Bart Van Assche <[email protected]> writes: > On 10/8/19 9:35 AM, Gabriel Krisman Bertazi wrote: >> André Almeida <[email protected]> writes: >> >>> diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h >>> index e0fce93ac127..8b745f229789 100644 >>> --- a/include/linux/blk-mq.h >>> +++ b/include/linux/blk-mq.h >>> @@ -10,74 +10,153 @@ struct blk_mq_tags; >>> struct blk_flush_queue; >>> /** >>> - * struct blk_mq_hw_ctx - State for a hardware queue facing the hardware >>> block device >>> + * struct blk_mq_hw_ctx - State for a hardware queue facing the hardware >>> + * block device >>> */ >>> struct blk_mq_hw_ctx { >>> struct { >>> + /** @lock: Lock for accessing dispatch queue */ >>> spinlock_t lock; >>> + /** >>> + * @dispatch: Queue of dispatched requests, waiting for >>> + * workers to send them to the hardware. >>> + */ >> >> It's been a few years since I looked at the block layer, but isn't >> this used to hold requests that were taken from the blk_mq_ctx, but >> couldn't be dispatched because the queue was full? > > I don't think so. I think that you are looking for the requeue_list > member of struct request_queue. >
Hmm, sorry, but I'm confused. I'm sure I'm missing something simple, since I haven't touched this in a while, so maybe you can quickly point me in the right direction? I see blk_mq_requeue_request() being used by device drivers to retry requests that failed, but if I read the code correctly, the flushed queue seems to be moved to hctx->dispatch when the device driver returned BLK_STS_RESOURCE or BLK_STS_DEV_RESOURCE to blk_mq_dispatch_rq_list(). I thought BLK_STS_RESOURCE was returned by the driver on .queue_rq() to signal there was no more resources on the hardware to service further requests. -- Gabriel Krisman Bertazi
