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(GC_NETBSD_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"
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 */
129 #if defined(GC_NETBSD_THREADS)
130 # include <sys/param.h>
131 # include <sys/sysctl.h>
136 #if defined(GC_DGUX386_THREADS)
137 # include <sys/dg_sys_info.h>
138 # include <sys/_int_psem.h>
139 /* sem_t is an uint in DG/UX */
140 typedef unsigned int sem_t;
141 #endif /* GC_DGUX386_THREADS */
147 #ifdef GC_USE_LD_WRAP
148 # define WRAP_FUNC(f) __wrap_##f
149 # define REAL_FUNC(f) __real_##f
151 # define WRAP_FUNC(f) GC_##f
152 # if !defined(GC_DGUX386_THREADS)
153 # define REAL_FUNC(f) f
154 # else /* GC_DGUX386_THREADS */
155 # define REAL_FUNC(f) __d10_##f
156 # endif /* GC_DGUX386_THREADS */
157 # undef pthread_create
158 # if !defined(GC_DARWIN_THREADS)
159 # undef pthread_sigmask
162 # undef pthread_detach
163 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
164 && !defined(_PTHREAD_USE_PTDNAM_)
165 /* Restore the original mangled names on Tru64 UNIX. */
166 # define pthread_create __pthread_create
167 # define pthread_join __pthread_join
168 # define pthread_detach __pthread_detach
174 static GC_bool parallel_initialized = FALSE;
176 void GC_init_parallel();
178 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
180 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
182 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
183 #undef PACKAGE_BUGREPORT
185 #undef PACKAGE_STRING
186 #undef PACKAGE_TARNAME
187 #undef PACKAGE_VERSION
188 #include "mono/utils/mono-compiler.h"
191 GC_key_t GC_thread_key;
193 #ifdef USE_COMPILER_TLS
195 * gcc errors out with /tmp/ccdPMFuq.s:2994: Error: symbol `.LTLS4' is already defined
196 * if the static is removed on ppc.
198 #if defined(__ppc__) || defined(__powerpc__)
201 __thread MONO_TLS_FAST void* GC_thread_tls;
204 static GC_bool keys_initialized;
206 #ifdef MONO_DEBUGGER_SUPPORTED
207 #include "include/libgc-mono-debugger.h"
210 /* Recover the contents of the freelist array fl into the global one gfl.*/
211 /* Note that the indexing scheme differs, in that gfl has finer size */
212 /* resolution, even if not all entries are used. */
213 /* We hold the allocator lock. */
214 static void return_freelists(ptr_t *fl, ptr_t *gfl)
220 for (i = 1; i < NFREELISTS; ++i) {
221 nwords = i * (GRANULARITY/sizeof(word));
224 if ((word)q >= HBLKSIZE) {
225 if (gfl[nwords] == 0) {
229 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
235 /* Clear fl[i], since the thread structure may hang around. */
236 /* Do it in a way that is likely to trap if we access it. */
237 fl[i] = (ptr_t)HBLKSIZE;
241 /* We statically allocate a single "size 0" object. It is linked to */
242 /* itself, and is thus repeatedly reused for all size 0 allocation */
243 /* requests. (Size 0 gcj allocation requests are incorrect, and */
244 /* we arrange for those to fault asap.) */
245 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
247 void GC_delete_gc_thread(pthread_t id, GC_thread gct);
248 void GC_destroy_thread_local(GC_thread p);
250 void GC_thread_deregister_foreign (void *data)
252 GC_thread me = (GC_thread)data;
253 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
254 if (me -> flags & FOREIGN_THREAD) {
256 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
257 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
258 GC_destroy_thread_local (me);
260 GC_delete_gc_thread(me->id, me);
265 /* Each thread structure must be initialized. */
266 /* This call must be made from the new thread. */
267 /* Caller holds allocation lock. */
268 void GC_init_thread_local(GC_thread p)
272 if (!keys_initialized) {
273 if (0 != GC_key_create(&GC_thread_key, GC_thread_deregister_foreign)) {
274 ABORT("Failed to create key for local allocator");
276 keys_initialized = TRUE;
278 if (0 != GC_setspecific(GC_thread_key, p)) {
279 ABORT("Failed to set thread specific allocation pointers");
281 for (i = 1; i < NFREELISTS; ++i) {
282 p -> ptrfree_freelists[i] = (ptr_t)1;
283 p -> normal_freelists[i] = (ptr_t)1;
284 # ifdef GC_GCJ_SUPPORT
285 p -> gcj_freelists[i] = (ptr_t)1;
288 /* Set up the size 0 free lists. */
289 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
290 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
291 # ifdef GC_GCJ_SUPPORT
292 p -> gcj_freelists[0] = (ptr_t)(-1);
296 #ifdef GC_GCJ_SUPPORT
297 extern ptr_t * GC_gcjobjfreelist;
300 /* We hold the allocator lock. */
301 void GC_destroy_thread_local(GC_thread p)
303 /* We currently only do this from the thread itself or from */
304 /* the fork handler for a child process. */
306 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
308 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
309 return_freelists(p -> normal_freelists, GC_objfreelist);
310 # ifdef GC_GCJ_SUPPORT
311 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
315 extern GC_PTR GC_generic_malloc_many();
317 GC_PTR GC_local_malloc(size_t bytes)
319 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
320 return(GC_malloc(bytes));
322 int index = INDEX_FROM_BYTES(bytes);
325 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
326 GC_key_t k = GC_thread_key;
330 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
331 if (EXPECT(0 == k, 0)) {
332 /* This can happen if we get called when the world is */
333 /* being initialized. Whether we can actually complete */
334 /* the initialization then is unclear. */
339 tsd = GC_getspecific(GC_thread_key);
340 # ifdef GC_ASSERTIONS
342 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
345 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
347 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
348 ptr_t next = obj_link(my_entry);
349 GC_PTR result = (GC_PTR)my_entry;
351 obj_link(my_entry) = 0;
352 PREFETCH_FOR_WRITE(next);
354 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
355 *my_fl = my_entry + index + 1;
356 return GC_malloc(bytes);
358 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
359 if (*my_fl == 0) return GC_oom_fn(bytes);
360 return GC_local_malloc(bytes);
365 GC_PTR GC_local_malloc_atomic(size_t bytes)
367 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
368 return(GC_malloc_atomic(bytes));
370 int index = INDEX_FROM_BYTES(bytes);
371 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
372 -> ptrfree_freelists + index;
373 ptr_t my_entry = *my_fl;
375 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
376 GC_PTR result = (GC_PTR)my_entry;
377 *my_fl = obj_link(my_entry);
379 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
380 *my_fl = my_entry + index + 1;
381 return GC_malloc_atomic(bytes);
383 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
384 /* *my_fl is updated while the collector is excluded; */
385 /* the free list is always visible to the collector as */
387 if (*my_fl == 0) return GC_oom_fn(bytes);
388 return GC_local_malloc_atomic(bytes);
393 #ifdef GC_GCJ_SUPPORT
395 #include "include/gc_gcj.h"
398 extern GC_bool GC_gcj_malloc_initialized;
401 extern int GC_gcj_kind;
403 GC_PTR GC_local_gcj_malloc(size_t bytes,
404 void * ptr_to_struct_containing_descr)
406 GC_ASSERT(GC_gcj_malloc_initialized);
407 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
408 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
410 int index = INDEX_FROM_BYTES(bytes);
411 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
412 -> gcj_freelists + index;
413 ptr_t my_entry = *my_fl;
414 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
415 GC_PTR result = (GC_PTR)my_entry;
416 GC_ASSERT(!GC_incremental);
417 /* We assert that any concurrent marker will stop us. */
418 /* Thus it is impossible for a mark procedure to see the */
419 /* allocation of the next object, but to see this object */
420 /* still containing a free list pointer. Otherwise the */
421 /* marker might find a random "mark descriptor". */
422 *(volatile ptr_t *)my_fl = obj_link(my_entry);
423 /* We must update the freelist before we store the pointer. */
424 /* Otherwise a GC at this point would see a corrupted */
426 /* A memory barrier is probably never needed, since the */
427 /* action of stopping this thread will cause prior writes */
429 GC_ASSERT(((void * volatile *)result)[1] == 0);
430 *(void * volatile *)result = ptr_to_struct_containing_descr;
432 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
433 if (!GC_incremental) *my_fl = my_entry + index + 1;
434 /* In the incremental case, we always have to take this */
435 /* path. Thus we leave the counter alone. */
436 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
438 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
439 if (*my_fl == 0) return GC_oom_fn(bytes);
440 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
445 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
446 /* adjust it. The size is assumed to be such that it can be */
447 /* allocated as a small object. */
448 void * GC_local_gcj_fast_malloc(size_t lw, void * ptr_to_struct_containing_descr)
450 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
451 -> gcj_freelists + lw;
452 ptr_t my_entry = *my_fl;
454 GC_ASSERT(GC_gcj_malloc_initialized);
456 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
457 GC_PTR result = (GC_PTR)my_entry;
458 GC_ASSERT(!GC_incremental);
459 /* We assert that any concurrent marker will stop us. */
460 /* Thus it is impossible for a mark procedure to see the */
461 /* allocation of the next object, but to see this object */
462 /* still containing a free list pointer. Otherwise the */
463 /* marker might find a random "mark descriptor". */
464 *(volatile ptr_t *)my_fl = obj_link(my_entry);
465 /* We must update the freelist before we store the pointer. */
466 /* Otherwise a GC at this point would see a corrupted */
468 /* A memory barrier is probably never needed, since the */
469 /* action of stopping this thread will cause prior writes */
471 GC_ASSERT(((void * volatile *)result)[1] == 0);
472 *(void * volatile *)result = ptr_to_struct_containing_descr;
474 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
475 if (!GC_incremental) *my_fl = my_entry + lw + 1;
476 /* In the incremental case, we always have to take this */
477 /* path. Thus we leave the counter alone. */
478 return GC_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
480 GC_generic_malloc_many(BYTES_FROM_INDEX(lw), GC_gcj_kind, my_fl);
481 if (*my_fl == 0) return GC_oom_fn(BYTES_FROM_INDEX(lw));
482 return GC_local_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
486 #endif /* GC_GCJ_SUPPORT */
488 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
490 # define GC_destroy_thread_local(t)
492 # endif /* !THREAD_LOCAL_ALLOC */
496 To make sure that we're using LinuxThreads and not some other thread
497 package, we generate a dummy reference to `pthread_kill_other_threads_np'
498 (was `__pthread_initial_thread_bos' but that disappeared),
499 which is a symbol defined in LinuxThreads, but (hopefully) not in other
502 We no longer do this, since this code is now portable enough that it might
503 actually work for something else.
505 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
508 long GC_nprocs = 1; /* Number of processors. We may not have */
509 /* access to all of them, but this is as good */
510 /* a guess as any ... */
515 # define MAX_MARKERS 16
518 static ptr_t marker_sp[MAX_MARKERS] = {0};
520 void * GC_mark_thread(void * id)
524 marker_sp[(word)id] = GC_approx_sp();
525 for (;; ++my_mark_no) {
526 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
527 /* promptly. This is important if it were called from the signal */
528 /* handler or from the GC lock acquisition code. Under Linux, it's */
529 /* not safe to call it from a signal handler, since it uses mutexes */
530 /* and condition variables. Since it is called only here, the */
531 /* argument is unnecessary. */
532 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
533 /* resynchronize if we get far off, e.g. because GC_mark_no */
535 my_mark_no = GC_mark_no;
537 # ifdef DEBUG_THREADS
538 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
540 GC_help_marker(my_mark_no);
544 extern long GC_markers; /* Number of mark threads we would */
545 /* like to have. Includes the */
546 /* initiating thread. */
548 pthread_t GC_mark_threads[MAX_MARKERS];
550 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
552 static void start_mark_threads()
557 if (GC_markers > MAX_MARKERS) {
558 WARN("Limiting number of mark threads\n", 0);
559 GC_markers = MAX_MARKERS;
561 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
563 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
564 ABORT("pthread_attr_setdetachstate failed");
566 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
567 /* Default stack size is usually too small: fix it. */
568 /* Otherwise marker threads or GC may run out of */
570 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
575 if (pthread_attr_getstacksize(&attr, &old_size) != 0)
576 ABORT("pthread_attr_getstacksize failed\n");
577 if (old_size < MIN_STACK_SIZE) {
578 if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
579 ABORT("pthread_attr_setstacksize failed\n");
582 # endif /* HPUX || GC_DGUX386_THREADS */
584 if (GC_print_stats) {
585 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
588 for (i = 0; i < GC_markers - 1; ++i) {
589 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
590 GC_mark_thread, (void *)(word)i)) {
591 WARN("Marker thread creation failed, errno = %ld.\n", errno);
596 #else /* !PARALLEL_MARK */
598 static __inline__ void start_mark_threads()
602 #endif /* !PARALLEL_MARK */
604 GC_bool GC_thr_initialized = FALSE;
606 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
609 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
610 * put it into a separate function.
612 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
613 static __attribute__((noinline)) unsigned char* get_gc_thread_key_addr GC_PROTO((void))
615 return (unsigned char*)&GC_thread_key;
618 void GC_push_thread_structures GC_PROTO((void))
620 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
621 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
622 GC_push_all((ptr_t)get_gc_thread_key_addr(),
623 (ptr_t)(get_gc_thread_key_addr())+sizeof(&GC_thread_key));
629 void GC_push_thread_structures GC_PROTO((void))
631 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
632 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
633 GC_push_all((ptr_t)(&GC_thread_key),
634 (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
640 #ifdef THREAD_LOCAL_ALLOC
641 /* We must explicitly mark ptrfree and gcj free lists, since the free */
642 /* list links wouldn't otherwise be found. We also set them in the */
643 /* normal free lists, since that involves touching less memory than if */
644 /* we scanned them normally. */
645 void GC_mark_thread_local_free_lists(void)
651 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
652 for (p = GC_threads[i]; 0 != p; p = p -> next) {
653 for (j = 1; j < NFREELISTS; ++j) {
654 q = p -> ptrfree_freelists[j];
655 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
656 q = p -> normal_freelists[j];
657 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
658 # ifdef GC_GCJ_SUPPORT
659 q = p -> gcj_freelists[j];
660 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
661 # endif /* GC_GCJ_SUPPORT */
666 #endif /* THREAD_LOCAL_ALLOC */
668 static struct GC_Thread_Rep first_thread;
670 /* Add a thread to GC_threads. We assume it wasn't already there. */
671 /* Caller holds allocation lock. */
672 GC_thread GC_new_thread(pthread_t id)
674 int hv = ((word)id) % THREAD_TABLE_SZ;
676 static GC_bool first_thread_used = FALSE;
678 if (!first_thread_used) {
679 result = &first_thread;
680 first_thread_used = TRUE;
682 result = (struct GC_Thread_Rep *)
683 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
685 if (result == 0) return(0);
687 result -> next = GC_threads[hv];
688 GC_threads[hv] = result;
689 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
693 /* Delete a thread from GC_threads. We assume it is there. */
694 /* (The code intentionally traps if it wasn't.) */
695 /* Caller holds allocation lock. */
696 void GC_delete_thread(pthread_t id)
698 int hv = ((word)id) % THREAD_TABLE_SZ;
699 register GC_thread p = GC_threads[hv];
700 register GC_thread prev = 0;
702 while (!pthread_equal(p -> id, id)) {
707 GC_threads[hv] = p -> next;
709 prev -> next = p -> next;
711 #ifdef MONO_DEBUGGER_SUPPORTED
712 if (gc_thread_vtable && gc_thread_vtable->thread_exited)
713 gc_thread_vtable->thread_exited (id, &p->stop_info.stack_ptr);
716 #ifdef GC_DARWIN_THREADS
717 mach_port_deallocate(mach_task_self(), p->stop_info.mach_thread);
723 /* If a thread has been joined, but we have not yet */
724 /* been notified, then there may be more than one thread */
725 /* in the table with the same pthread id. */
726 /* This is OK, but we need a way to delete a specific one. */
727 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
729 int hv = ((word)id) % THREAD_TABLE_SZ;
730 register GC_thread p = GC_threads[hv];
731 register GC_thread prev = 0;
738 GC_threads[hv] = p -> next;
740 prev -> next = p -> next;
743 #ifdef GC_DARWIN_THREADS
744 mach_port_deallocate(mach_task_self(), p->stop_info.mach_thread);
750 /* Return a GC_thread corresponding to a given pthread_t. */
751 /* Returns 0 if it's not there. */
752 /* Caller holds allocation lock or otherwise inhibits */
754 /* If there is more than one thread with the given id we */
755 /* return the most recent one. */
756 GC_thread GC_lookup_thread(pthread_t id)
758 int hv = ((word)id) % THREAD_TABLE_SZ;
759 register GC_thread p = GC_threads[hv];
761 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
765 int GC_thread_is_registered (void)
770 ptr = (void *)GC_lookup_thread(pthread_self());
777 /* Remove all entries from the GC_threads table, except the */
778 /* one for the current thread. We need to do this in the child */
779 /* process after a fork(), since only the current thread */
780 /* survives in the child. */
781 void GC_remove_all_threads_but_me(void)
783 pthread_t self = pthread_self();
785 GC_thread p, next, me;
787 for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
789 for (p = GC_threads[hv]; 0 != p; p = next) {
791 if (p -> id == self) {
795 # ifdef THREAD_LOCAL_ALLOC
796 if (!(p -> flags & FINISHED)) {
797 GC_destroy_thread_local(p);
799 # endif /* THREAD_LOCAL_ALLOC */
800 if (p != &first_thread) GC_INTERNAL_FREE(p);
807 #endif /* HANDLE_FORK */
809 #ifdef USE_PROC_FOR_LIBRARIES
810 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
815 # ifdef PARALLEL_MARK
816 for (i = 0; i < GC_markers; ++i) {
817 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
820 for (i = 0; i < THREAD_TABLE_SZ; i++) {
821 for (p = GC_threads[i]; p != 0; p = p -> next) {
822 if (0 != p -> stack_end) {
823 # ifdef STACK_GROWS_UP
824 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
825 # else /* STACK_GROWS_DOWN */
826 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
833 #endif /* USE_PROC_FOR_LIBRARIES */
835 #ifdef GC_LINUX_THREADS
836 /* Return the number of processors, or i<= 0 if it can't be determined. */
839 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
840 /* appears to be buggy in many cases. */
841 /* We look for lines "cpu<n>" in /proc/stat. */
842 # define STAT_BUF_SIZE 4096
843 # define STAT_READ read
844 /* If read is wrapped, this may need to be redefined to call */
846 char stat_buf[STAT_BUF_SIZE];
849 /* Some old kernels only have a single "cpu nnnn ..." */
850 /* entry in /proc/stat. We identify those as */
854 f = open("/proc/stat", O_RDONLY);
855 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
856 WARN("Couldn't read /proc/stat\n", 0);
859 for (i = 0; i < len - 100; ++i) {
860 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
861 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
862 int cpu_no = atoi(stat_buf + i + 4);
863 if (cpu_no >= result) result = cpu_no + 1;
869 #endif /* GC_LINUX_THREADS */
871 /* We hold the GC lock. Wait until an in-progress GC has finished. */
872 /* Repeatedly RELEASES GC LOCK in order to wait. */
873 /* If wait_for_all is true, then we exit with the GC lock held and no */
874 /* collection in progress; otherwise we just wait for the current GC */
876 extern GC_bool GC_collection_in_progress();
877 void GC_wait_for_gc_completion(GC_bool wait_for_all)
879 if (GC_incremental && GC_collection_in_progress()) {
880 int old_gc_no = GC_gc_no;
882 /* Make sure that no part of our stack is still on the mark stack, */
883 /* since it's about to be unmapped. */
884 while (GC_incremental && GC_collection_in_progress()
885 && (wait_for_all || old_gc_no == GC_gc_no)) {
887 GC_in_thread_creation = TRUE;
888 GC_collect_a_little_inner(1);
889 GC_in_thread_creation = FALSE;
899 /* Procedures called before and after a fork. The goal here is to make */
900 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
901 /* attainable, since the single UNIX spec seems to imply that one */
902 /* should only call async-signal-safe functions, and we probably can't */
903 /* quite guarantee that. But we give it our best shot. (That same */
904 /* spec also implies that it's not safe to call the system malloc */
905 /* between fork() and exec(). Thus we're doing no worse than it. */
907 /* Called before a fork() */
908 void GC_fork_prepare_proc(void)
910 /* Acquire all relevant locks, so that after releasing the locks */
911 /* the child will see a consistent state in which monitor */
912 /* invariants hold. Unfortunately, we can't acquire libc locks */
913 /* we might need, and there seems to be no guarantee that libc */
914 /* must install a suitable fork handler. */
915 /* Wait for an ongoing GC to finish, since we can't finish it in */
916 /* the (one remaining thread in) the child. */
918 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
919 GC_wait_for_reclaim();
921 GC_wait_for_gc_completion(TRUE);
922 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
923 GC_acquire_mark_lock();
927 /* Called in parent after a fork() */
928 void GC_fork_parent_proc(void)
930 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
931 GC_release_mark_lock();
936 /* Called in child after a fork() */
937 void GC_fork_child_proc(void)
939 /* Clean up the thread table, so that just our thread is left. */
940 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
941 GC_release_mark_lock();
943 GC_remove_all_threads_but_me();
944 # ifdef PARALLEL_MARK
945 /* Turn off parallel marking in the child, since we are probably */
946 /* just going to exec, and we would have to restart mark threads. */
949 # endif /* PARALLEL_MARK */
952 #endif /* HANDLE_FORK */
954 #if defined(GC_DGUX386_THREADS)
955 /* Return the number of processors, or i<= 0 if it can't be determined. */
958 /* <takis@XFree86.Org> */
960 struct dg_sys_info_pm_info pm_sysinfo;
963 status = dg_sys_info((long int *) &pm_sysinfo,
964 DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
966 /* set -1 for error */
970 numCpus = pm_sysinfo.idle_vp_count;
972 # ifdef DEBUG_THREADS
973 GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
977 #endif /* GC_DGUX386_THREADS */
979 /* We hold the allocation lock. */
982 # ifndef GC_DARWIN_THREADS
987 if (GC_thr_initialized) return;
988 GC_thr_initialized = TRUE;
991 /* Prepare for a possible fork. */
992 pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
994 # endif /* HANDLE_FORK */
995 /* Add the initial thread, so we can stop it. */
996 t = GC_new_thread(pthread_self());
997 # ifdef GC_DARWIN_THREADS
998 t -> stop_info.mach_thread = mach_thread_self();
1000 t -> stop_info.stack_ptr = (ptr_t)(&dummy);
1002 t -> flags = DETACHED | MAIN_THREAD;
1003 #ifdef MONO_DEBUGGER_SUPPORTED
1004 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1005 gc_thread_vtable->thread_created (pthread_self (), &t->stop_info.stack_ptr);
1010 /* Set GC_nprocs. */
1012 char * nprocs_string = GETENV("GC_NPROCS");
1014 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
1016 if (GC_nprocs <= 0) {
1017 # if defined(GC_HPUX_THREADS)
1018 GC_nprocs = pthread_num_processors_np();
1020 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
1021 GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
1022 if (GC_nprocs <= 0) GC_nprocs = 1;
1024 # if defined(GC_IRIX_THREADS)
1025 GC_nprocs = sysconf(_SC_NPROC_ONLN);
1026 if (GC_nprocs <= 0) GC_nprocs = 1;
1028 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS) || defined(GC_NETBSD_THREADS)
1030 size_t len = sizeof(ncpus);
1031 sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
1034 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1035 GC_nprocs = GC_get_nprocs();
1038 if (GC_nprocs <= 0) {
1039 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
1041 # ifdef PARALLEL_MARK
1045 # ifdef PARALLEL_MARK
1047 char * markers_string = GETENV("GC_MARKERS");
1048 if (markers_string != NULL) {
1049 GC_markers = atoi(markers_string);
1051 GC_markers = GC_nprocs;
1056 # ifdef PARALLEL_MARK
1058 if (GC_print_stats) {
1059 GC_printf2("Number of processors = %ld, "
1060 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
1063 if (GC_markers == 1) {
1064 GC_parallel = FALSE;
1066 if (GC_print_stats) {
1067 GC_printf0("Single marker thread, turning off parallel marking\n");
1072 /* Disable true incremental collection, but generational is OK. */
1073 GC_time_limit = GC_TIME_UNLIMITED;
1075 /* If we are using a parallel marker, actually start helper threads. */
1076 if (GC_parallel) start_mark_threads();
1081 /* Perform all initializations, including those that */
1082 /* may require allocation. */
1083 /* Called without allocation lock. */
1084 /* Must be called before a second thread is created. */
1085 /* Called without allocation lock. */
1086 void GC_init_parallel()
1088 if (parallel_initialized) return;
1089 parallel_initialized = TRUE;
1091 /* GC_init() calls us back, so set flag first. */
1092 if (!GC_is_initialized) GC_init();
1093 /* Initialize thread local free lists if used. */
1094 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1096 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1102 #if !defined(GC_DARWIN_THREADS)
1103 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
1105 sigset_t fudged_set;
1107 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
1109 sigdelset(&fudged_set, SIG_SUSPEND);
1112 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
1114 #endif /* !GC_DARWIN_THREADS */
1116 /* Wrappers for functions that are likely to block for an appreciable */
1117 /* length of time. Must be called in pairs, if at all. */
1118 /* Nothing much beyond the system call itself should be executed */
1119 /* between these. */
1121 void GC_start_blocking(void) {
1122 # define SP_SLOP 128
1125 me = GC_lookup_thread(pthread_self());
1126 GC_ASSERT(!(me -> thread_blocked));
1128 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
1130 # ifndef GC_DARWIN_THREADS
1131 me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
1135 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1137 /* Add some slop to the stack pointer, since the wrapped call may */
1138 /* end up pushing more callee-save registers. */
1139 # ifndef GC_DARWIN_THREADS
1140 # ifdef STACK_GROWS_UP
1141 me -> stop_info.stack_ptr += SP_SLOP;
1143 me -> stop_info.stack_ptr -= SP_SLOP;
1146 me -> thread_blocked = TRUE;
1150 void GC_end_blocking(void) {
1152 LOCK(); /* This will block if the world is stopped. */
1153 me = GC_lookup_thread(pthread_self());
1154 GC_ASSERT(me -> thread_blocked);
1155 me -> thread_blocked = FALSE;
1159 #if defined(GC_DGUX386_THREADS)
1160 #define __d10_sleep sleep
1161 #endif /* GC_DGUX386_THREADS */
1163 /* A wrapper for the standard C sleep function */
1164 int WRAP_FUNC(sleep) (unsigned int seconds)
1168 GC_start_blocking();
1169 result = REAL_FUNC(sleep)(seconds);
1175 void *(*start_routine)(void *);
1178 sem_t registered; /* 1 ==> in our thread table, but */
1179 /* parent hasn't yet noticed. */
1182 /* Called at thread exit. */
1183 /* Never called for main thread. That's OK, since it */
1184 /* results in at most a tiny one-time leak. And */
1185 /* linuxthreads doesn't reclaim the main threads */
1186 /* resources or id anyway. */
1187 void GC_thread_exit_proc(void *arg)
1192 me = GC_lookup_thread(pthread_self());
1193 GC_destroy_thread_local(me);
1194 if (me -> flags & DETACHED) {
1195 GC_delete_thread(pthread_self());
1197 me -> flags |= FINISHED;
1199 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1200 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1201 GC_remove_specific(GC_thread_key);
1203 /* The following may run the GC from "nonexistent" thread. */
1204 GC_wait_for_gc_completion(FALSE);
1208 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1211 GC_thread thread_gc_id;
1214 thread_gc_id = GC_lookup_thread(thread);
1215 /* This is guaranteed to be the intended one, since the thread id */
1216 /* cant have been recycled by pthreads. */
1218 result = REAL_FUNC(pthread_join)(thread, retval);
1219 # if defined (GC_FREEBSD_THREADS)
1220 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1221 appears to be) a spurious EINTR which caused the test and real code
1222 to gratuitously fail. Having looked at system pthread library source
1223 code, I see how this return code may be generated. In one path of
1224 code, pthread_join() just returns the errno setting of the thread
1225 being joined. This does not match the POSIX specification or the
1226 local man pages thus I have taken the liberty to catch this one
1227 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1228 if (result == EINTR) result = 0;
1232 /* Here the pthread thread id may have been recycled. */
1233 GC_delete_gc_thread(thread, thread_gc_id);
1240 WRAP_FUNC(pthread_detach)(pthread_t thread)
1243 GC_thread thread_gc_id;
1246 thread_gc_id = GC_lookup_thread(thread);
1248 result = REAL_FUNC(pthread_detach)(thread);
1251 thread_gc_id -> flags |= DETACHED;
1252 /* Here the pthread thread id may have been recycled. */
1253 if (thread_gc_id -> flags & FINISHED) {
1254 GC_delete_gc_thread(thread, thread_gc_id);
1261 GC_bool GC_in_thread_creation = FALSE;
1263 typedef void *(*ThreadStartFn)(void *);
1264 void * GC_start_routine_head(void * arg, void *base_addr,
1265 ThreadStartFn *start, void **start_arg )
1267 struct start_info * si = arg;
1270 pthread_t my_pthread;
1272 my_pthread = pthread_self();
1273 # ifdef DEBUG_THREADS
1274 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1275 GC_printf1("pid = %ld\n", (long) getpid());
1276 GC_printf1("sp = 0x%lx\n", (long) &arg);
1279 GC_in_thread_creation = TRUE;
1280 me = GC_new_thread(my_pthread);
1281 GC_in_thread_creation = FALSE;
1282 #ifdef GC_DARWIN_THREADS
1283 me -> stop_info.mach_thread = mach_thread_self();
1285 me -> stop_info.stack_ptr = 0;
1287 me -> flags = si -> flags;
1288 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1289 /* doesn't work because the stack base in /proc/self/stat is the */
1290 /* one for the main thread. There is a strong argument that that's */
1291 /* a kernel bug, but a pervasive one. */
1292 # ifdef STACK_GROWS_DOWN
1293 me -> stack_end = (ptr_t)(((word)(base_addr) + (GC_page_size - 1))
1294 & ~(GC_page_size - 1));
1295 # ifndef GC_DARWIN_THREADS
1296 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1298 /* Needs to be plausible, since an asynchronous stack mark */
1299 /* should not crash. */
1301 me -> stack_end = (ptr_t)((word)(base_addr) & ~(GC_page_size - 1));
1302 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1304 /* This is dubious, since we may be more than a page into the stack, */
1305 /* and hence skip some of it, though it's not clear that matters. */
1307 me -> backing_store_end = (ptr_t)
1308 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1309 /* This is also < 100% convincing. We should also read this */
1310 /* from /proc, but the hook to do so isn't there yet. */
1312 #ifdef MONO_DEBUGGER_SUPPORTED
1313 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1314 gc_thread_vtable->thread_created (my_pthread, &me->stop_info.stack_ptr);
1318 if (start) *start = si -> start_routine;
1319 if (start_arg) *start_arg = si -> arg;
1321 sem_post(&(si -> registered)); /* Last action on si. */
1322 /* OK to deallocate. */
1323 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1325 GC_init_thread_local(me);
1332 int GC_thread_register_foreign (void *base_addr)
1334 struct start_info si = { 0, }; /* stacked for legibility & locking */
1337 # ifdef DEBUG_THREADS
1338 GC_printf1( "GC_thread_register_foreign %p\n", &si );
1341 si.flags = FOREIGN_THREAD;
1343 if (!parallel_initialized) GC_init_parallel();
1345 if (!GC_thr_initialized) GC_thr_init();
1349 me = GC_start_routine_head(&si, base_addr, NULL, NULL);
1354 void * GC_start_routine(void * arg)
1357 struct start_info * si = arg;
1360 ThreadStartFn start;
1363 me = GC_start_routine_head (arg, &dummy, &start, &start_arg);
1365 pthread_cleanup_push(GC_thread_exit_proc, 0);
1366 # ifdef DEBUG_THREADS
1367 GC_printf1("start_routine = 0x%lx\n", start);
1369 result = (*start)(start_arg);
1371 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1373 me -> status = result;
1374 pthread_cleanup_pop(1);
1375 /* Cleanup acquires lock, ensuring that we can't exit */
1376 /* while a collection that thinks we're alive is trying to stop */
1382 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1383 const pthread_attr_t *attr,
1384 void *(*start_routine)(void *), void *arg)
1389 struct start_info * si;
1390 /* This is otherwise saved only in an area mmapped by the thread */
1391 /* library, which isn't visible to the collector. */
1393 /* We resist the temptation to muck with the stack size here, */
1394 /* even if the default is unreasonably small. That's the client's */
1395 /* responsibility. */
1398 si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
1401 if (!parallel_initialized) GC_init_parallel();
1402 if (0 == si) return(ENOMEM);
1403 sem_init(&(si -> registered), 0, 0);
1404 si -> start_routine = start_routine;
1407 if (!GC_thr_initialized) GC_thr_init();
1408 # ifdef GC_ASSERTIONS
1412 pthread_attr_t my_attr;
1413 pthread_attr_init(&my_attr);
1414 pthread_attr_getstacksize(&my_attr, &stack_size);
1416 pthread_attr_getstacksize(attr, &stack_size);
1418 # ifdef PARALLEL_MARK
1419 GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1421 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1422 /* HBLKSIZE=8192, sizeof(word)=8 */
1423 GC_ASSERT(stack_size >= 65536);
1425 /* Our threads may need to do some work for the GC. */
1426 /* Ridiculously small threads won't work, and they */
1427 /* probably wouldn't work anyway. */
1431 detachstate = PTHREAD_CREATE_JOINABLE;
1433 pthread_attr_getdetachstate(attr, &detachstate);
1435 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1436 si -> flags = my_flags;
1438 # ifdef DEBUG_THREADS
1439 GC_printf1("About to start new thread from thread 0x%X\n",
1443 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1445 # ifdef DEBUG_THREADS
1446 GC_printf1("Started thread 0x%X\n", *new_thread);
1448 /* Wait until child has been added to the thread table. */
1449 /* This also ensures that we hold onto si until the child is done */
1450 /* with it. Thus it doesn't matter whether it is otherwise */
1451 /* visible to the collector. */
1453 while (0 != sem_wait(&(si -> registered))) {
1454 if (EINTR != errno) ABORT("sem_wait failed");
1457 sem_destroy(&(si -> registered));
1459 GC_INTERNAL_FREE(si);
1465 #ifdef GENERIC_COMPARE_AND_SWAP
1466 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1468 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1469 GC_word old, GC_word new_val)
1472 pthread_mutex_lock(&GC_compare_and_swap_lock);
1479 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1483 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1486 pthread_mutex_lock(&GC_compare_and_swap_lock);
1488 *addr = old + how_much;
1489 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1493 #endif /* GENERIC_COMPARE_AND_SWAP */
1494 /* Spend a few cycles in a way that can't introduce contention with */
1495 /* othre threads. */
1499 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1500 volatile word dummy = 0;
1503 for (i = 0; i < 10; ++i) {
1504 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1505 __asm__ __volatile__ (" " : : : "memory");
1507 /* Something that's unlikely to be optimized away. */
1513 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1516 VOLATILE GC_bool GC_collecting = 0;
1517 /* A hint that we're in the collector and */
1518 /* holding the allocation lock for an */
1519 /* extended period. */
1521 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1522 /* If we don't want to use the below spinlock implementation, either */
1523 /* because we don't have a GC_test_and_set implementation, or because */
1524 /* we don't want to risk sleeping, we can still try spinning on */
1525 /* pthread_mutex_trylock for a while. This appears to be very */
1526 /* beneficial in many cases. */
1527 /* I suspect that under high contention this is nearly always better */
1528 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1529 /* Hence we still default to the spin lock. */
1530 /* This is also used to acquire the mark lock for the parallel */
1533 /* Here we use a strict exponential backoff scheme. I don't know */
1534 /* whether that's better or worse than the above. We eventually */
1535 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1536 /* explicitly sleep. */
1540 unsigned long GC_spin_count = 0;
1541 unsigned long GC_block_count = 0;
1542 unsigned long GC_unlocked_count = 0;
1545 void GC_generic_lock(pthread_mutex_t * lock)
1547 #ifndef NO_PTHREAD_TRYLOCK
1548 unsigned pause_length = 1;
1551 if (0 == pthread_mutex_trylock(lock)) {
1553 ++GC_unlocked_count;
1557 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1558 for (i = 0; i < pause_length; ++i) {
1561 switch(pthread_mutex_trylock(lock)) {
1570 ABORT("Unexpected error from pthread_mutex_trylock");
1573 #endif /* !NO_PTHREAD_TRYLOCK */
1577 pthread_mutex_lock(lock);
1580 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1582 #if defined(USE_SPIN_LOCK)
1584 /* Reasonably fast spin locks. Basically the same implementation */
1585 /* as STL alloc.h. This isn't really the right way to do this. */
1586 /* but until the POSIX scheduling mess gets straightened out ... */
1588 volatile unsigned int GC_allocate_lock = 0;
1593 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1594 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1595 static unsigned spin_max = low_spin_max;
1596 unsigned my_spin_max;
1597 static unsigned last_spins = 0;
1598 unsigned my_last_spins;
1601 if (!GC_test_and_set(&GC_allocate_lock)) {
1604 my_spin_max = spin_max;
1605 my_last_spins = last_spins;
1606 for (i = 0; i < my_spin_max; i++) {
1607 if (GC_collecting || GC_nprocs == 1) goto yield;
1608 if (i < my_last_spins/2 || GC_allocate_lock) {
1612 if (!GC_test_and_set(&GC_allocate_lock)) {
1615 * Spinning worked. Thus we're probably not being scheduled
1616 * against the other process with which we were contending.
1617 * Thus it makes sense to spin longer the next time.
1620 spin_max = high_spin_max;
1624 /* We are probably being scheduled against the other process. Sleep. */
1625 spin_max = low_spin_max;
1628 if (!GC_test_and_set(&GC_allocate_lock)) {
1631 # define SLEEP_THRESHOLD 12
1632 /* Under Linux very short sleeps tend to wait until */
1633 /* the current time quantum expires. On old Linux */
1634 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1635 /* (Under 2.4, this happens only for real-time */
1636 /* processes.) We want to minimize both behaviors */
1638 if (i < SLEEP_THRESHOLD) {
1644 /* Don't wait for more than about 15msecs, even */
1645 /* under extreme contention. */
1647 ts.tv_nsec = 1 << i;
1653 #else /* !USE_SPINLOCK */
1656 #ifndef NO_PTHREAD_TRYLOCK
1657 if (1 == GC_nprocs || GC_collecting) {
1658 pthread_mutex_lock(&GC_allocate_ml);
1660 GC_generic_lock(&GC_allocate_ml);
1662 #else /* !NO_PTHREAD_TRYLOCK */
1663 pthread_mutex_lock(&GC_allocate_ml);
1664 #endif /* !NO_PTHREAD_TRYLOCK */
1667 #endif /* !USE_SPINLOCK */
1669 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1671 #ifdef GC_ASSERTIONS
1672 pthread_t GC_mark_lock_holder = NO_THREAD;
1676 /* Ugly workaround for a linux threads bug in the final versions */
1677 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1678 /* field even when it fails to acquire the mutex. This causes */
1679 /* pthread_cond_wait to die. Remove for glibc2.2. */
1680 /* According to the man page, we should use */
1681 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1683 static pthread_mutex_t mark_mutex =
1684 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1686 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1689 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1691 void GC_acquire_mark_lock()
1694 if (pthread_mutex_lock(&mark_mutex) != 0) {
1695 ABORT("pthread_mutex_lock failed");
1698 GC_generic_lock(&mark_mutex);
1699 # ifdef GC_ASSERTIONS
1700 GC_mark_lock_holder = pthread_self();
1704 void GC_release_mark_lock()
1706 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1707 # ifdef GC_ASSERTIONS
1708 GC_mark_lock_holder = NO_THREAD;
1710 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1711 ABORT("pthread_mutex_unlock failed");
1715 /* Collector must wait for a freelist builders for 2 reasons: */
1716 /* 1) Mark bits may still be getting examined without lock. */
1717 /* 2) Partial free lists referenced only by locals may not be scanned */
1718 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1719 /* free-list link may be ignored. */
1720 void GC_wait_builder()
1722 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1723 # ifdef GC_ASSERTIONS
1724 GC_mark_lock_holder = NO_THREAD;
1726 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1727 ABORT("pthread_cond_wait failed");
1729 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1730 # ifdef GC_ASSERTIONS
1731 GC_mark_lock_holder = pthread_self();
1735 void GC_wait_for_reclaim()
1737 GC_acquire_mark_lock();
1738 while (GC_fl_builder_count > 0) {
1741 GC_release_mark_lock();
1744 void GC_notify_all_builder()
1746 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1747 if (pthread_cond_broadcast(&builder_cv) != 0) {
1748 ABORT("pthread_cond_broadcast failed");
1752 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1754 #ifdef PARALLEL_MARK
1756 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1758 void GC_wait_marker()
1760 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1761 # ifdef GC_ASSERTIONS
1762 GC_mark_lock_holder = NO_THREAD;
1764 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1765 ABORT("pthread_cond_wait failed");
1767 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1768 # ifdef GC_ASSERTIONS
1769 GC_mark_lock_holder = pthread_self();
1773 void GC_notify_all_marker()
1775 if (pthread_cond_broadcast(&mark_cv) != 0) {
1776 ABORT("pthread_cond_broadcast failed");
1780 #endif /* PARALLEL_MARK */
1782 # endif /* GC_LINUX_THREADS and friends */