gcc/
[official-gcc.git] / boehm-gc / pthread_support.c
blob6d8f020e076c13af730e219dd18f93bbf67d8628
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.
7 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
8 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
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.
17 * Support code for LinuxThreads, the clone()-based kernel
18 * thread package for Linux which is included in libc6.
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.
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.
29 /* DG/UX ix86 support <takis@xfree86.org> */
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.
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
47 /*#define DEBUG_THREADS 1*/
48 /*#define GC_ASSERTIONS*/
50 #include "gc_config.h"
52 #ifdef GC_PTHREAD_SYM_VERSION
53 #define _GNU_SOURCE
54 #include <dlfcn.h>
55 #endif
57 # include "gc.h"
58 # include "private/pthread_support.h"
60 # if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
61 && !defined(GC_WIN32_THREADS)
63 # if defined(GC_HPUX_THREADS) && !defined(USE_PTHREAD_SPECIFIC) \
64 && !defined(USE_COMPILER_TLS)
65 # ifdef __GNUC__
66 # define USE_PTHREAD_SPECIFIC
67 /* Empirically, as of gcc 3.3, USE_COMPILER_TLS doesn't work. */
68 # else
69 # define USE_COMPILER_TLS
70 # endif
71 # endif
73 # if defined USE_HPUX_TLS
74 --> Macro replaced by USE_COMPILER_TLS
75 # endif
77 # if (defined(GC_DGUX386_THREADS) || defined(GC_OSF1_THREADS) || \
78 defined(GC_DARWIN_THREADS) || defined(GC_AIX_THREADS)) \
79 && !defined(USE_PTHREAD_SPECIFIC)
80 # define USE_PTHREAD_SPECIFIC
81 # endif
83 # if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
84 # define _POSIX4A_DRAFT10_SOURCE 1
85 # endif
87 # if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
88 # define _USING_POSIX4A_DRAFT10 1
89 # endif
91 # ifdef THREAD_LOCAL_ALLOC
92 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
93 # include "private/specific.h"
94 # endif
95 # if defined(USE_PTHREAD_SPECIFIC)
96 # define GC_getspecific pthread_getspecific
97 # define GC_setspecific pthread_setspecific
98 # define GC_key_create pthread_key_create
99 typedef pthread_key_t GC_key_t;
100 # endif
101 # if defined(USE_COMPILER_TLS)
102 # define GC_getspecific(x) (x)
103 # define GC_setspecific(key, v) ((key) = (v), 0)
104 # define GC_key_create(key, d) 0
105 typedef void * GC_key_t;
106 # endif
107 # endif
108 # include <stdlib.h>
109 # include <pthread.h>
110 # include <sched.h>
111 # include <time.h>
112 # include <errno.h>
113 # include <unistd.h>
114 # include <sys/mman.h>
115 # include <sys/time.h>
116 # include <sys/types.h>
117 # include <sys/stat.h>
118 # include <fcntl.h>
119 # include <signal.h>
121 #if defined(GC_DARWIN_THREADS)
122 # include "private/darwin_semaphore.h"
123 #else
124 # include <semaphore.h>
125 #endif /* !GC_DARWIN_THREADS */
127 #if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
128 # include <sys/sysctl.h>
129 #endif /* GC_DARWIN_THREADS */
133 #if defined(GC_DGUX386_THREADS)
134 # include <sys/dg_sys_info.h>
135 # include <sys/_int_psem.h>
136 /* sem_t is an uint in DG/UX */
137 typedef unsigned int sem_t;
138 #endif /* GC_DGUX386_THREADS */
140 #ifndef __GNUC__
141 # define __inline__
142 #endif
144 #ifdef GC_USE_LD_WRAP
145 # define WRAP_FUNC(f) __wrap_##f
146 # define REAL_FUNC(f) __real_##f
147 #else
148 # define WRAP_FUNC(f) GC_##f
149 # if !defined(GC_DGUX386_THREADS)
150 # define REAL_FUNC(f) f
151 # else /* GC_DGUX386_THREADS */
152 # define REAL_FUNC(f) __d10_##f
153 # endif /* GC_DGUX386_THREADS */
154 # undef pthread_create
155 # if !defined(GC_DARWIN_THREADS)
156 # undef pthread_sigmask
157 # endif
158 # undef pthread_join
159 # undef pthread_detach
160 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
161 && !defined(_PTHREAD_USE_PTDNAM_)
162 /* Restore the original mangled names on Tru64 UNIX. */
163 # define pthread_create __pthread_create
164 # define pthread_join __pthread_join
165 # define pthread_detach __pthread_detach
166 # endif
167 #endif
169 void GC_thr_init();
171 static GC_bool parallel_initialized = FALSE;
173 void GC_init_parallel();
175 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
177 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
179 #ifdef USE_COMPILER_TLS
180 __thread
181 #endif
182 GC_key_t GC_thread_key;
184 static GC_bool keys_initialized;
186 /* Recover the contents of the freelist array fl into the global one gfl.*/
187 /* Note that the indexing scheme differs, in that gfl has finer size */
188 /* resolution, even if not all entries are used. */
189 /* We hold the allocator lock. */
190 static void return_freelists(ptr_t *fl, ptr_t *gfl)
192 int i;
193 ptr_t q, *qptr;
194 size_t nwords;
196 for (i = 1; i < NFREELISTS; ++i) {
197 nwords = i * (GRANULARITY/sizeof(word));
198 qptr = fl + i;
199 q = *qptr;
200 if ((word)q >= HBLKSIZE) {
201 if (gfl[nwords] == 0) {
202 gfl[nwords] = q;
203 } else {
204 /* Concatenate: */
205 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
206 GC_ASSERT(0 == q);
207 *qptr = gfl[nwords];
208 gfl[nwords] = fl[i];
211 /* Clear fl[i], since the thread structure may hang around. */
212 /* Do it in a way that is likely to trap if we access it. */
213 fl[i] = (ptr_t)HBLKSIZE;
217 /* We statically allocate a single "size 0" object. It is linked to */
218 /* itself, and is thus repeatedly reused for all size 0 allocation */
219 /* requests. (Size 0 gcj allocation requests are incorrect, and */
220 /* we arrange for those to fault asap.) */
221 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
223 /* Each thread structure must be initialized. */
224 /* This call must be made from the new thread. */
225 /* Caller holds allocation lock. */
226 void GC_init_thread_local(GC_thread p)
228 int i;
230 if (!keys_initialized) {
231 if (0 != GC_key_create(&GC_thread_key, 0)) {
232 ABORT("Failed to create key for local allocator");
234 keys_initialized = TRUE;
236 if (0 != GC_setspecific(GC_thread_key, p)) {
237 ABORT("Failed to set thread specific allocation pointers");
239 for (i = 1; i < NFREELISTS; ++i) {
240 p -> ptrfree_freelists[i] = (ptr_t)1;
241 p -> normal_freelists[i] = (ptr_t)1;
242 # ifdef GC_GCJ_SUPPORT
243 p -> gcj_freelists[i] = (ptr_t)1;
244 # endif
246 /* Set up the size 0 free lists. */
247 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
248 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
249 # ifdef GC_GCJ_SUPPORT
250 p -> gcj_freelists[0] = (ptr_t)(-1);
251 # endif
254 #ifdef GC_GCJ_SUPPORT
255 extern ptr_t * GC_gcjobjfreelist;
256 #endif
258 /* We hold the allocator lock. */
259 void GC_destroy_thread_local(GC_thread p)
261 /* We currently only do this from the thread itself or from */
262 /* the fork handler for a child process. */
263 # ifndef HANDLE_FORK
264 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
265 # endif
266 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
267 return_freelists(p -> normal_freelists, GC_objfreelist);
268 # ifdef GC_GCJ_SUPPORT
269 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
270 # endif
273 extern GC_PTR GC_generic_malloc_many();
275 GC_PTR GC_local_malloc(size_t bytes)
277 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
278 return(GC_malloc(bytes));
279 } else {
280 int index = INDEX_FROM_BYTES(bytes);
281 ptr_t * my_fl;
282 ptr_t my_entry;
283 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
284 GC_key_t k = GC_thread_key;
285 # endif
286 void * tsd;
288 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
289 if (EXPECT(0 == k, 0)) {
290 /* This can happen if we get called when the world is */
291 /* being initialized. Whether we can actually complete */
292 /* the initialization then is unclear. */
293 GC_init_parallel();
294 k = GC_thread_key;
296 # endif
297 tsd = GC_getspecific(GC_thread_key);
298 # ifdef GC_ASSERTIONS
299 LOCK();
300 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
301 UNLOCK();
302 # endif
303 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
304 my_entry = *my_fl;
305 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
306 ptr_t next = obj_link(my_entry);
307 GC_PTR result = (GC_PTR)my_entry;
308 *my_fl = next;
309 obj_link(my_entry) = 0;
310 PREFETCH_FOR_WRITE(next);
311 return result;
312 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
313 *my_fl = my_entry + index + 1;
314 return GC_malloc(bytes);
315 } else {
316 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
317 if (*my_fl == 0) return GC_oom_fn(bytes);
318 return GC_local_malloc(bytes);
323 GC_PTR GC_local_malloc_atomic(size_t bytes)
325 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
326 return(GC_malloc_atomic(bytes));
327 } else {
328 int index = INDEX_FROM_BYTES(bytes);
329 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
330 -> ptrfree_freelists + index;
331 ptr_t my_entry = *my_fl;
333 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
334 GC_PTR result = (GC_PTR)my_entry;
335 *my_fl = obj_link(my_entry);
336 return result;
337 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
338 *my_fl = my_entry + index + 1;
339 return GC_malloc_atomic(bytes);
340 } else {
341 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
342 /* *my_fl is updated while the collector is excluded; */
343 /* the free list is always visible to the collector as */
344 /* such. */
345 if (*my_fl == 0) return GC_oom_fn(bytes);
346 return GC_local_malloc_atomic(bytes);
351 #ifdef GC_GCJ_SUPPORT
353 #include "include/gc_gcj.h"
355 #ifdef GC_ASSERTIONS
356 extern GC_bool GC_gcj_malloc_initialized;
357 #endif
359 extern int GC_gcj_kind;
361 GC_PTR GC_local_gcj_malloc(size_t bytes,
362 void * ptr_to_struct_containing_descr)
364 GC_ASSERT(GC_gcj_malloc_initialized);
365 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
366 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
367 } else {
368 int index = INDEX_FROM_BYTES(bytes);
369 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
370 -> gcj_freelists + index;
371 ptr_t my_entry = *my_fl;
372 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
373 GC_PTR result = (GC_PTR)my_entry;
374 GC_ASSERT(!GC_incremental);
375 /* We assert that any concurrent marker will stop us. */
376 /* Thus it is impossible for a mark procedure to see the */
377 /* allocation of the next object, but to see this object */
378 /* still containing a free list pointer. Otherwise the */
379 /* marker might find a random "mark descriptor". */
380 *(volatile ptr_t *)my_fl = obj_link(my_entry);
381 /* We must update the freelist before we store the pointer. */
382 /* Otherwise a GC at this point would see a corrupted */
383 /* free list. */
384 /* A memory barrier is probably never needed, since the */
385 /* action of stopping this thread will cause prior writes */
386 /* to complete. */
387 GC_ASSERT(((void * volatile *)result)[1] == 0);
388 *(void * volatile *)result = ptr_to_struct_containing_descr;
389 return result;
390 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
391 if (!GC_incremental) *my_fl = my_entry + index + 1;
392 /* In the incremental case, we always have to take this */
393 /* path. Thus we leave the counter alone. */
394 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
395 } else {
396 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
397 if (*my_fl == 0) return GC_oom_fn(bytes);
398 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
403 #endif /* GC_GCJ_SUPPORT */
405 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
407 # define GC_destroy_thread_local(t)
409 # endif /* !THREAD_LOCAL_ALLOC */
411 #if 0
413 To make sure that we're using LinuxThreads and not some other thread
414 package, we generate a dummy reference to `pthread_kill_other_threads_np'
415 (was `__pthread_initial_thread_bos' but that disappeared),
416 which is a symbol defined in LinuxThreads, but (hopefully) not in other
417 thread packages.
419 We no longer do this, since this code is now portable enough that it might
420 actually work for something else.
422 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
423 #endif /* 0 */
425 long GC_nprocs = 1; /* Number of processors. We may not have */
426 /* access to all of them, but this is as good */
427 /* a guess as any ... */
429 #ifdef PARALLEL_MARK
431 # ifndef MAX_MARKERS
432 # define MAX_MARKERS 16
433 # endif
435 static ptr_t marker_sp[MAX_MARKERS] = {0};
437 void * GC_mark_thread(void * id)
439 word my_mark_no = 0;
441 marker_sp[(word)id] = GC_approx_sp();
442 for (;; ++my_mark_no) {
443 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
444 /* promptly. This is important if it were called from the signal */
445 /* handler or from the GC lock acquisition code. Under Linux, it's */
446 /* not safe to call it from a signal handler, since it uses mutexes */
447 /* and condition variables. Since it is called only here, the */
448 /* argument is unnecessary. */
449 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
450 /* resynchronize if we get far off, e.g. because GC_mark_no */
451 /* wrapped. */
452 my_mark_no = GC_mark_no;
454 # ifdef DEBUG_THREADS
455 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
456 # endif
457 GC_help_marker(my_mark_no);
461 extern long GC_markers; /* Number of mark threads we would */
462 /* like to have. Includes the */
463 /* initiating thread. */
465 pthread_t GC_mark_threads[MAX_MARKERS];
467 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
469 static void start_mark_threads()
471 unsigned i;
472 pthread_attr_t attr;
474 if (GC_markers > MAX_MARKERS) {
475 WARN("Limiting number of mark threads\n", 0);
476 GC_markers = MAX_MARKERS;
478 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
480 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
481 ABORT("pthread_attr_setdetachstate failed");
483 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
484 /* Default stack size is usually too small: fix it. */
485 /* Otherwise marker threads or GC may run out of */
486 /* space. */
487 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
489 size_t old_size;
490 int code;
492 if (pthread_attr_getstacksize(&attr, &old_size) != 0)
493 ABORT("pthread_attr_getstacksize failed\n");
494 if (old_size < MIN_STACK_SIZE) {
495 if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
496 ABORT("pthread_attr_setstacksize failed\n");
499 # endif /* HPUX || GC_DGUX386_THREADS */
500 # ifdef CONDPRINT
501 if (GC_print_stats) {
502 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
504 # endif
505 for (i = 0; i < GC_markers - 1; ++i) {
506 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
507 GC_mark_thread, (void *)(word)i)) {
508 WARN("Marker thread creation failed, errno = %ld.\n", errno);
513 #else /* !PARALLEL_MARK */
515 static __inline__ void start_mark_threads()
519 #endif /* !PARALLEL_MARK */
521 GC_bool GC_thr_initialized = FALSE;
523 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
525 void GC_push_thread_structures GC_PROTO((void))
527 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
528 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
529 GC_push_all((ptr_t)(&GC_thread_key),
530 (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
531 # endif
534 #ifdef THREAD_LOCAL_ALLOC
535 /* We must explicitly mark ptrfree and gcj free lists, since the free */
536 /* list links wouldn't otherwise be found. We also set them in the */
537 /* normal free lists, since that involves touching less memory than if */
538 /* we scanned them normally. */
539 void GC_mark_thread_local_free_lists(void)
541 int i, j;
542 GC_thread p;
543 ptr_t q;
545 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
546 for (p = GC_threads[i]; 0 != p; p = p -> next) {
547 for (j = 1; j < NFREELISTS; ++j) {
548 q = p -> ptrfree_freelists[j];
549 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
550 q = p -> normal_freelists[j];
551 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
552 # ifdef GC_GCJ_SUPPORT
553 q = p -> gcj_freelists[j];
554 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
555 # endif /* GC_GCJ_SUPPORT */
560 #endif /* THREAD_LOCAL_ALLOC */
562 static struct GC_Thread_Rep first_thread;
564 /* Add a thread to GC_threads. We assume it wasn't already there. */
565 /* Caller holds allocation lock. */
566 GC_thread GC_new_thread(pthread_t id)
568 int hv = ((word)id) % THREAD_TABLE_SZ;
569 GC_thread result;
570 static GC_bool first_thread_used = FALSE;
572 if (!first_thread_used) {
573 result = &first_thread;
574 first_thread_used = TRUE;
575 } else {
576 result = (struct GC_Thread_Rep *)
577 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
579 if (result == 0) return(0);
580 result -> id = id;
581 result -> next = GC_threads[hv];
582 GC_threads[hv] = result;
583 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
584 return(result);
587 /* Delete a thread from GC_threads. We assume it is there. */
588 /* (The code intentionally traps if it wasn't.) */
589 /* Caller holds allocation lock. */
590 void GC_delete_thread(pthread_t id)
592 int hv = ((word)id) % THREAD_TABLE_SZ;
593 register GC_thread p = GC_threads[hv];
594 register GC_thread prev = 0;
596 while (!pthread_equal(p -> id, id)) {
597 prev = p;
598 p = p -> next;
600 if (prev == 0) {
601 GC_threads[hv] = p -> next;
602 } else {
603 prev -> next = p -> next;
606 if (p != &first_thread)
607 GC_INTERNAL_FREE(p);
610 /* If a thread has been joined, but we have not yet */
611 /* been notified, then there may be more than one thread */
612 /* in the table with the same pthread id. */
613 /* This is OK, but we need a way to delete a specific one. */
614 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
616 int hv = ((word)id) % THREAD_TABLE_SZ;
617 register GC_thread p = GC_threads[hv];
618 register GC_thread prev = 0;
620 while (p != gc_id) {
621 prev = p;
622 p = p -> next;
624 if (prev == 0) {
625 GC_threads[hv] = p -> next;
626 } else {
627 prev -> next = p -> next;
629 GC_INTERNAL_FREE(p);
632 /* Return a GC_thread corresponding to a given pthread_t. */
633 /* Returns 0 if it's not there. */
634 /* Caller holds allocation lock or otherwise inhibits */
635 /* updates. */
636 /* If there is more than one thread with the given id we */
637 /* return the most recent one. */
638 GC_thread GC_lookup_thread(pthread_t id)
640 int hv = ((word)id) % THREAD_TABLE_SZ;
641 register GC_thread p = GC_threads[hv];
643 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
644 return(p);
647 #ifdef HANDLE_FORK
648 /* Remove all entries from the GC_threads table, except the */
649 /* one for the current thread. We need to do this in the child */
650 /* process after a fork(), since only the current thread */
651 /* survives in the child. */
652 void GC_remove_all_threads_but_me(void)
654 pthread_t self = pthread_self();
655 int hv;
656 GC_thread p, next, me;
658 for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
659 me = 0;
660 for (p = GC_threads[hv]; 0 != p; p = next) {
661 next = p -> next;
662 if (p -> id == self) {
663 me = p;
664 p -> next = 0;
665 } else {
666 # ifdef THREAD_LOCAL_ALLOC
667 if (!(p -> flags & FINISHED)) {
668 GC_destroy_thread_local(p);
670 # endif /* THREAD_LOCAL_ALLOC */
671 if (p != &first_thread) GC_INTERNAL_FREE(p);
674 GC_threads[hv] = me;
677 #endif /* HANDLE_FORK */
679 #ifdef USE_PROC_FOR_LIBRARIES
680 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
682 int i;
683 GC_thread p;
685 # ifdef PARALLEL_MARK
686 for (i = 0; i < GC_markers; ++i) {
687 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
689 # endif
690 for (i = 0; i < THREAD_TABLE_SZ; i++) {
691 for (p = GC_threads[i]; p != 0; p = p -> next) {
692 if (0 != p -> stack_end) {
693 # ifdef STACK_GROWS_UP
694 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
695 # else /* STACK_GROWS_DOWN */
696 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
697 # endif
701 return 0;
703 #endif /* USE_PROC_FOR_LIBRARIES */
705 #ifdef GC_LINUX_THREADS
706 /* Return the number of processors, or i<= 0 if it can't be determined. */
707 int GC_get_nprocs()
709 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
710 /* appears to be buggy in many cases. */
711 /* We look for lines "cpu<n>" in /proc/stat. */
712 # define STAT_BUF_SIZE 4096
713 # define STAT_READ read
714 /* If read is wrapped, this may need to be redefined to call */
715 /* the real one. */
716 char stat_buf[STAT_BUF_SIZE];
717 int f;
718 word result = 1;
719 /* Some old kernels only have a single "cpu nnnn ..." */
720 /* entry in /proc/stat. We identify those as */
721 /* uniprocessors. */
722 size_t i, len = 0;
724 f = open("/proc/stat", O_RDONLY);
725 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
726 WARN("Couldn't read /proc/stat\n", 0);
727 return -1;
729 for (i = 0; i < len - 100; ++i) {
730 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
731 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
732 int cpu_no = atoi(stat_buf + i + 4);
733 if (cpu_no >= result) result = cpu_no + 1;
736 close(f);
737 return result;
739 #endif /* GC_LINUX_THREADS */
741 /* We hold the GC lock. Wait until an in-progress GC has finished. */
742 /* Repeatedly RELEASES GC LOCK in order to wait. */
743 /* If wait_for_all is true, then we exit with the GC lock held and no */
744 /* collection in progress; otherwise we just wait for the current GC */
745 /* to finish. */
746 extern GC_bool GC_collection_in_progress();
747 void GC_wait_for_gc_completion(GC_bool wait_for_all)
749 if (GC_incremental && GC_collection_in_progress()) {
750 int old_gc_no = GC_gc_no;
752 /* Make sure that no part of our stack is still on the mark stack, */
753 /* since it's about to be unmapped. */
754 while (GC_incremental && GC_collection_in_progress()
755 && (wait_for_all || old_gc_no == GC_gc_no)) {
756 ENTER_GC();
757 GC_in_thread_creation = TRUE;
758 GC_collect_a_little_inner(1);
759 GC_in_thread_creation = FALSE;
760 EXIT_GC();
761 UNLOCK();
762 sched_yield();
763 LOCK();
768 #ifdef HANDLE_FORK
769 /* Procedures called before and after a fork. The goal here is to make */
770 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
771 /* attainable, since the single UNIX spec seems to imply that one */
772 /* should only call async-signal-safe functions, and we probably can't */
773 /* quite guarantee that. But we give it our best shot. (That same */
774 /* spec also implies that it's not safe to call the system malloc */
775 /* between fork() and exec(). Thus we're doing no worse than it. */
777 /* Called before a fork() */
778 void GC_fork_prepare_proc(void)
780 /* Acquire all relevant locks, so that after releasing the locks */
781 /* the child will see a consistent state in which monitor */
782 /* invariants hold. Unfortunately, we can't acquire libc locks */
783 /* we might need, and there seems to be no guarantee that libc */
784 /* must install a suitable fork handler. */
785 /* Wait for an ongoing GC to finish, since we can't finish it in */
786 /* the (one remaining thread in) the child. */
787 LOCK();
788 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
789 GC_wait_for_reclaim();
790 # endif
791 GC_wait_for_gc_completion(TRUE);
792 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
793 GC_acquire_mark_lock();
794 # endif
797 /* Called in parent after a fork() */
798 void GC_fork_parent_proc(void)
800 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
801 GC_release_mark_lock();
802 # endif
803 UNLOCK();
806 /* Called in child after a fork() */
807 void GC_fork_child_proc(void)
809 /* Clean up the thread table, so that just our thread is left. */
810 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
811 GC_release_mark_lock();
812 # endif
813 GC_remove_all_threads_but_me();
814 # ifdef PARALLEL_MARK
815 /* Turn off parallel marking in the child, since we are probably */
816 /* just going to exec, and we would have to restart mark threads. */
817 GC_markers = 1;
818 GC_parallel = FALSE;
819 # endif /* PARALLEL_MARK */
820 UNLOCK();
822 #endif /* HANDLE_FORK */
824 #if defined(GC_DGUX386_THREADS)
825 /* Return the number of processors, or i<= 0 if it can't be determined. */
826 int GC_get_nprocs()
828 /* <takis@XFree86.Org> */
829 int numCpus;
830 struct dg_sys_info_pm_info pm_sysinfo;
831 int status =0;
833 status = dg_sys_info((long int *) &pm_sysinfo,
834 DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
835 if (status < 0)
836 /* set -1 for error */
837 numCpus = -1;
838 else
839 /* Active CPUs */
840 numCpus = pm_sysinfo.idle_vp_count;
842 # ifdef DEBUG_THREADS
843 GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
844 # endif
845 return(numCpus);
847 #endif /* GC_DGUX386_THREADS */
849 /* We hold the allocation lock. */
850 void GC_thr_init()
852 # ifndef GC_DARWIN_THREADS
853 int dummy;
854 # endif
855 GC_thread t;
857 if (GC_thr_initialized) return;
858 GC_thr_initialized = TRUE;
860 # ifdef HANDLE_FORK
861 /* Prepare for a possible fork. */
862 pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
863 GC_fork_child_proc);
864 # endif /* HANDLE_FORK */
865 /* Add the initial thread, so we can stop it. */
866 t = GC_new_thread(pthread_self());
867 # ifdef GC_DARWIN_THREADS
868 t -> stop_info.mach_thread = mach_thread_self();
869 # else
870 t -> stop_info.stack_ptr = (ptr_t)(&dummy);
871 # endif
872 t -> flags = DETACHED | MAIN_THREAD;
874 GC_stop_init();
876 /* Set GC_nprocs. */
878 char * nprocs_string = GETENV("GC_NPROCS");
879 GC_nprocs = -1;
880 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
882 if (GC_nprocs <= 0) {
883 # if defined(GC_HPUX_THREADS)
884 GC_nprocs = pthread_num_processors_np();
885 # endif
886 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS) \
887 || defined(GC_SOLARIS_PTHREADS) || defined(GC_GNU_THREADS)
888 GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
889 if (GC_nprocs <= 0) GC_nprocs = 1;
890 # endif
891 # if defined(GC_IRIX_THREADS)
892 GC_nprocs = sysconf(_SC_NPROC_ONLN);
893 if (GC_nprocs <= 0) GC_nprocs = 1;
894 # endif
895 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
896 int ncpus = 1;
897 size_t len = sizeof(ncpus);
898 sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
899 GC_nprocs = ncpus;
900 # endif
901 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
902 GC_nprocs = GC_get_nprocs();
903 # endif
905 if (GC_nprocs <= 0) {
906 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
907 GC_nprocs = 2;
908 # ifdef PARALLEL_MARK
909 GC_markers = 1;
910 # endif
911 } else {
912 # ifdef PARALLEL_MARK
914 char * markers_string = GETENV("GC_MARKERS");
915 if (markers_string != NULL) {
916 GC_markers = atoi(markers_string);
917 } else {
918 GC_markers = GC_nprocs;
921 # endif
923 # ifdef PARALLEL_MARK
924 # ifdef CONDPRINT
925 if (GC_print_stats) {
926 GC_printf2("Number of processors = %ld, "
927 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
929 # endif
930 if (GC_markers == 1) {
931 GC_parallel = FALSE;
932 # ifdef CONDPRINT
933 if (GC_print_stats) {
934 GC_printf0("Single marker thread, turning off parallel marking\n");
936 # endif
937 } else {
938 GC_parallel = TRUE;
939 /* Disable true incremental collection, but generational is OK. */
940 GC_time_limit = GC_TIME_UNLIMITED;
942 /* If we are using a parallel marker, actually start helper threads. */
943 if (GC_parallel) start_mark_threads();
944 # endif
948 /* Perform all initializations, including those that */
949 /* may require allocation. */
950 /* Called without allocation lock. */
951 /* Must be called before a second thread is created. */
952 /* Called without allocation lock. */
953 void GC_init_parallel()
955 if (parallel_initialized) return;
956 parallel_initialized = TRUE;
958 /* GC_init() calls us back, so set flag first. */
959 if (!GC_is_initialized) GC_init();
960 /* Initialize thread local free lists if used. */
961 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
962 LOCK();
963 GC_init_thread_local(GC_lookup_thread(pthread_self()));
964 UNLOCK();
965 # endif
969 #if !defined(GC_DARWIN_THREADS)
970 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
972 sigset_t fudged_set;
974 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
975 fudged_set = *set;
976 sigdelset(&fudged_set, SIG_SUSPEND);
977 set = &fudged_set;
979 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
981 #endif /* !GC_DARWIN_THREADS */
983 /* Wrappers for functions that are likely to block for an appreciable */
984 /* length of time. Must be called in pairs, if at all. */
985 /* Nothing much beyond the system call itself should be executed */
986 /* between these. */
988 void GC_start_blocking(void) {
989 # define SP_SLOP 128
990 GC_thread me;
991 LOCK();
992 me = GC_lookup_thread(pthread_self());
993 GC_ASSERT(!(me -> thread_blocked));
994 # ifdef SPARC
995 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
996 # else
997 # ifndef GC_DARWIN_THREADS
998 me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
999 # endif
1000 # endif
1001 # ifdef IA64
1002 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1003 # endif
1004 /* Add some slop to the stack pointer, since the wrapped call may */
1005 /* end up pushing more callee-save registers. */
1006 # ifndef GC_DARWIN_THREADS
1007 # ifdef STACK_GROWS_UP
1008 me -> stop_info.stack_ptr += SP_SLOP;
1009 # else
1010 me -> stop_info.stack_ptr -= SP_SLOP;
1011 # endif
1012 # endif
1013 me -> thread_blocked = TRUE;
1014 UNLOCK();
1017 void GC_end_blocking(void) {
1018 GC_thread me;
1019 LOCK(); /* This will block if the world is stopped. */
1020 me = GC_lookup_thread(pthread_self());
1021 GC_ASSERT(me -> thread_blocked);
1022 me -> thread_blocked = FALSE;
1023 UNLOCK();
1026 #if defined(GC_DGUX386_THREADS)
1027 #define __d10_sleep sleep
1028 #endif /* GC_DGUX386_THREADS */
1030 /* A wrapper for the standard C sleep function */
1031 int WRAP_FUNC(sleep) (unsigned int seconds)
1033 int result;
1035 GC_start_blocking();
1036 result = REAL_FUNC(sleep)(seconds);
1037 GC_end_blocking();
1038 return result;
1041 struct start_info {
1042 void *(*start_routine)(void *);
1043 void *arg;
1044 word flags;
1045 sem_t registered; /* 1 ==> in our thread table, but */
1046 /* parent hasn't yet noticed. */
1049 /* Called at thread exit. */
1050 /* Never called for main thread. That's OK, since it */
1051 /* results in at most a tiny one-time leak. And */
1052 /* linuxthreads doesn't reclaim the main threads */
1053 /* resources or id anyway. */
1054 void GC_thread_exit_proc(void *arg)
1056 GC_thread me;
1058 LOCK();
1059 me = GC_lookup_thread(pthread_self());
1060 GC_destroy_thread_local(me);
1061 if (me -> flags & DETACHED) {
1062 GC_delete_thread(pthread_self());
1063 } else {
1064 me -> flags |= FINISHED;
1066 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1067 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1068 GC_remove_specific(GC_thread_key);
1069 # endif
1070 /* The following may run the GC from "nonexistent" thread. */
1071 GC_wait_for_gc_completion(FALSE);
1072 UNLOCK();
1075 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1077 int result;
1078 GC_thread thread_gc_id;
1080 LOCK();
1081 thread_gc_id = GC_lookup_thread(thread);
1082 /* This is guaranteed to be the intended one, since the thread id */
1083 /* cant have been recycled by pthreads. */
1084 UNLOCK();
1085 result = REAL_FUNC(pthread_join)(thread, retval);
1086 # if defined (GC_FREEBSD_THREADS)
1087 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1088 appears to be) a spurious EINTR which caused the test and real code
1089 to gratuitously fail. Having looked at system pthread library source
1090 code, I see how this return code may be generated. In one path of
1091 code, pthread_join() just returns the errno setting of the thread
1092 being joined. This does not match the POSIX specification or the
1093 local man pages thus I have taken the liberty to catch this one
1094 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1095 if (result == EINTR) result = 0;
1096 # endif
1097 if (result == 0) {
1098 LOCK();
1099 /* Here the pthread thread id may have been recycled. */
1100 GC_delete_gc_thread(thread, thread_gc_id);
1101 UNLOCK();
1103 return result;
1107 WRAP_FUNC(pthread_detach)(pthread_t thread)
1109 int result;
1110 GC_thread thread_gc_id;
1112 LOCK();
1113 thread_gc_id = GC_lookup_thread(thread);
1114 UNLOCK();
1115 result = REAL_FUNC(pthread_detach)(thread);
1116 if (result == 0) {
1117 LOCK();
1118 thread_gc_id -> flags |= DETACHED;
1119 /* Here the pthread thread id may have been recycled. */
1120 if (thread_gc_id -> flags & FINISHED) {
1121 GC_delete_gc_thread(thread, thread_gc_id);
1123 UNLOCK();
1125 return result;
1128 GC_bool GC_in_thread_creation = FALSE;
1130 GC_PTR GC_get_thread_stack_base()
1132 # ifdef HAVE_PTHREAD_GETATTR_NP
1133 pthread_t my_pthread;
1134 pthread_attr_t attr;
1135 ptr_t stack_addr;
1136 size_t stack_size;
1138 my_pthread = pthread_self();
1139 if (pthread_getattr_np (my_pthread, &attr) != 0)
1141 # ifdef DEBUG_THREADS
1142 GC_printf0("Can not determine stack base for attached thread");
1143 # endif
1144 return 0;
1146 pthread_attr_getstack (&attr, (void **) &stack_addr, &stack_size);
1147 pthread_attr_destroy (&attr);
1149 # ifdef DEBUG_THREADS
1150 GC_printf1("attached thread stack address: 0x%x\n", stack_addr);
1151 # endif
1153 # ifdef STACK_GROWS_DOWN
1154 return stack_addr + stack_size;
1155 # else
1156 return stack_addr;
1157 # endif
1159 # else
1160 # ifdef DEBUG_THREADS
1161 GC_printf0("Can not determine stack base for attached thread");
1162 # endif
1163 return 0;
1164 # endif
1167 void GC_register_my_thread()
1169 GC_thread me;
1170 pthread_t my_pthread;
1172 my_pthread = pthread_self();
1173 # ifdef DEBUG_THREADS
1174 GC_printf1("Attaching thread 0x%lx\n", my_pthread);
1175 GC_printf1("pid = %ld\n", (long) getpid());
1176 # endif
1178 /* Check to ensure this thread isn't attached already. */
1179 LOCK();
1180 me = GC_lookup_thread (my_pthread);
1181 UNLOCK();
1182 if (me != 0)
1184 # ifdef DEBUG_THREADS
1185 GC_printf1("Attempt to re-attach known thread 0x%lx\n", my_pthread);
1186 # endif
1187 return;
1190 LOCK();
1191 GC_in_thread_creation = TRUE;
1192 me = GC_new_thread(my_pthread);
1193 GC_in_thread_creation = FALSE;
1195 me -> flags |= DETACHED;
1197 #ifdef GC_DARWIN_THREADS
1198 me -> stop_info.mach_thread = mach_thread_self();
1199 #else
1200 me -> stack_end = GC_get_thread_stack_base();
1201 if (me -> stack_end == 0)
1202 GC_abort("Can not determine stack base for attached thread");
1204 # ifdef STACK_GROWS_DOWN
1205 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1206 # else
1207 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1208 # endif
1209 #endif
1211 # ifdef IA64
1212 me -> backing_store_end = (ptr_t)
1213 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1214 /* This is also < 100% convincing. We should also read this */
1215 /* from /proc, but the hook to do so isn't there yet. */
1216 # endif /* IA64 */
1218 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1219 GC_init_thread_local(me);
1220 # endif
1221 UNLOCK();
1224 void GC_unregister_my_thread()
1226 pthread_t my_pthread;
1228 my_pthread = pthread_self();
1230 # ifdef DEBUG_THREADS
1231 GC_printf1("Detaching thread 0x%lx\n", my_pthread);
1232 # endif
1234 GC_thread_exit_proc (0);
1237 void * GC_start_routine(void * arg)
1239 int dummy;
1240 struct start_info * si = arg;
1241 void * result;
1242 GC_thread me;
1243 pthread_t my_pthread;
1244 void *(*start)(void *);
1245 void *start_arg;
1247 my_pthread = pthread_self();
1248 # ifdef DEBUG_THREADS
1249 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1250 GC_printf1("pid = %ld\n", (long) getpid());
1251 GC_printf1("sp = 0x%lx\n", (long) &arg);
1252 # endif
1253 LOCK();
1254 GC_in_thread_creation = TRUE;
1255 me = GC_new_thread(my_pthread);
1256 GC_in_thread_creation = FALSE;
1257 #ifdef GC_DARWIN_THREADS
1258 me -> stop_info.mach_thread = mach_thread_self();
1259 #else
1260 me -> stop_info.stack_ptr = 0;
1261 #endif
1262 me -> flags = si -> flags;
1263 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1264 /* doesn't work because the stack base in /proc/self/stat is the */
1265 /* one for the main thread. There is a strong argument that that's */
1266 /* a kernel bug, but a pervasive one. */
1267 # ifdef STACK_GROWS_DOWN
1268 me -> stack_end = (ptr_t)(((word)(&dummy) + (GC_page_size - 1))
1269 & ~(GC_page_size - 1));
1270 # ifndef GC_DARWIN_THREADS
1271 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1272 # endif
1273 /* Needs to be plausible, since an asynchronous stack mark */
1274 /* should not crash. */
1275 # else
1276 me -> stack_end = (ptr_t)((word)(&dummy) & ~(GC_page_size - 1));
1277 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1278 # endif
1279 /* This is dubious, since we may be more than a page into the stack, */
1280 /* and hence skip some of it, though it's not clear that matters. */
1281 # ifdef IA64
1282 me -> backing_store_end = (ptr_t)
1283 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1284 /* This is also < 100% convincing. We should also read this */
1285 /* from /proc, but the hook to do so isn't there yet. */
1286 # endif /* IA64 */
1287 UNLOCK();
1288 start = si -> start_routine;
1289 # ifdef DEBUG_THREADS
1290 GC_printf1("start_routine = 0x%lx\n", start);
1291 # endif
1292 start_arg = si -> arg;
1293 sem_post(&(si -> registered)); /* Last action on si. */
1294 /* OK to deallocate. */
1295 pthread_cleanup_push(GC_thread_exit_proc, 0);
1296 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1297 LOCK();
1298 GC_init_thread_local(me);
1299 UNLOCK();
1300 # endif
1301 result = (*start)(start_arg);
1302 #if DEBUG_THREADS
1303 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1304 #endif
1305 me -> status = result;
1306 pthread_cleanup_pop(1);
1307 /* Cleanup acquires lock, ensuring that we can't exit */
1308 /* while a collection that thinks we're alive is trying to stop */
1309 /* us. */
1310 return(result);
1314 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1315 const pthread_attr_t *attr,
1316 void *(*start_routine)(void *), void *arg)
1318 int result;
1319 int detachstate;
1320 word my_flags = 0;
1321 struct start_info * si;
1322 /* This is otherwise saved only in an area mmapped by the thread */
1323 /* library, which isn't visible to the collector. */
1325 /* We resist the temptation to muck with the stack size here, */
1326 /* even if the default is unreasonably small. That's the client's */
1327 /* responsibility. */
1329 LOCK();
1330 si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
1331 NORMAL);
1332 UNLOCK();
1333 if (!parallel_initialized) GC_init_parallel();
1334 if (0 == si) return(ENOMEM);
1335 sem_init(&(si -> registered), 0, 0);
1336 si -> start_routine = start_routine;
1337 si -> arg = arg;
1338 LOCK();
1339 if (!GC_thr_initialized) GC_thr_init();
1340 # ifdef GC_ASSERTIONS
1342 size_t stack_size;
1343 if (NULL == attr) {
1344 pthread_attr_t my_attr;
1345 pthread_attr_init(&my_attr);
1346 pthread_attr_getstacksize(&my_attr, &stack_size);
1347 } else {
1348 pthread_attr_getstacksize(attr, &stack_size);
1350 # ifdef PARALLEL_MARK
1351 GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1352 # else
1353 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1354 /* HBLKSIZE=8192, sizeof(word)=8 */
1355 GC_ASSERT(stack_size >= 65536);
1356 # endif
1357 /* Our threads may need to do some work for the GC. */
1358 /* Ridiculously small threads won't work, and they */
1359 /* probably wouldn't work anyway. */
1361 # endif
1362 if (NULL == attr) {
1363 detachstate = PTHREAD_CREATE_JOINABLE;
1364 } else {
1365 pthread_attr_getdetachstate(attr, &detachstate);
1367 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1368 si -> flags = my_flags;
1369 UNLOCK();
1370 # ifdef DEBUG_THREADS
1371 GC_printf1("About to start new thread from thread 0x%X\n",
1372 pthread_self());
1373 # endif
1375 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1377 # ifdef DEBUG_THREADS
1378 GC_printf1("Started thread 0x%X\n", *new_thread);
1379 # endif
1380 /* Wait until child has been added to the thread table. */
1381 /* This also ensures that we hold onto si until the child is done */
1382 /* with it. Thus it doesn't matter whether it is otherwise */
1383 /* visible to the collector. */
1384 if (0 == result) {
1385 while (0 != sem_wait(&(si -> registered))) {
1386 if (EINTR != errno) ABORT("sem_wait failed");
1389 sem_destroy(&(si -> registered));
1390 LOCK();
1391 GC_INTERNAL_FREE(si);
1392 UNLOCK();
1394 return(result);
1397 #ifdef GENERIC_COMPARE_AND_SWAP
1398 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1400 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1401 GC_word old, GC_word new_val)
1403 GC_bool result;
1404 pthread_mutex_lock(&GC_compare_and_swap_lock);
1405 if (*addr == old) {
1406 *addr = new_val;
1407 result = TRUE;
1408 } else {
1409 result = FALSE;
1411 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1412 return result;
1415 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1417 GC_word old;
1418 pthread_mutex_lock(&GC_compare_and_swap_lock);
1419 old = *addr;
1420 *addr = old + how_much;
1421 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1422 return old;
1425 #endif /* GENERIC_COMPARE_AND_SWAP */
1426 /* Spend a few cycles in a way that can't introduce contention with */
1427 /* othre threads. */
1428 void GC_pause()
1430 int i;
1431 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1432 volatile word dummy = 0;
1433 # endif
1435 for (i = 0; i < 10; ++i) {
1436 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1437 __asm__ __volatile__ (" " : : : "memory");
1438 # else
1439 /* Something that's unlikely to be optimized away. */
1440 GC_noop(++dummy);
1441 # endif
1445 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1446 /* give up. */
1448 VOLATILE GC_bool GC_collecting = 0;
1449 /* A hint that we're in the collector and */
1450 /* holding the allocation lock for an */
1451 /* extended period. */
1453 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1454 /* If we don't want to use the below spinlock implementation, either */
1455 /* because we don't have a GC_test_and_set implementation, or because */
1456 /* we don't want to risk sleeping, we can still try spinning on */
1457 /* pthread_mutex_trylock for a while. This appears to be very */
1458 /* beneficial in many cases. */
1459 /* I suspect that under high contention this is nearly always better */
1460 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1461 /* Hence we still default to the spin lock. */
1462 /* This is also used to acquire the mark lock for the parallel */
1463 /* marker. */
1465 /* Here we use a strict exponential backoff scheme. I don't know */
1466 /* whether that's better or worse than the above. We eventually */
1467 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1468 /* explicitly sleep. */
1470 #define LOCK_STATS
1471 #ifdef LOCK_STATS
1472 unsigned long GC_spin_count = 0;
1473 unsigned long GC_block_count = 0;
1474 unsigned long GC_unlocked_count = 0;
1475 #endif
1477 void GC_generic_lock(pthread_mutex_t * lock)
1479 #ifndef NO_PTHREAD_TRYLOCK
1480 unsigned pause_length = 1;
1481 unsigned i;
1483 if (0 == pthread_mutex_trylock(lock)) {
1484 # ifdef LOCK_STATS
1485 ++GC_unlocked_count;
1486 # endif
1487 return;
1489 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1490 for (i = 0; i < pause_length; ++i) {
1491 GC_pause();
1493 switch(pthread_mutex_trylock(lock)) {
1494 case 0:
1495 # ifdef LOCK_STATS
1496 ++GC_spin_count;
1497 # endif
1498 return;
1499 case EBUSY:
1500 break;
1501 default:
1502 ABORT("Unexpected error from pthread_mutex_trylock");
1505 #endif /* !NO_PTHREAD_TRYLOCK */
1506 # ifdef LOCK_STATS
1507 ++GC_block_count;
1508 # endif
1509 pthread_mutex_lock(lock);
1512 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1514 #if defined(USE_SPIN_LOCK)
1516 /* Reasonably fast spin locks. Basically the same implementation */
1517 /* as STL alloc.h. This isn't really the right way to do this. */
1518 /* but until the POSIX scheduling mess gets straightened out ... */
1520 volatile unsigned int GC_allocate_lock = 0;
1523 void GC_lock()
1525 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1526 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1527 static unsigned spin_max = low_spin_max;
1528 unsigned my_spin_max;
1529 static unsigned last_spins = 0;
1530 unsigned my_last_spins;
1531 int i;
1533 if (!GC_test_and_set(&GC_allocate_lock)) {
1534 return;
1536 my_spin_max = spin_max;
1537 my_last_spins = last_spins;
1538 for (i = 0; i < my_spin_max; i++) {
1539 if (GC_collecting || GC_nprocs == 1) goto yield;
1540 if (i < my_last_spins/2 || GC_allocate_lock) {
1541 GC_pause();
1542 continue;
1544 if (!GC_test_and_set(&GC_allocate_lock)) {
1546 * got it!
1547 * Spinning worked. Thus we're probably not being scheduled
1548 * against the other process with which we were contending.
1549 * Thus it makes sense to spin longer the next time.
1551 last_spins = i;
1552 spin_max = high_spin_max;
1553 return;
1556 /* We are probably being scheduled against the other process. Sleep. */
1557 spin_max = low_spin_max;
1558 yield:
1559 for (i = 0;; ++i) {
1560 if (!GC_test_and_set(&GC_allocate_lock)) {
1561 return;
1563 # define SLEEP_THRESHOLD 12
1564 /* Under Linux very short sleeps tend to wait until */
1565 /* the current time quantum expires. On old Linux */
1566 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1567 /* (Under 2.4, this happens only for real-time */
1568 /* processes.) We want to minimize both behaviors */
1569 /* here. */
1570 if (i < SLEEP_THRESHOLD) {
1571 sched_yield();
1572 } else {
1573 struct timespec ts;
1575 if (i > 24) i = 24;
1576 /* Don't wait for more than about 15msecs, even */
1577 /* under extreme contention. */
1578 ts.tv_sec = 0;
1579 ts.tv_nsec = 1 << i;
1580 nanosleep(&ts, 0);
1585 #else /* !USE_SPINLOCK */
1586 void GC_lock()
1588 #ifndef NO_PTHREAD_TRYLOCK
1589 if (1 == GC_nprocs || GC_collecting) {
1590 pthread_mutex_lock(&GC_allocate_ml);
1591 } else {
1592 GC_generic_lock(&GC_allocate_ml);
1594 #else /* !NO_PTHREAD_TRYLOCK */
1595 pthread_mutex_lock(&GC_allocate_ml);
1596 #endif /* !NO_PTHREAD_TRYLOCK */
1599 #endif /* !USE_SPINLOCK */
1601 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1603 #ifdef GC_ASSERTIONS
1604 pthread_t GC_mark_lock_holder = NO_THREAD;
1605 #endif
1607 #if 0
1608 /* Ugly workaround for a linux threads bug in the final versions */
1609 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1610 /* field even when it fails to acquire the mutex. This causes */
1611 /* pthread_cond_wait to die. Remove for glibc2.2. */
1612 /* According to the man page, we should use */
1613 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1614 /* defined. */
1615 static pthread_mutex_t mark_mutex =
1616 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1617 #else
1618 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1619 #endif
1621 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1623 void GC_acquire_mark_lock()
1626 if (pthread_mutex_lock(&mark_mutex) != 0) {
1627 ABORT("pthread_mutex_lock failed");
1630 GC_generic_lock(&mark_mutex);
1631 # ifdef GC_ASSERTIONS
1632 GC_mark_lock_holder = pthread_self();
1633 # endif
1636 void GC_release_mark_lock()
1638 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1639 # ifdef GC_ASSERTIONS
1640 GC_mark_lock_holder = NO_THREAD;
1641 # endif
1642 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1643 ABORT("pthread_mutex_unlock failed");
1647 /* Collector must wait for a freelist builders for 2 reasons: */
1648 /* 1) Mark bits may still be getting examined without lock. */
1649 /* 2) Partial free lists referenced only by locals may not be scanned */
1650 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1651 /* free-list link may be ignored. */
1652 void GC_wait_builder()
1654 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1655 # ifdef GC_ASSERTIONS
1656 GC_mark_lock_holder = NO_THREAD;
1657 # endif
1658 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1659 ABORT("pthread_cond_wait failed");
1661 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1662 # ifdef GC_ASSERTIONS
1663 GC_mark_lock_holder = pthread_self();
1664 # endif
1667 void GC_wait_for_reclaim()
1669 GC_acquire_mark_lock();
1670 while (GC_fl_builder_count > 0) {
1671 GC_wait_builder();
1673 GC_release_mark_lock();
1676 void GC_notify_all_builder()
1678 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1679 if (pthread_cond_broadcast(&builder_cv) != 0) {
1680 ABORT("pthread_cond_broadcast failed");
1684 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1686 #ifdef PARALLEL_MARK
1688 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1690 void GC_wait_marker()
1692 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1693 # ifdef GC_ASSERTIONS
1694 GC_mark_lock_holder = NO_THREAD;
1695 # endif
1696 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1697 ABORT("pthread_cond_wait failed");
1699 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1700 # ifdef GC_ASSERTIONS
1701 GC_mark_lock_holder = pthread_self();
1702 # endif
1705 void GC_notify_all_marker()
1707 if (pthread_cond_broadcast(&mark_cv) != 0) {
1708 ABORT("pthread_cond_broadcast failed");
1712 #endif /* PARALLEL_MARK */
1714 # endif /* GC_LINUX_THREADS and friends */