sync race issue in nsexec.c

[hqhq]

When I use the latest runc with docker, got a lot of failure in integration tests, they all failed with exit status 25, full output looks like:

FAIL: docker_cli_run_test.go:3631: DockerSuite.TestRunAbsoluteInvalidCgroupParent

docker_cli_run_test.go:3642:
    // XXX: This may include a daemon crash.
    c.Fatalf("unexpected failure when running container with --cgroup-parent option - %s\n%v", string(out), err)
... Error: unexpected failure when running container with --cgroup-parent option - /go/src/github.com/docker/docker/bundles/1.11.2/test-integration-cli/../dynbinary/docker: Error response from daemon: rpc error: code = 2 desc = "oci runtime error: container_linux.go:247: starting container process caused \"process_linux.go:272: running exec setns process for init caused \\\"exit status 25\\\"\"\n".

exit status 125

To be explicit, we failed at: https://github.com/opencontainers/runc/blob/master/libcontainer/nsenter/nsexec.c#L579-L581

The code was added by: #1154
But I don't think it introduced this issue, it's probably because we are using the same socketpair to sync with child and grandchild, so there is race condition that parent read data in the wrong sequence.

[hqhq]

ping @cyphar , I'm working on a fix now.