Re: [PATCH] blk-mq: Avoid that I/O hangs in bt_get()

[Bart Van Assche <bvanassche@acm.org>]

On 10/06/14 20:53, Jens Axboe wrote:
> On 10/06/2014 11:40 AM, Jens Axboe wrote:
>> I've been able to reproduce this this morning, and your patch does seem
>> to fix it. The inc/add logic is making my head spin a bit. And we now
>> end up banging a lot more on the waitqueue lock through
>> prepare_to_wait(), so there's a substantial performance regression to go
>> with the change.
>>
>> I'll fiddle with this a bit and see if we can't retain existing
>> performance properties under tag contention, while still fixing the hang.
> 
> So I think your patch fixes the issue because it just keeps decrementing
> the wait counts, hence waking up a lot more than it should. This is also
> why I see a huge increase in wait queue spinlock time.
> 
> Does this work for you? I think the issue is plainly that we end up
> setting the batch counts too high. But tell me more about the number of
> queues, the depth (total or per queue?), and the fio job you are running.

Hello Jens,

Thanks for looking into this. I can't reproduce the I/O lockup after 
having reverted my patch and after having applied your patch. In the 
test I ran fio was started with the following command-line options:
 
fio --bs=512 --ioengine=libaio --rw=randread --buffered=0 --numjobs=12 
--iodepth=128 --iodepth_batch=64 --iodepth_batch_complete=64 --thread 
--norandommap --loops=2147483648 --runtime=3600 --group_reporting 
--gtod_reduce=1 --name=/dev/sdo --filename=/dev/sdo --invalidate=1

This job was run on a system with 12 CPU threads and against a SCSI 
initiator driver for which the number of hardware contexts had been set 
to 6. Queue depth per hardware queue was set to 127:
$ cat /sys/class/scsi_host/host10/can_queue
127

This is what fio reports about the average queue depth:

IOdepths: 1=0.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=100.0%
  submit: 0=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=100.0%, >=64=0.0%
complete: 0=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=100.0%, >=64=0.0%

While we are at it, how about the patch below ? That patch shouldn't
change any functionality but should make bt_clear_tag() slightly easier
to read.

Thanks,

Bart.

[PATCH] blk-mq: Make bt_clear_tag() easier to read

Eliminate a backwards goto statement from bt_clear_tag().

Signed-off-by: Bart Van Assche <bvanassche@acm.org>
---
 block/blk-mq-tag.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 3d1a956..2c63a2b 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -351,15 +351,12 @@ static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
 		return;
 
 	wait_cnt = atomic_dec_return(&bs->wait_cnt);
+	if (unlikely(wait_cnt < 0))
+		wait_cnt = atomic_inc_return(&bs->wait_cnt);
 	if (wait_cnt == 0) {
-wake:
 		atomic_add(bt->wake_cnt, &bs->wait_cnt);
 		bt_index_atomic_inc(&bt->wake_index);
 		wake_up(&bs->wait);
-	} else if (wait_cnt < 0) {
-		wait_cnt = atomic_inc_return(&bs->wait_cnt);
-		if (!wait_cnt)
-			goto wake;
 	}
 }
 
-- 
1.8.4.5