[LU-11063] RHEL7.[345] RCU breakage

• Type: Bug
• Resolution: Cannot Reproduce
• Priority: Major
• Affects Version/s: Lustre 2.13.0
• Severity: 3
• Bugzilla ID: 18,015
• Assignee: WC Triage
• Reporter: Oleg Drokin

Description

I finally traced my debug kernel problems with later rhel releases to RCU breakage of some sort.

ldlm_locks slab is declared as SLAB_DESTROY_BY_RCU if it's defined This is going back to bugzilla 18015 https://bugzilla.lustre.org/show_bug.cgi?id=18015 patch by BobiJam.

Now it appears that as we schedule a free in that slab and then destroy the slab, the actual free is delayed and is executed after the slab is already freed despite rcu_barrier() being present.

Clear bug that I will file rh bugzilla ticket for.

But in addition to that I wonder how much do we need that thing nowadays, esp. considering that newer kernels renamed the flag to SLAB_TYPESAFE_BY_RCU that we do not detect and just not set it in that case.

Should we just convert ldlm_locks into a normal slab again I wonder?

Issue Links

• is duplicated by: LU-12454 parallel-scale-nfsv3 test racer_on_nfs crashes with ‘BUG: unable to handle kernel paging request’
• is related to: LU-17097 RCU stall caused by osc_quota_cleanup
• is related to: LU-11568 Get rid of SLAB_DESTROY_BY_RCU
• is related to: LU-12374 client went down w/ panic during lustre_rmmod