Skip to content

Commit

Permalink
core/mount: use ptrace instead of go:linkname
Browse files Browse the repository at this point in the history 
The Go runtime has started to [lock down future uses of linkname][1] since
go1.23. In the go source code, containerd project has been marked in the
comment, [hall of shame][2]. Well, the go:linkname is used to fork no-op
subprocess efficiently. However, since that comment, I would like to use
ptrace and remove go:linkname in the whole repository.

With go1.22 `go:linkname`:

```bash
$ go test -bench=.  -benchmem ./ -exec sudo
goos: linux
goarch: amd64
pkg: github.com/containerd/containerd/v2/core/mount
cpu: AMD Ryzen 7 5800H with Radeon Graphics
BenchmarkBatchRunGetUsernsFD_Concurrent1-16                 2440            533320 ns/op            1145 B/op         43 allocs/op
BenchmarkBatchRunGetUsernsFD_Concurrent10-16                 342           3661616 ns/op           11562 B/op        421 allocs/op
PASS
ok      github.com/containerd/containerd/v2/core/mount  2.983s
```

With go1.22 `ptrace`:

```bash
$ go test -bench=.  -benchmem ./ -exec sudo
goos: linux
goarch: amd64
pkg: github.com/containerd/containerd/v2/core/mount
cpu: AMD Ryzen 7 5800H with Radeon Graphics
BenchmarkBatchRunGetUsernsFD_Concurrent1-16                 1785            739557 ns/op            3948 B/op         68 allocs/op
BenchmarkBatchRunGetUsernsFD_Concurrent10-16                 328           4024300 ns/op           39601 B/op        671 allocs/op
PASS
ok      github.com/containerd/containerd/v2/core/mount  3.104s
```

With go1.23 `ptrace`:

```bash
$ go test -bench=.  -benchmem ./ -exec sudo
goos: linux
goarch: amd64
pkg: github.com/containerd/containerd/v2/core/mount
cpu: AMD Ryzen 7 5800H with Radeon Graphics
BenchmarkBatchRunGetUsernsFD_Concurrent1-16                 1815            723252 ns/op            4220 B/op         69 allocs/op
BenchmarkBatchRunGetUsernsFD_Concurrent10-16                 319           3957157 ns/op           42351 B/op        682 allocs/op
PASS
ok      github.com/containerd/containerd/v2/core/mount  3.051s
```

Diff:

The `ptrace` is slower than `go:linkname` mode. However, it's accepctable.

```
goos: linux
goarch: amd64
pkg: github.com/containerd/containerd/v2/core/mount
cpu: AMD Ryzen 7 5800H with Radeon Graphics
                                    │ go122-golinkname │             go122-ptrace              │             go123-ptrace              │
                                    │      sec/op      │    sec/op     vs base                 │    sec/op     vs base                 │
BatchRunGetUsernsFD_Concurrent1-16        533.3µ ± ∞ ¹   739.6µ ± ∞ ¹        ~ (p=1.000 n=1) ²   723.3µ ± ∞ ¹        ~ (p=1.000 n=1) ²
BatchRunGetUsernsFD_Concurrent10-16       3.662m ± ∞ ¹   4.024m ± ∞ ¹        ~ (p=1.000 n=1) ²   3.957m ± ∞ ¹        ~ (p=1.000 n=1) ²
geomean                                   1.397m         1.725m        +23.45%                   1.692m        +21.06%
¹ need >= 6 samples for confidence interval at level 0.95
² need >= 4 samples to detect a difference at alpha level 0.05

                                    │ go122-golinkname │              go122-ptrace               │              go123-ptrace               │
                                    │       B/op       │     B/op       vs base                  │     B/op       vs base                  │
BatchRunGetUsernsFD_Concurrent1-16       1.118Ki ± ∞ ¹   3.855Ki ± ∞ ¹         ~ (p=1.000 n=1) ²   4.121Ki ± ∞ ¹         ~ (p=1.000 n=1) ²
BatchRunGetUsernsFD_Concurrent10-16      11.29Ki ± ∞ ¹   38.67Ki ± ∞ ¹         ~ (p=1.000 n=1) ²   41.36Ki ± ∞ ¹         ~ (p=1.000 n=1) ²
geomean                                  3.553Ki         12.21Ki        +243.65%                   13.06Ki        +267.43%
¹ need >= 6 samples for confidence interval at level 0.95
² need >= 4 samples to detect a difference at alpha level 0.05

                                    │ go122-golinkname │             go122-ptrace             │             go123-ptrace             │
                                    │    allocs/op     │  allocs/op   vs base                 │  allocs/op   vs base                 │
BatchRunGetUsernsFD_Concurrent1-16         43.00 ± ∞ ¹   68.00 ± ∞ ¹        ~ (p=1.000 n=1) ²   69.00 ± ∞ ¹        ~ (p=1.000 n=1) ²
BatchRunGetUsernsFD_Concurrent10-16        421.0 ± ∞ ¹   671.0 ± ∞ ¹        ~ (p=1.000 n=1) ²   682.0 ± ∞ ¹        ~ (p=1.000 n=1) ²
geomean                                    134.5         213.6        +58.76%                   216.9        +61.23%
¹ need >= 6 samples for confidence interval at level 0.95
² need >= 4 samples to detect a difference at alpha level 0.05
```

[1]: <golang/go#67401>
[2]: <https://github.com/golang/go/blob/release-branch.go1.23/src/runtime/proc.go#L4820>

Signed-off-by: Wei Fu <fuweid89@gmail.com>
  • Loading branch information
@fuweid
fuweid committed Aug 24, 2024
1 parent bcdf507 commit 9c91b2d
Show file tree
Hide file tree
Showing 7 changed files with 272 additions and 332 deletions.
100 changes: 0 additions & 100 deletions core/mount/mount_idmapped_linux.go
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,15 +19,11 @@ package mount
import (
"fmt"
"os"
"runtime"
"strconv"
"strings"
"sync"
"syscall"

"golang.org/x/sys/unix"

"github.com/containerd/containerd/v2/pkg/sys"
)

// TODO: Support multiple mappings in future
Expand Down Expand Up @@ -106,99 +102,3 @@ func GetUsernsFD(uidmap, gidmap string) (_usernsFD *os.File, _ error) {
}
return getUsernsFD(uidMaps, gidMaps)
}

func getUsernsFD(uidMaps, gidMaps []syscall.SysProcIDMap) (_usernsFD *os.File, retErr error) {
runtime.LockOSThread()
defer runtime.UnlockOSThread()

pid, pidfd, errno := sys.ForkUserns()
if errno != 0 {
return nil, errno
}

pidFD := os.NewFile(pidfd, "pidfd")
defer func() {
unix.PidfdSendSignal(int(pidFD.Fd()), unix.SIGKILL, nil, 0)

pidfdWaitid(pidFD)

pidFD.Close()
}()

// NOTE:
//
// The usernsFD will hold the userns reference in kernel. Even if the
// child process is reaped, the usernsFD is still valid.
usernsFD, err := os.Open(fmt.Sprintf("/proc/%d/ns/user", pid))
if err != nil {
return nil, fmt.Errorf("failed to get userns file descriptor for /proc/%d/user/ns: %w", pid, err)
}
defer func() {
if retErr != nil {
usernsFD.Close()
}
}()

uidmapFile, err := os.OpenFile(fmt.Sprintf("/proc/%d/%s", pid, "uid_map"), os.O_WRONLY, 0600)
if err != nil {
return nil, fmt.Errorf("failed to open /proc/%d/uid_map: %w", pid, err)
}
defer uidmapFile.Close()

gidmapFile, err := os.OpenFile(fmt.Sprintf("/proc/%d/%s", pid, "gid_map"), os.O_WRONLY, 0600)
if err != nil {
return nil, fmt.Errorf("failed to open /proc/%d/gid_map: %w", pid, err)
}
defer gidmapFile.Close()

testHookKillChildBeforePidfdSendSignal(pid, pidFD)

// Ensure the child process is still alive. If the err is ESRCH, we
// should return error because we can't guarantee the usernsFD and
// u[g]idmapFile are valid. It's safe to return error and retry.
if err := unix.PidfdSendSignal(int(pidFD.Fd()), 0, nil, 0); err != nil {
return nil, fmt.Errorf("failed to ensure child process is alive: %w", err)
}

testHookKillChildAfterPidfdSendSignal(pid, pidFD)

// NOTE:
//
// The u[g]id_map file descriptor is still valid if the child process
// is reaped.
writeMappings := func(f *os.File, idmap []syscall.SysProcIDMap) error {
mappings := ""
for _, m := range idmap {
mappings = fmt.Sprintf("%s%d %d %d\n", mappings, m.ContainerID, m.HostID, m.Size)
}

_, err := f.Write([]byte(mappings))
if err1 := f.Close(); err1 != nil && err == nil {
err = err1
}
return err
}

if err := writeMappings(uidmapFile, uidMaps); err != nil {
return nil, fmt.Errorf("failed to write uid_map: %w", err)
}

if err := writeMappings(gidmapFile, gidMaps); err != nil {
return nil, fmt.Errorf("failed to write gid_map: %w", err)
}
return usernsFD, nil
}

func pidfdWaitid(pidFD *os.File) error {
return sys.IgnoringEINTR(func() error {
return unix.Waitid(unix.P_PIDFD, int(pidFD.Fd()), nil, unix.WEXITED, nil)
})
}

var (
testHookLock sync.Mutex

testHookKillChildBeforePidfdSendSignal = func(_pid uintptr, _pidFD *os.File) {}

testHookKillChildAfterPidfdSendSignal = func(_pid uintptr, _pidFD *os.File) {}
)
125 changes: 17 additions & 108 deletions core/mount/mount_idmapped_linux_test.go
View file
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,6 @@ import (
kernel "github.com/containerd/containerd/v2/pkg/kernelversion"
"github.com/containerd/continuity/testutil"
"github.com/stretchr/testify/require"
"golang.org/x/sys/unix"
)

func BenchmarkBatchRunGetUsernsFD_Concurrent1(b *testing.B) {
Expand Down Expand Up @@ -72,7 +71,7 @@ var (
}
)

func TestGetUsernsFD(t *testing.T) {
func TestIdmappedMount(t *testing.T) {
testutil.RequiresRoot(t)

k512 := kernel.KernelVersion{Kernel: 5, Major: 12}
Expand All @@ -82,15 +81,12 @@ func TestGetUsernsFD(t *testing.T) {
t.Skip("GetUsernsFD requires kernel >= 5.12")
}

t.Run("basic", testGetUsernsFDBasic)

t.Run("when kill child process before write u[g]id maps", testGetUsernsFDKillChildWhenWriteUGIDMaps)

t.Run("when kill child process after open u[g]id_map file", testGetUsernsFDKillChildAfterOpenUGIDMapFiles)
t.Run("GetUsernsFD", testGetUsernsFD)

t.Run("IDMapMount", testIDMapMount)
}

func testGetUsernsFDBasic(t *testing.T) {
func testGetUsernsFD(t *testing.T) {
for idx, tc := range []struct {
uidMaps string
gidMaps string
Expand Down Expand Up @@ -128,108 +124,21 @@ func testGetUsernsFDBasic(t *testing.T) {
}
}

func testGetUsernsFDKillChildWhenWriteUGIDMaps(t *testing.T) {
hookFunc := func(reap bool) func(uintptr, *os.File) {
return func(_pid uintptr, pidFD *os.File) {
err := unix.PidfdSendSignal(int(pidFD.Fd()), unix.SIGKILL, nil, 0)
require.NoError(t, err)

if reap {
pidfdWaitid(pidFD)
}
}
}

for _, tcReap := range []bool{true, false} {
t.Run(fmt.Sprintf("#reap=%v", tcReap), func(t *testing.T) {
updateTestHookKillForGetUsernsFD(t, nil, hookFunc(tcReap))

usernsFD, err := getUsernsFD(testUIDMaps, testGIDMaps)
require.NoError(t, err)
defer usernsFD.Close()

srcDir, checkFunc := initIDMappedChecker(t, testUIDMaps, testGIDMaps)
destDir := t.TempDir()
defer func() {
require.NoError(t, UnmountAll(destDir, 0))
}()

err = IDMapMount(srcDir, destDir, int(usernsFD.Fd()))
usernsFD.Close()
require.NoError(t, err)
checkFunc(destDir)
})
}

}

func testGetUsernsFDKillChildAfterOpenUGIDMapFiles(t *testing.T) {
hookFunc := func(reap bool) func(uintptr, *os.File) {
return func(_pid uintptr, pidFD *os.File) {
err := unix.PidfdSendSignal(int(pidFD.Fd()), unix.SIGKILL, nil, 0)
require.NoError(t, err)

if reap {
pidfdWaitid(pidFD)
}
}
}

for _, tc := range []struct {
reap bool
expected error
}{
{
reap: false,
expected: nil,
},
{
reap: true,
expected: syscall.ESRCH,
},
} {
t.Run(fmt.Sprintf("#reap=%v", tc.reap), func(t *testing.T) {
updateTestHookKillForGetUsernsFD(t, hookFunc(tc.reap), nil)

usernsFD, err := getUsernsFD(testUIDMaps, testGIDMaps)
if tc.expected != nil {
require.Error(t, tc.expected, err)
return
}

require.NoError(t, err)
defer usernsFD.Close()

srcDir, checkFunc := initIDMappedChecker(t, testUIDMaps, testGIDMaps)
destDir := t.TempDir()
defer func() {
require.NoError(t, UnmountAll(destDir, 0))
}()

err = IDMapMount(srcDir, destDir, int(usernsFD.Fd()))
usernsFD.Close()
require.NoError(t, err)
checkFunc(destDir)
})
}
}
func testIDMapMount(t *testing.T) {
usernsFD, err := getUsernsFD(testUIDMaps, testGIDMaps)
require.NoError(t, err)
defer usernsFD.Close()

func updateTestHookKillForGetUsernsFD(t *testing.T, newBeforeFunc, newAfterFunc func(uintptr, *os.File)) {
testHookLock.Lock()
srcDir, checkFunc := initIDMappedChecker(t, testUIDMaps, testGIDMaps)
destDir := t.TempDir()
defer func() {
require.NoError(t, UnmountAll(destDir, 0))
}()

oldBefore := testHookKillChildBeforePidfdSendSignal
oldAfter := testHookKillChildAfterPidfdSendSignal
t.Cleanup(func() {
testHookKillChildBeforePidfdSendSignal = oldBefore
testHookKillChildAfterPidfdSendSignal = oldAfter
testHookLock.Unlock()
})
if newBeforeFunc != nil {
testHookKillChildBeforePidfdSendSignal = newBeforeFunc
}
if newAfterFunc != nil {
testHookKillChildAfterPidfdSendSignal = newAfterFunc
}
err = IDMapMount(srcDir, destDir, int(usernsFD.Fd()))
usernsFD.Close()
require.NoError(t, err)
checkFunc(destDir)
}

func initIDMappedChecker(t *testing.T, uidMaps, gidMaps []syscall.SysProcIDMap) (_srcDir string, _verifyFunc func(destDir string)) {
Expand Down
81 changes: 81 additions & 0 deletions core/mount/mount_idmapped_utils_linux.go
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,81 @@
//go:build go1.23 && linux

/*
Copyright The containerd 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 mount

import (
"fmt"
"os"
"syscall"

"github.com/containerd/containerd/v2/pkg/sys"

"golang.org/x/sys/unix"
)

// getUsernsFD returns pinnable user namespace's file descriptor.
//
// NOTE: The GO runtime uses pidfd to handle subprocess since go1.23. However,
// it has double close issue tracked by [1]. We can't use pidfd directly and
// the GO runtime doesn't export interface to show if it's using pidfd or not.
// So, we call `sys.SupportsPidFD` first and then use `os.Process` directly.
//
// [1]: https://github.com/golang/go/issues/68984
func getUsernsFD(uidMaps, gidMaps []syscall.SysProcIDMap) (*os.File, error) {
if !sys.SupportsPidFD() {
return nil, fmt.Errorf("failed to prevent pid reused issue because pidfd isn't supported")
}

proc, err := os.StartProcess("/proc/self/exe", []string{"containerd[getUsernsFD]"}, &os.ProcAttr{
Sys: &syscall.SysProcAttr{
Cloneflags: unix.CLONE_NEWUSER,
UidMappings: uidMaps,
GidMappings: gidMaps,
// NOTE: It's reexec but it's not heavy because subprocess
// be in PTRACE_TRACEME mode before performing execve.
Ptrace: true,
Pdeathsig: syscall.SIGKILL,
},
})
if err != nil {
return nil, fmt.Errorf("failed to start noop process for unshare: %w", err)
}

defer func() {
proc.Kill()
proc.Wait()
}()

// NOTE:
//
// The usernsFD will hold the userns reference in kernel. Even if the
// child process is reaped, the usernsFD is still valid.
usernsFD, err := os.Open(fmt.Sprintf("/proc/%d/ns/user", proc.Pid))
if err != nil {
return nil, fmt.Errorf("failed to get userns file descriptor for /proc/%d/user/ns: %w", proc.Pid, err)
}

// Ensure the child process is still alive. If the err is ESRCH, we
// should return error because we can't guarantee the usernsFD and
// u[g]idmapFile are valid. It's safe to return error and retry.
if err := proc.Signal(syscall.Signal(0)); err != nil {
usernsFD.Close()
return nil, fmt.Errorf("failed to ensure child process is alive: %w", err)
}
return usernsFD, nil
}
Loading

0 comments on commit 9c91b2d

Please sign in to comment.