WasmFS: ProxyWorker's 50 ms heartbeat outlives its thread and runs the freed queue on a pooled Worker (Assertion failed: pthread_self() in emscripten_proxy_execute_queue) #27858

emscripten::ProxyWorker (system/lib/wasmfs/thread_utils.h), which the OPFS and fetch backends use for their dedicated thread, starts a heartbeat on that thread: a 50 ms setInterval that calls emscripten_proxy_execute_queue() on the ProxyWorker's queue. Nothing stops it except ABORT.

~ProxyWorker() cancels and joins the thread, and then the queue is freed. But the thread's Worker goes back to the pool rather than being terminated, and its timer goes with it. So every 50 ms the Worker runs a freed queue: while it is idle, pthread_self() is null and, with assertions, emscripten_proxy_execute_queue() aborts; without assertions, or once the Worker hosts another thread, it reads freed memory. Backends are destroyed at exit with EXIT_RUNTIME, so a program with an OPFS or fetch backend can abort at exit this way.

Version of emscripten/emsdk:

emcc (Emscripten gcc/clang-like replacement + linker emulating GNU ld) 6.0.12-git (ce6d2aad8bbf1e508e5a6d112f986617198bdc0d)
clang version 24.0.0git (https:/github.com/llvm/llvm-project d3e9891b54da54deebe2a6df0eb67c8ec155fa65)

main at ce6d2aa with the toolchain from test/emsdk_version.txt (dbaebb43), Node.js v24.19.0, x86_64 Linux.

Test case (proxy_worker_exit.cpp, which includes the internal header as a backend does):

// A destroyed ProxyWorker must leave nothing running against its freed queue,
// both while its Worker is idle in the pool and once it hosts another thread.
// Each wait spans several of the thread's 50 ms heartbeats.

#include <assert.h>
#include <condition_variable>
#include <emscripten/emscripten.h>
#include <emscripten/eventloop.h>
#include <mutex>
#include <stdio.h>
#include <thread>

#include "thread_utils.h"

// A ProxyWorker can't be created on the main browser thread.
static void use_worker() {
  std::thread([] {
    emscripten::ProxyWorker worker;
    bool ran = false;
    worker([&] { ran = true; });
    assert(ran);
  }).join();
}

static void second_idle(void*) {
  printf("done\n");
  emscripten_force_exit(0);
}

static void first_idle(void*) {
  use_worker();
  emscripten_set_timeout(second_idle, 500, nullptr);
}

int main() {
  use_worker();
  emscripten_set_timeout(first_idle, 500, nullptr);
}

Failing command line in full:

$ em++ -pthread -sWASMFS -sPTHREAD_POOL_SIZE=2 -sEXIT_RUNTIME -I$EMSCRIPTEN/system/lib/wasmfs proxy_worker_exit.cpp -o t.js
$ node t.js
Aborted(Assertion failed: pthread_self(), at: /emsdk/emscripten/system/lib/pthread/proxying.c,113,emscripten_proxy_execute_queue)
worker sent an error! undefined:undefined: Aborted(Assertion failed: pthread_self(), at: /emsdk/emscripten/system/lib/pthread/proxying.c,113,emscripten_proxy_execute_queue)
node:internal/event_target:1136
  process.nextTick(() => { throw err; });
                           ^

RuntimeError: Aborted(Assertion failed: pthread_self(), at: /emsdk/emscripten/system/lib/pthread/proxying.c,113,emscripten_proxy_execute_queue)
    at abort (t.js:855:11)
    at ___assert_fail (t.js:1743:7)
    at wasm://wasm/0006d26e:wasm-function[189]:0x3aa3
    at t.js:887:12
    at Timeout._onTimeout (t.js:2262:13)
    at listOnTimeout (node:internal/timers:635:17)
    at process.processTimers (node:internal/timers:571:7)

Node.js v24.19.0
[exit status 1]

It aborts in 10 runs out of 10. At -O2 (no assertions) it prints done, since the freed queue is still readable and empty.

Where:

• The heartbeat (libwasmfs.js#L581-L591), started on the worker thread (thread_utils.h#L64).
• The destructor cancels and joins the thread; the queue member is destroyed after it (thread_utils.h#L91-L94).
• The exiting thread clears pthread_self() (pthread_create.c#L370), and its Worker goes back to the pool, not terminated (libpthread.js#L224-L236).
• The next tick fails here (proxying.c#L113).

I have a fix ready and will send a PR: the destructor stops the heartbeat on the thread, with a synchronous proxied call, before cancelling it.