Ceph : Issueshttps://tracker.ceph.com/https://tracker.ceph.com/favicon.ico2014-12-24T05:37:59ZCeph
Redmine Ceph - Bug #10431 (Resolved): PG can not finish peering due to mismatch between OSD peering queue...https://tracker.ceph.com/issues/104312014-12-24T05:37:59ZDong Leileidong@yahoo-inc.com
<p>We were debugging a PG stuck at peering problem. It may due to peering event lost or not been handled.</p>
<p>We found that some thread call osd->peering_queue.push_back without holding the osd_lock. It may cause a race condition when other threads (usually a dispatcher thread) push_back to peering_queue at the same time.</p>
<p>We found at least when handling an FlushedEvt, the thread will push_back osd peering_queue.</p>
<p>Can we add some checkers to assure the thread holds lock when doing osd->peering_wq.queue(PG*).</p>
<ul>
<li>firefly equivalent change <a class="changeset" title="osd: requeue PG when we skip handling a peering event If we don't handle the event, we need to p..." href="https://tracker.ceph.com/projects/ceph/repository/revisions/852d7b5b3c019c02c042b767fc88916088e1a94d">852d7b5b3c019c02c042b767fc88916088e1a94d</a></li>
</ul> Ceph - Bug #9485 (Resolved): Monitor crash due to wrong crush rule sethttps://tracker.ceph.com/issues/94852014-09-15T20:37:04ZDong Leileidong@yahoo-inc.com
<p>I create a customized crush rule for ec pool</p>
<p>1 set take default<br />2 choose firstn 6 type rack<br />3 chooseleaf firstn 2 type host<br />4 step emit</p>
<p>and it works fine, but if i change "firstn" to "indep" and set crushmap, the monitor crash and can not be restarted.</p>
<p>My cluster has only 6 rack. I guess that the "choose indep 6 type rack" may returns hole in the result, and at the next step CRUSH can not choose leaf from hole, so it fails.</p>
<p>Unfortunately, the wrong crush rule has been set, and every time I try to restart monitor, the monitor crash for the same reason "can not apply rule ... segment fault...". The whole cluster then become unusable.</p>
<p>Although I set a crush rule that can not be satisfied, the monitor should not crash forever.</p>