Re: [PATCH 3/4] drm: Check mode object lease status in all master ioctl paths

[Daniel Vetter]

On Sun, Apr 02, 2017 at 09:37:16AM -0700, Keith Packard wrote:
> Daniel Vetter <daniel@xxxxxxxx> writes:
> 
> > I think it'd be good if we could consolidate all the lease checking into
> > drm_mode_object_find (respectively __drm_mode_object_find). We'd need to
> > wire up the fpriv to be able to do that, but we could upstream that patch
> > right away before anything else. That should take care of most of the
> > checks in this patch here.
> 
> That's a good idea.
> 
> > There's a few things on top:
> > - filtering the various bitmasks. I think you have most, but we could
> >   perhaps upstream the helpers for these.
> 
> Yeah, would be nice to get hooks in place soon to avoid rebase
> adventures later. I guess that would involve shipping a stub drm_lease.h
> for now?

I think we should see how far wiring drm_file *fpriv through the lookup
functions gets us, but yeah we probably need a stub drm_lease.h for
filtering. For the interface maybe pass a drm_file *fpriv for all the
filtering functions, and deref into fpriv->master or whatever on the
implementation. That gives us all the freedom to filter however we want
really.

> > - filtering object lists (essentially getresources and getplanes ioctls).
> > - filtering implicit objects in the legacy ioctl. E.g. page_flip done
> >   through atomic doesn't just need the CRTC id, but also the id of the
> >   primary plane plus of the FB_ID atomic property. Similarly for all the
> >   other legacy ioctls. I think we want to make sure there's no difference
> >   here in behaviour.
> 
> Oh, all of the implicit resource access from the legacy ioctls. Yeah,
> that will take a bit of research to identify all of them.
> 
> > Especially for the last one it might be simplest to outright disallow all
> > legacy ioctl and require that sub-drm_master nodes only get access to the
> > read-only GET* ioctl (they get that anyway, even when they're not the
> > current master), plus atomic. Makes it a _lot_ easier to implement.
> > Downside is that amdgpu _really_ needs to land atomic asap :-)
> 
> I'd like to avoid that particular dependency as amdgpu is something of a
> requirement for this particular project...

Figured so :-) Otoh I've heard that atomic for amdgpu is happening for
real now. Although Harry told me I need to wait a few more weeks before
they're ready to present the rewritten stuff and get my feedback ...
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch