[rbd-mirror] Mirror daemon never recovers from being blacklisted
I can reproduce this rather reliably by:
- Restarting many OSDs (old nodes, slow spinning disks, likely exceeding default blacklist timeout).
- Sometimes, it also happens when restarting other RBD mirror daemons (we have 3).
The attached log is extracted from one blacklisted RBD mirror unable to recover at log level 15.
RBD volume names and domains are sanitized, otherwise the log is untouched.
#2 Updated by Mykola Golub about 1 year ago
- Status changed from In Progress to Fix Under Review
- Pull request ID set to 33411
In the provided log there are many messages like these ones:
2020-02-14 02:14:56.653 7f42f7ac1700 -1 rbd::mirror::InstanceReplayer: 0x55bab1dc3b80 start_image_replayer: global_image_id=446b538f-1f61-4daa-b05f-93f76cd5e652: blacklisted detected during image replay 2020-02-14 02:14:56.660 7f42f7ac1700 5 rbd::mirror::LeaderWatcher: 0x55bab29a9200 handle_rewatch_complete: r=-108
So both rbd-mirror's InstanceReplayer and LeaderWatcher detected the "blacklisted" state but it was not propagated on the higher level to restart the PoolReplayer.