pthread_support.c

/* 
 * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
 * Copyright (c) 1996 by Silicon Graphics.  All rights reserved.
 * Copyright (c) 1998 by Fergus Henderson.  All rights reserved.
 * Copyright (c) 2000-2004 by Hewlett-Packard Company.  All rights reserved.
 *
 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
 * OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
 *
 * Permission is hereby granted to use or copy this program
 * for any purpose,  provided the above notices are retained on all copies.
 * Permission to modify the code and to distribute modified code is granted,
 * provided the above notices are retained, and a notice that the code was
 * modified is included with the above copyright notice.
 */
/*
 * Support code for LinuxThreads, the clone()-based kernel
 * thread package for Linux which is included in libc6.
 *
 * This code relies on implementation details of LinuxThreads,
 * (i.e. properties not guaranteed by the Pthread standard),
 * though this version now does less of that than the other Pthreads
 * support code.
 *
 * Note that there is a lot of code duplication between linux_threads.c
 * and thread support for some of the other Posix platforms; any changes
 * made here may need to be reflected there too.
 */
 /* DG/UX ix86 support <takis@xfree86.org> */
/*
 * Linux_threads.c now also includes some code to support HPUX and
 * OSF1 (Compaq Tru64 Unix, really).  The OSF1 support is based on Eric Benson's
 * patch.
 *
 * Eric also suggested an alternate basis for a lock implementation in
 * his code:
 * + #elif defined(OSF1)
 * +    unsigned long GC_allocate_lock = 0;
 * +    msemaphore GC_allocate_semaphore;
 * + #  define GC_TRY_LOCK() \
 * +    ((msem_lock(&GC_allocate_semaphore, MSEM_IF_NOWAIT) == 0) \
 * +     ? (GC_allocate_lock = 1) \
 * +     : 0)
 * + #  define GC_LOCK_TAKEN GC_allocate_lock
 */

/*#define DEBUG_THREADS 1*/
/*#define GC_ASSERTIONS*/

#include "gc_config.h"

#ifdef GC_PTHREAD_SYM_VERSION
#define _GNU_SOURCE
#include <dlfcn.h>
#endif

# include "gc.h"
# include "private/pthread_support.h"

# if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
     && !defined(GC_WIN32_THREADS)

# if defined(GC_HPUX_THREADS) && !defined(USE_PTHREAD_SPECIFIC) \
     && !defined(USE_COMPILER_TLS)
#   ifdef __GNUC__
#     define USE_PTHREAD_SPECIFIC
      /* Empirically, as of gcc 3.3, USE_COMPILER_TLS doesn't work.	*/
#   else
#     define USE_COMPILER_TLS
#   endif
# endif

# if defined USE_HPUX_TLS
    --> Macro replaced by USE_COMPILER_TLS
# endif

# if (defined(GC_DGUX386_THREADS) || defined(GC_OSF1_THREADS) || \
      defined(GC_DARWIN_THREADS) || defined(GC_AIX_THREADS)) \
      && !defined(USE_PTHREAD_SPECIFIC)
#   define USE_PTHREAD_SPECIFIC
# endif

# if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
#   define _POSIX4A_DRAFT10_SOURCE 1
# endif

# if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
#   define _USING_POSIX4A_DRAFT10 1
# endif

# ifdef THREAD_LOCAL_ALLOC
#   if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
#     include "private/specific.h"
#   endif
#   if defined(USE_PTHREAD_SPECIFIC)
#     define GC_getspecific pthread_getspecific
#     define GC_setspecific pthread_setspecific
#     define GC_key_create pthread_key_create
      typedef pthread_key_t GC_key_t;
#   endif
#   if defined(USE_COMPILER_TLS)
#     define GC_getspecific(x) (x)
#     define GC_setspecific(key, v) ((key) = (v), 0)
#     define GC_key_create(key, d) 0
      typedef void * GC_key_t;
#   endif
# endif
# include <stdlib.h>
# include <pthread.h>
# include <sched.h>
# include <time.h>
# include <errno.h>
# include <unistd.h>
# include <sys/mman.h>
# include <sys/time.h>
# include <sys/types.h>
# include <sys/stat.h>
# include <fcntl.h>
# include <signal.h>

#if defined(GC_DARWIN_THREADS)
# include "private/darwin_semaphore.h"
#else
# include <semaphore.h>
#endif /* !GC_DARWIN_THREADS */

#if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
# include <sys/sysctl.h>
#endif /* GC_DARWIN_THREADS */



#if defined(GC_DGUX386_THREADS)
# include <sys/dg_sys_info.h>
# include <sys/_int_psem.h>
  /* sem_t is an uint in DG/UX */
  typedef unsigned int  sem_t;
#endif /* GC_DGUX386_THREADS */

#ifndef __GNUC__
#   define __inline__
#endif

#ifdef GC_USE_LD_WRAP
#   define WRAP_FUNC(f) __wrap_##f
#   define REAL_FUNC(f) __real_##f
#else
#   define WRAP_FUNC(f) GC_##f
#   if !defined(GC_DGUX386_THREADS)
#     define REAL_FUNC(f) f
#   else /* GC_DGUX386_THREADS */
#     define REAL_FUNC(f) __d10_##f
#   endif /* GC_DGUX386_THREADS */
#   undef pthread_create
#   if !defined(GC_DARWIN_THREADS)
#     undef pthread_sigmask
#   endif
#   undef pthread_join
#   undef pthread_detach
#   if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
       && !defined(_PTHREAD_USE_PTDNAM_)
/* Restore the original mangled names on Tru64 UNIX.  */
#     define pthread_create __pthread_create
#     define pthread_join __pthread_join
#     define pthread_detach __pthread_detach
#   endif
#endif

void GC_thr_init();

static GC_bool parallel_initialized = FALSE;

void GC_init_parallel();

# if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)

/* We don't really support thread-local allocation with DBG_HDRS_ALL */

#ifdef USE_COMPILER_TLS
  __thread
#endif
GC_key_t GC_thread_key;

static GC_bool keys_initialized;

/* Recover the contents of the freelist array fl into the global one gfl.*/
/* Note that the indexing scheme differs, in that gfl has finer size	*/
/* resolution, even if not all entries are used.			*/
/* We hold the allocator lock.						*/
static void return_freelists(ptr_t *fl, ptr_t *gfl)
{
    int i;
    ptr_t q, *qptr;
    size_t nwords;

    for (i = 1; i < NFREELISTS; ++i) {
	nwords = i * (GRANULARITY/sizeof(word));
        qptr = fl + i;	
	q = *qptr;
	if ((word)q >= HBLKSIZE) {
	  if (gfl[nwords] == 0) {
	    gfl[nwords] = q;
	  } else {
	    /* Concatenate: */
	    for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
	    GC_ASSERT(0 == q);
	    *qptr = gfl[nwords];
	    gfl[nwords] = fl[i];
	  }
	}
	/* Clear fl[i], since the thread structure may hang around.	*/
	/* Do it in a way that is likely to trap if we access it.	*/
	fl[i] = (ptr_t)HBLKSIZE;
    }
}

/* We statically allocate a single "size 0" object. It is linked to	*/
/* itself, and is thus repeatedly reused for all size 0 allocation	*/
/* requests.  (Size 0 gcj allocation requests are incorrect, and	*/
/* we arrange for those to fault asap.)					*/
static ptr_t size_zero_object = (ptr_t)(&size_zero_object);

/* Each thread structure must be initialized.	*/
/* This call must be made from the new thread.	*/
/* Caller holds allocation lock.		*/
void GC_init_thread_local(GC_thread p)
{
    int i;

    if (!keys_initialized) {
	if (0 != GC_key_create(&GC_thread_key, 0)) {
	    ABORT("Failed to create key for local allocator");
        }
	keys_initialized = TRUE;
    }
    if (0 != GC_setspecific(GC_thread_key, p)) {
	ABORT("Failed to set thread specific allocation pointers");
    }
    for (i = 1; i < NFREELISTS; ++i) {
	p -> ptrfree_freelists[i] = (ptr_t)1;
	p -> normal_freelists[i] = (ptr_t)1;
#	ifdef GC_GCJ_SUPPORT
	  p -> gcj_freelists[i] = (ptr_t)1;
#	endif
    }   
    /* Set up the size 0 free lists.	*/
    p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
    p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
#   ifdef GC_GCJ_SUPPORT
        p -> gcj_freelists[0] = (ptr_t)(-1);
#   endif
}

#ifdef GC_GCJ_SUPPORT
  extern ptr_t * GC_gcjobjfreelist;
#endif

/* We hold the allocator lock.	*/
void GC_destroy_thread_local(GC_thread p)
{
    /* We currently only do this from the thread itself or from	*/
    /* the fork handler for a child process.			*/
#   ifndef HANDLE_FORK
      GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
#   endif
    return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
    return_freelists(p -> normal_freelists, GC_objfreelist);
#   ifdef GC_GCJ_SUPPORT
   	return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
#   endif
}

extern GC_PTR GC_generic_malloc_many();

GC_PTR GC_local_malloc(size_t bytes)
{
    if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
        return(GC_malloc(bytes));
    } else {
	int index = INDEX_FROM_BYTES(bytes);
	ptr_t * my_fl;
	ptr_t my_entry;
#	if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
	GC_key_t k = GC_thread_key;
#	endif
	void * tsd;

#	if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
	    if (EXPECT(0 == k, 0)) {
		/* This can happen if we get called when the world is	*/
		/* being initialized.  Whether we can actually complete	*/
		/* the initialization then is unclear.			*/
		GC_init_parallel();
		k = GC_thread_key;
	    }
#	endif
	tsd = GC_getspecific(GC_thread_key);
#	ifdef GC_ASSERTIONS
	  LOCK();
	  GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
	  UNLOCK();
#	endif
	my_fl = ((GC_thread)tsd) -> normal_freelists + index;
	my_entry = *my_fl;
	if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
	    ptr_t next = obj_link(my_entry);
	    GC_PTR result = (GC_PTR)my_entry;
	    *my_fl = next;
	    obj_link(my_entry) = 0;
	    PREFETCH_FOR_WRITE(next);
	    return result;
	} else if ((word)my_entry - 1 < DIRECT_GRANULES) {
	    *my_fl = my_entry + index + 1;
            return GC_malloc(bytes);
	} else {
	    GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
	    if (*my_fl == 0) return GC_oom_fn(bytes);
	    return GC_local_malloc(bytes);
	}
    }
}

GC_PTR GC_local_malloc_atomic(size_t bytes)
{
    if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
        return(GC_malloc_atomic(bytes));
    } else {
	int index = INDEX_FROM_BYTES(bytes);
	ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
		        -> ptrfree_freelists + index;
	ptr_t my_entry = *my_fl;
    
	if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
	    GC_PTR result = (GC_PTR)my_entry;
	    *my_fl = obj_link(my_entry);
	    return result;
	} else if ((word)my_entry - 1 < DIRECT_GRANULES) {
	    *my_fl = my_entry + index + 1;
        return GC_malloc_atomic(bytes);
	} else {
	    GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
	    /* *my_fl is updated while the collector is excluded;	*/
	    /* the free list is always visible to the collector as 	*/
	    /* such.							*/
	    if (*my_fl == 0) return GC_oom_fn(bytes);
	    return GC_local_malloc_atomic(bytes);
	}
    }
}

#ifdef GC_GCJ_SUPPORT

#include "include/gc_gcj.h"

#ifdef GC_ASSERTIONS
  extern GC_bool GC_gcj_malloc_initialized;
#endif

extern int GC_gcj_kind;

GC_PTR GC_local_gcj_malloc(size_t bytes,
			   void * ptr_to_struct_containing_descr)
{
    GC_ASSERT(GC_gcj_malloc_initialized);
    if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
        return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
    } else {
	int index = INDEX_FROM_BYTES(bytes);
	ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
	                -> gcj_freelists + index;
	ptr_t my_entry = *my_fl;
	if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
	    GC_PTR result = (GC_PTR)my_entry;
	    GC_ASSERT(!GC_incremental);
	    /* We assert that any concurrent marker will stop us.	*/
	    /* Thus it is impossible for a mark procedure to see the 	*/
	    /* allocation of the next object, but to see this object 	*/
	    /* still containing a free list pointer.  Otherwise the 	*/
	    /* marker might find a random "mark descriptor".		*/
	    *(volatile ptr_t *)my_fl = obj_link(my_entry);
	    /* We must update the freelist before we store the pointer.	*/
	    /* Otherwise a GC at this point would see a corrupted	*/
	    /* free list.						*/
	    /* A memory barrier is probably never needed, since the 	*/
	    /* action of stopping this thread will cause prior writes	*/
	    /* to complete.						*/
	    GC_ASSERT(((void * volatile *)result)[1] == 0); 
	    *(void * volatile *)result = ptr_to_struct_containing_descr; 
	    return result;
	} else if ((word)my_entry - 1 < DIRECT_GRANULES) {
	    if (!GC_incremental) *my_fl = my_entry + index + 1;
	    	/* In the incremental case, we always have to take this */
	    	/* path.  Thus we leave the counter alone.		*/
            return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
	} else {
	    GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
	    if (*my_fl == 0) return GC_oom_fn(bytes);
	    return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
	}
    }
}

#endif /* GC_GCJ_SUPPORT */

# else  /* !THREAD_LOCAL_ALLOC  && !DBG_HDRS_ALL */

#   define GC_destroy_thread_local(t)

# endif /* !THREAD_LOCAL_ALLOC */

#if 0
/*
To make sure that we're using LinuxThreads and not some other thread
package, we generate a dummy reference to `pthread_kill_other_threads_np'
(was `__pthread_initial_thread_bos' but that disappeared),
which is a symbol defined in LinuxThreads, but (hopefully) not in other
thread packages.

We no longer do this, since this code is now portable enough that it might
actually work for something else.
*/
void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
#endif /* 0 */

long GC_nprocs = 1;	/* Number of processors.  We may not have	*/
			/* access to all of them, but this is as good	*/
			/* a guess as any ...				*/

#ifdef PARALLEL_MARK

# ifndef MAX_MARKERS
#   define MAX_MARKERS 16
# endif

static ptr_t marker_sp[MAX_MARKERS] = {0};

void * GC_mark_thread(void * id)
{
  word my_mark_no = 0;

  marker_sp[(word)id] = GC_approx_sp();
  for (;; ++my_mark_no) {
    /* GC_mark_no is passed only to allow GC_help_marker to terminate	*/
    /* promptly.  This is important if it were called from the signal	*/
    /* handler or from the GC lock acquisition code.  Under Linux, it's	*/
    /* not safe to call it from a signal handler, since it uses mutexes	*/
    /* and condition variables.  Since it is called only here, the 	*/
    /* argument is unnecessary.						*/
    if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
	/* resynchronize if we get far off, e.g. because GC_mark_no	*/
	/* wrapped.							*/
	my_mark_no = GC_mark_no;
    }
#   ifdef DEBUG_THREADS
	GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
#   endif
    GC_help_marker(my_mark_no);
  }
}

extern long GC_markers;		/* Number of mark threads we would	*/
				/* like to have.  Includes the 		*/
				/* initiating thread.			*/

pthread_t GC_mark_threads[MAX_MARKERS];

#define PTHREAD_CREATE REAL_FUNC(pthread_create)

static void start_mark_threads()
{
    unsigned i;
    pthread_attr_t attr;

    if (GC_markers > MAX_MARKERS) {
	WARN("Limiting number of mark threads\n", 0);
	GC_markers = MAX_MARKERS;
    }
    if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
	
    if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
	ABORT("pthread_attr_setdetachstate failed");

#   if defined(HPUX) || defined(GC_DGUX386_THREADS)
      /* Default stack size is usually too small: fix it. */
      /* Otherwise marker threads or GC may run out of	  */
      /* space.						  */
#     define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
      {
	size_t old_size;
	int code;

        if (pthread_attr_getstacksize(&attr, &old_size) != 0)
	  ABORT("pthread_attr_getstacksize failed\n");
	if (old_size < MIN_STACK_SIZE) {
	  if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
		  ABORT("pthread_attr_setstacksize failed\n");
	}
      }
#   endif /* HPUX || GC_DGUX386_THREADS */
#   ifdef CONDPRINT
      if (GC_print_stats) {
	GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
      }
#   endif
    for (i = 0; i < GC_markers - 1; ++i) {
      if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
			      GC_mark_thread, (void *)(word)i)) {
	WARN("Marker thread creation failed, errno = %ld.\n", errno);
      }
    }
}

#else  /* !PARALLEL_MARK */

static __inline__ void start_mark_threads()
{
}

#endif /* !PARALLEL_MARK */

GC_bool GC_thr_initialized = FALSE;

volatile GC_thread GC_threads[THREAD_TABLE_SZ];

void GC_push_thread_structures GC_PROTO((void))
{
    GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
#   if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
      GC_push_all((ptr_t)(&GC_thread_key),
	  (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
#   endif
}

#ifdef THREAD_LOCAL_ALLOC
/* We must explicitly mark ptrfree and gcj free lists, since the free 	*/
/* list links wouldn't otherwise be found.  We also set them in the 	*/
/* normal free lists, since that involves touching less memory than if	*/
/* we scanned them normally.						*/
void GC_mark_thread_local_free_lists(void)
{
    int i, j;
    GC_thread p;
    ptr_t q;
    
    for (i = 0; i < THREAD_TABLE_SZ; ++i) {
      for (p = GC_threads[i]; 0 != p; p = p -> next) {
	for (j = 1; j < NFREELISTS; ++j) {
	  q = p -> ptrfree_freelists[j];
	  if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
	  q = p -> normal_freelists[j];
	  if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
#	  ifdef GC_GCJ_SUPPORT
	    q = p -> gcj_freelists[j];
	    if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
#	  endif /* GC_GCJ_SUPPORT */
	}
      }
    }
}
#endif /* THREAD_LOCAL_ALLOC */

static struct GC_Thread_Rep first_thread;

/* Add a thread to GC_threads.  We assume it wasn't already there.	*/
/* Caller holds allocation lock.					*/
GC_thread GC_new_thread(pthread_t id)
{
    int hv = ((word)id) % THREAD_TABLE_SZ;
    GC_thread result;
    static GC_bool first_thread_used = FALSE;
    
    if (!first_thread_used) {
    	result = &first_thread;
    	first_thread_used = TRUE;
    } else {
        result = (struct GC_Thread_Rep *)
        	 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
    }
    if (result == 0) return(0);
    result -> id = id;
    result -> next = GC_threads[hv];
    GC_threads[hv] = result;
    GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
    return(result);
}

/* Delete a thread from GC_threads.  We assume it is there.	*/
/* (The code intentionally traps if it wasn't.)			*/
/* Caller holds allocation lock.				*/
void GC_delete_thread(pthread_t id)
{
    int hv = ((word)id) % THREAD_TABLE_SZ;
    register GC_thread p = GC_threads[hv];
    register GC_thread prev = 0;
    
    while (!pthread_equal(p -> id, id)) {
        prev = p;
        p = p -> next;
    }
    if (prev == 0) {
        GC_threads[hv] = p -> next;
    } else {
        prev -> next = p -> next;
    }

    if (p != &first_thread)
      GC_INTERNAL_FREE(p);
}

/* If a thread has been joined, but we have not yet		*/
/* been notified, then there may be more than one thread 	*/
/* in the table with the same pthread id.			*/
/* This is OK, but we need a way to delete a specific one.	*/
void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
{
    int hv = ((word)id) % THREAD_TABLE_SZ;
    register GC_thread p = GC_threads[hv];
    register GC_thread prev = 0;

    while (p != gc_id) {
        prev = p;
        p = p -> next;
    }
    if (prev == 0) {
        GC_threads[hv] = p -> next;
    } else {
        prev -> next = p -> next;
    }
    GC_INTERNAL_FREE(p);
}

/* Return a GC_thread corresponding to a given pthread_t.	*/
/* Returns 0 if it's not there.					*/
/* Caller holds  allocation lock or otherwise inhibits 		*/
/* updates.							*/
/* If there is more than one thread with the given id we 	*/
/* return the most recent one.					*/
GC_thread GC_lookup_thread(pthread_t id)
{
    int hv = ((word)id) % THREAD_TABLE_SZ;
    register GC_thread p = GC_threads[hv];
    
    while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
    return(p);
}

#ifdef HANDLE_FORK
/* Remove all entries from the GC_threads table, except the	*/
/* one for the current thread.  We need to do this in the child	*/
/* process after a fork(), since only the current thread 	*/
/* survives in the child.					*/
void GC_remove_all_threads_but_me(void)
{
    pthread_t self = pthread_self();
    int hv;
    GC_thread p, next, me;

    for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
      me = 0;
      for (p = GC_threads[hv]; 0 != p; p = next) {
	next = p -> next;
	if (p -> id == self) {
	  me = p;
	  p -> next = 0;
	} else {
#	  ifdef THREAD_LOCAL_ALLOC
	    if (!(p -> flags & FINISHED)) {
	      GC_destroy_thread_local(p);
	    }
#	  endif /* THREAD_LOCAL_ALLOC */
	  if (p != &first_thread) GC_INTERNAL_FREE(p);
	}
      }
      GC_threads[hv] = me;
    }
}
#endif /* HANDLE_FORK */

#ifdef USE_PROC_FOR_LIBRARIES
int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
{
    int i;
    GC_thread p;
    
#   ifdef PARALLEL_MARK
      for (i = 0; i < GC_markers; ++i) {
	if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
      }
#   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 1;
#	  else /* STACK_GROWS_DOWN */
            if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
#	  endif
	}
      }
    }
    return 0;
}
#endif /* USE_PROC_FOR_LIBRARIES */

#ifdef GC_LINUX_THREADS
/* Return the number of processors, or i<= 0 if it can't be determined.	*/
int GC_get_nprocs()
{
    /* 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 GC_wait_for_gc_completion(GC_bool wait_for_all)
{
    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()
{
    /* <takis@XFree86.Org> */
    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_printf1("Number of active CPUs in this system: %d\n", numCpus);
#  endif
    return(numCpus);
}
#endif /* GC_DGUX386_THREADS */

/* We hold the allocation lock.	*/
void GC_thr_init()
{
#   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 */
    /* 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_PTHREADS)
	  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_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 (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
#     ifdef CONDPRINT
        if (GC_print_stats) {
          GC_printf2("Number of processors = %ld, "
		 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
	}
#     endif
      if (GC_markers == 1) {
	GC_parallel = FALSE;
#	ifdef CONDPRINT
	  if (GC_print_stats) {
	    GC_printf0("Single marker thread, turning off parallel marking\n");
	  }
#	endif
      } 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.	*/
/* Called without allocation lock.			*/
void GC_init_parallel()
{
    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) && !defined(DBG_HDRS_ALL)
      LOCK();
      GC_init_thread_local(GC_lookup_thread(pthread_self()));
      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;
    
    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 */

/* Wrappers for functions that are likely to block for an appreciable	*/
/* length of time.  Must be called in pairs, if at all.			*/
/* Nothing much beyond the system call itself should be executed	*/
/* between these.							*/

void GC_start_blocking(void) {
#   define SP_SLOP 128
    GC_thread me;
    LOCK();
    me = GC_lookup_thread(pthread_self());
    GC_ASSERT(!(me -> thread_blocked));
#   ifdef SPARC
	me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
#   else
#   ifndef GC_DARWIN_THREADS
	me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
#   endif
#   endif