Switchable graphics and radeon PX runtime

[Takashi Iwai]

At Fri, 14 Nov 2014 12:29:17 -0500,
Alex Deucher wrote:
> 
> On Fri, Nov 14, 2014 at 11:14 AM, Takashi Iwai <tiwai at suse.de> wrote:
> > At Fri, 14 Nov 2014 10:40:08 -0500,
> > Alex Deucher wrote:
> >>
> >> On Fri, Nov 14, 2014 at 5:09 AM, Takashi Iwai <tiwai at suse.de> wrote:
> >> > At Fri, 14 Nov 2014 19:33:00 +1000,
> >> > Dave Airlie wrote:
> >> >>
> >> >> On 14 November 2014 18:12, Takashi Iwai <tiwai at suse.de> wrote:
> >> >> > Hi Alex,
> >> >> >
> >> >> > we've got a few bug reports about the behavior of radeon driver on
> >> >> > machines with Intel+AMD "switchable graphics" (no Muxless).  So far,
> >> >> > it seems that the sane only way to make the machine working is to get
> >> >> > back to the old vgaswitcheroo behavior via radeon.runpm=0.  Without
> >> >> > it, radeon GPU gives a spurious output as connected, eventually
> >> >> > crashes GNOME.  (Also, from the nature of the switchable graphics,
> >> >> > vgaswitcheroo looks more intuitive to me.)
> >> >>
> >> >> vgaswitcheroo only matters if there is a MUX, the point of it is to drive
> >> >> the MUX.
> >> >>
> >> >> dynamic poweroff makes more sense, switcheroo on/off switch was
> >> >> just a hack.
> >> >
> >> > Well, I find the current form fairly unintuitive, at least, for the
> >> > switchable (not optimus) graphics.
> >> > With dynamic PM, the card is activated on demand.  So you may enable
> >> > outputs of both cards at any time, right?
> >> >
> >> > Currently, all outputs from both cards are exposed in Xrandr,
> >> > e.g. LVDS1 DP1, HDMI1, VGA1, LVDS-1-1, HDMI-1-2, DisplayPort-1-2, and
> >> > VGA-1-1.  How can user-space know which one should be activated and
> >> > which not, when you can use effectively only a single card?
> >> >
> >> >> > How are such machines supposed to work with the recent system?  Is PX
> >> >> > wrongly detected on them, or something else missing?
> >> >>
> >> >> It sounds like the connector is wrongly detected and that should be what
> >> >> is fixed.
> >> >
> >> > Yeah, that's a problem indeed.  In the bug report, both LVDS1 and
> >> > LVDS-1-1 are reported to be connected at the same time while the
> >> > latter doesn't get any real size and position.  We didn't trace
> >> > whether this is the culprit of crash of GNOME, but at least, it looks
> >> > fairly weird.
> >> >
> >> > I forgot to give the original bug report:
> >> >   http://bugzilla.opensuse.org/show_bug.cgi?id=904417
> >> >
> >> > and the xrandr output is found at
> >> >   http://bugzilla.opensuse.org/show_bug.cgi?id=904417#c18
> >>
> >> It's not clear to me from the bug report what that problem is.  What
> >> exactly is gnome complaining about?
> >
> > It simply crashes by some reason, showing a sad face and complaining
> > something is wrong.  (And it's GNOME, not easy to get a proper log
> > like kernel :)
> > Some users complained about blank output, but I'm not sure whether
> > this is the same cause.
> >
> >> The X logs look fine.
> >
> > I couldn't see any errors there, too.  So it's just a wild guess, so
> > far...
> >
> >> From the
> >> xrandr output, LVDS1 (connected to the intel) is connected and active.
> >> LVDS-1-1 (connected to the radeon) is connected but not active.  That
> >> should be a perfectly reasonable configuration.  If LVDS-1-1 is not
> >> active, the radeon kernel driver will power down the GPU until the
> >> user either activates the panel or uses the dGPU as an offscreen
> >> renderer.
> >
> > Note that the xrandr output was taken on icewm or else.  So, right,
> > this might be non-issue.  But, I guess now that the issue will happen
> > when LVDS-1-1 is activated at the same time with LVDS.  GNOME tries to
> > activate all connected outputs at the same time as default.
> >
> >> There are plenty of cases when you may have a secondary GPU with
> >> attached displays that are not active. If gnome barfs on this it
> >> should be fixed in gnome.
> >
> > Yeah, GNOME has definitely a problem about it.  At least, it shouldn't
> > crash badly.
> >
> > But, it's still not clear to me how the activation of the radeon GPU
> > is supposed to work in switchable graphics case.  For PX or Otpimus,
> > it's clear.  But for switchable case, there is no offload rendering.
> > If you enable the output on radeon GPU while Intel output is being
> > used, what's going on?  Shouldn't they be handled exclusively, as user
> > expected?
> 
> You can still do offload rendering with a switchable system, it just
> happens to also have a mux which you can use or not use.  That's how
> it works today.
> 
> If you enable the outputs on both gpus, they will both power up the
> display hw on the GPU, but only the signals from the gpu selected by
> the mux will actually make it to the display.

Yes, that's the problem.  The desktop expects that the connected
output can be shown properly at the desired place and size. (GNOME
will try to place all connected outputs from left to right as default,
IIRC.) 
But, in this case, it's not informed that some connected outputs are
actually not visible.

> Does the attached hack help?  It basically attempts to determine
> whether the mux is switched to the radeon or not and report
> disconnected or not for the local displays.  To handle handle it
> properly, vgaswitcheroo would need to track the signal and ddc/hpd
> muxes and propagate them to the drivers.

Thanks!  I'll provide a test KMP including this patch to reporters to
give a try.

During debugging, I came up to a similar hack, too, but then changed
my mind later that the old good vgaswitcheroo might be more
straightforward (and indeed it "works" more or less). 

> >> Setting radeon.runpm=0 does not disable PX, it just stops the kernel
> >> driver from dynamically turning the dGPU on and off on demand.
> >
> > Yes.  The user seem more happy with explicit power control.  At least,
> > better than this kind of surprises...
> 
> It should behave exactly the same if the user starts X without
> explicitly powering down the dGPU.  The pci entry just happens to
> disappear without runpm so X doesn't try and load the radeon ddx.  The
> kernel driver is still loaded however.  So it's not really a fix per
> se.

That's true.  My intention was to avoid unexpected regressions.  It's
an old hardware, and users already are used to the old way.

If everything works, no reason to go back to old hacks :)


Takashi