blk-mq request allocation stalls [was: Re: [PATCH v3 0/8] dm: add request-based blk-mq support]

Commit Message

Jens Axboe Jan. 9, 2015, 9:07 p.m. UTC

On 01/09/2015 12:49 PM, Mike Snitzer wrote:
> On Wed, Jan 07 2015 at  3:40pm -0500,
> Keith Busch <[e-mail]> wrote:
>
>> On Wed, 7 Jan 2015, Bart Van Assche wrote:
>>> On 01/06/15 17:15, Jens Axboe wrote:
>>>> blk-mq request allocation is pretty much as optimized/fast as it can be.
>>>> The slowdown must be due to one of two reasons:
>>>>
>>>> - A bug related to running out of requests, perhaps a missing queue run
>>>> or something like that.
>>>> - A smaller number of available requests, due to the requested queue depth.
>>>>
>>>> Looking at Barts results, it looks like it's usually fast, but sometimes
>>>> very slow. That would seem to indicate it's option #1 above that is the
>>>> issue. Bart, since this seems to wait for quite a bit, would it be
>>>> possible to cat the 'tags' file for that queue when it is stuck like that?
>>>
>>> Hello Jens,
>>>
>>> Thanks for the assistance. Is this the output you were looking for
>>
>> I'm a little confused by the later comments given the below data. It says
>> multipath_clone_and_map() is stuck at bt_get, but that doesn't block
>> unless there are no tags available. The tags should be coming from one
>> of dm-1's path queues, and I'm assuming these queues are provided by sdc
>> and sdd. All their tags are free, so that looks like a missing wake_up
>> when the queue idles.
>
> Like I said in an earlier email, I cannot reproduce Bart's hangs running
> mkfs.xfs against a multipath device that is built ontop of a virtio
> device in a KVM guest.
>
> But I can hit __bt_get() failures on the virtio-blk device that I'm
> using for the root device on this guest.  Bart I'd be interested to see
> what you get when running the attached debug patch (likely will just
> echo the same type of info you've already provided).
>
> There does appear to be something weird going on with bt_get().  With
> the debug patch I'm seeing the following when I simply run "make install"
> of the kernel (it'll run dracut to build the initramfs, etc):
>
> You'll note that in all instances where __bt_get() returns -1 nr_free isn't 0.

Yeah, that doesn't look good. Can you try with this patch? The second 
hunk is the interesting bit, the first is more of a cleanup.

Patch

diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 60c9d4a93fe4..363d32d4bae6 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -143,7 +143,6 @@  static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
 static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
 {
 	int tag, org_last_tag, end;
-	bool wrap = last_tag != 0;
 
 	org_last_tag = last_tag;
 	end = bm->depth;
@@ -155,15 +154,16 @@  restart:
 			 * We started with an offset, start from 0 to
 			 * exhaust the map.
 			 */
-			if (wrap) {
-				wrap = false;
+			if (org_last_tag) {
 				end = org_last_tag;
-				last_tag = 0;
+				last_tag = org_last_tag = 0;
 				goto restart;
 			}
 			return -1;
 		}
 		last_tag = tag + 1;
+		if (last_tag >= bm->depth - 1)
+			last_tag = 0;
 	} while (test_and_set_bit(tag, &bm->word));
 
 	return tag;
@@ -199,9 +199,13 @@  static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
 			goto done;
 		}
 
-		last_tag = 0;
-		if (++index >= bt->map_nr)
+		index++;
+		last_tag += (index << bt->bits_per_word);
+
+		if (index >= bt->map_nr) {
 			index = 0;
+			last_tag = 0;
+		}
 	}
 
 	*tag_cache = 0;