ConvolverNode: SIGSEGV in Convolver::process caused by a race in ThreadPool::wait()

Start with common/cpp/audioapi/utils/ThreadPool.hpp, focusing on schedule() and wait(), then trace the captured processingBuffer references in ConvolverNode::processNode. Reproduce with the issue’s standalone harness and widened scheduling window. Done means wait() does not return before all tasks in the current batch complete, with a regression test covering the race and no dangling buffer access.

Description

We got a native crash in production (Android, arm64) inside Convolver::process:

#00 libc.so (__memcpy_aarch64_simd+232)
#01 libreact-native-audio-api.so (audioapi::Convolver::process(audioapi::AlignedAudioArray<64ul> const&, audioapi::AlignedAudioArray<64ul>&)+112)
#02 libreact-native-audio-api.so (audioapi::move_only_function<void ()>::callable_impl<... audioapi::ThreadPool::schedule<audioapi::ConvolverNode::performConvolution(std::shared_ptr<audioapi::AlignedAudioBuffer<64ul>> const&)::$_1 ...>::operator()())
#03 libreact-native-audio-api.so (audioapi::ThreadPool::workerThreadFunc(...)+220)
#04 libreact-native-audio-api.so (std::__thread_proxy ...)
#05 libc.so (__pthread_start)

It has happened once so far. I think it comes from a race in ThreadPool::wait() (common/cpp/audioapi/utils/ThreadPool.hpp). The code is the same in 0.13.2, 0.13.6 and on main as of today.

The interleaving:

1. Quantum N: the audio thread schedules the convolver tasks. All of them finish before the audio thread reaches wait(). The last worker has done tasksScheduled.fetch_sub (count is now 0), but it gets preempted before waitingForTasks.store(false) / notify_one().
2. The audio thread calls wait(): it stores true, sees tasksScheduled == 0, stores false and returns. That is correct for quantum N.
3. Quantum N+1: the audio thread schedules new tasks (count > 0), calls wait(), stores true and blocks in waitingForTasks.wait(true).
4. The worker from step 1 resumes and runs store(false) + notify_one(). wait() returns while the tasks from quantum N+1 are still pending or running.

After that, ConvolverNode::processNode returns early. The scheduled lambdas capture processingBuffer by reference ([&, i] in the mono branch, &processingBuffer in the stereo branch). That is a const std::shared_ptr<DSPAudioBuffer>& bound to a local in AudioNode::processAudio, so once the frame is gone the worker reads a dangling reference. Convolver::process then copies from it (_inputBuffer->copy(input, ...)), which matches the memcpy frame at process+112.

Steps to reproduce

The race window is a few instructions wide, so it does not show up in normal use. I reproduced it with a standalone harness that compiles ThreadPool.hpp as is, on macOS with clang++ -std=c++20.

• Stock header: with fast tasks, 2,000,000 quanta give 0 early returns. The window is too small to hit by chance.
• Widened window: I put a 2.5 ms sleep between fetch_sub reaching 0 and store(false) on 1 in 7 tasks. I also delay the scheduling thread for 3 ms before wait() on every third quantum, which simulates the audio thread being preempted. The sleep only stands in for a preempted worker and does not change the logic. In this setup, wait() returned with 0/2 tasks done in 142 of 3,000 quanta.
• With the fix below: the same widened setup gives 0 early returns.

Harness (the check after wait() is the only assertion):

#include <audioapi/utils/ThreadPool.hpp>
#include <atomic>
#include <chrono>
#include <cstdio>
#include <thread>

int main() {
  audioapi::ThreadPool pool(4);
  std::atomic<int> done{0};
  long early = 0;
  for (long q = 0; q < 3000; ++q) {
    done.store(0);
    for (int i = 0; i < 2; ++i)
      pool.schedule([&done] {
        std::this_thread::sleep_for(std::chrono::microseconds(1000));
        done.fetch_add(1);
      });
    if (q % 3 == 0) std::this_thread::sleep_for(std::chrono::microseconds(3000));  // scheduler preempted
    pool.wait();
    if (done.load() != 2) { ++early; while (done.load() != 2) {} }
  }
  std::printf("early_returns=%ld\n", early);
}

For the widened run, add this inside the bound task's if (left == 0) branch, before store(false):
if (counter++ % 7 == 0) std::this_thread::sleep_for(std::chrono::microseconds(2500));

Suggested fix: wait on the counter itself, so a late wake-up from a previous batch can't release the waiter:

// in schedule()'s bound task, after f(...)
if (cntrl->tasksScheduled.fetch_sub(1, std::memory_order_acq_rel) == 1) {
  cntrl->tasksScheduled.notify_all();
}

void wait() {
  size_t pending;
  while ((pending = controlBlock_->tasksScheduled.load(std::memory_order_acquire)) != 0) {
    controlBlock_->tasksScheduled.wait(pending, std::memory_order_acquire);
  }
}

With this, waitingForTasks is no longer needed. As a second safeguard, performConvolution could capture processingBuffer by value (it's a shared_ptr), so a task that outlives processNode doesn't read a dead stack slot.

Snack or a link to a repository

No public repo. The harness above is self-contained and needs only ThreadPool.hpp, SpscChannel.hpp, MoveOnlyFunction.hpp and core/utils/Constants.h on the include path.

React Native Audio API version

0.13.2 (code is unchanged in 0.13.6 and on main)

React Native version

0.86.2 (Expo SDK 57)

Platforms

Android

Architecture

New Architecture (bridgeless)

Build type

Release

Device

Real device

Device model

Android 16 (API 36), arm64

Acknowledgements

Yes, I searched the existing issues first.