Re: [PATCH v2 2/3] block: fix ordering between checking QUEUE_FLAG_QUIESCED and adding requests

[Muchun Song <muchun.song@linux.dev>]

> On Sep 11, 2024, at 11:59, Muchun Song <muchun.song@linux.dev> wrote:
> 
> 
> 
>> On Sep 11, 2024, at 11:54, Ming Lei <ming.lei@redhat.com> wrote:
>> 
>>> On Tue, Sep 10, 2024 at 07:22:16AM -0600, Jens Axboe wrote:
>>> On 9/3/24 2:16 AM, Muchun Song wrote:
>>>> Supposing the following scenario.
>>>> 
>>>> CPU0                                        CPU1
>>>> 
>>>> blk_mq_insert_request()         1) store    blk_mq_unquiesce_queue()
>>>> blk_mq_run_hw_queue()                       blk_queue_flag_clear(QUEUE_FLAG_QUIESCED)   3) store
>>>>   if (blk_queue_quiesced())   2) load         blk_mq_run_hw_queues()
>>>>       return                                      blk_mq_run_hw_queue()
>>>>   blk_mq_sched_dispatch_requests()                    if (!blk_mq_hctx_has_pending()) 4) load
>>>>                                                          return
>>>> 
>>>> The full memory barrier should be inserted between 1) and 2), as well as
>>>> between 3) and 4) to make sure that either CPU0 sees QUEUE_FLAG_QUIESCED is
>>>> cleared or CPU1 sees dispatch list or setting of bitmap of software queue.
>>>> Otherwise, either CPU will not re-run the hardware queue causing starvation.
>>>> 
>>>> So the first solution is to 1) add a pair of memory barrier to fix the
>>>> problem, another solution is to 2) use hctx->queue->queue_lock to synchronize
>>>> QUEUE_FLAG_QUIESCED. Here, we chose 2) to fix it since memory barrier is not
>>>> easy to be maintained.
>>> 
>>> Same comment here, 72-74 chars wide please.
>>> 
>>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>>> index b2d0f22de0c7f..ac39f2a346a52 100644
>>>> --- a/block/blk-mq.c
>>>> +++ b/block/blk-mq.c
>>>> @@ -2202,6 +2202,24 @@ void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
>>>> }
>>>> EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
>>>> 
>>>> +static inline bool blk_mq_hw_queue_need_run(struct blk_mq_hw_ctx *hctx)
>>>> +{
>>>> +    bool need_run;
>>>> +
>>>> +    /*
>>>> +     * When queue is quiesced, we may be switching io scheduler, or
>>>> +     * updating nr_hw_queues, or other things, and we can't run queue
>>>> +     * any more, even blk_mq_hctx_has_pending() can't be called safely.
>>>> +     *
>>>> +     * And queue will be rerun in blk_mq_unquiesce_queue() if it is
>>>> +     * quiesced.
>>>> +     */
>>>> +    __blk_mq_run_dispatch_ops(hctx->queue, false,
>>>> +                  need_run = !blk_queue_quiesced(hctx->queue) &&
>>>> +                  blk_mq_hctx_has_pending(hctx));
>>>> +    return need_run;
>>>> +}
>>> 
>>> This __blk_mq_run_dispatch_ops() is also way too wide, why didn't you
>>> just break it like where you copied it from?
>>> 
>>>> +
>>>> /**
>>>> * blk_mq_run_hw_queue - Start to run a hardware queue.
>>>> * @hctx: Pointer to the hardware queue to run.
>>>> @@ -2222,20 +2240,23 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
>>>> 
>>>> might_sleep_if(!async && hctx->flags & BLK_MQ_F_BLOCKING);
>>>> 
>>>> -    /*
>>>> -     * When queue is quiesced, we may be switching io scheduler, or
>>>> -     * updating nr_hw_queues, or other things, and we can't run queue
>>>> -     * any more, even __blk_mq_hctx_has_pending() can't be called safely.
>>>> -     *
>>>> -     * And queue will be rerun in blk_mq_unquiesce_queue() if it is
>>>> -     * quiesced.
>>>> -     */
>>>> -    __blk_mq_run_dispatch_ops(hctx->queue, false,
>>>> -        need_run = !blk_queue_quiesced(hctx->queue) &&
>>>> -        blk_mq_hctx_has_pending(hctx));
>>>> +    need_run = blk_mq_hw_queue_need_run(hctx);
>>>> +    if (!need_run) {
>>>> +        unsigned long flags;
>>>> 
>>>> -    if (!need_run)
>>>> -        return;
>>>> +    /*
>>>> +     * synchronize with blk_mq_unquiesce_queue(), becuase we check
>>>> +     * if hw queue is quiesced locklessly above, we need the use
>>>> +     * ->queue_lock to make sure we see the up-to-date status to
>>>> +     * not miss rerunning the hw queue.
>>>> +     */
>>>> +    spin_lock_irqsave(&hctx->queue->queue_lock, flags);
>>>> +    need_run = blk_mq_hw_queue_need_run(hctx);
>>>> +    spin_unlock_irqrestore(&hctx->queue->queue_lock, flags);
>>>> +
>>>> +    if (!need_run)
>>>> +        return;
>>>> +    }
>>> 
>>> Is this not solvable on the unquiesce side instead? It's rather a shame
>>> to add overhead to the fast path to avoid a race with something that's
>>> super unlikely, like quisce.
>> 
>> Yeah, it can be solved by adding synchronize_rcu()/srcu() in unquiesce
>> side, but SCSI may call it in non-sleepable context via scsi_internal_device_unblock_nowait().
> 
> Another approach will be like the fix for BLK_MQ_S_STOPPED (in patch 3),
> we could add a pair of mb into blk_queue_quiesced() and
> blk_mq_unquiesce_queue(). In which case, the fix will not affect any fast
> path, only slow path need the barrier overhead.

I misunderstood Jens’s question. I think Ming is right.
This approach only tries to reduce the overhead as 
much
as possible even for slow path compared to 
spinlock_based approach. 
Not solving the problem only from the unquiesce side.

Muchun,
Thanks.

> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index b2d0f22de0c7f..45588ddb08d6b 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -264,6 +264,12 @@ void blk_mq_unquiesce_queue(struct request_queue *q)
>                ;
>        } else if (!--q->quiesce_depth) {
>                blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
> +               /*
> +                * Pairs with the smp_mb() in blk_queue_quiesced() to order the
> +                * clearing of QUEUE_FLAG_QUIESCED above and the checking of
> +                * dispatch list in the subsequent routine.
> +                */
> +               smp_mb__after_atomic();
>                run_queue = true;
>        }
>        spin_unlock_irqrestore(&q->queue_lock, flags);
> diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
> index b8196e219ac22..7a71462892b66 100644
> --- a/include/linux/blkdev.h
> +++ b/include/linux/blkdev.h
> @@ -628,7 +628,25 @@ void blk_queue_flag_clear(unsigned int flag, struct request_queue *q);
> #define blk_noretry_request(rq) \
>        ((rq)->cmd_flags & (REQ_FAILFAST_DEV|REQ_FAILFAST_TRANSPORT| \
>                             REQ_FAILFAST_DRIVER))
> -#define blk_queue_quiesced(q)  test_bit(QUEUE_FLAG_QUIESCED, &(q)->queue_flags)
> +
> +static inline bool blk_queue_quiesced(struct request_queue *q)
> +{
> +       /* Fast path: hardware queue is unquiesced most of the time. */
> +       if (likely(!test_bit(QUEUE_FLAG_QUIESCED, &q->queue_flags)))
> +               return false;
> +
> +       /*
> +        * This barrier is used to order adding of dispatch list before and
> +        * the test of QUEUE_FLAG_QUIESCED below. Pairs with the memory barrier
> +        * in blk_mq_unquiesce_queue() so that dispatch code could either see
> +        * QUEUE_FLAG_QUIESCED is cleared or dispatch list is not  empty to
> +        * avoid missing dispatching requests.
> +        */
> +       smp_mb();
> +
> +       return test_bit(QUEUE_FLAG_QUIESCED, &q->queue_flags);
> +}
> +
> #define blk_queue_pm_only(q)   atomic_read(&(q)->pm_only)
> #define blk_queue_registered(q)        test_bit(QUEUE_FLAG_REGISTERED, &(q)->queue_flags)
> #define blk_queue_sq_sched(q)  test_bit(QUEUE_FLAG_SQ_SCHED, &(q)->queue_flags)
> 
> Muchun,
> Thanks.
> 
>> 
>> 
>> Thanks,
>> Ming
> 
>