Re: [PATCH] rwsem: reduce spinlock contention in wakeup code path

[Waiman Long <waiman.long@hp.com>]

On 09/27/2013 03:32 PM, Peter Hurley wrote:
> On 09/27/2013 03:00 PM, Waiman Long wrote:
>> With the 3.12-rc2 kernel, there is sizable spinlock contention on
>> the rwsem wakeup code path when running AIM7's high_systime workload
>> on a 8-socket 80-core DL980 (HT off) as reported by perf:
>>
>>    7.64%   reaim  [kernel.kallsyms]   [k] _raw_spin_lock_irqsave
>>               |--41.77%-- rwsem_wake
>>    1.61%   reaim  [kernel.kallsyms]   [k] _raw_spin_lock_irq
>>               |--92.37%-- rwsem_down_write_failed
>>
>> That was 4.7% of recorded CPU cycles.
>>
>> On a large NUMA machine, it is entirely possible that a fairly large
>> number of threads are queuing up in the ticket spinlock queue to do
>> the wakeup operation. In fact, only one will be needed.  This patch
>> tries to reduce spinlock contention by doing just that.
>>
>> A new wakeup field is added to the rwsem structure. This field is
>> set on entry to rwsem_wake() and __rwsem_do_wake() to mark that a
>> thread is pending to do the wakeup call. It is cleared on exit from
>> those functions.
>>
>> By checking if the wakeup flag is set, a thread can exit rwsem_wake()
>> immediately if another thread is pending to do the wakeup instead of
>> waiting to get the spinlock and find out that nothing need to be done.
>
> This will leave readers stranded if a former writer is in __rwsem_do_wake
> to wake up the readers and another writer steals the lock, but before
> the former writer exits without having woken up the readers, the locking
> stealing writer drops the lock and sees the wakeup flag is set, so
> doesn't bother to wake the readers.
>
> Regards,
> Peter Hurley
>

Yes, you are right. That can be a problem. Thank for pointing this out. 
The workloads that I used doesn't seem to exercise the readers. I will 
modify the patch to add code handle this failure case by resetting the 
wakeup flag, pushing it out and then retrying one more time to get the 
read lock. I  think that should address the problem.

Regards,
Longman