Add fs.NotifyOnOOM for out-of-memory notifications

Docker-DCO-1.1-Signed-off-by: Bernerd Schaefer <bj.schaefer@gmail.com> (github: bernerdschaefer)

cgroups/fs/notify_linux.go

@@ -0,0 +1,81 @@
+// +build linux
+
+package fs
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"syscall"
+
+	"github.com/docker/libcontainer/cgroups"
+)
+
+// NotifyOnOOM sends signals on the returned channel when the cgroup reaches
+// its memory limit. The channel is closed when the cgroup is removed.
+func NotifyOnOOM(c *cgroups.Cgroup) (chan struct{}, error) {
+	d, err := getCgroupData(c, 0)
+	if err != nil {
+		return nil, err
+	}
+
+	return notifyOnOOM(d)
+}
+
+func notifyOnOOM(d *data) (chan struct{}, error) {
+	dir, err := d.path("memory")
+	if err != nil {
+		return nil, err
+	}
+
+	fd, _, syserr := syscall.RawSyscall(syscall.SYS_EVENTFD2, 0, syscall.FD_CLOEXEC, 0)
+	if syserr != 0 {
+		return nil, syserr
+	}
+
+	eventfd := os.NewFile(fd, "eventfd")
+
+	oomControl, err := os.Open(filepath.Join(dir, "memory.oom_control"))
+	if err != nil {
+		eventfd.Close()
+		return nil, err
+	}
+
+	var (
+		eventControlPath = filepath.Join(dir, "cgroup.event_control")
+
+		data = fmt.Sprintf("%d %d", eventfd.Fd(), oomControl.Fd())
+	)
+
+	if err := writeFile(dir, "cgroup.event_control", data); err != nil {
+		eventfd.Close()
+		oomControl.Close()
+		return nil, err
+	}
+
+	ch := make(chan struct{})
+
+	go func() {
+		defer close(ch)
+		defer eventfd.Close()
+		defer oomControl.Close()
+
+		var buf = make([]byte, 8)
+
+		for {
+			if _, err := eventfd.Read(buf); err != nil {
+				return
+			}
+
+			// When a cgroup is destroyed, an event is sent to eventfd.
+			// So if the control path is gone, return instead of notifying.
+			if _, err := os.Lstat(eventControlPath); os.IsNotExist(err) {
+				return
+			}
+
+			ch <- struct{}{}
+		}
+	}()
+
+	return ch, nil
+}

cgroups/fs/notify_linux_test.go

@@ -0,0 +1,86 @@
+// +build linux
+
+package fs
+
+import (
+	"encoding/binary"
+	"fmt"
+	"syscall"
+	"testing"
+	"time"
+)
+
+func TestNotifyOnOOM(t *testing.T) {
+	helper := NewCgroupTestUtil("memory", t)
+	defer helper.cleanup()
+
+	helper.writeFileContents(map[string]string{
+		"memory.oom_control":   "",
+		"cgroup.event_control": "",
+	})
+
+	var eventFd, oomControlFd int
+
+	ooms, err := notifyOnOOM(helper.CgroupData)
+	if err != nil {
+		t.Fatal("expected no error, got:", err)
+	}
+
+	memoryPath, _ := helper.CgroupData.path("memory")
+	data, err := readFile(memoryPath, "cgroup.event_control")
+	if err != nil {
+		t.Fatal("couldn't read event control file:", err)
+	}
+
+	if _, err := fmt.Sscanf(data, "%d %d", &eventFd, &oomControlFd); err != nil {
+		t.Fatalf("invalid control data %q: %s", data, err)
+	}
+
+	// re-open the eventfd
+	efd, err := syscall.Dup(eventFd)
+	if err != nil {
+		t.Fatal("unable to reopen eventfd:", err)
+	}
+	defer syscall.Close(efd)
+
+	if err != nil {
+		t.Fatal("unable to dup event fd:", err)
+	}
+
+	buf := make([]byte, 8)
+	binary.LittleEndian.PutUint64(buf, 1)
+
+	if _, err := syscall.Write(efd, buf); err != nil {
+		t.Fatal("unable to write to eventfd:", err)
+	}
+
+	select {
+	case <-ooms:
+	case <-time.After(100 * time.Millisecond):
+		t.Fatal("no notification on oom channel after 100ms")
+	}
+
+	// simulate what happens when a cgroup is destroyed by cleaning up and then
+	// writing to the eventfd.
+	helper.cleanup()
+	if _, err := syscall.Write(efd, buf); err != nil {
+		t.Fatal("unable to write to eventfd:", err)
+	}
+
+	// give things a moment to shut down
+	select {
+	case _, ok := <-ooms:
+		if ok {
+			t.Fatal("expected no oom to be triggered")
+		}
+	case <-time.After(100 * time.Millisecond):
+	}
+
+	if _, _, err := syscall.Syscall(syscall.SYS_FCNTL, uintptr(oomControlFd), syscall.F_GETFD, 0); err != syscall.EBADF {
+		t.Error("expected oom control to be closed")
+	}
+
+	if _, _, err := syscall.Syscall(syscall.SYS_FCNTL, uintptr(eventFd), syscall.F_GETFD, 0); err != syscall.EBADF {
+		t.Error("expected event fd to be closed")
+	}
+}