On Fri, Aug 23, 2024 at 7:28 PM Ming Lei <ming.lei@xxxxxxxxxx> wrote: > > On Sun, Aug 11, 2024 at 06:19:21 PM +0800, Muchun Song wrote: > > Supposing the following scenario. > > > > CPU0 CPU1 > > > > blk_mq_request_issue_directly() blk_mq_unquiesce_queue() > > if (blk_queue_quiesced()) blk_queue_flag_clear(QUEUE_FLAG_QUIESCED) 3) store > > blk_mq_insert_request() blk_mq_run_hw_queues() > > /* blk_mq_run_hw_queue() > > * Add request to dispatch list or set bitmap of if (!blk_mq_hctx_has_pending()) 4) load > > * software queue. 1) store return > > */ > > blk_mq_run_hw_queue() > > if (blk_queue_quiesced()) 2) 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 QUEUE_FLAG_QUIESCED 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. > > Memory barrier shouldn't serve as bug fix for two slow code paths. > > One simple fix is to add helper of blk_queue_quiesced_lock(), and > call the following check on CPU0: > > if (blk_queue_quiesced_lock()) > blk_mq_run_hw_queue(); This only fixes blk_mq_request_issue_directly(), I think anywhere that matching this pattern (inserting a request to dispatch list and then running the hardware queue) should be fixed. And I think there are many places which match this pattern (E.g. blk_mq_submit_bio()). The above graph should be adjusted to the following. CPU0 CPU1 blk_mq_insert_request() 1) store blk_mq_unquiesce_queue() blk_mq_run_hw_queue() blk_queue_flag_clear(QUEUE_FLAG_QUIESCED) 3) store if (blk_queue_quiesced()) 2) load blk_mq_run_hw_queues() return blk_mq_run_hw_queue() blk_mq_sched_dispatch_requests() if (!blk_mq_hctx_has_pending()) 4) load return So I think fixing blk_mq_run_hw_queue() could cover all of the situations. Maybe I thought wrongly. Please correct me. Muchun, Thanks.