Re: [PATCH v5] locking/rwsem: Make handoff bit handling more consistent

[Waiman Long <longman@redhat.com>]

On 4/20/22 09:55, john.p.donnelly@oracle.com wrote:
> On 4/12/22 11:28 AM, john.p.donnelly@oracle.com wrote:
>> On 4/11/22 4:07 PM, Waiman Long wrote:
>>>
>>> On 4/11/22 17:03, john.p.donnelly@oracle.com wrote:
>>>>
>>>>>>
>>>>>> I have reached out to Waiman and he suggested this for our next 
>>>>>> test pass:
>>>>>>
>>>>>>
>>>>>> 1ee326196c6658 locking/rwsem: Always try to wake waiters in 
>>>>>> out_nolock path
>>>>>
>>>>> Does this commit help to avoid the lockup problem?
>>>>>
>>>>> Commit 1ee326196c6658 fixes a potential missed wakeup problem when 
>>>>> a reader first in the wait queue is interrupted out without 
>>>>> acquiring the lock. It is actually not a fix for commit 
>>>>> d257cc8cb8d5. However, this commit changes the out_nolock path 
>>>>> behavior of writers by leaving the handoff bit set when the wait 
>>>>> queue isn't empty. That likely makes the missed wakeup problem 
>>>>> easier to reproduce.
>>>>>
>>>>> Cheers,
>>>>> Longman
>>>>>
>>>>
>>>> Hi,
>>>>
>>>>
>>>> We are testing now
>>>>
>>>> ETA for fio soak test completion is  ~15hr from now.
>>>>
>>>> I wanted to share the stack traces for future reference + occurrences.
>>>>
>>> I am looking forward to your testing results tomorrow.
>>>
>>> Cheers,
>>> Longman
>>>
>> Hi
>>
>>   Our 24hr fio soak test with :
>>
>>   1ee326196c6658 locking/rwsem: Always try to wake waiters in 
>> out_nolock path
>>
>>
>>   applied to 5.15.30  passed.
>>
>>   I suggest you append  1ee326196c6658 with :
>>
>>
>>   cc: stable
>>
>>    Fixes: d257cc8cb8d5 ("locking/rwsem: Make handoff bit handling 
>> more consistent")
>>
>>
>> I'll leave the implementation details up to the core maintainers how 
>> to do that ;-)
>>
>> ...
>>
>> Thank you
>>
>> John.
>
> Hi ,
>
>
>  We have observed another panic with :
>
>  1ee326196c6658 locking/rwsem: Always try to wake waiters in out_nolock
>  path
>
>  Applied to 5.15.30 :
>
>
Sorry for the late reply as I was busy with other important tasks.

When you said panic, you mean a system hang, not an actual panic. Right?


> PID: 3789   TASK: ffff900fc409b300  CPU: 29  COMMAND: "dio/dm-0"
>  #0 [fffffe00006bce50] crash_nmi_callback at ffffffff97c772c3
>  #1 [fffffe00006bce58] nmi_handle at ffffffff97c40778
>  #2 [fffffe00006bcea0] default_do_nmi at ffffffff988161e2
>  #3 [fffffe00006bcec8] exc_nmi at ffffffff9881648d
>  #4 [fffffe00006bcef0] end_repeat_nmi at ffffffff98a0153b
>     [exception RIP: _raw_spin_lock_irq+35]
>     RIP: ffffffff98827333  RSP: ffffa9320917fc78  RFLAGS: 00000046
>     RAX: 0000000000000000  RBX: ffff900fc409b300  RCX: 0000000000000000
>     RDX: 0000000000000000  RSI: 0000000000000000  RDI: 0000000000000000
>     RBP: ffffa9320917fd20   R8: 0000000000000000   R9: 0000000000000000
>     R10: 0000000000000000  R11: 0000000000000000  R12: ffff90006259546c
>     R13: ffffa9320917fcb0  R14: ffff900062595458  R15: 0000000000000000
>     ORIG_RAX: ffffffffffffffff  CS: 0010  SS: 0018
> --- <NMI exception stack> ---
>  #5 [ffffa9320917fc78] _raw_spin_lock_irq at ffffffff98827333
>  #6 [ffffa9320917fc78] rwsem_down_write_slowpath at ffffffff97d25d49
>  #7 [ffffa9320917fd28] ext4_map_blocks at ffffffffc104b6dc [ext4]
>  #8 [ffffa9320917fd98] ext4_convert_unwritten_extents at 
> ffffffffc10369e0 [ext4]
>  #9 [ffffa9320917fdf0] ext4_dio_write_end_io at ffffffffc103b2aa [ext4]
> #10 [ffffa9320917fe18] iomap_dio_complete at ffffffff98013f45
> #11 [ffffa9320917fe48] iomap_dio_complete_work at ffffffff98014047
> #12 [ffffa9320917fe60] process_one_work at ffffffff97cd9191
> #13 [ffffa9320917fea8] rescuer_thread at ffffffff97cd991b
> #14 [ffffa9320917ff10] kthread at ffffffff97ce11f7
> #15 [ffffa9320917ff50] ret_from_fork at ffffffff97c04cf2
> crash>
>
>
> The failure is observed running "fio test suite"  as a 24 hour soak 
> test  on an LVM composed of four NVME devices, Intel 72 core server. 
> The test cycles through a variety of file-system types.
>
>
> This kernel has these commits
>
> 1ee326196c6658 locking/rwsem: Always try to wake waiters in 
> out_nolock  path
>
> d257cc8cb8d5 ("locking/rwsem: Make handoff bit handling more consistent")
>
> In earlier testing I had reverted d257cc8cb8d5 and did not observe 
> said panics.  I still feel d257cc8cb8d5 is  still the root cause.

So it is possible that 1ee326196c6658 does not completely eliminate the 
missed wakeup situation.

Regards,
Longman