[Nouveau] Regression in 5.15 in nouveau

[Daniel Vetter]

On Tue, Dec 07, 2021 at 06:32:06PM +0100, Karol Herbst wrote:
> On Tue, Dec 7, 2021 at 10:52 AM Christian König
> <christian.koenig at amd.com> wrote:
> >
> > Am 06.12.21 um 19:37 schrieb Dan Moulding:
> > > On 04.12.21 17:40, Stefan Fritsch wrote:
> > >> Hi,
> > >>
> > >> when updating from 5.14 to 5.15 on a system with NVIDIA GP108 [GeForce
> > >> GT 1030] (NV138) and Ryzen 9 3900XT using kde/plasma on X (not wayland),
> > >> there is a regression: There is now some annoying black flickering in
> > >> some applications, for example thunderbird, firefox, or mpv. It mostly
> > >> happens when scrolling or when playing video. Only the window of the
> > >> application flickers, not the whole screen. But the flickering is not
> > >> limited to the scrolled area: for example in firefox the url and
> > >> bookmark bars flicker, too, not only the web site. I have bisected the
> > >> issue to this commit:
> > >>
> > >> commit 3e1ad79bf66165bdb2baca3989f9227939241f11 (HEAD)
> > > I have been experiencing this same issue since switching to 5.15. I
> > > can confirm that reverting the above mentioned commit fixes the issue
> > > for me. I'm on GP104 hardware (GeForce GTX 1070), also running KDE
> > > Plasma on X.
> >
> > I'm still scratching my head what's going wrong here.
> >
> > Either we trigger some performance problem because we now wait twice for
> > submissions or nouveau is doing something very nasty and not syncing
> > it's memory accesses correctly.
> >
> > Attached is an only compile tested patch which might mitigate the first
> > problem.
> >
> > But if it's the second then nouveau has a really nasty design issue here
> > and somebody with more background on that driver design needs to take a
> > look.
> >
> 
> Ben mentioned a few times that fences might be busted but we all have
> no idea what's actually wrong. So it might be that your change is
> indeed triggering something which was always broken or something else.

Description sounds a bit like we're doing a clear before Xorg has had a
chance to copy the pixmap to the frontbuffer perhaps? That would point to
a fencing issue in userspace, and somehow ignoring fences ensures that the
Xorg copy/blt completes before we get around to clearing stuff.

I'm assuming we're rendering with glamour, so is nouveau relying on kernel
implicit sync or doing it's own fencing in userspace?
-Daniel

> 
> > Please test if that patch changes anything.
> >
> > Thanks,
> > Christian.
> >
> > >
> > > Cheers,
> > >
> > > -- Dan
> >
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch