|
e5619a5c
|
2024-10-15T18:27:00
|
|
Use EGL sync global lock for all EGL*sync entrypoint calls
To free EGL sync operations from waiting for other EGL calls
to finish, we use a separate global lock for EGL*sync entrypoints.
Below angle::SimpleMutex are added to protect resources that may
have race condition due to being accessed by EGL*sync calls
and non EGL*sync calls at the same time:
1. Display::mContextMapMutex that protects Display::mState.contextMap
2. static angle::base::NoDestructor<angle::SimpleMutex>
anglePlatformDisplayMapMutex that protects static
angle::base::NoDestructor<ANGLEPlatformDisplayMap> displays
3. static angle::base::NoDestructor<angle::SimpleMutex>
devicePlatformDisplayMapMutex that protects static
angle::base::NoDestructor<DevicePlatformDisplayMap> displays
EGL_Terminate() entry point takes both global lock and
global egl sync lock. This is to protect Display::mSyncMap,
Display::mSyncPools, and Display::mSyncHandleAllocator being
get cleared by thread 1 calling eglTerminate, while they are
still accessed by thread 2 through a call such as eglCreateSync.
So that we won't have thread 2 finish validating the sync object with
syncID exists in Display::mSyncMap, but then find the
mSyncMap.find(syncID) returns a nullptr due to the mSyncMap
is cleared by thread 1. Same applies to EGL_LabelObjectKHR(),
EGL_ReleaseThread(), ThreadCleanupCallback().
EGL_Initialize() writes to Display::mInitialized. This is read
by EGL Sync API validation functions. EGL_Initialize() also takes
both global lock and global sync lock to prevent race conditions
between EGL_Initialize() and EGL Sync APIs.
When ANGLE_CAPTURE_ENABLED is enabled, fall back to global lock,
as the CaptureEGLCallToFrameCapture() touches many resources
(e.g. mMaxAccessedResourceIDs, mFrameCalls)
that could lead to race conditions without a global lock.
Bug: b/362604439
Change-Id: Ic0d54a4cd66743bcd0f48f41f247dd223cff2f5e
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/5933570
Reviewed-by: Roman Lavrov <romanl@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Igor Nazarov <i.nazarov@samsung.com>
Commit-Queue: Yuxin Hu <yuxinhu@google.com>
|
|
25ce3dfe
|
2023-06-29T15:51:55
|
|
Use compare_exchange_weak() in AllocateGlobalMutexImpl()
Speculative fix of the crash.
New logic will ensure that `currentMutex` is not `nullptr` when returned
from `AllocateGlobalMutexImpl()`
Also removed explicit memory order for safety (no actual difference in
generated code on ARM64/x86-64, but ARM generates less instructions).
Bug: chromium:1457915
Change-Id: I8d932ee499f9d8ee4e38ab2173f4f1cefd0aa294
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/4650794
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
8b97c8d3
|
2023-03-31T15:54:40
|
|
Implement build option for load-time GlobalMutex allocation.
Performance difference on S906B for 1000'000 `eglGetError()` calls.
Before: 26.5 ms
After: 23.8 ms
(Assuming `angle::FastMutex2` is used).
Potentially can be enabled for other platforms.
Bug: angleproject:8101
Change-Id: Ic2037edaae6da82dbded71288bae250d7981e7d3
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/4386408
Reviewed-by: Charlie Lao <cclao@google.com>
Commit-Queue: Igor Nazarov <i.nazarov@samsung.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
b22124bd
|
2023-01-17T17:44:06
|
|
Implement new "Global Mutex" functionality.
This CL improves performance compared to the existing code as well as
allows using non-std mutex implementations. Also acts as a base for
future changes.
CL adds new build option:
angle_enable_global_mutex_recursion = is_android && angle_enable_vulkan
"mutex_recursion" work same way as `std::recursive_mutex` before. It
will help in situations when Vulkan API may return back to the ANGLE.
For example: RenderDoc layer EGL deadlock.
Automatic loading of "libVkLayer_GLES_RenderDoc.so" layer causes
deadlock in EGL.
Recursion stack:
#01 pc 000000000029ea80 /vendor/lib64/egl/libGLESv2_angle.so (egl::GlobalMutexHelper::lock(int)+596)
#02 pc 000000000029c59c /vendor/lib64/egl/libGLESv2_angle.so (EGL_GetError+32)
#04 pc 0000000000062368 /system/lib64/libEGL.so (eglQueryString+20)
#05 pc 0000000000508fec /data/local/debug/vulkan/libVkLayer_GLES_RenderDoc.so
#20 pc 0000000000016690 /system/lib64/libvulkan.so (vulkan::api::EnumerateInstanceLayerProperties(unsigned int*, VkLayerProperties*)+40)
#21 pc 00000000005aa030 /vendor/lib64/egl/libGLESv2_angle.so (rx::RendererVk::initialize(rx::DisplayVk*, egl::Display*, char const*, char const*)+292)
#26 pc 000000000029c7e8 /vendor/lib64/egl/libGLESv2_angle.so (EGL_Initialize+192)
Additionally, recursive mutex will partially solve Android
SurfaceTexture deadlock (angleproject:4354).
Some performance numbers for 1000'000 `eglGetError()` calls.
Mutex Time (ms.)
Android S906B
egl::GetGlobalMutex()(std::recursive_mutex) 41.4
(Default) GlobalMutex (std::recursive_mutex) 39.1
(Recursive) GlobalMutex (std::mutex) 34.9
(Debug) GlobalMutex (std::mutex) 34.7
(Default) GlobalMutex (std::mutex) 34.4
Windows
egl::GetGlobalMutex()(std::recursive_mutex) 20.5
(Default) GlobalMutex (std::recursive_mutex) 20.0
(Recursive) GlobalMutex (std::mutex) 21.9
(Debug) GlobalMutex (std::mutex) 20.5
(Default) GlobalMutex (std::mutex) 19.9
Note: Recursive GlobalMutex enabled only for Android Vulkan by default.
Original fix:
https://chromium-review.googlesource.com/c/angle/angle/+/2029218
Bug: angleproject:8101
Bug: angleproject:4354
Test: angle_unittests --gtest_filter=GlobalMutexTest.*
Change-Id: I9e9d9b5c598ad1177ffa147ea690bd955946a712
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/4401940
Commit-Queue: Igor Nazarov <i.nazarov@samsung.com>
Reviewed-by: Charlie Lao <cclao@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
|