[PATCH 01/13] drm/amdgpu: introduce and honour DRM_FORCE_AUTH workaround

Daniel Vetter daniel at ffwll.ch
Tue May 28 08:18:40 UTC 2019

On Tue, May 28, 2019 at 10:03 AM Koenig, Christian
<Christian.Koenig at amd.com> wrote:
> Am 28.05.19 um 09:38 schrieb Daniel Vetter:
> > [SNIP]
> >> Might be a good idea looking into reverting it partially, so that at
> >> least command submission and buffer allocation is still blocked.
> > I thought the issue is a lot more than vainfo, it's pretty much every
> > hacked up compositor under the sun getting this wrong one way or
> > another. Thinking about this some more, I also have no idea how you'd
> > want to deprecate rendering on primary nodes in general. Apparently
> > that breaks -modesetting already, and probably lots more compositors.
> > And it looks like we're finally achieve the goal kms set out to 10
> > years ago, and new compositors are sprouting up all the time. I guess
> > we could just break them all (on new hardware) and tell them to all
> > suck it up. But I don't think that's a great option. And just
> > deprecating this on amdgpu is going to be even harder, since then
> > everywhere else it'll keep working, and it's just amdgpu.ko that looks
> > broken.
> >
> > Aside: I'm not supporting Emil's idea here because it fixes any issues
> > Intel has - Intel doesn't care. I support it because reality sucks,
> > people get this render vs. primary vs. multi-gpu prime wrong all the
> > time (that's also why we have hardcoded display+gpu pairs in mesa for
> > the various soc combinations out there), and this looks like a
> > pragmatic solution. It'd be nice if every compositor and everything
> > else would perfectly support multi gpu and only use render nodes for
> > rendering, and only primary nodes for display. But reality is that
> > people hack on stuff until gears on screen and then move on to more
> > interesting things (to them). So I don't think we'll ever win this :-/
> Yeah, but this is a classic case of working around user space issues by
> making kernel changes instead of fixing user space.
> Having privileged (output control) and unprivileged (rendering control)
> functionality behind the same node is a mistake we have made a long time
> ago and render nodes finally seemed to be a way to fix that.
> I mean why are compositors using the primary node in the first place?
> Because they want to have access to privileged resources I think and in
> this case it is perfectly ok to do so.
> Now extending unprivileged access to the primary node actually sounds
> like a step into the wrong direction to me.
> I rather think that we should go down the route of completely dropping
> command submission and buffer allocation through the primary node for
> non master clients. And then as next step at some point drop support for
> authentication/flink.
> I mean we have done this with UMS as well and I don't see much other way
> to move forward and get rid of those ancient interface in the long term.

Well kms had some really good benefits that drove quick adoption, like
"suspend/resume actually has a chance of working" or "comes with
buffer management so you can run multiple gears".

The render node thing is a lot more niche use case (prime, better priv
separation), plus "it's cleaner design". And the "cleaner design" part
is something that empirically doesn't seem to matter :-/ Just two
- KHR_display/leases just iterated display resources on the fd needed
for rendering (and iirc there was even a patch to expose that for
render nodes too so it works with DRI3), because implementing
protocols is too hard. Barely managed to stop that one before it
- Various video players use the vblank ioctl on directly to schedule
frames, without telling the compositor. I discovered that when I
wanted to limite the vblank ioctl to master clients only. Again,
apparently too hard to use the existing extensions, or fix the bugs in
there, or whatever. One userspace got fixed last year, but it'll
probably get copypasted around forever :-/

So I don't think we'll ever manage to roll a clean split out, and best
we can do is give in and just hand userspace what it wants. As much as
that's misguided and unclean and all that. Maybe it'll result in a
least fewer stuff getting run as root to hack around this, because
fixing properly seems not to be on the table.

The beauty of kms is that we've achieved the mission, everyone's
writing their own thing. Which is also terrible, and I don't think
it'll get better.
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

More information about the amd-gfx mailing list