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
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
35 * Eric also suggested an alternate basis for a lock implementation in
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) \
44 * + # define GC_LOCK_TAKEN GC_allocate_lock
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)
58 # define USE_PTHREAD_SPECIFIC
59 /* Empirically, as of gcc 3.3, USE_COMPILER_TLS doesn't work. */
61 # define USE_COMPILER_TLS
65 # if defined USE_HPUX_TLS
66 --> Macro replaced by USE_COMPILER_TLS
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
75 # if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
76 # define _POSIX4A_DRAFT10_SOURCE 1
79 # if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
80 # define _USING_POSIX4A_DRAFT10 1
83 # ifdef THREAD_LOCAL_ALLOC
84 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
85 # include "private/specific.h"
88 /* Note that these macros should be used only to get/set the GC_thread pointer.
89 * We need to use both tls and pthread because we use the pthread_create function hook to
90 * free the data for foreign threads. When that doesn't happen, libgc could have old
91 * pthread_t that get reused...
93 # if defined(USE_PTHREAD_SPECIFIC)
94 # define GC_getspecific pthread_getspecific
95 # define GC_setspecific pthread_setspecific
96 # define GC_key_create pthread_key_create
97 typedef pthread_key_t GC_key_t;
99 # if defined(USE_COMPILER_TLS)
100 # define GC_getspecific(x) (GC_thread_tls)
101 # define GC_setspecific(key, v) (GC_thread_tls = (v), pthread_setspecific ((key), (v)))
102 # define GC_key_create pthread_key_create
103 typedef pthread_key_t GC_key_t;
107 # include <pthread.h>
112 # include <sys/mman.h>
113 # include <sys/time.h>
114 # include <sys/types.h>
115 # include <sys/stat.h>
119 #if defined(GC_DARWIN_THREADS)
120 # include "private/darwin_semaphore.h"
122 # include <semaphore.h>
123 #endif /* !GC_DARWIN_THREADS */
125 #if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
126 # include <sys/sysctl.h>
127 #endif /* GC_DARWIN_THREADS */
131 #if defined(GC_DGUX386_THREADS)
132 # include <sys/dg_sys_info.h>
133 # include <sys/_int_psem.h>
134 /* sem_t is an uint in DG/UX */
135 typedef unsigned int sem_t;
136 #endif /* GC_DGUX386_THREADS */
142 #ifdef GC_USE_LD_WRAP
143 # define WRAP_FUNC(f) __wrap_##f
144 # define REAL_FUNC(f) __real_##f
146 # define WRAP_FUNC(f) GC_##f
147 # if !defined(GC_DGUX386_THREADS)
148 # define REAL_FUNC(f) f
149 # else /* GC_DGUX386_THREADS */
150 # define REAL_FUNC(f) __d10_##f
151 # endif /* GC_DGUX386_THREADS */
152 # undef pthread_create
153 # if !defined(GC_DARWIN_THREADS)
154 # undef pthread_sigmask
157 # undef pthread_detach
158 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
159 && !defined(_PTHREAD_USE_PTDNAM_)
160 /* Restore the original mangled names on Tru64 UNIX. */
161 # define pthread_create __pthread_create
162 # define pthread_join __pthread_join
163 # define pthread_detach __pthread_detach
169 static GC_bool parallel_initialized = FALSE;
171 void GC_init_parallel();
173 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
175 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
177 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
178 #undef PACKAGE_BUGREPORT
180 #undef PACKAGE_STRING
181 #undef PACKAGE_TARNAME
182 #undef PACKAGE_VERSION
183 #include "mono/utils/mono-compiler.h"
186 GC_key_t GC_thread_key;
188 #ifdef USE_COMPILER_TLS
189 static __thread MONO_TLS_FAST void* GC_thread_tls;
192 static GC_bool keys_initialized;
194 /* Recover the contents of the freelist array fl into the global one gfl.*/
195 /* Note that the indexing scheme differs, in that gfl has finer size */
196 /* resolution, even if not all entries are used. */
197 /* We hold the allocator lock. */
198 static void return_freelists(ptr_t *fl, ptr_t *gfl)
204 for (i = 1; i < NFREELISTS; ++i) {
205 nwords = i * (GRANULARITY/sizeof(word));
208 if ((word)q >= HBLKSIZE) {
209 if (gfl[nwords] == 0) {
213 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
219 /* Clear fl[i], since the thread structure may hang around. */
220 /* Do it in a way that is likely to trap if we access it. */
221 fl[i] = (ptr_t)HBLKSIZE;
225 /* We statically allocate a single "size 0" object. It is linked to */
226 /* itself, and is thus repeatedly reused for all size 0 allocation */
227 /* requests. (Size 0 gcj allocation requests are incorrect, and */
228 /* we arrange for those to fault asap.) */
229 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
231 void GC_delete_gc_thread(pthread_t id, GC_thread gct);
232 void GC_destroy_thread_local(GC_thread p);
234 void GC_thread_deregister_foreign (void *data)
236 GC_thread me = (GC_thread)data;
237 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
238 if (me -> flags & FOREIGN_THREAD) {
240 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
241 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
242 GC_destroy_thread_local (me);
244 GC_delete_gc_thread(me->id, me);
249 /* Each thread structure must be initialized. */
250 /* This call must be made from the new thread. */
251 /* Caller holds allocation lock. */
252 void GC_init_thread_local(GC_thread p)
256 if (!keys_initialized) {
257 if (0 != GC_key_create(&GC_thread_key, GC_thread_deregister_foreign)) {
258 ABORT("Failed to create key for local allocator");
260 keys_initialized = TRUE;
262 if (0 != GC_setspecific(GC_thread_key, p)) {
263 ABORT("Failed to set thread specific allocation pointers");
265 for (i = 1; i < NFREELISTS; ++i) {
266 p -> ptrfree_freelists[i] = (ptr_t)1;
267 p -> normal_freelists[i] = (ptr_t)1;
268 # ifdef GC_GCJ_SUPPORT
269 p -> gcj_freelists[i] = (ptr_t)1;
272 /* Set up the size 0 free lists. */
273 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
274 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
275 # ifdef GC_GCJ_SUPPORT
276 p -> gcj_freelists[0] = (ptr_t)(-1);
280 #ifdef GC_GCJ_SUPPORT
281 extern ptr_t * GC_gcjobjfreelist;
284 /* We hold the allocator lock. */
285 void GC_destroy_thread_local(GC_thread p)
287 /* We currently only do this from the thread itself or from */
288 /* the fork handler for a child process. */
290 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
292 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
293 return_freelists(p -> normal_freelists, GC_objfreelist);
294 # ifdef GC_GCJ_SUPPORT
295 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
299 extern GC_PTR GC_generic_malloc_many();
301 GC_PTR GC_local_malloc(size_t bytes)
303 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
304 return(GC_malloc(bytes));
306 int index = INDEX_FROM_BYTES(bytes);
309 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
310 GC_key_t k = GC_thread_key;
314 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
315 if (EXPECT(0 == k, 0)) {
316 /* This can happen if we get called when the world is */
317 /* being initialized. Whether we can actually complete */
318 /* the initialization then is unclear. */
323 tsd = GC_getspecific(GC_thread_key);
324 # ifdef GC_ASSERTIONS
326 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
329 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
331 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
332 ptr_t next = obj_link(my_entry);
333 GC_PTR result = (GC_PTR)my_entry;
335 obj_link(my_entry) = 0;
336 PREFETCH_FOR_WRITE(next);
338 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
339 *my_fl = my_entry + index + 1;
340 return GC_malloc(bytes);
342 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
343 if (*my_fl == 0) return GC_oom_fn(bytes);
344 return GC_local_malloc(bytes);
349 GC_PTR GC_local_malloc_atomic(size_t bytes)
351 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
352 return(GC_malloc_atomic(bytes));
354 int index = INDEX_FROM_BYTES(bytes);
355 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
356 -> ptrfree_freelists + index;
357 ptr_t my_entry = *my_fl;
359 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
360 GC_PTR result = (GC_PTR)my_entry;
361 *my_fl = obj_link(my_entry);
363 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
364 *my_fl = my_entry + index + 1;
365 return GC_malloc_atomic(bytes);
367 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
368 /* *my_fl is updated while the collector is excluded; */
369 /* the free list is always visible to the collector as */
371 if (*my_fl == 0) return GC_oom_fn(bytes);
372 return GC_local_malloc_atomic(bytes);
377 #ifdef GC_GCJ_SUPPORT
379 #include "include/gc_gcj.h"
382 extern GC_bool GC_gcj_malloc_initialized;
385 extern int GC_gcj_kind;
387 GC_PTR GC_local_gcj_malloc(size_t bytes,
388 void * ptr_to_struct_containing_descr)
390 GC_ASSERT(GC_gcj_malloc_initialized);
391 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
392 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
394 int index = INDEX_FROM_BYTES(bytes);
395 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
396 -> gcj_freelists + index;
397 ptr_t my_entry = *my_fl;
398 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
399 GC_PTR result = (GC_PTR)my_entry;
400 GC_ASSERT(!GC_incremental);
401 /* We assert that any concurrent marker will stop us. */
402 /* Thus it is impossible for a mark procedure to see the */
403 /* allocation of the next object, but to see this object */
404 /* still containing a free list pointer. Otherwise the */
405 /* marker might find a random "mark descriptor". */
406 *(volatile ptr_t *)my_fl = obj_link(my_entry);
407 /* We must update the freelist before we store the pointer. */
408 /* Otherwise a GC at this point would see a corrupted */
410 /* A memory barrier is probably never needed, since the */
411 /* action of stopping this thread will cause prior writes */
413 GC_ASSERT(((void * volatile *)result)[1] == 0);
414 *(void * volatile *)result = ptr_to_struct_containing_descr;
416 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
417 if (!GC_incremental) *my_fl = my_entry + index + 1;
418 /* In the incremental case, we always have to take this */
419 /* path. Thus we leave the counter alone. */
420 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
422 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
423 if (*my_fl == 0) return GC_oom_fn(bytes);
424 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
429 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
430 /* adjust it. The size is assumed to be such that it can be */
431 /* allocated as a small object. */
432 void * GC_local_gcj_fast_malloc(size_t lw, void * ptr_to_struct_containing_descr)
434 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
435 -> gcj_freelists + lw;
436 ptr_t my_entry = *my_fl;
438 GC_ASSERT(GC_gcj_malloc_initialized);
440 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
441 GC_PTR result = (GC_PTR)my_entry;
442 GC_ASSERT(!GC_incremental);
443 /* We assert that any concurrent marker will stop us. */
444 /* Thus it is impossible for a mark procedure to see the */
445 /* allocation of the next object, but to see this object */
446 /* still containing a free list pointer. Otherwise the */
447 /* marker might find a random "mark descriptor". */
448 *(volatile ptr_t *)my_fl = obj_link(my_entry);
449 /* We must update the freelist before we store the pointer. */
450 /* Otherwise a GC at this point would see a corrupted */
452 /* A memory barrier is probably never needed, since the */
453 /* action of stopping this thread will cause prior writes */
455 GC_ASSERT(((void * volatile *)result)[1] == 0);
456 *(void * volatile *)result = ptr_to_struct_containing_descr;
458 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
459 if (!GC_incremental) *my_fl = my_entry + lw + 1;
460 /* In the incremental case, we always have to take this */
461 /* path. Thus we leave the counter alone. */
462 return GC_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
464 GC_generic_malloc_many(BYTES_FROM_INDEX(lw), GC_gcj_kind, my_fl);
465 if (*my_fl == 0) return GC_oom_fn(BYTES_FROM_INDEX(lw));
466 return GC_local_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
470 #endif /* GC_GCJ_SUPPORT */
472 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
474 # define GC_destroy_thread_local(t)
476 # endif /* !THREAD_LOCAL_ALLOC */
480 To make sure that we're using LinuxThreads and not some other thread
481 package, we generate a dummy reference to `pthread_kill_other_threads_np'
482 (was `__pthread_initial_thread_bos' but that disappeared),
483 which is a symbol defined in LinuxThreads, but (hopefully) not in other
486 We no longer do this, since this code is now portable enough that it might
487 actually work for something else.
489 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
492 long GC_nprocs = 1; /* Number of processors. We may not have */
493 /* access to all of them, but this is as good */
494 /* a guess as any ... */
499 # define MAX_MARKERS 16
502 static ptr_t marker_sp[MAX_MARKERS] = {0};
504 void * GC_mark_thread(void * id)
508 marker_sp[(word)id] = GC_approx_sp();
509 for (;; ++my_mark_no) {
510 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
511 /* promptly. This is important if it were called from the signal */
512 /* handler or from the GC lock acquisition code. Under Linux, it's */
513 /* not safe to call it from a signal handler, since it uses mutexes */
514 /* and condition variables. Since it is called only here, the */
515 /* argument is unnecessary. */
516 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
517 /* resynchronize if we get far off, e.g. because GC_mark_no */
519 my_mark_no = GC_mark_no;
521 # ifdef DEBUG_THREADS
522 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
524 GC_help_marker(my_mark_no);
528 extern long GC_markers; /* Number of mark threads we would */
529 /* like to have. Includes the */
530 /* initiating thread. */
532 pthread_t GC_mark_threads[MAX_MARKERS];
534 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
536 static void start_mark_threads()
541 if (GC_markers > MAX_MARKERS) {
542 WARN("Limiting number of mark threads\n", 0);
543 GC_markers = MAX_MARKERS;
545 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
547 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
548 ABORT("pthread_attr_setdetachstate failed");
550 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
551 /* Default stack size is usually too small: fix it. */
552 /* Otherwise marker threads or GC may run out of */
554 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
559 if (pthread_attr_getstacksize(&attr, &old_size) != 0)
560 ABORT("pthread_attr_getstacksize failed\n");
561 if (old_size < MIN_STACK_SIZE) {
562 if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
563 ABORT("pthread_attr_setstacksize failed\n");
566 # endif /* HPUX || GC_DGUX386_THREADS */
568 if (GC_print_stats) {
569 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
572 for (i = 0; i < GC_markers - 1; ++i) {
573 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
574 GC_mark_thread, (void *)(word)i)) {
575 WARN("Marker thread creation failed, errno = %ld.\n", errno);
580 #else /* !PARALLEL_MARK */
582 static __inline__ void start_mark_threads()
586 #endif /* !PARALLEL_MARK */
588 GC_bool GC_thr_initialized = FALSE;
590 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
593 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
594 * put it into a separate function.
596 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
597 static __attribute__((noinline)) unsigned char* get_gc_thread_key_addr GC_PROTO((void))
599 return (unsigned char*)&GC_thread_key;
602 void GC_push_thread_structures GC_PROTO((void))
604 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
605 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
606 GC_push_all((ptr_t)get_gc_thread_key_addr(),
607 (ptr_t)(get_gc_thread_key_addr())+sizeof(&GC_thread_key));
613 void GC_push_thread_structures GC_PROTO((void))
615 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
616 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
617 GC_push_all((ptr_t)(&GC_thread_key),
618 (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
624 #ifdef THREAD_LOCAL_ALLOC
625 /* We must explicitly mark ptrfree and gcj free lists, since the free */
626 /* list links wouldn't otherwise be found. We also set them in the */
627 /* normal free lists, since that involves touching less memory than if */
628 /* we scanned them normally. */
629 void GC_mark_thread_local_free_lists(void)
635 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
636 for (p = GC_threads[i]; 0 != p; p = p -> next) {
637 for (j = 1; j < NFREELISTS; ++j) {
638 q = p -> ptrfree_freelists[j];
639 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
640 q = p -> normal_freelists[j];
641 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
642 # ifdef GC_GCJ_SUPPORT
643 q = p -> gcj_freelists[j];
644 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
645 # endif /* GC_GCJ_SUPPORT */
650 #endif /* THREAD_LOCAL_ALLOC */
652 static struct GC_Thread_Rep first_thread;
654 /* Add a thread to GC_threads. We assume it wasn't already there. */
655 /* Caller holds allocation lock. */
656 GC_thread GC_new_thread(pthread_t id)
658 int hv = ((word)id) % THREAD_TABLE_SZ;
660 static GC_bool first_thread_used = FALSE;
662 if (!first_thread_used) {
663 result = &first_thread;
664 first_thread_used = TRUE;
666 result = (struct GC_Thread_Rep *)
667 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
669 if (result == 0) return(0);
671 result -> next = GC_threads[hv];
672 GC_threads[hv] = result;
673 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
677 /* Delete a thread from GC_threads. We assume it is there. */
678 /* (The code intentionally traps if it wasn't.) */
679 /* Caller holds allocation lock. */
680 void GC_delete_thread(pthread_t id)
682 int hv = ((word)id) % THREAD_TABLE_SZ;
683 register GC_thread p = GC_threads[hv];
684 register GC_thread prev = 0;
686 while (!pthread_equal(p -> id, id)) {
691 GC_threads[hv] = p -> next;
693 prev -> next = p -> next;
698 /* If a thread has been joined, but we have not yet */
699 /* been notified, then there may be more than one thread */
700 /* in the table with the same pthread id. */
701 /* This is OK, but we need a way to delete a specific one. */
702 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
704 int hv = ((word)id) % THREAD_TABLE_SZ;
705 register GC_thread p = GC_threads[hv];
706 register GC_thread prev = 0;
713 GC_threads[hv] = p -> next;
715 prev -> next = p -> next;
720 /* Return a GC_thread corresponding to a given pthread_t. */
721 /* Returns 0 if it's not there. */
722 /* Caller holds allocation lock or otherwise inhibits */
724 /* If there is more than one thread with the given id we */
725 /* return the most recent one. */
726 GC_thread GC_lookup_thread(pthread_t id)
728 int hv = ((word)id) % THREAD_TABLE_SZ;
729 register GC_thread p = GC_threads[hv];
731 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
735 int GC_thread_is_registered (void)
740 ptr = (void *)GC_lookup_thread(pthread_self());
747 /* Remove all entries from the GC_threads table, except the */
748 /* one for the current thread. We need to do this in the child */
749 /* process after a fork(), since only the current thread */
750 /* survives in the child. */
751 void GC_remove_all_threads_but_me(void)
753 pthread_t self = pthread_self();
755 GC_thread p, next, me;
757 for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
759 for (p = GC_threads[hv]; 0 != p; p = next) {
761 if (p -> id == self) {
765 # ifdef THREAD_LOCAL_ALLOC
766 if (!(p -> flags & FINISHED)) {
767 GC_destroy_thread_local(p);
769 # endif /* THREAD_LOCAL_ALLOC */
770 if (p != &first_thread) GC_INTERNAL_FREE(p);
776 #endif /* HANDLE_FORK */
778 #ifdef USE_PROC_FOR_LIBRARIES
779 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
784 # ifdef PARALLEL_MARK
785 for (i = 0; i < GC_markers; ++i) {
786 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
789 for (i = 0; i < THREAD_TABLE_SZ; i++) {
790 for (p = GC_threads[i]; p != 0; p = p -> next) {
791 if (0 != p -> stack_end) {
792 # ifdef STACK_GROWS_UP
793 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
794 # else /* STACK_GROWS_DOWN */
795 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
802 #endif /* USE_PROC_FOR_LIBRARIES */
804 #ifdef GC_LINUX_THREADS
805 /* Return the number of processors, or i<= 0 if it can't be determined. */
808 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
809 /* appears to be buggy in many cases. */
810 /* We look for lines "cpu<n>" in /proc/stat. */
811 # define STAT_BUF_SIZE 4096
812 # define STAT_READ read
813 /* If read is wrapped, this may need to be redefined to call */
815 char stat_buf[STAT_BUF_SIZE];
818 /* Some old kernels only have a single "cpu nnnn ..." */
819 /* entry in /proc/stat. We identify those as */
823 f = open("/proc/stat", O_RDONLY);
824 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
825 WARN("Couldn't read /proc/stat\n", 0);
828 for (i = 0; i < len - 100; ++i) {
829 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
830 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
831 int cpu_no = atoi(stat_buf + i + 4);
832 if (cpu_no >= result) result = cpu_no + 1;
838 #endif /* GC_LINUX_THREADS */
840 /* We hold the GC lock. Wait until an in-progress GC has finished. */
841 /* Repeatedly RELEASES GC LOCK in order to wait. */
842 /* If wait_for_all is true, then we exit with the GC lock held and no */
843 /* collection in progress; otherwise we just wait for the current GC */
845 extern GC_bool GC_collection_in_progress();
846 void GC_wait_for_gc_completion(GC_bool wait_for_all)
848 if (GC_incremental && GC_collection_in_progress()) {
849 int old_gc_no = GC_gc_no;
851 /* Make sure that no part of our stack is still on the mark stack, */
852 /* since it's about to be unmapped. */
853 while (GC_incremental && GC_collection_in_progress()
854 && (wait_for_all || old_gc_no == GC_gc_no)) {
856 GC_in_thread_creation = TRUE;
857 GC_collect_a_little_inner(1);
858 GC_in_thread_creation = FALSE;
868 /* Procedures called before and after a fork. The goal here is to make */
869 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
870 /* attainable, since the single UNIX spec seems to imply that one */
871 /* should only call async-signal-safe functions, and we probably can't */
872 /* quite guarantee that. But we give it our best shot. (That same */
873 /* spec also implies that it's not safe to call the system malloc */
874 /* between fork() and exec(). Thus we're doing no worse than it. */
876 /* Called before a fork() */
877 void GC_fork_prepare_proc(void)
879 /* Acquire all relevant locks, so that after releasing the locks */
880 /* the child will see a consistent state in which monitor */
881 /* invariants hold. Unfortunately, we can't acquire libc locks */
882 /* we might need, and there seems to be no guarantee that libc */
883 /* must install a suitable fork handler. */
884 /* Wait for an ongoing GC to finish, since we can't finish it in */
885 /* the (one remaining thread in) the child. */
887 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
888 GC_wait_for_reclaim();
890 GC_wait_for_gc_completion(TRUE);
891 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
892 GC_acquire_mark_lock();
896 /* Called in parent after a fork() */
897 void GC_fork_parent_proc(void)
899 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
900 GC_release_mark_lock();
905 /* Called in child after a fork() */
906 void GC_fork_child_proc(void)
908 /* Clean up the thread table, so that just our thread is left. */
909 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
910 GC_release_mark_lock();
912 GC_remove_all_threads_but_me();
913 # ifdef PARALLEL_MARK
914 /* Turn off parallel marking in the child, since we are probably */
915 /* just going to exec, and we would have to restart mark threads. */
918 # endif /* PARALLEL_MARK */
921 #endif /* HANDLE_FORK */
923 #if defined(GC_DGUX386_THREADS)
924 /* Return the number of processors, or i<= 0 if it can't be determined. */
927 /* <takis@XFree86.Org> */
929 struct dg_sys_info_pm_info pm_sysinfo;
932 status = dg_sys_info((long int *) &pm_sysinfo,
933 DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
935 /* set -1 for error */
939 numCpus = pm_sysinfo.idle_vp_count;
941 # ifdef DEBUG_THREADS
942 GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
946 #endif /* GC_DGUX386_THREADS */
948 /* We hold the allocation lock. */
951 # ifndef GC_DARWIN_THREADS
956 if (GC_thr_initialized) return;
957 GC_thr_initialized = TRUE;
960 /* Prepare for a possible fork. */
961 pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
963 # endif /* HANDLE_FORK */
964 /* Add the initial thread, so we can stop it. */
965 t = GC_new_thread(pthread_self());
966 # ifdef GC_DARWIN_THREADS
967 t -> stop_info.mach_thread = mach_thread_self();
969 t -> stop_info.stack_ptr = (ptr_t)(&dummy);
971 t -> flags = DETACHED | MAIN_THREAD;
977 char * nprocs_string = GETENV("GC_NPROCS");
979 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
981 if (GC_nprocs <= 0) {
982 # if defined(GC_HPUX_THREADS)
983 GC_nprocs = pthread_num_processors_np();
985 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
986 GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
987 if (GC_nprocs <= 0) GC_nprocs = 1;
989 # if defined(GC_IRIX_THREADS)
990 GC_nprocs = sysconf(_SC_NPROC_ONLN);
991 if (GC_nprocs <= 0) GC_nprocs = 1;
993 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
995 size_t len = sizeof(ncpus);
996 sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
999 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1000 GC_nprocs = GC_get_nprocs();
1003 if (GC_nprocs <= 0) {
1004 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
1006 # ifdef PARALLEL_MARK
1010 # ifdef PARALLEL_MARK
1012 char * markers_string = GETENV("GC_MARKERS");
1013 if (markers_string != NULL) {
1014 GC_markers = atoi(markers_string);
1016 GC_markers = GC_nprocs;
1021 # ifdef PARALLEL_MARK
1023 if (GC_print_stats) {
1024 GC_printf2("Number of processors = %ld, "
1025 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
1028 if (GC_markers == 1) {
1029 GC_parallel = FALSE;
1031 if (GC_print_stats) {
1032 GC_printf0("Single marker thread, turning off parallel marking\n");
1037 /* Disable true incremental collection, but generational is OK. */
1038 GC_time_limit = GC_TIME_UNLIMITED;
1040 /* If we are using a parallel marker, actually start helper threads. */
1041 if (GC_parallel) start_mark_threads();
1046 /* Perform all initializations, including those that */
1047 /* may require allocation. */
1048 /* Called without allocation lock. */
1049 /* Must be called before a second thread is created. */
1050 /* Called without allocation lock. */
1051 void GC_init_parallel()
1053 if (parallel_initialized) return;
1054 parallel_initialized = TRUE;
1056 /* GC_init() calls us back, so set flag first. */
1057 if (!GC_is_initialized) GC_init();
1058 /* Initialize thread local free lists if used. */
1059 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1061 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1067 #if !defined(GC_DARWIN_THREADS)
1068 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
1070 sigset_t fudged_set;
1072 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
1074 sigdelset(&fudged_set, SIG_SUSPEND);
1077 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
1079 #endif /* !GC_DARWIN_THREADS */
1081 /* Wrappers for functions that are likely to block for an appreciable */
1082 /* length of time. Must be called in pairs, if at all. */
1083 /* Nothing much beyond the system call itself should be executed */
1084 /* between these. */
1086 void GC_start_blocking(void) {
1087 # define SP_SLOP 128
1090 me = GC_lookup_thread(pthread_self());
1091 GC_ASSERT(!(me -> thread_blocked));
1093 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
1095 # ifndef GC_DARWIN_THREADS
1096 me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
1100 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1102 /* Add some slop to the stack pointer, since the wrapped call may */
1103 /* end up pushing more callee-save registers. */
1104 # ifndef GC_DARWIN_THREADS
1105 # ifdef STACK_GROWS_UP
1106 me -> stop_info.stack_ptr += SP_SLOP;
1108 me -> stop_info.stack_ptr -= SP_SLOP;
1111 me -> thread_blocked = TRUE;
1115 void GC_end_blocking(void) {
1117 LOCK(); /* This will block if the world is stopped. */
1118 me = GC_lookup_thread(pthread_self());
1119 GC_ASSERT(me -> thread_blocked);
1120 me -> thread_blocked = FALSE;
1124 #if defined(GC_DGUX386_THREADS)
1125 #define __d10_sleep sleep
1126 #endif /* GC_DGUX386_THREADS */
1128 /* A wrapper for the standard C sleep function */
1129 int WRAP_FUNC(sleep) (unsigned int seconds)
1133 GC_start_blocking();
1134 result = REAL_FUNC(sleep)(seconds);
1140 void *(*start_routine)(void *);
1143 sem_t registered; /* 1 ==> in our thread table, but */
1144 /* parent hasn't yet noticed. */
1147 /* Called at thread exit. */
1148 /* Never called for main thread. That's OK, since it */
1149 /* results in at most a tiny one-time leak. And */
1150 /* linuxthreads doesn't reclaim the main threads */
1151 /* resources or id anyway. */
1152 void GC_thread_exit_proc(void *arg)
1157 me = GC_lookup_thread(pthread_self());
1158 GC_destroy_thread_local(me);
1159 if (me -> flags & DETACHED) {
1160 GC_delete_thread(pthread_self());
1162 me -> flags |= FINISHED;
1164 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1165 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1166 GC_remove_specific(GC_thread_key);
1168 /* The following may run the GC from "nonexistent" thread. */
1169 GC_wait_for_gc_completion(FALSE);
1173 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1176 GC_thread thread_gc_id;
1179 thread_gc_id = GC_lookup_thread(thread);
1180 /* This is guaranteed to be the intended one, since the thread id */
1181 /* cant have been recycled by pthreads. */
1183 result = REAL_FUNC(pthread_join)(thread, retval);
1184 # if defined (GC_FREEBSD_THREADS)
1185 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1186 appears to be) a spurious EINTR which caused the test and real code
1187 to gratuitously fail. Having looked at system pthread library source
1188 code, I see how this return code may be generated. In one path of
1189 code, pthread_join() just returns the errno setting of the thread
1190 being joined. This does not match the POSIX specification or the
1191 local man pages thus I have taken the liberty to catch this one
1192 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1193 if (result == EINTR) result = 0;
1197 /* Here the pthread thread id may have been recycled. */
1198 GC_delete_gc_thread(thread, thread_gc_id);
1205 WRAP_FUNC(pthread_detach)(pthread_t thread)
1208 GC_thread thread_gc_id;
1211 thread_gc_id = GC_lookup_thread(thread);
1213 result = REAL_FUNC(pthread_detach)(thread);
1216 thread_gc_id -> flags |= DETACHED;
1217 /* Here the pthread thread id may have been recycled. */
1218 if (thread_gc_id -> flags & FINISHED) {
1219 GC_delete_gc_thread(thread, thread_gc_id);
1226 GC_bool GC_in_thread_creation = FALSE;
1228 typedef void *(*ThreadStartFn)(void *);
1229 void * GC_start_routine_head(void * arg, void *base_addr,
1230 ThreadStartFn *start, void **start_arg )
1232 struct start_info * si = arg;
1235 pthread_t my_pthread;
1237 my_pthread = pthread_self();
1238 # ifdef DEBUG_THREADS
1239 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1240 GC_printf1("pid = %ld\n", (long) getpid());
1241 GC_printf1("sp = 0x%lx\n", (long) &arg);
1244 GC_in_thread_creation = TRUE;
1245 me = GC_new_thread(my_pthread);
1246 GC_in_thread_creation = FALSE;
1247 #ifdef GC_DARWIN_THREADS
1248 me -> stop_info.mach_thread = mach_thread_self();
1250 me -> stop_info.stack_ptr = 0;
1252 me -> flags = si -> flags;
1253 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1254 /* doesn't work because the stack base in /proc/self/stat is the */
1255 /* one for the main thread. There is a strong argument that that's */
1256 /* a kernel bug, but a pervasive one. */
1257 # ifdef STACK_GROWS_DOWN
1258 me -> stack_end = (ptr_t)(((word)(base_addr) + (GC_page_size - 1))
1259 & ~(GC_page_size - 1));
1260 # ifndef GC_DARWIN_THREADS
1261 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1263 /* Needs to be plausible, since an asynchronous stack mark */
1264 /* should not crash. */
1266 me -> stack_end = (ptr_t)((word)(base_addr) & ~(GC_page_size - 1));
1267 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1269 /* This is dubious, since we may be more than a page into the stack, */
1270 /* and hence skip some of it, though it's not clear that matters. */
1272 me -> backing_store_end = (ptr_t)
1273 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1274 /* This is also < 100% convincing. We should also read this */
1275 /* from /proc, but the hook to do so isn't there yet. */
1279 if (start) *start = si -> start_routine;
1280 if (start_arg) *start_arg = si -> arg;
1282 sem_post(&(si -> registered)); /* Last action on si. */
1283 /* OK to deallocate. */
1284 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1286 GC_init_thread_local(me);
1293 int GC_thread_register_foreign (void *base_addr)
1295 struct start_info si = { 0, }; /* stacked for legibility & locking */
1298 # ifdef DEBUG_THREADS
1299 GC_printf1( "GC_thread_register_foreign %p\n", &si );
1302 si.flags = FOREIGN_THREAD;
1304 if (!parallel_initialized) GC_init_parallel();
1306 if (!GC_thr_initialized) GC_thr_init();
1310 me = GC_start_routine_head(&si, base_addr, NULL, NULL);
1315 void * GC_start_routine(void * arg)
1318 struct start_info * si = arg;
1321 ThreadStartFn start;
1324 me = GC_start_routine_head (arg, &dummy, &start, &start_arg);
1326 pthread_cleanup_push(GC_thread_exit_proc, 0);
1327 # ifdef DEBUG_THREADS
1328 GC_printf1("start_routine = 0x%lx\n", start);
1330 result = (*start)(start_arg);
1332 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1334 me -> status = result;
1335 pthread_cleanup_pop(1);
1336 /* Cleanup acquires lock, ensuring that we can't exit */
1337 /* while a collection that thinks we're alive is trying to stop */
1343 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1344 const pthread_attr_t *attr,
1345 void *(*start_routine)(void *), void *arg)
1350 struct start_info * si;
1351 /* This is otherwise saved only in an area mmapped by the thread */
1352 /* library, which isn't visible to the collector. */
1354 /* We resist the temptation to muck with the stack size here, */
1355 /* even if the default is unreasonably small. That's the client's */
1356 /* responsibility. */
1359 si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
1362 if (!parallel_initialized) GC_init_parallel();
1363 if (0 == si) return(ENOMEM);
1364 sem_init(&(si -> registered), 0, 0);
1365 si -> start_routine = start_routine;
1368 if (!GC_thr_initialized) GC_thr_init();
1369 # ifdef GC_ASSERTIONS
1373 pthread_attr_t my_attr;
1374 pthread_attr_init(&my_attr);
1375 pthread_attr_getstacksize(&my_attr, &stack_size);
1377 pthread_attr_getstacksize(attr, &stack_size);
1379 # ifdef PARALLEL_MARK
1380 GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1382 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1383 /* HBLKSIZE=8192, sizeof(word)=8 */
1384 GC_ASSERT(stack_size >= 65536);
1386 /* Our threads may need to do some work for the GC. */
1387 /* Ridiculously small threads won't work, and they */
1388 /* probably wouldn't work anyway. */
1392 detachstate = PTHREAD_CREATE_JOINABLE;
1394 pthread_attr_getdetachstate(attr, &detachstate);
1396 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1397 si -> flags = my_flags;
1399 # ifdef DEBUG_THREADS
1400 GC_printf1("About to start new thread from thread 0x%X\n",
1404 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1406 # ifdef DEBUG_THREADS
1407 GC_printf1("Started thread 0x%X\n", *new_thread);
1409 /* Wait until child has been added to the thread table. */
1410 /* This also ensures that we hold onto si until the child is done */
1411 /* with it. Thus it doesn't matter whether it is otherwise */
1412 /* visible to the collector. */
1414 while (0 != sem_wait(&(si -> registered))) {
1415 if (EINTR != errno) ABORT("sem_wait failed");
1418 sem_destroy(&(si -> registered));
1420 GC_INTERNAL_FREE(si);
1426 #ifdef GENERIC_COMPARE_AND_SWAP
1427 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1429 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1430 GC_word old, GC_word new_val)
1433 pthread_mutex_lock(&GC_compare_and_swap_lock);
1440 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1444 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1447 pthread_mutex_lock(&GC_compare_and_swap_lock);
1449 *addr = old + how_much;
1450 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1454 #endif /* GENERIC_COMPARE_AND_SWAP */
1455 /* Spend a few cycles in a way that can't introduce contention with */
1456 /* othre threads. */
1460 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1461 volatile word dummy = 0;
1464 for (i = 0; i < 10; ++i) {
1465 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1466 __asm__ __volatile__ (" " : : : "memory");
1468 /* Something that's unlikely to be optimized away. */
1474 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1477 VOLATILE GC_bool GC_collecting = 0;
1478 /* A hint that we're in the collector and */
1479 /* holding the allocation lock for an */
1480 /* extended period. */
1482 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1483 /* If we don't want to use the below spinlock implementation, either */
1484 /* because we don't have a GC_test_and_set implementation, or because */
1485 /* we don't want to risk sleeping, we can still try spinning on */
1486 /* pthread_mutex_trylock for a while. This appears to be very */
1487 /* beneficial in many cases. */
1488 /* I suspect that under high contention this is nearly always better */
1489 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1490 /* Hence we still default to the spin lock. */
1491 /* This is also used to acquire the mark lock for the parallel */
1494 /* Here we use a strict exponential backoff scheme. I don't know */
1495 /* whether that's better or worse than the above. We eventually */
1496 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1497 /* explicitly sleep. */
1501 unsigned long GC_spin_count = 0;
1502 unsigned long GC_block_count = 0;
1503 unsigned long GC_unlocked_count = 0;
1506 void GC_generic_lock(pthread_mutex_t * lock)
1508 #ifndef NO_PTHREAD_TRYLOCK
1509 unsigned pause_length = 1;
1512 if (0 == pthread_mutex_trylock(lock)) {
1514 ++GC_unlocked_count;
1518 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1519 for (i = 0; i < pause_length; ++i) {
1522 switch(pthread_mutex_trylock(lock)) {
1531 ABORT("Unexpected error from pthread_mutex_trylock");
1534 #endif /* !NO_PTHREAD_TRYLOCK */
1538 pthread_mutex_lock(lock);
1541 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1543 #if defined(USE_SPIN_LOCK)
1545 /* Reasonably fast spin locks. Basically the same implementation */
1546 /* as STL alloc.h. This isn't really the right way to do this. */
1547 /* but until the POSIX scheduling mess gets straightened out ... */
1549 volatile unsigned int GC_allocate_lock = 0;
1554 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1555 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1556 static unsigned spin_max = low_spin_max;
1557 unsigned my_spin_max;
1558 static unsigned last_spins = 0;
1559 unsigned my_last_spins;
1562 if (!GC_test_and_set(&GC_allocate_lock)) {
1565 my_spin_max = spin_max;
1566 my_last_spins = last_spins;
1567 for (i = 0; i < my_spin_max; i++) {
1568 if (GC_collecting || GC_nprocs == 1) goto yield;
1569 if (i < my_last_spins/2 || GC_allocate_lock) {
1573 if (!GC_test_and_set(&GC_allocate_lock)) {
1576 * Spinning worked. Thus we're probably not being scheduled
1577 * against the other process with which we were contending.
1578 * Thus it makes sense to spin longer the next time.
1581 spin_max = high_spin_max;
1585 /* We are probably being scheduled against the other process. Sleep. */
1586 spin_max = low_spin_max;
1589 if (!GC_test_and_set(&GC_allocate_lock)) {
1592 # define SLEEP_THRESHOLD 12
1593 /* Under Linux very short sleeps tend to wait until */
1594 /* the current time quantum expires. On old Linux */
1595 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1596 /* (Under 2.4, this happens only for real-time */
1597 /* processes.) We want to minimize both behaviors */
1599 if (i < SLEEP_THRESHOLD) {
1605 /* Don't wait for more than about 15msecs, even */
1606 /* under extreme contention. */
1608 ts.tv_nsec = 1 << i;
1614 #else /* !USE_SPINLOCK */
1617 #ifndef NO_PTHREAD_TRYLOCK
1618 if (1 == GC_nprocs || GC_collecting) {
1619 pthread_mutex_lock(&GC_allocate_ml);
1621 GC_generic_lock(&GC_allocate_ml);
1623 #else /* !NO_PTHREAD_TRYLOCK */
1624 pthread_mutex_lock(&GC_allocate_ml);
1625 #endif /* !NO_PTHREAD_TRYLOCK */
1628 #endif /* !USE_SPINLOCK */
1630 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1632 #ifdef GC_ASSERTIONS
1633 pthread_t GC_mark_lock_holder = NO_THREAD;
1637 /* Ugly workaround for a linux threads bug in the final versions */
1638 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1639 /* field even when it fails to acquire the mutex. This causes */
1640 /* pthread_cond_wait to die. Remove for glibc2.2. */
1641 /* According to the man page, we should use */
1642 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1644 static pthread_mutex_t mark_mutex =
1645 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1647 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1650 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1652 void GC_acquire_mark_lock()
1655 if (pthread_mutex_lock(&mark_mutex) != 0) {
1656 ABORT("pthread_mutex_lock failed");
1659 GC_generic_lock(&mark_mutex);
1660 # ifdef GC_ASSERTIONS
1661 GC_mark_lock_holder = pthread_self();
1665 void GC_release_mark_lock()
1667 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1668 # ifdef GC_ASSERTIONS
1669 GC_mark_lock_holder = NO_THREAD;
1671 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1672 ABORT("pthread_mutex_unlock failed");
1676 /* Collector must wait for a freelist builders for 2 reasons: */
1677 /* 1) Mark bits may still be getting examined without lock. */
1678 /* 2) Partial free lists referenced only by locals may not be scanned */
1679 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1680 /* free-list link may be ignored. */
1681 void GC_wait_builder()
1683 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1684 # ifdef GC_ASSERTIONS
1685 GC_mark_lock_holder = NO_THREAD;
1687 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1688 ABORT("pthread_cond_wait failed");
1690 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1691 # ifdef GC_ASSERTIONS
1692 GC_mark_lock_holder = pthread_self();
1696 void GC_wait_for_reclaim()
1698 GC_acquire_mark_lock();
1699 while (GC_fl_builder_count > 0) {
1702 GC_release_mark_lock();
1705 void GC_notify_all_builder()
1707 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1708 if (pthread_cond_broadcast(&builder_cv) != 0) {
1709 ABORT("pthread_cond_broadcast failed");
1713 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1715 #ifdef PARALLEL_MARK
1717 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1719 void GC_wait_marker()
1721 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1722 # ifdef GC_ASSERTIONS
1723 GC_mark_lock_holder = NO_THREAD;
1725 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1726 ABORT("pthread_cond_wait failed");
1728 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1729 # ifdef GC_ASSERTIONS
1730 GC_mark_lock_holder = pthread_self();
1734 void GC_notify_all_marker()
1736 if (pthread_cond_broadcast(&mark_cv) != 0) {
1737 ABORT("pthread_cond_broadcast failed");
1741 #endif /* PARALLEL_MARK */
1743 # endif /* GC_LINUX_THREADS and friends */