Re: [PATCH 2/4] locking/qrwlock: Reduce reader/writer to reader lock transfer latency

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

On 07/07/2015 07:49 AM, Will Deacon wrote:
> On Tue, Jul 07, 2015 at 12:17:31PM +0100, Peter Zijlstra wrote:
>> On Tue, Jul 07, 2015 at 10:17:11AM +0100, Will Deacon wrote:
>>>>> Thinking about it, can we kill _QW_WAITING altogether and set (cmpxchg
>>>>> from 0) wmode to _QW_LOCKED in the write_lock slowpath, polling (acquire)
>>>>> rmode until it hits zero?
>>>> No, this is how we make the lock fair so that an incoming streams of
>>>> later readers won't block a writer from getting the lock.
>>> But won't those readers effectively see that the lock is held for write
>>> (because we set wmode to _QW_LOCKED before the existing reader had drained)
>>> and therefore fall down the slow-path and get held up on the spinlock?
>> Yes, that's the entire point. Once there's a writer pending, new readers
>> should queue too.
> Agreed. My point was that we can achieve the same result without
> a separate _QW_WAITING flag afaict.
>
> Will
>

_QW_WAITING and _QW_LOCKED has different semantics and are necessary for 
the proper handshake between readers and writer. We set _QW_WAITING when 
readers own the lock and the writer is waiting for the readers to go 
away. The _QW_WAITING flag will force new readers to go to queuing while 
the writer is waiting. We set _QW_LOCKED when a writer own the lock and 
it can only be set atomically when no reader is present. Without the 
intermediate _QW_WAITING step, a continuous stream of incoming readers 
(which make the reader count never 0) could deny a writer from getting 
the lock indefinitely.

Cheers,
Longman