On Tue, Jan 23, 2018 at 04:54:02PM +0000, Bart Van Assche wrote:

On Wed, 2018-01-24 at 00:49 +0800, Ming Lei wrote:

quoted

On Tue, Jan 23, 2018 at 04:47:11PM +0000, Bart Van Assche wrote:

quoted

On Wed, 2018-01-24 at 00:41 +0800, Ming Lei wrote:

quoted

Could you explain where to call call_rcu()?  call_rcu() can't be used in
IO path at all.

Can you explain what makes you think that call_rcu() can't be used in the
I/O path? As you know call_rcu() invokes a function asynchronously. From
Documentation/RCU/Design/Requirements/Requirements.html:

  The <tt>call_rcu()</tt> function may be used in a number of
  situations where neither <tt>synchronize_rcu()</tt> nor
  <tt>synchronize_rcu_expedited()</tt> would be legal,
  including within preempt-disable code, <tt>local_bh_disable()</tt> code,
  interrupt-disable code, and interrupt handlers.

OK, suppose it is true, do you want to change most of STS_RESOURCE in
all drivers to this way? Why can't we use the generic and simple approach
in this patch?

Please reread my proposal. I did not propose to change any block drivers.
What I proposed is to change the blk_mq_delay_run_hw_queue() implementation
such that it uses call_rcu() instead of kblockd_mod_delayed_work_on().
That's a generic and simple approach.

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.

-- 
Ming