src/runtime/thread.c

   1 #include <stdlib.h>
   2 #include <stdio.h>
   3 #include <string.h>
   4 #include <sched.h>
   5 #include <signal.h>
   6 #include <stddef.h>
   7 #include <errno.h>
   8 #include <sys/types.h>
   9 #include <sys/wait.h>
  10
  11 #include "sbcl.h"
  12 #include "runtime.h"
  13 #include "validate.h"           /* for CONTROL_STACK_SIZE etc */
  14 #include "alloc.h"
  15 #include "thread.h"
  16 #include "arch.h"
  17 #include "target-arch-os.h"
  18 #include "os.h"
  19 #include "globals.h"
  20 #include "dynbind.h"
  21 #include "genesis/cons.h"
  22 #include "genesis/fdefn.h"
  23 #include "interr.h"             /* for lose() */
  24 #include "gc-internal.h"
  25
  26 #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
  27
  28 int dynamic_values_bytes=4096*sizeof(lispobj);  /* same for all threads */
  29 struct thread * volatile all_threads;
  30 volatile lispobj all_threads_lock;
  31 extern struct interrupt_data * global_interrupt_data;
  32 extern int linux_no_threads_p;
  33
  34 #ifdef LISP_FEATURE_SB_THREAD
  35 /* When trying to get all_threads_lock one should make sure that
  36  * sig_stop_for_gc is not blocked. Else there would be a possible
  37  * deadlock: gc locks it, other thread blocks signals, gc sends stop
  38  * request to other thread and waits, other thread blocks on lock. */
  39 void check_sig_stop_for_gc_can_arrive_or_lose()
  40 {
  41     /* Get the current sigmask, by blocking the empty set. */
  42     sigset_t empty,current;
  43     sigemptyset(&empty);
  44     thread_sigmask(SIG_BLOCK, &empty, &current);
  45     if (sigismember(&current,SIG_STOP_FOR_GC))
  46         lose("SIG_STOP_FOR_GC cannot arrive: it is blocked\n");
  47     if (SymbolValue(GC_INHIBIT,arch_os_get_current_thread()) != NIL)
  48         lose("SIG_STOP_FOR_GC cannot arrive: gc is inhibited\n");
  49     if (arch_pseudo_atomic_atomic(NULL))
  50         lose("SIG_STOP_FOR_GC cannot arrive: in pseudo atomic\n");
  51 }
  52
  53 #define GET_ALL_THREADS_LOCK(name) \
  54     { \
  55         sigset_t _newset,_oldset; \
  56         sigemptyset(&_newset); \
  57         sigaddset_deferrable(&_newset); \
  58         thread_sigmask(SIG_BLOCK, &_newset, &_oldset); \
  59         check_sig_stop_for_gc_can_arrive_or_lose(); \
  60         FSHOW_SIGNAL((stderr,"/%s:waiting on lock=%ld, thread=%lu\n",name, \
  61                all_threads_lock,arch_os_get_current_thread()->os_thread)); \
  62         get_spinlock(&all_threads_lock,(long)arch_os_get_current_thread()); \
  63         FSHOW_SIGNAL((stderr,"/%s:got lock, thread=%lu\n", \
  64                name,arch_os_get_current_thread()->os_thread));
  65
  66 #define RELEASE_ALL_THREADS_LOCK(name) \
  67         FSHOW_SIGNAL((stderr,"/%s:released lock\n",name)); \
  68         release_spinlock(&all_threads_lock); \
  69         thread_sigmask(SIG_SETMASK,&_oldset,0); \
  70     }
  71 #endif
  72
  73
  74 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  75 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
  76 #endif
  77
  78 static int
  79 initial_thread_trampoline(struct thread *th)
  80 {
  81     lispobj function;
  82 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  83     lispobj *args = NULL;
  84 #endif
  85     function = th->unbound_marker;
  86     th->unbound_marker = UNBOUND_MARKER_WIDETAG;
  87     if(arch_os_thread_init(th)==0) return 1;
  88
  89     if(th->os_thread < 1) lose("th->os_thread not set up right");
  90     th->state=STATE_RUNNING;
  91 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  92     return call_into_lisp_first_time(function,args,0);
  93 #else
  94     return funcall0(function);
  95 #endif
  96 }
  97
  98 #ifdef LISP_FEATURE_SB_THREAD
  99
 100 /* this is the first thing that runs in the child (which is why the
 101  * silly calling convention).  Basically it calls the user's requested
 102  * lisp function after doing arch_os_thread_init and whatever other
 103  * bookkeeping needs to be done
 104  */
 105 int
 106 new_thread_trampoline(struct thread *th)
 107 {
 108     lispobj function;
 109     int result;
 110     function = th->unbound_marker;
 111     th->unbound_marker = UNBOUND_MARKER_WIDETAG;
 112     if(arch_os_thread_init(th)==0) {
 113         /* FIXME: handle error */
 114         lose("arch_os_thread_init failed\n");
 115     }
 116
 117     /* wait here until our thread is linked into all_threads: see below */
 118     {
 119         volatile os_thread_t *tid=&th->os_thread;
 120         while(*tid<1) sched_yield();
 121     }
 122
 123     th->state=STATE_RUNNING;
 124     result = funcall0(function);
 125     th->state=STATE_DEAD;
 126     return result;
 127 }
 128 #endif /* LISP_FEATURE_SB_THREAD */
 129
 130 /* this is called from any other thread to create the new one, and
 131  * initialize all parts of it that can be initialized from another
 132  * thread
 133  */
 134
 135 struct thread * create_thread_struct(lispobj initial_function) {
 136     union per_thread_data *per_thread;
 137     struct thread *th=0;        /*  subdue gcc */
 138     void *spaces=0;
 139
 140     /* may as well allocate all the spaces at once: it saves us from
 141      * having to decide what to do if only some of the allocations
 142      * succeed */
 143     spaces=os_validate(0,
 144                        THREAD_CONTROL_STACK_SIZE+
 145                        BINDING_STACK_SIZE+
 146                        ALIEN_STACK_SIZE+
 147                        dynamic_values_bytes+
 148                        32*SIGSTKSZ);
 149     if(!spaces)
 150          return NULL;
 151     per_thread=(union per_thread_data *)
 152         (spaces+
 153          THREAD_CONTROL_STACK_SIZE+
 154          BINDING_STACK_SIZE+
 155          ALIEN_STACK_SIZE);
 156
 157     if(all_threads) {
 158         memcpy(per_thread,arch_os_get_current_thread(),
 159                dynamic_values_bytes);
 160     } else {
 161 #ifdef LISP_FEATURE_SB_THREAD
 162         int i;
 163         for(i=0;i<(dynamic_values_bytes/sizeof(lispobj));i++)
 164             per_thread->dynamic_values[i]=UNBOUND_MARKER_WIDETAG;
 165         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG)
 166             SetSymbolValue
 167                 (FREE_TLS_INDEX,
 168                  make_fixnum(MAX_INTERRUPTS+
 169                              sizeof(struct thread)/sizeof(lispobj)),
 170                  0);
 171 #define STATIC_TLS_INIT(sym,field) \
 172   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 173   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 174
 175         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 176         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 177         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 178         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 179         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 180 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 181         STATIC_TLS_INIT(PSEUDO_ATOMIC_ATOMIC,pseudo_atomic_atomic);
 182         STATIC_TLS_INIT(PSEUDO_ATOMIC_INTERRUPTED,pseudo_atomic_interrupted);
 183 #endif
 184 #undef STATIC_TLS_INIT
 185 #endif
 186     }
 187
 188     th=&per_thread->thread;
 189     th->control_stack_start = spaces;
 190     th->binding_stack_start=
 191         (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE);
 192     th->control_stack_end = th->binding_stack_start;
 193     th->alien_stack_start=
 194         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 195     th->binding_stack_pointer=th->binding_stack_start;
 196     th->this=th;
 197     th->os_thread=0;
 198     th->interrupt_fun=NIL;
 199     th->interrupt_fun_lock=0;
 200     th->state=STATE_STARTING;
 201 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 202     th->alien_stack_pointer=((void *)th->alien_stack_start
 203                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 204 #else
 205     th->alien_stack_pointer=((void *)th->alien_stack_start);
 206 #endif
 207 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 208     th->pseudo_atomic_interrupted=0;
 209     th->pseudo_atomic_atomic=0;
 210 #endif
 211 #ifdef LISP_FEATURE_GENCGC
 212     gc_set_region_empty(&th->alloc_region);
 213 #endif
 214
 215 #ifndef LISP_FEATURE_SB_THREAD
 216     /* the tls-points-into-struct-thread trick is only good for threaded
 217      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 218      * appropriate values from struct thread here, and make sure that
 219      * we use the appropriate SymbolValue macros to access any of the
 220      * variable quantities from the C runtime.  It's not quite OAOOM,
 221      * it just feels like it */
 222     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 223     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 224     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 225 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 226     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 227     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 228     SetSymbolValue(PSEUDO_ATOMIC_ATOMIC,(lispobj)th->pseudo_atomic_atomic,th);
 229     SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED,th->pseudo_atomic_interrupted,th);
 230 #else
 231     current_binding_stack_pointer=th->binding_stack_pointer;
 232     current_control_stack_pointer=th->control_stack_start;
 233 #endif
 234 #endif
 235     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 236     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 237     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 238     bind_variable(INTERRUPT_PENDING, NIL,th);
 239     bind_variable(INTERRUPTS_ENABLED,T,th);
 240     bind_variable(GC_PENDING,NIL,th);
 241 #ifdef LISP_FEATURE_SB_THREAD
 242     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 243 #endif
 244
 245     th->interrupt_data = (struct interrupt_data *)
 246         os_validate(0,(sizeof (struct interrupt_data)));
 247     if(all_threads)
 248         memcpy(th->interrupt_data,
 249                arch_os_get_current_thread()->interrupt_data,
 250                sizeof (struct interrupt_data));
 251     else
 252         memcpy(th->interrupt_data,global_interrupt_data,
 253                sizeof (struct interrupt_data));
 254
 255     th->unbound_marker=initial_function;
 256     return th;
 257 }
 258
 259 static void
 260 link_thread(struct thread *th,os_thread_t kid_tid)
 261 {
 262     if (all_threads) all_threads->prev=th;
 263     th->next=all_threads;
 264     th->prev=0;
 265     all_threads=th;
 266     /* note that th->os_thread is 0 at this time.  We rely on
 267      * all_threads_lock to ensure that we don't have >1 thread with
 268      * os_thread=0 on the list at once
 269      */
 270     protect_control_stack_guard_page(th,1);
 271     /* child will not start until this is set */
 272     th->os_thread=kid_tid;
 273     FSHOW((stderr,"/created thread %lu\n",kid_tid));
 274 }
 275
 276 void create_initial_thread(lispobj initial_function) {
 277     struct thread *th=create_thread_struct(initial_function);
 278     os_thread_t kid_tid=thread_self();
 279     if(th && kid_tid>0) {
 280         link_thread(th,kid_tid);
 281         initial_thread_trampoline(all_threads); /* no return */
 282     } else lose("can't create initial thread");
 283 }
 284
 285 #ifdef LISP_FEATURE_SB_THREAD
 286
 287 #ifndef __USE_XOPEN2K
 288 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 289                                   size_t __stacksize);
 290 #endif
 291
 292 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 293 {
 294     /* The new thread inherits the restrictive signal mask set here,
 295      * and enables signals again when it is set up properly. */
 296     pthread_attr_t attr;
 297     sigset_t newset,oldset;
 298     boolean r=1;
 299     sigemptyset(&newset);
 300     /* Blocking deferrable signals is enough, since gc_stop_the_world
 301      * waits until the child leaves STATE_STARTING. And why not let gc
 302      * proceed as soon as possible? */
 303     sigaddset_deferrable(&newset);
 304     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 305
 306     if((pthread_attr_init(&attr)) ||
 307        (pthread_attr_setstack(&attr,th->control_stack_start,
 308                               THREAD_CONTROL_STACK_SIZE-16)) ||
 309        (pthread_create
 310         (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th)))
 311         r=0;
 312     thread_sigmask(SIG_SETMASK,&oldset,0);
 313     return r;
 314 }
 315
 316 struct thread *create_thread(lispobj initial_function) {
 317     struct thread *th;
 318     os_thread_t kid_tid=0;
 319     boolean success;
 320
 321     if(linux_no_threads_p) return 0;
 322
 323     th=create_thread_struct(initial_function);
 324     if(th==0) return 0;
 325
 326     /* we must not be interrupted here after a successful
 327      * create_os_thread, because the kid will be waiting for its
 328      * thread struct to be linked */
 329     GET_ALL_THREADS_LOCK("create_thread")
 330
 331     success=create_os_thread(th,&kid_tid);
 332     if (success)
 333         link_thread(th,kid_tid);
 334     else
 335         os_invalidate((os_vm_address_t) th->control_stack_start,
 336                       ((sizeof (lispobj))
 337                        * (th->control_stack_end-th->control_stack_start)) +
 338                       BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+
 339                       32*SIGSTKSZ);
 340
 341     RELEASE_ALL_THREADS_LOCK("create_thread")
 342
 343     if (success)
 344         return th;
 345     else
 346         return 0;
 347 }
 348
 349 /* called from lisp from the thread object finalizer */
 350 void reap_dead_thread(struct thread *th)
 351 {
 352     if(th->state!=STATE_DEAD)
 353         lose("thread %p is not joinable, state=%d\n",th,th->state);
 354 #ifdef LISP_FEATURE_GENCGC
 355     {
 356         sigset_t newset,oldset;
 357         sigemptyset(&newset);
 358         sigaddset_blockable(&newset);
 359         thread_sigmask(SIG_BLOCK, &newset, &oldset);
 360         gc_alloc_update_page_tables(0, &th->alloc_region);
 361         release_spinlock(&all_threads_lock);
 362         thread_sigmask(SIG_SETMASK,&oldset,0);
 363     }
 364 #endif
 365     GET_ALL_THREADS_LOCK("reap_dead_thread")
 366     FSHOW((stderr,"/reap_dead_thread: reaping %lu\n",th->os_thread));
 367     if(th->prev)
 368         th->prev->next=th->next;
 369     else all_threads=th->next;
 370     if(th->next)
 371         th->next->prev=th->prev;
 372     RELEASE_ALL_THREADS_LOCK("reap_dead_thread")
 373     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 374     gc_assert(pthread_join(th->os_thread,NULL)==0);
 375     os_invalidate((os_vm_address_t) th->control_stack_start,
 376                   ((sizeof (lispobj))
 377                    * (th->control_stack_end-th->control_stack_start)) +
 378                   BINDING_STACK_SIZE+ALIEN_STACK_SIZE+dynamic_values_bytes+
 379                   32*SIGSTKSZ);
 380 }
 381
 382 /* Send the signo to os_thread, retry if the rt signal queue is
 383  * full. */
 384 static int kill_thread_safely(os_thread_t os_thread, int signo)
 385 {
 386     int r;
 387     /* The man page does not mention EAGAIN as a valid return value
 388      * for either pthread_kill or kill. But that's theory, this is
 389      * practice. By waiting here we assume that the delivery of this
 390      * signal is not necessary for the delivery of the signals in the
 391      * queue. In other words, we _assume_ there are no deadlocks. */
 392     while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
 393         /* wait a bit then try again in the hope of the rt signal
 394          * queue not being full */
 395         FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
 396         /* FIXME: some kind of backoff (random, exponential) would be
 397          * nice. */
 398         sleep(1);
 399     }
 400     return r;
 401 }
 402
 403 int interrupt_thread(struct thread *th, lispobj function)
 404 {
 405     /* In clone_threads, if A and B both interrupt C at approximately
 406      * the same time, it does not matter: the second signal will be
 407      * masked until the handler has returned from the first one.  In
 408      * pthreads though, we can't put the knowledge of what function to
 409      * call into the siginfo, so we have to store it in the
 410      * destination thread, and do it in such a way that A won't
 411      * clobber B's interrupt.  Hence, this stupid linked list.
 412      *
 413      * This does depend on SIG_INTERRUPT_THREAD being queued (as POSIX
 414      * RT signals are): we need to keep interrupt_fun data for exactly
 415      * as many signals as are going to be received by the destination
 416      * thread.
 417      */
 418     lispobj c=alloc_cons(function,NIL);
 419     sigset_t newset,oldset;
 420     sigemptyset(&newset);
 421     /* interrupt_thread_handler locks this spinlock with blockables
 422      * blocked (it does so for the sake of
 423      * arrange_return_to_lisp_function), so we must also block them or
 424      * else SIG_STOP_FOR_GC and all_threads_lock will find a way to
 425      * deadlock. */
 426     sigaddset_blockable(&newset);
 427     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 428     if (th == arch_os_get_current_thread())
 429         lose("cannot interrupt current thread");
 430     get_spinlock(&th->interrupt_fun_lock,
 431                  (long)arch_os_get_current_thread());
 432     ((struct cons *)native_pointer(c))->cdr=th->interrupt_fun;
 433     th->interrupt_fun=c;
 434     release_spinlock(&th->interrupt_fun_lock);
 435     thread_sigmask(SIG_SETMASK,&oldset,0);
 436     /* Called from lisp with the thread object as a parameter. Thus,
 437      * the object cannot be garbage collected and consequently reaped
 438      * and joined. Because it's not joined, kill should work (even if
 439      * the thread has died/exited). */
 440     {
 441         int status=kill_thread_safely(th->os_thread,SIG_INTERRUPT_THREAD);
 442         if (status==0) {
 443             return 0;
 444         } else if (status==ESRCH) {
 445             /* This thread has exited. */
 446             th->interrupt_fun=NIL;
 447             errno=ESRCH;
 448             return -1;
 449         } else {
 450             lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s",
 451                  th->os_thread,status,strerror(status));
 452         }
 453     }
 454 }
 455
 456 /* stopping the world is a two-stage process.  From this thread we signal
 457  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 458  * the usual pseudo-atomic checks (we don't want to stop a thread while
 459  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 460  */
 461
 462 /* To avoid deadlocks when gc stops the world all clients of each
 463  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 464  * holding the lock, but they must agree on which. */
 465 void gc_stop_the_world()
 466 {
 467     struct thread *p,*th=arch_os_get_current_thread();
 468     int status;
 469     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
 470                   th->os_thread));
 471     /* keep threads from starting while the world is stopped. */
 472     get_spinlock(&all_threads_lock,(long)th);
 473     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
 474                   th->os_thread));
 475     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 476     for(p=all_threads; p; p=p->next) {
 477         while(p->state==STATE_STARTING) sched_yield();
 478         if((p!=th) && (p->state==STATE_RUNNING)) {
 479             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 480             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %lu\n",
 481                           p->os_thread));
 482             if (status==ESRCH) {
 483                 /* This thread has exited. */
 484                 gc_assert(p->state==STATE_DEAD);
 485             } else if (status) {
 486                 lose("cannot send suspend thread=%lu: %d, %s",
 487                      p->os_thread,status,strerror(status));
 488             }
 489         }
 490     }
 491     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 492     /* wait for the running threads to stop or finish */
 493     for(p=all_threads;p;) {
 494         gc_assert(p->os_thread!=0);
 495         gc_assert(p->state!=STATE_STARTING);
 496         if((p==th) || (p->state==STATE_SUSPENDED) ||
 497            (p->state==STATE_DEAD)) {
 498             p=p->next;
 499         } else {
 500             sched_yield();
 501         }
 502     }
 503     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 504 }
 505
 506 void gc_start_the_world()
 507 {
 508     struct thread *p,*th=arch_os_get_current_thread();
 509     int status;
 510     /* if a resumed thread creates a new thread before we're done with
 511      * this loop, the new thread will get consed on the front of
 512      * all_threads, but it won't have been stopped so won't need
 513      * restarting */
 514     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 515     for(p=all_threads;p;p=p->next) {
 516         gc_assert(p->os_thread!=0);
 517         if((p!=th) && (p->state!=STATE_DEAD)) {
 518             if(p->state!=STATE_SUSPENDED) {
 519                 lose("gc_start_the_world: wrong thread state is %d\n",
 520                      fixnum_value(p->state));
 521             }
 522             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 523                           p->os_thread));
 524             p->state=STATE_RUNNING;
 525             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 526             if (status) {
 527                 lose("cannot resume thread=%lu: %d, %s",
 528                      p->os_thread,status,strerror(status));
 529             }
 530         }
 531     }
 532     /* If we waited here until all threads leave STATE_SUSPENDED, then
 533      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
 534      * performance implications, but does away with the 'rt signal
 535      * queue full' problem. */
 536     release_spinlock(&all_threads_lock);
 537     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 538 }
 539 #endif