On Mon, Aug 27, 2018 at 01:56:39PM +0800, jianchao.wang wrote:
> Hi Ming
>
> Currently, blk_mq_update_dispatch_busy is hooked in blk_mq_dispatch_rq_list
> and __blk_mq_issue_directly. blk_mq_update_dispatch_busy could be invoked on multiple
> cpus concurrently. But there is not any protection on the hctx->dispatch_busy. We cannot
> ensure the update on the dispatch_busy atomically.
The update itself is atomic given type of this variable is 'unsigned int'.
>
>
> Look at the test result after applied the debug patch below:
>
> fio-1761 [000] .... 227.246251: blk_mq_update_dispatch_busy.part.50: old 0 ewma 2 cur 2
> fio-1766 [004] .... 227.246252: blk_mq_update_dispatch_busy.part.50: old 2 ewma 1 cur 1
> fio-1755 [000] .... 227.246366: blk_mq_update_dispatch_busy.part.50: old 1 ewma 0 cur 0
> fio-1754 [003] .... 227.266050: blk_mq_update_dispatch_busy.part.50: old 2 ewma 3 cur 3
> fio-1763 [007] .... 227.266050: blk_mq_update_dispatch_busy.part.50: old 0 ewma 2 cur 2
> fio-1761 [000] .... 227.266051: blk_mq_update_dispatch_busy.part.50: old 3 ewma 2 cur 2
> fio-1766 [004] .... 227.266051: blk_mq_update_dispatch_busy.part.50: old 3 ewma 2 cur 2
> fio-1760 [005] .... 227.266165: blk_mq_update_dispatch_busy.part.50: old 2 ewma 1 cur 1
>
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -1088,11 +1088,12 @@ static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
> static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
> {
> unsigned int ewma;
> + unsigned int old;
>
> if (hctx->queue->elevator)
> return;
>
> - ewma = hctx->dispatch_busy;
> + old = ewma = hctx->dispatch_busy;
>
> if (!ewma && !busy)
> return;
> @@ -1103,6 +1104,8 @@ static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
> ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
>
> hctx->dispatch_busy = ewma;
> +
> + trace_printk("old %u ewma %u cur %u\n", old, ewma, READ_ONCE(hctx->dispatch_busy));
> }
>
>
> Is it expected ?
Yes, it won't be a issue in reality given hctx->dispatch_busy is used as
a hint, and it often works as expected and hctx->dispatch_busy is convergent
finally because it is exponential weighted moving average.
Thanks,
Ming
