Merge gvisor-containerd-shim

pkg/shim/runsc/utils.go

@@ -0,0 +1,44 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 runsc
+
+import (
+	"bytes"
+	"strings"
+	"sync"
+)
+
+var bytesBufferPool = sync.Pool{
+	New: func() interface{} {
+		return bytes.NewBuffer(nil)
+	},
+}
+
+func getBuf() *bytes.Buffer {
+	return bytesBufferPool.Get().(*bytes.Buffer)
+}
+
+func putBuf(b *bytes.Buffer) {
+	b.Reset()
+	bytesBufferPool.Put(b)
+}
+
+// FormatLogPath parses runsc config, and fill in %ID% in the log path.
+func FormatLogPath(id string, config map[string]string) {
+	if path, ok := config["debug-log"]; ok {
+		config["debug-log"] = strings.Replace(path, "%ID%", id, -1)
+	}
+}

pkg/shim/v1/proc/deleted_state.go

@@ -0,0 +1,49 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"context"
+
+	"github.com/containerd/console"
+	"github.com/containerd/containerd/errdefs"
+	"github.com/containerd/containerd/runtime/proc"
+	"github.com/pkg/errors"
+)
+
+type deletedState struct{}
+
+func (*deletedState) Resize(ws console.WinSize) error {
+	return errors.Errorf("cannot resize a deleted process")
+}
+
+func (*deletedState) Start(ctx context.Context) error {
+	return errors.Errorf("cannot start a deleted process")
+}
+
+func (*deletedState) Delete(ctx context.Context) error {
+	return errors.Wrap(errdefs.ErrNotFound, "cannot delete a deleted process")
+}
+
+func (*deletedState) Kill(ctx context.Context, sig uint32, all bool) error {
+	return errors.Wrap(errdefs.ErrNotFound, "cannot kill a deleted process")
+}
+
+func (*deletedState) SetExited(status int) {}
+
+func (*deletedState) Exec(ctx context.Context, path string, r *ExecConfig) (proc.Process, error) {
+	return nil, errors.Errorf("cannot exec in a deleted state")
+}

pkg/shim/v1/proc/exec.go

@@ -0,0 +1,282 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"sync"
+	"syscall"
+	"time"
+
+	"github.com/containerd/console"
+	"github.com/containerd/containerd/errdefs"
+	"github.com/containerd/containerd/runtime/proc"
+	"github.com/containerd/fifo"
+	runc "github.com/containerd/go-runc"
+	specs "github.com/opencontainers/runtime-spec/specs-go"
+	"github.com/pkg/errors"
+	"golang.org/x/sys/unix"
+
+	runsc "github.com/google/gvisor-containerd-shim/pkg/go-runsc"
+)
+
+type execProcess struct {
+	wg sync.WaitGroup
+
+	execState execState
+
+	mu          sync.Mutex
+	id          string
+	console     console.Console
+	io          runc.IO
+	status      int
+	exited      time.Time
+	pid         int
+	internalPid int
+	closers     []io.Closer
+	stdin       io.Closer
+	stdio       proc.Stdio
+	path        string
+	spec        specs.Process
+
+	parent    *Init
+	waitBlock chan struct{}
+}
+
+func (e *execProcess) Wait() {
+	<-e.waitBlock
+}
+
+func (e *execProcess) ID() string {
+	return e.id
+}
+
+func (e *execProcess) Pid() int {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.pid
+}
+
+func (e *execProcess) ExitStatus() int {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.status
+}
+
+func (e *execProcess) ExitedAt() time.Time {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.exited
+}
+
+func (e *execProcess) SetExited(status int) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	e.execState.SetExited(status)
+}
+
+func (e *execProcess) setExited(status int) {
+	e.status = status
+	e.exited = time.Now()
+	e.parent.Platform.ShutdownConsole(context.Background(), e.console)
+	close(e.waitBlock)
+}
+
+func (e *execProcess) Delete(ctx context.Context) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	return e.execState.Delete(ctx)
+}
+
+func (e *execProcess) delete(ctx context.Context) error {
+	e.wg.Wait()
+	if e.io != nil {
+		for _, c := range e.closers {
+			c.Close()
+		}
+		e.io.Close()
+	}
+	pidfile := filepath.Join(e.path, fmt.Sprintf("%s.pid", e.id))
+	// silently ignore error
+	os.Remove(pidfile)
+	internalPidfile := filepath.Join(e.path, fmt.Sprintf("%s-internal.pid", e.id))
+	// silently ignore error
+	os.Remove(internalPidfile)
+	return nil
+}
+
+func (e *execProcess) Resize(ws console.WinSize) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	return e.execState.Resize(ws)
+}
+
+func (e *execProcess) resize(ws console.WinSize) error {
+	if e.console == nil {
+		return nil
+	}
+	return e.console.Resize(ws)
+}
+
+func (e *execProcess) Kill(ctx context.Context, sig uint32, _ bool) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	return e.execState.Kill(ctx, sig, false)
+}
+
+func (e *execProcess) kill(ctx context.Context, sig uint32, _ bool) error {
+	internalPid := e.internalPid
+	if internalPid != 0 {
+		if err := e.parent.runtime.Kill(ctx, e.parent.id, int(sig), &runsc.KillOpts{
+			Pid: internalPid,
+		}); err != nil {
+			// If this returns error, consider the process has already stopped.
+			// TODO: Fix after signal handling is fixed.
+			return errors.Wrapf(errdefs.ErrNotFound, err.Error())
+		}
+	}
+	return nil
+}
+
+func (e *execProcess) Stdin() io.Closer {
+	return e.stdin
+}
+
+func (e *execProcess) Stdio() proc.Stdio {
+	return e.stdio
+}
+
+func (e *execProcess) Start(ctx context.Context) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	return e.execState.Start(ctx)
+}
+
+func (e *execProcess) start(ctx context.Context) (err error) {
+	var (
+		socket          *runc.Socket
+		pidfile         = filepath.Join(e.path, fmt.Sprintf("%s.pid", e.id))
+		internalPidfile = filepath.Join(e.path, fmt.Sprintf("%s-internal.pid", e.id))
+	)
+	if e.stdio.Terminal {
+		if socket, err = runc.NewTempConsoleSocket(); err != nil {
+			return errors.Wrap(err, "failed to create runc console socket")
+		}
+		defer socket.Close()
+	} else if e.stdio.IsNull() {
+		if e.io, err = runc.NewNullIO(); err != nil {
+			return errors.Wrap(err, "creating new NULL IO")
+		}
+	} else {
+		if e.io, err = runc.NewPipeIO(e.parent.IoUID, e.parent.IoGID, withConditionalIO(e.stdio)); err != nil {
+			return errors.Wrap(err, "failed to create runc io pipes")
+		}
+	}
+	opts := &runsc.ExecOpts{
+		PidFile:         pidfile,
+		InternalPidFile: internalPidfile,
+		IO:              e.io,
+		Detach:          true,
+	}
+	if socket != nil {
+		opts.ConsoleSocket = socket
+	}
+	eventCh := e.parent.Monitor.Subscribe()
+	defer func() {
+		// Unsubscribe if an error is returned.
+		if err != nil {
+			e.parent.Monitor.Unsubscribe(eventCh)
+		}
+	}()
+	if err := e.parent.runtime.Exec(ctx, e.parent.id, e.spec, opts); err != nil {
+		close(e.waitBlock)
+		return e.parent.runtimeError(err, "OCI runtime exec failed")
+	}
+	if e.stdio.Stdin != "" {
+		sc, err := fifo.OpenFifo(context.Background(), e.stdio.Stdin, syscall.O_WRONLY|syscall.O_NONBLOCK, 0)
+		if err != nil {
+			return errors.Wrapf(err, "failed to open stdin fifo %s", e.stdio.Stdin)
+		}
+		e.closers = append(e.closers, sc)
+		e.stdin = sc
+	}
+	var copyWaitGroup sync.WaitGroup
+	ctx, cancel := context.WithTimeout(ctx, 30*time.Second)
+	defer cancel()
+	if socket != nil {
+		console, err := socket.ReceiveMaster()
+		if err != nil {
+			return errors.Wrap(err, "failed to retrieve console master")
+		}
+		if e.console, err = e.parent.Platform.CopyConsole(ctx, console, e.stdio.Stdin, e.stdio.Stdout, e.stdio.Stderr, &e.wg, &copyWaitGroup); err != nil {
+			return errors.Wrap(err, "failed to start console copy")
+		}
+	} else if !e.stdio.IsNull() {
+		if err := copyPipes(ctx, e.io, e.stdio.Stdin, e.stdio.Stdout, e.stdio.Stderr, &e.wg, &copyWaitGroup); err != nil {
+			return errors.Wrap(err, "failed to start io pipe copy")
+		}
+	}
+	copyWaitGroup.Wait()
+	pid, err := runc.ReadPidFile(opts.PidFile)
+	if err != nil {
+		return errors.Wrap(err, "failed to retrieve OCI runtime exec pid")
+	}
+	e.pid = pid
+	internalPid, err := runc.ReadPidFile(opts.InternalPidFile)
+	if err != nil {
+		return errors.Wrap(err, "failed to retrieve OCI runtime exec internal pid")
+	}
+	e.internalPid = internalPid
+	go func() {
+		defer e.parent.Monitor.Unsubscribe(eventCh)
+		for event := range eventCh {
+			if event.Pid == e.pid {
+				ExitCh <- Exit{
+					Timestamp: event.Timestamp,
+					ID:        e.id,
+					Status:    event.Status,
+				}
+				break
+			}
+		}
+	}()
+	return nil
+}
+
+func (e *execProcess) Status(ctx context.Context) (string, error) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	// if we don't have a pid then the exec process has just been created
+	if e.pid == 0 {
+		return "created", nil
+	}
+	// if we have a pid and it can be signaled, the process is running
+	// TODO(random-liu): Use `runsc kill --pid`.
+	if err := unix.Kill(e.pid, 0); err == nil {
+		return "running", nil
+	}
+	// else if we have a pid but it can nolonger be signaled, it has stopped
+	return "stopped", nil
+}

pkg/shim/v1/proc/exec_state.go

@@ -0,0 +1,154 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"context"
+
+	"github.com/containerd/console"
+	"github.com/pkg/errors"
+)
+
+type execState interface {
+	Resize(console.WinSize) error
+	Start(context.Context) error
+	Delete(context.Context) error
+	Kill(context.Context, uint32, bool) error
+	SetExited(int)
+}
+
+type execCreatedState struct {
+	p *execProcess
+}
+
+func (s *execCreatedState) transition(name string) error {
+	switch name {
+	case "running":
+		s.p.execState = &execRunningState{p: s.p}
+	case "stopped":
+		s.p.execState = &execStoppedState{p: s.p}
+	case "deleted":
+		s.p.execState = &deletedState{}
+	default:
+		return errors.Errorf("invalid state transition %q to %q", stateName(s), name)
+	}
+	return nil
+}
+
+func (s *execCreatedState) Resize(ws console.WinSize) error {
+	return s.p.resize(ws)
+}
+
+func (s *execCreatedState) Start(ctx context.Context) error {
+	if err := s.p.start(ctx); err != nil {
+		return err
+	}
+	return s.transition("running")
+}
+
+func (s *execCreatedState) Delete(ctx context.Context) error {
+	if err := s.p.delete(ctx); err != nil {
+		return err
+	}
+	return s.transition("deleted")
+}
+
+func (s *execCreatedState) Kill(ctx context.Context, sig uint32, all bool) error {
+	return s.p.kill(ctx, sig, all)
+}
+
+func (s *execCreatedState) SetExited(status int) {
+	s.p.setExited(status)
+
+	if err := s.transition("stopped"); err != nil {
+		panic(err)
+	}
+}
+
+type execRunningState struct {
+	p *execProcess
+}
+
+func (s *execRunningState) transition(name string) error {
+	switch name {
+	case "stopped":
+		s.p.execState = &execStoppedState{p: s.p}
+	default:
+		return errors.Errorf("invalid state transition %q to %q", stateName(s), name)
+	}
+	return nil
+}
+
+func (s *execRunningState) Resize(ws console.WinSize) error {
+	return s.p.resize(ws)
+}
+
+func (s *execRunningState) Start(ctx context.Context) error {
+	return errors.Errorf("cannot start a running process")
+}
+
+func (s *execRunningState) Delete(ctx context.Context) error {
+	return errors.Errorf("cannot delete a running process")
+}
+
+func (s *execRunningState) Kill(ctx context.Context, sig uint32, all bool) error {
+	return s.p.kill(ctx, sig, all)
+}
+
+func (s *execRunningState) SetExited(status int) {
+	s.p.setExited(status)
+
+	if err := s.transition("stopped"); err != nil {
+		panic(err)
+	}
+}
+
+type execStoppedState struct {
+	p *execProcess
+}
+
+func (s *execStoppedState) transition(name string) error {
+	switch name {
+	case "deleted":
+		s.p.execState = &deletedState{}
+	default:
+		return errors.Errorf("invalid state transition %q to %q", stateName(s), name)
+	}
+	return nil
+}
+
+func (s *execStoppedState) Resize(ws console.WinSize) error {
+	return errors.Errorf("cannot resize a stopped container")
+}
+
+func (s *execStoppedState) Start(ctx context.Context) error {
+	return errors.Errorf("cannot start a stopped process")
+}
+
+func (s *execStoppedState) Delete(ctx context.Context) error {
+	if err := s.p.delete(ctx); err != nil {
+		return err
+	}
+	return s.transition("deleted")
+}
+
+func (s *execStoppedState) Kill(ctx context.Context, sig uint32, all bool) error {
+	return s.p.kill(ctx, sig, all)
+}
+
+func (s *execStoppedState) SetExited(status int) {
+	// no op
+}

pkg/shim/v1/proc/init.go

@@ -0,0 +1,462 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"path/filepath"
+	"strings"
+	"sync"
+	"syscall"
+	"time"
+
+	"github.com/containerd/console"
+	"github.com/containerd/containerd/errdefs"
+	"github.com/containerd/containerd/log"
+	"github.com/containerd/containerd/mount"
+	"github.com/containerd/containerd/runtime/proc"
+	"github.com/containerd/fifo"
+	runc "github.com/containerd/go-runc"
+	specs "github.com/opencontainers/runtime-spec/specs-go"
+	"github.com/pkg/errors"
+
+	runsc "github.com/google/gvisor-containerd-shim/pkg/go-runsc"
+)
+
+// InitPidFile name of the file that contains the init pid
+const InitPidFile = "init.pid"
+
+// Init represents an initial process for a container
+type Init struct {
+	wg        sync.WaitGroup
+	initState initState
+
+	// mu is used to ensure that `Start()` and `Exited()` calls return in
+	// the right order when invoked in separate go routines.
+	// This is the case within the shim implementation as it makes use of
+	// the reaper interface.
+	mu sync.Mutex
+
+	waitBlock chan struct{}
+
+	WorkDir string
+
+	id       string
+	Bundle   string
+	console  console.Console
+	Platform proc.Platform
+	io       runc.IO
+	runtime  *runsc.Runsc
+	status   int
+	exited   time.Time
+	pid      int
+	closers  []io.Closer
+	stdin    io.Closer
+	stdio    proc.Stdio
+	Rootfs   string
+	IoUID    int
+	IoGID    int
+	Sandbox  bool
+	UserLog  string
+	Monitor  ProcessMonitor
+}
+
+// NewRunsc returns a new runsc instance for a process
+func NewRunsc(root, path, namespace, runtime string, config map[string]string) *runsc.Runsc {
+	if root == "" {
+		root = RunscRoot
+	}
+	return &runsc.Runsc{
+		Command:      runtime,
+		PdeathSignal: syscall.SIGKILL,
+		Log:          filepath.Join(path, "log.json"),
+		LogFormat:    runc.JSON,
+		Root:         filepath.Join(root, namespace),
+		Config:       config,
+	}
+}
+
+// New returns a new init process
+func New(id string, runtime *runsc.Runsc, stdio proc.Stdio) *Init {
+	p := &Init{
+		id:        id,
+		runtime:   runtime,
+		stdio:     stdio,
+		status:    0,
+		waitBlock: make(chan struct{}),
+	}
+	p.initState = &createdState{p: p}
+	return p
+}
+
+// Create the process with the provided config
+func (p *Init) Create(ctx context.Context, r *CreateConfig) (err error) {
+	var socket *runc.Socket
+	if r.Terminal {
+		if socket, err = runc.NewTempConsoleSocket(); err != nil {
+			return errors.Wrap(err, "failed to create OCI runtime console socket")
+		}
+		defer socket.Close()
+	} else if hasNoIO(r) {
+		if p.io, err = runc.NewNullIO(); err != nil {
+			return errors.Wrap(err, "creating new NULL IO")
+		}
+	} else {
+		if p.io, err = runc.NewPipeIO(p.IoUID, p.IoGID, withConditionalIO(p.stdio)); err != nil {
+			return errors.Wrap(err, "failed to create OCI runtime io pipes")
+		}
+	}
+	pidFile := filepath.Join(p.Bundle, InitPidFile)
+	opts := &runsc.CreateOpts{
+		PidFile: pidFile,
+	}
+	if socket != nil {
+		opts.ConsoleSocket = socket
+	}
+	if p.Sandbox {
+		opts.IO = p.io
+		// UserLog is only useful for sandbox.
+		opts.UserLog = p.UserLog
+	}
+	if err := p.runtime.Create(ctx, r.ID, r.Bundle, opts); err != nil {
+		return p.runtimeError(err, "OCI runtime create failed")
+	}
+	if r.Stdin != "" {
+		sc, err := fifo.OpenFifo(context.Background(), r.Stdin, syscall.O_WRONLY|syscall.O_NONBLOCK, 0)
+		if err != nil {
+			return errors.Wrapf(err, "failed to open stdin fifo %s", r.Stdin)
+		}
+		p.stdin = sc
+		p.closers = append(p.closers, sc)
+	}
+	var copyWaitGroup sync.WaitGroup
+	ctx, cancel := context.WithTimeout(ctx, 30*time.Second)
+	defer cancel()
+	if socket != nil {
+		console, err := socket.ReceiveMaster()
+		if err != nil {
+			return errors.Wrap(err, "failed to retrieve console master")
+		}
+		console, err = p.Platform.CopyConsole(ctx, console, r.Stdin, r.Stdout, r.Stderr, &p.wg, &copyWaitGroup)
+		if err != nil {
+			return errors.Wrap(err, "failed to start console copy")
+		}
+		p.console = console
+	} else if !hasNoIO(r) {
+		if err := copyPipes(ctx, p.io, r.Stdin, r.Stdout, r.Stderr, &p.wg, &copyWaitGroup); err != nil {
+			return errors.Wrap(err, "failed to start io pipe copy")
+		}
+	}
+
+	copyWaitGroup.Wait()
+	pid, err := runc.ReadPidFile(pidFile)
+	if err != nil {
+		return errors.Wrap(err, "failed to retrieve OCI runtime container pid")
+	}
+	p.pid = pid
+	return nil
+}
+
+// Wait for the process to exit
+func (p *Init) Wait() {
+	<-p.waitBlock
+}
+
+// ID of the process
+func (p *Init) ID() string {
+	return p.id
+}
+
+// Pid of the process
+func (p *Init) Pid() int {
+	return p.pid
+}
+
+// ExitStatus of the process
+func (p *Init) ExitStatus() int {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.status
+}
+
+// ExitedAt at time when the process exited
+func (p *Init) ExitedAt() time.Time {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.exited
+}
+
+// Status of the process
+func (p *Init) Status(ctx context.Context) (string, error) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	c, err := p.runtime.State(ctx, p.id)
+	if err != nil {
+		if strings.Contains(err.Error(), "does not exist") {
+			return "stopped", nil
+		}
+		return "", p.runtimeError(err, "OCI runtime state failed")
+	}
+	return p.convertStatus(c.Status), nil
+}
+
+// Start the init process
+func (p *Init) Start(ctx context.Context) error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	return p.initState.Start(ctx)
+}
+
+func (p *Init) start(ctx context.Context) error {
+	var cio runc.IO
+	if !p.Sandbox {
+		cio = p.io
+	}
+	if err := p.runtime.Start(ctx, p.id, cio); err != nil {
+		return p.runtimeError(err, "OCI runtime start failed")
+	}
+	go func() {
+		status, err := p.runtime.Wait(context.Background(), p.id)
+		if err != nil {
+			log.G(ctx).WithError(err).Errorf("Failed to wait for container %q", p.id)
+			// TODO(random-liu): Handle runsc kill error.
+			if err := p.killAll(ctx); err != nil {
+				log.G(ctx).WithError(err).Errorf("Failed to kill container %q", p.id)
+			}
+			status = internalErrorCode
+		}
+		ExitCh <- Exit{
+			Timestamp: time.Now(),
+			ID:        p.id,
+			Status:    status,
+		}
+	}()
+	return nil
+}
+
+// SetExited of the init process with the next status
+func (p *Init) SetExited(status int) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	p.initState.SetExited(status)
+}
+
+func (p *Init) setExited(status int) {
+	p.exited = time.Now()
+	p.status = status
+	p.Platform.ShutdownConsole(context.Background(), p.console)
+	close(p.waitBlock)
+}
+
+// Delete the init process
+func (p *Init) Delete(ctx context.Context) error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	return p.initState.Delete(ctx)
+}
+
+func (p *Init) delete(ctx context.Context) error {
+	p.killAll(ctx)
+	p.wg.Wait()
+	err := p.runtime.Delete(ctx, p.id, nil)
+	// ignore errors if a runtime has already deleted the process
+	// but we still hold metadata and pipes
+	//
+	// this is common during a checkpoint, runc will delete the container state
+	// after a checkpoint and the container will no longer exist within runc
+	if err != nil {
+		if strings.Contains(err.Error(), "does not exist") {
+			err = nil
+		} else {
+			err = p.runtimeError(err, "failed to delete task")
+		}
+	}
+	if p.io != nil {
+		for _, c := range p.closers {
+			c.Close()
+		}
+		p.io.Close()
+	}
+	if err2 := mount.UnmountAll(p.Rootfs, 0); err2 != nil {
+		log.G(ctx).WithError(err2).Warn("failed to cleanup rootfs mount")
+		if err == nil {
+			err = errors.Wrap(err2, "failed rootfs umount")
+		}
+	}
+	return err
+}
+
+// Resize the init processes console
+func (p *Init) Resize(ws console.WinSize) error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	if p.console == nil {
+		return nil
+	}
+	return p.console.Resize(ws)
+}
+
+func (p *Init) resize(ws console.WinSize) error {
+	if p.console == nil {
+		return nil
+	}
+	return p.console.Resize(ws)
+}
+
+// Kill the init process
+func (p *Init) Kill(ctx context.Context, signal uint32, all bool) error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	return p.initState.Kill(ctx, signal, all)
+}
+
+func (p *Init) kill(context context.Context, signal uint32, all bool) error {
+	var (
+		killErr error
+		backoff = 100 * time.Millisecond
+	)
+	timeout := 1 * time.Second
+	for start := time.Now(); time.Now().Sub(start) < timeout; {
+		c, err := p.runtime.State(context, p.id)
+		if err != nil {
+			if strings.Contains(err.Error(), "does not exist") {
+				return errors.Wrapf(errdefs.ErrNotFound, "no such process")
+			}
+			return p.runtimeError(err, "OCI runtime state failed")
+		}
+		// For runsc, signal only works when container is running state.
+		// If the container is not in running state, directly return
+		// "no such process"
+		if p.convertStatus(c.Status) == "stopped" {
+			return errors.Wrapf(errdefs.ErrNotFound, "no such process")
+		}
+		killErr = p.runtime.Kill(context, p.id, int(signal), &runsc.KillOpts{
+			All: all,
+		})
+		if killErr == nil {
+			return nil
+		}
+		time.Sleep(backoff)
+		backoff *= 2
+	}
+	return p.runtimeError(killErr, "kill timeout")
+}
+
+// KillAll processes belonging to the init process
+func (p *Init) KillAll(context context.Context) error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.killAll(context)
+}
+
+func (p *Init) killAll(context context.Context) error {
+	p.runtime.Kill(context, p.id, int(syscall.SIGKILL), &runsc.KillOpts{
+		All: true,
+	})
+	// Ignore error handling for `runsc kill --all` for now.
+	// * If it doesn't return error, it is good;
+	// * If it returns error, consider the container has already stopped.
+	// TODO: Fix `runsc kill --all` error handling.
+	return nil
+}
+
+// Stdin of the process
+func (p *Init) Stdin() io.Closer {
+	return p.stdin
+}
+
+// Runtime returns the OCI runtime configured for the init process
+func (p *Init) Runtime() *runsc.Runsc {
+	return p.runtime
+}
+
+// Exec returns a new child process
+func (p *Init) Exec(ctx context.Context, path string, r *ExecConfig) (proc.Process, error) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	return p.initState.Exec(ctx, path, r)
+}
+
+// exec returns a new exec'd process
+func (p *Init) exec(ctx context.Context, path string, r *ExecConfig) (proc.Process, error) {
+	// process exec request
+	var spec specs.Process
+	if err := json.Unmarshal(r.Spec.Value, &spec); err != nil {
+		return nil, err
+	}
+	spec.Terminal = r.Terminal
+
+	e := &execProcess{
+		id:     r.ID,
+		path:   path,
+		parent: p,
+		spec:   spec,
+		stdio: proc.Stdio{
+			Stdin:    r.Stdin,
+			Stdout:   r.Stdout,
+			Stderr:   r.Stderr,
+			Terminal: r.Terminal,
+		},
+		waitBlock: make(chan struct{}),
+	}
+	e.execState = &execCreatedState{p: e}
+	return e, nil
+}
+
+// Stdio of the process
+func (p *Init) Stdio() proc.Stdio {
+	return p.stdio
+}
+
+func (p *Init) runtimeError(rErr error, msg string) error {
+	if rErr == nil {
+		return nil
+	}
+
+	rMsg, err := getLastRuntimeError(p.runtime)
+	switch {
+	case err != nil:
+		return errors.Wrapf(rErr, "%s: %s (%s)", msg, "unable to retrieve OCI runtime error", err.Error())
+	case rMsg == "":
+		return errors.Wrap(rErr, msg)
+	default:
+		return errors.Errorf("%s: %s", msg, rMsg)
+	}
+}
+
+func (p *Init) convertStatus(status string) string {
+	if status == "created" && !p.Sandbox && p.status == internalErrorCode {
+		// Treat start failure state for non-root container as stopped.
+		return "stopped"
+	}
+	return status
+}
+
+func withConditionalIO(c proc.Stdio) runc.IOOpt {
+	return func(o *runc.IOOption) {
+		o.OpenStdin = c.Stdin != ""
+		o.OpenStdout = c.Stdout != ""
+		o.OpenStderr = c.Stderr != ""
+	}
+}

pkg/shim/v1/proc/init_state.go

@@ -0,0 +1,182 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"context"
+
+	"github.com/containerd/console"
+	"github.com/containerd/containerd/errdefs"
+	"github.com/containerd/containerd/runtime/proc"
+	"github.com/pkg/errors"
+)
+
+type initState interface {
+	Resize(console.WinSize) error
+	Start(context.Context) error
+	Delete(context.Context) error
+	Exec(context.Context, string, *ExecConfig) (proc.Process, error)
+	Kill(context.Context, uint32, bool) error
+	SetExited(int)
+}
+
+type createdState struct {
+	p *Init
+}
+
+func (s *createdState) transition(name string) error {
+	switch name {
+	case "running":
+		s.p.initState = &runningState{p: s.p}
+	case "stopped":
+		s.p.initState = &stoppedState{p: s.p}
+	case "deleted":
+		s.p.initState = &deletedState{}
+	default:
+		return errors.Errorf("invalid state transition %q to %q", stateName(s), name)
+	}
+	return nil
+}
+
+func (s *createdState) Resize(ws console.WinSize) error {
+	return s.p.resize(ws)
+}
+
+func (s *createdState) Start(ctx context.Context) error {
+	if err := s.p.start(ctx); err != nil {
+		// Containerd doesn't allow deleting container in created state.
+		// However, for gvisor, a non-root container in created state can
+		// only go to running state. If the container can't be started,
+		// it can only stay in created state, and never be deleted.
+		// To work around that, we treat non-root container in start failure
+		// state as stopped.
+		if !s.p.Sandbox {
+			s.p.io.Close()
+			s.p.setExited(internalErrorCode)
+			if err := s.transition("stopped"); err != nil {
+				panic(err)
+			}
+		}
+		return err
+	}
+	return s.transition("running")
+}
+
+func (s *createdState) Delete(ctx context.Context) error {
+	if err := s.p.delete(ctx); err != nil {
+		return err
+	}
+	return s.transition("deleted")
+}
+
+func (s *createdState) Kill(ctx context.Context, sig uint32, all bool) error {
+	return s.p.kill(ctx, sig, all)
+}
+
+func (s *createdState) SetExited(status int) {
+	s.p.setExited(status)
+
+	if err := s.transition("stopped"); err != nil {
+		panic(err)
+	}
+}
+
+func (s *createdState) Exec(ctx context.Context, path string, r *ExecConfig) (proc.Process, error) {
+	return s.p.exec(ctx, path, r)
+}
+
+type runningState struct {
+	p *Init
+}
+
+func (s *runningState) transition(name string) error {
+	switch name {
+	case "stopped":
+		s.p.initState = &stoppedState{p: s.p}
+	default:
+		return errors.Errorf("invalid state transition %q to %q", stateName(s), name)
+	}
+	return nil
+}
+
+func (s *runningState) Resize(ws console.WinSize) error {
+	return s.p.resize(ws)
+}
+
+func (s *runningState) Start(ctx context.Context) error {
+	return errors.Errorf("cannot start a running process")
+}
+
+func (s *runningState) Delete(ctx context.Context) error {
+	return errors.Errorf("cannot delete a running process")
+}
+
+func (s *runningState) Kill(ctx context.Context, sig uint32, all bool) error {
+	return s.p.kill(ctx, sig, all)
+}
+
+func (s *runningState) SetExited(status int) {
+	s.p.setExited(status)
+
+	if err := s.transition("stopped"); err != nil {
+		panic(err)
+	}
+}
+
+func (s *runningState) Exec(ctx context.Context, path string, r *ExecConfig) (proc.Process, error) {
+	return s.p.exec(ctx, path, r)
+}
+
+type stoppedState struct {
+	p *Init
+}
+
+func (s *stoppedState) transition(name string) error {
+	switch name {
+	case "deleted":
+		s.p.initState = &deletedState{}
+	default:
+		return errors.Errorf("invalid state transition %q to %q", stateName(s), name)
+	}
+	return nil
+}
+
+func (s *stoppedState) Resize(ws console.WinSize) error {
+	return errors.Errorf("cannot resize a stopped container")
+}
+
+func (s *stoppedState) Start(ctx context.Context) error {
+	return errors.Errorf("cannot start a stopped process")
+}
+
+func (s *stoppedState) Delete(ctx context.Context) error {
+	if err := s.p.delete(ctx); err != nil {
+		return err
+	}
+	return s.transition("deleted")
+}
+
+func (s *stoppedState) Kill(ctx context.Context, sig uint32, all bool) error {
+	return errdefs.ToGRPCf(errdefs.ErrNotFound, "process %s not found", s.p.id)
+}
+
+func (s *stoppedState) SetExited(status int) {
+	// no op
+}
+
+func (s *stoppedState) Exec(ctx context.Context, path string, r *ExecConfig) (proc.Process, error) {
+	return nil, errors.Errorf("cannot exec in a stopped state")
+}

pkg/shim/v1/proc/io.go

@@ -0,0 +1,167 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"os"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"github.com/containerd/containerd/log"
+	"github.com/containerd/fifo"
+	runc "github.com/containerd/go-runc"
+)
+
+// TODO(random-liu): This file can be a util.
+
+var bufPool = sync.Pool{
+	New: func() interface{} {
+		buffer := make([]byte, 32<<10)
+		return &buffer
+	},
+}
+
+func copyPipes(ctx context.Context, rio runc.IO, stdin, stdout, stderr string, wg, cwg *sync.WaitGroup) error {
+	var sameFile *countingWriteCloser
+	for _, i := range []struct {
+		name string
+		dest func(wc io.WriteCloser, rc io.Closer)
+	}{
+		{
+			name: stdout,
+			dest: func(wc io.WriteCloser, rc io.Closer) {
+				wg.Add(1)
+				cwg.Add(1)
+				go func() {
+					cwg.Done()
+					p := bufPool.Get().(*[]byte)
+					defer bufPool.Put(p)
+					if _, err := io.CopyBuffer(wc, rio.Stdout(), *p); err != nil {
+						log.G(ctx).Warn("error copying stdout")
+					}
+					wg.Done()
+					wc.Close()
+					if rc != nil {
+						rc.Close()
+					}
+				}()
+			},
+		}, {
+			name: stderr,
+			dest: func(wc io.WriteCloser, rc io.Closer) {
+				wg.Add(1)
+				cwg.Add(1)
+				go func() {
+					cwg.Done()
+					p := bufPool.Get().(*[]byte)
+					defer bufPool.Put(p)
+					if _, err := io.CopyBuffer(wc, rio.Stderr(), *p); err != nil {
+						log.G(ctx).Warn("error copying stderr")
+					}
+					wg.Done()
+					wc.Close()
+					if rc != nil {
+						rc.Close()
+					}
+				}()
+			},
+		},
+	} {
+		ok, err := isFifo(i.name)
+		if err != nil {
+			return err
+		}
+		var (
+			fw io.WriteCloser
+			fr io.Closer
+		)
+		if ok {
+			if fw, err = fifo.OpenFifo(ctx, i.name, syscall.O_WRONLY, 0); err != nil {
+				return fmt.Errorf("gvisor-containerd-shim: opening %s failed: %s", i.name, err)
+			}
+			if fr, err = fifo.OpenFifo(ctx, i.name, syscall.O_RDONLY, 0); err != nil {
+				return fmt.Errorf("gvisor-containerd-shim: opening %s failed: %s", i.name, err)
+			}
+		} else {
+			if sameFile != nil {
+				sameFile.count++
+				i.dest(sameFile, nil)
+				continue
+			}
+			if fw, err = os.OpenFile(i.name, syscall.O_WRONLY|syscall.O_APPEND, 0); err != nil {
+				return fmt.Errorf("gvisor-containerd-shim: opening %s failed: %s", i.name, err)
+			}
+			if stdout == stderr {
+				sameFile = &countingWriteCloser{
+					WriteCloser: fw,
+					count:       1,
+				}
+			}
+		}
+		i.dest(fw, fr)
+	}
+	if stdin == "" {
+		return nil
+	}
+	f, err := fifo.OpenFifo(context.Background(), stdin, syscall.O_RDONLY|syscall.O_NONBLOCK, 0)
+	if err != nil {
+		return fmt.Errorf("gvisor-containerd-shim: opening %s failed: %s", stdin, err)
+	}
+	cwg.Add(1)
+	go func() {
+		cwg.Done()
+		p := bufPool.Get().(*[]byte)
+		defer bufPool.Put(p)
+
+		io.CopyBuffer(rio.Stdin(), f, *p)
+		rio.Stdin().Close()
+		f.Close()
+	}()
+	return nil
+}
+
+// countingWriteCloser masks io.Closer() until close has been invoked a certain number of times.
+type countingWriteCloser struct {
+	io.WriteCloser
+	count int64
+}
+
+func (c *countingWriteCloser) Close() error {
+	if atomic.AddInt64(&c.count, -1) > 0 {
+		return nil
+	}
+	return c.WriteCloser.Close()
+}
+
+// isFifo checks if a file is a fifo
+// if the file does not exist then it returns false
+func isFifo(path string) (bool, error) {
+	stat, err := os.Stat(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return false, nil
+		}
+		return false, err
+	}
+	if stat.Mode()&os.ModeNamedPipe == os.ModeNamedPipe {
+		return true, nil
+	}
+	return false, nil
+}

pkg/shim/v1/proc/process.go

@@ -0,0 +1,37 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"github.com/pkg/errors"
+)
+
+// RunscRoot is the path to the root runsc state directory
+const RunscRoot = "/run/containerd/runsc"
+
+func stateName(v interface{}) string {
+	switch v.(type) {
+	case *runningState, *execRunningState:
+		return "running"
+	case *createdState, *execCreatedState:
+		return "created"
+	case *deletedState:
+		return "deleted"
+	case *stoppedState:
+		return "stopped"
+	}
+	panic(errors.Errorf("invalid state %v", v))
+}

pkg/shim/v1/proc/types.go

@@ -0,0 +1,70 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"time"
+
+	google_protobuf "github.com/gogo/protobuf/types"
+
+	runc "github.com/containerd/go-runc"
+)
+
+// Mount holds filesystem mount configuration
+type Mount struct {
+	Type    string
+	Source  string
+	Target  string
+	Options []string
+}
+
+// CreateConfig hold task creation configuration
+type CreateConfig struct {
+	ID       string
+	Bundle   string
+	Runtime  string
+	Rootfs   []Mount
+	Terminal bool
+	Stdin    string
+	Stdout   string
+	Stderr   string
+	Options  *google_protobuf.Any
+}
+
+// ExecConfig holds exec creation configuration
+type ExecConfig struct {
+	ID       string
+	Terminal bool
+	Stdin    string
+	Stdout   string
+	Stderr   string
+	Spec     *google_protobuf.Any
+}
+
+// Exit is the type of exit events
+type Exit struct {
+	Timestamp time.Time
+	ID        string
+	Status    int
+}
+
+// ProcessMonitor monitors process exit changes
+type ProcessMonitor interface {
+	// Subscribe to process exit changes
+	Subscribe() chan runc.Exit
+	// Unsubscribe to process exit changes
+	Unsubscribe(c chan runc.Exit)
+}

pkg/shim/v1/proc/utils.go

@@ -0,0 +1,92 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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 proc
+
+import (
+	"encoding/json"
+	"io"
+	"os"
+	"strings"
+	"time"
+
+	runsc "github.com/google/gvisor-containerd-shim/pkg/go-runsc"
+)
+
+const (
+	internalErrorCode = 128
+	bufferSize        = 32
+)
+
+// ExitCh is the exit events channel for containers and exec processes
+// inside the sandbox.
+var ExitCh = make(chan Exit, bufferSize)
+
+// TODO(random-liu): This can be a utility.
+
+// TODO(mlaventure): move to runc package?
+func getLastRuntimeError(r *runsc.Runsc) (string, error) {
+	if r.Log == "" {
+		return "", nil
+	}
+
+	f, err := os.OpenFile(r.Log, os.O_RDONLY, 0400)
+	if err != nil {
+		return "", err
+	}
+
+	var (
+		errMsg string
+		log    struct {
+			Level string
+			Msg   string
+			Time  time.Time
+		}
+	)
+
+	dec := json.NewDecoder(f)
+	for err = nil; err == nil; {
+		if err = dec.Decode(&log); err != nil && err != io.EOF {
+			return "", err
+		}
+		if log.Level == "error" {
+			errMsg = strings.TrimSpace(log.Msg)
+		}
+	}
+
+	return errMsg, nil
+}
+
+func copyFile(to, from string) error {
+	ff, err := os.Open(from)
+	if err != nil {
+		return err
+	}
+	defer ff.Close()
+	tt, err := os.Create(to)
+	if err != nil {
+		return err
+	}
+	defer tt.Close()
+
+	p := bufPool.Get().(*[]byte)
+	defer bufPool.Put(p)
+	_, err = io.CopyBuffer(tt, ff, *p)
+	return err
+}
+
+func hasNoIO(r *CreateConfig) bool {
+	return r.Stdin == "" && r.Stdout == "" && r.Stderr == ""
+}

pkg/shim/v1/shim/platform.go

@@ -0,0 +1,110 @@
+// Copyright 2018 The containerd Authors.
+// 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
+//
+//     https://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 shim
+
+import (
+	"context"
+	"io"
+	"sync"
+	"syscall"
+
+	"github.com/containerd/console"
+	"github.com/containerd/fifo"
+	"github.com/pkg/errors"
+)
+
+type linuxPlatform struct {
+	epoller *console.Epoller
+}
+
+func (p *linuxPlatform) CopyConsole(ctx context.Context, console console.Console, stdin, stdout, stderr string, wg, cwg *sync.WaitGroup) (console.Console, error) {
+	if p.epoller == nil {
+		return nil, errors.New("uninitialized epoller")
+	}
+
+	epollConsole, err := p.epoller.Add(console)
+	if err != nil {
+		return nil, err
+	}
+
+	if stdin != "" {
+		in, err := fifo.OpenFifo(ctx, stdin, syscall.O_RDONLY, 0)
+		if err != nil {
+			return nil, err
+		}
+		cwg.Add(1)
+		go func() {
+			cwg.Done()
+			p := bufPool.Get().(*[]byte)
+			defer bufPool.Put(p)
+			io.CopyBuffer(epollConsole, in, *p)
+		}()
+	}
+
+	outw, err := fifo.OpenFifo(ctx, stdout, syscall.O_WRONLY, 0)
+	if err != nil {
+		return nil, err
+	}
+	outr, err := fifo.OpenFifo(ctx, stdout, syscall.O_RDONLY, 0)
+	if err != nil {
+		return nil, err
+	}
+	wg.Add(1)
+	cwg.Add(1)
+	go func() {
+		cwg.Done()
+		p := bufPool.Get().(*[]byte)
+		defer bufPool.Put(p)
+		io.CopyBuffer(outw, epollConsole, *p)
+		epollConsole.Close()
+		outr.Close()
+		outw.Close()
+		wg.Done()
+	}()
+	return epollConsole, nil
+}
+
+func (p *linuxPlatform) ShutdownConsole(ctx context.Context, cons console.Console) error {
+	if p.epoller == nil {
+		return errors.New("uninitialized epoller")
+	}
+	epollConsole, ok := cons.(*console.EpollConsole)
+	if !ok {
+		return errors.Errorf("expected EpollConsole, got %#v", cons)
+	}
+	return epollConsole.Shutdown(p.epoller.CloseConsole)
+}
+
+func (p *linuxPlatform) Close() error {
+	return p.epoller.Close()
+}
+
+// initialize a single epoll fd to manage our consoles. `initPlatform` should
+// only be called once.
+func (s *Service) initPlatform() error {
+	if s.platform != nil {
+		return nil
+	}
+	epoller, err := console.NewEpoller()
+	if err != nil {
+		return errors.Wrap(err, "failed to initialize epoller")
+	}
+	s.platform = &linuxPlatform{
+		epoller: epoller,
+	}
+	go epoller.Wait()
+	return nil
+}

pkg/shim/v1/utils/utils.go

@@ -0,0 +1,57 @@
+// Copyright 2018 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
+//
+//     https://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 utils
+
+import (
+	"encoding/json"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+
+	"github.com/containerd/cri/pkg/annotations"
+	specs "github.com/opencontainers/runtime-spec/specs-go"
+)
+
+// ReadSpec reads OCI spec from the bundle directory.
+func ReadSpec(bundle string) (*specs.Spec, error) {
+	f, err := os.Open(filepath.Join(bundle, "config.json"))
+	if err != nil {
+		return nil, err
+	}
+	b, err := ioutil.ReadAll(f)
+	if err != nil {
+		return nil, err
+	}
+	var spec specs.Spec
+	if err := json.Unmarshal(b, &spec); err != nil {
+		return nil, err
+	}
+	return &spec, nil
+}
+
+// IsSandbox checks whether a container is a sandbox container.
+func IsSandbox(spec *specs.Spec) bool {
+	t, ok := spec.Annotations[annotations.ContainerType]
+	return !ok || t == annotations.ContainerTypeSandbox
+}
+
+// UserLogPath gets user log path from OCI annotation.
+func UserLogPath(spec *specs.Spec) string {
+	sandboxLogDir := spec.Annotations[annotations.SandboxLogDir]
+	if sandboxLogDir == "" {
+		return ""
+	}
+	return filepath.Join(sandboxLogDir, "gvisor.log")
+}

pkg/shim/v1/utils/volumes.go

@@ -0,0 +1,156 @@
+// Copyright 2018 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
+//
+//     https://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 utils
+
+import (
+	"encoding/json"
+	"fmt"
+	"io/ioutil"
+	"path/filepath"
+	"strings"
+
+	"github.com/containerd/cri/pkg/annotations"
+	specs "github.com/opencontainers/runtime-spec/specs-go"
+	"github.com/pkg/errors"
+	"github.com/sirupsen/logrus"
+)
+
+const volumeKeyPrefix = "dev.gvisor.spec.mount."
+
+var kubeletPodsDir = "/var/lib/kubelet/pods"
+
+// volumeName gets volume name from volume annotation key, example:
+// dev.gvisor.spec.mount.NAME.share
+func volumeName(k string) string {
+	return strings.SplitN(strings.TrimPrefix(k, volumeKeyPrefix), ".", 2)[0]
+}
+
+// volumeFieldName gets volume field name from volume annotation key, example:
+// `type` is the field of dev.gvisor.spec.mount.NAME.type
+func volumeFieldName(k string) string {
+	parts := strings.Split(strings.TrimPrefix(k, volumeKeyPrefix), ".")
+	return parts[len(parts)-1]
+}
+
+// podUID gets pod UID from the pod log path.
+func podUID(s *specs.Spec) (string, error) {
+	sandboxLogDir := s.Annotations[annotations.SandboxLogDir]
+	if sandboxLogDir == "" {
+		return "", errors.New("no sandbox log path annotation")
+	}
+	fields := strings.Split(filepath.Base(sandboxLogDir), "_")
+	switch len(fields) {
+	case 1: // This is the old CRI logging path
+		return fields[0], nil
+	case 3: // This is the new CRI logging path
+		return fields[2], nil
+	}
+	return "", errors.Errorf("unexpected sandbox log path %q", sandboxLogDir)
+}
+
+// isVolumeKey checks whether an annotation key is for volume.
+func isVolumeKey(k string) bool {
+	return strings.HasPrefix(k, volumeKeyPrefix)
+}
+
+// volumeSourceKey constructs the annotation key for volume source.
+func volumeSourceKey(volume string) string {
+	return volumeKeyPrefix + volume + ".source"
+}
+
+// volumePath searches the volume path in the kubelet pod directory.
+func volumePath(volume, uid string) (string, error) {
+	// TODO: Support subpath when gvisor supports pod volume bind mount.
+	volumeSearchPath := fmt.Sprintf("%s/%s/volumes/*/%s", kubeletPodsDir, uid, volume)
+	dirs, err := filepath.Glob(volumeSearchPath)
+	if err != nil {
+		return "", err
+	}
+	if len(dirs) != 1 {
+		return "", errors.Errorf("unexpected matched volume list %v", dirs)
+	}
+	return dirs[0], nil
+}
+
+// isVolumePath checks whether a string is the volume path.
+func isVolumePath(volume, path string) (bool, error) {
+	// TODO: Support subpath when gvisor supports pod volume bind mount.
+	volumeSearchPath := fmt.Sprintf("%s/*/volumes/*/%s", kubeletPodsDir, volume)
+	return filepath.Match(volumeSearchPath, path)
+}
+
+// UpdateVolumeAnnotations add necessary OCI annotations for gvisor
+// volume optimization.
+func UpdateVolumeAnnotations(bundle string, s *specs.Spec) error {
+	var (
+		uid string
+		err error
+	)
+	if IsSandbox(s) {
+		uid, err = podUID(s)
+		if err != nil {
+			// Skip if we can't get pod UID, because this doesn't work
+			// for containerd 1.1.
+			logrus.WithError(err).Error("Can't get pod uid")
+			return nil
+		}
+	}
+	var updated bool
+	for k, v := range s.Annotations {
+		if !isVolumeKey(k) {
+			continue
+		}
+		if volumeFieldName(k) != "type" {
+			continue
+		}
+		volume := volumeName(k)
+		if uid != "" {
+			// This is a sandbox
+			path, err := volumePath(volume, uid)
+			if err != nil {
+				return errors.Wrapf(err, "get volume path for %q", volume)
+			}
+			s.Annotations[volumeSourceKey(volume)] = path
+			updated = true
+		} else {
+			// This is a container
+			for i := range s.Mounts {
+				// An error is returned for sandbox if source annotation
+				// is not successfully applied, so it is guaranteed that
+				// the source annotation for sandbox has already been
+				// successfully applied at this point.
+				// The volume name is unique inside a pod, so matching without
+				// podUID is fine here.
+				// TODO: Pass podUID down to shim for containers to do
+				// more accurate matching.
+				if yes, _ := isVolumePath(volume, s.Mounts[i].Source); yes {
+					// gVisor requires the container mount type to match
+					// sandbox mount type.
+					s.Mounts[i].Type = v
+					updated = true
+				}
+			}
+		}
+	}
+	if !updated {
+		return nil
+	}
+	// Update bundle
+	b, err := json.Marshal(s)
+	if err != nil {
+		return err
+	}
+	return ioutil.WriteFile(filepath.Join(bundle, "config.json"), b, 0666)
+}

pkg/shim/v1/utils/volumes_test.go

@@ -0,0 +1,309 @@
+// 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
+//
+//     https://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 utils
+
+import (
+	"encoding/json"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"reflect"
+	"testing"
+
+	"github.com/containerd/cri/pkg/annotations"
+	specs "github.com/opencontainers/runtime-spec/specs-go"
+)
+
+func TestUpdateVolumeAnnotations(t *testing.T) {
+	dir, err := ioutil.TempDir("", "test-update-volume-annotations")
+	if err != nil {
+		t.Fatalf("create tempdir: %v", err)
+	}
+	defer os.RemoveAll(dir)
+	kubeletPodsDir = dir
+
+	const (
+		testPodUID           = "testuid"
+		testVolumeName       = "testvolume"
+		testLogDirPath       = "/var/log/pods/testns_testname_" + testPodUID
+		testLegacyLogDirPath = "/var/log/pods/" + testPodUID
+	)
+	testVolumePath := fmt.Sprintf("%s/%s/volumes/kubernetes.io~empty-dir/%s", dir, testPodUID, testVolumeName)
+
+	if err := os.MkdirAll(testVolumePath, 0755); err != nil {
+		t.Fatalf("Create test volume: %v", err)
+	}
+
+	for _, test := range []struct {
+		desc         string
+		spec         *specs.Spec
+		expected     *specs.Spec
+		expectErr    bool
+		expectUpdate bool
+	}{
+		{
+			desc: "volume annotations for sandbox",
+			spec: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.SandboxLogDir:                              testLogDirPath,
+					annotations.ContainerType:                              annotations.ContainerTypeSandbox,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "pod",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "tmpfs",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+				},
+			},
+			expected: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.SandboxLogDir:                              testLogDirPath,
+					annotations.ContainerType:                              annotations.ContainerTypeSandbox,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "pod",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "tmpfs",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+					"dev.gvisor.spec.mount." + testVolumeName + ".source":  testVolumePath,
+				},
+			},
+			expectUpdate: true,
+		},
+		{
+			desc: "volume annotations for sandbox with legacy log path",
+			spec: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.SandboxLogDir:                              testLegacyLogDirPath,
+					annotations.ContainerType:                              annotations.ContainerTypeSandbox,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "pod",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "tmpfs",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+				},
+			},
+			expected: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.SandboxLogDir:                              testLegacyLogDirPath,
+					annotations.ContainerType:                              annotations.ContainerTypeSandbox,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "pod",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "tmpfs",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+					"dev.gvisor.spec.mount." + testVolumeName + ".source":  testVolumePath,
+				},
+			},
+			expectUpdate: true,
+		},
+		{
+			desc: "tmpfs: volume annotations for container",
+			spec: &specs.Spec{
+				Mounts: []specs.Mount{
+					{
+						Destination: "/test",
+						Type:        "bind",
+						Source:      testVolumePath,
+						Options:     []string{"ro"},
+					},
+					{
+						Destination: "/random",
+						Type:        "bind",
+						Source:      "/random",
+						Options:     []string{"ro"},
+					},
+				},
+				Annotations: map[string]string{
+					annotations.ContainerType:                              annotations.ContainerTypeContainer,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "pod",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "tmpfs",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+				},
+			},
+			expected: &specs.Spec{
+				Mounts: []specs.Mount{
+					{
+						Destination: "/test",
+						Type:        "tmpfs",
+						Source:      testVolumePath,
+						Options:     []string{"ro"},
+					},
+					{
+						Destination: "/random",
+						Type:        "bind",
+						Source:      "/random",
+						Options:     []string{"ro"},
+					},
+				},
+				Annotations: map[string]string{
+					annotations.ContainerType:                              annotations.ContainerTypeContainer,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "pod",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "tmpfs",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+				},
+			},
+			expectUpdate: true,
+		},
+		{
+			desc: "bind: volume annotations for container",
+			spec: &specs.Spec{
+				Mounts: []specs.Mount{
+					{
+						Destination: "/test",
+						Type:        "bind",
+						Source:      testVolumePath,
+						Options:     []string{"ro"},
+					},
+				},
+				Annotations: map[string]string{
+					annotations.ContainerType:                              annotations.ContainerTypeContainer,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "container",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "bind",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+				},
+			},
+			expected: &specs.Spec{
+				Mounts: []specs.Mount{
+					{
+						Destination: "/test",
+						Type:        "bind",
+						Source:      testVolumePath,
+						Options:     []string{"ro"},
+					},
+				},
+				Annotations: map[string]string{
+					annotations.ContainerType:                              annotations.ContainerTypeContainer,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "container",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "bind",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+				},
+			},
+			expectUpdate: true,
+		},
+		{
+			desc: "should not return error without pod log directory",
+			spec: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.ContainerType:                              annotations.ContainerTypeSandbox,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "pod",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "tmpfs",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+				},
+			},
+			expected: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.ContainerType:                              annotations.ContainerTypeSandbox,
+					"dev.gvisor.spec.mount." + testVolumeName + ".share":   "pod",
+					"dev.gvisor.spec.mount." + testVolumeName + ".type":    "tmpfs",
+					"dev.gvisor.spec.mount." + testVolumeName + ".options": "ro",
+				},
+			},
+		},
+		{
+			desc: "should return error if volume path does not exist",
+			spec: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.SandboxLogDir:                testLogDirPath,
+					annotations.ContainerType:                annotations.ContainerTypeSandbox,
+					"dev.gvisor.spec.mount.notexist.share":   "pod",
+					"dev.gvisor.spec.mount.notexist.type":    "tmpfs",
+					"dev.gvisor.spec.mount.notexist.options": "ro",
+				},
+			},
+			expectErr: true,
+		},
+		{
+			desc: "no volume annotations for sandbox",
+			spec: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.SandboxLogDir: testLogDirPath,
+					annotations.ContainerType: annotations.ContainerTypeSandbox,
+				},
+			},
+			expected: &specs.Spec{
+				Annotations: map[string]string{
+					annotations.SandboxLogDir: testLogDirPath,
+					annotations.ContainerType: annotations.ContainerTypeSandbox,
+				},
+			},
+		},
+		{
+			desc: "no volume annotations for container",
+			spec: &specs.Spec{
+				Mounts: []specs.Mount{
+					{
+						Destination: "/test",
+						Type:        "bind",
+						Source:      "/test",
+						Options:     []string{"ro"},
+					},
+					{
+						Destination: "/random",
+						Type:        "bind",
+						Source:      "/random",
+						Options:     []string{"ro"},
+					},
+				},
+				Annotations: map[string]string{
+					annotations.ContainerType: annotations.ContainerTypeContainer,
+				},
+			},
+			expected: &specs.Spec{
+				Mounts: []specs.Mount{
+					{
+						Destination: "/test",
+						Type:        "bind",
+						Source:      "/test",
+						Options:     []string{"ro"},
+					},
+					{
+						Destination: "/random",
+						Type:        "bind",
+						Source:      "/random",
+						Options:     []string{"ro"},
+					},
+				},
+				Annotations: map[string]string{
+					annotations.ContainerType: annotations.ContainerTypeContainer,
+				},
+			},
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			bundle, err := ioutil.TempDir(dir, "test-bundle")
+			if err != nil {
+				t.Fatalf("Create test bundle: %v", err)
+			}
+			err = UpdateVolumeAnnotations(bundle, test.spec)
+			if test.expectErr {
+				if err == nil {
+					t.Fatal("Expected error, but got nil")
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("Unexpected error: %v", err)
+			}
+			if !reflect.DeepEqual(test.expected, test.spec) {
+				t.Fatalf("Expected %+v, got %+v", test.expected, test.spec)
+			}
+			if test.expectUpdate {
+				b, err := ioutil.ReadFile(filepath.Join(bundle, "config.json"))
+				if err != nil {
+					t.Fatalf("Read spec from bundle: %v", err)
+				}
+				var spec specs.Spec
+				if err := json.Unmarshal(b, &spec); err != nil {
+					t.Fatalf("Unmarshal spec: %v", err)
+				}
+				if !reflect.DeepEqual(test.expected, &spec) {
+					t.Fatalf("Expected %+v, got %+v", test.expected, &spec)
+				}
+			}
+		})
+	}
+}

pkg/shim/v2/epoll.go

@@ -0,0 +1,128 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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.
+
+// +build linux
+
+package v2
+
+import (
+	"context"
+	"sync"
+
+	"github.com/containerd/cgroups"
+	eventstypes "github.com/containerd/containerd/api/events"
+	"github.com/containerd/containerd/events"
+	"github.com/containerd/containerd/runtime"
+	"github.com/sirupsen/logrus"
+	"golang.org/x/sys/unix"
+)
+
+func newOOMEpoller(publisher events.Publisher) (*epoller, error) {
+	fd, err := unix.EpollCreate1(unix.EPOLL_CLOEXEC)
+	if err != nil {
+		return nil, err
+	}
+	return &epoller{
+		fd:        fd,
+		publisher: publisher,
+		set:       make(map[uintptr]*item),
+	}, nil
+}
+
+type epoller struct {
+	mu sync.Mutex
+
+	fd        int
+	publisher events.Publisher
+	set       map[uintptr]*item
+}
+
+type item struct {
+	id string
+	cg cgroups.Cgroup
+}
+
+func (e *epoller) Close() error {
+	return unix.Close(e.fd)
+}
+
+func (e *epoller) run(ctx context.Context) {
+	var events [128]unix.EpollEvent
+	for {
+		select {
+		case <-ctx.Done():
+			e.Close()
+			return
+		default:
+			n, err := unix.EpollWait(e.fd, events[:], -1)
+			if err != nil {
+				if err == unix.EINTR {
+					continue
+				}
+				logrus.WithError(err).Error("cgroups: epoll wait")
+			}
+			for i := 0; i < n; i++ {
+				e.process(ctx, uintptr(events[i].Fd))
+			}
+		}
+	}
+}
+
+func (e *epoller) add(id string, cg cgroups.Cgroup) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	fd, err := cg.OOMEventFD()
+	if err != nil {
+		return err
+	}
+	e.set[fd] = &item{
+		id: id,
+		cg: cg,
+	}
+	event := unix.EpollEvent{
+		Fd:     int32(fd),
+		Events: unix.EPOLLHUP | unix.EPOLLIN | unix.EPOLLERR,
+	}
+	return unix.EpollCtl(e.fd, unix.EPOLL_CTL_ADD, int(fd), &event)
+}
+
+func (e *epoller) process(ctx context.Context, fd uintptr) {
+	flush(fd)
+	e.mu.Lock()
+	i, ok := e.set[fd]
+	if !ok {
+		e.mu.Unlock()
+		return
+	}
+	e.mu.Unlock()
+	if i.cg.State() == cgroups.Deleted {
+		e.mu.Lock()
+		delete(e.set, fd)
+		e.mu.Unlock()
+		unix.Close(int(fd))
+		return
+	}
+	if err := e.publisher.Publish(ctx, runtime.TaskOOMEventTopic, &eventstypes.TaskOOM{
+		ContainerID: i.id,
+	}); err != nil {
+		logrus.WithError(err).Error("publish OOM event")
+	}
+}
+
+func flush(fd uintptr) error {
+	var buf [8]byte
+	_, err := unix.Read(int(fd), buf[:])
+	return err
+}

pkg/shim/v2/options/options.go

@@ -0,0 +1,33 @@
+// Copyright 2018 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
+//
+//     https://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 options
+
+const OptionType = "io.containerd.runsc.v1.options"
+
+// Options is runtime options for io.containerd.runsc.v1.
+type Options struct {
+	// ShimCgroup is the cgroup the shim should be in.
+	ShimCgroup string `toml:"shim_cgroup"`
+	// IoUid is the I/O's pipes uid.
+	IoUid uint32 `toml:"io_uid"`
+	// IoUid is the I/O's pipes gid.
+	IoGid uint32 `toml:"io_gid"`
+	// BinaryName is the binary name of the runsc binary.
+	BinaryName string `toml:"binary_name"`
+	// Root is the runsc root directory.
+	Root string `toml:"root"`
+	// RunscConfig is a key/value map of all runsc flags.
+	RunscConfig map[string]string `toml:"runsc_config"`
+}

pkg/shim/v2/service_linux.go

@@ -0,0 +1,112 @@
+// Copyright 2018 The containerd Authors.
+// Copyright 2018 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
+//
+//     https://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.
+
+// +build linux
+
+package v2
+
+import (
+	"context"
+	"io"
+	"sync"
+	"syscall"
+
+	"github.com/containerd/console"
+	"github.com/containerd/fifo"
+	"github.com/pkg/errors"
+)
+
+type linuxPlatform struct {
+	epoller *console.Epoller
+}
+
+func (p *linuxPlatform) CopyConsole(ctx context.Context, console console.Console, stdin, stdout, stderr string, wg, cwg *sync.WaitGroup) (console.Console, error) {
+	if p.epoller == nil {
+		return nil, errors.New("uninitialized epoller")
+	}
+
+	epollConsole, err := p.epoller.Add(console)
+	if err != nil {
+		return nil, err
+	}
+
+	if stdin != "" {
+		in, err := fifo.OpenFifo(context.Background(), stdin, syscall.O_RDONLY|syscall.O_NONBLOCK, 0)
+		if err != nil {
+			return nil, err
+		}
+		cwg.Add(1)
+		go func() {
+			cwg.Done()
+			p := bufPool.Get().(*[]byte)
+			defer bufPool.Put(p)
+			io.CopyBuffer(epollConsole, in, *p)
+		}()
+	}
+
+	outw, err := fifo.OpenFifo(ctx, stdout, syscall.O_WRONLY, 0)
+	if err != nil {
+		return nil, err
+	}
+	outr, err := fifo.OpenFifo(ctx, stdout, syscall.O_RDONLY, 0)
+	if err != nil {
+		return nil, err
+	}
+	wg.Add(1)
+	cwg.Add(1)
+	go func() {
+		cwg.Done()
+		p := bufPool.Get().(*[]byte)
+		defer bufPool.Put(p)
+		io.CopyBuffer(outw, epollConsole, *p)
+		epollConsole.Close()
+		outr.Close()
+		outw.Close()
+		wg.Done()
+	}()
+	return epollConsole, nil
+}
+
+func (p *linuxPlatform) ShutdownConsole(ctx context.Context, cons console.Console) error {
+	if p.epoller == nil {
+		return errors.New("uninitialized epoller")
+	}
+	epollConsole, ok := cons.(*console.EpollConsole)
+	if !ok {
+		return errors.Errorf("expected EpollConsole, got %#v", cons)
+	}
+	return epollConsole.Shutdown(p.epoller.CloseConsole)
+}
+
+func (p *linuxPlatform) Close() error {
+	return p.epoller.Close()
+}
+
+// initialize a single epoll fd to manage our consoles. `initPlatform` should
+// only be called once.
+func (s *service) initPlatform() error {
+	if s.platform != nil {
+		return nil
+	}
+	epoller, err := console.NewEpoller()
+	if err != nil {
+		return errors.Wrap(err, "failed to initialize epoller")
+	}
+	s.platform = &linuxPlatform{
+		epoller: epoller,
+	}
+	go epoller.Wait()
+	return nil
+}