Data race in Halide profiler

[steven-johnson]

From running with TSAN locally, I get a data race reported for access to halide_profiler_get_state():

WARNING: ThreadSanitizer: data race (pid=51457)
  Read of size 4 at 0x7ff87234c548 by thread T5:
    #0 Halide::Runtime::Internal::sampling_profiler_thread(void*) <null> (libthird_Uparty_Shalide_Slibhalide_Ulibrary_Uruntime.so+0x19bc6)
    #1 Halide::Runtime::Internal::spawn_thread_helper(void*) <null> (libthird_Uparty_Shalide_Slibhalide_Ulibrary_Uruntime.so+0xe1cd)

  Previous write of size 4 at 0x7ff87234c548 by main thread:
    #0 halide_profiler_pipeline_end <null> (libthird_Uparty_Shalide_Slibhalide_Ulibrary_Uruntime.so+0x1a2fa)
    #1 <user-provided halide code>

  Location is global 'halide_profiler_get_state::s' of size 48 at 0x7ff87234c538 (libthird_Uparty_Shalide_Slibhalide_Ulibrary_Uruntime.so+0x00000001d548)

  Thread T5 (tid=51474, running) created by main thread at:
    #0 pthread_create third_party/llvm/llvm-project/compiler-rt/lib/tsan/rtl/tsan_interceptors_posix.cpp:962:3 (halide_target_map_test+0xf828a)
    #1 halide_spawn_thread <null> (libthird_Uparty_Shalide_Slibhalide_Ulibrary_Uruntime.so+0xe185)
    #2 halide_profiler_pipeline_start <null> (libthird_Uparty_Shalide_Slibhalide_Ulibrary_Uruntime.so+0x19df0)
    #3 <user-provided halide code>

[abadams]

The profiler works by seeing who wins a race condition on average, so this is WAI.

[abadams]

Note

Automated triage: this issue was flagged by Claude as likely safe to close, for the
reason below. The reasoning was checked against the current source, and a second pass
tried and failed to refute it, but no human verified this specific text.
If the reason is wrong, please reopen the issue and say so — some fraction of these
are expected to be wrong.

Reason (CLOSE_WONTFIX): The maintainer answered in-thread that the profiler is a sampling profiler and the race is the mechanism, not a bug. The profiler has since been rewritten and the sampler now holds the profiler state lock across each sample, so the only remaining unsynchronized accesses are the intentional per-instance sampling ones.

Evidence: Issue thread: abadams 2020-02-04 'The profiler works by seeing who wins a race condition on average, so this is WAI.' Current code: src/runtime/profiler_common.cpp:150-169 (sampling_profiler_thread takes halide_mutex_lock(&s->lock) around halide_profiler_sample), after the rewrite in 10162e8 'Rework profiler - initial PR (#9157)'.

Comment and close: 'Working as intended — the sampling profiler deliberately races with the pipeline. The profiler was also rewritten in #9157 and the sampler now holds the profiler state lock while sampling (src/runtime/profiler_common.cpp:150). Suppress it under TSAN if it is noisy for you.'