On Tue, Jan 23, 2018 at 10:01:37PM +0000, Bart Van Assche wrote:

On Wed, 2018-01-24 at 00:59 +0800, Ming Lei wrote:

quoted

How is that enough to fix the IO hang when driver returns STS_RESOURCE
and the queue is idle? If you want to follow previous dm-rq's way of
call blk_mq_delay_run_hw_queue() in .queue_rq(), the same trick need
to be applied to other drivers too, right?

Unfortunately most of STS_RESOURCE don't use this trick, but they need
to be handled.

The patch of 'blk-mq: introduce BLK_STS_DEV_RESOURCE' can fix all these
cases.

The goal of my proposal was to address the race between running the queue and
adding requests back to the dispatch queue only. Regarding the I/O hangs that
can occur if a block driver returns BLK_STS_RESOURCE: since all of these can
be addressed by inserting blk_mq_delay_run_hw_queue() calls in the affected
block drivers I prefer to modify the block drivers instead of making the
blk-mq core even more complicated.

IMO, this change doesn't make blk-mq code more complicated, also it is
well documented, see the change in block layer:

 block/blk-core.c             |  1 +
 block/blk-mq.c               | 19 +++++++++++++++----
 include/linux/blk_types.h    | 18 ++++++++++++++++++

Also 21 lines of them are comment, so only 17 lines code change needed
in block layer.

If inserting blk_mq_delay_run_hw_queue() to driver, the change can be
a bit complicated, since call_rcu has to be used, you need to figure out
one way to provide callback and the parameter. Even you have to document
the change in each driver.

[ming@ming linux]$ git grep -n BLK_STS_RESOURCE drivers/ | wc -l
42

There are at least 42 uses of BLK_STS_RESOURCE in drivers, in theory
you should insert call_rcu(blk_mq_delay_run_hw_queue) in every BLK_STS_RESOURCE
of drivers.

I leave the decisions to Jens and drivers maintainers.

Thanks,
Ming