[rocprofiler-sdk] Fix signal-handler re-entrancy deadlock and enable app-abort testsadd a thread re-entrancy guard

[minseobshin11]

Motivation

The rocprofv3-test-app-abort integration tests (4 tests) were disabled unconditionally as "currently unstable". The instability was a real deadlock in rocprofv3's abort signal handler that caused the profiled process to hang on every abort. This PR fixes that deadlock and re-enables the tests.

Technical Details

Root cause: rocprofv3 installs its signal handler before HSA does, so when HSA later calls sigaction() it receives rocprofv3's handler as its "previous" handler. On an abort signal, rocprofv3_error_signal_handler finalizes output and invokes the chained HSA handler, which calls back the previous handler (rocprofv3's), re-entering the handler on the same thread while still inside std::call_once. Recursive use of std::once_flag deadlocks, so the process hangs and CTest times out. (SA_RESETHAND doesn't help — the re-entry is a direct call, not a fresh signal delivery.)

Fix:

• source/lib/rocprofiler-sdk-tool/tool.cpp: add a thread_local re-entrancy guard at handler entry that breaks the chain loop and terminates instead of recursing into call_once. Preserves cross-thread call_once semantics and is async-signal-safe.
• tests/rocprofv3/aborted-app/CMakeLists.txt: remove the unconditional IS_DISABLED ON (still disabled under ThreadSanitizer).

JIRA ID

Resolves AIROCVAL-48

Test Plan

Built for gfx942 (MI300X, ROCm 7.2) and ran the app-abort tests under multiple conditions:

• ctest -R app-abort (execute + 3 validate steps).
• ctest --repeat until-fail:50 under heavy CPU contention (load ~205).
• ~470+ abort invocations total: serial loops, 192-way concurrent, and under CPU/memory contention, with hang detection.

Test Result

100% pass; the execute step completes in ~1s (previously a 45s hang). Zero hangs across all runs and no leftover/spinning processes. The guard also correctly handled an induced SIGABRT loop-back during stress, terminating cleanly instead of deadlocking.

Submission Checklist

• [V] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

Code Coverage Report

Code Coverage Report

Tests Only

Samples Only

Tests + Samples

[mcao59]

I see this is still in draft, but just wanted to flag that #6717 takes a broader approach to redesigning rocprofv3 signal handling.
May be worth coordinating the direction here so we don’t end up with overlapping or conflicting changes :)

[minseobshin11]

I see this is still in draft, but just wanted to flag that #6717 takes a broader approach to redesigning rocprofv3 signal handling. May be worth coordinating the direction here so we don’t end up with overlapping or conflicting changes :)

Thanks for the heads up! I agree there's overlap. #6717 reworks the whole handler to run finalization on a worker thread, which would address the same re-entrancy root cause and make this thread_local guard unnecessary. My PR was intended as a minimal fix :) I believe if #6717 can be merged, it will fix the app-abort flakiness as well!