SIGSEGV (exit 139) on Node worker_threads teardown after require('node-api-dotnet/net10.0') - pthread TLS destructor use-after-free

gdb-repro-run.sh
gdb-worker-teardown-bt.txt
gdb-worker-teardown-fault-mapping.txt
package.json

Summary

Loading node-api-dotnet inside a Node.js worker_threads Worker and then tearing that worker down - either via worker.terminate() from the parent or a graceful process.exit(0) inside the worker - reliably crashes the entire process with a native SIGSEGV (exit code 139) on Linux.

The crash requires zero interop calls: merely require('node-api-dotnet/net10.0') on the worker thread is enough. The same teardown on the main thread is clean, and a worker that does not load node-api-dotnet is clean.

A native backtrace shows the fault is in glibc's per-thread TLS-destructor cleanup (__nptl_deallocate_tsd) as the worker thread exits: a pthread_key destructor function pointer is dangling (points to unmapped memory), so calling it segfaults.

This is distinct from the thread-safe-function teardown issue in PR #480: applying that patch does not fix this crash, and neither does removing the JSThreadSafeFunction.Release() call entirely.

Environment

node-api-dotnet	0.9.21
.NET runtime	10.0.10 (Microsoft.NETCore.App), framework-dependent
Target framework moniker	net10.0
OS / libc	Linux x64, glibc (Debian 12 and Azure Linux 3.0 both affected)

Node.js version dependence (matches the PR #480 threshold)

Node.js	Result on worker teardown
22.17.0	SIGSEGV / exit 139
24.13.0	SIGSEGV / exit 139
24.18.1 (>= 24.14)	Deadlock / hang (process never exits)

Minimal reproduction

No Microsoft code required - just the npm package (npm install node-api-dotnet@0.9.21).

worker.cjs:

// Only load the host on the worker thread. No dotnet.require, no interop calls.
require('node-api-dotnet/net10.0');
require('worker_threads').parentPort.postMessage('ready');

main.cjs:

const { Worker } = require('worker_threads');
const w = new Worker(__dirname + '/worker.cjs');
w.on('message', async () => {
  await w.terminate();     // graceful process.exit(0) inside the worker also crashes
  process.exit(0);
});
w.on('exit', (code) => console.log('worker exit', code));

Run: node main.cjs ; echo "exit=$?" -> Observed exit=139.

Reproduce in Docker (Node 24.13.0)

docker build --build-arg NODE_VERSION=24.13.0 -t napd-repro:24.13.0 .
docker run --rm napd-repro:24.13.0
# -> v24.13.0 ... exit=139

Image = mcr.microsoft.com/dotnet/sdk:10.0 (provides the .NET 10 runtime) + Node.js 24.x. Building without NODE_VERSION installs the latest 24.x, which hangs instead of crashing.

Control cases (all clean)

• Main-thread teardown (even dotnet.require(<assembly>)) -> exit 0.
• Worker without node-api-dotnet -> exit 0.
• With interop objects (dotnet.require + managed objects) -> still 139 at teardown.

Native backtrace (gdb)

Thread N "node" received signal SIGSEGV, Segmentation fault.
#0  0x00007ffff4b3bbc0 in ?? ()                               <- dangling destructor pointer
#1  __GI___nptl_deallocate_tsd () at ./nptl/nptl_deallocate_tsd.c:73
#2  __GI___nptl_deallocate_tsd () at ./nptl/nptl_deallocate_tsd.c:22
#3  start_thread (arg=<optimized out>) at ./nptl/pthread_create.c:455
#4  clone3 () at ../sysdeps/unix/sysv/linux/x86_64/clone3.S:78

The faulting PC is not backed by any mapping (it sits in the guard gap above an 8 MB anonymous thread-stack region adjacent to the .NET Finalizer thread's stack). The value stored as the pthread_key destructor is garbage, so __nptl_deallocate_tsd jumps to unmapped memory. libcoreclr.so is mapped elsewhere, confirming the bad pointer does not point into the still-loaded CLR - it points at freed/reused memory.

Why PR #480 does not fix this

We built node-api-dotnet 0.9.21 with the PR #480 patch (verified the patched Microsoft.JavaScript.NodeApi.dll was actually loaded), and the crash still reproduces (139). Removing the _tsfn.Release() call entirely also still crashes. Consistent with the backtrace: the fault is in pthread TLS destructor dispatch, not the thread-safe-function release path.

What we're asking

1. Which pthread_key/TLS destructor is registered for CLR-hosting threads, and how its lifetime is ordered relative to napi environment (worker) teardown?
2. A fix (or supported guidance) so a Worker thread that loaded node-api-dotnet can be torn down without crashing the process.
3. If there is a supported pattern to safely detach the CLR from a Node worker thread before it exits (or keep the environment alive for the process lifetime), please document it.

Beginne damit, das angehängte gdb-repro-run.sh mit worker.cjs und main.cjs auszuführen, und untersuche anschließend gdb-worker-teardown-bt.txt und gdb-worker-teardown-fault-mapping.txt. Verfolge, welcher CLR-hosting-pthread-Key den Dangling Destructor registriert und wie seine Lebensdauer mit dem Teardown des Node-Workers zusammenhängt. Als abgeschlossen gilt die Aufgabe, wenn sowohl worker.terminate() als auch ein ordnungsgemäßes process.exit(0) in den gemeldeten Umgebungen ohne SIGSEGV oder Hängenbleiben abgeschlossen werden.