Merge branch 'master' into ip-forwarding

- Merges aleksej-paschenko's with HEAD
- Adds vfs2 support for ip_forward

10 files changed, 1242 insertions, 0 deletions

diff --git a/pkg/shim/v2/BUILD b/pkg/shim/v2/BUILD
new file mode 100644
index 000000000..7e0a114a0
--- /dev/null
+++ b/pkg/shim/v2/BUILD
@@ -0,0 +1,43 @@
+load("//tools:defs.bzl", "go_library")
+
+package(licenses = ["notice"])
+
+go_library(
+    name = "v2",
+    srcs = [
+        "api.go",
+        "epoll.go",
+        "service.go",
+        "service_linux.go",
+    ],
+    visibility = ["//shim:__subpackages__"],
+    deps = [
+        "//pkg/shim/runsc",
+        "//pkg/shim/v1/proc",
+        "//pkg/shim/v1/utils",
+        "//pkg/shim/v2/options",
+        "//pkg/shim/v2/runtimeoptions",
+        "//runsc/specutils",
+        "@com_github_burntsushi_toml//:go_default_library",
+        "@com_github_containerd_cgroups//:go_default_library",
+        "@com_github_containerd_console//:go_default_library",
+        "@com_github_containerd_containerd//api/events:go_default_library",
+        "@com_github_containerd_containerd//api/types/task:go_default_library",
+        "@com_github_containerd_containerd//errdefs:go_default_library",
+        "@com_github_containerd_containerd//events:go_default_library",
+        "@com_github_containerd_containerd//log:go_default_library",
+        "@com_github_containerd_containerd//mount:go_default_library",
+        "@com_github_containerd_containerd//namespaces:go_default_library",
+        "@com_github_containerd_containerd//pkg/process:go_default_library",
+        "@com_github_containerd_containerd//pkg/stdio:go_default_library",
+        "@com_github_containerd_containerd//runtime:go_default_library",
+        "@com_github_containerd_containerd//runtime/linux/runctypes:go_default_library",
+        "@com_github_containerd_containerd//runtime/v2/shim:go_default_library",
+        "@com_github_containerd_containerd//runtime/v2/task:go_default_library",
+        "@com_github_containerd_containerd//sys/reaper:go_default_library",
+        "@com_github_containerd_fifo//:go_default_library",
+        "@com_github_containerd_typeurl//:go_default_library",
+        "@com_github_gogo_protobuf//types:go_default_library",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/pkg/shim/v2/api.go b/pkg/shim/v2/api.go
new file mode 100644
index 000000000..dbe5c59f6
--- /dev/null
+++ b/pkg/shim/v2/api.go
@@ -0,0 +1,22 @@
+// 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 v2
+
+import (
+	"github.com/containerd/containerd/api/events"
+)
+
+type TaskOOM = events.TaskOOM
diff --git a/pkg/shim/v2/epoll.go b/pkg/shim/v2/epoll.go
new file mode 100644
index 000000000..41232cca8
--- /dev/null
+++ b/pkg/shim/v2/epoll.go
@@ -0,0 +1,129 @@
+// 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"
+	"fmt"
+	"sync"
+
+	"github.com/containerd/cgroups"
+	"github.com/containerd/containerd/events"
+	"github.com/containerd/containerd/runtime"
+	"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 || err == unix.EAGAIN {
+					continue
+				}
+				// Should not happen.
+				panic(fmt.Errorf("cgroups: epoll wait: %w", err))
+			}
+			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, &TaskOOM{
+		ContainerID: i.id,
+	}); err != nil {
+		// Should not happen.
+		panic(fmt.Errorf("publish OOM event: %w", err))
+	}
+}
+
+func flush(fd uintptr) error {
+	var buf [8]byte
+	_, err := unix.Read(int(fd), buf[:])
+	return err
+}
diff --git a/pkg/shim/v2/options/BUILD b/pkg/shim/v2/options/BUILD
new file mode 100644
index 000000000..ca212e874
--- /dev/null
+++ b/pkg/shim/v2/options/BUILD
@@ -0,0 +1,11 @@
+load("//tools:defs.bzl", "go_library")
+
+package(licenses = ["notice"])
+
+go_library(
+    name = "options",
+    srcs = [
+        "options.go",
+    ],
+    visibility = ["//:sandbox"],
+)
diff --git a/pkg/shim/v2/options/options.go b/pkg/shim/v2/options/options.go
new file mode 100644
index 000000000..de09f2f79
--- /dev/null
+++ b/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"`
+}
diff --git a/pkg/shim/v2/runtimeoptions/BUILD b/pkg/shim/v2/runtimeoptions/BUILD
new file mode 100644
index 000000000..01716034c
--- /dev/null
+++ b/pkg/shim/v2/runtimeoptions/BUILD
@@ -0,0 +1,20 @@
+load("//tools:defs.bzl", "go_library", "proto_library")
+
+package(licenses = ["notice"])
+
+proto_library(
+    name = "api",
+    srcs = [
+        "runtimeoptions.proto",
+    ],
+)
+
+go_library(
+    name = "runtimeoptions",
+    srcs = ["runtimeoptions.go"],
+    visibility = ["//pkg/shim/v2:__pkg__"],
+    deps = [
+        "//pkg/shim/v2/runtimeoptions:api_go_proto",
+        "@com_github_gogo_protobuf//proto:go_default_library",
+    ],
+)
diff --git a/pkg/shim/v2/runtimeoptions/runtimeoptions.go b/pkg/shim/v2/runtimeoptions/runtimeoptions.go
new file mode 100644
index 000000000..1c1a0c5d1
--- /dev/null
+++ b/pkg/shim/v2/runtimeoptions/runtimeoptions.go
@@ -0,0 +1,27 @@
+// 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 runtimeoptions
+
+import (
+	proto "github.com/gogo/protobuf/proto"
+	pb "gvisor.dev/gvisor/pkg/shim/v2/runtimeoptions/api_go_proto"
+)
+
+type Options = pb.Options
+
+func init() {
+	proto.RegisterType((*Options)(nil), "cri.runtimeoptions.v1.Options")
+}
diff --git a/pkg/shim/v2/runtimeoptions/runtimeoptions.proto b/pkg/shim/v2/runtimeoptions/runtimeoptions.proto
new file mode 100644
index 000000000..edb19020a
--- /dev/null
+++ b/pkg/shim/v2/runtimeoptions/runtimeoptions.proto
@@ -0,0 +1,25 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     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.
+
+syntax = "proto3";
+
+package runtimeoptions;
+
+// This is a version of the runtimeoptions CRI API that is vendored.
+//
+// Imported the full CRI package is a nightmare.
+message Options {
+  string type_url = 1;
+  string config_path = 2;
+}
diff --git a/pkg/shim/v2/service.go b/pkg/shim/v2/service.go
new file mode 100644
index 000000000..1534152fc
--- /dev/null
+++ b/pkg/shim/v2/service.go
@@ -0,0 +1,824 @@
+// 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 v2
+
+import (
+	"context"
+	"fmt"
+	"io/ioutil"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sync"
+	"syscall"
+	"time"
+
+	"github.com/BurntSushi/toml"
+	"github.com/containerd/cgroups"
+	"github.com/containerd/console"
+	"github.com/containerd/containerd/api/events"
+	"github.com/containerd/containerd/api/types/task"
+	"github.com/containerd/containerd/errdefs"
+	"github.com/containerd/containerd/log"
+	"github.com/containerd/containerd/mount"
+	"github.com/containerd/containerd/namespaces"
+	"github.com/containerd/containerd/pkg/process"
+	"github.com/containerd/containerd/pkg/stdio"
+	"github.com/containerd/containerd/runtime"
+	"github.com/containerd/containerd/runtime/linux/runctypes"
+	"github.com/containerd/containerd/runtime/v2/shim"
+	taskAPI "github.com/containerd/containerd/runtime/v2/task"
+	"github.com/containerd/containerd/sys/reaper"
+	"github.com/containerd/typeurl"
+	"github.com/gogo/protobuf/types"
+	"golang.org/x/sys/unix"
+
+	"gvisor.dev/gvisor/pkg/shim/runsc"
+	"gvisor.dev/gvisor/pkg/shim/v1/proc"
+	"gvisor.dev/gvisor/pkg/shim/v1/utils"
+	"gvisor.dev/gvisor/pkg/shim/v2/options"
+	"gvisor.dev/gvisor/pkg/shim/v2/runtimeoptions"
+	"gvisor.dev/gvisor/runsc/specutils"
+)
+
+var (
+	empty   = &types.Empty{}
+	bufPool = sync.Pool{
+		New: func() interface{} {
+			buffer := make([]byte, 32<<10)
+			return &buffer
+		},
+	}
+)
+
+var _ = (taskAPI.TaskService)(&service{})
+
+// configFile is the default config file name. For containerd 1.2,
+// we assume that a config.toml should exist in the runtime root.
+const configFile = "config.toml"
+
+// New returns a new shim service that can be used via GRPC.
+func New(ctx context.Context, id string, publisher shim.Publisher, cancel func()) (shim.Shim, error) {
+	ep, err := newOOMEpoller(publisher)
+	if err != nil {
+		return nil, err
+	}
+	go ep.run(ctx)
+	s := &service{
+		id:        id,
+		context:   ctx,
+		processes: make(map[string]process.Process),
+		events:    make(chan interface{}, 128),
+		ec:        proc.ExitCh,
+		oomPoller: ep,
+		cancel:    cancel,
+	}
+	go s.processExits()
+	runsc.Monitor = reaper.Default
+	if err := s.initPlatform(); err != nil {
+		cancel()
+		return nil, fmt.Errorf("failed to initialized platform behavior: %w", err)
+	}
+	go s.forward(publisher)
+	return s, nil
+}
+
+// service is the shim implementation of a remote shim over GRPC.
+type service struct {
+	mu sync.Mutex
+
+	context   context.Context
+	task      process.Process
+	processes map[string]process.Process
+	events    chan interface{}
+	platform  stdio.Platform
+	opts      options.Options
+	ec        chan proc.Exit
+	oomPoller *epoller
+
+	id     string
+	bundle string
+	cancel func()
+}
+
+func newCommand(ctx context.Context, containerdBinary, containerdAddress string) (*exec.Cmd, error) {
+	ns, err := namespaces.NamespaceRequired(ctx)
+	if err != nil {
+		return nil, err
+	}
+	self, err := os.Executable()
+	if err != nil {
+		return nil, err
+	}
+	cwd, err := os.Getwd()
+	if err != nil {
+		return nil, err
+	}
+	args := []string{
+		"-namespace", ns,
+		"-address", containerdAddress,
+		"-publish-binary", containerdBinary,
+	}
+	cmd := exec.Command(self, args...)
+	cmd.Dir = cwd
+	cmd.Env = append(os.Environ(), "GOMAXPROCS=2")
+	cmd.SysProcAttr = &syscall.SysProcAttr{
+		Setpgid: true,
+	}
+	return cmd, nil
+}
+
+func (s *service) StartShim(ctx context.Context, id, containerdBinary, containerdAddress, containerdTTRPCAddress string) (string, error) {
+	cmd, err := newCommand(ctx, containerdBinary, containerdAddress)
+	if err != nil {
+		return "", err
+	}
+	address, err := shim.SocketAddress(ctx, id)
+	if err != nil {
+		return "", err
+	}
+	socket, err := shim.NewSocket(address)
+	if err != nil {
+		return "", err
+	}
+	defer socket.Close()
+	f, err := socket.File()
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+
+	cmd.ExtraFiles = append(cmd.ExtraFiles, f)
+
+	if err := cmd.Start(); err != nil {
+		return "", err
+	}
+	defer func() {
+		if err != nil {
+			cmd.Process.Kill()
+		}
+	}()
+	// make sure to wait after start
+	go cmd.Wait()
+	if err := shim.WritePidFile("shim.pid", cmd.Process.Pid); err != nil {
+		return "", err
+	}
+	if err := shim.WriteAddress("address", address); err != nil {
+		return "", err
+	}
+	if err := shim.SetScore(cmd.Process.Pid); err != nil {
+		return "", fmt.Errorf("failed to set OOM Score on shim: %w", err)
+	}
+	return address, nil
+}
+
+func (s *service) Cleanup(ctx context.Context) (*taskAPI.DeleteResponse, error) {
+	path, err := os.Getwd()
+	if err != nil {
+		return nil, err
+	}
+	ns, err := namespaces.NamespaceRequired(ctx)
+	if err != nil {
+		return nil, err
+	}
+	runtime, err := s.readRuntime(path)
+	if err != nil {
+		return nil, err
+	}
+	r := proc.NewRunsc(s.opts.Root, path, ns, runtime, nil)
+	if err := r.Delete(ctx, s.id, &runsc.DeleteOpts{
+		Force: true,
+	}); err != nil {
+		log.L.Printf("failed to remove runc container: %v", err)
+	}
+	if err := mount.UnmountAll(filepath.Join(path, "rootfs"), 0); err != nil {
+		log.L.Printf("failed to cleanup rootfs mount: %v", err)
+	}
+	return &taskAPI.DeleteResponse{
+		ExitedAt:   time.Now(),
+		ExitStatus: 128 + uint32(unix.SIGKILL),
+	}, nil
+}
+
+func (s *service) readRuntime(path string) (string, error) {
+	data, err := ioutil.ReadFile(filepath.Join(path, "runtime"))
+	if err != nil {
+		return "", err
+	}
+	return string(data), nil
+}
+
+func (s *service) writeRuntime(path, runtime string) error {
+	return ioutil.WriteFile(filepath.Join(path, "runtime"), []byte(runtime), 0600)
+}
+
+// Create creates a new initial process and container with the underlying OCI
+// runtime.
+func (s *service) Create(ctx context.Context, r *taskAPI.CreateTaskRequest) (_ *taskAPI.CreateTaskResponse, err error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	ns, err := namespaces.NamespaceRequired(ctx)
+	if err != nil {
+		return nil, fmt.Errorf("create namespace: %w", err)
+	}
+
+	// Read from root for now.
+	var opts options.Options
+	if r.Options != nil {
+		v, err := typeurl.UnmarshalAny(r.Options)
+		if err != nil {
+			return nil, err
+		}
+		var path string
+		switch o := v.(type) {
+		case *runctypes.CreateOptions: // containerd 1.2.x
+			opts.IoUid = o.IoUid
+			opts.IoGid = o.IoGid
+			opts.ShimCgroup = o.ShimCgroup
+		case *runctypes.RuncOptions: // containerd 1.2.x
+			root := proc.RunscRoot
+			if o.RuntimeRoot != "" {
+				root = o.RuntimeRoot
+			}
+
+			opts.BinaryName = o.Runtime
+
+			path = filepath.Join(root, configFile)
+			if _, err := os.Stat(path); err != nil {
+				if !os.IsNotExist(err) {
+					return nil, fmt.Errorf("stat config file %q: %w", path, err)
+				}
+				// A config file in runtime root is not required.
+				path = ""
+			}
+		case *runtimeoptions.Options: // containerd 1.3.x+
+			if o.ConfigPath == "" {
+				break
+			}
+			if o.TypeUrl != options.OptionType {
+				return nil, fmt.Errorf("unsupported option type %q", o.TypeUrl)
+			}
+			path = o.ConfigPath
+		default:
+			return nil, fmt.Errorf("unsupported option type %q", r.Options.TypeUrl)
+		}
+		if path != "" {
+			if _, err = toml.DecodeFile(path, &opts); err != nil {
+				return nil, fmt.Errorf("decode config file %q: %w", path, err)
+			}
+		}
+	}
+
+	var mounts []proc.Mount
+	for _, m := range r.Rootfs {
+		mounts = append(mounts, proc.Mount{
+			Type:    m.Type,
+			Source:  m.Source,
+			Target:  m.Target,
+			Options: m.Options,
+		})
+	}
+
+	rootfs := filepath.Join(r.Bundle, "rootfs")
+	if err := os.Mkdir(rootfs, 0711); err != nil && !os.IsExist(err) {
+		return nil, err
+	}
+
+	config := &proc.CreateConfig{
+		ID:       r.ID,
+		Bundle:   r.Bundle,
+		Runtime:  opts.BinaryName,
+		Rootfs:   mounts,
+		Terminal: r.Terminal,
+		Stdin:    r.Stdin,
+		Stdout:   r.Stdout,
+		Stderr:   r.Stderr,
+		Options:  r.Options,
+	}
+	if err := s.writeRuntime(r.Bundle, opts.BinaryName); err != nil {
+		return nil, err
+	}
+	defer func() {
+		if err != nil {
+			if err := mount.UnmountAll(rootfs, 0); err != nil {
+				log.L.Printf("failed to cleanup rootfs mount: %v", err)
+			}
+		}
+	}()
+	for _, rm := range mounts {
+		m := &mount.Mount{
+			Type:    rm.Type,
+			Source:  rm.Source,
+			Options: rm.Options,
+		}
+		if err := m.Mount(rootfs); err != nil {
+			return nil, fmt.Errorf("failed to mount rootfs component %v: %w", m, err)
+		}
+	}
+	process, err := newInit(
+		ctx,
+		r.Bundle,
+		filepath.Join(r.Bundle, "work"),
+		ns,
+		s.platform,
+		config,
+		&opts,
+		rootfs,
+	)
+	if err != nil {
+		return nil, errdefs.ToGRPC(err)
+	}
+	if err := process.Create(ctx, config); err != nil {
+		return nil, errdefs.ToGRPC(err)
+	}
+	// Save the main task id and bundle to the shim for additional
+	// requests.
+	s.id = r.ID
+	s.bundle = r.Bundle
+
+	// Set up OOM notification on the sandbox's cgroup. This is done on
+	// sandbox create since the sandbox process will be created here.
+	pid := process.Pid()
+	if pid > 0 {
+		cg, err := cgroups.Load(cgroups.V1, cgroups.PidPath(pid))
+		if err != nil {
+			return nil, fmt.Errorf("loading cgroup for %d: %w", pid, err)
+		}
+		if err := s.oomPoller.add(s.id, cg); err != nil {
+			return nil, fmt.Errorf("add cg to OOM monitor: %w", err)
+		}
+	}
+	s.task = process
+	s.opts = opts
+	return &taskAPI.CreateTaskResponse{
+		Pid: uint32(process.Pid()),
+	}, nil
+
+}
+
+// Start starts a process.
+func (s *service) Start(ctx context.Context, r *taskAPI.StartRequest) (*taskAPI.StartResponse, error) {
+	p, err := s.getProcess(r.ExecID)
+	if err != nil {
+		return nil, err
+	}
+	if err := p.Start(ctx); err != nil {
+		return nil, err
+	}
+	// TODO: Set the cgroup and oom notifications on restore.
+	// https://github.com/google/gvisor-containerd-shim/issues/58
+	return &taskAPI.StartResponse{
+		Pid: uint32(p.Pid()),
+	}, nil
+}
+
+// Delete deletes the initial process and container.
+func (s *service) Delete(ctx context.Context, r *taskAPI.DeleteRequest) (*taskAPI.DeleteResponse, error) {
+	p, err := s.getProcess(r.ExecID)
+	if err != nil {
+		return nil, err
+	}
+	if p == nil {
+		return nil, errdefs.ToGRPCf(errdefs.ErrFailedPrecondition, "container must be created")
+	}
+	if err := p.Delete(ctx); err != nil {
+		return nil, err
+	}
+	isTask := r.ExecID == ""
+	if !isTask {
+		s.mu.Lock()
+		delete(s.processes, r.ExecID)
+		s.mu.Unlock()
+	}
+	if isTask && s.platform != nil {
+		s.platform.Close()
+	}
+	return &taskAPI.DeleteResponse{
+		ExitStatus: uint32(p.ExitStatus()),
+		ExitedAt:   p.ExitedAt(),
+		Pid:        uint32(p.Pid()),
+	}, nil
+}
+
+// Exec spawns an additional process inside the container.
+func (s *service) Exec(ctx context.Context, r *taskAPI.ExecProcessRequest) (*types.Empty, error) {
+	s.mu.Lock()
+	p := s.processes[r.ExecID]
+	s.mu.Unlock()
+	if p != nil {
+		return nil, errdefs.ToGRPCf(errdefs.ErrAlreadyExists, "id %s", r.ExecID)
+	}
+	p = s.task
+	if p == nil {
+		return nil, errdefs.ToGRPCf(errdefs.ErrFailedPrecondition, "container must be created")
+	}
+	process, err := p.(*proc.Init).Exec(ctx, s.bundle, &proc.ExecConfig{
+		ID:       r.ExecID,
+		Terminal: r.Terminal,
+		Stdin:    r.Stdin,
+		Stdout:   r.Stdout,
+		Stderr:   r.Stderr,
+		Spec:     r.Spec,
+	})
+	if err != nil {
+		return nil, errdefs.ToGRPC(err)
+	}
+	s.mu.Lock()
+	s.processes[r.ExecID] = process
+	s.mu.Unlock()
+	return empty, nil
+}
+
+// ResizePty resizes the terminal of a process.
+func (s *service) ResizePty(ctx context.Context, r *taskAPI.ResizePtyRequest) (*types.Empty, error) {
+	p, err := s.getProcess(r.ExecID)
+	if err != nil {
+		return nil, err
+	}
+	ws := console.WinSize{
+		Width:  uint16(r.Width),
+		Height: uint16(r.Height),
+	}
+	if err := p.Resize(ws); err != nil {
+		return nil, errdefs.ToGRPC(err)
+	}
+	return empty, nil
+}
+
+// State returns runtime state information for a process.
+func (s *service) State(ctx context.Context, r *taskAPI.StateRequest) (*taskAPI.StateResponse, error) {
+	p, err := s.getProcess(r.ExecID)
+	if err != nil {
+		return nil, err
+	}
+	st, err := p.Status(ctx)
+	if err != nil {
+		return nil, err
+	}
+	status := task.StatusUnknown
+	switch st {
+	case "created":
+		status = task.StatusCreated
+	case "running":
+		status = task.StatusRunning
+	case "stopped":
+		status = task.StatusStopped
+	}
+	sio := p.Stdio()
+	return &taskAPI.StateResponse{
+		ID:         p.ID(),
+		Bundle:     s.bundle,
+		Pid:        uint32(p.Pid()),
+		Status:     status,
+		Stdin:      sio.Stdin,
+		Stdout:     sio.Stdout,
+		Stderr:     sio.Stderr,
+		Terminal:   sio.Terminal,
+		ExitStatus: uint32(p.ExitStatus()),
+		ExitedAt:   p.ExitedAt(),
+	}, nil
+}
+
+// Pause the container.
+func (s *service) Pause(ctx context.Context, r *taskAPI.PauseRequest) (*types.Empty, error) {
+	return empty, errdefs.ToGRPC(errdefs.ErrNotImplemented)
+}
+
+// Resume the container.
+func (s *service) Resume(ctx context.Context, r *taskAPI.ResumeRequest) (*types.Empty, error) {
+	return empty, errdefs.ToGRPC(errdefs.ErrNotImplemented)
+}
+
+// Kill a process with the provided signal.
+func (s *service) Kill(ctx context.Context, r *taskAPI.KillRequest) (*types.Empty, error) {
+	p, err := s.getProcess(r.ExecID)
+	if err != nil {
+		return nil, err
+	}
+	if p == nil {
+		return nil, errdefs.ToGRPCf(errdefs.ErrFailedPrecondition, "container must be created")
+	}
+	if err := p.Kill(ctx, r.Signal, r.All); err != nil {
+		return nil, errdefs.ToGRPC(err)
+	}
+	return empty, nil
+}
+
+// Pids returns all pids inside the container.
+func (s *service) Pids(ctx context.Context, r *taskAPI.PidsRequest) (*taskAPI.PidsResponse, error) {
+	pids, err := s.getContainerPids(ctx, r.ID)
+	if err != nil {
+		return nil, errdefs.ToGRPC(err)
+	}
+	var processes []*task.ProcessInfo
+	for _, pid := range pids {
+		pInfo := task.ProcessInfo{
+			Pid: pid,
+		}
+		for _, p := range s.processes {
+			if p.Pid() == int(pid) {
+				d := &runctypes.ProcessDetails{
+					ExecID: p.ID(),
+				}
+				a, err := typeurl.MarshalAny(d)
+				if err != nil {
+					return nil, fmt.Errorf("failed to marshal process %d info: %w", pid, err)
+				}
+				pInfo.Info = a
+				break
+			}
+		}
+		processes = append(processes, &pInfo)
+	}
+	return &taskAPI.PidsResponse{
+		Processes: processes,
+	}, nil
+}
+
+// CloseIO closes the I/O context of a process.
+func (s *service) CloseIO(ctx context.Context, r *taskAPI.CloseIORequest) (*types.Empty, error) {
+	p, err := s.getProcess(r.ExecID)
+	if err != nil {
+		return nil, err
+	}
+	if stdin := p.Stdin(); stdin != nil {
+		if err := stdin.Close(); err != nil {
+			return nil, fmt.Errorf("close stdin: %w", err)
+		}
+	}
+	return empty, nil
+}
+
+// Checkpoint checkpoints the container.
+func (s *service) Checkpoint(ctx context.Context, r *taskAPI.CheckpointTaskRequest) (*types.Empty, error) {
+	return empty, errdefs.ToGRPC(errdefs.ErrNotImplemented)
+}
+
+// Connect returns shim information such as the shim's pid.
+func (s *service) Connect(ctx context.Context, r *taskAPI.ConnectRequest) (*taskAPI.ConnectResponse, error) {
+	var pid int
+	if s.task != nil {
+		pid = s.task.Pid()
+	}
+	return &taskAPI.ConnectResponse{
+		ShimPid: uint32(os.Getpid()),
+		TaskPid: uint32(pid),
+	}, nil
+}
+
+func (s *service) Shutdown(ctx context.Context, r *taskAPI.ShutdownRequest) (*types.Empty, error) {
+	s.cancel()
+	os.Exit(0)
+	return empty, nil
+}
+
+func (s *service) Stats(ctx context.Context, r *taskAPI.StatsRequest) (*taskAPI.StatsResponse, error) {
+	path, err := os.Getwd()
+	if err != nil {
+		return nil, err
+	}
+	ns, err := namespaces.NamespaceRequired(ctx)
+	if err != nil {
+		return nil, err
+	}
+	runtime, err := s.readRuntime(path)
+	if err != nil {
+		return nil, err
+	}
+	rs := proc.NewRunsc(s.opts.Root, path, ns, runtime, nil)
+	stats, err := rs.Stats(ctx, s.id)
+	if err != nil {
+		return nil, err
+	}
+
+	// gvisor currently (as of 2020-03-03) only returns the total memory
+	// usage and current PID value[0]. However, we copy the common fields here
+	// so that future updates will propagate correct information.  We're
+	// using the cgroups.Metrics structure so we're returning the same type
+	// as runc.
+	//
+	// [0]: https://github.com/google/gvisor/blob/277a0d5a1fbe8272d4729c01ee4c6e374d047ebc/runsc/boot/events.go#L61-L81
+	data, err := typeurl.MarshalAny(&cgroups.Metrics{
+		CPU: &cgroups.CPUStat{
+			Usage: &cgroups.CPUUsage{
+				Total:  stats.Cpu.Usage.Total,
+				Kernel: stats.Cpu.Usage.Kernel,
+				User:   stats.Cpu.Usage.User,
+				PerCPU: stats.Cpu.Usage.Percpu,
+			},
+			Throttling: &cgroups.Throttle{
+				Periods:          stats.Cpu.Throttling.Periods,
+				ThrottledPeriods: stats.Cpu.Throttling.ThrottledPeriods,
+				ThrottledTime:    stats.Cpu.Throttling.ThrottledTime,
+			},
+		},
+		Memory: &cgroups.MemoryStat{
+			Cache: stats.Memory.Cache,
+			Usage: &cgroups.MemoryEntry{
+				Limit:   stats.Memory.Usage.Limit,
+				Usage:   stats.Memory.Usage.Usage,
+				Max:     stats.Memory.Usage.Max,
+				Failcnt: stats.Memory.Usage.Failcnt,
+			},
+			Swap: &cgroups.MemoryEntry{
+				Limit:   stats.Memory.Swap.Limit,
+				Usage:   stats.Memory.Swap.Usage,
+				Max:     stats.Memory.Swap.Max,
+				Failcnt: stats.Memory.Swap.Failcnt,
+			},
+			Kernel: &cgroups.MemoryEntry{
+				Limit:   stats.Memory.Kernel.Limit,
+				Usage:   stats.Memory.Kernel.Usage,
+				Max:     stats.Memory.Kernel.Max,
+				Failcnt: stats.Memory.Kernel.Failcnt,
+			},
+			KernelTCP: &cgroups.MemoryEntry{
+				Limit:   stats.Memory.KernelTCP.Limit,
+				Usage:   stats.Memory.KernelTCP.Usage,
+				Max:     stats.Memory.KernelTCP.Max,
+				Failcnt: stats.Memory.KernelTCP.Failcnt,
+			},
+		},
+		Pids: &cgroups.PidsStat{
+			Current: stats.Pids.Current,
+			Limit:   stats.Pids.Limit,
+		},
+	})
+	if err != nil {
+		return nil, err
+	}
+	return &taskAPI.StatsResponse{
+		Stats: data,
+	}, nil
+}
+
+// Update updates a running container.
+func (s *service) Update(ctx context.Context, r *taskAPI.UpdateTaskRequest) (*types.Empty, error) {
+	return empty, errdefs.ToGRPC(errdefs.ErrNotImplemented)
+}
+
+// Wait waits for a process to exit.
+func (s *service) Wait(ctx context.Context, r *taskAPI.WaitRequest) (*taskAPI.WaitResponse, error) {
+	p, err := s.getProcess(r.ExecID)
+	if err != nil {
+		return nil, err
+	}
+	if p == nil {
+		return nil, errdefs.ToGRPCf(errdefs.ErrFailedPrecondition, "container must be created")
+	}
+	p.Wait()
+
+	return &taskAPI.WaitResponse{
+		ExitStatus: uint32(p.ExitStatus()),
+		ExitedAt:   p.ExitedAt(),
+	}, nil
+}
+
+func (s *service) processExits() {
+	for e := range s.ec {
+		s.checkProcesses(e)
+	}
+}
+
+func (s *service) checkProcesses(e proc.Exit) {
+	// TODO(random-liu): Add `shouldKillAll` logic if container pid
+	// namespace is supported.
+	for _, p := range s.allProcesses() {
+		if p.ID() == e.ID {
+			if ip, ok := p.(*proc.Init); ok {
+				// Ensure all children are killed.
+				if err := ip.KillAll(s.context); err != nil {
+					log.G(s.context).WithError(err).WithField("id", ip.ID()).
+						Error("failed to kill init's children")
+				}
+			}
+			p.SetExited(e.Status)
+			s.events <- &events.TaskExit{
+				ContainerID: s.id,
+				ID:          p.ID(),
+				Pid:         uint32(p.Pid()),
+				ExitStatus:  uint32(e.Status),
+				ExitedAt:    p.ExitedAt(),
+			}
+			return
+		}
+	}
+}
+
+func (s *service) allProcesses() (o []process.Process) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for _, p := range s.processes {
+		o = append(o, p)
+	}
+	if s.task != nil {
+		o = append(o, s.task)
+	}
+	return o
+}
+
+func (s *service) getContainerPids(ctx context.Context, id string) ([]uint32, error) {
+	s.mu.Lock()
+	p := s.task
+	s.mu.Unlock()
+	if p == nil {
+		return nil, fmt.Errorf("container must be created: %w", errdefs.ErrFailedPrecondition)
+	}
+	ps, err := p.(*proc.Init).Runtime().Ps(ctx, id)
+	if err != nil {
+		return nil, err
+	}
+	pids := make([]uint32, 0, len(ps))
+	for _, pid := range ps {
+		pids = append(pids, uint32(pid))
+	}
+	return pids, nil
+}
+
+func (s *service) forward(publisher shim.Publisher) {
+	for e := range s.events {
+		ctx, cancel := context.WithTimeout(s.context, 5*time.Second)
+		err := publisher.Publish(ctx, getTopic(e), e)
+		cancel()
+		if err != nil {
+			// Should not happen.
+			panic(fmt.Errorf("post event: %w", err))
+		}
+	}
+}
+
+func (s *service) getProcess(execID string) (process.Process, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if execID == "" {
+		return s.task, nil
+	}
+	p := s.processes[execID]
+	if p == nil {
+		return nil, errdefs.ToGRPCf(errdefs.ErrNotFound, "process does not exist %s", execID)
+	}
+	return p, nil
+}
+
+func getTopic(e interface{}) string {
+	switch e.(type) {
+	case *events.TaskCreate:
+		return runtime.TaskCreateEventTopic
+	case *events.TaskStart:
+		return runtime.TaskStartEventTopic
+	case *events.TaskOOM:
+		return runtime.TaskOOMEventTopic
+	case *events.TaskExit:
+		return runtime.TaskExitEventTopic
+	case *events.TaskDelete:
+		return runtime.TaskDeleteEventTopic
+	case *events.TaskExecAdded:
+		return runtime.TaskExecAddedEventTopic
+	case *events.TaskExecStarted:
+		return runtime.TaskExecStartedEventTopic
+	default:
+		log.L.Printf("no topic for type %#v", e)
+	}
+	return runtime.TaskUnknownTopic
+}
+
+func newInit(ctx context.Context, path, workDir, namespace string, platform stdio.Platform, r *proc.CreateConfig, options *options.Options, rootfs string) (*proc.Init, error) {
+	spec, err := utils.ReadSpec(r.Bundle)
+	if err != nil {
+		return nil, fmt.Errorf("read oci spec: %w", err)
+	}
+	if err := utils.UpdateVolumeAnnotations(r.Bundle, spec); err != nil {
+		return nil, fmt.Errorf("update volume annotations: %w", err)
+	}
+	runsc.FormatLogPath(r.ID, options.RunscConfig)
+	runtime := proc.NewRunsc(options.Root, path, namespace, options.BinaryName, options.RunscConfig)
+	p := proc.New(r.ID, runtime, stdio.Stdio{
+		Stdin:    r.Stdin,
+		Stdout:   r.Stdout,
+		Stderr:   r.Stderr,
+		Terminal: r.Terminal,
+	})
+	p.Bundle = r.Bundle
+	p.Platform = platform
+	p.Rootfs = rootfs
+	p.WorkDir = workDir
+	p.IoUID = int(options.IoUid)
+	p.IoGID = int(options.IoGid)
+	p.Sandbox = specutils.SpecContainerType(spec) == specutils.ContainerTypeSandbox
+	p.UserLog = utils.UserLogPath(spec)
+	p.Monitor = reaper.Default
+	return p, nil
+}
diff --git a/pkg/shim/v2/service_linux.go b/pkg/shim/v2/service_linux.go
new file mode 100644
index 000000000..1800ab90b
--- /dev/null
+++ b/pkg/shim/v2/service_linux.go
@@ -0,0 +1,108 @@
+// 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"
+	"fmt"
+	"io"
+	"sync"
+	"syscall"
+
+	"github.com/containerd/console"
+	"github.com/containerd/fifo"
+)
+
+type linuxPlatform struct {
+	epoller *console.Epoller
+}
+
+func (p *linuxPlatform) CopyConsole(ctx context.Context, console console.Console, stdin, stdout, stderr string, wg *sync.WaitGroup) (console.Console, error) {
+	if p.epoller == nil {
+		return nil, fmt.Errorf("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
+		}
+		go func() {
+			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)
+	go func() {
+		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 fmt.Errorf("uninitialized epoller")
+	}
+	epollConsole, ok := cons.(*console.EpollConsole)
+	if !ok {
+		return fmt.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 fmt.Errorf("failed to initialize epoller: %w", err)
+	}
+	s.platform = &linuxPlatform{
+		epoller: epoller,
+	}
+	go epoller.Wait()
+	return nil
+}