[Mesa-dev] [PATCH 3/3] egl/android: Dequeue buffers inside EGL calls (v2)

[Tomasz Figa]

On Wed, Apr 19, 2017 at 11:51 PM, Emil Velikov <emil.l.velikov at gmail.com> wrote:
> Hi Tomasz,
>
> On 19 April 2017 at 08:00, Tomasz Figa <tfiga at chromium.org> wrote:
>> Android buffer queues can be abandoned, which results in failing to
>> dequeue next buffer. Currently this would fail somewhere deep within
>> the DRI stack calling loader's getBuffers*(), without any error
>> reporting to the client app. However Android framework code relies on
>> proper signaling of this event, so we move buffer dequeue to
>> createWindowSurface() and swapBuffers() call, which can generate proper
>> EGL errors. To keep the performance benefits of delayed buffer handling,
>> if any, fence wait and DRI image creation is kept delayed until
>> getBuffers*() is called by the DRI driver.
>>
>> v2:
>>  - add missing fence wait in DRI loader case,
>>  - split simple code moving to a separate patch (Emil),
>>  - fix previous rebase error.
>>
>> Signed-off-by: Tomasz Figa <tfiga at chromium.org>
>> ---
>>  src/egl/drivers/dri2/egl_dri2.h         |  1 +
>>  src/egl/drivers/dri2/platform_android.c | 94 +++++++++++++++++++--------------
>>  2 files changed, 54 insertions(+), 41 deletions(-)
>>
>> diff --git a/src/egl/drivers/dri2/egl_dri2.h b/src/egl/drivers/dri2/egl_dri2.h
>> index f16663712d..92b234d622 100644
>> --- a/src/egl/drivers/dri2/egl_dri2.h
>> +++ b/src/egl/drivers/dri2/egl_dri2.h
>> @@ -288,6 +288,7 @@ struct dri2_egl_surface
>>  #ifdef HAVE_ANDROID_PLATFORM
>>     struct ANativeWindow *window;
>>     struct ANativeWindowBuffer *buffer;
>> +   int acquire_fence_fd;
>>     __DRIimage *dri_image_back;
>>     __DRIimage *dri_image_front;
>>
>> diff --git a/src/egl/drivers/dri2/platform_android.c b/src/egl/drivers/dri2/platform_android.c
>> index 9a84a4c43d..0a75d790c5 100644
>> --- a/src/egl/drivers/dri2/platform_android.c
>> +++ b/src/egl/drivers/dri2/platform_android.c
>> @@ -189,15 +189,9 @@ droid_free_local_buffers(struct dri2_egl_surface *dri2_surf)
>>     }
>>  }
>>
>> -static EGLBoolean
>> -droid_window_dequeue_buffer(struct dri2_egl_surface *dri2_surf)
>> +static void
>> +wait_and_close_acquire_fence(struct dri2_egl_surface *dri2_surf)
>>  {
>> -   int fence_fd;
>> -
>> -   if (dri2_surf->window->dequeueBuffer(dri2_surf->window, &dri2_surf->buffer,
>> -                                        &fence_fd))
>> -      return EGL_FALSE;
>> -
>>     /* If access to the buffer is controlled by a sync fence, then block on the
>>      * fence.
>>      *
>> @@ -215,15 +209,25 @@ droid_window_dequeue_buffer(struct dri2_egl_surface *dri2_surf)
>>      *    any value except -1) then the caller is responsible for closing the
>>      *    file descriptor.
>>      */
>> -    if (fence_fd >= 0) {
>> +    if (dri2_surf->acquire_fence_fd >= 0) {
>>         /* From the SYNC_IOC_WAIT documentation in <linux/sync.h>:
>>          *
>>          *    Waits indefinitely if timeout < 0.
>>          */
>>          int timeout = -1;
>> -        sync_wait(fence_fd, timeout);
>> -        close(fence_fd);
>> +        sync_wait(dri2_surf->acquire_fence_fd, timeout);
>> +        close(dri2_surf->acquire_fence_fd);
>> +        dri2_surf->acquire_fence_fd = -1;
>>     }
>> +}
>> +
>> +static EGLBoolean
>> +droid_window_dequeue_buffer(_EGLDisplay *disp,
>> +                            struct dri2_egl_surface *dri2_surf)
>> +{
>> +   if (dri2_surf->window->dequeueBuffer(dri2_surf->window, &dri2_surf->buffer,
>> +                                        &dri2_surf->acquire_fence_fd))
>> +      return EGL_FALSE;
>>
>>     dri2_surf->buffer->common.incRef(&dri2_surf->buffer->common);
>>
>> @@ -254,6 +258,14 @@ droid_window_dequeue_buffer(struct dri2_egl_surface *dri2_surf)
>>        dri2_surf->back = &dri2_surf->color_buffers[0];
>>     }
>>
>> +   /* free outdated buffers and update the surface size */
>> +   if (dri2_surf->base.Width != dri2_surf->buffer->width ||
>> +       dri2_surf->base.Height != dri2_surf->buffer->height) {
>> +      droid_free_local_buffers(dri2_surf);
>> +      dri2_surf->base.Width = dri2_surf->buffer->width;
>> +      dri2_surf->base.Height = dri2_surf->buffer->height;
>> +   }
>> +
>>     return EGL_TRUE;
>>  }
>>
>> @@ -263,6 +275,9 @@ droid_window_enqueue_buffer(_EGLDisplay *disp,
>>  {
>>     struct dri2_egl_display *dri2_dpy = dri2_egl_display(disp);
>>
>> +   /* In case we haven't done any rendering. */
>> +   wait_and_close_acquire_fence(dri2_surf);
>> +
>>     /* To avoid blocking other EGL calls, release the display mutex before
>>      * we enter droid_window_enqueue_buffer() and re-acquire the mutex upon
>>      * return.
>> @@ -319,6 +334,7 @@ droid_create_surface(_EGLDriver *drv, _EGLDisplay *disp, EGLint type,
>>        _eglError(EGL_BAD_ALLOC, "droid_create_surface");
>>        return NULL;
>>     }
>> +   dri2_surf->acquire_fence_fd = -1;
>>
>>     if (!_eglInitSurface(&dri2_surf->base, disp, type, conf, attrib_list))
>>        goto cleanup_surface;
>> @@ -360,10 +376,18 @@ droid_create_surface(_EGLDriver *drv, _EGLDisplay *disp, EGLint type,
>>     if (window) {
>>        window->common.incRef(&window->common);
>>        dri2_surf->window = window;
>> +      if (!droid_window_dequeue_buffer(disp, dri2_surf)) {
> Haven't checked the refcounting too close, mostly a gut feeling.
>
> Do we need to refcount twice - once prior to calling
> droid_window_dequeue_buffer and a second time within?
> Hmm actually it's consistent with the teardown side - once in
> droid_destroy_surface itself and again in droid_window_enqueue_buffer.

These are different refcounts, one for the window and one the buffer.
However according to the ANativeWindow spec, it might not be necessary
to increment the refcount if the driver references the buffer only
between dequeue and queue. Still, I'd say it's something for a
separate cleanup.

>
> For the series
> Cc: <mesa-stable at lists.freedesktop.org>
> Reviewed-by: Emil Velikov <emil.velikov at collabora.com>

Thanks!

I'd also like to hear from Mauro if this version, combined with the
patch to use cancelBuffer instead of queueBuffer for destroy surface,
by any chance helps with his wallpaper issue.

Best regards,
Tomasz