If the block layer has entered requests and gets a CPU hot plug event
prior to the resume event, it will wait for those requests to exit. If
the nvme driver is shutting down, it will not start the queues back up,
preventing forward progress.
To fix that, this patch freezes the request queues when the driver intends
to shut down the controller so that no new requests may enter. After the
controller has been disabled, the queues will be restarted to force all
entered requests to end in failure so that blk-mq's hot cpu notifier may
progress. To ensure the queue usage count is 0 on a shutdown, the driver
waits for freeze to complete before completing the controller shutdown.
Keith, can you explain (again) for me why is the freeze_wait must happen
after the controller has been disabled, instead of starting the queues
and waiting right after freeze start?
Yeah, the driver needs to make forward progress even if the controller
isn't functioning. If we do the freeze wait before disabling the
controller, there's no way to reclaim missing completions. If the
controller is working perfectly, it'd be okay, but the driver would be
stuck if there's a problem.
OK, I think we can get it for fabrics too, need to figure out how to
handle it there too.
Do you have a reproducer?
