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 #ifdef MONO_DEBUGGER_SUPPORTED
195 #include "include/libgc-mono-debugger.h"
198 /* Recover the contents of the freelist array fl into the global one gfl.*/
199 /* Note that the indexing scheme differs, in that gfl has finer size */
200 /* resolution, even if not all entries are used. */
201 /* We hold the allocator lock. */
202 static void return_freelists(ptr_t *fl, ptr_t *gfl)
208 for (i = 1; i < NFREELISTS; ++i) {
209 nwords = i * (GRANULARITY/sizeof(word));
212 if ((word)q >= HBLKSIZE) {
213 if (gfl[nwords] == 0) {
217 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
223 /* Clear fl[i], since the thread structure may hang around. */
224 /* Do it in a way that is likely to trap if we access it. */
225 fl[i] = (ptr_t)HBLKSIZE;
229 /* We statically allocate a single "size 0" object. It is linked to */
230 /* itself, and is thus repeatedly reused for all size 0 allocation */
231 /* requests. (Size 0 gcj allocation requests are incorrect, and */
232 /* we arrange for those to fault asap.) */
233 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
235 void GC_delete_gc_thread(pthread_t id, GC_thread gct);
236 void GC_destroy_thread_local(GC_thread p);
238 void GC_thread_deregister_foreign (void *data)
240 GC_thread me = (GC_thread)data;
241 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
242 if (me -> flags & FOREIGN_THREAD) {
244 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
245 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
246 GC_destroy_thread_local (me);
248 GC_delete_gc_thread(me->id, me);
253 /* Each thread structure must be initialized. */
254 /* This call must be made from the new thread. */
255 /* Caller holds allocation lock. */
256 void GC_init_thread_local(GC_thread p)
260 if (!keys_initialized) {
261 if (0 != GC_key_create(&GC_thread_key, GC_thread_deregister_foreign)) {
262 ABORT("Failed to create key for local allocator");
264 keys_initialized = TRUE;
266 if (0 != GC_setspecific(GC_thread_key, p)) {
267 ABORT("Failed to set thread specific allocation pointers");
269 for (i = 1; i < NFREELISTS; ++i) {
270 p -> ptrfree_freelists[i] = (ptr_t)1;
271 p -> normal_freelists[i] = (ptr_t)1;
272 # ifdef GC_GCJ_SUPPORT
273 p -> gcj_freelists[i] = (ptr_t)1;
276 /* Set up the size 0 free lists. */
277 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
278 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
279 # ifdef GC_GCJ_SUPPORT
280 p -> gcj_freelists[0] = (ptr_t)(-1);
284 #ifdef GC_GCJ_SUPPORT
285 extern ptr_t * GC_gcjobjfreelist;
288 /* We hold the allocator lock. */
289 void GC_destroy_thread_local(GC_thread p)
291 /* We currently only do this from the thread itself or from */
292 /* the fork handler for a child process. */
294 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
296 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
297 return_freelists(p -> normal_freelists, GC_objfreelist);
298 # ifdef GC_GCJ_SUPPORT
299 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
303 extern GC_PTR GC_generic_malloc_many();
305 GC_PTR GC_local_malloc(size_t bytes)
307 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
308 return(GC_malloc(bytes));
310 int index = INDEX_FROM_BYTES(bytes);
313 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
314 GC_key_t k = GC_thread_key;
318 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
319 if (EXPECT(0 == k, 0)) {
320 /* This can happen if we get called when the world is */
321 /* being initialized. Whether we can actually complete */
322 /* the initialization then is unclear. */
327 tsd = GC_getspecific(GC_thread_key);
328 # ifdef GC_ASSERTIONS
330 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
333 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
335 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
336 ptr_t next = obj_link(my_entry);
337 GC_PTR result = (GC_PTR)my_entry;
339 obj_link(my_entry) = 0;
340 PREFETCH_FOR_WRITE(next);
342 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
343 *my_fl = my_entry + index + 1;
344 return GC_malloc(bytes);
346 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
347 if (*my_fl == 0) return GC_oom_fn(bytes);
348 return GC_local_malloc(bytes);
353 GC_PTR GC_local_malloc_atomic(size_t bytes)
355 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
356 return(GC_malloc_atomic(bytes));
358 int index = INDEX_FROM_BYTES(bytes);
359 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
360 -> ptrfree_freelists + index;
361 ptr_t my_entry = *my_fl;
363 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
364 GC_PTR result = (GC_PTR)my_entry;
365 *my_fl = obj_link(my_entry);
367 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
368 *my_fl = my_entry + index + 1;
369 return GC_malloc_atomic(bytes);
371 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
372 /* *my_fl is updated while the collector is excluded; */
373 /* the free list is always visible to the collector as */
375 if (*my_fl == 0) return GC_oom_fn(bytes);
376 return GC_local_malloc_atomic(bytes);
381 #ifdef GC_GCJ_SUPPORT
383 #include "include/gc_gcj.h"
386 extern GC_bool GC_gcj_malloc_initialized;
389 extern int GC_gcj_kind;
391 GC_PTR GC_local_gcj_malloc(size_t bytes,
392 void * ptr_to_struct_containing_descr)
394 GC_ASSERT(GC_gcj_malloc_initialized);
395 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
396 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
398 int index = INDEX_FROM_BYTES(bytes);
399 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
400 -> gcj_freelists + index;
401 ptr_t my_entry = *my_fl;
402 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
403 GC_PTR result = (GC_PTR)my_entry;
404 GC_ASSERT(!GC_incremental);
405 /* We assert that any concurrent marker will stop us. */
406 /* Thus it is impossible for a mark procedure to see the */
407 /* allocation of the next object, but to see this object */
408 /* still containing a free list pointer. Otherwise the */
409 /* marker might find a random "mark descriptor". */
410 *(volatile ptr_t *)my_fl = obj_link(my_entry);
411 /* We must update the freelist before we store the pointer. */
412 /* Otherwise a GC at this point would see a corrupted */
414 /* A memory barrier is probably never needed, since the */
415 /* action of stopping this thread will cause prior writes */
417 GC_ASSERT(((void * volatile *)result)[1] == 0);
418 *(void * volatile *)result = ptr_to_struct_containing_descr;
420 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
421 if (!GC_incremental) *my_fl = my_entry + index + 1;
422 /* In the incremental case, we always have to take this */
423 /* path. Thus we leave the counter alone. */
424 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
426 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
427 if (*my_fl == 0) return GC_oom_fn(bytes);
428 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
433 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
434 /* adjust it. The size is assumed to be such that it can be */
435 /* allocated as a small object. */
436 void * GC_local_gcj_fast_malloc(size_t lw, void * ptr_to_struct_containing_descr)
438 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
439 -> gcj_freelists + lw;
440 ptr_t my_entry = *my_fl;
442 GC_ASSERT(GC_gcj_malloc_initialized);
444 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
445 GC_PTR result = (GC_PTR)my_entry;
446 GC_ASSERT(!GC_incremental);
447 /* We assert that any concurrent marker will stop us. */
448 /* Thus it is impossible for a mark procedure to see the */
449 /* allocation of the next object, but to see this object */
450 /* still containing a free list pointer. Otherwise the */
451 /* marker might find a random "mark descriptor". */
452 *(volatile ptr_t *)my_fl = obj_link(my_entry);
453 /* We must update the freelist before we store the pointer. */
454 /* Otherwise a GC at this point would see a corrupted */
456 /* A memory barrier is probably never needed, since the */
457 /* action of stopping this thread will cause prior writes */
459 GC_ASSERT(((void * volatile *)result)[1] == 0);
460 *(void * volatile *)result = ptr_to_struct_containing_descr;
462 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
463 if (!GC_incremental) *my_fl = my_entry + lw + 1;
464 /* In the incremental case, we always have to take this */
465 /* path. Thus we leave the counter alone. */
466 return GC_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
468 GC_generic_malloc_many(BYTES_FROM_INDEX(lw), GC_gcj_kind, my_fl);
469 if (*my_fl == 0) return GC_oom_fn(BYTES_FROM_INDEX(lw));
470 return GC_local_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
474 #endif /* GC_GCJ_SUPPORT */
476 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
478 # define GC_destroy_thread_local(t)
480 # endif /* !THREAD_LOCAL_ALLOC */
484 To make sure that we're using LinuxThreads and not some other thread
485 package, we generate a dummy reference to `pthread_kill_other_threads_np'
486 (was `__pthread_initial_thread_bos' but that disappeared),
487 which is a symbol defined in LinuxThreads, but (hopefully) not in other
490 We no longer do this, since this code is now portable enough that it might
491 actually work for something else.
493 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
496 long GC_nprocs = 1; /* Number of processors. We may not have */
497 /* access to all of them, but this is as good */
498 /* a guess as any ... */
503 # define MAX_MARKERS 16
506 static ptr_t marker_sp[MAX_MARKERS] = {0};
508 void * GC_mark_thread(void * id)
512 marker_sp[(word)id] = GC_approx_sp();
513 for (;; ++my_mark_no) {
514 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
515 /* promptly. This is important if it were called from the signal */
516 /* handler or from the GC lock acquisition code. Under Linux, it's */
517 /* not safe to call it from a signal handler, since it uses mutexes */
518 /* and condition variables. Since it is called only here, the */
519 /* argument is unnecessary. */
520 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
521 /* resynchronize if we get far off, e.g. because GC_mark_no */
523 my_mark_no = GC_mark_no;
525 # ifdef DEBUG_THREADS
526 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
528 GC_help_marker(my_mark_no);
532 extern long GC_markers; /* Number of mark threads we would */
533 /* like to have. Includes the */
534 /* initiating thread. */
536 pthread_t GC_mark_threads[MAX_MARKERS];
538 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
540 static void start_mark_threads()
545 if (GC_markers > MAX_MARKERS) {
546 WARN("Limiting number of mark threads\n", 0);
547 GC_markers = MAX_MARKERS;
549 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
551 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
552 ABORT("pthread_attr_setdetachstate failed");
554 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
555 /* Default stack size is usually too small: fix it. */
556 /* Otherwise marker threads or GC may run out of */
558 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
563 if (pthread_attr_getstacksize(&attr, &old_size) != 0)
564 ABORT("pthread_attr_getstacksize failed\n");
565 if (old_size < MIN_STACK_SIZE) {
566 if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
567 ABORT("pthread_attr_setstacksize failed\n");
570 # endif /* HPUX || GC_DGUX386_THREADS */
572 if (GC_print_stats) {
573 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
576 for (i = 0; i < GC_markers - 1; ++i) {
577 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
578 GC_mark_thread, (void *)(word)i)) {
579 WARN("Marker thread creation failed, errno = %ld.\n", errno);
584 #else /* !PARALLEL_MARK */
586 static __inline__ void start_mark_threads()
590 #endif /* !PARALLEL_MARK */
592 GC_bool GC_thr_initialized = FALSE;
594 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
597 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
598 * put it into a separate function.
600 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
601 static __attribute__((noinline)) unsigned char* get_gc_thread_key_addr GC_PROTO((void))
603 return (unsigned char*)&GC_thread_key;
606 void GC_push_thread_structures GC_PROTO((void))
608 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
609 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
610 GC_push_all((ptr_t)get_gc_thread_key_addr(),
611 (ptr_t)(get_gc_thread_key_addr())+sizeof(&GC_thread_key));
617 void GC_push_thread_structures GC_PROTO((void))
619 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
620 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
621 GC_push_all((ptr_t)(&GC_thread_key),
622 (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
628 #ifdef THREAD_LOCAL_ALLOC
629 /* We must explicitly mark ptrfree and gcj free lists, since the free */
630 /* list links wouldn't otherwise be found. We also set them in the */
631 /* normal free lists, since that involves touching less memory than if */
632 /* we scanned them normally. */
633 void GC_mark_thread_local_free_lists(void)
639 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
640 for (p = GC_threads[i]; 0 != p; p = p -> next) {
641 for (j = 1; j < NFREELISTS; ++j) {
642 q = p -> ptrfree_freelists[j];
643 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
644 q = p -> normal_freelists[j];
645 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
646 # ifdef GC_GCJ_SUPPORT
647 q = p -> gcj_freelists[j];
648 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
649 # endif /* GC_GCJ_SUPPORT */
654 #endif /* THREAD_LOCAL_ALLOC */
656 static struct GC_Thread_Rep first_thread;
658 /* Add a thread to GC_threads. We assume it wasn't already there. */
659 /* Caller holds allocation lock. */
660 GC_thread GC_new_thread(pthread_t id)
662 int hv = ((word)id) % THREAD_TABLE_SZ;
664 static GC_bool first_thread_used = FALSE;
666 if (!first_thread_used) {
667 result = &first_thread;
668 first_thread_used = TRUE;
670 result = calloc (1, sizeof (struct GC_Thread_Rep));
672 if (result == 0) return(0);
674 result -> next = GC_threads[hv];
675 GC_threads[hv] = result;
676 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
680 /* Delete a thread from GC_threads. We assume it is there. */
681 /* (The code intentionally traps if it wasn't.) */
682 /* Caller holds allocation lock. */
683 void GC_delete_thread(pthread_t id)
685 int hv = ((word)id) % THREAD_TABLE_SZ;
686 register GC_thread p = GC_threads[hv];
687 register GC_thread prev = 0;
689 while (!pthread_equal(p -> id, id)) {
694 GC_threads[hv] = p -> next;
696 prev -> next = p -> next;
698 #ifdef MONO_DEBUGGER_SUPPORTED
699 if (gc_thread_vtable && gc_thread_vtable->thread_exited)
700 gc_thread_vtable->thread_exited (id, &p->stop_info.stack_ptr);
705 /* If a thread has been joined, but we have not yet */
706 /* been notified, then there may be more than one thread */
707 /* in the table with the same pthread id. */
708 /* This is OK, but we need a way to delete a specific one. */
709 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
711 int hv = ((word)id) % THREAD_TABLE_SZ;
712 register GC_thread p = GC_threads[hv];
713 register GC_thread prev = 0;
720 GC_threads[hv] = p -> next;
722 prev -> next = p -> next;
727 /* Return a GC_thread corresponding to a given pthread_t. */
728 /* Returns 0 if it's not there. */
729 /* Caller holds allocation lock or otherwise inhibits */
731 /* If there is more than one thread with the given id we */
732 /* return the most recent one. */
733 GC_thread GC_lookup_thread(pthread_t id)
735 int hv = ((word)id) % THREAD_TABLE_SZ;
736 register GC_thread p = GC_threads[hv];
738 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
742 int GC_thread_is_registered (void)
747 ptr = (void *)GC_lookup_thread(pthread_self());
754 /* Remove all entries from the GC_threads table, except the */
755 /* one for the current thread. We need to do this in the child */
756 /* process after a fork(), since only the current thread */
757 /* survives in the child. */
758 void GC_remove_all_threads_but_me(void)
760 pthread_t self = pthread_self();
762 GC_thread p, next, me;
764 for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
766 for (p = GC_threads[hv]; 0 != p; p = next) {
768 if (p -> id == self) {
772 # ifdef THREAD_LOCAL_ALLOC
773 if (!(p -> flags & FINISHED)) {
774 GC_destroy_thread_local(p);
776 # endif /* THREAD_LOCAL_ALLOC */
777 if (p != &first_thread) free(p);
783 #endif /* HANDLE_FORK */
785 #ifdef USE_PROC_FOR_LIBRARIES
786 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
791 # ifdef PARALLEL_MARK
792 for (i = 0; i < GC_markers; ++i) {
793 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
796 for (i = 0; i < THREAD_TABLE_SZ; i++) {
797 for (p = GC_threads[i]; p != 0; p = p -> next) {
798 if (0 != p -> stack_end) {
799 # ifdef STACK_GROWS_UP
800 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
801 # else /* STACK_GROWS_DOWN */
802 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
809 #endif /* USE_PROC_FOR_LIBRARIES */
811 #ifdef GC_LINUX_THREADS
812 /* Return the number of processors, or i<= 0 if it can't be determined. */
815 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
816 /* appears to be buggy in many cases. */
817 /* We look for lines "cpu<n>" in /proc/stat. */
818 # define STAT_BUF_SIZE 4096
819 # define STAT_READ read
820 /* If read is wrapped, this may need to be redefined to call */
822 char stat_buf[STAT_BUF_SIZE];
825 /* Some old kernels only have a single "cpu nnnn ..." */
826 /* entry in /proc/stat. We identify those as */
830 f = open("/proc/stat", O_RDONLY);
831 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
832 WARN("Couldn't read /proc/stat\n", 0);
835 for (i = 0; i < len - 100; ++i) {
836 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
837 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
838 int cpu_no = atoi(stat_buf + i + 4);
839 if (cpu_no >= result) result = cpu_no + 1;
845 #endif /* GC_LINUX_THREADS */
847 /* We hold the GC lock. Wait until an in-progress GC has finished. */
848 /* Repeatedly RELEASES GC LOCK in order to wait. */
849 /* If wait_for_all is true, then we exit with the GC lock held and no */
850 /* collection in progress; otherwise we just wait for the current GC */
852 extern GC_bool GC_collection_in_progress();
853 void GC_wait_for_gc_completion(GC_bool wait_for_all)
855 if (GC_incremental && GC_collection_in_progress()) {
856 int old_gc_no = GC_gc_no;
858 /* Make sure that no part of our stack is still on the mark stack, */
859 /* since it's about to be unmapped. */
860 while (GC_incremental && GC_collection_in_progress()
861 && (wait_for_all || old_gc_no == GC_gc_no)) {
863 GC_in_thread_creation = TRUE;
864 GC_collect_a_little_inner(1);
865 GC_in_thread_creation = FALSE;
875 /* Procedures called before and after a fork. The goal here is to make */
876 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
877 /* attainable, since the single UNIX spec seems to imply that one */
878 /* should only call async-signal-safe functions, and we probably can't */
879 /* quite guarantee that. But we give it our best shot. (That same */
880 /* spec also implies that it's not safe to call the system malloc */
881 /* between fork() and exec(). Thus we're doing no worse than it. */
883 /* Called before a fork() */
884 void GC_fork_prepare_proc(void)
886 /* Acquire all relevant locks, so that after releasing the locks */
887 /* the child will see a consistent state in which monitor */
888 /* invariants hold. Unfortunately, we can't acquire libc locks */
889 /* we might need, and there seems to be no guarantee that libc */
890 /* must install a suitable fork handler. */
891 /* Wait for an ongoing GC to finish, since we can't finish it in */
892 /* the (one remaining thread in) the child. */
894 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
895 GC_wait_for_reclaim();
897 GC_wait_for_gc_completion(TRUE);
898 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
899 GC_acquire_mark_lock();
903 /* Called in parent after a fork() */
904 void GC_fork_parent_proc(void)
906 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
907 GC_release_mark_lock();
912 /* Called in child after a fork() */
913 void GC_fork_child_proc(void)
915 /* Clean up the thread table, so that just our thread is left. */
916 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
917 GC_release_mark_lock();
919 GC_remove_all_threads_but_me();
920 # ifdef PARALLEL_MARK
921 /* Turn off parallel marking in the child, since we are probably */
922 /* just going to exec, and we would have to restart mark threads. */
925 # endif /* PARALLEL_MARK */
928 #endif /* HANDLE_FORK */
930 #if defined(GC_DGUX386_THREADS)
931 /* Return the number of processors, or i<= 0 if it can't be determined. */
934 /* <takis@XFree86.Org> */
936 struct dg_sys_info_pm_info pm_sysinfo;
939 status = dg_sys_info((long int *) &pm_sysinfo,
940 DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
942 /* set -1 for error */
946 numCpus = pm_sysinfo.idle_vp_count;
948 # ifdef DEBUG_THREADS
949 GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
953 #endif /* GC_DGUX386_THREADS */
955 /* We hold the allocation lock. */
958 # ifndef GC_DARWIN_THREADS
963 if (GC_thr_initialized) return;
964 GC_thr_initialized = TRUE;
967 /* Prepare for a possible fork. */
968 pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
970 # endif /* HANDLE_FORK */
971 /* Add the initial thread, so we can stop it. */
972 t = GC_new_thread(pthread_self());
973 # ifdef GC_DARWIN_THREADS
974 t -> stop_info.mach_thread = mach_thread_self();
976 t -> stop_info.stack_ptr = (ptr_t)(&dummy);
978 t -> flags = DETACHED | MAIN_THREAD;
979 #ifdef MONO_DEBUGGER_SUPPORTED
980 if (gc_thread_vtable && gc_thread_vtable->thread_created)
981 gc_thread_vtable->thread_created (pthread_self (), &t->stop_info.stack_ptr);
988 char * nprocs_string = GETENV("GC_NPROCS");
990 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
992 if (GC_nprocs <= 0) {
993 # if defined(GC_HPUX_THREADS)
994 GC_nprocs = pthread_num_processors_np();
996 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
997 GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
998 if (GC_nprocs <= 0) GC_nprocs = 1;
1000 # if defined(GC_IRIX_THREADS)
1001 GC_nprocs = sysconf(_SC_NPROC_ONLN);
1002 if (GC_nprocs <= 0) GC_nprocs = 1;
1004 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
1006 size_t len = sizeof(ncpus);
1007 sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
1010 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1011 GC_nprocs = GC_get_nprocs();
1014 if (GC_nprocs <= 0) {
1015 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
1017 # ifdef PARALLEL_MARK
1021 # ifdef PARALLEL_MARK
1023 char * markers_string = GETENV("GC_MARKERS");
1024 if (markers_string != NULL) {
1025 GC_markers = atoi(markers_string);
1027 GC_markers = GC_nprocs;
1032 # ifdef PARALLEL_MARK
1034 if (GC_print_stats) {
1035 GC_printf2("Number of processors = %ld, "
1036 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
1039 if (GC_markers == 1) {
1040 GC_parallel = FALSE;
1042 if (GC_print_stats) {
1043 GC_printf0("Single marker thread, turning off parallel marking\n");
1048 /* Disable true incremental collection, but generational is OK. */
1049 GC_time_limit = GC_TIME_UNLIMITED;
1051 /* If we are using a parallel marker, actually start helper threads. */
1052 if (GC_parallel) start_mark_threads();
1057 /* Perform all initializations, including those that */
1058 /* may require allocation. */
1059 /* Called without allocation lock. */
1060 /* Must be called before a second thread is created. */
1061 /* Called without allocation lock. */
1062 void GC_init_parallel()
1064 if (parallel_initialized) return;
1065 parallel_initialized = TRUE;
1067 /* GC_init() calls us back, so set flag first. */
1068 if (!GC_is_initialized) GC_init();
1069 /* Initialize thread local free lists if used. */
1070 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1072 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1078 #if !defined(GC_DARWIN_THREADS)
1079 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
1081 sigset_t fudged_set;
1083 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
1085 sigdelset(&fudged_set, SIG_SUSPEND);
1088 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
1090 #endif /* !GC_DARWIN_THREADS */
1092 /* Wrappers for functions that are likely to block for an appreciable */
1093 /* length of time. Must be called in pairs, if at all. */
1094 /* Nothing much beyond the system call itself should be executed */
1095 /* between these. */
1097 void GC_start_blocking(void) {
1098 # define SP_SLOP 128
1101 me = GC_lookup_thread(pthread_self());
1102 GC_ASSERT(!(me -> thread_blocked));
1104 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
1106 # ifndef GC_DARWIN_THREADS
1107 me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
1111 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1113 /* Add some slop to the stack pointer, since the wrapped call may */
1114 /* end up pushing more callee-save registers. */
1115 # ifndef GC_DARWIN_THREADS
1116 # ifdef STACK_GROWS_UP
1117 me -> stop_info.stack_ptr += SP_SLOP;
1119 me -> stop_info.stack_ptr -= SP_SLOP;
1122 me -> thread_blocked = TRUE;
1126 void GC_end_blocking(void) {
1128 LOCK(); /* This will block if the world is stopped. */
1129 me = GC_lookup_thread(pthread_self());
1130 GC_ASSERT(me -> thread_blocked);
1131 me -> thread_blocked = FALSE;
1135 #if defined(GC_DGUX386_THREADS)
1136 #define __d10_sleep sleep
1137 #endif /* GC_DGUX386_THREADS */
1139 /* A wrapper for the standard C sleep function */
1140 int WRAP_FUNC(sleep) (unsigned int seconds)
1144 GC_start_blocking();
1145 result = REAL_FUNC(sleep)(seconds);
1151 void *(*start_routine)(void *);
1154 sem_t registered; /* 1 ==> in our thread table, but */
1155 /* parent hasn't yet noticed. */
1158 /* Called at thread exit. */
1159 /* Never called for main thread. That's OK, since it */
1160 /* results in at most a tiny one-time leak. And */
1161 /* linuxthreads doesn't reclaim the main threads */
1162 /* resources or id anyway. */
1163 void GC_thread_exit_proc(void *arg)
1168 me = GC_lookup_thread(pthread_self());
1169 GC_destroy_thread_local(me);
1170 if (me -> flags & DETACHED) {
1171 GC_delete_thread(pthread_self());
1173 me -> flags |= FINISHED;
1175 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1176 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1177 GC_remove_specific(GC_thread_key);
1179 /* The following may run the GC from "nonexistent" thread. */
1180 GC_wait_for_gc_completion(FALSE);
1184 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1187 GC_thread thread_gc_id;
1190 thread_gc_id = GC_lookup_thread(thread);
1191 /* This is guaranteed to be the intended one, since the thread id */
1192 /* cant have been recycled by pthreads. */
1194 result = REAL_FUNC(pthread_join)(thread, retval);
1195 # if defined (GC_FREEBSD_THREADS)
1196 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1197 appears to be) a spurious EINTR which caused the test and real code
1198 to gratuitously fail. Having looked at system pthread library source
1199 code, I see how this return code may be generated. In one path of
1200 code, pthread_join() just returns the errno setting of the thread
1201 being joined. This does not match the POSIX specification or the
1202 local man pages thus I have taken the liberty to catch this one
1203 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1204 if (result == EINTR) result = 0;
1208 /* Here the pthread thread id may have been recycled. */
1209 GC_delete_gc_thread(thread, thread_gc_id);
1216 WRAP_FUNC(pthread_detach)(pthread_t thread)
1219 GC_thread thread_gc_id;
1222 thread_gc_id = GC_lookup_thread(thread);
1224 result = REAL_FUNC(pthread_detach)(thread);
1227 thread_gc_id -> flags |= DETACHED;
1228 /* Here the pthread thread id may have been recycled. */
1229 if (thread_gc_id -> flags & FINISHED) {
1230 GC_delete_gc_thread(thread, thread_gc_id);
1237 GC_bool GC_in_thread_creation = FALSE;
1239 typedef void *(*ThreadStartFn)(void *);
1240 void * GC_start_routine_head(void * arg, void *base_addr,
1241 ThreadStartFn *start, void **start_arg )
1243 struct start_info * si = arg;
1246 pthread_t my_pthread;
1248 my_pthread = pthread_self();
1249 # ifdef DEBUG_THREADS
1250 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1251 GC_printf1("pid = %ld\n", (long) getpid());
1252 GC_printf1("sp = 0x%lx\n", (long) &arg);
1255 GC_in_thread_creation = TRUE;
1256 me = GC_new_thread(my_pthread);
1257 GC_in_thread_creation = FALSE;
1258 #ifdef GC_DARWIN_THREADS
1259 me -> stop_info.mach_thread = mach_thread_self();
1261 me -> stop_info.stack_ptr = 0;
1263 me -> flags = si -> flags;
1264 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1265 /* doesn't work because the stack base in /proc/self/stat is the */
1266 /* one for the main thread. There is a strong argument that that's */
1267 /* a kernel bug, but a pervasive one. */
1268 # ifdef STACK_GROWS_DOWN
1269 me -> stack_end = (ptr_t)(((word)(base_addr) + (GC_page_size - 1))
1270 & ~(GC_page_size - 1));
1271 # ifndef GC_DARWIN_THREADS
1272 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1274 /* Needs to be plausible, since an asynchronous stack mark */
1275 /* should not crash. */
1277 me -> stack_end = (ptr_t)((word)(base_addr) & ~(GC_page_size - 1));
1278 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1280 /* This is dubious, since we may be more than a page into the stack, */
1281 /* and hence skip some of it, though it's not clear that matters. */
1283 me -> backing_store_end = (ptr_t)
1284 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1285 /* This is also < 100% convincing. We should also read this */
1286 /* from /proc, but the hook to do so isn't there yet. */
1288 #ifdef MONO_DEBUGGER_SUPPORTED
1289 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1290 gc_thread_vtable->thread_created (my_pthread, &me->stop_info.stack_ptr);
1294 if (start) *start = si -> start_routine;
1295 if (start_arg) *start_arg = si -> arg;
1297 sem_post(&(si -> registered)); /* Last action on si. */
1298 /* OK to deallocate. */
1299 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1301 GC_init_thread_local(me);
1308 int GC_thread_register_foreign (void *base_addr)
1310 struct start_info si = { 0, }; /* stacked for legibility & locking */
1313 # ifdef DEBUG_THREADS
1314 GC_printf1( "GC_thread_register_foreign %p\n", &si );
1317 si.flags = FOREIGN_THREAD;
1319 if (!parallel_initialized) GC_init_parallel();
1321 if (!GC_thr_initialized) GC_thr_init();
1325 me = GC_start_routine_head(&si, base_addr, NULL, NULL);
1330 void * GC_start_routine(void * arg)
1333 struct start_info * si = arg;
1336 ThreadStartFn start;
1339 me = GC_start_routine_head (arg, &dummy, &start, &start_arg);
1341 pthread_cleanup_push(GC_thread_exit_proc, 0);
1342 # ifdef DEBUG_THREADS
1343 GC_printf1("start_routine = 0x%lx\n", start);
1345 result = (*start)(start_arg);
1347 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1349 me -> status = result;
1350 pthread_cleanup_pop(1);
1351 /* Cleanup acquires lock, ensuring that we can't exit */
1352 /* while a collection that thinks we're alive is trying to stop */
1358 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1359 const pthread_attr_t *attr,
1360 void *(*start_routine)(void *), void *arg)
1365 struct start_info * si;
1366 /* This is otherwise saved only in an area mmapped by the thread */
1367 /* library, which isn't visible to the collector. */
1369 /* We resist the temptation to muck with the stack size here, */
1370 /* even if the default is unreasonably small. That's the client's */
1371 /* responsibility. */
1374 si = calloc (1, sizeof (struct start_info));
1376 if (!parallel_initialized) GC_init_parallel();
1377 if (0 == si) return(ENOMEM);
1378 sem_init(&(si -> registered), 0, 0);
1379 si -> start_routine = start_routine;
1382 if (!GC_thr_initialized) GC_thr_init();
1383 # ifdef GC_ASSERTIONS
1387 pthread_attr_t my_attr;
1388 pthread_attr_init(&my_attr);
1389 pthread_attr_getstacksize(&my_attr, &stack_size);
1391 pthread_attr_getstacksize(attr, &stack_size);
1393 # ifdef PARALLEL_MARK
1394 GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1396 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1397 /* HBLKSIZE=8192, sizeof(word)=8 */
1398 GC_ASSERT(stack_size >= 65536);
1400 /* Our threads may need to do some work for the GC. */
1401 /* Ridiculously small threads won't work, and they */
1402 /* probably wouldn't work anyway. */
1406 detachstate = PTHREAD_CREATE_JOINABLE;
1408 pthread_attr_getdetachstate(attr, &detachstate);
1410 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1411 si -> flags = my_flags;
1413 # ifdef DEBUG_THREADS
1414 GC_printf1("About to start new thread from thread 0x%X\n",
1418 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1420 # ifdef DEBUG_THREADS
1421 GC_printf1("Started thread 0x%X\n", *new_thread);
1423 /* Wait until child has been added to the thread table. */
1424 /* This also ensures that we hold onto si until the child is done */
1425 /* with it. Thus it doesn't matter whether it is otherwise */
1426 /* visible to the collector. */
1428 while (0 != sem_wait(&(si -> registered))) {
1429 if (EINTR != errno) ABORT("sem_wait failed");
1432 sem_destroy(&(si -> registered));
1440 #ifdef GENERIC_COMPARE_AND_SWAP
1441 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1443 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1444 GC_word old, GC_word new_val)
1447 pthread_mutex_lock(&GC_compare_and_swap_lock);
1454 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1458 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1461 pthread_mutex_lock(&GC_compare_and_swap_lock);
1463 *addr = old + how_much;
1464 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1468 #endif /* GENERIC_COMPARE_AND_SWAP */
1469 /* Spend a few cycles in a way that can't introduce contention with */
1470 /* othre threads. */
1474 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1475 volatile word dummy = 0;
1478 for (i = 0; i < 10; ++i) {
1479 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1480 __asm__ __volatile__ (" " : : : "memory");
1482 /* Something that's unlikely to be optimized away. */
1488 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1491 VOLATILE GC_bool GC_collecting = 0;
1492 /* A hint that we're in the collector and */
1493 /* holding the allocation lock for an */
1494 /* extended period. */
1496 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1497 /* If we don't want to use the below spinlock implementation, either */
1498 /* because we don't have a GC_test_and_set implementation, or because */
1499 /* we don't want to risk sleeping, we can still try spinning on */
1500 /* pthread_mutex_trylock for a while. This appears to be very */
1501 /* beneficial in many cases. */
1502 /* I suspect that under high contention this is nearly always better */
1503 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1504 /* Hence we still default to the spin lock. */
1505 /* This is also used to acquire the mark lock for the parallel */
1508 /* Here we use a strict exponential backoff scheme. I don't know */
1509 /* whether that's better or worse than the above. We eventually */
1510 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1511 /* explicitly sleep. */
1515 unsigned long GC_spin_count = 0;
1516 unsigned long GC_block_count = 0;
1517 unsigned long GC_unlocked_count = 0;
1520 void GC_generic_lock(pthread_mutex_t * lock)
1522 #ifndef NO_PTHREAD_TRYLOCK
1523 unsigned pause_length = 1;
1526 if (0 == pthread_mutex_trylock(lock)) {
1528 ++GC_unlocked_count;
1532 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1533 for (i = 0; i < pause_length; ++i) {
1536 switch(pthread_mutex_trylock(lock)) {
1545 ABORT("Unexpected error from pthread_mutex_trylock");
1548 #endif /* !NO_PTHREAD_TRYLOCK */
1552 pthread_mutex_lock(lock);
1555 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1557 #if defined(USE_SPIN_LOCK)
1559 /* Reasonably fast spin locks. Basically the same implementation */
1560 /* as STL alloc.h. This isn't really the right way to do this. */
1561 /* but until the POSIX scheduling mess gets straightened out ... */
1563 volatile unsigned int GC_allocate_lock = 0;
1568 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1569 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1570 static unsigned spin_max = low_spin_max;
1571 unsigned my_spin_max;
1572 static unsigned last_spins = 0;
1573 unsigned my_last_spins;
1576 if (!GC_test_and_set(&GC_allocate_lock)) {
1579 my_spin_max = spin_max;
1580 my_last_spins = last_spins;
1581 for (i = 0; i < my_spin_max; i++) {
1582 if (GC_collecting || GC_nprocs == 1) goto yield;
1583 if (i < my_last_spins/2 || GC_allocate_lock) {
1587 if (!GC_test_and_set(&GC_allocate_lock)) {
1590 * Spinning worked. Thus we're probably not being scheduled
1591 * against the other process with which we were contending.
1592 * Thus it makes sense to spin longer the next time.
1595 spin_max = high_spin_max;
1599 /* We are probably being scheduled against the other process. Sleep. */
1600 spin_max = low_spin_max;
1603 if (!GC_test_and_set(&GC_allocate_lock)) {
1606 # define SLEEP_THRESHOLD 12
1607 /* Under Linux very short sleeps tend to wait until */
1608 /* the current time quantum expires. On old Linux */
1609 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1610 /* (Under 2.4, this happens only for real-time */
1611 /* processes.) We want to minimize both behaviors */
1613 if (i < SLEEP_THRESHOLD) {
1619 /* Don't wait for more than about 15msecs, even */
1620 /* under extreme contention. */
1622 ts.tv_nsec = 1 << i;
1628 #else /* !USE_SPINLOCK */
1631 #ifndef NO_PTHREAD_TRYLOCK
1632 if (1 == GC_nprocs || GC_collecting) {
1633 pthread_mutex_lock(&GC_allocate_ml);
1635 GC_generic_lock(&GC_allocate_ml);
1637 #else /* !NO_PTHREAD_TRYLOCK */
1638 pthread_mutex_lock(&GC_allocate_ml);
1639 #endif /* !NO_PTHREAD_TRYLOCK */
1642 #endif /* !USE_SPINLOCK */
1644 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1646 #ifdef GC_ASSERTIONS
1647 pthread_t GC_mark_lock_holder = NO_THREAD;
1651 /* Ugly workaround for a linux threads bug in the final versions */
1652 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1653 /* field even when it fails to acquire the mutex. This causes */
1654 /* pthread_cond_wait to die. Remove for glibc2.2. */
1655 /* According to the man page, we should use */
1656 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1658 static pthread_mutex_t mark_mutex =
1659 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1661 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1664 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1666 void GC_acquire_mark_lock()
1669 if (pthread_mutex_lock(&mark_mutex) != 0) {
1670 ABORT("pthread_mutex_lock failed");
1673 GC_generic_lock(&mark_mutex);
1674 # ifdef GC_ASSERTIONS
1675 GC_mark_lock_holder = pthread_self();
1679 void GC_release_mark_lock()
1681 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1682 # ifdef GC_ASSERTIONS
1683 GC_mark_lock_holder = NO_THREAD;
1685 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1686 ABORT("pthread_mutex_unlock failed");
1690 /* Collector must wait for a freelist builders for 2 reasons: */
1691 /* 1) Mark bits may still be getting examined without lock. */
1692 /* 2) Partial free lists referenced only by locals may not be scanned */
1693 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1694 /* free-list link may be ignored. */
1695 void GC_wait_builder()
1697 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1698 # ifdef GC_ASSERTIONS
1699 GC_mark_lock_holder = NO_THREAD;
1701 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1702 ABORT("pthread_cond_wait failed");
1704 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1705 # ifdef GC_ASSERTIONS
1706 GC_mark_lock_holder = pthread_self();
1710 void GC_wait_for_reclaim()
1712 GC_acquire_mark_lock();
1713 while (GC_fl_builder_count > 0) {
1716 GC_release_mark_lock();
1719 void GC_notify_all_builder()
1721 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1722 if (pthread_cond_broadcast(&builder_cv) != 0) {
1723 ABORT("pthread_cond_broadcast failed");
1727 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1729 #ifdef PARALLEL_MARK
1731 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1733 void GC_wait_marker()
1735 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1736 # ifdef GC_ASSERTIONS
1737 GC_mark_lock_holder = NO_THREAD;
1739 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1740 ABORT("pthread_cond_wait failed");
1742 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1743 # ifdef GC_ASSERTIONS
1744 GC_mark_lock_holder = pthread_self();
1748 void GC_notify_all_marker()
1750 if (pthread_cond_broadcast(&mark_cv) != 0) {
1751 ABORT("pthread_cond_broadcast failed");
1755 #endif /* PARALLEL_MARK */
1757 # endif /* GC_LINUX_THREADS and friends */