root/gc/pthread_support.c

DEFINITIONS

1. return_freelists
2. GC_init_thread_local
3. GC_destroy_thread_local
4. GC_local_malloc
5. GC_local_malloc_atomic
6. GC_local_gcj_malloc
7. GC_mark_thread
8. start_mark_threads
9. start_mark_threads
10. GC_PROTO
11. GC_mark_thread_local_free_lists
12. GC_new_thread
13. GC_delete_thread
14. GC_delete_gc_thread
15. GC_lookup_thread
16. GC_remove_all_threads_but_me
17. GC_segment_is_thread_stack
18. GC_get_nprocs
19. GC_wait_for_gc_completion
20. GC_fork_prepare_proc
21. GC_fork_parent_proc
22. GC_fork_child_proc
23. GC_get_nprocs
24. GC_thr_init
25. GC_init_parallel
26. GC_start_blocking
27. GC_end_blocking
28. WRAP_FUNC
29. GC_thread_exit_proc
30. WRAP_FUNC
31. GC_start_routine
32. GC_compare_and_exchange
33. GC_atomic_add
34. GC_pause
35. GC_generic_lock
36. GC_lock
37. GC_lock
38. GC_acquire_mark_lock
39. GC_release_mark_lock
40. GC_wait_builder
41. GC_wait_for_reclaim
42. GC_notify_all_builder
43. GC_wait_marker
44. GC_notify_all_marker

   1 /* 
   2  * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
   3  * Copyright (c) 1996 by Silicon Graphics.  All rights reserved.
   4  * Copyright (c) 1998 by Fergus Henderson.  All rights reserved.
   5  * Copyright (c) 2000-2004 by Hewlett-Packard Company.  All rights reserved.
   6  *
   7  * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
   8  * OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
   9  *
  10  * Permission is hereby granted to use or copy this program
  11  * for any purpose,  provided the above notices are retained on all copies.
  12  * Permission to modify the code and to distribute modified code is granted,
  13  * provided the above notices are retained, and a notice that the code was
  14  * modified is included with the above copyright notice.
  15  */
  16 /*
  17  * Support code for LinuxThreads, the clone()-based kernel
  18  * thread package for Linux which is included in libc6.
  19  *
  20  * This code relies on implementation details of LinuxThreads,
  21  * (i.e. properties not guaranteed by the Pthread standard),
  22  * though this version now does less of that than the other Pthreads
  23  * support code.
  24  *
  25  * Note that there is a lot of code duplication between linux_threads.c
  26  * and thread support for some of the other Posix platforms; any changes
  27  * made here may need to be reflected there too.
  28  */
  29  /* DG/UX ix86 support <takis@xfree86.org> */
  30 /*
  31  * Linux_threads.c now also includes some code to support HPUX and
  32  * OSF1 (Compaq Tru64 Unix, really).  The OSF1 support is based on Eric Benson's
  33  * patch.
  34  *
  35  * Eric also suggested an alternate basis for a lock implementation in
  36  * his code:
  37  * + #elif defined(OSF1)
  38  * +    unsigned long GC_allocate_lock = 0;
  39  * +    msemaphore GC_allocate_semaphore;
  40  * + #  define GC_TRY_LOCK() \
  41  * +    ((msem_lock(&GC_allocate_semaphore, MSEM_IF_NOWAIT) == 0) \
  42  * +     ? (GC_allocate_lock = 1) \
  43  * +     : 0)
  44  * + #  define GC_LOCK_TAKEN GC_allocate_lock
  45  */
  47 /*#define DEBUG_THREADS 1*/
  48 /*#define GC_ASSERTIONS*/
  50 # include "private/pthread_support.h"
  52 # if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
  53      && !defined(GC_WIN32_THREADS)
  55 # if defined(GC_HPUX_THREADS) && !defined(USE_PTHREAD_SPECIFIC) \
  56      && !defined(USE_COMPILER_TLS)
  57 #   ifdef __GNUC__
  58 #     define USE_PTHREAD_SPECIFIC
  59       /* Empirically, as of gcc 3.3, USE_COMPILER_TLS doesn't work.     */
  60 #   else
  61 #     define USE_COMPILER_TLS
  62 #   endif
  63 # endif
  65 # if defined USE_HPUX_TLS
  66     --> Macro replaced by USE_COMPILER_TLS
  67 # endif
  69 # if (defined(GC_DGUX386_THREADS) || defined(GC_OSF1_THREADS) || \
  70       defined(GC_DARWIN_THREADS) || defined(GC_AIX_THREADS)) \
  71       && !defined(USE_PTHREAD_SPECIFIC)
  72 #   define USE_PTHREAD_SPECIFIC
  73 # endif
  75 # if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
  76 #   define _POSIX4A_DRAFT10_SOURCE 1
  77 # endif
  79 # if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
  80 #   define _USING_POSIX4A_DRAFT10 1
  81 # endif
  83 # ifdef THREAD_LOCAL_ALLOC
  84 #   if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
  85 #     include "private/specific.h"
  86 #   endif
  87 #   if defined(USE_PTHREAD_SPECIFIC)
  88 #     define GC_getspecific pthread_getspecific
  89 #     define GC_setspecific pthread_setspecific
  90 #     define GC_key_create pthread_key_create
  91       typedef pthread_key_t GC_key_t;
  92 #   endif
  93 #   if defined(USE_COMPILER_TLS)
  94 #     define GC_getspecific(x) (x)
  95 #     define GC_setspecific(key, v) ((key) = (v), 0)
  96 #     define GC_key_create(key, d) 0
  97       typedef void * GC_key_t;
  98 #   endif
  99 # endif
 100 # include <stdlib.h>
 101 # include <pthread.h>
 102 # include <sched.h>
 103 # include <time.h>
 104 # include <errno.h>
 105 # include <unistd.h>
 106 # include <sys/mman.h>
 107 # include <sys/time.h>
 108 # include <sys/types.h>
 109 # include <sys/stat.h>
 110 # include <fcntl.h>
 111 # include <signal.h>
 113 #if defined(GC_DARWIN_THREADS)
 114 # include "private/darwin_semaphore.h"
 115 #else
 116 # include <semaphore.h>
 117 #endif /* !GC_DARWIN_THREADS */
 119 #if defined(GC_DARWIN_THREADS)
 120 # include <sys/sysctl.h>
 121 #endif /* GC_DARWIN_THREADS */
 125 #if defined(GC_DGUX386_THREADS)
 126 # include <sys/dg_sys_info.h>
 127 # include <sys/_int_psem.h>
 128   /* sem_t is an uint in DG/UX */
 129   typedef unsigned int  sem_t;
 130 #endif /* GC_DGUX386_THREADS */
 132 #ifndef __GNUC__
 133 #   define __inline__
 134 #endif
 136 #ifdef GC_USE_LD_WRAP
 137 #   define WRAP_FUNC(f) __wrap_##f
 138 #   define REAL_FUNC(f) __real_##f
 139 #else
 140 #   define WRAP_FUNC(f) GC_##f
 141 #   if !defined(GC_DGUX386_THREADS)
 142 #     define REAL_FUNC(f) f
 143 #   else /* GC_DGUX386_THREADS */
 144 #     define REAL_FUNC(f) __d10_##f
 145 #   endif /* GC_DGUX386_THREADS */
 146 #   undef pthread_create
 147 #   if !defined(GC_DARWIN_THREADS)
 148 #     undef pthread_sigmask
 149 #   endif
 150 #   undef pthread_join
 151 #   undef pthread_detach
 152 #   if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
 153        && !defined(_PTHREAD_USE_PTDNAM_)
 154 /* Restore the original mangled names on Tru64 UNIX.  */
 155 #     define pthread_create __pthread_create
 156 #     define pthread_join __pthread_join
 157 #     define pthread_detach __pthread_detach
 158 #   endif
 159 #endif
 161 void GC_thr_init();
 163 static GC_bool parallel_initialized = FALSE;
 165 void GC_init_parallel();
 167 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
 169 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
 171 #ifdef USE_COMPILER_TLS
 172   __thread
 173 #endif
GC_key_t GC_thread_key;
 176 static GC_bool keys_initialized;
 178 /* Recover the contents of the freelist array fl into the global one gfl.*/
 179 /* Note that the indexing scheme differs, in that gfl has finer size    */
 180 /* resolution, even if not all entries are used.                        */
 181 /* We hold the allocator lock.                                          */
 182 static void return_freelists(ptr_t *fl, ptr_t *gfl)
 183 {
 184     int i;
ptr_t q, *qptr;
size_t nwords;
 188     for (i = 1; i < NFREELISTS; ++i) {
nwords = i * (GRANULARITY/sizeof(word));
qptr = fl + i;  
q = *qptr;
 192         if ((word)q >= HBLKSIZE) {
 193           if (gfl[nwords] == 0) {
gfl[nwords] = q;
 195           } else {
 196             /* Concatenate: */
 197             for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
GC_ASSERT(0 == q);
 199             *qptr = gfl[nwords];
gfl[nwords] = fl[i];
 201           }
 202         }
 203         /* Clear fl[i], since the thread structure may hang around.     */
 204         /* Do it in a way that is likely to trap if we access it.       */
fl[i] = (ptr_t)HBLKSIZE;
 206     }
 207 }
 209 /* We statically allocate a single "size 0" object. It is linked to     */
 210 /* itself, and is thus repeatedly reused for all size 0 allocation      */
 211 /* requests.  (Size 0 gcj allocation requests are incorrect, and        */
 212 /* we arrange for those to fault asap.)                                 */
 213 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
 215 /* Each thread structure must be initialized.   */
 216 /* This call must be made from the new thread.  */
 217 /* Caller holds allocation lock.                */
 218 void GC_init_thread_local(GC_thread p)
 219 {
 220     int i;
 222     if (!keys_initialized) {
 223         if (0 != GC_key_create(&GC_thread_key, 0)) {
ABORT("Failed to create key for local allocator");
 225         }
keys_initialized = TRUE;
 227     }
 228     if (0 != GC_setspecific(GC_thread_key, p)) {
ABORT("Failed to set thread specific allocation pointers");
 230     }
 231     for (i = 1; i < NFREELISTS; ++i) {
p -> ptrfree_freelists[i] = (ptr_t)1;
p -> normal_freelists[i] = (ptr_t)1;
 234 #       ifdef GC_GCJ_SUPPORT
p -> gcj_freelists[i] = (ptr_t)1;
 236 #       endif
 237     }   
 238     /* 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);
 241 #   ifdef GC_GCJ_SUPPORT
p -> gcj_freelists[0] = (ptr_t)(-1);
 243 #   endif
 244 }
 246 #ifdef GC_GCJ_SUPPORT
 247   extern ptr_t * GC_gcjobjfreelist;
 248 #endif
 250 /* We hold the allocator lock.  */
 251 void GC_destroy_thread_local(GC_thread p)
 252 {
 253     /* We currently only do this from the thread itself or from */
 254     /* the fork handler for a child process.                    */
 255 #   ifndef HANDLE_FORK
GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
 257 #   endif
return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
return_freelists(p -> normal_freelists, GC_objfreelist);
 260 #   ifdef GC_GCJ_SUPPORT
return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
 262 #   endif
 263 }
 265 extern GC_PTR GC_generic_malloc_many();
GC_PTR GC_local_malloc(size_t bytes)
 268 {
 269     if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
 270         return(GC_malloc(bytes));
 271     } else {
 272         int index = INDEX_FROM_BYTES(bytes);
ptr_t * my_fl;
ptr_t my_entry;
 275 #       if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
GC_key_t k = GC_thread_key;
 277 #       endif
 278         void * tsd;
 280 #       if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
 281             if (EXPECT(0 == k, 0)) {
 282                 /* This can happen if we get called when the world is   */
 283                 /* being initialized.  Whether we can actually complete */
 284                 /* the initialization then is unclear.                  */
GC_init_parallel();
k = GC_thread_key;
 287             }
 288 #       endif
tsd = GC_getspecific(GC_thread_key);
 290 #       ifdef GC_ASSERTIONS
LOCK();
GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
UNLOCK();
 294 #       endif
my_fl = ((GC_thread)tsd) -> normal_freelists + index;
my_entry = *my_fl;
 297         if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
ptr_t next = obj_link(my_entry);
GC_PTR result = (GC_PTR)my_entry;
 300             *my_fl = next;
obj_link(my_entry) = 0;
PREFETCH_FOR_WRITE(next);
 303             return result;
 304         } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
 305             *my_fl = my_entry + index + 1;
 306             return GC_malloc(bytes);
 307         } else {
GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
 309             if (*my_fl == 0) return GC_oom_fn(bytes);
 310             return GC_local_malloc(bytes);
 311         }
 312     }
 313 }
GC_PTR GC_local_malloc_atomic(size_t bytes)
 316 {
 317     if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
 318         return(GC_malloc_atomic(bytes));
 319     } else {
 320         int index = INDEX_FROM_BYTES(bytes);
ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
 322                         -> ptrfree_freelists + index;
ptr_t my_entry = *my_fl;
 325         if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
GC_PTR result = (GC_PTR)my_entry;
 327             *my_fl = obj_link(my_entry);
 328             return result;
 329         } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
 330             *my_fl = my_entry + index + 1;
 331         return GC_malloc_atomic(bytes);
 332         } else {
GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
 334             /* *my_fl is updated while the collector is excluded;       */
 335             /* the free list is always visible to the collector as      */
 336             /* such.                                                    */
 337             if (*my_fl == 0) return GC_oom_fn(bytes);
 338             return GC_local_malloc_atomic(bytes);
 339         }
 340     }
 341 }
 343 #ifdef GC_GCJ_SUPPORT
 345 #include "include/gc_gcj.h"
 347 #ifdef GC_ASSERTIONS
 348   extern GC_bool GC_gcj_malloc_initialized;
 349 #endif
 351 extern int GC_gcj_kind;
GC_PTR GC_local_gcj_malloc(size_t bytes,
 354                            void * ptr_to_struct_containing_descr)
 355 {
GC_ASSERT(GC_gcj_malloc_initialized);
 357     if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
 358         return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
 359     } else {
 360         int index = INDEX_FROM_BYTES(bytes);
ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
 362                         -> gcj_freelists + index;
ptr_t my_entry = *my_fl;
 364         if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
GC_PTR result = (GC_PTR)my_entry;
GC_ASSERT(!GC_incremental);
 367             /* We assert that any concurrent marker will stop us.       */
 368             /* Thus it is impossible for a mark procedure to see the    */
 369             /* allocation of the next object, but to see this object    */
 370             /* still containing a free list pointer.  Otherwise the     */
 371             /* marker might find a random "mark descriptor".            */
 372             *(volatile ptr_t *)my_fl = obj_link(my_entry);
 373             /* We must update the freelist before we store the pointer. */
 374             /* Otherwise a GC at this point would see a corrupted       */
 375             /* free list.                                               */
 376             /* A memory barrier is probably never needed, since the     */
 377             /* action of stopping this thread will cause prior writes   */
 378             /* to complete.                                             */
GC_ASSERT(((void * volatile *)result)[1] == 0); 
 380             *(void * volatile *)result = ptr_to_struct_containing_descr; 
 381             return result;
 382         } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
 383             if (!GC_incremental) *my_fl = my_entry + index + 1;
 384                 /* In the incremental case, we always have to take this */
 385                 /* path.  Thus we leave the counter alone.              */
 386             return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
 387         } else {
GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
 389             if (*my_fl == 0) return GC_oom_fn(bytes);
 390             return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
 391         }
 392     }
 393 }
 395 #endif /* GC_GCJ_SUPPORT */
 397 # else  /* !THREAD_LOCAL_ALLOC  && !DBG_HDRS_ALL */
 399 #   define GC_destroy_thread_local(t)
 401 # endif /* !THREAD_LOCAL_ALLOC */
 403 #if 0
 404 /*
 405 To make sure that we're using LinuxThreads and not some other thread
 406 package, we generate a dummy reference to `pthread_kill_other_threads_np'
 407 (was `__pthread_initial_thread_bos' but that disappeared),
 408 which is a symbol defined in LinuxThreads, but (hopefully) not in other
 409 thread packages.
 410 
 411 We no longer do this, since this code is now portable enough that it might
 412 actually work for something else.
 413 */
 414 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
 415 #endif /* 0 */
 417 long GC_nprocs = 1;     /* Number of processors.  We may not have       */
 418                         /* access to all of them, but this is as good   */
 419                         /* a guess as any ...                           */
 421 #ifdef PARALLEL_MARK
 423 # ifndef MAX_MARKERS
 424 #   define MAX_MARKERS 16
 425 # endif
 427 static ptr_t marker_sp[MAX_MARKERS] = {0};
 429 void * GC_mark_thread(void * id)
 430 {
word my_mark_no = 0;
marker_sp[(word)id] = GC_approx_sp();
 434   for (;; ++my_mark_no) {
 435     /* GC_mark_no is passed only to allow GC_help_marker to terminate   */
 436     /* promptly.  This is important if it were called from the signal   */
 437     /* handler or from the GC lock acquisition code.  Under Linux, it's */
 438     /* not safe to call it from a signal handler, since it uses mutexes */
 439     /* and condition variables.  Since it is called only here, the      */
 440     /* argument is unnecessary.                                         */
 441     if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
 442         /* resynchronize if we get far off, e.g. because GC_mark_no     */
 443         /* wrapped.                                                     */
my_mark_no = GC_mark_no;
 445     }
 446 #   ifdef DEBUG_THREADS
GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
 448 #   endif
GC_help_marker(my_mark_no);
 450   }
 451 }
 453 extern long GC_markers;         /* Number of mark threads we would      */
 454                                 /* like to have.  Includes the          */
 455                                 /* initiating thread.                   */
pthread_t GC_mark_threads[MAX_MARKERS];
 459 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
 461 static void start_mark_threads()
 462 {
 463     unsigned i;
pthread_attr_t attr;
 466     if (GC_markers > MAX_MARKERS) {
WARN("Limiting number of mark threads\n", 0);
GC_markers = MAX_MARKERS;
 469     }
 470     if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
 472     if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
ABORT("pthread_attr_setdetachstate failed");
 475 #   if defined(HPUX) || defined(GC_DGUX386_THREADS)
 476       /* Default stack size is usually too small: fix it. */
 477       /* Otherwise marker threads or GC may run out of    */
 478       /* space.                                           */
 479 #     define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
 480       {
size_t old_size;
 482         int code;
 484         if (pthread_attr_getstacksize(&attr, &old_size) != 0)
ABORT("pthread_attr_getstacksize failed\n");
 486         if (old_size < MIN_STACK_SIZE) {
 487           if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
ABORT("pthread_attr_setstacksize failed\n");
 489         }
 490       }
 491 #   endif /* HPUX || GC_DGUX386_THREADS */
 492 #   ifdef CONDPRINT
 493       if (GC_print_stats) {
GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
 495       }
 496 #   endif
 497     for (i = 0; i < GC_markers - 1; ++i) {
 498       if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
GC_mark_thread, (void *)(word)i)) {
WARN("Marker thread creation failed, errno = %ld.\n", errno);
 501       }
 502     }
 503 }
 505 #else  /* !PARALLEL_MARK */
 507 static __inline__ void start_mark_threads()
 508 {
 509 }
 511 #endif /* !PARALLEL_MARK */
GC_bool GC_thr_initialized = FALSE;
 515 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
 517 void GC_push_thread_structures GC_PROTO((void))
 518 {
GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
 520 #   if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
GC_push_all((ptr_t)(&GC_thread_key),
 522           (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
 523 #   endif
 524 }
 526 #ifdef THREAD_LOCAL_ALLOC
 527 /* We must explicitly mark ptrfree and gcj free lists, since the free   */
 528 /* list links wouldn't otherwise be found.  We also set them in the     */
 529 /* normal free lists, since that involves touching less memory than if  */
 530 /* we scanned them normally.                                            */
 531 void GC_mark_thread_local_free_lists(void)
 532 {
 533     int i, j;
GC_thread p;
ptr_t q;
 537     for (i = 0; i < THREAD_TABLE_SZ; ++i) {
 538       for (p = GC_threads[i]; 0 != p; p = p -> next) {
 539         for (j = 1; j < NFREELISTS; ++j) {
q = p -> ptrfree_freelists[j];
 541           if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
q = p -> normal_freelists[j];
 543           if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
 544 #         ifdef GC_GCJ_SUPPORT
q = p -> gcj_freelists[j];
 546             if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
 547 #         endif /* GC_GCJ_SUPPORT */
 548         }
 549       }
 550     }
 551 }
 552 #endif /* THREAD_LOCAL_ALLOC */
 554 static struct GC_Thread_Rep first_thread;
 556 /* Add a thread to GC_threads.  We assume it wasn't already there.      */
 557 /* Caller holds allocation lock.                                        */
GC_thread GC_new_thread(pthread_t id)
 559 {
 560     int hv = ((word)id) % THREAD_TABLE_SZ;
GC_thread result;
 562     static GC_bool first_thread_used = FALSE;
 564     if (!first_thread_used) {
result = &first_thread;
first_thread_used = TRUE;
 567     } else {
result = (struct GC_Thread_Rep *)
GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
 570     }
 571     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);
 576     return(result);
 577 }
 579 /* Delete a thread from GC_threads.  We assume it is there.     */
 580 /* (The code intentionally traps if it wasn't.)                 */
 581 /* Caller holds allocation lock.                                */
 582 void GC_delete_thread(pthread_t id)
 583 {
 584     int hv = ((word)id) % THREAD_TABLE_SZ;
 585     register GC_thread p = GC_threads[hv];
 586     register GC_thread prev = 0;
 588     while (!pthread_equal(p -> id, id)) {
prev = p;
p = p -> next;
 591     }
 592     if (prev == 0) {
GC_threads[hv] = p -> next;
 594     } else {
prev -> next = p -> next;
 596     }
GC_INTERNAL_FREE(p);
 598 }
 600 /* If a thread has been joined, but we have not yet             */
 601 /* been notified, then there may be more than one thread        */
 602 /* in the table with the same pthread id.                       */
 603 /* This is OK, but we need a way to delete a specific one.      */
 604 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
 605 {
 606     int hv = ((word)id) % THREAD_TABLE_SZ;
 607     register GC_thread p = GC_threads[hv];
 608     register GC_thread prev = 0;
 610     while (p != gc_id) {
prev = p;
p = p -> next;
 613     }
 614     if (prev == 0) {
GC_threads[hv] = p -> next;
 616     } else {
prev -> next = p -> next;
 618     }
GC_INTERNAL_FREE(p);
 620 }
 622 /* Return a GC_thread corresponding to a given pthread_t.       */
 623 /* Returns 0 if it's not there.                                 */
 624 /* Caller holds  allocation lock or otherwise inhibits          */
 625 /* updates.                                                     */
 626 /* If there is more than one thread with the given id we        */
 627 /* return the most recent one.                                  */
GC_thread GC_lookup_thread(pthread_t id)
 629 {
 630     int hv = ((word)id) % THREAD_TABLE_SZ;
 631     register GC_thread p = GC_threads[hv];
 633     while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
 634     return(p);
 635 }
 637 #ifdef HANDLE_FORK
 638 /* Remove all entries from the GC_threads table, except the     */
 639 /* one for the current thread.  We need to do this in the child */
 640 /* process after a fork(), since only the current thread        */
 641 /* survives in the child.                                       */
 642 void GC_remove_all_threads_but_me(void)
 643 {
pthread_t self = pthread_self();
 645     int hv;
GC_thread p, next, me;
 648     for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
me = 0;
 650       for (p = GC_threads[hv]; 0 != p; p = next) {
next = p -> next;
 652         if (p -> id == self) {
me = p;
p -> next = 0;
 655         } else {
 656 #         ifdef THREAD_LOCAL_ALLOC
 657             if (!(p -> flags & FINISHED)) {
GC_destroy_thread_local(p);
 659             }
 660 #         endif /* THREAD_LOCAL_ALLOC */
 661           if (p != &first_thread) GC_INTERNAL_FREE(p);
 662         }
 663       }
GC_threads[hv] = me;
 665     }
 666 }
 667 #endif /* HANDLE_FORK */
 669 #ifdef USE_PROC_FOR_LIBRARIES
 670 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
 671 {
 672     int i;
GC_thread p;
 675 #   ifdef PARALLEL_MARK
 676       for (i = 0; i < GC_markers; ++i) {
 677         if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
 678       }
 679 #   endif
 680     for (i = 0; i < THREAD_TABLE_SZ; i++) {
 681       for (p = GC_threads[i]; p != 0; p = p -> next) {
 682         if (0 != p -> stack_end) {
 683 #         ifdef STACK_GROWS_UP
 684             if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
 685 #         else /* STACK_GROWS_DOWN */
 686             if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
 687 #         endif
 688         }
 689       }
 690     }
 691     return 0;
 692 }
 693 #endif /* USE_PROC_FOR_LIBRARIES */
 695 #ifdef GC_LINUX_THREADS
 696 /* Return the number of processors, or i<= 0 if it can't be determined. */
 697 int GC_get_nprocs()
 698 {
 699     /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that       */
 700     /* appears to be buggy in many cases.                               */
 701     /* We look for lines "cpu<n>" in /proc/stat.                        */
 702 #   define STAT_BUF_SIZE 4096
 703 #   define STAT_READ read
 704         /* If read is wrapped, this may need to be redefined to call    */
 705         /* the real one.                                                */
 706     char stat_buf[STAT_BUF_SIZE];
 707     int f;
word result = 1;
 709         /* Some old kernels only have a single "cpu nnnn ..."   */
 710         /* entry in /proc/stat.  We identify those as           */
 711         /* uniprocessors.                                       */
size_t i, len = 0;
f = open("/proc/stat", O_RDONLY);
 715     if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
WARN("Couldn't read /proc/stat\n", 0);
 717         return -1;
 718     }
 719     for (i = 0; i < len - 100; ++i) {
 720         if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
 721             && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
 722             int cpu_no = atoi(stat_buf + i + 4);
 723             if (cpu_no >= result) result = cpu_no + 1;
 724         }
 725     }
close(f);
 727     return result;
 728 }
 729 #endif /* GC_LINUX_THREADS */
 731 /* We hold the GC lock.  Wait until an in-progress GC has finished.     */
 732 /* Repeatedly RELEASES GC LOCK in order to wait.                        */
 733 /* If wait_for_all is true, then we exit with the GC lock held and no   */
 734 /* collection in progress; otherwise we just wait for the current GC    */
 735 /* to finish.                                                           */
 736 extern GC_bool GC_collection_in_progress();
 737 void GC_wait_for_gc_completion(GC_bool wait_for_all)
 738 {
 739     if (GC_incremental && GC_collection_in_progress()) {
 740         int old_gc_no = GC_gc_no;
 742         /* Make sure that no part of our stack is still on the mark stack, */
 743         /* since it's about to be unmapped.                                */
 744         while (GC_incremental && GC_collection_in_progress()
 745                && (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();
 754         }
 755     }
 756 }
 758 #ifdef HANDLE_FORK
 759 /* Procedures called before and after a fork.  The goal here is to make */
 760 /* it safe to call GC_malloc() in a forked child.  It's unclear that is */
 761 /* attainable, since the single UNIX spec seems to imply that one       */
 762 /* should only call async-signal-safe functions, and we probably can't  */
 763 /* quite guarantee that.  But we give it our best shot.  (That same     */
 764 /* spec also implies that it's not safe to call the system malloc       */
 765 /* between fork() and exec().  Thus we're doing no worse than it.       */
 767 /* Called before a fork()               */
 768 void GC_fork_prepare_proc(void)
 769 {
 770     /* Acquire all relevant locks, so that after releasing the locks    */
 771     /* the child will see a consistent state in which monitor           */
 772     /* invariants hold.  Unfortunately, we can't acquire libc locks     */
 773     /* we might need, and there seems to be no guarantee that libc      */
 774     /* must install a suitable fork handler.                            */
 775     /* Wait for an ongoing GC to finish, since we can't finish it in    */
 776     /* the (one remaining thread in) the child.                         */
LOCK();
 778 #     if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
GC_wait_for_reclaim();
 780 #     endif
GC_wait_for_gc_completion(TRUE);
 782 #     if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
GC_acquire_mark_lock();
 784 #     endif
 785 }
 787 /* Called in parent after a fork()      */
 788 void GC_fork_parent_proc(void)
 789 {
 790 #   if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
GC_release_mark_lock();
 792 #   endif
UNLOCK();
 794 }
 796 /* Called in child after a fork()       */
 797 void GC_fork_child_proc(void)
 798 {
 799     /* Clean up the thread table, so that just our thread is left. */
 800 #   if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
GC_release_mark_lock();
 802 #   endif
GC_remove_all_threads_but_me();
 804 #   ifdef PARALLEL_MARK
 805       /* Turn off parallel marking in the child, since we are probably  */
 806       /* just going to exec, and we would have to restart mark threads. */
GC_markers = 1;
GC_parallel = FALSE;
 809 #   endif /* PARALLEL_MARK */
UNLOCK();
 811 }
 812 #endif /* HANDLE_FORK */
 814 #if defined(GC_DGUX386_THREADS)
 815 /* Return the number of processors, or i<= 0 if it can't be determined. */
 816 int GC_get_nprocs()
 817 {
 818     /* <takis@XFree86.Org> */
 819     int numCpus;
 820     struct dg_sys_info_pm_info pm_sysinfo;
 821     int status =0;
status = dg_sys_info((long int *) &pm_sysinfo,
DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
 825     if (status < 0)
 826        /* set -1 for error */
numCpus = -1;
 828     else
 829       /* Active CPUs */
numCpus = pm_sysinfo.idle_vp_count;
 832 #  ifdef DEBUG_THREADS
GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
 834 #  endif
 835     return(numCpus);
 836 }
 837 #endif /* GC_DGUX386_THREADS */
 839 /* We hold the allocation lock. */
 840 void GC_thr_init()
 841 {
 842 #   ifndef GC_DARWIN_THREADS
 843       int dummy;
 844 #   endif
GC_thread t;
 847     if (GC_thr_initialized) return;
GC_thr_initialized = TRUE;
 850 #   ifdef HANDLE_FORK
 851       /* Prepare for a possible fork.   */
pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
GC_fork_child_proc);
 854 #   endif /* HANDLE_FORK */
 855     /* Add the initial thread, so we can stop it.       */
t = GC_new_thread(pthread_self());
 857 #     ifdef GC_DARWIN_THREADS
t -> stop_info.mach_thread = mach_thread_self();
 859 #     else
t -> stop_info.stack_ptr = (ptr_t)(&dummy);
 861 #     endif
t -> flags = DETACHED | MAIN_THREAD;
GC_stop_init();
 866     /* Set GC_nprocs.  */
 867       {
 868         char * nprocs_string = GETENV("GC_NPROCS");
GC_nprocs = -1;
 870         if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
 871       }
 872       if (GC_nprocs <= 0) {
 873 #       if defined(GC_HPUX_THREADS)
GC_nprocs = pthread_num_processors_np();
 875 #       endif
 876 #       if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
 878           if (GC_nprocs <= 0) GC_nprocs = 1;
 879 #       endif
 880 #       if defined(GC_FREEBSD_THREADS) || defined(GC_IRIX_THREADS)
 881           /* FIXME: For Irix, that's a ridiculous assumption.   */
GC_nprocs = 1;
 883 #       endif
 884 #       if defined(GC_DARWIN_THREADS)
 885           int ncpus = 1;
size_t len = sizeof(ncpus);
sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
GC_nprocs = ncpus;
 889 #       endif
 890 #       if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
GC_nprocs = GC_get_nprocs();
 892 #       endif
 893       }
 894       if (GC_nprocs <= 0) {
WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
GC_nprocs = 2;
 897 #       ifdef PARALLEL_MARK
GC_markers = 1;
 899 #       endif
 900       } else {
 901 #       ifdef PARALLEL_MARK
 902           {
 903             char * markers_string = GETENV("GC_MARKERS");
 904             if (markers_string != NULL) {
GC_markers = atoi(markers_string);
 906             } else {
GC_markers = GC_nprocs;
 908             }
 909           }
 910 #       endif
 911       }
 912 #   ifdef PARALLEL_MARK
 913 #     ifdef CONDPRINT
 914         if (GC_print_stats) {
GC_printf2("Number of processors = %ld, "
 916                  "number of marker threads = %ld\n", GC_nprocs, GC_markers);
 917         }
 918 #     endif
 919       if (GC_markers == 1) {
GC_parallel = FALSE;
 921 #       ifdef CONDPRINT
 922           if (GC_print_stats) {
GC_printf0("Single marker thread, turning off parallel marking\n");
 924           }
 925 #       endif
 926       } else {
GC_parallel = TRUE;
 928         /* Disable true incremental collection, but generational is OK. */
GC_time_limit = GC_TIME_UNLIMITED;
 930       }
 931       /* If we are using a parallel marker, actually start helper threads.  */
 932         if (GC_parallel) start_mark_threads();
 933 #   endif
 934 }
 937 /* Perform all initializations, including those that    */
 938 /* may require allocation.                              */
 939 /* Called without allocation lock.                      */
 940 /* Must be called before a second thread is created.    */
 941 /* Called without allocation lock.                      */
 942 void GC_init_parallel()
 943 {
 944     if (parallel_initialized) return;
parallel_initialized = TRUE;
 947     /* GC_init() calls us back, so set flag first.      */
 948     if (!GC_is_initialized) GC_init();
 949     /* Initialize thread local free lists if used.      */
 950 #   if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
LOCK();
GC_init_thread_local(GC_lookup_thread(pthread_self()));
UNLOCK();
 954 #   endif
 955 }
 958 #if !defined(GC_DARWIN_THREADS)
 959 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
 960 {
sigset_t fudged_set;
 963     if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
fudged_set = *set;
sigdelset(&fudged_set, SIG_SUSPEND);
set = &fudged_set;
 967     }
 968     return(REAL_FUNC(pthread_sigmask)(how, set, oset));
 969 }
 970 #endif /* !GC_DARWIN_THREADS */
 972 /* Wrappers for functions that are likely to block for an appreciable   */
 973 /* length of time.  Must be called in pairs, if at all.                 */
 974 /* Nothing much beyond the system call itself should be executed        */
 975 /* between these.                                                       */
 977 void GC_start_blocking(void) {
 978 #   define SP_SLOP 128
GC_thread me;
LOCK();
me = GC_lookup_thread(pthread_self());
GC_ASSERT(!(me -> thread_blocked));
 983 #   ifdef SPARC
me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
 985 #   else
 986 #   ifndef GC_DARWIN_THREADS
me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
 988 #   endif
 989 #   endif
 990 #   ifdef IA64
me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
 992 #   endif
 993     /* Add some slop to the stack pointer, since the wrapped call may   */
 994     /* end up pushing more callee-save registers.                       */
 995 #   ifndef GC_DARWIN_THREADS
 996 #   ifdef STACK_GROWS_UP
me -> stop_info.stack_ptr += SP_SLOP;
 998 #   else
me -> stop_info.stack_ptr -= SP_SLOP;
1000 #   endif
1001 #   endif
me -> thread_blocked = TRUE;
UNLOCK();
1004 }
1006 void GC_end_blocking(void) {
GC_thread me;
LOCK();   /* This will block if the world is stopped.       */
me = GC_lookup_thread(pthread_self());
GC_ASSERT(me -> thread_blocked);
me -> thread_blocked = FALSE;
UNLOCK();
1013 }
1015 #if defined(GC_DGUX386_THREADS)
1016 #define __d10_sleep sleep
1017 #endif /* GC_DGUX386_THREADS */
1019 /* A wrapper for the standard C sleep function  */
1020 int WRAP_FUNC(sleep) (unsigned int seconds)
1021 {
1022     int result;
GC_start_blocking();
result = REAL_FUNC(sleep)(seconds);
GC_end_blocking();
1027     return result;
1028 }
1030 struct start_info {
1031     void *(*start_routine)(void *);
1032     void *arg;
word flags;
sem_t registered;           /* 1 ==> in our thread table, but       */
1035                                 /* parent hasn't yet noticed.           */
1036 };
1038 /* Called at thread exit.                               */
1039 /* Never called for main thread.  That's OK, since it   */
1040 /* results in at most a tiny one-time leak.  And        */
1041 /* linuxthreads doesn't reclaim the main threads        */
1042 /* resources or id anyway.                              */
1043 void GC_thread_exit_proc(void *arg)
1044 {
GC_thread me;
LOCK();
me = GC_lookup_thread(pthread_self());
GC_destroy_thread_local(me);
1050     if (me -> flags & DETACHED) {
GC_delete_thread(pthread_self());
1052     } else {
me -> flags |= FINISHED;
1054     }
1055 #   if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1056        && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
GC_remove_specific(GC_thread_key);
1058 #   endif
1059     /* The following may run the GC from "nonexistent" thread.  */
GC_wait_for_gc_completion(FALSE);
UNLOCK();
1062 }
1064 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1065 {
1066     int result;
GC_thread thread_gc_id;
LOCK();
thread_gc_id = GC_lookup_thread(thread);
1071     /* This is guaranteed to be the intended one, since the thread id   */
1072     /* cant have been recycled by pthreads.                             */
UNLOCK();
result = REAL_FUNC(pthread_join)(thread, retval);
1075 # if defined (GC_FREEBSD_THREADS)
1076     /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1077        appears to be) a spurious EINTR which caused the test and real code
1078        to gratuitously fail.  Having looked at system pthread library source
1079        code, I see how this return code may be generated.  In one path of
1080        code, pthread_join() just returns the errno setting of the thread
1081        being joined.  This does not match the POSIX specification or the
1082        local man pages thus I have taken the liberty to catch this one
1083        spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1084     if (result == EINTR) result = 0;
1085 # endif
1086     if (result == 0) {
LOCK();
1088         /* Here the pthread thread id may have been recycled. */
GC_delete_gc_thread(thread, thread_gc_id);
UNLOCK();
1091     }
1092     return result;
1093 }
1095 int
WRAP_FUNC(pthread_detach)(pthread_t thread)
1097 {
1098     int result;
GC_thread thread_gc_id;
LOCK();
thread_gc_id = GC_lookup_thread(thread);
UNLOCK();
result = REAL_FUNC(pthread_detach)(thread);
1105     if (result == 0) {
LOCK();
thread_gc_id -> flags |= DETACHED;
1108       /* Here the pthread thread id may have been recycled. */
1109       if (thread_gc_id -> flags & FINISHED) {
GC_delete_gc_thread(thread, thread_gc_id);
1111       }
UNLOCK();
1113     }
1114     return result;
1115 }
GC_bool GC_in_thread_creation = FALSE;
1119 void * GC_start_routine(void * arg)
1120 {
1121     int dummy;
1122     struct start_info * si = arg;
1123     void * result;
GC_thread me;
pthread_t my_pthread;
1126     void *(*start)(void *);
1127     void *start_arg;
my_pthread = pthread_self();
1130 #   ifdef DEBUG_THREADS
GC_printf1("Starting thread 0x%lx\n", my_pthread);
GC_printf1("pid = %ld\n", (long) getpid());
GC_printf1("sp = 0x%lx\n", (long) &arg);
1134 #   endif
LOCK();
GC_in_thread_creation = TRUE;
me = GC_new_thread(my_pthread);
GC_in_thread_creation = FALSE;
1139 #ifdef GC_DARWIN_THREADS
me -> stop_info.mach_thread = mach_thread_self();
1141 #else
me -> stop_info.stack_ptr = 0;
1143 #endif
me -> flags = si -> flags;
1145     /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99)    */
1146     /* doesn't work because the stack base in /proc/self/stat is the    */
1147     /* one for the main thread.  There is a strong argument that that's */
1148     /* a kernel bug, but a pervasive one.                               */
1149 #   ifdef STACK_GROWS_DOWN
me -> stack_end = (ptr_t)(((word)(&dummy) + (GC_page_size - 1))
1151                                 & ~(GC_page_size - 1));
1152 #         ifndef GC_DARWIN_THREADS
me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1154 #         endif
1155         /* Needs to be plausible, since an asynchronous stack mark      */
1156         /* should not crash.                                            */
1157 #   else
me -> stack_end = (ptr_t)((word)(&dummy) & ~(GC_page_size - 1));
me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1160 #   endif
1161     /* This is dubious, since we may be more than a page into the stack, */
1162     /* and hence skip some of it, though it's not clear that matters.    */
1163 #   ifdef IA64
me -> backing_store_end = (ptr_t)
1165                         (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1166       /* This is also < 100% convincing.  We should also read this      */
1167       /* from /proc, but the hook to do so isn't there yet.             */
1168 #   endif /* IA64 */
UNLOCK();
start = si -> start_routine;
1171 #   ifdef DEBUG_THREADS
GC_printf1("start_routine = 0x%lx\n", start);
1173 #   endif
start_arg = si -> arg;
sem_post(&(si -> registered));      /* Last action on si.   */
1176                                         /* OK to deallocate.    */
pthread_cleanup_push(GC_thread_exit_proc, 0);
1178 #   if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
LOCK();
GC_init_thread_local(me);
UNLOCK();
1182 #   endif
result = (*start)(start_arg);
1184 #if DEBUG_THREADS
GC_printf1("Finishing thread 0x%x\n", pthread_self());
1186 #endif
me -> status = result;
pthread_cleanup_pop(1);
1189     /* Cleanup acquires lock, ensuring that we can't exit               */
1190     /* while a collection that thinks we're alive is trying to stop     */
1191     /* us.                                                              */
1192     return(result);
1193 }
1195 int
WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1197                   const pthread_attr_t *attr,
1198                   void *(*start_routine)(void *), void *arg)
1199 {
1200     int result;
1201     int detachstate;
word my_flags = 0;
1203     struct start_info * si; 
1204         /* This is otherwise saved only in an area mmapped by the thread */
1205         /* library, which isn't visible to the collector.                */
1207     /* We resist the temptation to muck with the stack size here,       */
1208     /* even if the default is unreasonably small.  That's the client's  */
1209     /* responsibility.                                                  */
LOCK();
si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
NORMAL);
UNLOCK();
1215     if (!parallel_initialized) GC_init_parallel();
1216     if (0 == si) return(ENOMEM);
sem_init(&(si -> registered), 0, 0);
si -> start_routine = start_routine;
si -> arg = arg;
LOCK();
1221     if (!GC_thr_initialized) GC_thr_init();
1222 #   ifdef GC_ASSERTIONS
1223       {
size_t stack_size;
1225         if (NULL == attr) {
pthread_attr_t my_attr;
pthread_attr_init(&my_attr);
pthread_attr_getstacksize(&my_attr, &stack_size);
1229         } else {
pthread_attr_getstacksize(attr, &stack_size);
1231         }
1232 #       ifdef PARALLEL_MARK
GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1234 #       else
1235           /* FreeBSD-5.3/Alpha: default pthread stack is 64K,   */
1236           /* HBLKSIZE=8192, sizeof(word)=8                      */
GC_ASSERT(stack_size >= 65536);
1238 #       endif
1239         /* Our threads may need to do some work for the GC.     */
1240         /* Ridiculously small threads won't work, and they      */
1241         /* probably wouldn't work anyway.                       */
1242       }
1243 #   endif
1244     if (NULL == attr) {
detachstate = PTHREAD_CREATE_JOINABLE;
1246     } else { 
pthread_attr_getdetachstate(attr, &detachstate);
1248     }
1249     if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
si -> flags = my_flags;
UNLOCK();
1252 #   ifdef DEBUG_THREADS
GC_printf1("About to start new thread from thread 0x%X\n",
pthread_self());
1255 #   endif
result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1259 #   ifdef DEBUG_THREADS
GC_printf1("Started thread 0x%X\n", *new_thread);
1261 #   endif
1262     /* Wait until child has been added to the thread table.             */
1263     /* This also ensures that we hold onto si until the child is done   */
1264     /* with it.  Thus it doesn't matter whether it is otherwise         */
1265     /* visible to the collector.                                        */
1266     if (0 == result) {
1267         while (0 != sem_wait(&(si -> registered))) {
1268             if (EINTR != errno) ABORT("sem_wait failed");
1269         }
1270     }
sem_destroy(&(si -> registered));
LOCK();
GC_INTERNAL_FREE(si);
UNLOCK();
1276     return(result);
1277 }
1279 #ifdef GENERIC_COMPARE_AND_SWAP
pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
GC_bool GC_compare_and_exchange(volatile GC_word *addr,
GC_word old, GC_word new_val)
1284   {
GC_bool result;
pthread_mutex_lock(&GC_compare_and_swap_lock);
1287     if (*addr == old) {
1288       *addr = new_val;
result = TRUE;
1290     } else {
result = FALSE;
1292     }
pthread_mutex_unlock(&GC_compare_and_swap_lock);
1294     return result;
1295   }
GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1298   {
GC_word old;
pthread_mutex_lock(&GC_compare_and_swap_lock);
old = *addr;
1302     *addr = old + how_much;
pthread_mutex_unlock(&GC_compare_and_swap_lock);
1304     return old;
1305   }
1307 #endif /* GENERIC_COMPARE_AND_SWAP */
1308 /* Spend a few cycles in a way that can't introduce contention with     */
1309 /* othre threads.                                                       */
1310 void GC_pause()
1311 {
1312     int i;
1313 #   if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1314       volatile word dummy = 0;
1315 #   endif
1317     for (i = 0; i < 10; ++i) { 
1318 #     if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1319         __asm__ __volatile__ (" " : : : "memory");
1320 #     else
1321         /* Something that's unlikely to be optimized away. */
GC_noop(++dummy);
1323 #     endif
1324     }
1325 }
1327 #define SPIN_MAX 128    /* Maximum number of calls to GC_pause before   */
1328                         /* give up.                                     */
VOLATILE GC_bool GC_collecting = 0;
1331                         /* A hint that we're in the collector and       */
1332                         /* holding the allocation lock for an           */
1333                         /* extended period.                             */
1335 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1336 /* If we don't want to use the below spinlock implementation, either    */
1337 /* because we don't have a GC_test_and_set implementation, or because   */
1338 /* we don't want to risk sleeping, we can still try spinning on         */
1339 /* pthread_mutex_trylock for a while.  This appears to be very          */
1340 /* beneficial in many cases.                                            */
1341 /* I suspect that under high contention this is nearly always better    */
1342 /* than the spin lock.  But it's a bit slower on a uniprocessor.        */
1343 /* Hence we still default to the spin lock.                             */
1344 /* This is also used to acquire the mark lock for the parallel          */
1345 /* marker.                                                              */
1347 /* Here we use a strict exponential backoff scheme.  I don't know       */
1348 /* whether that's better or worse than the above.  We eventually        */
1349 /* yield by calling pthread_mutex_lock(); it never makes sense to       */
1350 /* explicitly sleep.                                                    */
1352 #define LOCK_STATS
1353 #ifdef LOCK_STATS
1354   unsigned long GC_spin_count = 0;
1355   unsigned long GC_block_count = 0;
1356   unsigned long GC_unlocked_count = 0;
1357 #endif
1359 void GC_generic_lock(pthread_mutex_t * lock)
1360 {
1361 #ifndef NO_PTHREAD_TRYLOCK
1362     unsigned pause_length = 1;
1363     unsigned i;
1365     if (0 == pthread_mutex_trylock(lock)) {
1366 #       ifdef LOCK_STATS
1367             ++GC_unlocked_count;
1368 #       endif
1369         return;
1370     }
1371     for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1372         for (i = 0; i < pause_length; ++i) {
GC_pause();
1374         }
1375         switch(pthread_mutex_trylock(lock)) {
1376             case 0:
1377 #               ifdef LOCK_STATS
1378                     ++GC_spin_count;
1379 #               endif
1380                 return;
1381             case EBUSY:
1382                 break;
1383             default:
ABORT("Unexpected error from pthread_mutex_trylock");
1385         }
1386     }
1387 #endif /* !NO_PTHREAD_TRYLOCK */
1388 #   ifdef LOCK_STATS
1389         ++GC_block_count;
1390 #   endif
pthread_mutex_lock(lock);
1392 }
1394 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1396 #if defined(USE_SPIN_LOCK)
1398 /* Reasonably fast spin locks.  Basically the same implementation */
1399 /* as STL alloc.h.  This isn't really the right way to do this.   */
1400 /* but until the POSIX scheduling mess gets straightened out ...  */
1402 volatile unsigned int GC_allocate_lock = 0;
1405 void GC_lock()
1406 {
1407 #   define low_spin_max 30  /* spin cycles if we suspect uniprocessor */
1408 #   define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1409     static unsigned spin_max = low_spin_max;
1410     unsigned my_spin_max;
1411     static unsigned last_spins = 0;
1412     unsigned my_last_spins;
1413     int i;
1415     if (!GC_test_and_set(&GC_allocate_lock)) {
1416         return;
1417     }
my_spin_max = spin_max;
my_last_spins = last_spins;
1420     for (i = 0; i < my_spin_max; i++) {
1421         if (GC_collecting || GC_nprocs == 1) goto yield;
1422         if (i < my_last_spins/2 || GC_allocate_lock) {
GC_pause();
1424             continue;
1425         }
1426         if (!GC_test_and_set(&GC_allocate_lock)) {
1427             /*
1428              * got it!
1429              * Spinning worked.  Thus we're probably not being scheduled
1430              * against the other process with which we were contending.
1431              * Thus it makes sense to spin longer the next time.
1432              */
last_spins = i;
spin_max = high_spin_max;
1435             return;
1436         }
1437     }
1438     /* We are probably being scheduled against the other process.  Sleep. */
spin_max = low_spin_max;
yield:
1441     for (i = 0;; ++i) {
1442         if (!GC_test_and_set(&GC_allocate_lock)) {
1443             return;
1444         }
1445 #       define SLEEP_THRESHOLD 12
1446                 /* Under Linux very short sleeps tend to wait until     */
1447                 /* the current time quantum expires.  On old Linux      */
1448                 /* kernels nanosleep(<= 2ms) just spins under Linux.    */
1449                 /* (Under 2.4, this happens only for real-time          */
1450                 /* processes.)  We want to minimize both behaviors      */
1451                 /* here.                                                */
1452         if (i < SLEEP_THRESHOLD) {
sched_yield();
1454         } else {
1455             struct timespec ts;
1457             if (i > 24) i = 24;
1458                         /* Don't wait for more than about 15msecs, even */
1459                         /* under extreme contention.                    */
ts.tv_sec = 0;
ts.tv_nsec = 1 << i;
nanosleep(&ts, 0);
1463         }
1464     }
1465 }
1467 #else  /* !USE_SPINLOCK */
1468 void GC_lock()
1469 {
1470 #ifndef NO_PTHREAD_TRYLOCK
1471     if (1 == GC_nprocs || GC_collecting) {
pthread_mutex_lock(&GC_allocate_ml);
1473     } else {
GC_generic_lock(&GC_allocate_ml);
1475     }
1476 #else  /* !NO_PTHREAD_TRYLOCK */
pthread_mutex_lock(&GC_allocate_ml);
1478 #endif /* !NO_PTHREAD_TRYLOCK */
1479 }
1481 #endif /* !USE_SPINLOCK */
1483 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1485 #ifdef GC_ASSERTIONS
pthread_t GC_mark_lock_holder = NO_THREAD;
1487 #endif
1489 #if 0
1490   /* Ugly workaround for a linux threads bug in the final versions      */
1491   /* of glibc2.1.  Pthread_mutex_trylock sets the mutex owner           */
1492   /* field even when it fails to acquire the mutex.  This causes        */
1493   /* pthread_cond_wait to die.  Remove for glibc2.2.                    */
1494   /* According to the man page, we should use                           */
1495   /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually   */
1496   /* defined.                                                           */
1497   static pthread_mutex_t mark_mutex =
1498         {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1499 #else
1500   static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1501 #endif
1503 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1505 void GC_acquire_mark_lock()
1506 {
1507 /*
1508     if (pthread_mutex_lock(&mark_mutex) != 0) {
1509         ABORT("pthread_mutex_lock failed");
1510     }
1511 */
GC_generic_lock(&mark_mutex);
1513 #   ifdef GC_ASSERTIONS
GC_mark_lock_holder = pthread_self();
1515 #   endif
1516 }
1518 void GC_release_mark_lock()
1519 {
GC_ASSERT(GC_mark_lock_holder == pthread_self());
1521 #   ifdef GC_ASSERTIONS
GC_mark_lock_holder = NO_THREAD;
1523 #   endif
1524     if (pthread_mutex_unlock(&mark_mutex) != 0) {
ABORT("pthread_mutex_unlock failed");
1526     }
1527 }
1529 /* Collector must wait for a freelist builders for 2 reasons:           */
1530 /* 1) Mark bits may still be getting examined without lock.             */
1531 /* 2) Partial free lists referenced only by locals may not be scanned   */
1532 /*    correctly, e.g. if they contain "pointer-free" objects, since the */
1533 /*    free-list link may be ignored.                                    */
1534 void GC_wait_builder()
1535 {
GC_ASSERT(GC_mark_lock_holder == pthread_self());
1537 #   ifdef GC_ASSERTIONS
GC_mark_lock_holder = NO_THREAD;
1539 #   endif
1540     if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
ABORT("pthread_cond_wait failed");
1542     }
GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1544 #   ifdef GC_ASSERTIONS
GC_mark_lock_holder = pthread_self();
1546 #   endif
1547 }
1549 void GC_wait_for_reclaim()
1550 {
GC_acquire_mark_lock();
1552     while (GC_fl_builder_count > 0) {
GC_wait_builder();
1554     }
GC_release_mark_lock();
1556 }
1558 void GC_notify_all_builder()
1559 {
GC_ASSERT(GC_mark_lock_holder == pthread_self());
1561     if (pthread_cond_broadcast(&builder_cv) != 0) {
ABORT("pthread_cond_broadcast failed");
1563     }
1564 }
1566 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1568 #ifdef PARALLEL_MARK
1570 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1572 void GC_wait_marker()
1573 {
GC_ASSERT(GC_mark_lock_holder == pthread_self());
1575 #   ifdef GC_ASSERTIONS
GC_mark_lock_holder = NO_THREAD;
1577 #   endif
1578     if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
ABORT("pthread_cond_wait failed");
1580     }
GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1582 #   ifdef GC_ASSERTIONS
GC_mark_lock_holder = pthread_self();
1584 #   endif
1585 }
1587 void GC_notify_all_marker()
1588 {
1589     if (pthread_cond_broadcast(&mark_cv) != 0) {
ABORT("pthread_cond_broadcast failed");
1591     }
1592 }
1594 #endif /* PARALLEL_MARK */
1596 # endif /* GC_LINUX_THREADS and friends */