write-barrier handling with multiple threads

svn: r16880

src/mzscheme/gc2/newgc.c

@@ -1698,8 +1698,13 @@ static int designate_modified_gc(NewGC *gc, void *p)
       page->mprotected = 0;
       vm_protect_pages(page->addr, (page->size_class > 1) ? round_to_apage_size(page->size) : APAGE_SIZE, 1);
       page->back_pointers = 1;
-      return 1;
     }
+    /* For a single mutator thread, we shouldn't get here
+       (and a `return 1' in the braces above would make more
+       sense). With multiple mutators, though, two threads might
+       hit the same page at effectively the same time, and only
+       the first one of them will handle the signal. */
+    return 1;
   } else {
     if (gc->primoridal_gc) {
       return designate_modified_gc(gc->primoridal_gc, p);

src/mzscheme/src/future.c

@@ -50,7 +50,7 @@ void scheme_init_futures(Scheme_Env *env)
 
 extern void *on_demand_jit_code;
 
-#define THREAD_POOL_SIZE 3
+#define THREAD_POOL_SIZE 1
 #define INITIAL_C_STACK_SIZE 500000
 static pthread_t g_pool_threads[THREAD_POOL_SIZE];
 static int *g_fuel_pointers[THREAD_POOL_SIZE];
@@ -59,6 +59,7 @@ static int g_num_avail_threads = 0;
 static unsigned long g_cur_cpu_mask = 1;
 static void *g_signal_handle = NULL;
 
+static struct NewGC *g_shared_GC;
 future_t *g_future_queue = NULL;
 int g_next_futureid = 0;
 pthread_t g_rt_threadid = 0;
@@ -86,7 +87,7 @@ static void register_traversers(void);
 extern void scheme_on_demand_generate_lambda(Scheme_Native_Closure *nc, int argc, Scheme_Object **argv);
 
 static void start_gc_not_ok();
-static void end_gc_not_ok();
+static void end_gc_not_ok(future_t *ft);
 
 THREAD_LOCAL_DECL(static future_t *current_ft);
 
@@ -310,6 +311,7 @@ void futures_init(void)
   for (i = 0; i < THREAD_POOL_SIZE; i++)
   {
       gc_counter_ptr = &scheme_did_gc_count;
+      g_shared_GC = GC;
       pthread_create(&threadid, &attr, worker_thread_future_loop, &i);
       sema_wait(&ready_sema);
 	
@@ -337,8 +339,14 @@ static void start_gc_not_ok()
 #endif
 }
 
-static void end_gc_not_ok()
+static void end_gc_not_ok(future_t *ft)
 {
+  if (ft) {
+    scheme_set_runstack_limits(ft->runstack_start, 
+                               ft->runstack_size,
+                               ft->runstack - ft->runstack_start,
+                               ft->runstack_size);
+  }
   pthread_mutex_lock(&gc_ok_m);
   --gc_not_ok;
   pthread_cond_signal(&gc_ok_c);
@@ -465,6 +473,7 @@ Scheme_Object *future(int argc, Scheme_Object *argv[])
 			rs = rs_start XFORM_OK_PLUS init_runstack_size;
 			ft->runstack_start = rs_start;
     	ft->runstack = rs;
+    	ft->runstack_size = init_runstack_size;
 		}   
 
     //pthread_mutex_unlock(&g_future_queue_mutex);
@@ -603,6 +612,8 @@ void *worker_thread_future_loop(void *arg)
 
         scheme_init_os_thread();
 
+        GC = g_shared_GC;
+
 	//Set processor affinity
 	/*pthread_mutex_lock(&g_future_queue_mutex);
 	if (pthread_setaffinity_np(pthread_self(), sizeof(g_cur_cpu_mask), &g_cur_cpu_mask))
@@ -635,7 +646,7 @@ void *worker_thread_future_loop(void *arg)
         pthread_mutex_lock(&g_future_queue_mutex);
 				while (!(ft = get_pending_future()))
 				{
-          end_gc_not_ok();
+          end_gc_not_ok(NULL);
 					pthread_cond_wait(&g_future_pending_cv, &g_future_queue_mutex);
           start_gc_not_ok();
 				}
@@ -679,12 +690,15 @@ void *worker_thread_future_loop(void *arg)
         ft->work_completed = 1;
         ft->retval = v;
 
+        ft->runstack = NULL;
+        ft->runstack_start = NULL;
+
 				//Update the status 
 				ft->status = FINISHED;
 				scheme_signal_received_at(g_signal_handle);
         pthread_mutex_unlock(&g_future_queue_mutex);
 
-        end_gc_not_ok();
+        end_gc_not_ok(NULL);
 
         goto wait_for_work;
 
@@ -737,7 +751,7 @@ int future_do_runtimecall(
 
     //Wait for the signal that the RT call is finished
                 future->can_continue_cv = &worker_can_continue_cv;
-    end_gc_not_ok();
+    end_gc_not_ok(future);
     pthread_cond_wait(&worker_can_continue_cv, &g_future_queue_mutex);
     start_gc_not_ok();
 

src/mzscheme/src/future.h

@@ -80,6 +80,7 @@ typedef struct future {
 	int work_completed;
 	pthread_cond_t *can_continue_cv;
 
+        long runstack_size;
 	Scheme_Object **runstack;
 	Scheme_Object **runstack_start;
 	Scheme_Object *orig_lambda;

src/mzscheme/src/salloc.c

@@ -65,6 +65,9 @@ THREAD_LOCAL Thread_Local_Variables scheme_thread_locals;
 extern int scheme_num_copied_stacks;
 static unsigned long scheme_primordial_os_thread_stack_base;
 THREAD_LOCAL_DECL(static unsigned long scheme_os_thread_stack_base);
+#ifdef USE_THREAD_LOCAL
+Thread_Local_Variables *scheme_vars; /* for debugging */
+#endif
 
 static Scheme_Report_Out_Of_Memory_Proc more_report_out_of_memory;
 
@@ -164,7 +167,12 @@ int scheme_main_stack_setup(int no_auto_statics, Scheme_Nested_Main _main, void
     fprintf(stderr, "pthread key create failed");
     abort();
   }
+#endif
+
   scheme_init_os_thread();
+
+#ifdef USE_THREAD_LOCAL
+  scheme_vars = scheme_get_thread_local_variables();
 #endif
 
   scheme_set_stack_base(PROMPT_STACK(stack_start), no_auto_statics);