On Thu, Jan 18, 2018 at 05:13:53PM +0000, Bart Van Assche wrote:

On Thu, 2018-01-18 at 11:50 -0500, Mike Snitzer wrote:

quoted

The issue you say it was originally intended to fix _should_ be
addressed with this change:
https://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=dm-4.16&id=4dd6edd23e7ea971efddc303f9e67eb79e95808e

Hello Mike,

Sorry but I'm not convinced that that patch is sufficient. That patch helps
if .end_io() is called with status BLK_STS_RESOURCE and also if
blk_insert_cloned_request() returns the .queue_rq() return value. It does not
help if .queue_rq() returns BLK_STS_RESOURCE and that return value gets
ignored.

The return value from .queue_rq() is handled by blk-mq, why do you think
it can be ignored? Please see blk_mq_dispatch_rq_list().

I think that can happen as follows:
- Request cloning in multipath_clone_and_map() succeeds and that function
  returns DM_MAPIO_REMAPPED.
- dm_dispatch_clone_request() calls blk_insert_cloned_request().
- blk_insert_cloned_request() calls blk_mq_request_direct_issue(), which
  results in a call of __blk_mq_try_issue_directly().
- __blk_mq_try_issue_directly() calls blk_mq_sched_insert_request(). In this

This only happens iff queue is stopped or quiesced, then we return
BLK_STS_OK to blk-mq via .queue_rq(), please see __blk_mq_try_issue_directly(),
how does this cause IO hang? 

  case the BLK_STS_RESOURCE returned by the .queue_rq() implementation of the
  underlying path will be ignored.

No, this case won't return BLK_STS_RESOURCE.

-- 
Ming