[LU-1587] watchdog detected hard LOCKUP on OSS under heavy client reconnects Created: 30/Jun/12  Updated: 13/Apr/14  Resolved: 13/Apr/14

Status: Resolved
Project: Lustre
Component/s: None
Affects Version/s: Lustre 2.1.1
Fix Version/s: None

Type: Bug Priority: Blocker
Reporter: Ned Bass Assignee: Niu Yawei (Inactive)
Resolution: Cannot Reproduce Votes: 0
Labels: None
Environment:

https://github.com/chaos/lustre/commits/2.1.1-4chaos


Severity: 3
Rank (Obsolete): 3981

 Description   

We are experiencing a severe disruption of service on our classified network. Beginning around 5pm yesterday (Friday Jun 29) we began to have many evictions and reconnections across multiple servers and across multiple lustre clusters on our classified network. Some sort of network event may have triggered this, but we're not sure at this point. Lustre servers started getting STONITH'd by their failover partners.

Many servers have been getting panics due to CPU hard lockup. The stack traces aren't helpful; they just show the swapper task handling the watchdog NMI. We're trying to quiesce things by stopping lnet on the compute cluster routers, getting lustre started on all the servers, then bringing client clusters back online one at a time. However, we are still having sporadic CPU hard lockups on OSS's as clients reconnect. Typically the lockup follows several thousand client reconnects.

We have some complete crash dumps to look at, but there's not much to go on. Some process must be leaving interrupts disabled, but I'm not sure how to go about identifying with the data we have.



 Comments   
Comment by Peter Jones [ 01/Jul/12 ]

Niu could you please look into this one Peter

Comment by Niu Yawei (Inactive) [ 02/Jul/12 ]

Could you post the stack trace here as well? Though it might not be helpful. Thanks.

Comment by Prakash Surya (Inactive) [ 02/Jul/12 ]

Hi Niu. I'll post the stack trace as soon as I can. Although, I briefly talked to Ned and he didn't think the stack trace would be helpful.

Comment by Prakash Surya (Inactive) [ 02/Jul/12 ]

Here's a stack trace from one of the nodes which crashed. Please excuse any typo's, I had to copy it by hand:

Kernel panic - not syncing: Watchdog detected hard LOCKUP on cpu 4
PID: 0   TASK: ...   CPU: 4   COMMAND: "swapper"
0.  machine_kexec
1.  crash_kexec
2.  panic
3.  watchdog_overflow_callback
4.  __perf_event_overflow
5.  perf_event_overlfow
6.  x86_pmu_handle_irq
7.  perf_event_nmi_handler
8.  notifier_call_chain
9.  atomic_notifier_call_chain
10. notify_die
11. do_nmi
12. nmi 
13. ktime_get
14. tick_check_idle
15. irq_enter
16. smp_apic_timer_interrupt
17. apic_timer_interrupt
18. apic_timer_interrupt
19. notifier_call_chain
20. cpu_idle
Comment by Prakash Surya (Inactive) [ 02/Jul/12 ]

This curiously lines up with many 'leap second' issues I've read about over the weekend. Although, it's pure speculation if this is issue is related.

Comment by Andreas Dilger [ 13/Apr/14 ]

Haven't seen any more information about this bug, and unable to make any progress based on the description here. Based on the later comments, this might have been caused by the leap second bug.

Generated at Sat Feb 10 01:17:57 UTC 2024 using Jira 9.4.14#940014-sha1:734e6822bbf0d45eff9af51f82432957f73aa32c.