On Fri, Aug 25, 2017 at 12:32:33PM -0400, Mike Snitzer wrote:

On Fri, Aug 25 2017 at 12:08pm -0400,
Ming Lei [off-list ref] wrote:

quoted

On Fri, Aug 25, 2017 at 11:48:39AM -0400, Mike Snitzer wrote:

quoted

On Fri, Aug 25 2017 at 11:27am -0400,
Ming Lei [off-list ref] wrote:

quoted

We don't need to update orignal dm request partially
when ending each cloned bio, and this patch just
updates orignal dm request once when the whole
cloned request is finished.

Partial request update can be a bit expensive, so
we should try to avoid it, especially it is run
in softirq context.

After this patch is applied, both hard lockup and
soft lockup aren't reproduced any more in one hour
of running Laurence's test[1] on IB/SRP. Without
this patch, the lockup can be reproduced in several
minutes.

BTW, after d4acf3650c7c(block: Make blk_mq_delay_kick_requeue_list()
rerun the queue at a quiet time), we need to make the
test more aggressive for reproducing the lockup:

	1) run hammer_write.sh 32 or 64 concurrently.
	2) write 8M each time

[1] https://marc.info/?l=linux-block&m=150220185510245&w=2

Bart said he cannot reproduce the lockups with his patchset applied.
Have you tested using Bart's patchset?

d4acf3650c7c(block: Make blk_mq_delay_kick_requeue_list() rerun the
queue at a quiet time) has been in linus tree.

For other patches, I didn't test it yet. Because every time
when the lockup is triggered, it is always in blk_recalc_rq_segments(),
and not see any patch is dealing with that.

Please test with all of Bart's patches applied!

Just done the test with Bart's patch, still can
see soft lockup when running the test described
in commit log for a couple of minutes. BTW, my
test is much more aggressive than Laurence's, I
write 8M each time, and run 64 hammer_write.sh
concurrently.

I don't think the two are contradictory. Anyway,
this patch will decrease CPU utilization of SOFTIRQ, and
it is a improvement.

quoted

quoted

Comments inlined below.

quoted

---
 drivers/md/dm-rq.c | 15 +++++----------
 drivers/md/dm-rq.h |  1 +
 2 files changed, 6 insertions(+), 10 deletions(-)

diff --git a/drivers/md/dm-rq.c b/drivers/md/dm-rq.c
index c6ebc5b1e00e..50cd96c7de45 100644
--- a/drivers/md/dm-rq.c
+++ b/drivers/md/dm-rq.c

@@ -117,9 +117,9 @@ static void end_clone_bio(struct bio *clone)
 	struct dm_rq_clone_bio_info *info =
 		container_of(clone, struct dm_rq_clone_bio_info, clone);
 	struct dm_rq_target_io *tio = info->tio;
-	struct bio *bio = info->orig;
 	unsigned int nr_bytes = info->orig->bi_iter.bi_size;
 	blk_status_t error = clone->bi_status;
+	bool is_last = !clone->bi_next;
 
 	bio_put(clone);
 

@@ -144,21 +144,15 @@ static void end_clone_bio(struct bio *clone)
 	 * I/O for the bio successfully completed.
 	 * Notice the data completion to the upper layer.
 	 */
-
-	/*
-	 * bios are processed from the head of the list.
-	 * So the completing bio should always be rq->bio.
-	 * If it's not, something wrong is happening.
-	 */
-	if (tio->orig->bio != bio)
-		DMERR("bio completion is going in the middle of the request");

Why did you remove this check?

The check isn't valid any more because this patch only update
original dm rq in .end_bio() of the last cloned bio.

Fair enough, so just a side-effect of your all or nothing completion handling.

It is not all or nothing, and still can do partial update like
current way.

quoted

quoted

quoted

+	tio->completed += nr_bytes;
 
 	/*
 	 * Update the original request.
 	 * Do not use blk_end_request() here, because it may complete
 	 * the original request before the clone, and break the ordering.
 	 */
-	blk_update_request(tio->orig, BLK_STS_OK, nr_bytes);
+	if (is_last)
+		blk_update_request(tio->orig, BLK_STS_OK, tio->completed);
 }

Partial completion support is important given the potential for path
failures interrupting requests.  Why do you think it is OK to avoid it
by switching to an all or nothing approach?

If the cloned rq is partially completed, this dm rq is still partially
completed. This patch only update dm rq in the last cloned bio's
.end_io().

Which is exactly what we do _not_ want because it removed partial
completion.  Which is an important feature of DM multipath.  Christoph
echoed its importance some time ago:
https://www.redhat.com/archives/dm-devel/2015-May/msg00228.html

quoted

Also if one middle cloned bio is completed with error, the current
implementation doesn't update dm rq any more from that bio, so
looks the following patch is consistent with current
implementation, what do you think of it?

Not seeing how.  Yes, the current implementation will not account for a
part of the request that failed.  That is fine.. the bio failed, so
nothing to update!

So no, I don't understand why you've added the 'exit' goto to update the
request on error.  If the request is to be failed upgrades it'll get

It is for updating the previous completed(successful) bios because
'tio->completed' records the completed/successful bytes, that is
same with the current implementation.

The difference is that this patch calls one blk_update_request()
to do that and it is a whole rq update most of times, but the
current way calls one blk_update_request() for each cloned bio
to do that. Since blk_update_request() on partial completion need
to recalculate segments, and it is very expensive, we should avoid that.

blk_update_request() is called in SOFTIRQ context, that is why
I believe it is the main reason of the lockup, and my test shows
that no lockup any more with this patch.

Hope I explain it clearly, :-)

-- 
Ming