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) 2000-2001 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_IRIX_THREADS) && !defined(GC_WIN32_THREADS) \
54 && !defined(GC_AIX_THREADS)
56 # if defined(GC_HPUX_THREADS) && !defined(USE_PTHREAD_SPECIFIC) \
57 && !defined(USE_COMPILER_TLS)
59 # define USE_PTHREAD_SPECIFIC
60 /* Empirically, as of gcc 3.3, USE_COMPILER_TLS doesn't work. */
62 # define USE_COMPILER_TLS
66 # if defined USE_HPUX_TLS
67 --> Macro replaced by USE_COMPILER_TLS
70 # if (defined(GC_DGUX386_THREADS) || defined(GC_OSF1_THREADS) || \
71 defined(GC_DARWIN_THREADS)) && !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"
87 # if defined(USE_PTHREAD_SPECIFIC)
88 # define GC_getspecific pthread_getspecific
89 # define GC_setspecific pthread_setspecific
90 # define GC_key_create pthread_key_create
91 typedef pthread_key_t GC_key_t;
93 # if defined(USE_COMPILER_TLS)
94 # define GC_getspecific(x) (x)
95 # define GC_setspecific(key, v) ((key) = (v), 0)
96 # define GC_key_create(key, d) 0
97 typedef void * GC_key_t;
101 # include <pthread.h>
106 # include <sys/mman.h>
107 # include <sys/time.h>
108 # include <sys/types.h>
109 # include <sys/stat.h>
113 #if defined(GC_DARWIN_THREADS)
114 # include "private/darwin_semaphore.h"
116 # include <semaphore.h>
117 #endif /* !GC_DARWIN_THREADS */
119 #if defined(GC_DARWIN_THREADS)
120 # include <sys/sysctl.h>
121 #endif /* GC_DARWIN_THREADS */
125 #if defined(GC_DGUX386_THREADS)
126 # include <sys/dg_sys_info.h>
127 # include <sys/_int_psem.h>
128 /* sem_t is an uint in DG/UX */
129 typedef unsigned int sem_t;
130 #endif /* GC_DGUX386_THREADS */
136 #ifdef GC_USE_LD_WRAP
137 # define WRAP_FUNC(f) __wrap_##f
138 # define REAL_FUNC(f) __real_##f
140 # define WRAP_FUNC(f) GC_##f
141 # if !defined(GC_DGUX386_THREADS)
142 # define REAL_FUNC(f) f
143 # else /* GC_DGUX386_THREADS */
144 # define REAL_FUNC(f) __d10_##f
145 # endif /* GC_DGUX386_THREADS */
146 # undef pthread_create
147 # if !defined(GC_DARWIN_THREADS)
148 # undef pthread_sigmask
151 # undef pthread_detach
152 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
153 && !defined(_PTHREAD_USE_PTDNAM_)
154 /* Restore the original mangled names on Tru64 UNIX. */
155 # define pthread_create __pthread_create
156 # define pthread_join __pthread_join
157 # define pthread_detach __pthread_detach
163 static GC_bool parallel_initialized = FALSE;
165 void GC_init_parallel();
167 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
169 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
171 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
172 #undef PACKAGE_BUGREPORT
174 #undef PACKAGE_STRING
175 #undef PACKAGE_TARNAME
176 #undef PACKAGE_VERSION
177 #include "mono/utils/mono-compiler.h"
180 #ifdef USE_COMPILER_TLS
181 __thread MONO_TLS_FAST
183 GC_key_t GC_thread_key;
185 static GC_bool keys_initialized;
187 /* Recover the contents of the freelist array fl into the global one gfl.*/
188 /* Note that the indexing scheme differs, in that gfl has finer size */
189 /* resolution, even if not all entries are used. */
190 /* We hold the allocator lock. */
191 static void return_freelists(ptr_t *fl, ptr_t *gfl)
197 for (i = 1; i < NFREELISTS; ++i) {
198 nwords = i * (GRANULARITY/sizeof(word));
201 if ((word)q >= HBLKSIZE) {
202 if (gfl[nwords] == 0) {
206 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
212 /* Clear fl[i], since the thread structure may hang around. */
213 /* Do it in a way that is likely to trap if we access it. */
214 fl[i] = (ptr_t)HBLKSIZE;
218 /* We statically allocate a single "size 0" object. It is linked to */
219 /* itself, and is thus repeatedly reused for all size 0 allocation */
220 /* requests. (Size 0 gcj allocation requests are incorrect, and */
221 /* we arrange for those to fault asap.) */
222 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
224 /* Each thread structure must be initialized. */
225 /* This call must be made from the new thread. */
226 /* Caller holds allocation lock. */
227 void GC_init_thread_local(GC_thread p)
231 if (!keys_initialized) {
232 if (0 != GC_key_create(&GC_thread_key, 0)) {
233 ABORT("Failed to create key for local allocator");
235 keys_initialized = TRUE;
237 if (0 != GC_setspecific(GC_thread_key, p)) {
238 ABORT("Failed to set thread specific allocation pointers");
240 for (i = 1; i < NFREELISTS; ++i) {
241 p -> ptrfree_freelists[i] = (ptr_t)1;
242 p -> normal_freelists[i] = (ptr_t)1;
243 # ifdef GC_GCJ_SUPPORT
244 p -> gcj_freelists[i] = (ptr_t)1;
247 /* Set up the size 0 free lists. */
248 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
249 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
250 # ifdef GC_GCJ_SUPPORT
251 p -> gcj_freelists[0] = (ptr_t)(-1);
255 #ifdef GC_GCJ_SUPPORT
256 extern ptr_t * GC_gcjobjfreelist;
259 /* We hold the allocator lock. */
260 void GC_destroy_thread_local(GC_thread p)
262 /* We currently only do this from the thread itself or from */
263 /* the fork handler for a child process. */
265 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
267 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
268 return_freelists(p -> normal_freelists, GC_objfreelist);
269 # ifdef GC_GCJ_SUPPORT
270 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
274 extern GC_PTR GC_generic_malloc_many();
276 GC_PTR GC_local_malloc(size_t bytes)
278 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
279 return(GC_malloc(bytes));
281 int index = INDEX_FROM_BYTES(bytes);
284 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
285 GC_key_t k = GC_thread_key;
289 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
290 if (EXPECT(0 == k, 0)) {
291 /* This can happen if we get called when the world is */
292 /* being initialized. Whether we can actually complete */
293 /* the initialization then is unclear. */
298 tsd = GC_getspecific(GC_thread_key);
299 # ifdef GC_ASSERTIONS
301 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
304 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
306 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
307 ptr_t next = obj_link(my_entry);
308 GC_PTR result = (GC_PTR)my_entry;
310 obj_link(my_entry) = 0;
311 PREFETCH_FOR_WRITE(next);
313 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
314 *my_fl = my_entry + index + 1;
315 return GC_malloc(bytes);
317 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
318 if (*my_fl == 0) return GC_oom_fn(bytes);
319 return GC_local_malloc(bytes);
324 GC_PTR GC_local_malloc_atomic(size_t bytes)
326 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
327 return(GC_malloc_atomic(bytes));
329 int index = INDEX_FROM_BYTES(bytes);
330 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
331 -> ptrfree_freelists + index;
332 ptr_t my_entry = *my_fl;
334 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
335 GC_PTR result = (GC_PTR)my_entry;
336 *my_fl = obj_link(my_entry);
338 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
339 *my_fl = my_entry + index + 1;
340 return GC_malloc_atomic(bytes);
342 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
343 /* *my_fl is updated while the collector is excluded; */
344 /* the free list is always visible to the collector as */
346 if (*my_fl == 0) return GC_oom_fn(bytes);
347 return GC_local_malloc_atomic(bytes);
352 #ifdef GC_GCJ_SUPPORT
354 #include "include/gc_gcj.h"
357 extern GC_bool GC_gcj_malloc_initialized;
360 extern int GC_gcj_kind;
362 GC_PTR GC_local_gcj_malloc(size_t bytes,
363 void * ptr_to_struct_containing_descr)
365 GC_ASSERT(GC_gcj_malloc_initialized);
366 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
367 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
369 int index = INDEX_FROM_BYTES(bytes);
370 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
371 -> gcj_freelists + index;
372 ptr_t my_entry = *my_fl;
373 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
374 GC_PTR result = (GC_PTR)my_entry;
375 GC_ASSERT(!GC_incremental);
376 /* We assert that any concurrent marker will stop us. */
377 /* Thus it is impossible for a mark procedure to see the */
378 /* allocation of the next object, but to see this object */
379 /* still containing a free list pointer. Otherwise the */
380 /* marker might find a random "mark descriptor". */
381 *(volatile ptr_t *)my_fl = obj_link(my_entry);
382 /* We must update the freelist before we store the pointer. */
383 /* Otherwise a GC at this point would see a corrupted */
385 /* A memory barrier is probably never needed, since the */
386 /* action of stopping this thread will cause prior writes */
388 GC_ASSERT(((void * volatile *)result)[1] == 0);
389 *(void * volatile *)result = ptr_to_struct_containing_descr;
391 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
392 if (!GC_incremental) *my_fl = my_entry + index + 1;
393 /* In the incremental case, we always have to take this */
394 /* path. Thus we leave the counter alone. */
395 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
397 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
398 if (*my_fl == 0) return GC_oom_fn(bytes);
399 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
404 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
405 /* adjust it. The size is assumed to be such that it can be */
406 /* allocated as a small object. */
407 void * GC_local_gcj_fast_malloc(size_t lw, void * ptr_to_struct_containing_descr)
409 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
410 -> gcj_freelists + lw;
411 ptr_t my_entry = *my_fl;
413 GC_ASSERT(GC_gcj_malloc_initialized);
415 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
416 GC_PTR result = (GC_PTR)my_entry;
417 GC_ASSERT(!GC_incremental);
418 /* We assert that any concurrent marker will stop us. */
419 /* Thus it is impossible for a mark procedure to see the */
420 /* allocation of the next object, but to see this object */
421 /* still containing a free list pointer. Otherwise the */
422 /* marker might find a random "mark descriptor". */
423 *(volatile ptr_t *)my_fl = obj_link(my_entry);
424 /* We must update the freelist before we store the pointer. */
425 /* Otherwise a GC at this point would see a corrupted */
427 /* A memory barrier is probably never needed, since the */
428 /* action of stopping this thread will cause prior writes */
430 GC_ASSERT(((void * volatile *)result)[1] == 0);
431 *(void * volatile *)result = ptr_to_struct_containing_descr;
433 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
434 if (!GC_incremental) *my_fl = my_entry + lw + 1;
435 /* In the incremental case, we always have to take this */
436 /* path. Thus we leave the counter alone. */
437 return GC_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
439 GC_generic_malloc_many(BYTES_FROM_INDEX(lw), GC_gcj_kind, my_fl);
440 if (*my_fl == 0) return GC_oom_fn(BYTES_FROM_INDEX(lw));
441 return GC_local_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
445 #endif /* GC_GCJ_SUPPORT */
447 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
449 # define GC_destroy_thread_local(t)
451 # endif /* !THREAD_LOCAL_ALLOC */
455 To make sure that we're using LinuxThreads and not some other thread
456 package, we generate a dummy reference to `pthread_kill_other_threads_np'
457 (was `__pthread_initial_thread_bos' but that disappeared),
458 which is a symbol defined in LinuxThreads, but (hopefully) not in other
461 We no longer do this, since this code is now portable enough that it might
462 actually work for something else.
464 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
467 long GC_nprocs = 1; /* Number of processors. We may not have */
468 /* access to all of them, but this is as good */
469 /* a guess as any ... */
474 # define MAX_MARKERS 16
477 static ptr_t marker_sp[MAX_MARKERS] = {0};
479 void * GC_mark_thread(void * id)
483 marker_sp[(word)id] = GC_approx_sp();
484 for (;; ++my_mark_no) {
485 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
486 /* promptly. This is important if it were called from the signal */
487 /* handler or from the GC lock acquisition code. Under Linux, it's */
488 /* not safe to call it from a signal handler, since it uses mutexes */
489 /* and condition variables. Since it is called only here, the */
490 /* argument is unnecessary. */
491 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
492 /* resynchronize if we get far off, e.g. because GC_mark_no */
494 my_mark_no = GC_mark_no;
496 # ifdef DEBUG_THREADS
497 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
499 GC_help_marker(my_mark_no);
503 extern long GC_markers; /* Number of mark threads we would */
504 /* like to have. Includes the */
505 /* initiating thread. */
507 pthread_t GC_mark_threads[MAX_MARKERS];
509 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
511 static void start_mark_threads()
516 if (GC_markers > MAX_MARKERS) {
517 WARN("Limiting number of mark threads\n", 0);
518 GC_markers = MAX_MARKERS;
520 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
522 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
523 ABORT("pthread_attr_setdetachstate failed");
525 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
526 /* Default stack size is usually too small: fix it. */
527 /* Otherwise marker threads or GC may run out of */
529 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
534 if (pthread_attr_getstacksize(&attr, &old_size) != 0)
535 ABORT("pthread_attr_getstacksize failed\n");
536 if (old_size < MIN_STACK_SIZE) {
537 if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
538 ABORT("pthread_attr_setstacksize failed\n");
541 # endif /* HPUX || GC_DGUX386_THREADS */
543 if (GC_print_stats) {
544 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
547 for (i = 0; i < GC_markers - 1; ++i) {
548 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
549 GC_mark_thread, (void *)(word)i)) {
550 WARN("Marker thread creation failed, errno = %ld.\n", errno);
555 #else /* !PARALLEL_MARK */
557 static __inline__ void start_mark_threads()
561 #endif /* !PARALLEL_MARK */
563 GC_bool GC_thr_initialized = FALSE;
565 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
568 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
569 * put it into a separate function.
571 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
572 static __attribute__((noinline)) unsigned char* get_gc_thread_key_addr GC_PROTO((void))
574 return (unsigned char*)&GC_thread_key;
577 void GC_push_thread_structures GC_PROTO((void))
579 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
580 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
581 GC_push_all((ptr_t)get_gc_thread_key_addr(),
582 (ptr_t)(get_gc_thread_key_addr())+sizeof(&GC_thread_key));
588 void GC_push_thread_structures GC_PROTO((void))
590 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
591 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
592 GC_push_all((ptr_t)(&GC_thread_key),
593 (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
599 #ifdef THREAD_LOCAL_ALLOC
600 /* We must explicitly mark ptrfree and gcj free lists, since the free */
601 /* list links wouldn't otherwise be found. We also set them in the */
602 /* normal free lists, since that involves touching less memory than if */
603 /* we scanned them normally. */
604 void GC_mark_thread_local_free_lists(void)
610 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
611 for (p = GC_threads[i]; 0 != p; p = p -> next) {
612 for (j = 1; j < NFREELISTS; ++j) {
613 q = p -> ptrfree_freelists[j];
614 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
615 q = p -> normal_freelists[j];
616 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
617 # ifdef GC_GCJ_SUPPORT
618 q = p -> gcj_freelists[j];
619 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
620 # endif /* GC_GCJ_SUPPORT */
625 #endif /* THREAD_LOCAL_ALLOC */
627 static struct GC_Thread_Rep first_thread;
629 /* Add a thread to GC_threads. We assume it wasn't already there. */
630 /* Caller holds allocation lock. */
631 GC_thread GC_new_thread(pthread_t id)
633 int hv = ((word)id) % THREAD_TABLE_SZ;
635 static GC_bool first_thread_used = FALSE;
637 if (!first_thread_used) {
638 result = &first_thread;
639 first_thread_used = TRUE;
641 result = (struct GC_Thread_Rep *)
642 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
644 if (result == 0) return(0);
646 result -> next = GC_threads[hv];
647 GC_threads[hv] = result;
648 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
652 /* Delete a thread from GC_threads. We assume it is there. */
653 /* (The code intentionally traps if it wasn't.) */
654 /* Caller holds allocation lock. */
655 void GC_delete_thread(pthread_t id)
657 int hv = ((word)id) % THREAD_TABLE_SZ;
658 register GC_thread p = GC_threads[hv];
659 register GC_thread prev = 0;
661 while (!pthread_equal(p -> id, id)) {
666 GC_threads[hv] = p -> next;
668 prev -> next = p -> next;
673 /* If a thread has been joined, but we have not yet */
674 /* been notified, then there may be more than one thread */
675 /* in the table with the same pthread id. */
676 /* This is OK, but we need a way to delete a specific one. */
677 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
679 int hv = ((word)id) % THREAD_TABLE_SZ;
680 register GC_thread p = GC_threads[hv];
681 register GC_thread prev = 0;
688 GC_threads[hv] = p -> next;
690 prev -> next = p -> next;
695 /* Return a GC_thread corresponding to a given pthread_t. */
696 /* Returns 0 if it's not there. */
697 /* Caller holds allocation lock or otherwise inhibits */
699 /* If there is more than one thread with the given id we */
700 /* return the most recent one. */
701 GC_thread GC_lookup_thread(pthread_t id)
703 int hv = ((word)id) % THREAD_TABLE_SZ;
704 register GC_thread p = GC_threads[hv];
706 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
710 int GC_thread_is_registered (void)
715 ptr = (void *)GC_lookup_thread(pthread_self());
722 /* Remove all entries from the GC_threads table, except the */
723 /* one for the current thread. We need to do this in the child */
724 /* process after a fork(), since only the current thread */
725 /* survives in the child. */
726 void GC_remove_all_threads_but_me(void)
728 pthread_t self = pthread_self();
730 GC_thread p, next, me;
732 for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
734 for (p = GC_threads[hv]; 0 != p; p = next) {
736 if (p -> id == self) {
740 # ifdef THREAD_LOCAL_ALLOC
741 if (!(p -> flags & FINISHED)) {
742 GC_destroy_thread_local(p);
744 # endif /* THREAD_LOCAL_ALLOC */
745 if (p != &first_thread) GC_INTERNAL_FREE(p);
751 #endif /* HANDLE_FORK */
753 #ifdef USE_PROC_FOR_LIBRARIES
754 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
759 # ifdef PARALLEL_MARK
760 for (i = 0; i < GC_markers; ++i) {
761 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
764 for (i = 0; i < THREAD_TABLE_SZ; i++) {
765 for (p = GC_threads[i]; p != 0; p = p -> next) {
766 if (0 != p -> stack_end) {
767 # ifdef STACK_GROWS_UP
768 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
769 # else /* STACK_GROWS_DOWN */
770 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
777 #endif /* USE_PROC_FOR_LIBRARIES */
779 #ifdef GC_LINUX_THREADS
780 /* Return the number of processors, or i<= 0 if it can't be determined. */
783 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
784 /* appears to be buggy in many cases. */
785 /* We look for lines "cpu<n>" in /proc/stat. */
786 # define STAT_BUF_SIZE 4096
787 # define STAT_READ read
788 /* If read is wrapped, this may need to be redefined to call */
790 char stat_buf[STAT_BUF_SIZE];
793 /* Some old kernels only have a single "cpu nnnn ..." */
794 /* entry in /proc/stat. We identify those as */
798 f = open("/proc/stat", O_RDONLY);
799 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
800 WARN("Couldn't read /proc/stat\n", 0);
803 for (i = 0; i < len - 100; ++i) {
804 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
805 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
806 int cpu_no = atoi(stat_buf + i + 4);
807 if (cpu_no >= result) result = cpu_no + 1;
813 #endif /* GC_LINUX_THREADS */
815 /* We hold the GC lock. Wait until an in-progress GC has finished. */
816 /* Repeatedly RELEASES GC LOCK in order to wait. */
817 /* If wait_for_all is true, then we exit with the GC lock held and no */
818 /* collection in progress; otherwise we just wait for the current GC */
820 extern GC_bool GC_collection_in_progress();
821 void GC_wait_for_gc_completion(GC_bool wait_for_all)
823 if (GC_incremental && GC_collection_in_progress()) {
824 int old_gc_no = GC_gc_no;
826 /* Make sure that no part of our stack is still on the mark stack, */
827 /* since it's about to be unmapped. */
828 while (GC_incremental && GC_collection_in_progress()
829 && (wait_for_all || old_gc_no == GC_gc_no)) {
831 GC_in_thread_creation = TRUE;
832 GC_collect_a_little_inner(1);
833 GC_in_thread_creation = FALSE;
843 /* Procedures called before and after a fork. The goal here is to make */
844 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
845 /* attainable, since the single UNIX spec seems to imply that one */
846 /* should only call async-signal-safe functions, and we probably can't */
847 /* quite guarantee that. But we give it our best shot. (That same */
848 /* spec also implies that it's not safe to call the system malloc */
849 /* between fork() and exec(). Thus we're doing no worse than it. */
851 /* Called before a fork() */
852 void GC_fork_prepare_proc(void)
854 /* Acquire all relevant locks, so that after releasing the locks */
855 /* the child will see a consistent state in which monitor */
856 /* invariants hold. Unfortunately, we can't acquire libc locks */
857 /* we might need, and there seems to be no guarantee that libc */
858 /* must install a suitable fork handler. */
859 /* Wait for an ongoing GC to finish, since we can't finish it in */
860 /* the (one remaining thread in) the child. */
862 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
863 GC_wait_for_reclaim();
865 GC_wait_for_gc_completion(TRUE);
866 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
867 GC_acquire_mark_lock();
871 /* Called in parent after a fork() */
872 void GC_fork_parent_proc(void)
874 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
875 GC_release_mark_lock();
880 /* Called in child after a fork() */
881 void GC_fork_child_proc(void)
883 /* Clean up the thread table, so that just our thread is left. */
884 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
885 GC_release_mark_lock();
887 GC_remove_all_threads_but_me();
888 # ifdef PARALLEL_MARK
889 /* Turn off parallel marking in the child, since we are probably */
890 /* just going to exec, and we would have to restart mark threads. */
893 # endif /* PARALLEL_MARK */
896 #endif /* HANDLE_FORK */
898 #if defined(GC_DGUX386_THREADS)
899 /* Return the number of processors, or i<= 0 if it can't be determined. */
902 /* <takis@XFree86.Org> */
904 struct dg_sys_info_pm_info pm_sysinfo;
907 status = dg_sys_info((long int *) &pm_sysinfo,
908 DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
910 /* set -1 for error */
914 numCpus = pm_sysinfo.idle_vp_count;
916 # ifdef DEBUG_THREADS
917 GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
921 #endif /* GC_DGUX386_THREADS */
923 /* We hold the allocation lock. */
926 # ifndef GC_DARWIN_THREADS
931 if (GC_thr_initialized) return;
932 GC_thr_initialized = TRUE;
935 /* Prepare for a possible fork. */
936 pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
938 # endif /* HANDLE_FORK */
939 /* Add the initial thread, so we can stop it. */
940 t = GC_new_thread(pthread_self());
941 # ifdef GC_DARWIN_THREADS
942 t -> stop_info.mach_thread = mach_thread_self();
944 t -> stop_info.stack_ptr = (ptr_t)(&dummy);
946 t -> flags = DETACHED | MAIN_THREAD;
952 char * nprocs_string = GETENV("GC_NPROCS");
954 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
956 if (GC_nprocs <= 0) {
957 # if defined(GC_HPUX_THREADS)
958 GC_nprocs = pthread_num_processors_np();
960 # if defined(GC_OSF1_THREADS)
961 GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
962 if (GC_nprocs <= 0) GC_nprocs = 1;
964 # if defined(GC_FREEBSD_THREADS)
967 # if defined(GC_DARWIN_THREADS)
969 size_t len = sizeof(ncpus);
970 sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
973 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
974 GC_nprocs = GC_get_nprocs();
977 if (GC_nprocs <= 0) {
978 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
980 # ifdef PARALLEL_MARK
984 # ifdef PARALLEL_MARK
986 char * markers_string = GETENV("GC_MARKERS");
987 if (markers_string != NULL) {
988 GC_markers = atoi(markers_string);
990 GC_markers = GC_nprocs;
995 # ifdef PARALLEL_MARK
997 if (GC_print_stats) {
998 GC_printf2("Number of processors = %ld, "
999 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
1002 if (GC_markers == 1) {
1003 GC_parallel = FALSE;
1005 if (GC_print_stats) {
1006 GC_printf0("Single marker thread, turning off parallel marking\n");
1011 /* Disable true incremental collection, but generational is OK. */
1012 GC_time_limit = GC_TIME_UNLIMITED;
1018 /* Perform all initializations, including those that */
1019 /* may require allocation. */
1020 /* Called without allocation lock. */
1021 /* Must be called before a second thread is created. */
1022 /* Called without allocation lock. */
1023 void GC_init_parallel()
1025 if (parallel_initialized) return;
1026 parallel_initialized = TRUE;
1028 /* GC_init() calls us back, so set flag first. */
1029 if (!GC_is_initialized) GC_init();
1030 /* If we are using a parallel marker, start the helper threads. */
1031 # ifdef PARALLEL_MARK
1032 if (GC_parallel) start_mark_threads();
1034 /* Initialize thread local free lists if used. */
1035 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1037 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1043 #if !defined(GC_DARWIN_THREADS)
1044 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
1046 sigset_t fudged_set;
1048 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
1050 sigdelset(&fudged_set, SIG_SUSPEND);
1053 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
1055 #endif /* !GC_DARWIN_THREADS */
1057 /* Wrappers for functions that are likely to block for an appreciable */
1058 /* length of time. Must be called in pairs, if at all. */
1059 /* Nothing much beyond the system call itself should be executed */
1060 /* between these. */
1062 void GC_start_blocking(void) {
1063 # define SP_SLOP 128
1066 me = GC_lookup_thread(pthread_self());
1067 GC_ASSERT(!(me -> thread_blocked));
1069 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
1071 # ifndef GC_DARWIN_THREADS
1072 me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
1076 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1078 /* Add some slop to the stack pointer, since the wrapped call may */
1079 /* end up pushing more callee-save registers. */
1080 # ifndef GC_DARWIN_THREADS
1081 # ifdef STACK_GROWS_UP
1082 me -> stop_info.stack_ptr += SP_SLOP;
1084 me -> stop_info.stack_ptr -= SP_SLOP;
1087 me -> thread_blocked = TRUE;
1091 void GC_end_blocking(void) {
1093 LOCK(); /* This will block if the world is stopped. */
1094 me = GC_lookup_thread(pthread_self());
1095 GC_ASSERT(me -> thread_blocked);
1096 me -> thread_blocked = FALSE;
1100 #if defined(GC_DGUX386_THREADS)
1101 #define __d10_sleep sleep
1102 #endif /* GC_DGUX386_THREADS */
1104 /* A wrapper for the standard C sleep function */
1105 int WRAP_FUNC(sleep) (unsigned int seconds)
1109 GC_start_blocking();
1110 result = REAL_FUNC(sleep)(seconds);
1116 void *(*start_routine)(void *);
1119 sem_t registered; /* 1 ==> in our thread table, but */
1120 /* parent hasn't yet noticed. */
1123 /* Called at thread exit. */
1124 /* Never called for main thread. That's OK, since it */
1125 /* results in at most a tiny one-time leak. And */
1126 /* linuxthreads doesn't reclaim the main threads */
1127 /* resources or id anyway. */
1128 void GC_thread_exit_proc(void *arg)
1133 me = GC_lookup_thread(pthread_self());
1134 GC_destroy_thread_local(me);
1135 if (me -> flags & DETACHED) {
1136 GC_delete_thread(pthread_self());
1138 me -> flags |= FINISHED;
1140 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1141 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1142 GC_remove_specific(GC_thread_key);
1144 /* The following may run the GC from "nonexistent" thread. */
1145 GC_wait_for_gc_completion(FALSE);
1149 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1152 GC_thread thread_gc_id;
1155 thread_gc_id = GC_lookup_thread(thread);
1156 /* This is guaranteed to be the intended one, since the thread id */
1157 /* cant have been recycled by pthreads. */
1159 result = REAL_FUNC(pthread_join)(thread, retval);
1160 # if defined (GC_FREEBSD_THREADS)
1161 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1162 appears to be) a spurious EINTR which caused the test and real code
1163 to gratuitously fail. Having looked at system pthread library source
1164 code, I see how this return code may be generated. In one path of
1165 code, pthread_join() just returns the errno setting of the thread
1166 being joined. This does not match the POSIX specification or the
1167 local man pages thus I have taken the liberty to catch this one
1168 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1169 if (result == EINTR) result = 0;
1173 /* Here the pthread thread id may have been recycled. */
1174 GC_delete_gc_thread(thread, thread_gc_id);
1181 WRAP_FUNC(pthread_detach)(pthread_t thread)
1184 GC_thread thread_gc_id;
1187 thread_gc_id = GC_lookup_thread(thread);
1189 result = REAL_FUNC(pthread_detach)(thread);
1192 thread_gc_id -> flags |= DETACHED;
1193 /* Here the pthread thread id may have been recycled. */
1194 if (thread_gc_id -> flags & FINISHED) {
1195 GC_delete_gc_thread(thread, thread_gc_id);
1202 GC_bool GC_in_thread_creation = FALSE;
1204 void * GC_start_routine(void * arg)
1207 struct start_info * si = arg;
1210 pthread_t my_pthread;
1211 void *(*start)(void *);
1214 my_pthread = pthread_self();
1215 # ifdef DEBUG_THREADS
1216 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1217 GC_printf1("pid = %ld\n", (long) getpid());
1218 GC_printf1("sp = 0x%lx\n", (long) &arg);
1221 GC_in_thread_creation = TRUE;
1222 me = GC_new_thread(my_pthread);
1223 GC_in_thread_creation = FALSE;
1224 #ifdef GC_DARWIN_THREADS
1225 me -> stop_info.mach_thread = mach_thread_self();
1227 me -> stop_info.stack_ptr = 0;
1229 me -> flags = si -> flags;
1230 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1231 /* doesn't work because the stack base in /proc/self/stat is the */
1232 /* one for the main thread. There is a strong argument that that's */
1233 /* a kernel bug, but a pervasive one. */
1234 # ifdef STACK_GROWS_DOWN
1235 me -> stack_end = (ptr_t)(((word)(&dummy) + (GC_page_size - 1))
1236 & ~(GC_page_size - 1));
1237 # ifndef GC_DARWIN_THREADS
1238 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1240 /* Needs to be plausible, since an asynchronous stack mark */
1241 /* should not crash. */
1243 me -> stack_end = (ptr_t)((word)(&dummy) & ~(GC_page_size - 1));
1244 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1246 /* This is dubious, since we may be more than a page into the stack, */
1247 /* and hence skip some of it, though it's not clear that matters. */
1249 me -> backing_store_end = (ptr_t)
1250 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1251 /* This is also < 100% convincing. We should also read this */
1252 /* from /proc, but the hook to do so isn't there yet. */
1255 start = si -> start_routine;
1256 # ifdef DEBUG_THREADS
1257 GC_printf1("start_routine = 0x%lx\n", start);
1259 start_arg = si -> arg;
1260 sem_post(&(si -> registered)); /* Last action on si. */
1261 /* OK to deallocate. */
1262 pthread_cleanup_push(GC_thread_exit_proc, 0);
1263 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1265 GC_init_thread_local(me);
1268 result = (*start)(start_arg);
1270 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1272 me -> status = result;
1273 pthread_cleanup_pop(1);
1274 /* Cleanup acquires lock, ensuring that we can't exit */
1275 /* while a collection that thinks we're alive is trying to stop */
1281 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1282 const pthread_attr_t *attr,
1283 void *(*start_routine)(void *), void *arg)
1288 struct start_info * si;
1289 /* This is otherwise saved only in an area mmapped by the thread */
1290 /* library, which isn't visible to the collector. */
1292 /* We resist the temptation to muck with the stack size here, */
1293 /* even if the default is unreasonably small. That's the client's */
1294 /* responsibility. */
1297 si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
1300 if (!parallel_initialized) GC_init_parallel();
1301 if (0 == si) return(ENOMEM);
1302 sem_init(&(si -> registered), 0, 0);
1303 si -> start_routine = start_routine;
1306 if (!GC_thr_initialized) GC_thr_init();
1307 # ifdef GC_ASSERTIONS
1311 pthread_attr_t my_attr;
1312 pthread_attr_init(&my_attr);
1313 pthread_attr_getstacksize(&my_attr, &stack_size);
1315 pthread_attr_getstacksize(attr, &stack_size);
1317 GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1318 /* Our threads may need to do some work for the GC. */
1319 /* Ridiculously small threads won't work, and they */
1320 /* probably wouldn't work anyway. */
1324 detachstate = PTHREAD_CREATE_JOINABLE;
1326 pthread_attr_getdetachstate(attr, &detachstate);
1328 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1329 si -> flags = my_flags;
1331 # ifdef DEBUG_THREADS
1332 GC_printf1("About to start new thread from thread 0x%X\n",
1336 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1338 # ifdef DEBUG_THREADS
1339 GC_printf1("Started thread 0x%X\n", *new_thread);
1341 /* Wait until child has been added to the thread table. */
1342 /* This also ensures that we hold onto si until the child is done */
1343 /* with it. Thus it doesn't matter whether it is otherwise */
1344 /* visible to the collector. */
1346 while (0 != sem_wait(&(si -> registered))) {
1347 if (EINTR != errno) ABORT("sem_wait failed");
1350 sem_destroy(&(si -> registered));
1352 GC_INTERNAL_FREE(si);
1358 #ifdef GENERIC_COMPARE_AND_SWAP
1359 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1361 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1362 GC_word old, GC_word new_val)
1365 pthread_mutex_lock(&GC_compare_and_swap_lock);
1372 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1376 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1379 pthread_mutex_lock(&GC_compare_and_swap_lock);
1381 *addr = old + how_much;
1382 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1386 #endif /* GENERIC_COMPARE_AND_SWAP */
1387 /* Spend a few cycles in a way that can't introduce contention with */
1388 /* othre threads. */
1392 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1393 volatile word dummy = 0;
1396 for (i = 0; i < 10; ++i) {
1397 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1398 __asm__ __volatile__ (" " : : : "memory");
1400 /* Something that's unlikely to be optimized away. */
1406 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1409 VOLATILE GC_bool GC_collecting = 0;
1410 /* A hint that we're in the collector and */
1411 /* holding the allocation lock for an */
1412 /* extended period. */
1414 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1415 /* If we don't want to use the below spinlock implementation, either */
1416 /* because we don't have a GC_test_and_set implementation, or because */
1417 /* we don't want to risk sleeping, we can still try spinning on */
1418 /* pthread_mutex_trylock for a while. This appears to be very */
1419 /* beneficial in many cases. */
1420 /* I suspect that under high contention this is nearly always better */
1421 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1422 /* Hence we still default to the spin lock. */
1423 /* This is also used to acquire the mark lock for the parallel */
1426 /* Here we use a strict exponential backoff scheme. I don't know */
1427 /* whether that's better or worse than the above. We eventually */
1428 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1429 /* explicitly sleep. */
1433 unsigned long GC_spin_count = 0;
1434 unsigned long GC_block_count = 0;
1435 unsigned long GC_unlocked_count = 0;
1438 void GC_generic_lock(pthread_mutex_t * lock)
1440 #ifndef NO_PTHREAD_TRYLOCK
1441 unsigned pause_length = 1;
1444 if (0 == pthread_mutex_trylock(lock)) {
1446 ++GC_unlocked_count;
1450 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1451 for (i = 0; i < pause_length; ++i) {
1454 switch(pthread_mutex_trylock(lock)) {
1463 ABORT("Unexpected error from pthread_mutex_trylock");
1466 #endif /* !NO_PTHREAD_TRYLOCK */
1470 pthread_mutex_lock(lock);
1473 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1475 #if defined(USE_SPIN_LOCK)
1477 /* Reasonably fast spin locks. Basically the same implementation */
1478 /* as STL alloc.h. This isn't really the right way to do this. */
1479 /* but until the POSIX scheduling mess gets straightened out ... */
1481 volatile unsigned int GC_allocate_lock = 0;
1486 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1487 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1488 static unsigned spin_max = low_spin_max;
1489 unsigned my_spin_max;
1490 static unsigned last_spins = 0;
1491 unsigned my_last_spins;
1494 if (!GC_test_and_set(&GC_allocate_lock)) {
1497 my_spin_max = spin_max;
1498 my_last_spins = last_spins;
1499 for (i = 0; i < my_spin_max; i++) {
1500 if (GC_collecting || GC_nprocs == 1) goto yield;
1501 if (i < my_last_spins/2 || GC_allocate_lock) {
1505 if (!GC_test_and_set(&GC_allocate_lock)) {
1508 * Spinning worked. Thus we're probably not being scheduled
1509 * against the other process with which we were contending.
1510 * Thus it makes sense to spin longer the next time.
1513 spin_max = high_spin_max;
1517 /* We are probably being scheduled against the other process. Sleep. */
1518 spin_max = low_spin_max;
1521 if (!GC_test_and_set(&GC_allocate_lock)) {
1524 # define SLEEP_THRESHOLD 12
1525 /* Under Linux very short sleeps tend to wait until */
1526 /* the current time quantum expires. On old Linux */
1527 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1528 /* (Under 2.4, this happens only for real-time */
1529 /* processes.) We want to minimize both behaviors */
1531 if (i < SLEEP_THRESHOLD) {
1537 /* Don't wait for more than about 15msecs, even */
1538 /* under extreme contention. */
1540 ts.tv_nsec = 1 << i;
1546 #else /* !USE_SPINLOCK */
1549 #ifndef NO_PTHREAD_TRYLOCK
1550 if (1 == GC_nprocs || GC_collecting) {
1551 pthread_mutex_lock(&GC_allocate_ml);
1553 GC_generic_lock(&GC_allocate_ml);
1555 #else /* !NO_PTHREAD_TRYLOCK */
1556 pthread_mutex_lock(&GC_allocate_ml);
1557 #endif /* !NO_PTHREAD_TRYLOCK */
1560 #endif /* !USE_SPINLOCK */
1562 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1564 #ifdef GC_ASSERTIONS
1565 pthread_t GC_mark_lock_holder = NO_THREAD;
1569 /* Ugly workaround for a linux threads bug in the final versions */
1570 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1571 /* field even when it fails to acquire the mutex. This causes */
1572 /* pthread_cond_wait to die. Remove for glibc2.2. */
1573 /* According to the man page, we should use */
1574 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1576 static pthread_mutex_t mark_mutex =
1577 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1579 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1582 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1584 void GC_acquire_mark_lock()
1587 if (pthread_mutex_lock(&mark_mutex) != 0) {
1588 ABORT("pthread_mutex_lock failed");
1591 GC_generic_lock(&mark_mutex);
1592 # ifdef GC_ASSERTIONS
1593 GC_mark_lock_holder = pthread_self();
1597 void GC_release_mark_lock()
1599 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1600 # ifdef GC_ASSERTIONS
1601 GC_mark_lock_holder = NO_THREAD;
1603 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1604 ABORT("pthread_mutex_unlock failed");
1608 /* Collector must wait for a freelist builders for 2 reasons: */
1609 /* 1) Mark bits may still be getting examined without lock. */
1610 /* 2) Partial free lists referenced only by locals may not be scanned */
1611 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1612 /* free-list link may be ignored. */
1613 void GC_wait_builder()
1615 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1616 # ifdef GC_ASSERTIONS
1617 GC_mark_lock_holder = NO_THREAD;
1619 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1620 ABORT("pthread_cond_wait failed");
1622 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1623 # ifdef GC_ASSERTIONS
1624 GC_mark_lock_holder = pthread_self();
1628 void GC_wait_for_reclaim()
1630 GC_acquire_mark_lock();
1631 while (GC_fl_builder_count > 0) {
1634 GC_release_mark_lock();
1637 void GC_notify_all_builder()
1639 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1640 if (pthread_cond_broadcast(&builder_cv) != 0) {
1641 ABORT("pthread_cond_broadcast failed");
1645 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1647 #ifdef PARALLEL_MARK
1649 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1651 void GC_wait_marker()
1653 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1654 # ifdef GC_ASSERTIONS
1655 GC_mark_lock_holder = NO_THREAD;
1657 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1658 ABORT("pthread_cond_wait failed");
1660 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1661 # ifdef GC_ASSERTIONS
1662 GC_mark_lock_holder = pthread_self();
1666 void GC_notify_all_marker()
1668 if (pthread_cond_broadcast(&mark_cv) != 0) {
1669 ABORT("pthread_cond_broadcast failed");
1673 #endif /* PARALLEL_MARK */
1675 # endif /* GC_LINUX_THREADS and friends */