On Wed, Mar 04, 2026 at 11:11:08AM +0000, John Garry wrote:

On 04/03/2026 06:13, Benjamin Marzinski wrote:

quoted

quoted

+	scsi_mpath_end_request(req);
+
  	/*
  	 * In the MQ case the command gets freed by __blk_mq_end_request,
  	 * so we have to do all cleanup that depends on it earlier.

This looks wrong. We start accounting in scsi_queue_rq(), and we need to
end it whenever we complete or requeue the request, otherwise the
accounting will get off. But not all requests go through
scsi_end_request(). scsi_mpath_failover_req(), for instance, calls
blk_mq_end_request() directly, and other functions, like
scsi_queue_insert() call blk_mq_requeue_request(). I'm pretty sure that
this should go in scsi_complete(), as well in the error path of
scsi_queue_rq().

ok, let me check that further.

I think I was a little hasty here. Looking at sd_mpath_start_command()
and sd_mpath_end_command() in patch 17, I can see that they protect
against repeat calls, so requeueing the request should be o.k. There's
still a problem when scsi_mpath_failover_req() calls blk_mq_end_request()
directly, and when scsi_queue_rq() exits with a failure where the
request won't requeued (all the returns except BLK_STS_OK,
BLK_STS_RESOURCE, and BLK_STS_DEV_RESOURCE).

-Ben

Thanks for the notice.