CI: sigprof_race regression test crashes on macOS even with SIG_IGN fix applied

[korniltsev-grafanista-yolo-vibecoder239]

Summary

The sigprof_race regression test (added in PR #26) crashes on macOS CI even with the SIG_IGN fix applied. Both macOS test jobs fail:

• Test (macos-latest, stable, x86_64-apple-darwin) — signal crash
• Test (macos-latest, nightly, x86_64-apple-darwin) — signal crash

Linux CI and local Linux runs pass consistently (42 runs, 0 failures).

Context

PR #26 adds a regression test for the SIGPROF race condition fixed in 978d3aa. The test rapidly cycles ProfilerGuard (start/stop profiler) 8000 times at 999 Hz with background CPU-burning threads. Without the fix, a pending SIGPROF delivered after SIG_DFL is restored crashes the process.

The fix uses SIG_IGN instead of restoring SIG_DFL, which should safely ignore any pending signals. This works on Linux but apparently not on macOS.

Possible causes to investigate

1. macOS signal delivery semantics differ from Linux — macOS may deliver SIGPROF differently (e.g., to specific threads, or with different timing relative to sigaction() calls)
2. SIG_IGN may not take effect atomically on macOS — there could be a window where the old handler is gone but SIG_IGN isn't yet installed
3. Timer behavior — setitimer(ITIMER_PROF) may behave differently on macOS, possibly delivering signals after the timer is stopped
4. The test itself may be too aggressive for macOS — the 8000 iterations + 4 threads + 999 Hz combination may need platform-specific tuning

CI run

https://github.com/grafana/pprof-rs/actions/runs/22701001347

Related

• PR test: add regression test for SIGPROF race crash in profiler start/stop cycles #26
• Original fix: 978d3aa
• unregister_signal_handler restores SIG_DFL, causing process termination from stray SIGPROF tikv/pprof-rs#288

[korniltsev-grafanista-yolo-vibecoder239]

CI Failure Logs

Run 22700026022 (commit dead34f — fix applied)

macOS stable (x86_64-apple-darwin)

Job: 65814855824

Running tests/sigprof_race.rs (target/x86_64-apple-darwin/debug/deps/sigprof_race-36c691bbc596f681)
test result: ok. 13 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 6.58s

error: test failed, to rerun pass `--test sigprof_race`
test test_sigprof_race_crash ... 
  process didn't exit successfully: `.../sigprof_race-36c691bbc596f681 --test-threads 1` (signal: 27)
Process completed with exit code 101.

macOS nightly (x86_64-apple-darwin)

Job: 65814855813

Running tests/sigprof_race.rs (target/x86_64-apple-darwin/debug/deps/sigprof_race-36716b58ad20b1d7)
test result: ok. 13 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 5.98s

error: test failed, to rerun pass `--test sigprof_race`
test test_sigprof_race_crash ... 
  process didn't exit successfully: `.../sigprof_race-36716b58ad20b1d7 --test-threads 1` (signal: 27)
Process completed with exit code 101.

Both macOS jobs crash with signal 27 (SIGPROF) almost immediately (~1 second into the test). All 13 unit tests pass before the integration test runs. The SIG_IGN fix is applied in this commit — it prevents crashes on Linux but not on macOS.

[korniltsev-grafanista-yolo-vibecoder239]

Correction: Logs from the correct run (commit 13150e3, fix applied)

The previous comment linked logs from commit dead34f which had the fix reverted — those crashes were expected. Here are the logs from run 22701001347 for commit 13150e3 which has the SIG_IGN fix applied:

macOS stable (x86_64-apple-darwin)

Job: 65817904978

Running tests/sigprof_race.rs (target/x86_64-apple-darwin/debug/deps/sigprof_race-36c691bbc596f681)
test result: ok. 13 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 6.66s

error: test failed, to rerun pass `--test sigprof_race`
test test_sigprof_race_crash ... 
  process didn't exit successfully: `.../sigprof_race-36c691bbc596f681 --test-threads 1` (signal: 10, SIGBUS: access to undefined memory)
Process completed with exit code 101.

macOS nightly (x86_64-apple-darwin)

Job: 65817904962

Running tests/sigprof_race.rs (target/x86_64-apple-darwin/debug/deps/sigprof_race-36716b58ad20b1d7)
test result: ok. 13 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 5.66s

error: test failed, to rerun pass `--test sigprof_race`
test test_sigprof_race_crash ... 
  process didn't exit successfully: `.../sigprof_race-36716b58ad20b1d7 --test-threads 1` (signal: 10, SIGBUS: access to undefined memory)
Process completed with exit code 101.

Key finding

The crash is SIGBUS (signal 10: access to undefined memory), NOT SIGPROF (signal 27). This suggests the issue on macOS is not the same SIG_DFL race — it may be a memory corruption or signal stack issue triggered by the aggressive profiler cycling in the test.