Supposing the following scenario with a virtio_blk driver. CPU0 CPU1 /* * Add request to dispatch list or set bitmap of * software queue. 1) store virtblk_done() */ blk_mq_run_hw_queues()/blk_mq_delay_run_hw_queues() blk_mq_start_stopped_hw_queues() if (blk_mq_hctx_stopped()) 2) load blk_mq_start_stopped_hw_queue() continue clear_bit(BLK_MQ_S_STOPPED) 3) store blk_mq_run_hw_queue()/blk_mq_delay_run_hw_queue() blk_mq_run_hw_queue() if (!blk_mq_hctx_has_pending()) 4) load return blk_mq_sched_dispatch_requests() 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 BLK_MQ_S_STOPPED 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. Signed-off-by: Muchun Song <songmuchun@xxxxxxxxxxxxx> --- block/blk-mq.c | 24 ++++++++++++++++++++++++ 1 file changed, 24 insertions(+) diff --git a/block/blk-mq.c b/block/blk-mq.c index 6f18993b8f454..385a74e566874 100644 --- a/block/blk-mq.c +++ b/block/blk-mq.c @@ -2299,6 +2299,18 @@ void blk_mq_run_hw_queues(struct request_queue *q, bool async) sq_hctx = NULL; if (blk_queue_sq_sched(q)) sq_hctx = blk_mq_get_sq_hctx(q); + + /* + * This barrier is used to order adding of dispatch list or setting + * of bitmap of any software queue outside of this function and the + * test of BLK_MQ_S_STOPPED in the following routine. Pairs with the + * barrier in blk_mq_start_stopped_hw_queue(). So dispatch code could + * either see BLK_MQ_S_STOPPED is cleared or dispatch list or setting + * of bitmap of any software queue to avoid missing dispatching + * requests. + */ + smp_mb(); + queue_for_each_hw_ctx(q, hctx, i) { if (blk_mq_hctx_stopped(hctx)) continue; @@ -2327,6 +2339,18 @@ void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs) sq_hctx = NULL; if (blk_queue_sq_sched(q)) sq_hctx = blk_mq_get_sq_hctx(q); + + /* + * This barrier is used to order adding of dispatch list or setting + * of bitmap of any software queue outside of this function and the + * test of BLK_MQ_S_STOPPED in the following routine. Pairs with the + * barrier in blk_mq_start_stopped_hw_queue(). So dispatch code could + * either see BLK_MQ_S_STOPPED is cleared or dispatch list or setting + * of bitmap of any software queue to avoid missing dispatching + * requests. + */ + smp_mb(); + queue_for_each_hw_ctx(q, hctx, i) { if (blk_mq_hctx_stopped(hctx)) continue; -- 2.20.1