pthread_support.c

Boost provides free peer-reviewed portable C++ source libraries. We emphasize libraries that work

    GC_ASSERT(I_HOLD_LOCK());
#   ifdef PARALLEL_MARK
      for (i = 0; i < GC_markers; ++i) {
	if (marker_sp[i] > lo & marker_sp[i] < hi) return TRUE;
#       ifdef IA64
	  if (marker_bsp[i] > lo & marker_bsp[i] < hi) return TRUE;
#	endif
      }
#   endif
    for (i = 0; i < THREAD_TABLE_SZ; i++) {
      for (p = GC_threads[i]; p != 0; p = p -> next) {
	if (0 != p -> stack_end) {
#	  ifdef STACK_GROWS_UP
            if (p -> stack_end >= lo && p -> stack_end < hi) return TRUE;
#	  else /* STACK_GROWS_DOWN */
            if (p -> stack_end > lo && p -> stack_end <= hi) return TRUE;
#	  endif
	}
      }
    }
    return FALSE;
}
#endif /* USE_PROC_FOR_LIBRARIES */

#ifdef IA64
/* Find the largest stack_base smaller than bound.  May be used	*/
/* to find the boundary between a register stack and adjacent	*/
/* immediately preceding memory stack.				*/
ptr_t GC_greatest_stack_base_below(ptr_t bound)
{
    int i;
    GC_thread p;
    ptr_t result = 0;
    
    GC_ASSERT(I_HOLD_LOCK());
#   ifdef PARALLEL_MARK
      for (i = 0; i < GC_markers; ++i) {
	if (marker_sp[i] > result && marker_sp[i] < bound)
	  result = marker_sp[i];
      }
#   endif
    for (i = 0; i < THREAD_TABLE_SZ; i++) {
      for (p = GC_threads[i]; p != 0; p = p -> next) {
	if (p -> stack_end > result && p -> stack_end < bound) {
	  result = p -> stack_end;
	}
      }
    }
    return result;
}
#endif /* IA64 */

#ifdef GC_LINUX_THREADS
/* Return the number of processors, or i<= 0 if it can't be determined.	*/
int GC_get_nprocs(void)
{
    /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that	*/
    /* appears to be buggy in many cases.				*/
    /* We look for lines "cpu<n>" in /proc/stat.			*/
#   define STAT_BUF_SIZE 4096
#   define STAT_READ read
	/* If read is wrapped, this may need to be redefined to call 	*/
	/* the real one.						*/
    char stat_buf[STAT_BUF_SIZE];
    int f;
    word result = 1;
	/* Some old kernels only have a single "cpu nnnn ..."	*/
	/* entry in /proc/stat.  We identify those as 		*/
	/* uniprocessors.					*/
    size_t i, len = 0;

    f = open("/proc/stat", O_RDONLY);
    if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
	WARN("Couldn't read /proc/stat\n", 0);
	return -1;
    }
    for (i = 0; i < len - 100; ++i) {
        if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
	    && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
	    int cpu_no = atoi(stat_buf + i + 4);
	    if (cpu_no >= result) result = cpu_no + 1;
	}
    }
    close(f);
    return result;
}
#endif /* GC_LINUX_THREADS */

/* We hold the GC lock.  Wait until an in-progress GC has finished.	*/
/* Repeatedly RELEASES GC LOCK in order to wait.			*/
/* If wait_for_all is true, then we exit with the GC lock held and no	*/
/* collection in progress; otherwise we just wait for the current GC	*/
/* to finish.								*/
extern GC_bool GC_collection_in_progress(void);
void GC_wait_for_gc_completion(GC_bool wait_for_all)
{
    GC_ASSERT(I_HOLD_LOCK());
    if (GC_incremental && GC_collection_in_progress()) {
	int old_gc_no = GC_gc_no;

	/* Make sure that no part of our stack is still on the mark stack, */
	/* since it's about to be unmapped.				   */
	while (GC_incremental && GC_collection_in_progress()
	       && (wait_for_all || old_gc_no == GC_gc_no)) {
	    ENTER_GC();
	    GC_in_thread_creation = TRUE;
            GC_collect_a_little_inner(1);
	    GC_in_thread_creation = FALSE;
	    EXIT_GC();
	    UNLOCK();
	    sched_yield();
	    LOCK();
	}
    }
}

#ifdef HANDLE_FORK
/* Procedures called before and after a fork.  The goal here is to make */
/* it safe to call GC_malloc() in a forked child.  It's unclear that is	*/
/* attainable, since the single UNIX spec seems to imply that one 	*/
/* should only call async-signal-safe functions, and we probably can't	*/
/* quite guarantee that.  But we give it our best shot.  (That same	*/
/* spec also implies that it's not safe to call the system malloc	*/
/* between fork() and exec().  Thus we're doing no worse than it.	*/

/* Called before a fork()		*/
void GC_fork_prepare_proc(void)
{
    /* Acquire all relevant locks, so that after releasing the locks	*/
    /* the child will see a consistent state in which monitor 		*/
    /* invariants hold.	 Unfortunately, we can't acquire libc locks	*/
    /* we might need, and there seems to be no guarantee that libc	*/
    /* must install a suitable fork handler.				*/
    /* Wait for an ongoing GC to finish, since we can't finish it in	*/
    /* the (one remaining thread in) the child.				*/
      LOCK();
#     if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
        GC_wait_for_reclaim();
#     endif
      GC_wait_for_gc_completion(TRUE);
#     if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
        GC_acquire_mark_lock();
#     endif
}

/* Called in parent after a fork()	*/
void GC_fork_parent_proc(void)
{
#   if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
      GC_release_mark_lock();
#   endif
    UNLOCK();
}

/* Called in child after a fork()	*/
void GC_fork_child_proc(void)
{
    /* Clean up the thread table, so that just our thread is left. */
#   if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
      GC_release_mark_lock();
#   endif
    GC_remove_all_threads_but_me();
#   ifdef PARALLEL_MARK
      /* Turn off parallel marking in the child, since we are probably 	*/
      /* just going to exec, and we would have to restart mark threads.	*/
        GC_markers = 1;
        GC_parallel = FALSE;
#   endif /* PARALLEL_MARK */
    UNLOCK();
}
#endif /* HANDLE_FORK */

#if defined(GC_DGUX386_THREADS)
/* Return the number of processors, or i<= 0 if it can't be determined. */
int GC_get_nprocs(void)
{
    /* <[e-mail]> */
    int numCpus;
    struct dg_sys_info_pm_info pm_sysinfo;
    int status =0;

    status = dg_sys_info((long int *) &pm_sysinfo,
	DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
    if (status < 0)
       /* set -1 for error */
       numCpus = -1;
    else
      /* Active CPUs */
      numCpus = pm_sysinfo.idle_vp_count;

#  ifdef DEBUG_THREADS
    GC_printf("Number of active CPUs in this system: %d\n", numCpus);
#  endif
    return(numCpus);
}
#endif /* GC_DGUX386_THREADS */

#if defined(GC_NETBSD_THREADS)
static int get_ncpu(void)
{
    int mib[] = {CTL_HW,HW_NCPU};
    int res;
    size_t len = sizeof(res);

    sysctl(mib, sizeof(mib)/sizeof(int), &res, &len, NULL, 0);
    return res;
}
#endif	/* GC_NETBSD_THREADS */

# if defined(GC_LINUX_THREADS) && defined(INCLUDE_LINUX_THREAD_DESCR)
__thread int dummy_thread_local;
# endif

/* We hold the allocation lock.	*/
void GC_thr_init(void)
{
#   ifndef GC_DARWIN_THREADS
        int dummy;
#   endif
    GC_thread t;

    if (GC_thr_initialized) return;
    GC_thr_initialized = TRUE;
    
#   ifdef HANDLE_FORK
      /* Prepare for a possible fork.	*/
        pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
	  	       GC_fork_child_proc);
#   endif /* HANDLE_FORK */
#   if defined(INCLUDE_LINUX_THREAD_DESCR)
      /* Explicitly register the region including the address 		*/
      /* of a thread local variable.  This should included thread	*/
      /* locals for the main thread, except for those allocated		*/
      /* in response to dlopen calls.					*/  
	{
	  ptr_t thread_local_addr = (ptr_t)(&dummy_thread_local);
	  ptr_t main_thread_start, main_thread_end;
          if (!GC_enclosing_mapping(thread_local_addr, &main_thread_start,
				    &main_thread_end)) {
	    ABORT("Failed to find mapping for main thread thread locals");
	  }
	  GC_add_roots_inner(main_thread_start, main_thread_end, FALSE);
	}
#   endif
    /* Add the initial thread, so we can stop it.	*/
      t = GC_new_thread(pthread_self());
#     ifdef GC_DARWIN_THREADS
         t -> stop_info.mach_thread = mach_thread_self();
#     else
         t -> stop_info.stack_ptr = (ptr_t)(&dummy);
#     endif
      t -> flags = DETACHED | MAIN_THREAD;

    GC_stop_init();

    /* Set GC_nprocs.  */
      {
	char * nprocs_string = GETENV("GC_NPROCS");
	GC_nprocs = -1;
	if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
      }
      if (GC_nprocs <= 0) {
#       if defined(GC_HPUX_THREADS)
	  GC_nprocs = pthread_num_processors_np();
#       endif
#	if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS) \
	   || defined(GC_SOLARIS_THREADS)
	  GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
	  if (GC_nprocs <= 0) GC_nprocs = 1;
#	endif
#       if defined(GC_IRIX_THREADS)
	  GC_nprocs = sysconf(_SC_NPROC_ONLN);
	  if (GC_nprocs <= 0) GC_nprocs = 1;
#       endif
#       if defined(GC_NETBSD_THREADS)
	  GC_nprocs = get_ncpu();
#       endif
#       if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
	  int ncpus = 1;
	  size_t len = sizeof(ncpus);
	  sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
	  GC_nprocs = ncpus;
#       endif
#	if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
          GC_nprocs = GC_get_nprocs();
#	endif
#       if defined(GC_GNU_THREADS)
	  if (GC_nprocs <= 0) GC_nprocs = 1;
#       endif
      }
      if (GC_nprocs <= 0) {
	WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
	GC_nprocs = 2;
#	ifdef PARALLEL_MARK
	  GC_markers = 1;
#	endif
      } else {
#	ifdef PARALLEL_MARK
          {
	    char * markers_string = GETENV("GC_MARKERS");
	    if (markers_string != NULL) {
	      GC_markers = atoi(markers_string);
	    } else {
	      GC_markers = GC_nprocs;
	    }
          }
#	endif
      }
#   ifdef PARALLEL_MARK
      if (GC_print_stats) {
          GC_log_printf("Number of processors = %ld, "
		 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
      }
      if (GC_markers == 1) {
	GC_parallel = FALSE;
	if (GC_print_stats) {
	    GC_log_printf(
		"Single marker thread, turning off parallel marking\n");
	}
      } else {
	GC_parallel = TRUE;
	/* Disable true incremental collection, but generational is OK.	*/
	GC_time_limit = GC_TIME_UNLIMITED;
      }
      /* If we are using a parallel marker, actually start helper threads.  */
        if (GC_parallel) start_mark_threads();
#   endif
}


/* Perform all initializations, including those that	*/
/* may require allocation.				*/
/* Called without allocation lock.			*/
/* Must be called before a second thread is created.	*/
/* Did we say it's called without the allocation lock?	*/
void GC_init_parallel(void)
{
    if (parallel_initialized) return;
    parallel_initialized = TRUE;

    /* GC_init() calls us back, so set flag first.	*/
    if (!GC_is_initialized) GC_init();
    /* Initialize thread local free lists if used.	*/
#   if defined(THREAD_LOCAL_ALLOC)
      LOCK();
      GC_init_thread_local(&(GC_lookup_thread(pthread_self())->tlfs));
      UNLOCK();
#   endif
}


#if !defined(GC_DARWIN_THREADS)
int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
{
    sigset_t fudged_set;
    
    INIT_REAL_SYMS();
    if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
        fudged_set = *set;
        sigdelset(&fudged_set, SIG_SUSPEND);
        set = &fudged_set;
    }
    return(REAL_FUNC(pthread_sigmask)(how, set, oset));
}
#endif /* !GC_DARWIN_THREADS */

/* Wrapper for functions that are likely to block for an appreciable	*/
/* length of time.							*/

struct blocking_data {
    void (*fn)(void *);
    void *arg;
};

static void GC_do_blocking_inner(ptr_t data, void * context) {
    struct blocking_data * d = (struct blocking_data *) data;
    GC_thread me;
    LOCK();
    me = GC_lookup_thread(pthread_self());
    GC_ASSERT(!(me -> thread_blocked));
#   ifdef SPARC
	me -> stop_info.stack_ptr = GC_save_regs_in_stack();
#   elif !defined(GC_DARWIN_THREADS)
	me -> stop_info.stack_ptr = GC_approx_sp();
#   endif
#   ifdef IA64
	me -> backing_store_ptr = GC_save_regs_in_stack();
#   endif
    me -> thread_blocked = TRUE;
    /* Save context here if we want to support precise stack marking */
    UNLOCK();
    (d -> fn)(d -> arg);
    LOCK();   /* This will block if the world is stopped.	*/
    me -> thread_blocked = FALSE;
    UNLOCK();
}

void GC_do_blocking(void (*fn)(void *), void *arg) {
    struct blocking_data my_data;

    my_data.fn = fn;
    my_data.arg = arg;
    GC_with_callee_saves_pushed(GC_do_blocking_inner, (ptr_t)(&my_data));
}
    
struct start_info {
    void *(*start_routine)(void *);
    void *arg;
    word flags;
    sem_t registered;   	/* 1 ==> in our thread table, but 	*/
				/* parent hasn't yet noticed.		*/
};

int GC_unregister_my_thread(void)
{
    GC_thread me;

    LOCK();
    /* Wait for any GC that may be marking from our stack to	*/
    /* complete before we remove this thread.			*/
    GC_wait_for_gc_completion(FALSE);
    me = GC_lookup_thread(pthread_self());
#   if defined(THREAD_LOCAL_ALLOC)
      GC_destroy_thread_local(&(me->tlfs));
#   endif
    if (me -> flags & DETACHED) {
    	GC_delete_thread(pthread_self());
    } else {
	me -> flags |= FINISHED;
    }
#   if defined(THREAD_LOCAL_ALLOC)
      GC_remove_specific(GC_thread_key);
#   endif
    UNLOCK();
    return GC_SUCCESS;
}

/* Called at thread exit.				*/
/* Never called for main thread.  That's OK, since it	*/
/* results in at most a tiny one-time leak.  And 	*/
/* linuxthreads doesn't reclaim the main threads 	*/
/* resources or id anyway.				*/
void GC_thread_exit_proc(void *arg)
{
    GC_unregister_my_thread();
}

int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
{
    int result;
    GC_thread thread_gc_id;
    
    INIT_REAL_SYMS();
    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();
    result = REAL_FUNC(pthread_join)(thread, retval);
# if defined (GC_FREEBSD_THREADS)
    /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
       appears to be) a spurious EINTR which caused the test and real code
       to gratuitously fail.  Having looked at system pthread library source
       code, I see how this return code may be generated.  In one path of
       code, pthread_join() just returns the errno setting of the thread
       being joined.  This does not match the POSIX specification or the
       local man pages thus I have taken the liberty to catch this one
       spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
    if (result == EINTR) result = 0;
# endif
    if (result == 0) {
        LOCK();
        /* Here the pthread thread id may have been recycled. */
        GC_delete_gc_thread(thread_gc_id);
        UNLOCK();
    }
    return result;
}

int
WRAP_FUNC(pthread_detach)(pthread_t thread)
{
    int result;
    GC_thread thread_gc_id;
    
    INIT_REAL_SYMS();
    LOCK();
    thread_gc_id = GC_lookup_thread(thread);
    UNLOCK();
    result = REAL_FUNC(pthread_detach)(thread);
    if (result == 0) {
      LOCK();
      thread_gc_id -> flags |= DETACHED;
      /* Here the pthread thread id may have been recycled. */
      if (thread_gc_id -> flags & FINISHED) {
        GC_delete_gc_thread(thread_gc_id);
      }
      UNLOCK();
    }