Hi Keith,

On Tue, May 8, 2018 at 11:30 PM, Keith Busch [off-list ref] wrote:

On Sat, Apr 28, 2018 at 11:50:17AM +0800, Ming Lei wrote:

quoted

This sync may be raced with one timed-out request, which may be handled
as BLK_EH_HANDLED or BLK_EH_RESET_TIMER, so the above sync queues can't
work reliably.

Ming,

As proposed, that scenario is impossible to encounter. Resetting the
controller inline with the timeout reaps all the commands, and then
sets the controller state to RESETTING. While blk-mq may not allow the
driver to complete those requests, having the driver sync with the queues
will hold the controller in the reset state until blk-mq is done with
its timeout work; therefore, it is impossible for the NVMe driver to
return "BLK_EH_RESET_TIMER", and all commands will be completed through
nvme_timeout's BLK_EH_HANDLED exactly as desired.

That isn't true for multiple namespace case,  each request queue has its
own timeout work, and all these timeout work can be triggered concurrently.

Could you please recheck my suggestion? The alternatives proposed are
far too risky for a 4.17 consideration, and I'm hoping we can stabilize
this behavior in the current release if possible.

Bart's rework on timeout won't cover the case of multiple namespace, so
we have to sync timeout inside driver.

The approach taken in my V4 has been simpler, could you take a look at it?

Thanks,
Ming Lei