On 2018/11/21 11:11, Jens Axboe wrote:

On 11/20/18 4:58 PM, Damien Le Moal wrote:

quoted

On 2018/11/21 2:31, Jens Axboe wrote:

quoted

I think the below should fix it, we haven't necessarily setup an
ioc if we're just doing as passthrough request.

diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
index 13b8dc332541..f096d8989773 100644
--- a/block/blk-mq-sched.c
+++ b/block/blk-mq-sched.c

@@ -34,9 +34,16 @@ EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
 void blk_mq_sched_assign_ioc(struct request *rq)
 {
 	struct request_queue *q = rq->q;
-	struct io_context *ioc = current->io_context;
+	struct io_context *ioc;
 	struct io_cq *icq;
 
+	/*
+	 * May not have an IO context if it's a passthrough request
+	 */
+	ioc = current->io_context;
+	if (!ioc)
+		return;
+
 	spin_lock_irq(&q->queue_lock);
 	icq = ioc_lookup_icq(ioc, q);
 	spin_unlock_irq(&q->queue_lock);

This seems reasonable to me, but I wonder why this problem was not triggering
before. The previous code getting the ioc with the rq_ioc(bio) call was
essentially the same and there was no "if (!ioc) return;" in
blk_mq_sched_assign_ioc() before the patch.
Any idea why this is popping up now ?

Ming,

Is this a new test your are running ? If this same problem triggers on stable
kernels, Jens patch needs to go to stable too.

No, it's definitely introduced in your patches:

-                       if (e->type->icq_cache && rq_ioc(bio))
-                               blk_mq_sched_assign_ioc(rq, bio);
+                       if (e->type->icq_cache)
+                               blk_mq_sched_assign_ioc(rq);

Arg ! Yes, I missed this. My apologies.

Please run blktests on a series. Always. There's no excuse not to.

I did run my usual tests exercising drives with various fio workloads. But I did
not run blktests itself. I will fix my workflow to include it.

Thanks.


-- 
Damien Le Moal
Western Digital Research