1.0.1.1:
[sbcl.git] / src / runtime / thread.c
index 3614c80..2c3367c 100644 (file)
@@ -1,14 +1,36 @@
+/*
+ * This software is part of the SBCL system. See the README file for
+ * more information.
+ *
+ * This software is derived from the CMU CL system, which was
+ * written at Carnegie Mellon University and released into the
+ * public domain. The software is in the public domain and is
+ * provided with absolutely no warranty. See the COPYING and CREDITS
+ * files for more information.
+ */
+
+#include "sbcl.h"
+
 #include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
+#ifndef LISP_FEATURE_WIN32
 #include <sched.h>
+#endif
 #include <signal.h>
 #include <stddef.h>
 #include <errno.h>
 #include <sys/types.h>
+#ifndef LISP_FEATURE_WIN32
 #include <sys/wait.h>
+#endif
+
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+#include <mach/mach.h>
+#include <mach/mach_error.h>
+#include <mach/mach_types.h>
+#endif
 
-#include "sbcl.h"
 #include "runtime.h"
 #include "validate.h"           /* for CONTROL_STACK_SIZE etc */
 #include "alloc.h"
 #include "interr.h"             /* for lose() */
 #include "gc-internal.h"
 
+#ifdef LISP_FEATURE_WIN32
+/*
+ * Win32 doesn't have SIGSTKSZ, and we're not switching stacks anyway,
+ * so define it arbitrarily
+ */
+#define SIGSTKSZ 1024
+#endif
+
+#if defined(LISP_FEATURE_DARWIN) && defined(LISP_FEATURE_SB_THREAD)
+#define QUEUE_FREEABLE_THREAD_STACKS
+#endif
+
+#ifdef LISP_FEATURE_FREEBSD
+#define CREATE_CLEANUP_THREAD
+#define LOCK_CREATE_THREAD
+#endif
+
 #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
 
 struct freeable_stack {
+#ifdef QUEUE_FREEABLE_THREAD_STACKS
+    struct freeable_stack *next;
+#endif
     os_thread_t os_thread;
     os_vm_address_t stack;
 };
 
+
+#ifdef QUEUE_FREEABLE_THREAD_STACKS
+static struct freeable_stack * volatile freeable_stack_queue = 0;
+static int freeable_stack_count = 0;
+pthread_mutex_t freeable_stack_lock = PTHREAD_MUTEX_INITIALIZER;
+#else
 static struct freeable_stack * volatile freeable_stack = 0;
+#endif
 
 int dynamic_values_bytes=4096*sizeof(lispobj);  /* same for all threads */
 struct thread * volatile all_threads;
 extern struct interrupt_data * global_interrupt_data;
-extern int linux_no_threads_p;
 
 #ifdef LISP_FEATURE_SB_THREAD
-
 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
+#ifdef LOCK_CREATE_THREAD
+static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
+#endif
+#endif
 
 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
 #endif
 
-#endif
-
 static void
 link_thread(struct thread *th)
 {
-    th->os_thread=thread_self();
     if (all_threads) all_threads->prev=th;
     th->next=all_threads;
     th->prev=0;
     all_threads=th;
-    protect_control_stack_guard_page(1);
 }
 
 #ifdef LISP_FEATURE_SB_THREAD
@@ -82,6 +129,10 @@ initial_thread_trampoline(struct thread *th)
     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
     if(arch_os_thread_init(th)==0) return 1;
     link_thread(th);
+    th->os_thread=thread_self();
+#ifndef LISP_FEATURE_WIN32
+    protect_control_stack_guard_page(1);
+#endif
 
 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
     return call_into_lisp_first_time(function,args,0);
@@ -96,6 +147,100 @@ initial_thread_trampoline(struct thread *th)
 
 #ifdef LISP_FEATURE_SB_THREAD
 
+#ifdef QUEUE_FREEABLE_THREAD_STACKS
+
+static void
+queue_freeable_thread_stack(struct thread *thread_to_be_cleaned_up)
+{
+     if (thread_to_be_cleaned_up) {
+        pthread_mutex_lock(&freeable_stack_lock);
+        if (freeable_stack_queue) {
+            struct freeable_stack *new_freeable_stack = 0, *next;
+            next = freeable_stack_queue;
+            while (next->next) {
+                next = next->next;
+            }
+            new_freeable_stack = (struct freeable_stack *)
+                os_validate(0, sizeof(struct freeable_stack));
+            new_freeable_stack->next = NULL;
+            new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
+            new_freeable_stack->stack = (os_vm_address_t)
+                thread_to_be_cleaned_up->control_stack_start;
+            next->next = new_freeable_stack;
+            freeable_stack_count++;
+        } else {
+            struct freeable_stack *new_freeable_stack = 0;
+            new_freeable_stack = (struct freeable_stack *)
+                os_validate(0, sizeof(struct freeable_stack));
+            new_freeable_stack->next = NULL;
+            new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
+            new_freeable_stack->stack = (os_vm_address_t)
+                thread_to_be_cleaned_up->control_stack_start;
+            freeable_stack_queue = new_freeable_stack;
+            freeable_stack_count++;
+        }
+        pthread_mutex_unlock(&freeable_stack_lock);
+    }
+}
+
+#define FREEABLE_STACK_QUEUE_SIZE 4
+
+static void
+free_freeable_stacks() {
+    if (freeable_stack_queue && (freeable_stack_count > FREEABLE_STACK_QUEUE_SIZE)) {
+        struct freeable_stack* old;
+        pthread_mutex_lock(&freeable_stack_lock);
+        old = freeable_stack_queue;
+        freeable_stack_queue = old->next;
+        freeable_stack_count--;
+        gc_assert(pthread_join(old->os_thread, NULL) == 0);
+        FSHOW((stderr, "freeing thread %x stack\n", old->os_thread));
+        os_invalidate(old->stack, THREAD_STRUCT_SIZE);
+        os_invalidate((os_vm_address_t)old, sizeof(struct freeable_stack));
+        pthread_mutex_unlock(&freeable_stack_lock);
+    }
+}
+
+#elif defined(CREATE_CLEANUP_THREAD)
+static void *
+cleanup_thread(void *arg)
+{
+    struct freeable_stack *freeable = arg;
+    pthread_t self = pthread_self();
+
+    FSHOW((stderr, "/cleaner thread(%p): joining %p\n",
+           self, freeable->os_thread));
+    gc_assert(pthread_join(freeable->os_thread, NULL) == 0);
+    FSHOW((stderr, "/cleaner thread(%p): free stack %p\n",
+           self, freeable->stack));
+    os_invalidate(freeable->stack, THREAD_STRUCT_SIZE);
+    free(freeable);
+
+    pthread_detach(self);
+
+    return NULL;
+}
+
+static void
+create_cleanup_thread(struct thread *thread_to_be_cleaned_up)
+{
+    pthread_t thread;
+    int result;
+
+    if (thread_to_be_cleaned_up) {
+        struct freeable_stack *freeable =
+            malloc(sizeof(struct freeable_stack));
+        gc_assert(freeable != NULL);
+        freeable->os_thread = thread_to_be_cleaned_up->os_thread;
+        freeable->stack =
+            (os_vm_address_t) thread_to_be_cleaned_up->control_stack_start;
+        result = pthread_create(&thread, NULL, cleanup_thread, freeable);
+        gc_assert(result == 0);
+        sched_yield();
+    }
+}
+
+#else
 static void
 free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
 {
@@ -111,7 +256,7 @@ free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
         swap_lispobjs((lispobj *)(void *)&freeable_stack,
                       (lispobj)new_freeable_stack);
     if (new_freeable_stack) {
-        FSHOW((stderr,"/reaping %lu\n", new_freeable_stack->os_thread));
+        FSHOW((stderr,"/reaping %p\n", (void*) new_freeable_stack->os_thread));
         /* Under NPTL pthread_join really waits until the thread
          * exists and the stack can be safely freed. This is sadly not
          * mandated by the pthread spec. */
@@ -121,6 +266,7 @@ free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
                       sizeof(struct freeable_stack));
     }
 }
+#endif
 
 /* this is the first thing that runs in the child (which is why the
  * silly calling convention).  Basically it calls the user's requested
@@ -131,7 +277,8 @@ int
 new_thread_trampoline(struct thread *th)
 {
     lispobj function;
-    int result;
+    int result, lock_ret;
+
     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
     function = th->no_tls_value_marker;
     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
@@ -140,28 +287,58 @@ new_thread_trampoline(struct thread *th)
         lose("arch_os_thread_init failed\n");
     }
 
+    th->os_thread=thread_self();
+    protect_control_stack_guard_page(1);
     /* Since GC can only know about this thread from the all_threads
      * list and we're just adding this thread to it there is no danger
-     * of deadlocking even with SIG_STOP_FOR_GC blocked. */
-    pthread_mutex_lock(&all_threads_lock);
+     * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
+     * not). */
+    lock_ret = pthread_mutex_lock(&all_threads_lock);
+    gc_assert(lock_ret == 0);
     link_thread(th);
-    pthread_mutex_unlock(&all_threads_lock);
+    lock_ret = pthread_mutex_unlock(&all_threads_lock);
+    gc_assert(lock_ret == 0);
 
     result = funcall0(function);
+
+    /* Block GC */
+    block_blockable_signals();
     th->state=STATE_DEAD;
 
     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
      * thread, but since we are already dead it won't wait long. */
-    pthread_mutex_lock(&all_threads_lock);
+    lock_ret = pthread_mutex_lock(&all_threads_lock);
+    gc_assert(lock_ret == 0);
+
+    gc_alloc_update_page_tables(0, &th->alloc_region);
     unlink_thread(th);
     pthread_mutex_unlock(&all_threads_lock);
+    gc_assert(lock_ret == 0);
 
-    gc_alloc_update_page_tables(0, &th->alloc_region);
     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
     os_invalidate((os_vm_address_t)th->interrupt_data,
                   (sizeof (struct interrupt_data)));
+
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+    FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
+    mach_port_move_member(mach_task_self(),
+                          THREAD_STRUCT_TO_EXCEPTION_PORT(th),
+                          MACH_PORT_NULL);
+    mach_port_deallocate(mach_task_self(),
+                         THREAD_STRUCT_TO_EXCEPTION_PORT(th));
+    mach_port_destroy(mach_task_self(),
+                      THREAD_STRUCT_TO_EXCEPTION_PORT(th));
+#endif
+
+#ifdef QUEUE_FREEABLE_THREAD_STACKS
+    queue_freeable_thread_stack(th);
+#elif defined(CREATE_CLEANUP_THREAD)
+    create_cleanup_thread(th);
+#else
     free_thread_stack_later(th);
-    FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
+#endif
+
+    FSHOW((stderr,"/exiting thread %p\n", thread_self()));
     return result;
 }
 
@@ -187,7 +364,14 @@ create_thread_struct(lispobj initial_function) {
     union per_thread_data *per_thread;
     struct thread *th=0;        /*  subdue gcc */
     void *spaces=0;
+#ifdef LISP_FEATURE_SB_THREAD
+    int i;
+#endif
 
+#ifdef CREATE_CLEANUP_THREAD
+    /* Give a chance for cleanup threads to run. */
+    sched_yield();
+#endif
     /* may as well allocate all the spaces at once: it saves us from
      * having to decide what to do if only some of the allocations
      * succeed */
@@ -200,17 +384,14 @@ create_thread_struct(lispobj initial_function) {
          BINDING_STACK_SIZE+
          ALIEN_STACK_SIZE);
 
-    if(all_threads) {
-        memcpy(per_thread,arch_os_get_current_thread(),
-               dynamic_values_bytes);
-    } else {
 #ifdef LISP_FEATURE_SB_THREAD
-        int i;
-        for(i=0;i<(dynamic_values_bytes/sizeof(lispobj));i++)
-            per_thread->dynamic_values[i]=NO_TLS_VALUE_MARKER_WIDETAG;
+    for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
+        per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
+    if (all_threads == 0) {
         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
             SetSymbolValue
                 (FREE_TLS_INDEX,
+                 /* FIXME: should be MAX_INTERRUPTS -1 ? */
                  make_fixnum(MAX_INTERRUPTS+
                              sizeof(struct thread)/sizeof(lispobj)),
                  0);
@@ -226,12 +407,11 @@ create_thread_struct(lispobj initial_function) {
         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
-        STATIC_TLS_INIT(PSEUDO_ATOMIC_ATOMIC,pseudo_atomic_atomic);
-        STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted);
+        STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 #endif
 #undef STATIC_TLS_INIT
-#endif
     }
+#endif
 
     th=&per_thread->thread;
     th->control_stack_start = spaces;
@@ -251,8 +431,7 @@ create_thread_struct(lispobj initial_function) {
     th->alien_stack_pointer=((void *)th->alien_stack_start);
 #endif
 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
-    th->pseudo_atomic_interrupted=0;
-    th->pseudo_atomic_atomic=0;
+    th->pseudo_atomic_bits=0;
 #endif
 #ifdef LISP_FEATURE_GENCGC
     gc_set_region_empty(&th->alloc_region);
@@ -271,8 +450,7 @@ create_thread_struct(lispobj initial_function) {
 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
-    SetSymbolValue(PSEUDO_ATOMIC_ATOMIC,(lispobj)th->pseudo_atomic_atomic,th);
-    SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED,th->pseudo_atomic_interrupted,th);
+    SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 #else
     current_binding_stack_pointer=th->binding_stack_pointer;
     current_control_stack_pointer=th->control_stack_start;
@@ -296,14 +474,27 @@ create_thread_struct(lispobj initial_function) {
     }
     th->interrupt_data->pending_handler = 0;
     th->no_tls_value_marker=initial_function;
+
+    th->stepping = NIL;
     return th;
 }
 
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+mach_port_t setup_mach_exception_handling_thread();
+kern_return_t mach_thread_init(mach_port_t thread_exception_port);
+
+#endif
+
 void create_initial_thread(lispobj initial_function) {
     struct thread *th=create_thread_struct(initial_function);
     if(th) {
+#ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
+        kern_return_t ret;
+
+        setup_mach_exception_handling_thread();
+#endif
         initial_thread_trampoline(th); /* no return */
-    } else lose("can't create initial thread");
+    } else lose("can't create initial thread\n");
 }
 
 #ifdef LISP_FEATURE_SB_THREAD
@@ -320,6 +511,15 @@ boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
     pthread_attr_t attr;
     sigset_t newset,oldset;
     boolean r=1;
+    int retcode, initcode, sizecode, addrcode;
+
+    FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
+
+#ifdef LOCK_CREATE_THREAD
+    retcode = pthread_mutex_lock(&create_thread_lock);
+    gc_assert(retcode == 0);
+    FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
+#endif
     sigemptyset(&newset);
     /* Blocking deferrable signals is enough, no need to block
      * SIG_STOP_FOR_GC because the child process is not linked onto
@@ -327,13 +527,38 @@ boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
     sigaddset_deferrable(&newset);
     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 
-    if((pthread_attr_init(&attr)) ||
+#if defined(LISP_FEATURE_DARWIN)
+#define CONTROL_STACK_ADJUST 8192 /* darwin wants page-aligned stacks */
+#else
+#define CONTROL_STACK_ADJUST 16
+#endif
+
+    if((initcode = pthread_attr_init(&attr)) ||
+       /* FIXME: why do we even have this in the first place? */
        (pthread_attr_setstack(&attr,th->control_stack_start,
-                              THREAD_CONTROL_STACK_SIZE-16)) ||
-       (pthread_create
-        (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th)))
+                              THREAD_CONTROL_STACK_SIZE-CONTROL_STACK_ADJUST)) ||
+#undef CONTROL_STACK_ADJUST
+       (retcode = pthread_create
+        (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) {
+        FSHOW_SIGNAL((stderr, "init, size, addr = %d, %d, %d\n", initcode, sizecode, addrcode));
+        FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
+        FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n",
+                      THREAD_CONTROL_STACK_SIZE-16, PTHREAD_STACK_MIN));
+        if(retcode < 0) {
+            perror("create_os_thread");
+        }
         r=0;
+    }
+
+#ifdef QUEUE_FREEABLE_THREAD_STACKS
+    free_freeable_stacks();
+#endif
     thread_sigmask(SIG_SETMASK,&oldset,0);
+#ifdef LOCK_CREATE_THREAD
+    retcode = pthread_mutex_unlock(&create_thread_lock);
+    gc_assert(retcode == 0);
+    FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
+#endif
     return r;
 }
 
@@ -341,9 +566,11 @@ os_thread_t create_thread(lispobj initial_function) {
     struct thread *th;
     os_thread_t kid_tid;
 
-    if(linux_no_threads_p) return 0;
-
-    th=create_thread_struct(initial_function);
+    /* Assuming that a fresh thread struct has no lisp objects in it,
+     * linking it to all_threads can be left to the thread itself
+     * without fear of gc lossage. initial_function violates this
+     * assumption and must stay pinned until the child starts up. */
+    th = create_thread_struct(initial_function);
     if(th==0) return 0;
 
     if (create_os_thread(th,&kid_tid)) {
@@ -356,7 +583,8 @@ os_thread_t create_thread(lispobj initial_function) {
 
 /* Send the signo to os_thread, retry if the rt signal queue is
  * full. */
-static int kill_thread_safely(os_thread_t os_thread, int signo)
+int
+kill_thread_safely(os_thread_t os_thread, int signo)
 {
     int r;
     /* The man page does not mention EAGAIN as a valid return value
@@ -383,7 +611,7 @@ int signal_interrupt_thread(os_thread_t os_thread)
     } else if (status == ESRCH) {
         return -1;
     } else {
-        lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s",
+        lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n",
              os_thread, status, strerror(status));
     }
 }
@@ -400,24 +628,38 @@ int signal_interrupt_thread(os_thread_t os_thread)
 void gc_stop_the_world()
 {
     struct thread *p,*th=arch_os_get_current_thread();
-    int status;
+    int status, lock_ret;
+#ifdef LOCK_CREATE_THREAD
+    /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
+     * on FreeBSD. */
+    FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock, thread=%lu\n",
+                  th->os_thread));
+    lock_ret = pthread_mutex_lock(&create_thread_lock);
+    gc_assert(lock_ret == 0);
+    FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n",
+                  th->os_thread));
+#endif
     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
                   th->os_thread));
     /* keep threads from starting while the world is stopped. */
-    pthread_mutex_lock(&all_threads_lock); \
+    lock_ret = pthread_mutex_lock(&all_threads_lock);      \
+    gc_assert(lock_ret == 0);
+
     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
                   th->os_thread));
     /* stop all other threads by sending them SIG_STOP_FOR_GC */
     for(p=all_threads; p; p=p->next) {
+        gc_assert(p->os_thread != 0);
+        FSHOW_SIGNAL((stderr,"/gc_stop_the_world: p->state: %x\n", p->state));
         if((p!=th) && ((p->state==STATE_RUNNING))) {
-            FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %lu\n",
-                          p->os_thread));
+            FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n",
+                          p, p->os_thread));
             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
             if (status==ESRCH) {
                 /* This thread has exited. */
                 gc_assert(p->state==STATE_DEAD);
             } else if (status) {
-                lose("cannot send suspend thread=%lu: %d, %s",
+                lose("cannot send suspend thread=%lu: %d, %s\n",
                      p->os_thread,status,strerror(status));
             }
         }
@@ -425,12 +667,11 @@ void gc_stop_the_world()
     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
     /* wait for the running threads to stop or finish */
     for(p=all_threads;p;) {
-        gc_assert(p->os_thread!=0);
-        if((p==th) || (p->state==STATE_SUSPENDED) ||
-           (p->state==STATE_DEAD)) {
-            p=p->next;
-        } else {
+        FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p));
+        if((p!=th) && (p->state==STATE_RUNNING)) {
             sched_yield();
+        } else {
+            p=p->next;
         }
     }
     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
@@ -439,7 +680,7 @@ void gc_stop_the_world()
 void gc_start_the_world()
 {
     struct thread *p,*th=arch_os_get_current_thread();
-    int status;
+    int status, lock_ret;
     /* if a resumed thread creates a new thread before we're done with
      * this loop, the new thread will get consed on the front of
      * all_threads, but it won't have been stopped so won't need
@@ -455,9 +696,14 @@ void gc_start_the_world()
             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
                           p->os_thread));
             p->state=STATE_RUNNING;
+
+#if defined(SIG_RESUME_FROM_GC)
+            status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC);
+#else
             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
+#endif
             if (status) {
-                lose("cannot resume thread=%lu: %d, %s",
+                lose("cannot resume thread=%lu: %d, %s\n",
                      p->os_thread,status,strerror(status));
             }
         }
@@ -466,7 +712,14 @@ void gc_start_the_world()
      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
      * performance implications, but does away with the 'rt signal
      * queue full' problem. */
-    pthread_mutex_unlock(&all_threads_lock); \
+
+    lock_ret = pthread_mutex_unlock(&all_threads_lock);
+    gc_assert(lock_ret == 0);
+#ifdef LOCK_CREATE_THREAD
+    lock_ret = pthread_mutex_unlock(&create_thread_lock);
+    gc_assert(lock_ret == 0);
+#endif
+
     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 }
 #endif