Supposing first scenario with a virtio_blk driver.

CPU0 CPU1

blk_mq_try_issue_directly() __blk_mq_issue_directly() q->mq_ops->queue_rq() virtio_queue_rq() blk_mq_stop_hw_queue() virtblk_done() blk_mq_request_bypass_insert() 1) store blk_mq_start_stopped_hw_queue() clear_bit(BLK_MQ_S_STOPPED) 3) store blk_mq_run_hw_queue() if (!blk_mq_hctx_has_pending()) 4) load return blk_mq_sched_dispatch_requests() blk_mq_run_hw_queue() if (!blk_mq_hctx_has_pending()) return blk_mq_sched_dispatch_requests() if (blk_mq_hctx_stopped()) 2) load return __blk_mq_sched_dispatch_requests()

Supposing another scenario.

CPU0 CPU1

blk_mq_requeue_work() blk_mq_insert_request() 1) store virtblk_done() blk_mq_start_stopped_hw_queue() blk_mq_run_hw_queues() clear_bit(BLK_MQ_S_STOPPED) 3) store blk_mq_run_hw_queue() if (!blk_mq_hctx_has_pending()) 4) load return blk_mq_sched_dispatch_requests() if (blk_mq_hctx_stopped()) 2) load continue blk_mq_run_hw_queue()

Both scenarios are similar, 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. Otherwise, either CPU will not rerun the hardware queue causing starvation of the request.

The easy way to fix it is to add the essential full memory barrier into helper of blk_mq_hctx_stopped(). In order to not affect the fast path (hardware queue is not stopped most of the time), we only insert the barrier into the slow path. Actually, only slow path needs to care about missing of dispatching the request to the low-level device driver.

Fixes: 320ae51feed5 ("blk-mq: new multi-queue block IO queueing mechanism") Cc: stable@vger.kernel.org Cc: Muchun Song muchun.song@linux.dev Signed-off-by: Muchun Song songmuchun@bytedance.com Reviewed-by: Ming Lei ming.lei@redhat.com --- block/blk-mq.c | 6 ++++++ block/blk-mq.h | 13 +++++++++++++ 2 files changed, 19 insertions(+)

diff --git a/block/blk-mq.c b/block/blk-mq.c index ff6df6c7eeb25..b90c1680cb780 100644 --- a/block/blk-mq.c +++ b/block/blk-mq.c @@ -2413,6 +2413,12 @@ void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async) return;

clear_bit(BLK_MQ_S_STOPPED, &hctx->state); + /* + * Pairs with the smp_mb() in blk_mq_hctx_stopped() to order the + * clearing of BLK_MQ_S_STOPPED above and the checking of dispatch + * list in the subsequent routine. + */ + smp_mb__after_atomic(); blk_mq_run_hw_queue(hctx, async); } EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue); diff --git a/block/blk-mq.h b/block/blk-mq.h index 260beea8e332c..f36f3bff70d86 100644 --- a/block/blk-mq.h +++ b/block/blk-mq.h @@ -228,6 +228,19 @@ static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data

static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx) { + /* Fast path: hardware queue is not stopped most of the time. */ + if (likely(!test_bit(BLK_MQ_S_STOPPED, &hctx->state))) + return false; + + /* + * This barrier is used to order adding of dispatch list before and + * the test of BLK_MQ_S_STOPPED below. Pairs with the memory barrier + * in blk_mq_start_stopped_hw_queue() so that dispatch code could + * either see BLK_MQ_S_STOPPED is cleared or dispatch list is not + * empty to avoid missing dispatching requests. + */ + smp_mb(); + return test_bit(BLK_MQ_S_STOPPED, &hctx->state); }