[PATCH] drm: support up to 128 drm devices

[Simon Ser]

On Monday, July 17th, 2023 at 15:24, Emil Velikov <emil.l.velikov at gmail.com> wrote:

> > > For going forward, here is one way we can shave this yak:
> > >  - update libdrm to max 64 nodes
> > >  - roll libdrm release, nag distributions to update to it // could be
> > > folded with the next release below
> > >  - update libdrm to use name driven type detection
> > >  - roll libdrm release, nag distributions to update to it
> > >  - once ^^ release hits most distributions, merge the above proposed
> > > kernel patch
> > >    - the commit message should explain the caveats and fixed libdrm version
> > >    - we should be prepared to revert the commit, if it causes user
> > > space regression - fix (see below) and re-introduce the kernel patch
> > > 1-2 releases later
> >
> > That sounds really scary to me. I'd really prefer to try not to break the
> > kernel uAPI here.
> >
> 
> With part in particular? Mind you I'm not particularly happy either,
> since in essence it's like a controlled explosion.

I believe there are ways to extend the uAPI to support more devices without
breaking the uAPI. Michał Winiarski's patch for instance tried something to
this effect.

> > The kernel rule is "do not break user-space".
> 
> Yes, in a perfect world. In practice, there have been multiple kernel
> changes breaking user-space. Some got reverted, some remained.
> AFAICT the above will get us out of the sticky situation we're in with
> the least amount of explosion.
> 
> If there is a concrete proposal, please go ahead and sorry if I've
> missed it. I'm supposed to be off, having fun with family when I saw
> this whole thing explode.
> 
> Small note: literally all the users I've seen will stop on a missing
> node (card or render) aka if the kernel creates card0...63 and then
> card200... then (hard wavy estimate) 80% of the apps will be broken.

That's fine, because that's not a kernel regression. Supporting more than 64
devices is a new kernel feature, and if some user-space ignores the new nodes,
that's not a kernel regression. A regression only happens when a use-case which
works with an older kernel is broken by a newer kernel.