mds: we should wait messenger when MDSDaemon suicide #6996
Conversation
|
msgr->wait() is called from ceph_mds.cc -- I don't think you can wait() in suicide() because you might call suicide() while handling a message (i.e. you are inside the messenger handling code, so would deadlock). Unless I'm missing something? What case is this meant to handle? |
|
MonClient will failed to authenticate when mds miss keying, then MDSDaemon will suicide and MDSDaemon::init will return (r < 0), ceph_mds.cc will goto shutdown, msgr->wait() is't called before delete. |
|
OK, so the correct thing in that case is probably to call msgr->wait() from the error handling branch in main() |
|
Can you review this? I add msgr->wait() in main() before msgr be deleted. |
| @@ -244,6 +244,7 @@ int main(int argc, const char **argv) | |||
| // detection, etc.). don't bother if it was a suicide. | |||
| if (mds->is_clean_shutdown()) { | |||
| delete mds; | |||
| msgr->wait(); | |||
There was a problem hiding this comment.
For the case you're dealing with, you would want to add the msgr->wait() up at line 216, before "goto shutdown". Adding it here doesn't work (is_clean_shutdown is something completely different)
Signed-off-by: Wei Feng <feng.wei@h3c.com>
|
Can you review this? |
mds: we should wait messenger when MDSDaemon suicide Reviewed-by: John Spray <john.spray@redhat.com>
Signed-off-by: Wei Feng feng.wei@h3c.com