[PATCH 04/10] blk-mq: kill undead requests during CPU hotplug notify

[Keith Busch]

On Sun, 27 Sep 2015, Christoph Hellwig wrote:
> REQ_NO_TIMEOUT requests can linger out basically forever, so we need
> to kill them before waiting for all requests to complete in the
> hotplug CPU notification.  Note that we really should be doing an
> abort here, but that will require much more invasive changes to allow
> certain preemptable requests to be started while the queue has already
> been frozen.  As this new code isn't any worse than the early request
> freeing done currently by the nvme driver I'd like to get it in for now
> and sort the full hornet's nest later.

The command is still owned by the device and breaks if the controller
happens to complete the command after a cpu hot event. This was 'ok'
when the driver provided special completion handling.

We'd have to reset the controller to reliably recover the command,
but that's a bit heavy handed.