Capture and publish OOM events (#57)

* Capture and publish OOM events

We use oom notifiers on the cgroup to publish oom events to containerd.
This is passed back via CRI to Kubernetes etc. for more helpful error
reporting.

Fixes #56

1 files changed, 129 insertions, 0 deletions

diff --git a/pkg/v2/epoll.go b/pkg/v2/epoll.go
new file mode 100644
index 000000000..76c7b54d6
--- /dev/null
+++ b/pkg/v2/epoll.go
@@ -0,0 +1,129 @@
+// +build linux
+
+/*
+   Copyright The containerd 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
+
+       http://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"
+	"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
+}