Move sleep slicing from musl's __wait to the lower level emscripten_futex_wait. NFC

sbc100 · sbc100 · commit 56c46cb1aef3 · 2026-03-18T16:25:40.000-07:00
This means we only need to do this breaking up on our `wait` operations
in a single place.  It also means that other users the
`emscripten_futex_wait` API don't break pthread proxying or async
cancellation.

The moved code is only included in pthread-enabled builds so should not
effect Wasm Workers builders.

This change also paves the way for enabling musl's `__wait` to work
with `WASM_WORKERS`.
diff --git a/system/lib/libc/musl/src/thread/__wait.c b/system/lib/libc/musl/src/thread/__wait.c
@@ -15,26 +15,11 @@ void __wait(volatile int *addr, volatile int *waiters, int val, int priv)
 	}
 	if (waiters) a_inc(waiters);
 #ifdef __EMSCRIPTEN__
-	int is_runtime_thread = emscripten_is_main_runtime_thread();
-
-	// Main runtime thread may need to run proxied calls, so sleep in very small slices to be responsive.
-	double max_ms_slice_to_sleep = is_runtime_thread ? 1 : 100;
-
-	while (*addr==val) {
-		if (is_runtime_thread || pthread_self()->cancelasync == PTHREAD_CANCEL_ASYNCHRONOUS) {
-			int e;
-			do {
-				if (pthread_self()->cancel) {
-					if (waiters) a_dec(waiters);
-					return;
-				}
-				// Must wait in slices in case this thread is cancelled in between.
-				e = emscripten_futex_wait((void*)addr, val, max_ms_slice_to_sleep);
-			} while (e == -ETIMEDOUT);
-		} else {
-			// Can wait in one go.
-			emscripten_futex_wait((void*)addr, val, INFINITY);
-		}
+	// loop here to handler spurious wakeups from the underlying
+	// emscripten_futex_wait.
+	int ret = 0;
+	while (*addr==val && ret == 0) {
+		ret = emscripten_futex_wait((void*)addr, val, INFINITY);
 	}
 #else
 	while (*addr==val) {
diff --git a/system/lib/pthread/emscripten_futex_wait.c b/system/lib/pthread/emscripten_futex_wait.c
@@ -19,7 +19,7 @@ extern void* _emscripten_main_thread_futex;
 
 static int futex_wait_main_browser_thread(volatile void* addr,
                                           uint32_t val,
-                                          double timeout) {
+                                          double timeout, bool cancelable) {
   // Atomics.wait is not available in the main browser thread, so simulate it
   // via busy spinning. Only the main browser thread is allowed to call into
   // this function. It is not thread-safe to be called from any other thread.
@@ -45,6 +45,12 @@ static int futex_wait_main_browser_thread(volatile void* addr,
   assert(last_addr == 0);
 
   while (1) {
+#ifdef __EMSCRIPTEN_PTHREADS__
+    // We if we were cancelled
+    if (cancelable && pthread_self()->cancel) {
+      return -ETIMEDOUT;
+    }
+#endif
     // Check for a timeout.
     now = emscripten_get_now();
     if (now > end) {
@@ -119,48 +125,90 @@ int emscripten_futex_wait(volatile void *addr, uint32_t val, double max_wait_ms)
     return -EINVAL;
   }
 
-  // Pass 0 here, which means we don't have access to the current time in this
-  // function.  This tells _emscripten_yield to call emscripten_get_now if (and
-  // only if) it needs to know the time.
-  _emscripten_yield(0);
-
   int ret;
   emscripten_conditional_set_current_thread_status(EM_THREAD_STATUS_RUNNING, EM_THREAD_STATUS_WAITFUTEX);
 
+#ifdef __EMSCRIPTEN_PTHREADS__
+  bool cancelable = pthread_self()->cancelasync == PTHREAD_CANCEL_ASYNCHRONOUS;
+#else
+  bool cancelable = false;
+#endif
+
   // For the main browser thread and audio worklets we can't use
   // __builtin_wasm_memory_atomic_wait32 so we have busy wait instead.
   if (!_emscripten_thread_supports_atomics_wait()) {
-    ret = futex_wait_main_browser_thread(addr, val, max_wait_ms);
+    ret = futex_wait_main_browser_thread(addr, val, max_wait_ms, cancelable);
     emscripten_conditional_set_current_thread_status(EM_THREAD_STATUS_WAITFUTEX, EM_THREAD_STATUS_RUNNING);
     return ret;
   }
 
   // -1 (or any negative number) means wait indefinitely.
   int64_t max_wait_ns = ATOMICS_WAIT_DURATION_INFINITE;
   if (max_wait_ms != INFINITY) {
-    max_wait_ns = (int64_t)(max_wait_ms*1000*1000);
+    max_wait_ns = (int64_t)(max_wait_ms * 1e6);
   }
-#ifdef EMSCRIPTEN_DYNAMIC_LINKING
-  // After the main thread queues dlopen events, it checks if the target threads
-  // are sleeping.
-  // If `sleeping` is set then the main thread knows that event will be
-  // processed after the sleep (before any other user code).  In this case the
-  // main thread does not wait for any kind of response form the thread.
-  // If `sleeping` is not set then we know we should wait for the thread process
-  // the queue, either from the call here directly after setting `sleeping` to
-  // 1, or from another callsite (e.g. the one in `emscripten_yield`).
-  int is_runtime_thread = emscripten_is_main_runtime_thread();
-  if (!is_runtime_thread) {
-    __pthread_self()->sleeping = 1;
-    _emscripten_process_dlopen_queue();
+
+#ifdef __EMSCRIPTEN_PTHREADS__
+  // When building with pthread support there are two conditions under which we
+  // need to limit the amount of time we spend in atomic.wait.
+  // 1. We are the main runtime thread.  In this case we need to be able to
+  //    process proxied events from workers.  Note that this is not always
+  //    the same as being the main browser thread.  For example, when running
+  //    under node or when launching and emscripten-built program in a Web
+  //    Worker.  This this case we limit our wait slices to 1ms intervals.
+  // 2. When the current thread has async cancellation enabled.  In this case
+  //    we limit the wait duration to 100ms intervals.
+  int64_t wakeup_interval = 0;
+  bool is_runtime_thread = emscripten_is_main_runtime_thread();
+  if (is_runtime_thread) {
+    // If the current thread is the main runtime thread then only wait in 1ms slices.
+    wakeup_interval = 1 * 1e6;
+  }
+  else if (cancelable) {
+    // If the current thread is async cancellable then only wait in 100ms slices.
+    wakeup_interval = 100 * 1e6;
+  }
+
+  int64_t interations;
+  if (wakeup_interval) {
+    interations = max_wait_ns / wakeup_interval;
+    max_wait_ns = wakeup_interval;
   }
+
+  do {
 #endif
-  ret = __builtin_wasm_memory_atomic_wait32((int*)addr, val, max_wait_ns);
+    // Pass 0 here, which means we don't have access to the current time in this
+    // function.  This tells _emscripten_yield to call emscripten_get_now if (and
+    // only if) it needs to know the time.
+    _emscripten_yield(0);
+
 #ifdef EMSCRIPTEN_DYNAMIC_LINKING
-  if (!is_runtime_thread) {
-    __pthread_self()->sleeping = 0;
-    _emscripten_process_dlopen_queue();
-  }
+    // After the main thread queues dlopen events, it checks if the target threads
+    // are sleeping.
+    // If `sleeping` is set then the main thread knows that event will be
+    // processed after the sleep (before any other user code).  In this case the
+    // main thread does not wait for any kind of response form the thread.
+    // If `sleeping` is not set then we know we should wait for the thread process
+    // the queue, either from the call here directly after setting `sleeping` to
+    // 1, or from another callsite (e.g. the one in `emscripten_yield`).
+    if (!is_runtime_thread) {
+      __pthread_self()->sleeping = 1;
+      _emscripten_process_dlopen_queue();
+    }
+#endif
+    ret = __builtin_wasm_memory_atomic_wait32((int*)addr, val, max_wait_ns);
+#ifdef EMSCRIPTEN_DYNAMIC_LINKING
+    if (!is_runtime_thread) {
+      __pthread_self()->sleeping = 0;
+      _emscripten_process_dlopen_queue();
+    }
+#endif
+#ifdef __EMSCRIPTEN_PTHREADS__
+    if (cancelable && ret == ATOMICS_WAIT_TIMED_OUT && pthread_self()->cancel) {
+      // We were cancelled
+      break;
+    }
+  } while (wakeup_interval && ret == ATOMICS_WAIT_TIMED_OUT && (max_wait_ms == INFINITY || interations--));
 #endif
 
   emscripten_conditional_set_current_thread_status(EM_THREAD_STATUS_WAITFUTEX, EM_THREAD_STATUS_RUNNING);
diff --git a/system/lib/pthread/library_pthread.c b/system/lib/pthread/library_pthread.c
@@ -87,8 +87,10 @@ void emscripten_thread_sleep(double msecs) {
   // If we have less than this many msecs left to wait, busy spin that instead.
   double min_ms_slice_to_sleep = 0.1;
 
-  // runtime thread may need to run proxied calls, so sleep in very small slices to be responsive.
-  double max_ms_slice_to_sleep = emscripten_is_main_runtime_thread() ? 1 : 100;
+  // Break up sleeping so that we process proxied work at regular intervals.
+  // TODO(sbc): This should be remove and/or moved down into
+  // `emscripten_futex_wait`.
+  double max_ms_slice_to_sleep = 100;
 
   emscripten_conditional_set_current_thread_status(
     EM_THREAD_STATUS_RUNNING, EM_THREAD_STATUS_SLEEPING);
diff --git a/test/codesize/test_codesize_minimal_pthreads.json b/test/codesize/test_codesize_minimal_pthreads.json
@@ -1,10 +1,10 @@
 {
   "a.out.js": 7884,
   "a.out.js.gz": 3855,
-  "a.out.nodebug.wasm": 19707,
-  "a.out.nodebug.wasm.gz": 9122,
-  "total": 27591,
-  "total_gz": 12977,
+  "a.out.nodebug.wasm": 19775,
+  "a.out.nodebug.wasm.gz": 9202,
+  "total": 27659,
+  "total_gz": 13057,
   "sent": [
     "a (memory)",
     "b (emscripten_get_now)",
diff --git a/test/codesize/test_codesize_minimal_pthreads_memgrowth.json b/test/codesize/test_codesize_minimal_pthreads_memgrowth.json
@@ -1,10 +1,10 @@
 {
   "a.out.js": 8306,
   "a.out.js.gz": 4060,
-  "a.out.nodebug.wasm": 19708,
-  "a.out.nodebug.wasm.gz": 9123,
-  "total": 28014,
-  "total_gz": 13183,
+  "a.out.nodebug.wasm": 19776,
+  "a.out.nodebug.wasm.gz": 9203,
+  "total": 28082,
+  "total_gz": 13263,
   "sent": [
     "a (memory)",
     "b (emscripten_get_now)",
