[waffle] [Piglit] [ANNOUNCE] waffle-1.2.2

[Chad Versace]

On 11/28/2012 12:22 PM, Marek Olšák wrote:
> On Wed, Nov 28, 2012 at 8:42 PM, Chad Versace
> <chad.versace at linux.intel.com> wrote:
>> On 11/28/2012 10:40 AM, Marek Olšák wrote:
>>> On Wed, Nov 28, 2012 at 7:26 PM, Chad Versace
>>> <chad.versace at linux.intel.com> wrote:
>>>> On 11/27/2012 07:47 PM, Marek Olšák wrote:
>>>>> HI Chad,
>>>>>
>>>>> FYI, the core GL support in the GLX backend of waffle and/or piglit
>>>>> seems to be broken. All piglit tests that specify the core GL version
>>>>> and are run on a driver which cannot create a core GL context (that is
>>>>> any open driver except i965), always fail. The tests report
>>>>> GLXBadFBConfig and just die. I think the piglit_init function isn't
>>>>> even reached.
>>>>>
>>>>> I don't use EGL.
>>>>>
>>>>> Marek
>>>>
>>>> Marek,
>>>>
>>>> I tried some experiments with i965 and older versions of Mesa in an
>>>> attempt to reproduce your problem, but I can't. In the two experiments
>>>> I tried below, the behavior was exactly as expected.
>>>>
>>>> * Test 1: Against i965 from mesa-7.11, which supports only GL 2.1 and no
>>>> GLX_ARB_create_context.
>>>>
>>>>> PIGLIT_DEBUG=1 PIGLIT_PLATFORM=glx ./gl-3.1-minmax
>>>> piglit: debug: waffle_config_choose failed due to WAFFLE_ERROR_UNSUPPORTED_ON_PLATFORM: GLX_ARB_create_context is required in order to request a GL version not equal to the default value 1.0
>>>> piglit: info: Failed to create GL 3.1 core context
>>>> Mesa: Initializing x86-64 optimizations
>>>> Test requires GL version 3.1
>>>> PIGLIT: {'result': 'skip' }
>>>>
>>>>
>>>> * Test 2: Against i965 from mesa-9.0.1, which supports GL 3.1 core, and an Xserver
>>>> without GLX_ARB_create_context.
>>>>
>>>>> PIGLIT_DEBUG=1 PIGLIT_PLATFORM=glx ./gl-3.1-minmax
>>>> piglit: debug: waffle_config_choose failed due to WAFFLE_ERROR_UNSUPPORTED_ON_PLATFORM: GLX_ARB_create_context is required in order to request a GL version not equal to the default value 1.0
>>>> piglit: info: Failed to create GL 3.1 core context
>>>> Test requires GL version 3.1
>>>> PIGLIT: {'result': 'skip' }
>>>>
>>>>
>>>> Please run, with PIGLIT_DEBUG=1, the piglit test that's causing you trouble
>>>> and send me the output. Hopefully that will be enough for me to identify
>>>> and fix the bug.
>>>
>>> PIGLIT_DEBUG doesn't print anything. All I get is:
>>>
>>> X Error of failed request:  GLXBadFBConfig
>>>   Major opcode of failed request:  154 (GLX)
>>>   Minor opcode of failed request:  34 ()
>>>   Serial number of failed request:  22
>>>   Current serial number in output stream:  20
>>>
>>> I get the same error with gallium and classic swrast.
>>>
>>> Marek
>>
>> Even with swrast, I'm not getting GLXBadFBConfig.
>>
>>> PIGLIT_DEBUG=1 PIGLIT_PLATFORM=glx LIBGL_ALWAYS_SOFTWARE=1 bin/gl-3.1-minmax
>> piglit: debug: waffle_config_choose failed due to WAFFLE_ERROR_UNSUPPORTED_ON_PLATFORM: GLX_ARB_create_context is required in order to request a GL version not equal to the default value 1.0
>> piglit: info: Failed to create GL 3.1 core context
>> Test requires GL version 3.1
>> PIGLIT: {'result': 'skip' }
>>
>>
>> The GLX call that likely emits GLXBadFBConfig is the call to glXCreateNewContext
>> at glx_context.c:144 (lines relative to waffle-1.2.2). That call succeeds for me,
>> so I don't know why it's failing for you. The following information would help:
>>
>> 1. Break at glx_context.c:144.
>> 2. Verify that the call to glXCreateNewContext there emits GLXBadFBConfig.
>> 3. Give me the backtrace there.
>> 4. In gdb, give me the output of `print *config`.
>> 5. Give me the output of glxinfo.
> 
> After some investigation, it's actually glXCreateContextAttribsARB
> that fails, which is what I would expect to fail if it can't create a
> core context. The implementation is in mesa/src/glx/create_context.c.
> I'm cc'in Ian, because he implemented it.
> 
> glXCreateContextAttribsARB falls back to indirect rendering if it
> fails to create a direct context and then
> xcb_glx_create_context_attribs_arb_checked returns GLXBadFBConfig.
> 
> I have 2 options there:
> - don't try indirect rendering and return NULL
> - don't call __glXSendErrorForXcb (which kills the app) and return NULL
> 
> Both options yield the behavior waffle expects. My guess is that we
> should remove the call to __glXSendErrorForXcb, but I'm not sure if
> it's a correct fix.
> 
> BTW I have X Server 1.13.
> 
> Marek

I third option is to add a new waffle function that allows clients
to register various error handlers.

When writing waffle, I completely
forgot that the default handler for X errors simply exits. I was working
from Wayland's model, where unhandled errors are ignored. Another lesson
learned for waffle 2.0.