diff options
Diffstat (limited to 'absl/container/internal/hashtablez_sampler.h')
| -rw-r--r-- | absl/container/internal/hashtablez_sampler.h | 19 |
1 files changed, 15 insertions, 4 deletions
diff --git a/absl/container/internal/hashtablez_sampler.h b/absl/container/internal/hashtablez_sampler.h index ee4d293e..0fd9349f 100644 --- a/absl/container/internal/hashtablez_sampler.h +++ b/absl/container/internal/hashtablez_sampler.h @@ -82,16 +82,27 @@ struct HashtablezInfo : public profiling_internal::Sample<HashtablezInfo> { std::atomic<size_t> hashes_bitwise_xor; std::atomic<size_t> max_reserve; + // One could imagine that inline_element_size could be non-atomic, since it + // *almost* follows the rules for the fields that are set by + // `PrepareForSampling`. However, TSAN reports a race (see b/207323922) in + // which + // A: Thread 1: Register() returns, unlocking init_mu. + // B: Thread 2: Iterate() is called, locking init_mu. + // C: Thread 1: inlined_element_size is stored. + // D: Thread 2: inlined_element_size is accessed (a race). + // A simple solution is to make inline_element_size atomic so that we treat it + // at as we do the other atomic fields. + std::atomic<size_t> inline_element_size; + // All of the fields below are set by `PrepareForSampling`, they must not be // mutated in `Record*` functions. They are logically `const` in that sense. - // These are guarded by init_mu, but that is not externalized to clients, who - // can only read them during `HashtablezSampler::Iterate` which will hold the - // lock. + // These are guarded by init_mu, but that is not externalized to clients, + // which can read them only during `SampleRecorder::Iterate` which will hold + // the lock. static constexpr int kMaxStackDepth = 64; absl::Time create_time; int32_t depth; void* stack[kMaxStackDepth]; - size_t inline_element_size; }; inline void RecordRehashSlow(HashtablezInfo* info, size_t total_probe_length) { |