When the GPU process crashes repeatedly, usually 3 crashes in a short timeframe, a less desirable but hopefully more stable mode will be chosen for the next startup. Where available, we want to use Vulkan over GL, and GL over software. Vulkan is more experimental than GL and doesn‘t have the same level of support or maturity. Software rendering should be more stable than accelerated graphics because it doesn’t rely on third party drivers, or on graphics hardware outside our control.
GPU process crashes are counted in GpuProcessHost::RecordProcessCrash(). Occasional crashes are permitted without triggering the fallback behaviour, so one crash is forgiven per interval elapsed since the last crash. For example, if the GPU crashes for a second time after N minutes, the previous crash is forgiven. This interval is determined by GetForgiveMinutes(). If the crash limit is hit, then fallback occurs.
The crash limit, determined by kGpuFallbackCrashCount, is higher on Android and ChromeOS because these platforms don't support software rendering. Additionally, on Android, the OS can kill the GPU process arbitrarily. So the crash tolerance is a little bit wider here.
The order of GPU process modes is determined at startup based on the platform, command line switches, and enabled features. This is then stored as a stack in GpuDataManagerImplPrivate::fallback_modes_.
For example, on Linux with all options available, the stack of GPU modes will look like this after initialization:
+--------------------+ | HARDWARE_VULKAN | <-- +--------------------+ | HARDWARE_GL | +--------------------+ | SWIFTSHADER | +--------------------+ | DISPLAY_COMPOSITOR | +--------------------+
After the order is determined, or when fallback occurs, the top element from the stack is popped and used as the next GPU mode. If the stack is empty when fallback occurs, then the GPU process is too unstable to use, and the Browser process intentionally crashes.
HARDWARE_GLThe GPU process is running with OpenGL hardware acceleration enabled.
HARDWARE_METAL and HARDWARE_VULKANThe GPU process is running with OpenGL hardware acceleration enabled, as well as Metal or Vulkan.
This doesn't necessarily determine what Metal or Vulkan are being used for, just that they are initialized. In particular, for Vulkan, --enable-features=Vulkan will cause Vulkan to be used for compositing and rasterization, whereas --use-vulkan by itself will only initialize Vulkan so that it can be used for other purposes, such as WebGPU.
SWIFTSHADERThe GPU process is running with hardware acceleration disabled, but SwiftShader will be initialized for software-backed WebGL.
DISPLAY_COMPOSITORThe GPU process is running for the display compositor only, no acceleration is enabled.
There are a few platforms that expect hardware acceleration, with some exceptions for certain circumstances.
Android requires hardware acceleration, except in the case of Chromecast audio-only builds. These run with the flag --disable-gpu, and the GPU process is launched in DISPLAY_COMPOSITOR mode.
Fuchsia always expects Vulkan to be available, and doesn't support falling back to using GL or software. For testing purposes, the flag --disable-gpu is allowed, and the GPU process will be launched in SWIFTSHADER mode.