Files
buildkit/executor/runcexecutor/executor.go
Akihiro Suda c54f4a986d support --oci-worker-no-process-sandbox
Note that this mode allows build executor containers to kill (and potentially ptrace) an arbitrary process in the BuildKit host namespace.
This mode should be enabled only when the BuildKit is running in a container as an unprivileged user.

Signed-off-by: Akihiro Suda <suda.akihiro@lab.ntt.co.jp>
2019-01-08 10:42:52 +09:00

328 lines
7.8 KiB
Go

package runcexecutor
import (
"context"
"encoding/json"
"io"
"io/ioutil"
"os"
"os/exec"
"path/filepath"
"strconv"
"strings"
"syscall"
"time"
"github.com/containerd/containerd/contrib/seccomp"
"github.com/containerd/containerd/mount"
containerdoci "github.com/containerd/containerd/oci"
"github.com/containerd/continuity/fs"
runc "github.com/containerd/go-runc"
"github.com/moby/buildkit/cache"
"github.com/moby/buildkit/executor"
"github.com/moby/buildkit/executor/oci"
"github.com/moby/buildkit/identity"
"github.com/moby/buildkit/solver/pb"
"github.com/moby/buildkit/util/network"
rootlessspecconv "github.com/moby/buildkit/util/rootless/specconv"
"github.com/moby/buildkit/util/system"
specs "github.com/opencontainers/runtime-spec/specs-go"
"github.com/pkg/errors"
"github.com/sirupsen/logrus"
)
type Opt struct {
// root directory
Root string
CommandCandidates []string
// without root privileges (has nothing to do with Opt.Root directory)
Rootless bool
// DefaultCgroupParent is the cgroup-parent name for executor
DefaultCgroupParent string
// ProcessMode
ProcessMode oci.ProcessMode
}
var defaultCommandCandidates = []string{"buildkit-runc", "runc"}
type runcExecutor struct {
runc *runc.Runc
root string
cmd string
cgroupParent string
rootless bool
networkProviders map[pb.NetMode]network.Provider
processMode oci.ProcessMode
}
func New(opt Opt, networkProviders map[pb.NetMode]network.Provider) (executor.Executor, error) {
cmds := opt.CommandCandidates
if cmds == nil {
cmds = defaultCommandCandidates
}
var cmd string
var found bool
for _, cmd = range cmds {
if _, err := exec.LookPath(cmd); err == nil {
found = true
break
}
}
if !found {
return nil, errors.Errorf("failed to find %s binary", cmd)
}
root := opt.Root
if err := os.MkdirAll(root, 0700); err != nil {
return nil, errors.Wrapf(err, "failed to create %s", root)
}
root, err := filepath.Abs(root)
if err != nil {
return nil, err
}
root, err = filepath.EvalSymlinks(root)
if err != nil {
return nil, err
}
// clean up old hosts/resolv.conf file. ignore errors
os.RemoveAll(filepath.Join(root, "hosts"))
os.RemoveAll(filepath.Join(root, "resolv.conf"))
runtime := &runc.Runc{
Command: cmd,
Log: filepath.Join(root, "runc-log.json"),
LogFormat: runc.JSON,
PdeathSignal: syscall.SIGKILL, // this can still leak the process
Setpgid: true,
// we don't execute runc with --rootless=(true|false) explicitly,
// so as to support non-runc runtimes
}
w := &runcExecutor{
runc: runtime,
root: root,
cgroupParent: opt.DefaultCgroupParent,
rootless: opt.Rootless,
networkProviders: networkProviders,
processMode: opt.ProcessMode,
}
return w, nil
}
func (w *runcExecutor) Exec(ctx context.Context, meta executor.Meta, root cache.Mountable, mounts []executor.Mount, stdin io.ReadCloser, stdout, stderr io.WriteCloser) error {
provider, ok := w.networkProviders[meta.NetMode]
if !ok {
return errors.Errorf("unknown network mode %s", meta.NetMode)
}
namespace, err := provider.New()
if err != nil {
return err
}
defer namespace.Close()
if meta.NetMode == pb.NetMode_HOST {
logrus.Info("enabling HostNetworking")
}
resolvConf, err := oci.GetResolvConf(ctx, w.root)
if err != nil {
return err
}
hostsFile, clean, err := oci.GetHostsFile(ctx, w.root, meta.ExtraHosts)
if err != nil {
return err
}
if clean != nil {
defer clean()
}
mountable, err := root.Mount(ctx, false)
if err != nil {
return err
}
rootMount, err := mountable.Mount()
if err != nil {
return err
}
defer mountable.Release()
id := identity.NewID()
bundle := filepath.Join(w.root, id)
if err := os.Mkdir(bundle, 0700); err != nil {
return err
}
defer os.RemoveAll(bundle)
rootFSPath := filepath.Join(bundle, "rootfs")
if err := os.Mkdir(rootFSPath, 0700); err != nil {
return err
}
if err := mount.All(rootMount, rootFSPath); err != nil {
return err
}
defer mount.Unmount(rootFSPath, 0)
uid, gid, sgids, err := oci.GetUser(ctx, rootFSPath, meta.User)
if err != nil {
return err
}
f, err := os.Create(filepath.Join(bundle, "config.json"))
if err != nil {
return err
}
defer f.Close()
opts := []containerdoci.SpecOpts{oci.WithUIDGID(uid, gid, sgids)}
if system.SeccompSupported() {
opts = append(opts, seccomp.WithDefaultProfile())
}
if meta.ReadonlyRootFS {
opts = append(opts, containerdoci.WithRootFSReadonly())
}
if w.cgroupParent != "" {
var cgroupsPath string
lastSeparator := w.cgroupParent[len(w.cgroupParent)-1:]
if strings.Contains(w.cgroupParent, ".slice") && lastSeparator == ":" {
cgroupsPath = w.cgroupParent + id
} else {
cgroupsPath = filepath.Join("/", w.cgroupParent, "buildkit", id)
}
opts = append(opts, containerdoci.WithCgroup(cgroupsPath))
}
spec, cleanup, err := oci.GenerateSpec(ctx, meta, mounts, id, resolvConf, hostsFile, namespace, w.processMode, opts...)
if err != nil {
return err
}
defer cleanup()
spec.Root.Path = rootFSPath
if _, ok := root.(cache.ImmutableRef); ok { // TODO: pass in with mount, not ref type
spec.Root.Readonly = true
}
newp, err := fs.RootPath(rootFSPath, meta.Cwd)
if err != nil {
return errors.Wrapf(err, "working dir %s points to invalid target", newp)
}
if err := os.MkdirAll(newp, 0755); err != nil {
return errors.Wrapf(err, "failed to create working directory %s", newp)
}
if err := setOOMScoreAdj(spec); err != nil {
return err
}
if w.rootless {
if err := rootlessspecconv.ToRootless(spec); err != nil {
return err
}
}
if err := json.NewEncoder(f).Encode(spec); err != nil {
return err
}
// runCtx/killCtx is used for extra check in case the kill command blocks
runCtx, cancelRun := context.WithCancel(context.Background())
defer cancelRun()
done := make(chan struct{})
go func() {
for {
select {
case <-ctx.Done():
killCtx, timeout := context.WithTimeout(context.Background(), 7*time.Second)
if err := w.runc.Kill(killCtx, id, int(syscall.SIGKILL), nil); err != nil {
logrus.Errorf("failed to kill runc %s: %+v", id, err)
select {
case <-killCtx.Done():
timeout()
cancelRun()
return
default:
}
}
timeout()
select {
case <-time.After(50 * time.Millisecond):
case <-done:
return
}
case <-done:
return
}
}
}()
logrus.Debugf("> creating %s %v", id, meta.Args)
status, err := w.runc.Run(runCtx, id, bundle, &runc.CreateOpts{
IO: &forwardIO{stdin: stdin, stdout: stdout, stderr: stderr},
})
close(done)
if err != nil {
return err
}
if status != 0 {
err := errors.Errorf("exit code: %d", status)
select {
case <-ctx.Done():
return errors.Wrapf(ctx.Err(), err.Error())
default:
return err
}
}
return nil
}
type forwardIO struct {
stdin io.ReadCloser
stdout, stderr io.WriteCloser
}
func (s *forwardIO) Close() error {
return nil
}
func (s *forwardIO) Set(cmd *exec.Cmd) {
cmd.Stdin = s.stdin
cmd.Stdout = s.stdout
cmd.Stderr = s.stderr
}
func (s *forwardIO) Stdin() io.WriteCloser {
return nil
}
func (s *forwardIO) Stdout() io.ReadCloser {
return nil
}
func (s *forwardIO) Stderr() io.ReadCloser {
return nil
}
// setOOMScoreAdj comes from https://github.com/genuinetools/img/blob/2fabe60b7dc4623aa392b515e013bbc69ad510ab/executor/runc/executor.go#L182-L192
func setOOMScoreAdj(spec *specs.Spec) error {
// Set the oom_score_adj of our children containers to that of the current process.
b, err := ioutil.ReadFile("/proc/self/oom_score_adj")
if err != nil {
return errors.Wrap(err, "failed to read /proc/self/oom_score_adj")
}
s := strings.TrimSpace(string(b))
oom, err := strconv.Atoi(s)
if err != nil {
return errors.Wrapf(err, "failed to parse %s as int", s)
}
spec.Process.OOMScoreAdj = &oom
return nil
}