[PATCH 0/3] NVMe: Introduce CMB allocation scheme

[Jon Derrick]

On Wed, Jan 13, 2016 at 12:08:39AM +0000, Keith Busch wrote:
> On Wed, Jan 06, 2016 at 08:13:08PM +0000, Keith Busch wrote:
> > On Wed, Dec 30, 2015 at 10:47:56AM -0700, Jon Derrick wrote:
> > > This patchset changes the CMB allocation scheme. Instead of reserving
> > > the entire range for SQes and automatically placing the SQes in the CMB,
> > > this set creates sysfs knobs to manage it. This allows partial usage of
> > > the CMB for SQes, so that the remainder can be mmapped or used for other
> > > NVMe-defined CMB usages.
> > 
> > Thanks, Jon.
> > 
> > In general, I like the control and visibility this provides. With this,
> > I think we're on the verge of warranting a new drivers/nvme/host/sysfs.c
> > file to keep all the nvme sysfs management in one place.
> > 
> > I don't think anyone is in a rush to see this in the 4.5 merge window,
> > so I'll load this on a CMB capable machine and provide a more thorough
> > review in the next week or so.
> 
> One potential problem, this allows the queue depth to change on a live
> request_queue. The block layer won't allow that right now. That can
> also theoretically happen on a controller reset, but doesn't happen in
> real life (that I know of). For this CMB proposal to really be useful,
> we need a new blk-mq API to let us change the depth at runtime.
> 

Yep. I noticed the same thing today. It works coinicidentally -if- the
depth remains the same before and after CMB allocation.

> I thought the dynamic h/w contexts patch[1] was 'ok' and may have
> provided a route to change the depth as well. That patch, though, saw a
> 0-day boot failure. The failure was incorrectly sent to my spam folder,
> and I only noticed it this week. If that can be acceptably resolved,
> then it looks possible to alter the queue depth with a bit more work.

I've cleaned up the nvme-side depth dependencies but still don't have
anything to modify the hardware contexts. I'll take a look at your patch
in the next couple of days.