analyze-build doesn’t honor cgroup CPU limitations if launched inside a container

[dsilakov]

When launched without debug, analyze-build creates a pool of size None:

https://github.com/llvm/llvm-project/blob/main/clang/tools/scan-build-py/lib/libscanbuild/analyze.py#L238

In this case, python creates a pool of os.cpu_count() items. Unfortunately, cpu_count() doesn’t honor cgroup limitations – this is a known cpython issue (python/cpython#80235) and it doesn’t look like it will be fixed in the near future.

This provides significant server load if you launch analyzer inside a container on some powerful machine.

At the same time, pylint has developed its own implementation of cpu counter which is aware of cgroups: _query_cpu(), https://github.com/PyCQA/pylint/blob/main/pylint/lint/run.py#L34

It would be nice to take _query_cpu() from pylint (or provide a similar solution) and call it instead of using None in https://github.com/llvm/llvm-project/blob/main/clang/tools/scan-build-py/lib/libscanbuild/analyze.py#L238

[llvmbot]

@llvm/issue-subscribers-clang-static-analyzer

[steakhal]

I rarely use scanbuild, but I think what you suggests makes sense. Put me on the review if you plan to propose a change.

[dsilakov]

Ok. Though after some experiments _cpu_count() implementation in pylint turned out to be not ideal and can work incorrectly on a real hardware - see pylint-dev/pylint#7338. So currently I use its truncated variant without a hack for AWS. Will monitor how it works and create a PR if everything is fine.