Set OOMKilled state on any OOM event

This restores the behavior that existed prior to #16235 for setting
OOMKilled, while retaining the additional benefits it introduced around
emitting the oom event.

This also adds a test for the most obvious OOM cases which would have
caught this regression.

Fixes #18510

Signed-off-by: Euan <euank@amazon.com>

container/monitor.go

@@ -304,8 +304,7 @@ func (m *containerMonitor) shouldRestart(exitCode int) bool {
 // received ack from the execution drivers
 func (m *containerMonitor) callback(processConfig *execdriver.ProcessConfig, pid int, chOOM <-chan struct{}) error {
 	go func() {
-		_, ok := <-chOOM
-		if ok {
+		for range chOOM {
 			m.logEvent("oom")
 		}
 	}()

daemon/execdriver/native/driver.go

@@ -171,7 +171,10 @@ func (d *Driver) Run(c *execdriver.Command, pipes *execdriver.Pipes, hooks execd
 		return execdriver.ExitStatus{ExitCode: -1}, err
 	}
 
+	// 'oom' is used to emit 'oom' events to the eventstream, 'oomKilled' is used
+	// to set the 'OOMKilled' flag in state
 	oom := notifyOnOOM(cont)
+	oomKilled := notifyOnOOM(cont)
 	if hooks.Start != nil {
 		pid, err := p.Pid()
 		if err != nil {
@@ -198,7 +201,21 @@ func (d *Driver) Run(c *execdriver.Command, pipes *execdriver.Pipes, hooks execd
 	}
 	cont.Destroy()
 	destroyed = true
-	_, oomKill := <-oom
+	// oomKilled will have an oom event if any process within the container was
+	// OOM killed at any time, not only if the init process OOMed.
+	//
+	// Perhaps we only want the OOMKilled flag to be set if the OOM
+	// resulted in a container death, but there isn't a good way to do this
+	// because the kernel's cgroup oom notification does not provide information
+	// such as the PID. This could be heuristically done by checking that the OOM
+	// happened within some very small time slice for the container dying (and
+	// optionally exit-code 137), but I don't think the cgroup oom notification
+	// can be used to reliably determine this
+	//
+	// Even if there were multiple OOMs, it's sufficient to read one value
+	// because libcontainer's oom notify will discard the channel after the
+	// cgroup is destroyed
+	_, oomKill := <-oomKilled
 	return execdriver.ExitStatus{ExitCode: utils.ExitStatus(ps.Sys().(syscall.WaitStatus)), OOMKilled: oomKill}, nil
 }
 

integration-cli/docker_cli_oom_killed_test.go

@@ -0,0 +1,32 @@
+// +build !windows
+
+package main
+
+import (
+	"github.com/docker/docker/pkg/integration/checker"
+	"github.com/go-check/check"
+)
+
+func (s *DockerSuite) TestInspectOomKilledTrue(c *check.C) {
+	testRequires(c, DaemonIsLinux, memoryLimitSupport)
+
+	name := "testoomkilled"
+	_, exitCode, _ := dockerCmdWithError("run", "--name", name, "-m", "10MB", "busybox", "sh", "-c", "x=a; while true; do x=$x$x$x$x; done")
+
+	c.Assert(exitCode, checker.Equals, 137, check.Commentf("OOM exit should be 137"))
+
+	oomKilled, err := inspectField(name, "State.OOMKilled")
+	c.Assert(oomKilled, checker.Equals, "true")
+	c.Assert(err, checker.IsNil)
+}
+
+func (s *DockerSuite) TestInspectOomKilledFalse(c *check.C) {
+	testRequires(c, DaemonIsLinux, memoryLimitSupport)
+
+	name := "testoomkilled"
+	dockerCmd(c, "run", "--name", name, "-m", "10MB", "busybox", "sh", "-c", "echo hello world")
+
+	oomKilled, err := inspectField(name, "State.OOMKilled")
+	c.Assert(oomKilled, checker.Equals, "false")
+	c.Assert(err, checker.IsNil)
+}