Re: [Feedback] hybrid GPU + multimonitor: partial success with Gnome-shell / Mutter 3.27.1 in Wayland mode

[Jonas Ådahl <jadahl gmail com>]

On Tue, Feb 06, 2018 at 05:06:02PM +0100, Cyrille Chépélov wrote:
> Hi Jonas,
>
> responding slightly out of order (both below an inline); plus we've pretty
> much already started to bring this onto the issue tracker where this
> belongs:

Thanks for filing those issues.

> > It'd also be good if you used the top of the master branch instead of
> > 3.27.1 (both mutter and gnome-shell), as otherwise you might hit a set
> > of bugs fixed during the development cycle.
>
> Now rebuilding to work top-of-master for Mesa, mutter & gnome-shell:
>
>  * mesa: master at a5053ba27e
>  * mutter: master at 70fcf745
>  * gnome-shell: master at 0b51ead00
>  * gnome-shell-extensions: master at ae65a82 /(not strictly necessary
>    but to keep dpkg happy)/

... snip ...

> > >   * When /un/pluggin an external display through HDMI, the machine keeps
> > >     working, and could withstand an unlimited number of plug/unplug cycles.
> > This is also indeed expected to work.
> >
> > Note that when I did this work, I hit a nouveau bug that triggered
> > freezes. I have not checked whether those have been fixed.
>
> Noted. As an aside, I noticed that when shutting down gnome-session + the
> gdm3 service (going back to text mode), the "nouveau" kernel module remains
> "used" (it increases by 1 each time gdm starts), which makes difficult
> removing it for the purpose of powering down the GPU. Would eventually
> report it on 
> https://bugs.freedesktop.org/buglist.cgi?component=Drivers%2FDRI%2Fnouveau&list_id=636975&product=Mesa&resolution=---

Should probably double check that we close nouveau KMS fd when
terminating, but either way, the driver should be able to handle a
process with a KMS fd open crashing either way.

>

... snip ...

> > FWIW, when rendering the content for your nouveau connected monitor, we
> > do so using the Intel GPU. After eglSwapBuffers(), we take the result,
> > blit it onto a buffer on the nvidia GPU using nouveau, and hand that
> > over to the monitor.
> Ah, I wasn't aware of this. So basically for now 'hybrid mode' is
> effectively using the iGPU for most tasks, while the dGPU is used as
> effectively a PCI-to-HDMI bitmap-shipping dongle.
>
> So perhaps there may be an interesting alternative to GPU hotplug, if there
> is a way to start nouveau (and the GPU) permanently but shut down most of
> the engines (https://nouveau.freedesktop.org/wiki/KernelModuleParameters/ #
> grep for "list of engines). Paying a /few /watts at idle to gain easy enough
> monitor plug&play might be a good enough option for a while (I'll
> investigate separately)

FWIW, we still use OpenGL etc for the bitmap-shipping as long as the
driver for the dedicated GPU supports the GLES 3 features we depend on.

>

... snip ...

> > >      3. I had once a crash in native code called by javascript code, but
> > >         lost the backtrace before I could save it
> > If it was a segmentation fault, you could set the SHELL_DEBUG env
> > variable to "backtrace-segfaults" (in a way that gnome-shell will see it
> > when starting). This will make gnome-shell to print Javascript
> > backtraces to the journal when it hits a segmentation fault.
> It was a segfault indeed. Added SHELL_DEBUG=backtrace-segfaults for next
> time
>
> >   * I once managed to move a Firefox (Quantum) window with a moving
> >     YouTube video from the main (i915) to the external (nouveau) screen,
> >     without an obvious causality on the crash that happens seconds later.
> > FWIW, this wouldn't cause Firefox to switch to using the nvidia GPU for
> > rendering, as we don't have a way to tell applications to change what to
> > render with yet.
> Makes sense. Again, the iGPU is plenty adequate for my needs, I really wish
> GTX 1060 didn't omit the hardware muxer (or Gigabyte provided a no-nonsense
> option without unnecessary hardware).

You can still use the dGPU for rendering application content. IIRC using
"DRI_PRIME=1 some-egl-application" worked as expected. There are
optimizations to be done (e.g. skip going via the iGPU driver when
possible) but should at least be possible to use it.


Jonas