Really disable Vulkan API calls when running with -feature -Vulkan [356896486]

Fixed

Bug

Status Update

No update yet.

Description

ru...@gmail.com

created issue #1

Aug 2, 2024 12:35PM

The following Vulkan related problems were found on emulator 35.1.18 and emu-master-dev.

While running android emulators with big parallelism (10+ instances in parallel) i've noticed that occasionally there are errors like:

ERROR   | emuglConfig_get_vulkan_hardware_gpu: Failed to create vulkan instance error code: -9

which do seem to affect the succeeding sessions. Most likely it could be a problem in the gpu vulkan library or smth else, but the issue is that I'm running the emulators with -feature -Vulkan to turn of Vulkan in order to be able to use snapshots, but nevertheless some Vulkan APIs are called (as seen from the error) which seems to cause issues at my scale. The problematic place seems to be emuglConfig_get_vulkan_hardware_gpu in external/qemu/android/android-ui/modules/aemu-gl-init/src/android/opengl/emugl_config.cpp (https://android.googlesource.com/platform/external/qemu/+/refs/heads/emu-master-dev/android/android-ui/modules/aemu-gl-init/src/android/opengl/emugl_config.cpp#246)

stack:

#0  emuglConfig_get_vulkan_hardware_gpu (vendor=0x7fffffffaca0, major=0x7fffffffac9c, minor=0x7fffffffac98, patch=0x7fffffffac94) at /home/rudolfs/Development/Google/aosp/external/qemu/android/android-ui/modules/aemu-gl-init/src/android/opengl/emugl_config.cpp:263
#1  0x00007ffff50936b0 in emuglConfig_init (config=0x7fffffffaf50, gpu_enabled=true, gpu_mode=0x7ffff3ff9856 "host", gpu_option=0x0, bitness=64, no_window=false, blacklisted=false, has_guest_renderer=true, uiPreferredBackend=0, use_host_vulkan=false) at /home/rudolfs/Development/Google/aosp/external/qemu/android/android-ui/modules/aemu-gl-init/src/android/opengl/emugl_config.cpp:687
#2  0x0000555558301de9 in androidEmuglConfigInit (config=0x7fffffffaf50, avdName=0x5555594724b0 "pixel6_13.0_1", avdArch=0x5555596c43a0 "x86_64", apiLevel=33, hasGoogleApis=true, gpuOption=0x0, hwGpuModePtr=0x555558c243f8 <s_hwConfig+312>, wantedBitness=0, noWindow=false, uiPreferredBackend=WINSYS_GLESBACKEND_PREFERENCE_AUTO, hostGpuVulkanBlacklisted=0x7fffffffaf0e, forceUseHostGpuVulkan=false) at /home/rudolfs/Development/Google/aosp/external/qemu/android/android-ui/modules/aemu-gl-init/src/android/main-emugl.cpp:103
#3  0x0000555557ff315d in configAndStartRenderer (uiPreferredBackend=WINSYS_GLESBACKEND_PREFERENCE_AUTO, config_out=0x7fffffffb808) at /home/rudolfs/Development/Google/aosp/external/qemu/android/android-ui/modules/aemu-ui-window/src/android/main-common-ui.c:286
#4  0x0000555556ee6743 in main (argc=0, argv=0x7fffffffd5f8) at /home/rudolfs/Development/Google/aosp/external/qemu/android-qemu2-glue/main.cpp:3258

For now I've been able to get around this by renaming the vulkan folder under lib64 so that no Vulkan API calls are ever made (the flow just stops since it cannot load the library), but could it be possible to adjust this so that -feature -Vulkan really disabled Vulkan.

Another much smaller related issues is that even with -feature -Vulkan, the following lines are logged:

Deciding if GLDirectMem/Vulkan should be enabled. Selected renderer: 1 API level: 34 host GPU on the denylist? 0
Host GPU selected, enabling Vulkan.

This seems to come from main-common-ui.c (https://android.googlesource.com/platform/external/qemu/+/refs/heads/emu-master-dev/android/android-ui/modules/aemu-ui-window/src/android/main-common-ui.c#315) which does not seem to honour the feature flag (eventually Vulkan is not used since lower in the logic it does check the override state, but that could have been done at the very start to avoid the confusing log messages).

Comments

jp...@google.com <jp...@google.com> Aug 2, 2024 03:53PM

Assigned to ya...@google.com.

ya...@google.com <ya...@google.com> #2Aug 2, 2024 05:44PM

https://bit.ly/3UxSQmb

ya...@google.com <ya...@google.com> Oct 1, 2024 09:59PM

Reassigned to ko...@google.com.

ko...@google.com <ko...@google.com> Oct 4, 2024 09:52AM

Accepted by ko...@google.com.

ko...@google.com <ko...@google.com> #3Oct 8, 2024 06:11PM

Marked as fixed.

Here is what I found: On API 31+, time was not updated after the host was waken up. While on API 30 and under, time was updated after host was waken up. The main difference is caused by switching the modem simulator. In API 30 and under, time will be updated whenever signal strength query is received. Reference

Looking at the code in API 31+ for time update when signal strength query is received, I think it is possible that timeUpdate is not invoked

void NetworkService::HandleSignalStrength(const Client& client) {
  std::vector<std::string> responses;
  std::stringstream ss;

  if (WakeupFromSleep()) {
    misc_service_->TimeUpdate();
  } else if (first_signal_strength_request_) {
    first_signal_strength_request_ = false;
    misc_service_->TimeUpdate();
  }

  android_last_signal_time_ = time(0);

  auto response = BuildCSQCommandResponse(GetCurrentSignalStrength());

  responses.push_back(response);
  responses.push_back("OK");
  client.SendCommandResponse(responses);
}

Issue 356896486