[dm-devel] [PATCH 7/7] dm-mpath: Fix a race condition in __multipath_map()

[Bart Van Assche]

On 11/21/2016 04:34 PM, Mike Snitzer wrote:
> But you WARN_ON_ONCE(clone && q->mq_ops) will trigger with sq-on-mq.

Hello Mike,

This behavior is indeed triggered by the sq-on-mq test. After having 
added the following code in __bind():

	if (old_map &&
	    dm_table_all_blk_mq_devices(old_map) !=
	    dm_table_all_blk_mq_devices(t))
		pr_debug("%s: old all_blk_mq %d <> new all_blk_mq %d\n",
			 dm_device_name(md),
			 dm_table_all_blk_mq_devices(old_map),
			 dm_table_all_blk_mq_devices(t));

I see the following output appear frequently in the kernel log:

dm_mod:__bind: 254:0: old all_blk_mq 1 <> new all_blk_mq 0

Could these all_blk_mq state changes explain that WARN_ON_ONCE(clone && 
q->mq_ops) is triggered in __multipath_map()? Does this mean that the 
comment in patch http://marc.info/?l=dm-devel&m=147925314306752 is correct?

Thanks,

Bart.