On Thu, Sep 06, 2018 at 10:19:27PM +0800, jianchao.wang wrote: > > > On 09/06/2018 09:07 PM, Ming Lei wrote: > > On Thu, Sep 06, 2018 at 09:28:46AM +0800, jianchao.wang wrote: > >> Hi Ming > >> > >> On 09/06/2018 06:09 AM, Ming Lei wrote: > >>> nvme_dev_add() still may call freeze & unfreeze queue, so your patch > >>> can't avoid draining queue completely here. > >> > >> Yes, I know this. We still need to freeze queue when update nr_hw_queues. > >> But we move forward a step at least. :) > >> We don't need to drain request queue in normal case of nvme_reset_work. > > > > It is hard to say who is the normal case. In case of CPU hotplug, it is quite > > easy to trigger updating nr_hw_queues. > > > >> > >> As for updating nr_hw_queues, we could try some other method on it next. > > > > The thing is that draining queue may be inevitable inside reset work > > function because of updating nr_hw_queues, that means the approach in > > this patchset is just a partial solution > > > > Or it may not make sense to do that because we may never remove the draining > > queue in the code path of updating nr_hw_queues. > > > > Given they belong to same issue, I suggest to solve them all. > > I still think if just prevent new IO, freeze queue is over-kill. > > With respect to update nr_hw_queues, we could try to take the bios down from the > queued requests as the blk_steal_bios. Then re-issue these bios after updating the > nr_hw_queues. It is always easier to say than done, :-) Please go ahead and post patches, and I am happy to review. Thanks, Ming
