Re: [PATCH v2] blk-mq: Fix a race condition in blk_mq_mark_tag_wait()

[Jens Axboe <axboe@kernel.dk>]

On 6/27/18 6:00 PM, Bart Van Assche wrote:
> On 06/27/18 16:27, Ming Lei wrote:
>> On Wed, Jun 27, 2018 at 01:02:12PM -0700, Bart Van Assche wrote:
>>> Because the hctx lock is not held around the only
>>> blk_mq_tag_wakeup_all() call in the block layer, the wait queue
>>> entry removal in blk_mq_dispatch_wake() is protected by the wait
>>> queue lock only. Since the hctx->dispatch_wait entry can occur on
>>> any of the SBQ_WAIT_QUEUES, the wait queue presence check, adding
>>> .dispatch_wait to a wait queue and removing the wait queue entry
>>> must all be protected by both the hctx lock and the wait queue
>>> lock.
>>
>> Actually we don't need to use hctx->lock for protecting
>> hctx->dispatch_wait, and one new lock of hctx->dispatch_wait_lock is
>> enough, please see the following patch:
>>
>> 	https://marc.info/?l=linux-block&m=152998658713265&w=2
>>
>> Then we can avoid to disable irq when acquiring hctx->lock.
> 
> I think it's more a matter of taste than a technical decision to choose 
> which patch goes upstream.

I do think the split lock is cleaner in this case, since it avoids
making hctx->lock irq disabling.

-- 
Jens Axboe