libct: use pidfd and epoll to wait the init process exit

Signed-off-by: lifubang <[email protected]>

Author: lifubang

delete.go

@@ -5,23 +5,16 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
-	"time"
 
 	"github.com/opencontainers/runc/libcontainer"
 	"github.com/urfave/cli"
-
-	"golang.org/x/sys/unix"
 )
 
-func killContainer(container *libcontainer.Container) error {
-	_ = container.Signal(unix.SIGKILL)
-	for i := 0; i < 100; i++ {
-		time.Sleep(100 * time.Millisecond)
-		if err := container.Signal(unix.Signal(0)); err != nil {
-			return container.Destroy()
-		}
+func killAndDestroy(container *libcontainer.Container) error {
+	if err := container.EnsureKilled(); err != nil {
+		return err
 	}
-	return errors.New("container init still running")
+	return container.Destroy()
 }
 
 var deleteCommand = cli.Command{
@@ -71,7 +64,7 @@ status of "ubuntu01" as "stopped" the following will delete resources held for
 		// namespace) there may be some leftover processes in the
 		// container's cgroup.
 		if force {
-			return killContainer(container)
+			return killAndDestroy(container)
 		}
 		s, err := container.Status()
 		if err != nil {
@@ -81,7 +74,7 @@ status of "ubuntu01" as "stopped" the following will delete resources held for
 		case libcontainer.Stopped:
 			return container.Destroy()
 		case libcontainer.Created:
-			return killContainer(container)
+			return killAndDestroy(container)
 		default:
 			return fmt.Errorf("cannot delete container %s that is not stopped: %s", id, s)
 		}

libcontainer/README.md

@@ -230,6 +230,9 @@ container.Resume()
 // send signal to container's init process.
 container.Signal(signal)
 
+// send signal to container's init process and waits for the kernel to finish killing it.
+container.EnsureKilled()
+
 // update container resource constraints.
 container.Set(config)
 

libcontainer/container_linux.go

@@ -377,9 +377,13 @@ func (c *Container) start(process *Process) (retErr error) {
 
 // Signal sends a specified signal to container's init.
 //
-// When s is SIGKILL and the container does not have its own PID namespace, all
-// the container's processes are killed. In this scenario, the libcontainer
+// When s is SIGKILL:
+// 1. If the container does not have its own PID namespace, all the
+// container's processes are killed. In this scenario, the libcontainer
 // user may be required to implement a proper child reaper.
+// 2. Otherwise, we just send the SIGKILL signal to the init process,
+// but we don't wait for the init process to disappear. If you want to
+// wait, please use c.KillAndWaitExit instead.
 func (c *Container) Signal(s os.Signal) error {
 	c.m.Lock()
 	defer c.m.Unlock()
@@ -431,6 +435,81 @@ func (c *Container) signal(s os.Signal) error {
 	return nil
 }
 
+func (c *Container) killViaPidfd() error {
+	pidfd, err := unix.PidfdOpen(c.initProcess.pid(), 0)
+	if err != nil {
+		return err
+	}
+	defer unix.Close(pidfd)
+
+	epollfd, err := unix.EpollCreate1(unix.EPOLL_CLOEXEC)
+	if err != nil {
+		return err
+	}
+	defer unix.Close(epollfd)
+
+	event := unix.EpollEvent{
+		Events: unix.EPOLLIN,
+		Fd:     int32(pidfd),
+	}
+	if err := unix.EpollCtl(epollfd, unix.EPOLL_CTL_ADD, pidfd, &event); err != nil {
+		return err
+	}
+
+	if err := unix.PidfdSendSignal(pidfd, unix.SIGKILL, nil, 0); err != nil {
+		return err
+	}
+
+	events := make([]unix.EpollEvent, 1)
+	for {
+		// Set the timeout to 10s, the same as in kill below.
+		n, err := unix.EpollWait(epollfd, events, 10000)
+		if err != nil {
+			if err == unix.EINTR {
+				continue
+			}
+			return err
+		}
+
+		if n == 0 {
+			return errors.New("container init still running")
+		}
+
+		if n > 0 {
+			event := events[0]
+			if event.Fd == int32(pidfd) {
+				return nil
+			}
+		}
+	}
+}
+
+func (c *Container) kill() error {
+	_ = c.Signal(unix.SIGKILL)
+	for i := 0; i < 100; i++ {
+		time.Sleep(100 * time.Millisecond)
+		if err := c.Signal(unix.Signal(0)); err != nil {
+			return nil
+		}
+	}
+	return errors.New("container init still running")
+}
+
+// EnsureKilled kills the container and waits for the kernel to finish killing it.
+func (c *Container) EnsureKilled() error {
+	// When a container doesn't have a private pidns, we have to kill all processes
+	// in the cgroup, it's more simpler to use `cgroup.kill` or `unix.Kill`.
+	if c.config.Namespaces.IsPrivate(configs.NEWPID) {
+		err := c.killViaPidfd()
+		if err == nil {
+			return nil
+		}
+
+		logrus.Debugf("pidfd & epoll failed, falling back to unix.Signal: %v", err)
+	}
+	return c.kill()
+}
+
 func (c *Container) createExecFifo() (retErr error) {
 	rootuid, err := c.config.HostRootUID()
 	if err != nil {