On Fri, Jun 15, 2018 at 10:22 AM, jianchao.wang <jianchao.w.wang@xxxxxxxxxx> wrote: > Hi Ming > > On 06/15/2018 10:17 AM, Ming Lei wrote: >> On Fri, Jun 15, 2018 at 9:57 AM, Jianchao Wang >> <jianchao.w.wang@xxxxxxxxxx> wrote: >>> After f6e7d48 (block: remove BLK_EH_HANDLED), LLDD is responsible >>> to complete the timed out request, however, for blk-legacy, the >>> 'complete' is still marked, blk_complete_request will do nothing, >>> we export __blk_complete_request for LLDD to complete the request >>> in timeout path. >>> >>> Signed-off-by: Jianchao Wang <jianchao.w.wang@xxxxxxxxxx> >>> --- >>> block/blk-softirq.c | 1 + >>> 1 file changed, 1 insertion(+) >>> >>> diff --git a/block/blk-softirq.c b/block/blk-softirq.c >>> index 01e2b35..15c1f5e 100644 >>> --- a/block/blk-softirq.c >>> +++ b/block/blk-softirq.c >>> @@ -144,6 +144,7 @@ void __blk_complete_request(struct request *req) >>> >>> local_irq_restore(flags); >>> } >>> +EXPORT_SYMBOL(__blk_complete_request); >>> >>> /** >>> * blk_complete_request - end I/O on a request >>> -- >>> 2.7.4 >>> >> >> Looks non-blk-mq timeout code need to convert to ref-counter >> based approach too? > > IMO, ref-counter is just to fix the blk-mq req life recycle issue. Just thought of that, it is one blk-mq specific issue. > It cannot replace the blk_mark_rq_complete which could avoid the race between > timeout and io completion path. > Or maybe my understanding is wrong ... I didn't mean that this patch is unnecessary. But the question is that given driver has to deal with race between timeout and normal completion, why don't you follow blk-mq's way to move the atomic state change into __blk_complete_request()? Thanks, Ming Lei
