[compiler-rt] [ASan] Prevent ASan/LSan deadlock by preloading modules before error reporting (PR #131756)

via llvm-commits llvm-commits at lists.llvm.org
Sat Apr 19 03:43:18 PDT 2025


Camsyn wrote:

> I'm also noticing flakiness of this test on some of my x64 linux builders as well:
> 
> * https://lab.llvm.org/buildbot/#/builders/202/builds/765
> * https://lab.llvm.org/buildbot/#/builders/202/builds/766
> * https://lab.llvm.org/buildbot/#/builders/174/builds/16460
> * https://lab.llvm.org/buildbot/#/builders/174/builds/16432
> * https://lab.llvm.org/buildbot/#/builders/174/builds/16386
> * https://lab.llvm.org/buildbot/#/builders/174/builds/16339
> * https://lab.llvm.org/buildbot/#/builders/174/builds/16310
> * https://lab.llvm.org/buildbot/#/builders/174/builds/16292
> * https://lab.llvm.org/buildbot/#/builders/174/builds/16286

In my opinion, either the flakiness hits the flaw of the incomplete fix (`dlopen`/`dlclose` between the module preloading and the ASan error printing), or there is another deadlock path.

I managed to replicate the build locally, but unfortunately couldn't reproduce the flakiness (run `make check-asan` 10 times and execute `xxx/asan_lsan_deadlock.cpp.tmp` 100 times). In my environment, this flackiness only appears if the patch is commented out :(.

I will still actively look for the cause, and it would be better if there could be more information about the test failure.

Perhaps we need to temporarily disable the test case before fixing the flakiness? @vitalybuka 



https://github.com/llvm/llvm-project/pull/131756


More information about the llvm-commits mailing list