[Intel-gfx] VGA arbiter support for Intel HD?

[Ville Syrjälä]

On Wed, Aug 14, 2013 at 01:39:55PM -0600, Alex Williamson wrote:
> On Wed, 2013-08-14 at 21:30 +0300, Ville Syrjälä wrote:
> > On Wed, Aug 14, 2013 at 11:14:48AM -0600, Alex Williamson wrote:
> > > On Wed, 2013-08-14 at 17:47 +0300, Ville Syrjälä wrote:
> > > > On Wed, Aug 14, 2013 at 07:23:57AM -0600, Alex Williamson wrote:
> > > > > Hi,
> > > > > 
> > > > > I'm trying to add support for device assignment of PCI VGA devices with
> > > > > VFIO and QEMU.  For normal, discrete discrete graphics the Linux VGA
> > > > > arbiter works fairly well, disabling VGA on one bridge and adding it to
> > > > > another (though I wish all the kernel VGA drivers made use of it).  The
> > > > > i915 driver only seems to support disabling VGA on really old GMCH
> > > > > devices (see intel_modeset_vga_set_state).  This means that if I boot
> > > > > with IGD as the primary graphics and attempt to assign a discrete
> > > > > graphics device, all the VGA range accesses are still routed to IGD, I
> > > > > end up getting some error messages from the IGD interrupt handler, and
> > > > > the discrete card never initializes.
> > > > > 
> > > > > I spent some time looking through the Sand Bridge, Ivy Bridge, and
> > > > > Haswell datasheets, and I'm a bit concerned whether the hardware even
> > > > > provides a reasonable way to disable VGA anymore.  Quoting 2.17 from the
> > > > > Haswell docs:
> > > > > 
> > > > >         Accesses to the VGA memory range are directed to IGD depend on
> > > > >         the configuration.  The configuration is specified by:
> > > > >               * Internal graphics controller in Device 2 is enabled
> > > > >                 (DEVEN.D2EN bit 4)
> > > > >               * Internal graphics VGA in Device 0 Function 0 is enabled
> > > > >                 through register GGC bit 1.
> > > > >               * IGD's memory accesses (PCICMD2 04 – 05h, MAE bit 1) in
> > > > >                 Device 2 configuration space are enabled.
> > > > >               * VGA compatibility memory accesses (VGA Miscellaneous
> > > > >                 Output register – MSR Register, bit 1) are enabled.
> > > > >               * Software sets the proper value for VGA Memory Map Mode
> > > > >                 register (VGA GR06 Register, bits 3-2). See the
> > > > >                 following table for translations.
> > > > > 
> > > > > (There's a similar list for VGA I/O range)  I've found that if I disable
> > > > > memory and I/O in the PCI command register for IGD then I do get VGA
> > > > > routing to the PEG device and the discrete VBIOS works.  This obviously
> > > > > isn't a good option for the VGA arbiter since it entirely disables IGD.
> > > > > 
> > > > > The GGC registers aren't meant for runtime switching and are actually
> > > > > locked.  Disabling IGD via the device 2 enable bit doesn't seem like and
> > > > > option.  I don't quite understand the VGA miscellaneous output register
> > > > > and VGA memory map mode, but the table provided for the latter makes me
> > > > > think they just augment the VGA ranges and don't disable them.
> > > > 
> > > > Bit 1 of MSR (0x3c2/0x3cc) should allow you to turn off VGA mem
> > > > access while leaving other memory space access working.
> > > > 
> > > > As for VGA I/O decode, IIRC there's no standard bit for that in VGA
> > > > or PCI config registers, and I can't see any other  bit for it in the
> > > > docs. But I guess you could just turn off I/O space completely
> > > > via the PCI_COMMAND register. We shouldn't need it for anything beyond
> > > > i915_disable_vga() and that has the necessary vgaarb calls already.
> > > 
> > > Thanks Ville.  The MSR seems to work for VGA memory.  Disabling I/O via
> > > PCI_COMMAND does works, but something is re-enabling it after
> > > intel_modeset_vga_set_state().  If I manually disable I/O with setpci
> > > then I do have VGA routing to PEG and can still interact with the KMS
> > > console on IGD.  It's unfortunate that the MSR bit for I/O only disables
> > > pieces of the range.  If we have no other options, I'll try to hunt down
> > > where I/O is being re-enabled and see how feasible it is to avoid.
> > > Thanks,
> > 
> > Hmm. Now that I look at vgaarb more it seems I misunderstood the way it
> > works. Based on the code it looks like it will permanently remove the
> > device from the arbiration if set_vga_decode indicates that it doesn't
> > decode legacy resources. And it calls set_vga_decode w/ decode=false
> > if there are more than two VGA cards in the system. That means
> 
> s/more than two/two or more/
> 
> > i915_disable_vga() is actually broken whenever there is another
> > VGA card in the system.
> 
> I didn't follow why i915_disable_vga is broken.  It seems like the
> intention is to disable VGA regardless of how many VGA devices are
> present.

I suppose that's the original intention, but i915_disable_vga() needs
VGA I/O access, so it is currently broken if VGA I/O is disabled.
Originally i915_disable_vga() used MMIO access, but there's a hardware
workaround (since ILK IIRC) stating that VGA registers must be
accessed w/ port I/O. That's progress for you :(

-- 
Ville Syrjälä
Intel OTC