Re: [PATCH] nvme-core: Fix subsystem instance mismatches
From: Logan Gunthorpe
Date: Tue Sep 03 2019 - 14:13:51 EST
On 2019-09-03 10:46 a.m., Keith Busch wrote:
> On Tue, Sep 03, 2019 at 10:08:01AM -0600, Logan Gunthorpe wrote:
>> On 2019-08-31 9:29 a.m., Keith Busch wrote:
>>> On Fri, Aug 30, 2019 at 06:01:39PM -0600, Logan Gunthorpe wrote:
>>>> To fix this, assign the subsystem's instance based on the instance
>>>> number of the controller's instance that first created it. There should
>>>> always be fewer subsystems than controllers so the should not be a need
>>>> to create extra subsystems that overlap existing controllers.
>>> The subsystem's lifetime is not tied to the controller's. When the
>>> controller is removed and releases its instance, the next controller
>>> to take that available instance will create naming collisions with the
>>> subsystem still using it.
>> Hmm, yes, ok.
>> So perhaps we can just make the subsystem prefer the ctrl's instance
>> when allocating the ID? Then at least, in the common case, the
>> controller numbers will match the subsystem numbers. Only when there's
>> random hot-plugs would the numbers get out of sync.
> I really don't know about a patch that works only on static
> configurations. Connects and disconnects do happen on live systems,
> so the numerals will inevitably get out of sync.
Well this depends on how big a problem we think the number mismatch is.
Right now it's pretty annoying because numbers aren't matching for
non-CMIC controllers in simple setups on boot. I think having a small
patch that makes it more consistent for the static would be worth it and
if CMIC controllers with significant hot-plug events have mismatches
that seems more understandable to me.
> Could we possibly make /dev/nvmeX be a subsystem handle without causing
> trouble for anyone? This would essentially be the same thing as today
> for non-CMIC controllers with a device-per-controller and only affects
> the CMIC ones.
Well then we'd have to be able to do everything that's possible with a
controller via the subsystem and it would have to multiplex all admin
commands for CMIC ones, etc to a sensible controller. This might make
sense in the long term but it sounds like a larger project than I have
time to take on.