On 11/30/2017 05:08 PM, Bart Van Assche wrote:
> The following code at the end of blk_mq_dispatch_rq_list() detects
> whether or not wake_up(&hctx->dispatch_wait) has been called
> concurrently with pushing back requests onto the dispatch list:
>
> list_empty_careful(&hctx->dispatch_wait.entry)
>
> Since blk_mq_dispatch_wake() is protected by another lock than the
> dispatch list and since blk_mq_run_hw_queue() does not acquire any
> lock if it notices that no requests are pending,
> blk_mq_dispatch_wake() is not ordered against the code that pushes
> back requests onto the dispatch list. Avoid that the dispatch_wait
> empty check fails due to load/store reordering by serializing it
> against the dispatch_wait queue wakeup. This patch fixes a queue
> stall I ran into while testing a SCSI initiator driver with the
> maximum target depth set to one.
>
> Signed-off-by: Bart Van Assche <bart.vanassche@xxxxxxx>
> Cc: Ming Lei <ming.lei@xxxxxxxxxx>
> Cc: Omar Sandoval <osandov@xxxxxx>
> Cc: Christoph Hellwig <hch@xxxxxx>
> Cc: Hannes Reinecke <hare@xxxxxxx>
> Cc: Johannes Thumshirn <jthumshirn@xxxxxxx>
> ---
> block/blk-mq.c | 16 +++++++++++++++-
> 1 file changed, 15 insertions(+), 1 deletion(-)
>
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index b4225f606737..a11767a4d95c 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -1074,6 +1074,20 @@ static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
> return ret;
> }
>
> +static bool blk_mq_dispatch_list_empty(struct blk_mq_hw_ctx *hctx)
> +{
> + struct sbq_wait_state *ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
> + struct wait_queue_head *wq_head = &ws->wait;
> + unsigned long flags;
> + bool result;
> +
> + spin_lock_irqsave(&wq_head->lock, flags);
> + result = list_empty(&hctx->dispatch_wait.entry);
> + spin_unlock_irqrestore(&wq_head->lock, flags);
> +
> + return result;
> +}
This can't fix anything, since you're still depending on the state
outside the lock. You probably just changed the window slightly.
--
Jens Axboe
