Re: [PATCH] block: Fix a race in the runtime power management code

[Bart Van Assche <bvanassche@acm.org>]

On 2020-08-25 02:11, Stanley Chu wrote:
>> diff --git a/block/blk-pm.c b/block/blk-pm.c
>> index b85234d758f7..17bd020268d4 100644
>> --- a/block/blk-pm.c
>> +++ b/block/blk-pm.c
>> @@ -67,6 +67,10 @@ int blk_pre_runtime_suspend(struct request_queue *q)
>>  
>>  	WARN_ON_ONCE(q->rpm_status != RPM_ACTIVE);
>>  
>> +	spin_lock_irq(&q->queue_lock);
>> +	q->rpm_status = RPM_SUSPENDING;
>> +	spin_unlock_irq(&q->queue_lock);
>> +
> 
> Has below alternative way been considered that RPM_SUSPENDING is set
> after blk_freeze_queue_start()?
> 
> 	blk_freeze_queue_start(q);
> 
> +	spin_lock_irq(&q->queue_lock);
> +	q->rpm_status = RPM_SUSPENDING;
> +	spin_unlock_irq(&q->queue_lock);
> 
> 
> Otherwise requests can enter queue while rpm_status is RPM_SUSPENDING
> during a small window, i.e., before blk_set_pm_only() is invoked. This
> would make the definition of rpm_status ambiguous.
> 
> In this way, the racing could be also solved:
> 
> - Before blk_freeze_queue_start(), any requests shall be allowed to
> enter queue
> - blk_freeze_queue_start() freezes the queue and blocks all upcoming
> requests (make them wait_event(q->mq_freeze_wq))
> - rpm_status is set as RPM_SUSPENDING
> - blk_mq_unfreeze_queue() wakes up q->mq_freeze_wq and then
> blk_pm_request_resume() can be executed

Hi Stanley,

I prefer the order from the patch. I think it is important to change
q->rpm_status into RPM_SUSPENDING before blk_queue_enter() calls
blk_queue_pm_only(). Otherwise it could happen that blk_queue_enter()
calls blk_pm_request_resume() while q->rpm_status == RPM_ACTIVE, resulting
in blk_queue_enter() not resuming a queue although that queue should be
resumed.

Thanks,

Bart.