Add thread-local governing hardware thread ownership.

Author: insertinterestingnamehere

include/qt_threadpool.h

@@ -18,6 +18,6 @@ typedef enum {
 
 hw_pool_init_status hw_pool_init(uint32_t num_threads);
 void hw_pool_destroy();
-void hw_pool_run_on_all(qt_threadpool_func_type func, void *arg);
+void run_on_current_pool(qt_threadpool_func_type func, void *arg);
 
 #endif

src/threadpool.c

@@ -75,6 +75,8 @@ typedef struct {
 #endif
 } pool_header;
 
+_Thread_local pool_header *owned_pool;
+
 typedef struct {
   // 16 byte aligned to allow loading it in one atomic instruction
   // on architectures where that makes sense (most of them).
@@ -306,6 +308,7 @@ API_FUNC hw_pool_init_status hw_pool_init(uint32_t num_threads) {
 #ifdef QPOOL_USE_PTHREADS
   pthread_attr_destroy(&attr);
 #endif
+  owned_pool = &hw_pool;
   return POOL_INIT_SUCCESS;
 cleanup_threads:
   if (i) {
@@ -348,6 +351,7 @@ API_FUNC hw_pool_init_status hw_pool_init(uint32_t num_threads) {
 }
 
 API_FUNC __attribute__((no_sanitize("memory"))) void hw_pool_destroy() {
+  owned_pool = NULL;
   uint32_t num_threads =
     atomic_load_explicit(&hw_pool.num_threads, memory_order_relaxed);
   char *buffer = atomic_load_explicit(&hw_pool.threads, memory_order_relaxed);
@@ -400,6 +404,10 @@ pool_run_on_all(pool_header *pool, qt_threadpool_func_type func, void *arg) {
   suspend_main_while_working(pool);
 }
 
+API_FUNC void run_on_current_pool(qt_threadpool_func_type func, void *arg) {
+  pool_run_on_all(owned_pool, func, arg);
+}
+
 API_FUNC void hw_pool_run_on_all(qt_threadpool_func_type func, void *arg) {
   pool_run_on_all(&hw_pool, func, arg);
 }

test/internal/threadpool.c

@@ -10,7 +10,7 @@ int main() {
   hw_pool_init(2);
   hw_pool_destroy();
   hw_pool_init(2);
-  hw_pool_run_on_all(&on_thread_test, NULL);
+  run_on_current_pool(&on_thread_test, NULL);
   hw_pool_destroy();
   printf("exited successfully\n");
   fflush(stdout);