[LU-4330] LustreError: 46336:0:(events.c:433:ptlrpc_master_callback()) ASSERTION( callback == request_out_callback || callback == reply_in_callback || callback == client_bulk_callback || callback == request_in_callback || callback == reply_out_callback ... ) failed

• Type: Bug
• Resolution: Unresolved
• Priority: Major
• Affects Version/s: Lustre 2.1.6
• Severity: 3
• Assignee: Bruno Faccini (Inactive)
• Reporter: Sebastien Buisson (Inactive)

Description

Hi,

More and more compute nodes of several different customer clusters are hitting an LBUG on this 'assertion failed' issue:

2013-11-21 14:06:54 LustreError: 46336:0:(events.c:433:ptlrpc_master_callback()) ASSERTION( callback == request_out_callback || callback == reply_in_callback || callback == client_bulk_callback || callback == request_in_callback || callback == reply_out_callback || callback == server_bulk_callback ) failed:
2013-11-21 14:06:54 LustreError: 46336:0:(events.c:433:ptlrpc_master_callback()) LBUG
2013-11-21 14:06:54 Nov 21 14:06:54 Pid: 46336, comm: kiblnd_sd_00
2013-11-21 14:06:54 compute5666 kernel
2013-11-21 14:06:54 : LustreError: 4Call Trace:
2013-11-21 14:06:54 6336:0:(events.c [<ffffffffa041c7f5>] libcfs_debug_dumpstack+0x55/0x80 [libcfs]
2013-11-21 14:06:54 :433:ptlrpc_mast [<ffffffffa041ce07>] lbug_with_loc+0x47/0xb0 [libcfs]
2013-11-21 14:06:54 er_callback()) A [<ffffffffa06a426c>] ptlrpc_master_callback+0xcc/0xd0 [ptlrpc]
2013-11-21 14:06:54 SSERTION( callba [<ffffffffa048ebd2>] lnet_enq_event_locked+0x62/0xd0 [lnet]
2013-11-21 14:06:54 ck == request_ou [<ffffffffa048ecdb>] lnet_finalize+0x9b/0x2f0 [lnet]
2013-11-21 14:06:54 t_callback || ca [<ffffffffa083d073>] kiblnd_recv+0x103/0x570 [ko2iblnd]
2013-11-21 14:06:54 llback == reply_ [<ffffffffa04928dd>] lnet_ni_recv+0xad/0x2f0 [lnet]
2013-11-21 14:06:54 in_callback || c [<ffffffffa0492c06>] lnet_recv_put+0xe6/0x120 [lnet]
2013-11-21 14:06:54 allback == clien [<ffffffffa0499c33>] lnet_parse+0x1273/0x1b80 [lnet]
2013-11-21 14:06:54 t_bulk_callback [<ffffffff81042ca3>] ? enqueue_task+0x43/0x90
2013-11-21 14:06:54 || callback == r [<ffffffffa083d7ab>] kiblnd_handle_rx+0x2cb/0x680 [ko2iblnd]
2013-11-21 14:06:54 equest_in_callba [<ffffffffa083e590>] kiblnd_rx_complete+0x2d0/0x440 [ko2iblnd]
2013-11-21 14:06:54 ck || callback = [<ffffffff81042a63>] ? __wake_up+0x53/0x70
2013-11-21 14:06:54 = reply_out_call [<ffffffffa083e762>] kiblnd_complete+0x62/0xe0 [ko2iblnd]
2013-11-21 14:06:54 back || callback [<ffffffffa083eb19>] kiblnd_scheduler+0x339/0x7a0 [ko2iblnd]
2013-11-21 14:06:54 == server_bulk_ [<ffffffff8104a320>] ? default_wake_function+0x0/0x20
2013-11-21 14:06:54 callback ) faile [<ffffffffa083e7e0>] ? kiblnd_scheduler+0x0/0x7a0 [ko2iblnd]
2013-11-21 14:06:54 d:
2013-11-21 14:06:54 Nov 21 14:06 [<ffffffff8100412a>] child_rip+0xa/0x20
2013-11-21 14:06:54 :54 compute5666 ke [<ffffffffa083e7e0>] ? kiblnd_scheduler+0x0/0x7a0 [ko2iblnd]
2013-11-21 14:06:54 rnel: LustreErro [<ffffffffa083e7e0>] ? kiblnd_scheduler+0x0/0x7a0 [ko2iblnd]
2013-11-21 14:06:54 r: 46336:0:(even [<ffffffff81004120>] ? child_rip+0x0/0x20
2013-11-21 14:06:54 ts.c:433:ptlrpc_
2013-11-21 14:06:54 master_callback()) LBUG
2013-11-21 14:06:54 Nov 21 14:06:54 compute566Kernel panic - not syncing: LBUG

For information, systems are running with kernel boot parameter 'tolerant=1' set.

We have a crash dump that we will upload to ftp.

Sebastien.

Issue Links

• is related to: LU-8362 page fault: exception RIP: lnet_mt_match_md+135
• is related to: LU-3848 Compute node crashes due to error in Lustre : ASSERTION( me == md->md_me ) failed

Comments

[Peter Jones]

Regardless of what upstream think of this fix, I think that we no longer need to track it for 2.9 so I will remove the fix version

[Bruno Faccini (Inactive)]

James, sorry to be late to comment/answer you.
In fact, I should have re-phrase my commit msg before submitting patch upstream, seems that Greg thought that this patch has been created to fix the corruotions issues, when it has been created to pursue with the corruption isolation, by first avoiding LNet MEs/small-MDs to be impacted. Then the idea came up that moving LNet MEs/small-MDs to a specific kmem_cache has been developed and seen as a possible performance enhancement.
BTW, Greg seems curious about getting some numbers/stats about this enhancement!

[James A Simmons]

Attempted to merge the patch for LU-4330 upstream and it was rejected since according to Greg it will not fix the problem. See email

http://driverdev.linuxdriverproject.org/pipermail/driverdev-devel/2016-June/090787.html

Bruno can you reproduce this problem? Have you tried debugging with slub?

[Peter Jones]

Landed for 2.9

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch http://review.whamcloud.com/18586/
Subject: LU-4330 lnet: Allocate MEs and small MDs in own kmem_caches
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 9d9bb678d6b3707623845e0ce67dd7fd07a12fe9

[Bruno Faccini (Inactive)]

James,
Thanks for your input.
But well, I can not tell for this old case/crash, but on my side I have seen recent <size-128> Slabs corruption cases that are ext4/ldiskfs related (see LU-7980).

[James A Simmons]

I think I might know what the problem is. A recent patch for ko2iblnd in the upstream kernel landed that exposed a serious memory corruption.
The commit is 3d1477309806459d39e13d8c3206ba35d183c34a "Replace sg++ with sg = sg_next(sg)" The scatter gather list is from
tx->tx_frags which is IBLND_MAX_RDMA_FRAGS in size. Since you write at an offset into the tx_frags that mean you really need IBLND_MAX_RDMA_FRAGS + 1 in size for the frags. Currently the upstream clients will crash when you attempt to access scatter list entry IBLND_MAX_RDMA_FRAGS + 1.

[Andreas Dilger]

Bruno, if you update those patches for newer kernels, please submit the new patches into lustre/kernel_patches/patches so they are available for use in the future, since bugzilla may disappear at some point.

[Andreas Dilger]

The scary thing is that there would continue to be random memory corruptions in the size-128 slab, but they will just be corrupting some other part of memory.

If this problem can be found in a relatively short amount of testing time, then there are debugging patches available that could be applied to the kernel to make all kmalloc() calls actually map to vmalloc() internally and have vmalloc() always use a new memory address, and then when the memory is freed the page is unmapped and the address never used again. If another thread is incorrectly accessing an unmapped address (use after free) it will fault and then the source of the corruption may be found. Unfortunately, this impacts the performance and can only be used for debugging and not in production.

Patches are available in https://bugzilla.lustre.org/show_bug.cgi?id=22471 but the would likely need to be updated for newer kernels. They can definitely help find memory corruption problems that are otherwise very difficult to find.

[Gerrit Updater]

Faccini Bruno ([e-mail]) uploaded a new patch: http://review.whamcloud.com/18586
Subject: LU-4330 lnet: Allocate MEs and small MDs in own kmem_caches
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 11e4b34d75505476f363c1cf4400755a7f30f766