WARNING: CPU: 2 PID: 207 at drivers/nvme/host/core.c:527 nvme_setup_cmd+0x3d3

[Keith Busch]

On Thu, Feb 01, 2018 at 10:58:23AM -0700, Jens Axboe wrote:
> I was able to reproduce on a test box, pretty trivially in fact:
> 
> # echo mq-deadline > /sys/block/nvme2n1/queue/scheduler
> # mkfs.ext4 /dev/nvme2n1
> # mount /dev/nvme2n1 /data -o discard
> # dd if=/dev/zero of=/data/10g bs=1M count=10k
> # sync
> # rm /data/10g
> # sync <- triggered
> 
> Your patch still doesn't work, but mainly because we init the segments
> to 0 when setting up a discard. The below works for me, and cleans up
> the merge path a bit, since your patch was missing various adjustments
> on both the merged and freed request.

I'm still finding cases not accounted even your patch. I had to use the
following on top of that, and this pattern looks like it needs to be
repeated for all schedulers:

---
diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
index 55c0a745b427..25c14c58385c 100644
--- a/block/blk-mq-sched.c
+++ b/block/blk-mq-sched.c
@@ -259,6 +259,8 @@ bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
 		if (!*merged_request)
 			elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
 		return true;
+	case ELEVATOR_DISCARD_MERGE:
+		return bio_attempt_discard_merge(q, rq, bio);
 	default:
 		return false;
 	}
diff --git a/block/mq-deadline.c b/block/mq-deadline.c
index c56f211c8440..a0f5752b6858 100644
--- a/block/mq-deadline.c
+++ b/block/mq-deadline.c
@@ -451,7 +451,7 @@ static int dd_request_merge(struct request_queue *q, struct request **rq,
 
 		if (elv_bio_merge_ok(__rq, bio)) {
 			*rq = __rq;
-			return ELEVATOR_FRONT_MERGE;
+			return blk_try_merge(__rq, bio);
 		}
 	}
 
--