[LU-10931] failed peer discovery still taking too long

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.13.0
• Affects Version/s: Lustre 2.13.0
• Severity: 3
• Assignee: Amir Shehata (Inactive)
• Reporter: John Hammond

Description

On master, when running conf-sanity I often see mount stuck in the following stack
trace:

n:lustre-release# stack1 llog
29833 llog_process_th
[<ffffffffc06be64b>] lnet_discover_peer_locked+0x10b/0x380 [lnet]
[<ffffffffc06be930>] LNetPrimaryNID+0x70/0x1a0 [lnet]
[<ffffffffc0990ade>] ptlrpc_connection_get+0x3e/0x450 [ptlrpc]
[<ffffffffc098518c>] ptlrpc_uuid_to_connection+0xec/0x1a0 [ptlrpc]
[<ffffffffc09580c2>] import_set_conn+0xb2/0x7a0 [ptlrpc]
[<ffffffffc09587c3>] client_import_add_conn+0x13/0x20 [ptlrpc]
[<ffffffffc074efa9>] class_add_conn+0x419/0x680 [obdclass]
[<ffffffffc0750bc6>] class_process_config+0x19b6/0x27e0 [obdclass]
[<ffffffffc0753644>] class_config_llog_handler+0x934/0x14d0 [obdclass]
[<ffffffffc0717904>] llog_process_thread+0x834/0x1550 [obdclass]
[<ffffffffc071902f>] llog_process_thread_daemonize+0x9f/0xe0 [obdclass]
[<ffffffff810b252f>] kthread+0xcf/0xe0
[<ffffffff816b8798>] ret_from_fork+0x58/0x90
[<ffffffffffffffff>] 0xffffffffffffffff

conf-sanity has some tests that use bogus NIDs like 1.2.3.4 and 4.3.2.1.These are actually real IPv4 addresses but AFAICT they just discard all packets.I can see that the discovery thread cancels discovery on these peers but the llog_process_thread seems to stay in lnet_discover_peer_locked() for upto 60 seconds after. Looking at the code I can't see how it would get worken up in this case. Why doesn't lnet_peer_cancel_discovery() wake up the waiters on lp_dc_waitq? Or why don't we use schedule_timeout() with the discovery/transaction timeout in lnet_discover_peer_locked()?

Issue Links

• is duplicated by: LU-12416 NMI watchdog: BUG: soft lockup - CPU#0 stuck for 22s! [mount.lustre:11956]
• is related to: LU-12442 recovery-small test_136: mounts stuck in lnet_discover_peer_locked()
• is related to: LU-12519 sanity-sec test 31 crashes with ASSERTION( list_empty(&lp->lp_peer_nets) )
• is related to: LU-12424 LNet MR routing: possible loop when discovery is off

Comments

[Peter Jones]

Landed for 2.13

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/35444/
Subject: LU-10931 lnet: handle unlink before send completes
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: d8fc5c23fe541e0ff6ce5bec6302957714c3f69f

[Gerrit Updater]

Amir Shehata ([e-mail]) uploaded a new patch: https://review.whamcloud.com/35444
Subject: LU-10931 lnet: handle LNetMDUnlink case
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 97594d111f7bee96e056025dcad7f0d9bac91369

[Amir Shehata (Inactive)]

If LNetMDUnlink() is called on an md with md->md_refcount > 0 then the eq callback isn't called. What's happening here is the response times out before the send completes. So we have a refcount on the MD. The Unlink gets dropped on the floor. Send completes, but because we've already timed out the REPLY for the GET is dropped. Now we're left with a peer that is in:
LNET_PEER_MULTI_RAIL
LNET_PEER_DISCOVERING
LNET_PEER_PING_SENT
But no more events are coming to it, and the discovery never completes

[Gerrit Updater]

James Nunez ([e-mail]) uploaded a new patch: https://review.whamcloud.com/35408
Subject: LU-10931 tests: checking debug status for recovery-small
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 492cebf6912af3840cdbd39a815e5fc5e41bb96c

[James Nunez (Inactive)]

Patch 35325 put recovery-small test 136 on the ALWAYS_EXCEPT list and does not fix this issue. Please keep this issue open until a fix for this issue is found and the test is removed from the ALWAYS_EXCEPT list.

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/35325/
Subject: LU-10931 tests: stop running recovery-small 136
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: b511c65f468698aeb6e68f091ee3f30ec27830d0

[James Nunez (Inactive)]

I tested the patch for LU-12424, https://review.whamcloud.com/#/c/35191/, but we are still seeing this issue.

The patch https://review.whamcloud.com/#/c/35301/ is just a duplicate of Amir's LU-12424 patch and I ran recovery-small several times and still see this hang. Two failures with the LU-12424 patch are at
https://testing.whamcloud.com/test_sessions/6a97872a-fc69-444f-afbf-d44bb93d11cf
https://testing.whamcloud.com/test_sessions/6c9c2bb5-43c9-4ddd-a3a7-e5da94403e1a

Amir - Would you please comment on this issue and if the 35191 patch should fix this issue?

[Andreas Dilger]

Close as a duplicate because patch https://review.whamcloud.com/35191 "LU-12424 lnet: prevent loop in LNetPrimaryNID()" has landed and should resolve the issue.

[Gerrit Updater]

James Nunez ([e-mail]) uploaded a new patch: https://review.whamcloud.com/35325
Subject: LU-10931 tests: stop running recovery-small 136
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 38b6a60c213999d03e4ae5000a2bfdb7db1121b6

[Gerrit Updater]

James Nunez ([e-mail]) uploaded a new patch: https://review.whamcloud.com/35301
Subject: LU-10931 lnet: run testing against patch
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 08e1512871834b218f4f321743f9679e0a1b55e7