aix_irix_threads.c

A garbage collector for C and C

    pthread_mutex_unlock(&GC_suspend_lock);
    pthread_cond_broadcast(&GC_continue_cv);
}

#endif /* GC_AIX_THREADS */


/* We hold allocation lock.  Should do exactly the right thing if the	*/
/* world is stopped.  Should not fail if it isn't.			*/
void GC_push_all_stacks()
{
    register int i;
    register GC_thread p;
    register ptr_t hot, cold;
    pthread_t me = pthread_self();
    
    /* GC_init() should have been called before GC_push_all_stacks is
     * invoked, and GC_init calls GC_thr_init(), which sets
     * GC_thr_initialized. */
    GC_ASSERT(GC_thr_initialized);

    /* GC_printf1("Pushing stacks from thread 0x%x\n", me); */
    GC_ASSERT(I_HOLD_LOCK());
    for (i = 0; i < THREAD_TABLE_SZ; i++) {
      for (p = GC_threads[i]; p != 0; p = p -> next) {
        if (p -> flags & FINISHED) continue;
	cold = p->stack_cold;
	if (!cold) cold=GC_stackbottom; /* 0 indicates 'original stack' */
        if (pthread_equal(p -> id, me)) {
	    hot = GC_approx_sp();
	} else {
#        ifdef GC_AIX_THREADS
          /* AIX doesn't use signals to suspend, so we need to get an */
	  /* accurate hot stack pointer.			      */
	  /* See http://publib16.boulder.ibm.com/pseries/en_US/libs/basetrf1/pthread_getthrds_np.htm */
          pthread_t id = p -> id;
          struct __pthrdsinfo pinfo;
          int regbuf[64];
          int val = sizeof(regbuf);
          int retval = pthread_getthrds_np(&id, PTHRDSINFO_QUERY_ALL, &pinfo,
			  		   sizeof(pinfo), regbuf, &val);
          if (retval != 0) {
	    printf("ERROR: pthread_getthrds_np() failed in GC\n");
	    abort();
	  }
	  /* according to the AIX ABI, 
	     "the lowest possible valid stack address is 288 bytes (144 + 144)
	     less than the current value of the stack pointer.  Functions may
	     use this stack space as volatile storage which is not preserved
	     across function calls."
	     ftp://ftp.penguinppc64.org/pub/people/amodra/PPC-elf64abi.txt.gz
	  */
          hot = (ptr_t)(unsigned long)pinfo.__pi_ustk-288;
	  cold = (ptr_t)pinfo.__pi_stackend; /* more precise */
          /* push the registers too, because they won't be on stack */
          GC_push_all_eager((ptr_t)&pinfo.__pi_context,
			    (ptr_t)((&pinfo.__pi_context)+1));
          GC_push_all_eager((ptr_t)regbuf, ((ptr_t)regbuf)+val);
#	 else
              hot = p -> stack_hot;
#	 endif
	}
#	ifdef STACK_GROWS_UP
          GC_push_all_stack(cold, hot);
#	else
 /* printf("thread 0x%x: hot=0x%08x cold=0x%08x\n", p -> id, hot, cold); */
          GC_push_all_stack(hot, cold);
#	endif
      }
    }
}


/* We hold the allocation lock.	*/
void GC_thr_init()
{
    GC_thread t;
    struct sigaction act;

    if (GC_thr_initialized) return;
    GC_ASSERT(I_HOLD_LOCK());
    GC_thr_initialized = TRUE;
#ifndef GC_AIX_THREADS
    (void) sigaction(SIG_SUSPEND, 0, &act);
    if (act.sa_handler != SIG_DFL)
    	ABORT("Previously installed SIG_SUSPEND handler");
    /* Install handler.	*/
	act.sa_handler = GC_suspend_handler;
	act.sa_flags = SA_RESTART;
	(void) sigemptyset(&act.sa_mask);
        if (0 != sigaction(SIG_SUSPEND, &act, 0))
	    ABORT("Failed to install SIG_SUSPEND handler");
#endif
    /* Add the initial thread, so we can stop it.	*/
      t = GC_new_thread(pthread_self());
      /* use '0' to indicate GC_stackbottom, since GC_init() has not
       * completed by the time we are called (from GC_init_inner()) */
      t -> stack_cold = 0; /* the original stack. */
      t -> stack_hot = (ptr_t)(&t);
      t -> flags = DETACHED;
}

int GC_pthread_sigmask(int how, const sigset_t *set, sigset_t *oset)
{
    sigset_t fudged_set;
    
#ifdef GC_AIX_THREADS
    return(pthread_sigmask(how, set, oset));
#endif

    if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
        fudged_set = *set;
        sigdelset(&fudged_set, SIG_SUSPEND);
        set = &fudged_set;
    }
    return(pthread_sigmask(how, set, oset));
}

struct start_info {
    void *(*start_routine)(void *);
    void *arg;
    word flags;
    pthread_mutex_t registeredlock;
    pthread_cond_t registered;     
    int volatile registereddone;
};

void GC_thread_exit_proc(void *arg)
{
    GC_thread me;

    LOCK();
    me = GC_lookup_thread(pthread_self());
    me -> flags |= FINISHED;
    /* reclaim DETACHED thread right away; otherwise wait until join() */
    if (me -> flags & DETACHED) {
	GC_delete_gc_thread(pthread_self(), me);
    }
    UNLOCK();
}

int GC_pthread_join(pthread_t thread, void **retval)
{
    int result;
    GC_thread thread_gc_id;
    
    LOCK();
    thread_gc_id = GC_lookup_thread(thread);
    /* This is guaranteed to be the intended one, since the thread id	*/
    /* cant have been recycled by pthreads.				*/
    UNLOCK();
    GC_ASSERT(!(thread_gc_id->flags & DETACHED));
    result = pthread_join(thread, retval);
    /* Some versions of the Irix pthreads library can erroneously 	*/
    /* return EINTR when the call succeeds.				*/
	if (EINTR == result) result = 0;
    GC_ASSERT(thread_gc_id->flags & FINISHED);
    LOCK();
    /* Here the pthread thread id may have been recycled. */
    GC_delete_gc_thread(thread, thread_gc_id);
    UNLOCK();
    return result;
}

void * GC_start_routine(void * arg)
{
    int dummy;
    struct start_info * si = arg;
    void * result;
    GC_thread me;
    pthread_t my_pthread;
    void *(*start)(void *);
    void *start_arg;

    my_pthread = pthread_self();
    /* If a GC occurs before the thread is registered, that GC will	*/
    /* ignore this thread.  That's fine, since it will block trying to  */
    /* acquire the allocation lock, and won't yet hold interesting 	*/
    /* pointers.							*/
    LOCK();
    /* We register the thread here instead of in the parent, so that	*/
    /* we don't need to hold the allocation lock during pthread_create. */
    /* Holding the allocation lock there would make REDIRECT_MALLOC	*/
    /* impossible.  It probably still doesn't work, but we're a little  */
    /* closer ...							*/
    /* This unfortunately means that we have to be careful the parent	*/
    /* doesn't try to do a pthread_join before we're registered.	*/
    me = GC_new_thread(my_pthread);
    me -> flags = si -> flags;
    me -> stack_cold = (ptr_t) &dummy; /* this now the 'start of stack' */
    me -> stack_hot = me->stack_cold;/* this field should always be sensible */
    UNLOCK();
    start = si -> start_routine;
    start_arg = si -> arg;

    pthread_mutex_lock(&(si->registeredlock));
    si->registereddone = 1;
    pthread_cond_signal(&(si->registered));
    pthread_mutex_unlock(&(si->registeredlock));
    /* si went away as soon as we did this unlock */

    pthread_cleanup_push(GC_thread_exit_proc, 0);
    result = (*start)(start_arg);
    me -> status = result;
    pthread_cleanup_pop(1);
	/* This involves acquiring the lock, ensuring that we can't exit */
	/* while a collection that thinks we're alive is trying to stop  */
	/* us.								 */
    return(result);
}

int
GC_pthread_create(pthread_t *new_thread,
		  const pthread_attr_t *attr,
                  void *(*start_routine)(void *), void *arg)
{
    int result;
    GC_thread t;
    int detachstate;
    word my_flags = 0;
    struct start_info * si;
    	/* This is otherwise saved only in an area mmapped by the thread */
    	/* library, which isn't visible to the collector.		 */

    LOCK();
    /* GC_INTERNAL_MALLOC implicitly calls GC_init() if required */
    si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
						 NORMAL);
    GC_ASSERT(GC_thr_initialized); /* initialized by GC_init() */
    UNLOCK();
    if (0 == si) return(ENOMEM);
    pthread_mutex_init(&(si->registeredlock), NULL);
    pthread_cond_init(&(si->registered),NULL);
    pthread_mutex_lock(&(si->registeredlock));
    si -> start_routine = start_routine;
    si -> arg = arg;

    pthread_attr_getdetachstate(attr, &detachstate);
    if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
    si -> flags = my_flags;
    result = pthread_create(new_thread, attr, GC_start_routine, si); 

    /* Wait until child has been added to the thread table.		*/
    /* This also ensures that we hold onto si until the child is done	*/
    /* with it.  Thus it doesn't matter whether it is otherwise		*/
    /* visible to the collector.					*/

    if (0 == result) {
      si->registereddone = 0;
      while (!si->registereddone) 
        pthread_cond_wait(&(si->registered), &(si->registeredlock));
    }
    pthread_mutex_unlock(&(si->registeredlock));

    pthread_cond_destroy(&(si->registered));
    pthread_mutex_destroy(&(si->registeredlock));
    LOCK();
    GC_INTERNAL_FREE(si);
    UNLOCK();

    return(result);
}

/* For now we use the pthreads locking primitives on HP/UX */

VOLATILE GC_bool GC_collecting = 0; /* A hint that we're in the collector and       */
                        /* holding the allocation lock for an           */
                        /* extended period.                             */

/* Reasonably fast spin locks.  Basically the same implementation */
/* as STL alloc.h.						  */

#define SLEEP_THRESHOLD 3

volatile unsigned int GC_allocate_lock = 0;
#define GC_TRY_LOCK() !GC_test_and_set(&GC_allocate_lock)
#define GC_LOCK_TAKEN GC_allocate_lock

void GC_lock()
{
#   define low_spin_max 30  /* spin cycles if we suspect uniprocessor */
#   define high_spin_max 1000 /* spin cycles for multiprocessor */
    static unsigned spin_max = low_spin_max;
    unsigned my_spin_max;
    static unsigned last_spins = 0;
    unsigned my_last_spins;
    volatile unsigned junk;
#   define PAUSE junk *= junk; junk *= junk; junk *= junk; junk *= junk
    int i;

    if (GC_TRY_LOCK()) {
        return;
    }
    junk = 0;
    my_spin_max = spin_max;
    my_last_spins = last_spins;
    for (i = 0; i < my_spin_max; i++) {
        if (GC_collecting) goto yield;
        if (i < my_last_spins/2 || GC_LOCK_TAKEN) {
            PAUSE; 
            continue;
        }
        if (GC_TRY_LOCK()) {
	    /*
             * got it!
             * Spinning worked.  Thus we're probably not being scheduled
             * against the other process with which we were contending.
             * Thus it makes sense to spin longer the next time.
	     */
            last_spins = i;
            spin_max = high_spin_max;
            return;
        }
    }
    /* We are probably being scheduled against the other process.  Sleep. */
    spin_max = low_spin_max;
yield:
    for (i = 0;; ++i) {
        if (GC_TRY_LOCK()) {
            return;
        }
        if (i < SLEEP_THRESHOLD) {
            sched_yield();
	} else {
	    struct timespec ts;
	
	    if (i > 26) i = 26;
			/* Don't wait for more than about 60msecs, even	*/
			/* under extreme contention.			*/
	    ts.tv_sec = 0;
	    ts.tv_nsec = 1 << i;
	    nanosleep(&ts, 0);
	}
    }
}

# else  /* !GC_IRIX_THREADS && !GC_AIX_THREADS */

#ifndef LINT
  int GC_no_Irix_threads;
#endif

# endif /* IRIX_THREADS */