On 2023/11/10 02:13, Bart Van Assche wrote: > On 11/8/23 17:27, Chengming Zhou wrote: >> CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION is enabled in the kernel config, >> so null_queue_rq() will return BLK_STS_RESOURCE or BLK_STS_DEV_RESOURCE >> for some requests, which have been marked as IN_FLIGHT status. >> >> Then null_queue_rqs() put these requests in the rqlist and return back, >> blk-mq will try to queue them individually once again, caused the warning >> "WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE)" in blk_mq_start_request(). >> >> So handling of return value of null_queue_rq() in null_queue_rqs() is wrong, >> maybe we should __blk_mq_requeue_request() for these requests, before >> adding them in the rqlist? > > Please follow the example of virtio_queue_rqs() and send any requests > that need to be requeued back to the block layer core instead of > handling these directly in null_queue_rqs(). > Ok, I reviewed the code of virtio_queue_rqs(), found the main difference is that request won't fail after blk_mq_start_request(). But in null_blk case, the request will fail after blk_mq_start_request(), return BLK_STS_RESOURCE or BLK_STS_DEV_RESOURCE. If we return these rqs back to the block layer core, they will be queued individually once again. So caused the warning. Thanks!
