AMD/AMD hybrid graphics

[Dave Airlie]

On Fri, Feb 21, 2014 at 12:25 PM, Michel Dänzer <michel at daenzer.net> wrote:
> On Don, 2014-02-20 at 08:44 +0100, Boszormenyi Zoltan wrote:
>> 2014-02-20 06:47 keltezéssel, Michel Dänzer írta:
>> > On Don, 2014-02-20 at 06:09 +0100, Boszormenyi Zoltan wrote:
>> >> 2014-02-20 04:20 keltezéssel, Michel Dänzer írta:
>> >>> On Mit, 2014-02-19 at 11:56 +0100, Boszormenyi Zoltan wrote:
>> >>>> 2014-02-19 10:59 keltezéssel, Michel Dänzer írta:
>> >>>>> On Mit, 2014-02-19 at 09:11 +0100, Boszormenyi Zoltan wrote:
>> >>>>>
>> >>>>>> Can Mesa/Xorg use both r600g and radeonsi at the same time?
>> >>>>> Yes, that seems to work fine for others. You may need Mesa 10.1 or newer
>> >>>>> though.
>> >>>> Do you mean mean with Mesa 9.2.5 and Xorg server 1.14.4 in
>> >>>> Fedora 20 at this time, it's not possible unless I compile my own
>> >>>> llvm-3.5 SVN, Mesa 10.1 or 10.2 GIT and Xorg 1.15 GIT?
>> >>> I don't think Xorg 1.15 is necessary, but it shouldn't hurt either.
>> >>>
>> >>>
>> >>>> Attached is the log from both 1.14.4 (FC20) and 1.15.0 (rawhide), [...]
>> >>> The log files end abruptly, so we need to see the X server stderr
>> >>> output. Assuming you're using gdm, it should be captured
>> >>> in /var/log/gdm*/:0.log .
>> >> FC20 has lightdm by default, here's /var/log/lightdm/x-0.log
>> > [...]
>> >
>> >> X: ../../../include/privates.h:122: dixGetPrivateAddr: Assertion
>> >> `key->initialized' failed.
>> > Can you get a backtrace for this assertion failure with gdb? See
>> > http://wiki.x.org/wiki/Development/Documentation/ServerDebugging/
>
> [...]
>
>> #0  0x00007fadc165f1c9 in __GI_raise (sig=sig at entry=6) at
>> ../nptl/sysdeps/unix/sysv/linux/raise.c:56
>> #1  0x00007fadc16608d8 in __GI_abort () at abort.c:89
>> #2  0x00007fadc1658126 in __assert_fail_base (fmt=0x7fadc17a98a0 "%s%s%s:%u: %s%sAssertion
>> `%s' failed.\n%n", assertion=assertion at entry=0x5a4c53 "key->initialized",
>>      file=file at entry=0x5a8dfc "../../../include/privates.h", line=line at entry=122,
>> function=function at entry=0x5aaa90 <__PRETTY_FUNCTION__.8544> "dixGetPrivateAddr") at
>> assert.c:92
>> #3  0x00007fadc16581d2 in __GI___assert_fail (assertion=assertion at entry=0x5a4c53
>> "key->initialized", file=file at entry=0x5a8dfc "../../../include/privates.h",
>> line=line at entry=122,
>>      function=function at entry=0x5aaa90 <__PRETTY_FUNCTION__.8544> "dixGetPrivateAddr") at
>> assert.c:101
>> #4  0x0000000000424f70 in dixGetPrivateAddr (key=<optimized out>, key=<optimized out>,
>> privates=0x169d558) at ../../../include/privates.h:122
>> #5  0x00000000004810eb in dixGetPrivateAddr (key=<optimized out>, key=<optimized out>,
>> privates=0x169d558) at xf86cmap.c:239
>> #6  dixSetPrivate (val=<optimized out>, key=0x822f80 <CMapScreenKeyRec>,
>> privates=0x169d558) at ../../../include/privates.h:148
>> #7  xf86HandleColormaps (pScreen=pScreen at entry=0x169d180, maxColors=maxColors at entry=256,
>> sigRGBbits=10, loadPalette=loadPalette at entry=0x7fadbcd37ea0 <drmmode_load_palette>,
>>      setOverscan=setOverscan at entry=0x0, flags=flags at entry=3) at xf86cmap.c:184
>> #8  0x00007fadbcd3b32c in drmmode_setup_colormap (pScreen=pScreen at entry=0x169d180,
>> pScrn=pScrn at entry=0x169c930) at drmmode_display.c:1990
>> #9  0x00007fadbcd370ef in RADEONScreenInit_KMS (pScreen=pScreen at entry=0x169d180,
>> argc=argc at entry=1, argv=argv at entry=0x7fff83c684c8) at radeon_kms.c:1366
>> #10 0x0000000000438b4d in AddGPUScreen (pfnInit=0x7fadbcd36c60 <RADEONScreenInit_KMS>,
>> argc=argc at entry=1, argv=argv at entry=0x7fff83c684c8) at dispatch.c:3874
>> #11 0x000000000047aa9f in InitOutput (pScreenInfo=pScreenInfo at entry=0x82dd60 <screenInfo>,
>> argc=argc at entry=1, argv=argv at entry=0x7fff83c684c8) at xf86Init.c:886
>
> Hmm, looks like there's confusion about how colormaps are supposed to be
> handled.
>
>
> Dave, any ideas?

xf86_crtc_supports_gamma probably stops us from ever getting into that
function, but in the case where we have 0 crtcs I could see fallout.

Either fix the DDX to not call colormap handling if we have 0 crtcs
and/or fix the server to not install bother with colormaps if we have
0 crtcs.

Dave.