Incorrect detection number threads in multisocket motherboards on Windows

[GermanAizek]

Running Git on multi-CPU motherboard. I have in one CPU 18 cores and 36 threads, but there are two of them installed.

Found it in graphical Git client:
gitextensions/gitextensions#11523

Same behavior is in regular git CLI usage.

Setup

• Which version of Git for Windows are you using? Is it 32-bit or 64-bit?

$ git --version --build-options

** git version 2.42.0.windows.1
cpu: x86_64
built from commit: a2e49ec355ae22e74288fa74f50821f3cc1a1551
sizeof-long: 4
sizeof-size_t: 8
shell-path: /bin/sh
feature: fsmonitor--daemon **

• Which version of Windows are you running? Vista, 7, 8, 10? Is it 32-bit or 64-bit?

$ cmd.exe /c ver

** Microsoft Windows [Version 10.0.19045.3086] **

Details

• What commands did you run to trigger this issue?

** git clone <url> **

[dscho]

This is the code location responsible for determining the number of "CPUs" (which really means "number of cores"):

git/thread-utils.c

Lines 30 to 35 in 4991b61

	#ifdef GIT_WINDOWS_NATIVE
	SYSTEM_INFO info;
	GetSystemInfo(&info);
	if ((int)info.dwNumberOfProcessors > 0)
	return (int)info.dwNumberOfProcessors;

@GermanAizek any idea how you could detect the correct number instead, using the Win32 API?

[GermanAizek]

@dscho I have already fixed this with mrexodia in x64dbg x64dbg/x64dbg#3272
and this OpenXRay/xray-16@02aaaf9
and this https://github.com/llvm/llvm-project/ issues/72267

[GermanAizek]

@dscho GermanAizek/llvm-project@d1fa25f here is cleanest code after checking by llvm-clang contributors

[GermanAizek]

@dscho what is your C++ standard? I can do PR with an older C++11 standard

[rimrul]

@dscho what is your C++ standard? I can do PR with an older C++11 standard

somewhere between C89 and C99 (no ++).

[rimrul]

This is the code location responsible for determining the number of "CPUs" (which really means "number of cores"):

git/thread-utils.c

Lines 30 to 35 in 4991b61

	#ifdef GIT_WINDOWS_NATIVE
	SYSTEM_INFO info;
	GetSystemInfo(&info);
	if ((int)info.dwNumberOfProcessors > 0)
	return (int)info.dwNumberOfProcessors;

Could we even set processor group affinity for a thread without moving away from pthreads?

[GermanAizek]

somewhere between C89 and C99 (no ++).

I haven't written in pure С for a long time, but in theory remove lambda function and replace std:: functions should solve problem.

[dscho]

Could we even set processor group affinity for a thread without moving away from pthreads?

I guess that this GNU-specific extension is available to us: https://www.man7.org/linux/man-pages/man3/pthread_attr_setaffinity_np.3.html. However, Git's source code tries to shy away from GNU-specific extensions, in general...

[GermanAizek]

@dscho, I found solution for your project https://projects.blender.org/blender/blender/commit/8f1a558

[GermanAizek]

@dscho, not relevant, I found a solution more suitable for you, so as not to rewrite a lot and make support for regular motherboards and NUMA. More info here: #4896

[GermanAizek]

@dscho, screenshot from SDK with builded git branch fix-numa-cores

Now configuration correctly defines 72 threads and both CPU is fully loaded at 100%, and not as before 50-55%.