[Mesa-dev] [PATCH 0/4] egl/*: Set disp->DriverData to NULL on error

[Martin Peres]

On 04/08/16 15:17, Nicolas Boichat wrote:
> On Thu, Aug 4, 2016 at 7:17 PM, Martin Peres
> <martin.peres at linux.intel.com> wrote:
>>
>>
>> On 04/08/16 13:08, Nicolas Boichat wrote:
>>>
>>> On Thu, Aug 4, 2016 at 5:52 PM, Martin Peres
>>> <martin.peres at linux.intel.com> wrote:
>>>>
>>>> On 04/08/16 05:07, Nicolas Boichat wrote:
>>>>>
>>>>>
>>>>> egl/x11 fixup has already be sent here:
>>>>> https://patchwork.freedesktop.org/patch/101934/
>>>>>
>>>>> Compile-tested only (on android only).
>>>>>
>>>>> I was first thinking of modifying most functions in egl_dri2.c to take
>>>>> in
>>>>> dri2_egl_display *dri2_dpy instead of _EGLDisplay *disp as parameter, so
>>>>> we do not need to set disp->DriverData early in the dri_initialize_*
>>>>> functions.
>>>>>
>>>>> But there are some functions that do need disp as well:
>>>>>  - dri2_add_config, to call _eglInitConfig (disp->DriverData does not
>>>>> need
>>>>> to be
>>>>>    set), and to check if extensions are supported.
>>>>>  - dri2_setup_screen (similar reasons)
>>>>>
>>>>> Ideas to untangle this are welcome. In the mean time, this patch series
>>>>> should
>>>>> fix crashes and a piglit test.
>>>>>
>>>>> Nicolas Boichat (4):
>>>>>   egl/wayland: Set disp->DriverData to NULL on error
>>>>>   egl/surfaceless: Set disp->DriverData to NULL on error
>>>>>   egl/drm: Set disp->DriverData to NULL on error
>>>>>   egl/android: Set dpy->DriverData to NULL on error
>>>>>
>>>>
>>>> This patch series fixes egl_khr_fence_sync for me:
>>>> Tested-by: Martin Peres <martin.peres at linux.intel.com>
>>>
>>>
>>> Thanks!
>>>
>>>> However, egl-create-context-verify-gl-flavor is still crashing, even with
>>>> "egl/dri2: dri2_initialize: Do not reference-count TestOnly display"
>>>> applied. The issue is that eglGetProcAddress("glGetString") returns NULL.
>>>
>>>
>>> That's surprising, Michel said that my patch fixed that issue. Did you
>>> try reverting "9ee683f877 (egl/dri2: Add reference count for
>>> dri2_egl_display)" to confirm that my original patch broke it?
>>
>>
>> Yes, reverting the patch fixes the issue.
>
> Ok, let's see how I understand eglGetProcAddress should work in this case:
> - eglapi.c:eglGetProcAddress
> - egldriver.c:_eglGetDriverProc
> - egldriver.c:_eglMatchDriver, dpy->Options.TestOnly = EGL_TRUE
> - egldriver.c:_eglMatchAndInitialize
> - Calls drv->API.Initialize = egl_dri2.c:dri2_initialize
> - Returns EGL_TRUE as TestOnly is true (before "egl/dri2:
> dri2_initialize: Do not reference-count TestOnly display", it would
> return EGL_FALSE, so eglGetProcAddress would fail)
>
> So I don't really see how 9ee683f877 + "egl/dri2: dri2_initialize: Do
> not reference-count TestOnly display" would fail (while reverting
> 9ee683f877 would work).
>
> Any chance you could help tracing through the code? Or at least enable
> debugging output?

Sorry for the noise, a make clean fixed the issue :s

I can confirm that "egl/dri2: dri2_initialize: Do not reference-count 
TestOnly display" fixes the issue, you can add my Tested-by!