Check if the container is running if no event

When there is no event for the container it can happen because of a
crash and the container state on the persistent disk will have a
mismatch between what was in `/run` ( machine crash ).

This situation will create an unkillable container in docker because
containerd does not see it and it is not running but docker thinks it is
and you cannot tell it anything different.

This fixes the issue by checking if containerd has the container running
if we do not have an event instead of just returning.

Signed-off-by: Michael Crosby <crosbymichael@gmail.com>

1 files changed, 13 insertions, 5 deletions

diff --git a/libcontainerd/client_linux.go b/libcontainerd/client_linux.go
index afb68b4a45..99d299f555 100644
--- a/libcontainerd/client_linux.go
+++ b/libcontainerd/client_linux.go
@@ -505,12 +505,20 @@ func (clnt *client) Restore(containerID string, options ...CreateOption) error {
 			return err
 		}
 
-		// If ev is nil, then we already consumed all the event of the
-		// container, included the "exit" one.
-		// Thus we return to avoid overriding the Exit Code.
 		if ev == nil {
-			logrus.Warnf("libcontainerd: restore was called on a fully synced container (%s)", containerID)
-			return nil
+			if _, err := clnt.getContainer(containerID); err == nil {
+				// If ev is nil and the container is running in containerd,
+				// we already consumed all the event of the
+				// container, included the "exit" one.
+				// Thus we return to avoid overriding the Exit Code.
+				logrus.Warnf("libcontainerd: restore was called on a fully synced container (%s)", containerID)
+				return nil
+			}
+			// the container is not running so we need to fix the state within docker
+			ev = &containerd.Event{
+				Type:   StateExit,
+				Status: 1,
+			}
 		}
 
 		// get the exit status for this container