[Intel-gfx] [PATCH v2 00/22] Enable gpu switching on the MacBook Pro

[Daniel Vetter]

On Tue, Aug 11, 2015 at 12:29:17PM +0200, Lukas Wunner wrote:
> This is a follow-up to the v1 posted in April:
> http://lists.freedesktop.org/archives/dri-devel/2015-April/081515.html
> 
> 
> Patches 1 - 17 enable GPU switching on the pre-retina MacBook Pro.
> These were tested successfully by multiple people and solve two
> tickets in Bugzilla:
> https://bugzilla.kernel.org/show_bug.cgi?id=88861
> https://bugs.freedesktop.org/show_bug.cgi?id=61115
> 
> Patches 18 - 22 are a preview of how we're tackling retina support.
> Those are marked experimental and are NOT ready to be merged yet.
> Feedback on them is welcome.
> 
> The patches are based on drm-next.
> 
> They were tested on the following hardware (thanks a lot everyone!):
> Tested-by: Paul Hordiienko <pvt.gord at gmail.com>
>     [MBP  6,2 2010  intel ILK + nvidia GT216  pre-retina]
> Tested-by: William Brown <william at blackhats.net.au>
>     [MBP  8,2 2011  intel SNB + amd turks     pre-retina]
> Tested-by: Lukas Wunner <lukas at wunner.de>
>     [MBP  9,1 2012  intel IVB + nvidia GK107  pre-retina]
> Tested-by: Bruno Bierbaumer <bruno at bierbaumer.net>
>     [MBP 11,3 2013  intel HSW + nvidia GK107  retina -- work in progress]
> 
> 
> What's new:
> 
> * By default the MBP boots with the display switched to the discrete GPU
>   but it can be forced to the integrated GPU with an EFI boot variable.
>   Here's a handy tool for that: https://github.com/0xbb/gpu-switch
>   v1 didn't work in this configuration, v2 does.
> 
> * Reprobing if the inactive GPU initializes before the apple-gmux module:
>   v1 used Matthew Garrett's approach of adding a driver callback.
>   v2 simply generates a hotplug event instead. nouveau polls its outputs
>   every 10 seconds so we want it to poll immediately once apple-gmux
>   registers. That is achieved by the hotplug event. The i915 driver is
>   changed to behave identically to nouveau. (Right now it deletes LVDS
>   and eDP connectors from the mode configuration if they can't be probed,
>   deeming them to be ghosts.)

I thought -EDEFERREDPROBE is what we should be using if drivers don't get
loaded in the right order? Hand-rolling depency avoidance stuff is imo a
horrible idea.

> * Framebuffer recreation if the inactive GPU initializes before the
>   apple-gmux module (i.e. discarding the default 1024x768 fb and replacing
>   with a new one with the actual panel resolution): v1 only supported this
>   for i915, v2 has a generic solution which works with nouveau and radeon
>   as well. (Necessary if the discrete GPU is forced to be the inactive one
>   on boot via the EFI variable.)

Would completely remove the need for this ;-)

> * Generally lots of rough edges were smoothed to hopefully make the
>   patches more suitable for merging. E.g. there's a bug in i915 where
>   the SSC status set by BIOS is preserved too late and v1 only contained
>   a workaround, whereas v2 contains a proper fix in a separate commit.
> 
> 
> The long journey towards retina support:
> 
> The pixel clock required for retina resolution is not supported by LVDS
> (which was used on pre-retinas), necessitating eDP instead. Problem is,
> the gmux register which switches the DDC lines on pre-retinas doesn't
> switch the AUX channel on retinas. Disassembling the OS X driver revealed
> that the gmux in retina MBPs gained an additional register 0x7f which gets
> written to when setting up the eDP configuration. There was some hope that
> this might switch the AUX channel. Alas, we tried writing various values
> to that register but were unable to get the inactive GPU to talk to the
> panel. The purpose of register 0x7f remains a mystery.
> 
> Teardowns of the first generation retina MBP name the NXP CBTL06142 and
> TI HD3SS212 as multiplexers and according to the data sheets I've found,
> neither supports switching the AUX channel separately from the main link.
> 
> Matthew Garrett had the idea of having the active GPU stash the EDID and
> the first 8 bytes of the DPCD (receiver capabilities) and letting the
> inactive GPU retrieve that data. I rebased and rewrote his patches and
> got everything working, only to discover that the drivers are unhappy
> with just 8 bytes of DPCD. They need full access to the DPCD to set up
> their outputs. We could stash the entire DPCD but some parts of it are
> mutable so the stashed data may become stale when the active GPU performs
> writes to the DPCD.
> 
> So I had the idea of using the active GPU as a proxy to talk to the panel,
> thus emulating switching of the AUX channel in software. We can leverage
> the struct drm_dp_aux and i2c_adapter (for DDC) to achieve this, swapping
> the inactive GPU's structs with those of the active GPU on the fly.
> That approach is implemented in patches 18 - 22 but there are still some
> driver issues that I'm debugging. The current status as per the latest
> logs Bruno sent me is that i915 rejects the mode retrieved via proxying
> with CLOCK_HIGH and nouveau aborts link training halfway through.
> Bottom line is that it's not yet working but we're getting closer.
> 
> As a side effect, the pre-retinas gain a second way to initialize their
> outputs: They can either use gmux to switch the DDC lines, or use the
> active GPU as a proxy for the DDC communication. Which method gets used
> depends on the order in which the drivers initialize, the inactive GPU
> will happily use whatever is available and it automatically receives
> a hotplug event once either method becomes ready for use.
> 
> But, once again, the patches implementing proxying (patches 18 - 22)
> are still in a state of flux and not ready for prime time, unlike the
> prior ones which seem stable. Folks are hereby invited to poke holes
> into them and I'm looking forward to your feedback.

You can't share the dp aux like that. It's true that we need a bit more
data (there's a few eDP related feature blocsk we need), but sharing the
aux channel entirely is no-go. If you do you get drivers trying to link
train and at best this fails and at worst you'll upset the configuration
of the other driver and piss of the panel enough to need a hard reset
until it works again.

I think just reading edid and the relevant dp aux data in apple-gmux or
somewhere like that and stalling driver load until that's ready is the
only clean option. And of course we must ensure that inactive drivers
don't try to use the epd link at all since that will piss off the panel.

I think the real tricky bit here with vgaswitcheroo is locking, I need to
take a separate lock at the patches for that.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch