On Thu, 2018-12-20 at 06:07 -0700, Jens Axboe wrote: +AD4 On 12/20/18 6:02 AM, Jens Axboe wrote: +AD4 +AD4 +AD4 I'm afraid this cannot work. +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 The 'tags' here could be the hctx-+AD4-sched+AF8-tags, but what we need to +AD4 +AD4 +AD4 clear is hctx-+AD4-tags-+AD4-rqs+AFsAXQ. +AD4 +AD4 +AD4 +AD4 You are right, of course, a bit too quick on the trigger. This one +AD4 +AD4 should work better, and also avoids that silly quadratic loop. I don't +AD4 +AD4 think we need the tag +AD0APQ -1 check, but probably best to be safe. +AD4 +AD4 Sent out the wrong one, here it is. Bart, if you can reproduce, can you +AD4 give it a whirl? +AD4 +AD4 +AD4 diff --git a/block/blk-mq.c b/block/blk-mq.c +AD4 index 2de972857496..fc04bb648f18 100644 +AD4 --- a/block/blk-mq.c +AD4 +-+-+- b/block/blk-mq.c +AD4 +AEAAQA -2025,7 +-2025,7 +AEAAQA void blk+AF8-mq+AF8-free+AF8-rqs(struct blk+AF8-mq+AF8-tag+AF8-set +ACo-set, struct blk+AF8-mq+AF8-tags +ACo-tags, +AD4 +AHs +AD4 struct page +ACo-page+ADs +AD4 +AD4 - if (tags-+AD4-rqs +ACYAJg set-+AD4-ops-+AD4-exit+AF8-request) +AHs +AD4 +- if (tags-+AD4-rqs) +AHs +AD4 int i+ADs +AD4 +AD4 for (i +AD0 0+ADs i +ADw tags-+AD4-nr+AF8-tags+ADs i+-+-) +AHs +AD4 +AEAAQA -2033,8 +-2033,14 +AEAAQA void blk+AF8-mq+AF8-free+AF8-rqs(struct blk+AF8-mq+AF8-tag+AF8-set +ACo-set, struct blk+AF8-mq+AF8-tags +ACo-tags, +AD4 +AD4 if (+ACE-rq) +AD4 continue+ADs +AD4 - set-+AD4-ops-+AD4-exit+AF8-request(set, rq, hctx+AF8-idx)+ADs +AD4 +- if (set-+AD4-ops-+AD4-exit+AF8-request) +AD4 +- set-+AD4-ops-+AD4-exit+AF8-request(set, rq, hctx+AF8-idx)+ADs +AD4 tags-+AD4-static+AF8-rqs+AFs-i+AF0 +AD0 NULL+ADs +AD4 +- +AD4 +- if (rq-+AD4-tag +AD0APQ -1) +AD4 +- continue+ADs +AD4 +- if (set-+AD4-tags+AFs-hctx+AF8-idx+AF0--+AD4-rqs+AFs-rq-+AD4-tag+AF0 +AD0APQ rq) +AD4 +- set-+AD4-tags+AFs-hctx+AF8-idx+AF0--+AD4-rqs+AFs-rq-+AD4-tag+AF0 +AD0 NULL+ADs +AD4 +AH0 +AD4 +AH0 +AD4 +AD4 +AEAAQA -2113,6 +-2119,7 +AEAAQA static int blk+AF8-mq+AF8-init+AF8-request(struct blk+AF8-mq+AF8-tag+AF8-set +ACo-set, struct request +ACo-rq, +AD4 return ret+ADs +AD4 +AH0 +AD4 +AD4 +- rq-+AD4-tag +AD0 -1+ADs +AD4 WRITE+AF8-ONCE(rq-+AD4-state, MQ+AF8-RQ+AF8-IDLE)+ADs +AD4 return 0+ADs +AD4 +AH0 Hi Jens, Are you sure this is sufficient? My understanding is that blk+AF8-mq+AF8-queue+AF8-tag+AF8-busy+AF8-iter() iterates over all tags in the tag set. So if the request queue on which part+AF8-in+AF8-flight() is called and the request queue for which blk+AF8-mq+AF8-free+AF8-rqs() is called share their tag set then part+AF8-in+AF8-flight() and blk+AF8-mq+AF8-free+AF8-rqs() can run concurrently. That can cause ugly race conditions. Do you think it would be a good idea to modify the inflight accounting code such that it only considers the requests of a single request queue instead of all requests for a given tag set? Thanks, Bart.
