** Description changed: - The upstream bug is https://tracker.ceph.com/issues/53002 + [Impact] - It's been backported to relevant releases upstream (Octopus, Pacific, and Quincy). - Octopus 15.2.16 has the fix. So does Quincy 17.2.0. However, the latest Pacific release missed out this fix. So needed to be SRU'ed for Pacific only. + OSDs crash at randomly due to race condition that can occur + at times. + + This was observed when onode's removal is followed by reading + and the latter causes object release before the removal is finalized. + The root cause is an improper 'pinned' state assessment in Onode::get(). - Master tracker: https://tracker.ceph.com/issues/53002 + [Test Plan] + + Deploy a ceph cluster and do write some data to the cluster. + While performing some reads again from the cluster, no crashes + are seen in any OSDs. The race condition can be mimicked + by holding one thread (under debugger) while the other one + continues to update 'nput' counter. + + [Where problems could occur] + + Despite the new atomic counter it might not be cover cases + and still introduce further data race and/or crashes continue + to happen. + + [Other Info] + + The upstream bug is https://tracker.ceph.com/issues/53002 + + It's been backported to relevant releases upstream (Octopus, Pacific, and + Quincy). Octopus 15.2.16 has the fix. So does Quincy 17.2.0. However, + the latest Pacific release missed out this fix. So SRU is needed for + Pacific (only). Pacific tracker: https://tracker.ceph.com/issues/53608 Pacific PR: https://github.com/ceph/ceph/pull/44723
** Description changed: [Impact] - OSDs crash at randomly due to race condition that can occur - at times. - - This was observed when onode's removal is followed by reading - and the latter causes object release before the removal is finalized. - The root cause is an improper 'pinned' state assessment in Onode::get(). + OSDs crash at randomly due to race condition that can occur + at times. + + This was observed when onode's removal is followed by reading + and the latter causes object release before the removal is finalized. + The root cause is an improper 'pinned' state assessment in Onode::get(). [Test Plan] - Deploy a ceph cluster and do write some data to the cluster. - While performing some reads again from the cluster, no crashes - are seen in any OSDs. The race condition can be mimicked - by holding one thread (under debugger) while the other one - continues to update 'nput' counter. + Deploy a ceph cluster and do write some data to the cluster. + While performing some reads again from the cluster, no crashes + are seen in any OSDs. The race condition can be mimicked + by holding one thread (under debugger) while the other one + continues to update 'nput' counter. [Where problems could occur] - Despite the new atomic counter it might not be cover cases - and still introduce further data race and/or crashes continue - to happen. - + Despite the new atomic counter it might not be cover cases + and still introduce further data race and/or crashes continue + to happen. + [Other Info] - - The upstream bug is https://tracker.ceph.com/issues/53002 It's been backported to relevant releases upstream (Octopus, Pacific, and Quincy). Octopus 15.2.16 has the fix. So does Quincy 17.2.0. However, the latest Pacific release missed out this fix. So SRU is needed for Pacific (only). + Master tracker: https://tracker.ceph.com/issues/53002 + Pacific tracker: https://tracker.ceph.com/issues/53608 Pacific PR: https://github.com/ceph/ceph/pull/44723 -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/1970460 Title: [SRU] Avoid premature onode release To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/ceph/+bug/1970460/+subscriptions -- ubuntu-bugs mailing list ubuntu-bugs@lists.ubuntu.com https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
