Fix event-loop hang when dlopen'd Go/cgo lib coexists with WASM on Linux

src/jsc/bindings/BunProcess.cpp

@@ -386,6 +386,7 @@ extern "C" void Bun__unlink(const char*, size_t);
 
 extern "C" void CrashHandler__setDlOpenAction(const char* action);
 extern "C" bool Bun__VM__allowAddons(void* vm);
+extern "C" void Bun__repairJscGcSignalAfterDlopen();
 
 JSC_DEFINE_HOST_FUNCTION(Process_functionDlopen, (JSC::JSGlobalObject * globalObject_, JSC::CallFrame* callFrame))
 {
@@ -534,6 +535,8 @@ JSC_DEFINE_HOST_FUNCTION(Process_functionDlopen, (JSC::JSGlobalObject * globalOb
     CrashHandler__setDlOpenAction(utf8.data());
     void* handle = dlopen(utf8.data(), RTLD_LAZY);
     CrashHandler__setDlOpenAction(nullptr);
+    if (handle)
+        Bun__repairJscGcSignalAfterDlopen();
 
     tryToDeleteIfNecessary();
 #endif

src/sys/lib.rs

@@ -6091,7 +6091,12 @@ pub fn dlopen(filename: &ZStr, flags: i32) -> Option<*mut c_void> {
     {
         // SAFETY: filename is NUL-terminated.
         let p = unsafe { libc::dlopen(filename.as_ptr(), flags) };
-        if p.is_null() { None } else { Some(p) }
+        if !p.is_null() {
+            repair_jsc_gc_signal_after_dlopen();
+            Some(p)
+        } else {
+            None
+        }
     }
     #[cfg(windows)]
     {
@@ -6101,6 +6106,72 @@ pub fn dlopen(filename: &ZStr, flags: i32) -> Option<*mut c_void> {
         if p.is_null() { None } else { Some(p.cast()) }
     }
 }
+
+/// Undo `SA_ONSTACK` that a just-`dlopen`ed library may have slapped onto JSC's
+/// thread-suspend signal.
+///
+/// JavaScriptCore suspends threads for GC/code-cache flushes by delivering a
+/// per-thread signal (`SIGPWR` on Linux in Bun's WebKit fork; `SIGUSR1`
+/// elsewhere) and having the handler snapshot the interrupted thread's
+/// context. The handler's sanity check — "am I running on the tracked thread's
+/// stack?" — uses its own frame pointer and therefore only holds when the
+/// signal arrives on the normal stack.
+///
+/// Several language runtimes iterate every signal in `initsig()`-style hooks
+/// when loaded as a shared library and, for any signal whose handler they
+/// don't want to own, force `SA_ONSTACK` onto the inherited disposition so
+/// their own threads' synchronous faults stay on a managed alt stack. Go's
+/// cgo runtime is the canonical offender — `runtime.setsigstack` sets
+/// `SA_ONSTACK` on *every* signal whose handler is non-default, including
+/// ours — but any runtime that follows the same recipe (Mono, some JNI
+/// layouts, Rust async runtimes compiled as cdylibs) trips the same wire.
+///
+/// When that happens in a process that already has an alternate signal stack
+/// installed (ASAN's runtime, Bun's own crash handler, libbacktrace, …) the
+/// next JSC suspend delivers the signal onto the alt stack; the stack check
+/// fails on every attempt; `Thread::suspend()` retries forever and the event
+/// loop stops. See oven-sh/bun#31158.
+///
+/// Fixing this inside WTF (use the interrupted SP from ucontext instead of
+/// the handler's own SP) is the clean path and should go upstream too;
+/// clearing `SA_ONSTACK` here is the belt-and-suspenders patch so the
+/// current prebuilt WebKit doesn't deadlock in the meantime.
+#[cfg(unix)]
+pub fn repair_jsc_gc_signal_after_dlopen() {
+    // Only Linux reproduces this: upstream WTF uses SIGUSR1 everywhere but
+    // Bun remapped to SIGPWR on Linux so userland SIGUSR1 stays available.
+    // `SA_ONSTACK` is a POSIX concept so the clearing itself works anywhere,
+    // but the wrongness we're correcting (`SA_ONSTACK` on our GC signal, plus
+    // an alt stack) only lines up on Linux in practice.
+    #[cfg(target_os = "linux")]
+    {
+        const SIGPWR: i32 = 30;
+        // SAFETY: both `oldact` and `newact` are stack-local, properly sized,
+        // and the kernel writes/reads only their POD contents.
+        unsafe {
+            let mut oldact: libc::sigaction = core::mem::zeroed();
+            if libc::sigaction(SIGPWR, core::ptr::null(), &raw mut oldact) != 0 {
+                return;
+            }
+            if (oldact.sa_flags & libc::SA_ONSTACK) == 0 {
+                return;
+            }
+            let mut newact = oldact;
+            newact.sa_flags &= !libc::SA_ONSTACK;
+            let _ = libc::sigaction(SIGPWR, &raw const newact, core::ptr::null_mut());
+        }
+    }
+}
+
+#[cfg(not(unix))]
+pub fn repair_jsc_gc_signal_after_dlopen() {}
+
+/// C-ABI wrapper so call sites outside the Rust tree (BunProcess.cpp's
+/// `process.dlopen`) can use the same post-dlopen repair.
+#[unsafe(no_mangle)]
+pub extern "C" fn Bun__repairJscGcSignalAfterDlopen() {
+    repair_jsc_gc_signal_after_dlopen();
+}
 /// sys.zig:4565 — `dlsym(handle, name)`.
 pub fn dlsym_impl(handle: Option<*mut c_void>, name: &ZStr) -> Option<*mut c_void> {
     #[cfg(unix)]

test/regression/issue/31158.test.ts

@@ -0,0 +1,87 @@
+// Regression test for https://github.com/oven-sh/bun/issues/31158.
+//
+// When a shared library (e.g. Go's cgo runtime) force-sets `SA_ONSTACK` on
+// JavaScriptCore's thread-suspend signal (SIGPWR on Linux in Bun's WebKit
+// fork), every delivery of that signal on a thread that has an alternate
+// signal stack (ASAN runtime, Bun's crash handler, libbacktrace) runs the
+// handler on the alt stack. WTF's stack-check in `signalHandlerSuspendResume`
+// used `currentStackPointer()` — the handler's own SP — so the sanity
+// check failed every time, `Thread::suspend()` retried forever, and the
+// event loop stopped advancing on the next WASM compile/install.
+//
+// Triggers the exact shape of the bug without needing Go installed in CI:
+// force `SA_ONSTACK` onto SIGPWR via `bun:ffi`, then call `dlopen` once
+// more so the repair fires, then compile+call a WASM function (which
+// triggers `resetInstructionCacheOnAllThreads`). Without the fix,
+// `setTimeout` after the WASM call never fires and the child hangs.
+import { expect, test } from "bun:test";
+import { bunEnv, bunExe, isGlibc } from "harness";
+
+// The bug is only reachable on Linux: SIGPWR is only used as the suspend/
+// resume signal on Linux (see `vendor/WebKit/Source/WTF/wtf/posix/
+// ThreadingPOSIX.cpp` — USE(BUN_JSC_ADDITIONS) branch). The reproduction
+// dlopens `libc.so.6` directly so it needs glibc; musl names its libc
+// differently and static-musl builds of Bun don't even have a dynamic
+// loader in the room.
+test.skipIf(!isGlibc)("event loop survives SA_ONSTACK on SIGPWR + WASM (oven-sh/bun#31158)", async () => {
+  const script = /* ts */ `
+      import { dlopen, FFIType, ptr } from "bun:ffi";
+
+      // 1. Open libc so we can flip SA_ONSTACK on SIGPWR the same way Go's
+      //    cgo 'initsig' does on every inherited handler.
+      const libc = dlopen("libc.so.6", {
+        sigaction: { args: [FFIType.i32, FFIType.ptr, FFIType.ptr], returns: FFIType.i32 },
+      });
+
+      // x86_64 glibc layout: sa_handler (8), sa_mask (128), sa_flags (4), sa_restorer (8).
+      // We only touch sa_flags so this layout holds on aarch64 too (same offsets).
+      const SIGACTION_SIZE = 152;
+      const FLAGS_OFFSET = 8 + 128;
+      const SIGPWR = 30;
+      const SA_ONSTACK = 0x08000000;
+
+      const buf = new ArrayBuffer(SIGACTION_SIZE);
+      const view = new DataView(buf);
+
+      libc.symbols.sigaction(SIGPWR, null, ptr(buf));
+      view.setInt32(FLAGS_OFFSET, view.getInt32(FLAGS_OFFSET, true) | SA_ONSTACK, true);
+      libc.symbols.sigaction(SIGPWR, ptr(buf), null);
+
+      // 2. Trigger a fresh dlopen — the fix clears SA_ONSTACK here.
+      dlopen("libc.so.6", { getpid: { args: [], returns: FFIType.i32 } });
+
+      // 3. Exercise the WASM path that calls resetInstructionCacheOnAllThreads.
+      //    The bytes below are a minimal WASM module with one exported function
+      //    that loops 10_000 times and returns — enough to trigger compilation.
+      const bytes = new Uint8Array([
+        0, 97, 115, 109, 1, 0, 0, 0, 1, 5, 1, 96, 0, 1, 127, 3, 2, 1, 0, 7, 5,
+        1, 1, 102, 0, 0, 10, 34, 1, 32, 1, 1, 127, 65, 0, 33, 0, 2, 64, 3, 64,
+        32, 0, 65, 1, 106, 33, 0, 32, 0, 65, 144, 206, 0, 72, 13, 0, 11, 11, 32,
+        0, 11,
+      ]);
+      const inst = new WebAssembly.Instance(new WebAssembly.Module(bytes));
+      (inst.exports.f as () => number)();
+
+      // 4. A setTimeout that must fire — the bug makes this hang forever.
+      await new Promise(r => setTimeout(r, 10));
+      console.log("EVENT_LOOP_ALIVE");
+    `;
+
+  await using proc = Bun.spawn({
+    cmd: [bunExe(), "-e", script],
+    env: bunEnv,
+    stdout: "pipe",
+    stderr: "pipe",
+  });
+
+  // Kill the child after 10 s so a regression shows up as "hung, no output"
+  // rather than a stuck test harness.
+  const watchdog = setTimeout(() => proc.kill("SIGKILL"), 10_000);
+
+  const [stdout, stderr, exitCode] = await Promise.all([proc.stdout.text(), proc.stderr.text(), proc.exited]);
+  clearTimeout(watchdog);
+
+  expect(stderr).toBe("");
+  expect(stdout).toContain("EVENT_LOOP_ALIVE");
+  expect(exitCode).toBe(0);
+});