[PATCH] blk-mq: auto-start hw queue in requeue

[PATCH] blk-mq: auto-start hw queue in requeue

[Ming Lei]

Requests often need to be retried because of timeout, hardware
busy or sort of reasons via blk_mq_requeue_request(), but at
that time dispatch queue may have been stopped.

This patch starts the dispatch queue automatically in requeue's
work handler, and fixes scsi-mq's timeout issue, which can be
triggered if there are two or more requests timedout.

Suggested-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
---
 block/blk-mq.c |   19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 80ade9f..edd0ece 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -33,6 +33,8 @@ static DEFINE_MUTEX(all_q_mutex);
 static LIST_HEAD(all_q_list);
 
 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
+static void __blk_mq_run_queues(struct request_queue *q, bool async,
+				bool restart);
 
 /*
  * Check if any of the ctx's have pending work in this hardware queue
@@ -473,7 +475,7 @@ static void blk_mq_requeue_work(struct work_struct *work)
 		blk_mq_insert_request(rq, false, false, false);
 	}
 
-	blk_mq_run_queues(q, false);
+	__blk_mq_run_queues(q, false, true);
 }
 
 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
@@ -806,15 +808,21 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 	}
 }
 
-void blk_mq_run_queues(struct request_queue *q, bool async)
+static void __blk_mq_run_queues(struct request_queue *q, bool async,
+				bool restart)
 {
 	struct blk_mq_hw_ctx *hctx;
 	int i;
 
 	queue_for_each_hw_ctx(q, hctx, i) {
+		bool stopped = test_bit(BLK_MQ_S_STOPPED, &hctx->state);
+
+		if (restart && stopped)
+		    clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
+
 		if ((!blk_mq_hctx_has_pending(hctx) &&
 		    list_empty_careful(&hctx->dispatch)) ||
-		    test_bit(BLK_MQ_S_STOPPED, &hctx->state))
+		    (stopped && !restart))
 			continue;
 
 		preempt_disable();
@@ -822,6 +830,11 @@ void blk_mq_run_queues(struct request_queue *q, bool async)
 		preempt_enable();
 	}
 }
+
+void blk_mq_run_queues(struct request_queue *q, bool async)
+{
+	__blk_mq_run_queues(q, async, false);
+}
 EXPORT_SYMBOL(blk_mq_run_queues);
 
 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
-- 
1.7.9.5

Re: [PATCH] blk-mq: auto-start hw queue in requeue

[Christoph Hellwig]

On Fri, Sep 19, 2014 at 10:22:50PM +0800, Ming Lei wrote:
> Requests often need to be retried because of timeout, hardware
> busy or sort of reasons via blk_mq_requeue_request(), but at
> that time dispatch queue may have been stopped.
> 
> This patch starts the dispatch queue automatically in requeue's
> work handler, and fixes scsi-mq's timeout issue, which can be
> triggered if there are two or more requests timedout.

Sounds like blk_mq_requeue_work should simply call void
blk_mq_start_hw_queues where it calls blk_mq_run_queues currently.

Re: [PATCH] blk-mq: auto-start hw queue in requeue

[Jens Axboe]

On 09/19/2014 11:48 AM, Christoph Hellwig wrote:
> On Fri, Sep 19, 2014 at 10:22:50PM +0800, Ming Lei wrote:
>> Requests often need to be retried because of timeout, hardware
>> busy or sort of reasons via blk_mq_requeue_request(), but at
>> that time dispatch queue may have been stopped.
>>
>> This patch starts the dispatch queue automatically in requeue's
>> work handler, and fixes scsi-mq's timeout issue, which can be
>> triggered if there are two or more requests timedout.
> 
> Sounds like blk_mq_requeue_work should simply call void
> blk_mq_start_hw_queues where it calls blk_mq_run_queues currently.

I agree. I think Ming wanted to avoid an unnecessary bitflip if
possible, but the fix is more convoluted than I would like.


-- 
Jens Axboe

Re: [PATCH] blk-mq: auto-start hw queue in requeue

[Christoph Hellwig]

On Fri, Sep 19, 2014 at 11:52:59AM -0600, Jens Axboe wrote:
> I agree. I think Ming wanted to avoid an unnecessary bitflip if
> possible, but the fix is more convoluted than I would like.

If the requeue from irq context ever ends up in the fast path we
have much worse problems..

Re: [PATCH] blk-mq: auto-start hw queue in requeue

[Jens Axboe]

On 09/19/2014 11:54 AM, Christoph Hellwig wrote:
> On Fri, Sep 19, 2014 at 11:52:59AM -0600, Jens Axboe wrote:
>> I agree. I think Ming wanted to avoid an unnecessary bitflip if
>> possible, but the fix is more convoluted than I would like.
> 
> If the requeue from irq context ever ends up in the fast path we
> have much worse problems..

Oh yes certainly, my point was purely that I think this is why Mings fix
ended up being more involved than it need be.

-- 
Jens Axboe