Re: [RFC PATCH v2 2/2] blk-mq: Lockout tagset iter when freeing rqs

From: John Garry
Date: Wed Dec 23 2020 - 06:12:21 EST


On 22/12/2020 16:16, Bart Van Assche wrote:
On 12/22/20 3:15 AM, John Garry wrote:
So then we could have something like this:

---8<---

  -435,9 +444,13 @@ void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
     if (!blk_mq_hw_queue_mapped(hctx))
             continue;

+    while (!atomic_inc_not_zero(&tags->iter_usage_counter));
+
     if (tags->nr_reserved_tags)
         bt_for_each(hctx, tags->breserved_tags, fn, priv, true);
     bt_for_each(hctx, tags->bitmap_tags, fn, priv, false);

+    atomic_dec(&tags->iter_usage_counter);
}

blk_queue_exit(q);

--->8---

And similar for blk_mq_tagset_busy_iter(). How about it?


Hi Bart,

Are there any blk_mq_tagset_busy_iter() calls that happen from a context where the tag set can disappear while that function is in progress?


So isn't the blk_mq_tag_set always a member of the host driver data for those cases, and, since blk_mq_tagset_busy_iter() is for iter'ing block driver tags and called from block driver or hctx_busy_show(), it would exist for the lifetime of the host device.

Some blk_mq_tagset_busy_iter() calls happen from a context where it is not allowed to sleep but also where it is guaranteed that the tag set won't disappear, e.g. the call from inside sdk_mq_queue_rq().

You're talking about skd_mq_queue_rq() -> skd_in_flight() -> blk_mq_tagset_busy_iter(), right?

So I would expect any .queue_rq calls to complete before the associated request queue and tagset may be unregistered.


How about using a mutex inside blk_mq_queue_tag_busy_iter() instead? As far as I can see all blk_mq_queue_tag_busy_iter() happen from a context where it is allowed to sleep.

Well then it seems sensible to add might_sleep() also.

And we still have the blk_mq_queue_tag_busy_iter() problem. As Ming mentioned yesterday, we know contexts where from where it is called which may not sleep.

Thanks,
John