blk-mq: fix mismatch between IO scheduler insert and finish

Commit Message

Chengming Zhou Aug. 19, 2023, 3:12 a.m. UTC

From: Chengming Zhou <[e-mail]>

IO scheduler has requirement that one request which has been inserted
must call finish_request() only once.

Now we have three special cases to consider:
1. rq has not insert, has complete: e.g. empty preflush
2. rq has insert, has not complete: e.g. merged requests will be freed
3. rq has insert, has twice complete: e.g. postflushes

Note case 1 which existed before, has been no problem since all the
schedulers will check in their finish_request() if the rq has been
inserted or not, like checking "rq->elv.priv[0]".

Then case 2 and case 3 are the introduced regression, we moved the
scheduler finish_request() from free phase to complete phase to solve
a deadlock problem. But it caused no finish_request() for request in
case 2, and double finish_request() for request in case 3.

So we still need finish_request() in blk_mq_free_request() to cover
case 2. And clear RQF_USE_SCHED flag to avoid double finish_request().
It should be fine since we're freeing the request now anyway.

Of course, we can also make all schedulers' finish_request() to clear
"rq->elv.priv[0]" to avoid double finish. Or clear it in blk-mq, make
the rq like not inserted as case 1.

FYI it's easy to reproduce warning in mq-deadline using this:
```
DEV=sdb
echo mq-deadline > /sys/block/$DEV/queue/scheduler
mkfs.ext4 /dev/$DEV
mount /dev/$DEV /mnt
cd /mnt
stress-ng --symlink 4 --timeout 60
echo none > /sys/block/$DEV/queue/scheduler
```

Reported-by: kernel test robot <[e-mail]>
Closes: https://lore.kernel.org/oe-lkp/[e-mail]
Signed-off-by: Chengming Zhou <[e-mail]>
---
 block/blk-mq.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

Comments

Jens Axboe Aug. 19, 2023, 1:49 p.m. UTC | #1

On 8/18/23 9:12 PM, [e-mail] wrote:
> From: Chengming Zhou <[e-mail]>
> 
> IO scheduler has requirement that one request which has been inserted
> must call finish_request() only once.
> 
> Now we have three special cases to consider:
> 1. rq has not insert, has complete: e.g. empty preflush
> 2. rq has insert, has not complete: e.g. merged requests will be freed
> 3. rq has insert, has twice complete: e.g. postflushes
> 
> Note case 1 which existed before, has been no problem since all the
> schedulers will check in their finish_request() if the rq has been
> inserted or not, like checking "rq->elv.priv[0]".
> 
> Then case 2 and case 3 are the introduced regression, we moved the
> scheduler finish_request() from free phase to complete phase to solve
> a deadlock problem. But it caused no finish_request() for request in
> case 2, and double finish_request() for request in case 3.
> 
> So we still need finish_request() in blk_mq_free_request() to cover
> case 2. And clear RQF_USE_SCHED flag to avoid double finish_request().
> It should be fine since we're freeing the request now anyway.
> 
> Of course, we can also make all schedulers' finish_request() to clear
> "rq->elv.priv[0]" to avoid double finish. Or clear it in blk-mq, make
> the rq like not inserted as case 1.
> 
> FYI it's easy to reproduce warning in mq-deadline using this:
> ```
> DEV=sdb
> echo mq-deadline > /sys/block/$DEV/queue/scheduler
> mkfs.ext4 /dev/$DEV
> mount /dev/$DEV /mnt
> cd /mnt
> stress-ng --symlink 4 --timeout 60
> echo none > /sys/block/$DEV/queue/scheduler
> ```
> 
> Reported-by: kernel test robot <[e-mail]>
> Closes: https://lore.kernel.org/oe-lkp/[e-mail]
> Signed-off-by: Chengming Zhou <[e-mail]>

I folded in this one and added a link to it as well, final result is here:

https://git.kernel.dk/cgit/linux/commit/?h=block-6.5&id=e5c0ca13659e9d18f53368d651ed7e6e433ec1cf

I'll get this sent off today.

Patch

diff --git a/block/blk-mq.c b/block/blk-mq.c
index a6d59320e034..953f08354c8c 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -685,8 +685,15 @@  static void blk_mq_finish_request(struct request *rq)
 {
 	struct request_queue *q = rq->q;
 
-	if (rq->rq_flags & RQF_USE_SCHED)
+	if (rq->rq_flags & RQF_USE_SCHED) {
 		q->elevator->type->ops.finish_request(rq);
+		/*
+		 * For postflush request that may need to be
+		 * completed twice, we should clear this flag
+		 * to avoid double finish_request() on the rq.
+		 */
+		rq->rq_flags &= ~RQF_USE_SCHED;
+	}
 }
 
 static void __blk_mq_free_request(struct request *rq)
@@ -715,6 +722,8 @@  void blk_mq_free_request(struct request *rq)
 {
 	struct request_queue *q = rq->q;
 
+	blk_mq_finish_request(rq);
+
 	if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 		laptop_io_completion(q->disk->bdi);