder + filePathBuilder.WriteByte('/') + + // Create and mount the submount. + root := mntns.Root() + defer root.DecRef() + 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() + 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() + 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() + 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() + 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 BenchmarkVFS2MemfsMountStat(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.New() + 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(vfsObj) + + var filePathBuilder strings.Builder + filePathBuilder.WriteByte('/') + + // Create the mount point. + root := mntns.Root() + defer root.DecRef() + 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() + // 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() + vd = nextVD + filePathBuilder.WriteString(name) + filePathBuilder.WriteByte('/') + } + + // Verify that we didn't create any directories under the mount + // point (i.e. they were all created on the submount). + firstDirName := fmt.Sprintf("%d", depth) + if child := mountPoint.Dentry().Child(firstDirName); child != nil { + b.Fatalf("created directory %q under root mount, not submount", firstDirName) + } + + // 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() + if err != nil { + b.Fatalf("failed to create file %q: %v", filename, err) + } + fd.DecRef() + 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/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go new file mode 100644 index 000000000..887ca2619 --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/directory.go @@ -0,0 +1,187 @@ +// 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/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +type directory struct { + inode inode + + // 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 filesystem.mu. + childList dentryList +} + +func (fs *filesystem) newDirectory(creds *auth.Credentials, mode linux.FileMode) *inode { + dir := &directory{} + dir.inode.init(dir, fs, creds, mode) + dir.inode.nlink = 2 // from "." and parent directory or ".." for root + return &dir.inode +} + +func (i *inode) isDir() bool { + _, ok := i.impl.(*directory) + return ok +} + +type directoryFD struct { + fileDescription + vfs.DirectoryFileDescriptionDefaultImpl + + // Protected by filesystem.mu. + iter *dentry + off int64 +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *directoryFD) Release() { + if fd.iter != nil { + fs := fd.filesystem() + dir := fd.inode().impl.(*directory) + fs.mu.Lock() + dir.childList.Remove(fd.iter) + fs.mu.Unlock() + fd.iter = nil + } +} + +// IterDirents implements vfs.FileDescriptionImpl.IterDirents. +func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { + fs := fd.filesystem() + vfsd := fd.vfsfd.VirtualDentry().Dentry() + + fs.mu.Lock() + defer fs.mu.Unlock() + + if fd.off == 0 { + if !cb.Handle(vfs.Dirent{ + Name: ".", + Type: linux.DT_DIR, + Ino: vfsd.Impl().(*dentry).inode.ino, + NextOff: 1, + }) { + return nil + } + fd.off++ + } + if fd.off == 1 { + parentInode := vfsd.ParentOrSelf().Impl().(*dentry).inode + if !cb.Handle(vfs.Dirent{ + Name: "..", + Type: parentInode.direntType(), + Ino: parentInode.ino, + NextOff: 2, + }) { + return nil + } + fd.off++ + } + + dir := vfsd.Impl().(*dentry).inode.impl.(*directory) + 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 !cb.Handle(vfs.Dirent{ + Name: child.vfsd.Name(), + Type: child.inode.direntType(), + Ino: child.inode.ino, + NextOff: fd.off + 1, + }) { + dir.childList.InsertBefore(child, fd.iter) + return nil + } + 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) { + fs := fd.filesystem() + fs.mu.Lock() + defer fs.mu.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 + } + + dir := fd.inode().impl.(*directory) + + // 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..26979729e --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -0,0 +1,698 @@ +// 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/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// 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(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { + if !d.inode.isDir() { + return nil, syserror.ENOTDIR + } + if err := d.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); err != nil { + return nil, err + } +afterSymlink: + nextVFSD, err := rp.ResolveComponent(&d.vfsd) + if err != nil { + return nil, err + } + if nextVFSD == nil { + // Since the Dentry tree is the sole source of truth for tmpfs, if it's + // not in the Dentry tree, it doesn't exist. + return nil, syserror.ENOENT + } + next := nextVFSD.Impl().(*dentry) + if symlink, ok := next.inode.impl.(*symlink); ok && rp.ShouldFollowSymlink() { + // TODO: symlink traversals update access time + if err := rp.HandleSymlink(symlink.target); err != nil { + return nil, err + } + goto afterSymlink // don't check the current directory again + } + rp.Advance() + return next, 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(rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { + for !rp.Final() { + next, err := stepLocked(rp, d) + if err != nil { + return nil, err + } + d = next + } + if !d.inode.isDir() { + return nil, syserror.ENOTDIR + } + return d, 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(rp *vfs.ResolvingPath) (*dentry, error) { + d := rp.Start().Impl().(*dentry) + for !rp.Done() { + next, err := stepLocked(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(rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string) error) error { + fs.mu.Lock() + defer fs.mu.Unlock() + parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return err + } + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + name := rp.Component() + if name == "." || name == ".." { + return syserror.EEXIST + } + // Call parent.vfsd.Child() instead of stepLocked() or rp.ResolveChild(), + // because if the child exists we want to return EEXIST immediately instead + // of attempting symlink/mount traversal. + if parent.vfsd.Child(name) != nil { + return syserror.EEXIST + } + if !dir && rp.MustBeDir() { + return syserror.ENOENT + } + // In memfs, the only way to cause a dentry to be disowned is by removing + // it from the filesystem, so this check is equivalent to checking if + // parent has been removed. + if parent.vfsd.IsDisowned() { + return syserror.ENOENT + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + return create(parent, name) +} + +// 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(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, true /* isDir */); 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() + d, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return nil, err + } + d.IncRef() + return &d.vfsd, nil +} + +// LinkAt implements vfs.FilesystemImpl.LinkAt. +func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { + return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + if rp.Mount() != vd.Mount() { + return syserror.EXDEV + } + d := vd.Dentry().Impl().(*dentry) + if d.inode.isDir() { + return syserror.EPERM + } + if d.inode.nlink == 0 { + return syserror.ENOENT + } + if d.inode.nlink == maxLinks { + return syserror.EMLINK + } + d.inode.incLinksLocked() + child := fs.newDentry(d.inode) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + }) +} + +// MkdirAt implements vfs.FilesystemImpl.MkdirAt. +func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { + return fs.doCreateAt(rp, true /* dir */, func(parent *dentry, name string) error { + if parent.inode.nlink == maxLinks { + return syserror.EMLINK + } + parent.inode.incLinksLocked() // from child's ".." + child := fs.newDentry(fs.newDirectory(rp.Credentials(), opts.Mode)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + }) +} + +// MknodAt implements vfs.FilesystemImpl.MknodAt. +func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { + return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + switch opts.Mode.FileType() { + case 0, linux.S_IFREG: + child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + case linux.S_IFIFO: + child := fs.newDentry(fs.newNamedPipe(rp.Credentials(), opts.Mode)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return nil + case linux.S_IFBLK, linux.S_IFCHR, linux.S_IFSOCK: + // Not yet supported. + return syserror.EPERM + default: + return syserror.EINVAL + } + }) +} + +// 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(rp) + if err != nil { + return nil, err + } + return d.open(ctx, rp, opts.Flags, 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 directories with O_CREAT. + if rp.MustBeDir() { + return nil, syserror.EISDIR + } + if mustCreate { + return nil, syserror.EEXIST + } + return start.open(ctx, rp, opts.Flags, false /* afterCreate */) + } +afterTrailingSymlink: + parent, err := walkParentDirLocked(rp, start) + if err != nil { + return nil, err + } + // Check for search permission in the parent directory. + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayExec, true); 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 + } + // Determine whether or not we need to create a file. + child, err := stepLocked(rp, parent) + if err == syserror.ENOENT { + // Already checked for searchability above; now check for writability. + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true); err != nil { + return nil, err + } + if err := rp.Mount().CheckBeginWrite(); err != nil { + return nil, err + } + defer rp.Mount().EndWrite() + // Create and open the child. + child := fs.newDentry(fs.newRegularFile(rp.Credentials(), opts.Mode)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + return child.open(ctx, rp, opts.Flags, true) + } + if err != nil { + return nil, err + } + // Do we need to resolve a trailing symlink? + if !rp.Done() { + start = parent + goto afterTrailingSymlink + } + // Open existing file. + if mustCreate { + return nil, syserror.EEXIST + } + return child.open(ctx, rp, opts.Flags, false) +} + +func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, flags uint32, afterCreate bool) (*vfs.FileDescription, error) { + ats := vfs.AccessTypesForOpenFlags(flags) + if !afterCreate { + if err := d.inode.checkPermissions(rp.Credentials(), ats, d.inode.isDir()); err != nil { + return nil, err + } + } + mnt := rp.Mount() + switch impl := d.inode.impl.(type) { + case *regularFile: + var fd regularFileFD + fd.readable = vfs.MayReadFileWithOpenFlags(flags) + fd.writable = vfs.MayWriteFileWithOpenFlags(flags) + if fd.writable { + if err := mnt.CheckBeginWrite(); err != nil { + return nil, err + } + // mnt.EndWrite() is called by regularFileFD.Release(). + } + fd.vfsfd.Init(&fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}) + if flags&linux.O_TRUNC != 0 { + impl.mu.Lock() + impl.data.Truncate(0, impl.memFile) + atomic.StoreUint64(&impl.size, 0) + impl.mu.Unlock() + } + return &fd.vfsfd, nil + case *directory: + // Can't open directories writably. + if ats&vfs.MayWrite != 0 { + return nil, syserror.EISDIR + } + var fd directoryFD + fd.vfsfd.Init(&fd, flags, mnt, &d.vfsd, &vfs.FileDescriptionOptions{}) + return &fd.vfsfd, nil + case *symlink: + // Can't open symlinks without O_PATH (which is unimplemented). + return nil, syserror.ELOOP + case *namedPipe: + return newNamedPipeFD(ctx, impl, rp, &d.vfsd, flags) + 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(rp) + if err != nil { + return "", err + } + symlink, ok := d.inode.impl.(*symlink) + if !ok { + return "", syserror.EINVAL + } + 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() + newParent, err := walkParentDirLocked(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() + + oldParent := oldParentVD.Dentry().Impl().(*dentry) + if err := oldParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + // Call vfs.Dentry.Child() instead of stepLocked() or rp.ResolveChild(), + // because if the existing child is a symlink or mount point then we want + // to rename over it rather than follow it. + renamedVFSD := oldParent.vfsd.Child(oldName) + if renamedVFSD == nil { + return syserror.ENOENT + } + renamed := renamedVFSD.Impl().(*dentry) + if renamed.inode.isDir() { + if renamed == newParent || renamedVFSD.IsAncestorOf(&newParent.vfsd) { + return syserror.EINVAL + } + if oldParent != newParent { + // Writability is needed to change renamed's "..". + if err := renamed.inode.checkPermissions(rp.Credentials(), vfs.MayWrite, true /* isDir */); err != nil { + return err + } + } + } else { + if opts.MustBeDir || rp.MustBeDir() { + return syserror.ENOTDIR + } + } + + if err := newParent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + replacedVFSD := newParent.vfsd.Child(newName) + var replaced *dentry + if replacedVFSD != nil { + replaced = replacedVFSD.Impl().(*dentry) + if replaced.inode.isDir() { + if !renamed.inode.isDir() { + return syserror.EISDIR + } + if replaced.vfsd.HasChildren() { + return syserror.ENOTEMPTY + } + } else { + if rp.MustBeDir() { + return syserror.ENOTDIR + } + if renamed.inode.isDir() { + return syserror.ENOTDIR + } + } + } else { + if renamed.inode.isDir() && newParent.inode.nlink == maxLinks { + return syserror.EMLINK + } + } + if newParent.vfsd.IsDisowned() { + 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 renamedVFSD == replacedVFSD { + return nil + } + vfsObj := rp.VirtualFilesystem() + oldParentDir := oldParent.inode.impl.(*directory) + newParentDir := newParent.inode.impl.(*directory) + if err := vfsObj.PrepareRenameDentry(vfs.MountNamespaceFromContext(ctx), renamedVFSD, replacedVFSD); err != nil { + return err + } + if replaced != nil { + newParentDir.childList.Remove(replaced) + if replaced.inode.isDir() { + newParent.inode.decLinksLocked() // from replaced's ".." + } + replaced.inode.decLinksLocked() + } + oldParentDir.childList.Remove(renamed) + newParentDir.childList.PushBack(renamed) + if renamed.inode.isDir() { + oldParent.inode.decLinksLocked() + newParent.inode.incLinksLocked() + } + // TODO: update timestamps and parent directory sizes + vfsObj.CommitRenameReplaceDentry(renamedVFSD, &newParent.vfsd, newName, replacedVFSD) + 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() + parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return err + } + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + name := rp.Component() + if name == "." { + return syserror.EINVAL + } + if name == ".." { + return syserror.ENOTEMPTY + } + childVFSD := parent.vfsd.Child(name) + if childVFSD == nil { + return syserror.ENOENT + } + child := childVFSD.Impl().(*dentry) + if !child.inode.isDir() { + return syserror.ENOTDIR + } + if childVFSD.HasChildren() { + return syserror.ENOTEMPTY + } + mnt := rp.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + vfsObj := rp.VirtualFilesystem() + if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + return err + } + parent.inode.impl.(*directory).childList.Remove(child) + parent.inode.decLinksLocked() // from child's ".." + child.inode.decLinksLocked() + vfsObj.CommitDeleteDentry(childVFSD) + return nil +} + +// SetStatAt implements vfs.FilesystemImpl.SetStatAt. +func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return err + } + if opts.Stat.Mask == 0 { + return nil + } + // TODO: implement inode.setStat + return syserror.EPERM +} + +// 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(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() + _, err := resolveLocked(rp) + if err != nil { + return linux.Statfs{}, err + } + // TODO: actually implement statfs + return linux.Statfs{}, syserror.ENOSYS +} + +// SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. +func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { + return fs.doCreateAt(rp, false /* dir */, func(parent *dentry, name string) error { + child := fs.newDentry(fs.newSymlink(rp.Credentials(), target)) + parent.vfsd.InsertChild(&child.vfsd, name) + parent.inode.impl.(*directory).childList.PushBack(child) + 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() + parent, err := walkParentDirLocked(rp, rp.Start().Impl().(*dentry)) + if err != nil { + return err + } + if err := parent.inode.checkPermissions(rp.Credentials(), vfs.MayWrite|vfs.MayExec, true /* isDir */); err != nil { + return err + } + name := rp.Component() + if name == "." || name == ".." { + return syserror.EISDIR + } + childVFSD := parent.vfsd.Child(name) + if childVFSD == nil { + return syserror.ENOENT + } + child := childVFSD.Impl().(*dentry) + 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() + if err := vfsObj.PrepareDeleteDentry(vfs.MountNamespaceFromContext(ctx), childVFSD); err != nil { + return err + } + parent.inode.impl.(*directory).childList.Remove(child) + child.inode.decLinksLocked() + vfsObj.CommitDeleteDentry(childVFSD) + return nil +} + +// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. +func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath) ([]string, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return nil, err + } + // TODO(b/127675828): support extended attributes + return nil, syserror.ENOTSUP +} + +// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. +func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) (string, error) { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return "", err + } + // TODO(b/127675828): support extended attributes + return "", syserror.ENOTSUP +} + +// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. +func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return err + } + // TODO(b/127675828): support extended attributes + return syserror.ENOTSUP +} + +// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. +func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { + fs.mu.RLock() + defer fs.mu.RUnlock() + _, err := resolveLocked(rp) + if err != nil { + return err + } + // TODO(b/127675828): support extended attributes + return syserror.ENOTSUP +} + +// 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() + return vfs.GenericPrependPath(vfsroot, vd, b) +} diff --git a/pkg/sentry/fsimpl/tmpfs/named_pipe.go b/pkg/sentry/fsimpl/tmpfs/named_pipe.go new file mode 100644 index 000000000..40bde54de --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/named_pipe.go @@ -0,0 +1,60 @@ +// 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/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +type namedPipe struct { + inode inode + + pipe *pipe.VFSPipe +} + +// Preconditions: +// * fs.mu must be locked. +// * rp.Mount().CheckBeginWrite() has been called successfully. +func (fs *filesystem) newNamedPipe(creds *auth.Credentials, mode linux.FileMode) *inode { + file := &namedPipe{pipe: pipe.NewVFSPipe(pipe.DefaultPipeSize, usermem.PageSize)} + file.inode.init(file, fs, creds, mode) + file.inode.nlink = 1 // Only the parent has a link. + return &file.inode +} + +// namedPipeFD implements vfs.FileDescriptionImpl. Methods are implemented +// entirely via struct embedding. +type namedPipeFD struct { + fileDescription + + *pipe.VFSPipeFD +} + +func newNamedPipeFD(ctx context.Context, np *namedPipe, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, flags uint32) (*vfs.FileDescription, error) { + var err error + var fd namedPipeFD + fd.VFSPipeFD, err = np.pipe.NewVFSPipeFD(ctx, rp, vfsd, &fd.vfsfd, flags) + if err != nil { + return nil, err + } + mnt := rp.Mount() + fd.vfsfd.Init(&fd, flags, mnt, vfsd, &vfs.FileDescriptionOptions{}) + return &fd.vfsfd, nil +} diff --git a/pkg/sentry/fsimpl/tmpfs/pipe_test.go b/pkg/sentry/fsimpl/tmpfs/pipe_test.go new file mode 100644 index 000000000..70b42a6ec --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/pipe_test.go @@ -0,0 +1,235 @@ +// 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/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/context/contexttest" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +const fileName = "mypipe" + +func TestSeparateFDs(t *testing.T) { + ctx, creds, vfsObj, root := setup(t) + defer root.DecRef() + + // Open the read side. This is done in a concurrently because opening + // One end the pipe blocks until the other end is opened. + pop := vfs.PathOperation{ + Root: root, + Start: root, + 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() + + rfd, ok := <-rfdchan + if !ok { + t.Fatalf("failed to open pipe for reading %q", fileName) + } + defer rfd.DecRef() + + const msg = "vamos azul" + checkEmpty(ctx, t, rfd) + checkWrite(ctx, t, wfd, msg) + checkRead(ctx, t, rfd, msg) +} + +func TestNonblockingRead(t *testing.T) { + ctx, creds, vfsObj, root := setup(t) + defer root.DecRef() + + // Open the read side as nonblocking. + pop := vfs.PathOperation{ + Root: root, + Start: root, + 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() + + // Open the write side. + openOpts = vfs.OpenOptions{Flags: linux.O_WRONLY} + wfd, err := vfsObj.OpenAt(ctx, creds, &pop, &openOpts) + if err != nil { + t.Fatalf("failed to open pipe for writing %q: %v", fileName, err) + } + defer wfd.DecRef() + + const msg = "geh blau" + checkEmpty(ctx, t, rfd) + checkWrite(ctx, t, wfd, msg) + checkRead(ctx, t, rfd, msg) +} + +func TestNonblockingWriteError(t *testing.T) { + ctx, creds, vfsObj, root := setup(t) + defer root.DecRef() + + // Open the write side as nonblocking, which should return ENXIO. + pop := vfs.PathOperation{ + Root: root, + Start: root, + 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() + + // 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() + + const msg = "forza blu" + checkEmpty(ctx, t, fd) + checkWrite(ctx, t, fd, msg) + checkRead(ctx, t, fd, msg) +} + +// setup creates a VFS with a pipe in the root directory at path fileName. The +// returned VirtualDentry must be DecRef()'d be the caller. It calls t.Fatal +// upon failure. +func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesystem, vfs.VirtualDentry) { + ctx := contexttest.Context(t) + creds := auth.CredentialsFromContext(ctx) + + // Create VFS. + vfsObj := vfs.New() + vfsObj.MustRegisterFilesystemType("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..f51e247a7 --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go @@ -0,0 +1,357 @@ +// 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 ( + "io" + "math" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "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/safemem" + "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +type regularFile struct { + inode inode + + // memFile is a platform.File used to allocate pages to this regularFile. + memFile *pgalloc.MemoryFile + + // mu protects the fields below. + mu sync.RWMutex + + // data maps offsets into the file to offsets into memFile that store + // the file's data. + data fsutil.FileRangeSet + + // size is the size of data, but accessed using atomic memory + // operations to avoid locking in inode.stat(). + size uint64 + + // seals represents file seals on this inode. + seals uint32 +} + +func (fs *filesystem) newRegularFile(creds *auth.Credentials, mode linux.FileMode) *inode { + file := ®ularFile{ + memFile: fs.memFile, + } + file.inode.init(file, fs, creds, mode) + file.inode.nlink = 1 // from parent directory + return &file.inode +} + +type regularFileFD struct { + fileDescription + + // These are immutable. + readable bool + writable bool + + // 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() { + if fd.writable { + fd.vfsfd.VirtualDentry().Mount().EndWrite() + } +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + if !fd.readable { + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + if dst.NumBytes() == 0 { + return 0, nil + } + f := fd.inode().impl.(*regularFile) + rw := getRegularFileReadWriter(f, offset) + n, err := dst.CopyOutFrom(ctx, rw) + putRegularFileReadWriter(rw) + return int64(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) { + if !fd.writable { + return 0, syserror.EINVAL + } + if offset < 0 { + return 0, syserror.EINVAL + } + srclen := src.NumBytes() + if srclen == 0 { + return 0, nil + } + f := fd.inode().impl.(*regularFile) + end := offset + srclen + if end < offset { + // Overflow. + return 0, syserror.EFBIG + } + rw := getRegularFileReadWriter(f, offset) + n, err := src.CopyInTo(ctx, rw) + putRegularFileReadWriter(rw) + return n, 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, err := fd.PWrite(ctx, src, fd.off, opts) + fd.off += n + 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 +} + +// Sync implements vfs.FileDescriptionImpl.Sync. +func (fd *regularFileFD) Sync(ctx context.Context) error { + return nil +} + +// 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 ®ularFileReadWriter{} + }, +} + +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.mu.RLock() + + // Compute the range to read (limited by file size and overflow-checked). + if rw.off >= rw.file.size { + rw.file.mu.RUnlock() + return 0, io.EOF + } + end := rw.file.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 { + rw.file.mu.RUnlock() + 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 { + rw.file.mu.RUnlock() + 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 { + rw.file.mu.RUnlock() + return done, err + } + + // Continue. + seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{} + } + } + rw.file.mu.RUnlock() + return done, nil +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + rw.file.mu.Lock() + + // 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 + rw.file.mu.Unlock() + 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. + rw.file.mu.Unlock() + 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 { + atomic.StoreUint64(&rw.file.size, rw.off) + } + + rw.file.mu.Unlock() + return done, retErr +} 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..3731c5b6f --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go @@ -0,0 +1,224 @@ +// 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/fspath" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/kernel/contexttest" + "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/vfs" +) + +// 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, filename string) (*vfs.FileDescription, func(), error) { + creds := auth.CredentialsFromContext(ctx) + + vfsObj := vfs.New() + vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) + if err != nil { + return nil, nil, fmt.Errorf("failed to create tmpfs root mount: %v", err) + } + root := mntns.Root() + + // Create the file that will be write/read. + fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filename), + FollowFinalSymlink: true, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL, + Mode: 0644, + }) + if err != nil { + root.DecRef() + mntns.DecRef(vfsObj) + return nil, nil, fmt.Errorf("failed to create file %q: %v", filename, err) + } + + return fd, func() { + root.DecRef() + mntns.DecRef(vfsObj) + }, nil +} + +// 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, "simpleReadWrite") + 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, "PRead") + 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 TestPRead(t *testing.T) { + ctx := contexttest.Context(t) + fd, cleanup, err := newFileFD(ctx, "PRead") + 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)) + } + }) + } + } +} diff --git a/pkg/sentry/fsimpl/tmpfs/symlink.go b/pkg/sentry/fsimpl/tmpfs/symlink.go new file mode 100644 index 000000000..5246aca84 --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/symlink.go @@ -0,0 +1,36 @@ +// 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/sentry/kernel/auth" +) + +type symlink struct { + inode inode + target string // immutable +} + +func (fs *filesystem) newSymlink(creds *auth.Credentials, target string) *inode { + link := &symlink{ + target: target, + } + link.inode.init(link, fs, creds, 0777) + 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..7be6faa5b --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -0,0 +1,299 @@ +// 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 a filesystem implementation that behaves like tmpfs: +// the Dentry tree is the sole source of truth for the state of the filesystem. +// +// Lock order: +// +// filesystem.mu +// regularFileFD.offMu +// regularFile.mu +// inode.mu +package tmpfs + +import ( + "fmt" + "math" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// 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 + + // mu serializes changes to the Dentry tree. + mu sync.RWMutex + + nextInoMinusOne uint64 // accessed using atomic memory operations +} + +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. +func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { + memFileProvider := pgalloc.MemoryFileProviderFromContext(ctx) + if memFileProvider == nil { + panic("MemoryFileProviderFromContext returned nil") + } + fs := filesystem{ + memFile: memFileProvider.MemoryFile(), + } + fs.vfsfs.Init(vfsObj, &fs) + root := fs.newDentry(fs.newDirectory(creds, 01777)) + return &fs.vfsfs, &root.vfsd, nil +} + +// Release implements vfs.FilesystemImpl.Release. +func (fs *filesystem) Release() { +} + +// dentry implements vfs.DentryImpl. +type dentry struct { + vfsd vfs.Dentry + + // inode is the inode represented by this dentry. Multiple Dentries may + // share a single non-directory inode (with hard links). inode is + // immutable. + inode *inode + + // 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. + // (tmpfs doesn't really have resources to release, but we implement + // reference counting because tmpfs regular files will.) + + // dentryEntry (ugh) links dentries into their parent directory.childList. + dentryEntry +} + +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() { + d.inode.decRef() +} + +// inode represents a filesystem object. +type inode struct { + // refs is a reference count. refs is accessed using atomic memory + // operations. + // + // A reference is held on all inodes that are reachable in the filesystem + // tree. For non-directories (which may have multiple hard links), this + // means that a reference is dropped when nlink reaches 0. For directories, + // nlink never reaches 0 due to the "." entry; instead, + // filesystem.RmdirAt() drops the reference. + refs int64 + + // Inode metadata; protected by mu and accessed using atomic memory + // operations unless otherwise specified. + mu sync.RWMutex + mode uint32 // excluding file type bits, which are based on impl + 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 + + impl interface{} // immutable +} + +const maxLinks = math.MaxUint32 + +func (i *inode) init(impl interface{}, fs *filesystem, creds *auth.Credentials, mode linux.FileMode) { + i.refs = 1 + i.mode = uint32(mode) + i.uid = uint32(creds.EffectiveKUID) + i.gid = uint32(creds.EffectiveKGID) + i.ino = atomic.AddUint64(&fs.nextInoMinusOne, 1) + // i.nlink initialized by caller + i.impl = impl +} + +// 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("memfs.inode.incLinksLocked() called with maximum link count") + } + atomic.AddUint32(&i.nlink, 1) +} + +// decLinksLocked decrements i's link count. +// +// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0. +func (i *inode) decLinksLocked() { + if i.nlink == 0 { + panic("tmpfs.inode.decLinksLocked() called with no existing links") + } + atomic.AddUint32(&i.nlink, ^uint32(0)) +} + +func (i *inode) incRef() { + if atomic.AddInt64(&i.refs, 1) <= 1 { + panic("tmpfs.inode.incRef() called without holding a reference") + } +} + +func (i *inode) tryIncRef() bool { + for { + refs := atomic.LoadInt64(&i.refs) + if refs == 0 { + return false + } + if atomic.CompareAndSwapInt64(&i.refs, refs, refs+1) { + return true + } + } +} + +func (i *inode) decRef() { + if refs := atomic.AddInt64(&i.refs, -1); refs == 0 { + // This is unnecessary; it's mostly to simulate what tmpfs would do. + if regFile, ok := i.impl.(*regularFile); ok { + regFile.mu.Lock() + regFile.data.DropAll(regFile.memFile) + atomic.StoreUint64(®File.size, 0) + regFile.mu.Unlock() + } + } else if refs < 0 { + panic("tmpfs.inode.decRef() called without holding a reference") + } +} + +func (i *inode) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes, isDir bool) error { + return vfs.GenericCheckPermissions(creds, ats, isDir, uint16(atomic.LoadUint32(&i.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. +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 + stat.Blksize = 1 // usermem.PageSize in tmpfs + 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 + // TODO: device number + switch impl := i.impl.(type) { + case *regularFile: + stat.Mode |= linux.S_IFREG + stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS + stat.Size = uint64(atomic.LoadUint64(&impl.size)) + // In tmpfs, this will be FileRangeSet.Span() / 512 (but also cached in + // a uint64 accessed using atomic memory operations to avoid taking + // locks). + stat.Blocks = allocatedBlocksForSize(stat.Size) + case *directory: + stat.Mode |= linux.S_IFDIR + case *symlink: + stat.Mode |= linux.S_IFLNK + stat.Mask |= linux.STATX_SIZE | linux.STATX_BLOCKS + stat.Size = uint64(len(impl.target)) + stat.Blocks = allocatedBlocksForSize(stat.Size) + case *namedPipe: + stat.Mode |= linux.S_IFIFO + default: + panic(fmt.Sprintf("unknown inode type: %T", i.impl)) + } +} + +// 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 i.impl.(type) { + case *regularFile: + return linux.DT_REG + case *directory: + return linux.DT_DIR + case *symlink: + return linux.DT_LNK + default: + panic(fmt.Sprintf("unknown inode type: %T", i.impl)) + } +} + +// fileDescription is embedded by tmpfs implementations of +// vfs.FileDescriptionImpl. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl +} + +func (fd *fileDescription) filesystem() *filesystem { + return fd.vfsfd.Mount().Filesystem().Impl().(*filesystem) +} + +func (fd *fileDescription) inode() *inode { + return fd.vfsfd.Dentry().Impl().(*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 { + if opts.Stat.Mask == 0 { + return nil + } + // TODO: implement inode.setStat + return syserror.EPERM +} -- cgit v1.2.3