Impeller Vulkan: Infinite timeout in vkWaitForFences / acquireNextImageKHR causes ANR on Android when GPU is under pressure #184871

Steps to reproduce

1. Run a Flutter application on an Android device with Adreno GPU (Impeller/Vulkan backend)
2. While the Flutter app is in the foreground, trigger GPU pressure through one of:
   • Launch a GPU-intensive game (e.g., cold start of a game that performs heavy shader compilation and texture uploads)
   • Allow DMA-BUF memory leak to accumulate (observed 1.9GB~3.1GB in production)
3. While GPU is under pressure, switch away from the Flutter app (press Home or switch to another app), triggering Activity.onStop() → onSurfaceDestroyed()
4. The app ANRs within 5-10 seconds

Expected results

The Flutter engine should gracefully handle GPU pressure during surface teardown. When the GPU cannot complete rendering within a reasonable time, the engine should:

• Time out the fence wait
• Abandon the current frame (dropping one frame is acceptable)
• Free the raster thread to process the teardown task
• Allow the platform thread to complete the Activity lifecycle without blocking indefinitely

Actual results

The platform (main) thread is blocked indefinitely, causing an ANR (Application Not Responding). The blocking chain:

Platform thread:
  Shell::OnPlatformViewDestroyed()     [shell.cc:1189]
    → latch.Wait()                     // no timeout, waits forever
      → waiting for raster thread to complete teardown

Raster thread (blocked, cannot process teardown):
  KHRSwapchainImplVK::AcquireNextDrawable()
    → KHRFrameSynchronizerVK::WaitForFence()  [khr_swapchain_impl_vk.cc:54]
      → vkWaitForFences(timeout=UINT64_MAX)   // infinite wait on GPU fence

The raster thread is stuck in vkWaitForFences with UINT64_MAX timeout because the GPU cannot signal the fence (due to contention or memory pressure). Since the raster thread never completes the current draw, it never picks up the teardown task from its queue, so the latch is never signaled, and the platform thread blocks forever.

Code sample

bool WaitForFence(const vk::Device& device) {
    if (!acquire_fence_pending) {
      return true;
    }
    if (auto result = device.waitForFences(
            *acquire,                             // fence
            true,                                 // wait all
            std::numeric_limits<uint64_t>::max()  // timeout (ns) ← INFINITE
        );
        result != vk::Result::eSuccess) {
      VALIDATION_LOG << "Fence wait failed: " << vk::to_string(result);
      return false;
    }
    // ...
}

// @bug  Non-infinite timeouts are not supported on some older Android
//       devices and the only indication we get is log spam which serves to
//       add confusion. Just use an infinite timeout instead of being
//       defensive.
auto [acq_result, index] = context.GetDevice().acquireNextImageKHR(
    *swapchain_,                           // swapchain
    std::numeric_limits<uint64_t>::max(),  // timeout (ns) ← INFINITE
    *sync->render_ready,                   // signal semaphore
    nullptr                                // fence
);

// Shell::OnPlatformViewDestroyed()
fml::TaskRunner::RunNowOrPostTask(task_runners_.GetRasterTaskRunner(),
                                  raster_task);
latch.Wait();  // ← no timeout, blocks platform thread forever

Screenshots or Video

Logs

Flutter Doctor output

Flutter 3.38
Android devices: Qualcomm Adreno GPU (Snapdragon platforms)
Affected devices observed: Xiaomi Mix Filp2  REDMI K90