[client] egl: force DMA copy into texture in on_frame

This commit forces the DMA'd memory to be copied into the texture in
the EGL on_frame handler. This avoids tearing when the LG host inevitably
updates the underlying memory. We need an additional copy inside the GPU,
but this is cheap compared to copying from system memory.

We could have used logic to lock the memory buffer, but that would require
performing DMA on every frame, which wastes memory bandwidth. This
manifests as reduced frame rate when moving the mouse compared to the
non-DMA implementation.

We also keep multiple EGLImages, one for each DMA fd, to avoid issues
with the OpenGL driver.

client/renderers/EGL/texture.c

@@ -79,6 +79,18 @@ struct EGL_Texture
   int             bufferCount;
   GLuint          tex;
   struct Buffer   buf[BUFFER_COUNT];
+
+  size_t dmaImageCount;
+  size_t dmaImageUsed;
+  struct
+  {
+    int fd;
+    EGLImage image;
+  }
+  * dmaImages;
+
+  GLuint dmaFBO;
+  GLuint dmaTex;
 };
 
 bool egl_texture_init(EGL_Texture ** texture, EGLDisplay * display)
@@ -121,6 +133,10 @@ void egl_texture_free(EGL_Texture ** texture)
   glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
   glDeleteTextures(1, &(*texture)->tex);
 
+  for (size_t i = 0; i < (*texture)->dmaImageUsed; ++i)
+    eglDestroyImage((*texture)->display, (*texture)->dmaImages[i].image);
+  free((*texture)->dmaImages);
+
   free(*texture);
   *texture = NULL;
 }
@@ -248,7 +264,15 @@ bool egl_texture_setup(EGL_Texture * texture, enum EGL_PixelFormat pixFmt, size_
   }
 
   if (useDMA)
+  {
+    if (texture->dmaFBO)
+      glDeleteFramebuffers(1, &texture->dmaFBO);
+    if (texture->dmaTex)
+      glDeleteTextures(1, &texture->dmaTex);
+    glGenFramebuffers(1, &texture->dmaFBO);
+    glGenTextures(1, &texture->dmaTex);
     return true;
+  }
 
   glBindTexture(GL_TEXTURE_2D, texture->tex);
   glTexImage2D(GL_TEXTURE_2D, 0, texture->intFormat, texture->width,
@@ -365,41 +389,93 @@ bool egl_texture_update_from_dma(EGL_Texture * texture, const FrameBuffer * fram
     return true;
   }
 
-  EGLAttrib const attribs[] =
-  {
-    EGL_WIDTH                    , texture->width,
-    EGL_HEIGHT                   , texture->height,
-    EGL_LINUX_DRM_FOURCC_EXT     , texture->fourcc,
-    EGL_DMA_BUF_PLANE0_FD_EXT    , dmaFd,
-    EGL_DMA_BUF_PLANE0_OFFSET_EXT, 0,
-    EGL_DMA_BUF_PLANE0_PITCH_EXT , texture->stride,
-    EGL_NONE                     , EGL_NONE
-  };
+  EGLImage image = EGL_NO_IMAGE;
 
-  /* create the image backed by the dma buffer */
-  EGLImage image = eglCreateImage(
-    texture->display,
-    EGL_NO_CONTEXT,
-    EGL_LINUX_DMA_BUF_EXT,
-    (EGLClientBuffer)NULL,
-    attribs
-  );
+  for (int i = 0; i < texture->dmaImageUsed; ++i)
+  {
+    if (texture->dmaImages[i].fd == dmaFd)
+    {
+      image = texture->dmaImages[i].image;
+      break;
+    }
+  }
 
   if (image == EGL_NO_IMAGE)
   {
-    DEBUG_EGL_ERROR("Failed to create ELGImage for DMA transfer");
-    return false;
-  }
+    EGLAttrib const attribs[] =
+    {
+      EGL_WIDTH                    , texture->width,
+      EGL_HEIGHT                   , texture->height,
+      EGL_LINUX_DRM_FOURCC_EXT     , texture->fourcc,
+      EGL_DMA_BUF_PLANE0_FD_EXT    , dmaFd,
+      EGL_DMA_BUF_PLANE0_OFFSET_EXT, 0,
+      EGL_DMA_BUF_PLANE0_PITCH_EXT , texture->stride,
+      EGL_NONE                     , EGL_NONE
+    };
 
-  /* bind the texture and initiate the transfer */
-  glBindTexture(GL_TEXTURE_2D, texture->tex);
-  g_egl_dynProcs.glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, image);
+    /* create the image backed by the dma buffer */
+    image = eglCreateImage(
+      texture->display,
+      EGL_NO_CONTEXT,
+      EGL_LINUX_DMA_BUF_EXT,
+      (EGLClientBuffer)NULL,
+      attribs
+    );
+
+    if (image == EGL_NO_IMAGE)
+    {
+      DEBUG_EGL_ERROR("Failed to create ELGImage for DMA transfer");
+      return false;
+    }
+
+    if (texture->dmaImageUsed == texture->dmaImageCount)
+    {
+      size_t newCount = texture->dmaImageCount * 2 + 2;
+      void * new = realloc(texture->dmaImages, newCount * sizeof *texture->dmaImages);
+      if (!new)
+      {
+        DEBUG_EGL_ERROR("Failed to allocate memory");
+        eglDestroyImage(texture->display, image);
+        return false;
+      }
+      texture->dmaImageCount = newCount;
+      texture->dmaImages     = new;
+    }
+
+    const size_t index = texture->dmaImageUsed++;
+    texture->dmaImages[index].fd    = dmaFd;
+    texture->dmaImages[index].image = image;
+  }
 
   /* wait for completion */
   framebuffer_wait(frame, texture->height * texture->stride);
 
-  /* destroy the image to prevent future writes corrupting the display image */
-  eglDestroyImage(texture->display, image);
+  glBindTexture(GL_TEXTURE_2D, texture->dmaTex);
+  g_egl_dynProcs.glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, image);
+
+  glBindFramebuffer(GL_FRAMEBUFFER, texture->dmaFBO);
+  glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture->dmaTex, 0);
+
+  glBindTexture(GL_TEXTURE_2D, texture->tex);
+  glCopyTexImage2D(GL_TEXTURE_2D, 0, texture->intFormat, 0, 0, texture->width, texture->height, 0);
+
+  GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+  glFlush();
+
+  switch (glClientWaitSync(fence, 0, 10000000)) // 10ms
+  {
+    case GL_ALREADY_SIGNALED:
+    case GL_CONDITION_SATISFIED:
+      break;
+
+    case GL_TIMEOUT_EXPIRED:
+      egl_warn_slow();
+      break;
+
+    case GL_WAIT_FAILED:
+    case GL_INVALID_VALUE:
+      DEBUG_EGL_ERROR("glClientWaitSync failed");
+  }
 
   atomic_fetch_add_explicit(&texture->state.w, 1, memory_order_release);
   return true;