On 2023/7/10 21:33, Christoph Hellwig wrote: > On Mon, Jul 10, 2023 at 02:47:05PM +0800, chengming.zhou@xxxxxxxxx wrote: >> From: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx> >> >> Now we unconditionally blk_rq_init_flush() to replace rq->end_io to >> make rq return twice back to the flush state machine for post-flush. >> >> Obviously, non post-flush requests don't need it, they don't need to >> end request twice, so they don't need to replace rq->end_io callback. >> And the same for requests with the FUA bit on hardware with FUA support. >> >> So we move blk_rq_init_flush() to REQ_FSEQ_DATA stage and only replace >> rq->end_io if it needs post-flush. Otherwise, it can end like normal >> request and doesn't need to return back to the flush state machine. > > I really like the idea behind this optimization, but I kinda hate > adding more magic to the already way too magic flush sequence. Yes, agree. > > I wonder if a better idea would be to kill the flush sequence entirely, > and just split the flush_queue into a preflush and a postflush queue. > This would remove a field from struct request and lead to more readable > code as well. I have thought about this for some time, it seems feasible. So I implement it today and test it using blktests, it works. I will send the patchset soon. Thanks.
