On Tue, Feb 4, 2020 at 1:20 AM Ming Lei [off-list ref] wrote:

On Mon, Feb 03, 2020 at 12:59:50PM -0800, Salman Qazi wrote:

quoted

We observed that it is possible for a flush to bypass the I/O
scheduler and get added to hctx->dispatch in blk_mq_sched_bypass_insert.

We always bypass io scheduler for flush request.

quoted

This can happen while a kworker is running blk_mq_do_dispatch_sched call
in blk_mq_sched_dispatch_requests.

However, the blk_mq_do_dispatch_sched call doesn't end in bounded time.
As a result, the flush can sit there indefinitely, as the I/O scheduler
feeds an arbitrary number of requests to the hardware.

blk-mq supposes to handle requests in hctx->dispatch with higher
priority, see blk_mq_sched_dispatch_requests().

However, there is single hctx->run_work, so async run queue for dispatching
flush request can only be run until another async run queue from scheduler
is done.

quoted

The solution is to periodically poll hctx->dispatch in
blk_mq_do_dispatch_sched, to put a bound on the latency of the commands
sitting there.

Signed-off-by: Salman Qazi <redacted>
---
 block/blk-mq-sched.c   |  6 ++++++
 block/blk-mq.c         |  4 ++++
 block/blk-sysfs.c      | 33 +++++++++++++++++++++++++++++++++
 include/linux/blkdev.h |  2 ++
 4 files changed, 45 insertions(+)

diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
index ca22afd47b3d..75cdec64b9c7 100644
--- a/block/blk-mq-sched.c
+++ b/block/blk-mq-sched.c

@@ -90,6 +90,7 @@ static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
      struct request_queue *q = hctx->queue;
      struct elevator_queue *e = q->elevator;
      LIST_HEAD(rq_list);
+     int count = 0;

      do {
              struct request *rq;

@@ -97,6 +98,10 @@ static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
              if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
                      break;

+             if (count > 0 && count % q->max_sched_batch == 0 &&
+                 !list_empty_careful(&hctx->dispatch))
+                     break;

q->max_sched_batch may not be needed, and it is reasonable to always
disaptch requests in hctx->dispatch first.

Also such trick is missed in dispatch from sw queue.

I will update the commit message, drop max_sched_batch and just turn
it into a simple check and add the same
thing to the dispatch from the sw queue.

Thanks,
Ming