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 /* Note sles9 gcc on powerpc gets confused by the define to set GC_thread_tls and pthread_setspecific
101 * so we actually use a static inline function decalred below that is equivalent to:
102 * define GC_setspecific(key, v) (GC_thread_tls = (v), pthread_setspecific ((key), (v)))
104 # define GC_getspecific(x) (GC_thread_tls)
105 # define GC_key_create pthread_key_create
106 typedef pthread_key_t GC_key_t;
110 # include <pthread.h>
115 # include <sys/mman.h>
116 # include <sys/time.h>
117 # include <sys/types.h>
118 # include <sys/stat.h>
122 #if defined(GC_DARWIN_THREADS)
123 # include "private/darwin_semaphore.h"
125 # include <semaphore.h>
126 #endif /* !GC_DARWIN_THREADS */
128 #if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
129 # include <sys/sysctl.h>
130 #endif /* GC_DARWIN_THREADS */
132 #if defined(GC_NETBSD_THREADS)
133 # include <sys/param.h>
134 # include <sys/sysctl.h>
139 #if defined(GC_DGUX386_THREADS)
140 # include <sys/dg_sys_info.h>
141 # include <sys/_int_psem.h>
142 /* sem_t is an uint in DG/UX */
143 typedef unsigned int sem_t;
144 #endif /* GC_DGUX386_THREADS */
150 #ifdef GC_USE_LD_WRAP
151 # define WRAP_FUNC(f) __wrap_##f
152 # define REAL_FUNC(f) __real_##f
154 # define WRAP_FUNC(f) GC_##f
155 # if !defined(GC_DGUX386_THREADS)
156 # define REAL_FUNC(f) f
157 # else /* GC_DGUX386_THREADS */
158 # define REAL_FUNC(f) __d10_##f
159 # endif /* GC_DGUX386_THREADS */
160 # undef pthread_create
161 # if !defined(GC_DARWIN_THREADS)
162 # undef pthread_sigmask
165 # undef pthread_detach
166 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
167 && !defined(_PTHREAD_USE_PTDNAM_)
168 /* Restore the original mangled names on Tru64 UNIX. */
169 # define pthread_create __pthread_create
170 # define pthread_join __pthread_join
171 # define pthread_detach __pthread_detach
177 static GC_bool parallel_initialized = FALSE;
179 void GC_init_parallel();
181 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
183 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
185 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
186 #undef PACKAGE_BUGREPORT
188 #undef PACKAGE_STRING
189 #undef PACKAGE_TARNAME
190 #undef PACKAGE_VERSION
191 #include "mono/utils/mono-compiler.h"
194 GC_key_t GC_thread_key;
196 #ifdef USE_COMPILER_TLS
197 __thread MONO_TLS_FAST void* GC_thread_tls;
200 * gcc errors out with /tmp/ccdPMFuq.s:2994: Error: symbol `.LTLS4' is already defined
201 * if the inline is added on powerpc
203 #if !defined(__ppc__) && !defined(__powerpc__)
206 static int GC_setspecific (GC_key_t key, void *value) {
207 GC_thread_tls = value;
208 return pthread_setspecific (key, value);
212 static GC_bool keys_initialized;
214 #ifdef MONO_DEBUGGER_SUPPORTED
215 #include "include/libgc-mono-debugger.h"
218 /* Recover the contents of the freelist array fl into the global one gfl.*/
219 /* Note that the indexing scheme differs, in that gfl has finer size */
220 /* resolution, even if not all entries are used. */
221 /* We hold the allocator lock. */
222 static void return_freelists(ptr_t *fl, ptr_t *gfl)
228 for (i = 1; i < NFREELISTS; ++i) {
229 nwords = i * (GRANULARITY/sizeof(word));
232 if ((word)q >= HBLKSIZE) {
233 if (gfl[nwords] == 0) {
237 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
243 /* Clear fl[i], since the thread structure may hang around. */
244 /* Do it in a way that is likely to trap if we access it. */
245 fl[i] = (ptr_t)HBLKSIZE;
249 /* We statically allocate a single "size 0" object. It is linked to */
250 /* itself, and is thus repeatedly reused for all size 0 allocation */
251 /* requests. (Size 0 gcj allocation requests are incorrect, and */
252 /* we arrange for those to fault asap.) */
253 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
255 void GC_delete_gc_thread(pthread_t id, GC_thread gct);
256 void GC_destroy_thread_local(GC_thread p);
258 void GC_thread_deregister_foreign (void *data)
260 GC_thread me = (GC_thread)data;
261 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
262 if (me -> flags & FOREIGN_THREAD) {
264 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
265 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
266 GC_destroy_thread_local (me);
268 GC_delete_gc_thread(me->id, me);
273 /* Each thread structure must be initialized. */
274 /* This call must be made from the new thread. */
275 /* Caller holds allocation lock. */
276 void GC_init_thread_local(GC_thread p)
280 if (!keys_initialized) {
281 if (0 != GC_key_create(&GC_thread_key, GC_thread_deregister_foreign)) {
282 ABORT("Failed to create key for local allocator");
284 keys_initialized = TRUE;
286 if (0 != GC_setspecific(GC_thread_key, p)) {
287 ABORT("Failed to set thread specific allocation pointers");
289 for (i = 1; i < NFREELISTS; ++i) {
290 p -> ptrfree_freelists[i] = (ptr_t)1;
291 p -> normal_freelists[i] = (ptr_t)1;
292 # ifdef GC_GCJ_SUPPORT
293 p -> gcj_freelists[i] = (ptr_t)1;
296 /* Set up the size 0 free lists. */
297 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
298 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
299 # ifdef GC_GCJ_SUPPORT
300 p -> gcj_freelists[0] = (ptr_t)(-1);
304 #ifdef GC_GCJ_SUPPORT
305 extern ptr_t * GC_gcjobjfreelist;
308 /* We hold the allocator lock. */
309 void GC_destroy_thread_local(GC_thread p)
311 /* We currently only do this from the thread itself or from */
312 /* the fork handler for a child process. */
314 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
316 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
317 return_freelists(p -> normal_freelists, GC_objfreelist);
318 # ifdef GC_GCJ_SUPPORT
319 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
323 extern GC_PTR GC_generic_malloc_many();
325 GC_PTR GC_local_malloc(size_t bytes)
327 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
328 return(GC_malloc(bytes));
330 int index = INDEX_FROM_BYTES(bytes);
333 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
334 GC_key_t k = GC_thread_key;
338 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
339 if (EXPECT(0 == k, 0)) {
340 /* This can happen if we get called when the world is */
341 /* being initialized. Whether we can actually complete */
342 /* the initialization then is unclear. */
347 tsd = GC_getspecific(GC_thread_key);
348 # ifdef GC_ASSERTIONS
350 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
353 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
355 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
356 ptr_t next = obj_link(my_entry);
357 GC_PTR result = (GC_PTR)my_entry;
359 obj_link(my_entry) = 0;
360 PREFETCH_FOR_WRITE(next);
362 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
363 *my_fl = my_entry + index + 1;
364 return GC_malloc(bytes);
366 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
367 if (*my_fl == 0) return GC_oom_fn(bytes);
368 return GC_local_malloc(bytes);
373 GC_PTR GC_local_malloc_atomic(size_t bytes)
375 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
376 return(GC_malloc_atomic(bytes));
378 int index = INDEX_FROM_BYTES(bytes);
379 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
380 -> ptrfree_freelists + index;
381 ptr_t my_entry = *my_fl;
383 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
384 GC_PTR result = (GC_PTR)my_entry;
385 *my_fl = obj_link(my_entry);
387 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
388 *my_fl = my_entry + index + 1;
389 return GC_malloc_atomic(bytes);
391 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
392 /* *my_fl is updated while the collector is excluded; */
393 /* the free list is always visible to the collector as */
395 if (*my_fl == 0) return GC_oom_fn(bytes);
396 return GC_local_malloc_atomic(bytes);
401 #ifdef GC_GCJ_SUPPORT
403 #include "include/gc_gcj.h"
406 extern GC_bool GC_gcj_malloc_initialized;
409 extern int GC_gcj_kind;
411 GC_PTR GC_local_gcj_malloc(size_t bytes,
412 void * ptr_to_struct_containing_descr)
414 GC_ASSERT(GC_gcj_malloc_initialized);
415 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
416 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
418 int index = INDEX_FROM_BYTES(bytes);
419 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
420 -> gcj_freelists + index;
421 ptr_t my_entry = *my_fl;
422 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
423 GC_PTR result = (GC_PTR)my_entry;
424 GC_ASSERT(!GC_incremental);
425 /* We assert that any concurrent marker will stop us. */
426 /* Thus it is impossible for a mark procedure to see the */
427 /* allocation of the next object, but to see this object */
428 /* still containing a free list pointer. Otherwise the */
429 /* marker might find a random "mark descriptor". */
430 *(volatile ptr_t *)my_fl = obj_link(my_entry);
431 /* We must update the freelist before we store the pointer. */
432 /* Otherwise a GC at this point would see a corrupted */
434 /* A memory barrier is probably never needed, since the */
435 /* action of stopping this thread will cause prior writes */
437 GC_ASSERT(((void * volatile *)result)[1] == 0);
438 *(void * volatile *)result = ptr_to_struct_containing_descr;
440 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
441 if (!GC_incremental) *my_fl = my_entry + index + 1;
442 /* In the incremental case, we always have to take this */
443 /* path. Thus we leave the counter alone. */
444 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
446 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
447 if (*my_fl == 0) return GC_oom_fn(bytes);
448 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
453 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
454 /* adjust it. The size is assumed to be such that it can be */
455 /* allocated as a small object. */
456 void * GC_local_gcj_fast_malloc(size_t lw, void * ptr_to_struct_containing_descr)
458 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
459 -> gcj_freelists + lw;
460 ptr_t my_entry = *my_fl;
462 GC_ASSERT(GC_gcj_malloc_initialized);
464 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
465 GC_PTR result = (GC_PTR)my_entry;
466 GC_ASSERT(!GC_incremental);
467 /* We assert that any concurrent marker will stop us. */
468 /* Thus it is impossible for a mark procedure to see the */
469 /* allocation of the next object, but to see this object */
470 /* still containing a free list pointer. Otherwise the */
471 /* marker might find a random "mark descriptor". */
472 *(volatile ptr_t *)my_fl = obj_link(my_entry);
473 /* We must update the freelist before we store the pointer. */
474 /* Otherwise a GC at this point would see a corrupted */
476 /* A memory barrier is probably never needed, since the */
477 /* action of stopping this thread will cause prior writes */
479 GC_ASSERT(((void * volatile *)result)[1] == 0);
480 *(void * volatile *)result = ptr_to_struct_containing_descr;
482 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
483 if (!GC_incremental) *my_fl = my_entry + lw + 1;
484 /* In the incremental case, we always have to take this */
485 /* path. Thus we leave the counter alone. */
486 return GC_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
488 GC_generic_malloc_many(BYTES_FROM_INDEX(lw), GC_gcj_kind, my_fl);
489 if (*my_fl == 0) return GC_oom_fn(BYTES_FROM_INDEX(lw));
490 return GC_local_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
494 #endif /* GC_GCJ_SUPPORT */
496 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
498 # define GC_destroy_thread_local(t)
500 # endif /* !THREAD_LOCAL_ALLOC */
504 To make sure that we're using LinuxThreads and not some other thread
505 package, we generate a dummy reference to `pthread_kill_other_threads_np'
506 (was `__pthread_initial_thread_bos' but that disappeared),
507 which is a symbol defined in LinuxThreads, but (hopefully) not in other
510 We no longer do this, since this code is now portable enough that it might
511 actually work for something else.
513 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
516 long GC_nprocs = 1; /* Number of processors. We may not have */
517 /* access to all of them, but this is as good */
518 /* a guess as any ... */
523 # define MAX_MARKERS 16
526 static ptr_t marker_sp[MAX_MARKERS] = {0};
528 void * GC_mark_thread(void * id)
532 marker_sp[(word)id] = GC_approx_sp();
533 for (;; ++my_mark_no) {
534 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
535 /* promptly. This is important if it were called from the signal */
536 /* handler or from the GC lock acquisition code. Under Linux, it's */
537 /* not safe to call it from a signal handler, since it uses mutexes */
538 /* and condition variables. Since it is called only here, the */
539 /* argument is unnecessary. */
540 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
541 /* resynchronize if we get far off, e.g. because GC_mark_no */
543 my_mark_no = GC_mark_no;
545 # ifdef DEBUG_THREADS
546 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
548 GC_help_marker(my_mark_no);
552 extern long GC_markers; /* Number of mark threads we would */
553 /* like to have. Includes the */
554 /* initiating thread. */
556 pthread_t GC_mark_threads[MAX_MARKERS];
558 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
560 static void start_mark_threads()
565 if (GC_markers > MAX_MARKERS) {
566 WARN("Limiting number of mark threads\n", 0);
567 GC_markers = MAX_MARKERS;
569 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
571 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
572 ABORT("pthread_attr_setdetachstate failed");
574 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
575 /* Default stack size is usually too small: fix it. */
576 /* Otherwise marker threads or GC may run out of */
578 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
583 if (pthread_attr_getstacksize(&attr, &old_size) != 0)
584 ABORT("pthread_attr_getstacksize failed\n");
585 if (old_size < MIN_STACK_SIZE) {
586 if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
587 ABORT("pthread_attr_setstacksize failed\n");
590 # endif /* HPUX || GC_DGUX386_THREADS */
592 if (GC_print_stats) {
593 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
596 for (i = 0; i < GC_markers - 1; ++i) {
597 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
598 GC_mark_thread, (void *)(word)i)) {
599 WARN("Marker thread creation failed, errno = %ld.\n", errno);
604 #else /* !PARALLEL_MARK */
606 static __inline__ void start_mark_threads()
610 #endif /* !PARALLEL_MARK */
612 GC_bool GC_thr_initialized = FALSE;
614 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
617 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
618 * put it into a separate function.
620 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
621 static __attribute__((noinline)) unsigned char* get_gc_thread_key_addr GC_PROTO((void))
623 return (unsigned char*)&GC_thread_key;
626 void GC_push_thread_structures GC_PROTO((void))
628 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
629 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
630 GC_push_all((ptr_t)get_gc_thread_key_addr(),
631 (ptr_t)(get_gc_thread_key_addr())+sizeof(&GC_thread_key));
637 void GC_push_thread_structures GC_PROTO((void))
639 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
640 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
641 GC_push_all((ptr_t)(&GC_thread_key),
642 (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
648 #ifdef THREAD_LOCAL_ALLOC
649 /* We must explicitly mark ptrfree and gcj free lists, since the free */
650 /* list links wouldn't otherwise be found. We also set them in the */
651 /* normal free lists, since that involves touching less memory than if */
652 /* we scanned them normally. */
653 void GC_mark_thread_local_free_lists(void)
659 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
660 for (p = GC_threads[i]; 0 != p; p = p -> next) {
661 for (j = 1; j < NFREELISTS; ++j) {
662 q = p -> ptrfree_freelists[j];
663 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
664 q = p -> normal_freelists[j];
665 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
666 # ifdef GC_GCJ_SUPPORT
667 q = p -> gcj_freelists[j];
668 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
669 # endif /* GC_GCJ_SUPPORT */
674 #endif /* THREAD_LOCAL_ALLOC */
676 static struct GC_Thread_Rep first_thread;
678 /* Add a thread to GC_threads. We assume it wasn't already there. */
679 /* Caller holds allocation lock. */
680 GC_thread GC_new_thread(pthread_t id)
682 int hv = ((word)id) % THREAD_TABLE_SZ;
684 static GC_bool first_thread_used = FALSE;
686 if (!first_thread_used) {
687 result = &first_thread;
688 first_thread_used = TRUE;
690 result = (struct GC_Thread_Rep *)
691 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
693 if (result == 0) return(0);
695 #ifdef PLATFORM_ANDROID
696 result -> kernel_id = gettid();
698 result -> next = GC_threads[hv];
699 GC_threads[hv] = result;
700 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
704 /* Delete a thread from GC_threads. We assume it is there. */
705 /* (The code intentionally traps if it wasn't.) */
706 /* Caller holds allocation lock. */
707 void GC_delete_thread(pthread_t id)
709 int hv = ((word)id) % THREAD_TABLE_SZ;
710 register GC_thread p = GC_threads[hv];
711 register GC_thread prev = 0;
713 while (!pthread_equal(p -> id, id)) {
718 GC_threads[hv] = p -> next;
720 prev -> next = p -> next;
722 #ifdef MONO_DEBUGGER_SUPPORTED
723 if (gc_thread_vtable && gc_thread_vtable->thread_exited)
724 gc_thread_vtable->thread_exited (id, &p->stop_info.stack_ptr);
727 #ifdef GC_DARWIN_THREADS
728 mach_port_deallocate(mach_task_self(), p->stop_info.mach_thread);
734 /* If a thread has been joined, but we have not yet */
735 /* been notified, then there may be more than one thread */
736 /* in the table with the same pthread id. */
737 /* This is OK, but we need a way to delete a specific one. */
738 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
740 int hv = ((word)id) % THREAD_TABLE_SZ;
741 register GC_thread p = GC_threads[hv];
742 register GC_thread prev = 0;
749 GC_threads[hv] = p -> next;
751 prev -> next = p -> next;
754 #ifdef GC_DARWIN_THREADS
755 mach_port_deallocate(mach_task_self(), p->stop_info.mach_thread);
761 /* Return a GC_thread corresponding to a given pthread_t. */
762 /* Returns 0 if it's not there. */
763 /* Caller holds allocation lock or otherwise inhibits */
765 /* If there is more than one thread with the given id we */
766 /* return the most recent one. */
767 GC_thread GC_lookup_thread(pthread_t id)
769 int hv = ((word)id) % THREAD_TABLE_SZ;
770 register GC_thread p = GC_threads[hv];
772 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
776 int GC_thread_is_registered (void)
781 ptr = (void *)GC_lookup_thread(pthread_self());
788 /* Remove all entries from the GC_threads table, except the */
789 /* one for the current thread. We need to do this in the child */
790 /* process after a fork(), since only the current thread */
791 /* survives in the child. */
792 void GC_remove_all_threads_but_me(void)
794 pthread_t self = pthread_self();
796 GC_thread p, next, me;
798 for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
800 for (p = GC_threads[hv]; 0 != p; p = next) {
802 if (p -> id == self) {
806 # ifdef THREAD_LOCAL_ALLOC
807 if (!(p -> flags & FINISHED)) {
808 GC_destroy_thread_local(p);
810 # endif /* THREAD_LOCAL_ALLOC */
811 if (p != &first_thread) GC_INTERNAL_FREE(p);
818 #endif /* HANDLE_FORK */
820 #ifdef USE_PROC_FOR_LIBRARIES
821 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
826 # ifdef PARALLEL_MARK
827 for (i = 0; i < GC_markers; ++i) {
828 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
831 for (i = 0; i < THREAD_TABLE_SZ; i++) {
832 for (p = GC_threads[i]; p != 0; p = p -> next) {
833 if (0 != p -> stack_end) {
834 # ifdef STACK_GROWS_UP
835 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
836 # else /* STACK_GROWS_DOWN */
837 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
844 #endif /* USE_PROC_FOR_LIBRARIES */
846 #ifdef GC_LINUX_THREADS
847 /* Return the number of processors, or i<= 0 if it can't be determined. */
850 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
851 /* appears to be buggy in many cases. */
852 /* We look for lines "cpu<n>" in /proc/stat. */
853 # define STAT_BUF_SIZE 4096
854 # define STAT_READ read
855 /* If read is wrapped, this may need to be redefined to call */
857 char stat_buf[STAT_BUF_SIZE];
860 /* Some old kernels only have a single "cpu nnnn ..." */
861 /* entry in /proc/stat. We identify those as */
865 f = open("/proc/stat", O_RDONLY);
866 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
867 WARN("Couldn't read /proc/stat\n", 0);
870 for (i = 0; i < len - 100; ++i) {
871 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
872 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
873 int cpu_no = atoi(stat_buf + i + 4);
874 if (cpu_no >= result) result = cpu_no + 1;
880 #endif /* GC_LINUX_THREADS */
882 /* We hold the GC lock. Wait until an in-progress GC has finished. */
883 /* Repeatedly RELEASES GC LOCK in order to wait. */
884 /* If wait_for_all is true, then we exit with the GC lock held and no */
885 /* collection in progress; otherwise we just wait for the current GC */
887 extern GC_bool GC_collection_in_progress();
888 void GC_wait_for_gc_completion(GC_bool wait_for_all)
890 if (GC_incremental && GC_collection_in_progress()) {
891 int old_gc_no = GC_gc_no;
893 /* Make sure that no part of our stack is still on the mark stack, */
894 /* since it's about to be unmapped. */
895 while (GC_incremental && GC_collection_in_progress()
896 && (wait_for_all || old_gc_no == GC_gc_no)) {
898 GC_in_thread_creation = TRUE;
899 GC_collect_a_little_inner(1);
900 GC_in_thread_creation = FALSE;
910 /* Procedures called before and after a fork. The goal here is to make */
911 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
912 /* attainable, since the single UNIX spec seems to imply that one */
913 /* should only call async-signal-safe functions, and we probably can't */
914 /* quite guarantee that. But we give it our best shot. (That same */
915 /* spec also implies that it's not safe to call the system malloc */
916 /* between fork() and exec(). Thus we're doing no worse than it. */
918 /* Called before a fork() */
919 void GC_fork_prepare_proc(void)
921 /* Acquire all relevant locks, so that after releasing the locks */
922 /* the child will see a consistent state in which monitor */
923 /* invariants hold. Unfortunately, we can't acquire libc locks */
924 /* we might need, and there seems to be no guarantee that libc */
925 /* must install a suitable fork handler. */
926 /* Wait for an ongoing GC to finish, since we can't finish it in */
927 /* the (one remaining thread in) the child. */
929 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
930 GC_wait_for_reclaim();
932 GC_wait_for_gc_completion(TRUE);
933 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
934 GC_acquire_mark_lock();
938 /* Called in parent after a fork() */
939 void GC_fork_parent_proc(void)
941 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
942 GC_release_mark_lock();
947 /* Called in child after a fork() */
948 void GC_fork_child_proc(void)
950 /* Clean up the thread table, so that just our thread is left. */
951 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
952 GC_release_mark_lock();
954 GC_remove_all_threads_but_me();
955 # ifdef PARALLEL_MARK
956 /* Turn off parallel marking in the child, since we are probably */
957 /* just going to exec, and we would have to restart mark threads. */
960 # endif /* PARALLEL_MARK */
963 #endif /* HANDLE_FORK */
965 #if defined(GC_DGUX386_THREADS)
966 /* Return the number of processors, or i<= 0 if it can't be determined. */
969 /* <takis@XFree86.Org> */
971 struct dg_sys_info_pm_info pm_sysinfo;
974 status = dg_sys_info((long int *) &pm_sysinfo,
975 DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
977 /* set -1 for error */
981 numCpus = pm_sysinfo.idle_vp_count;
983 # ifdef DEBUG_THREADS
984 GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
988 #endif /* GC_DGUX386_THREADS */
990 /* We hold the allocation lock. */
993 # ifndef GC_DARWIN_THREADS
998 if (GC_thr_initialized) return;
999 GC_thr_initialized = TRUE;
1002 /* Prepare for a possible fork. */
1003 pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
1004 GC_fork_child_proc);
1005 # endif /* HANDLE_FORK */
1006 /* Add the initial thread, so we can stop it. */
1007 t = GC_new_thread(pthread_self());
1008 # ifdef GC_DARWIN_THREADS
1009 t -> stop_info.mach_thread = mach_thread_self();
1011 t -> stop_info.stack_ptr = (ptr_t)(&dummy);
1013 t -> flags = DETACHED | MAIN_THREAD;
1014 #ifdef MONO_DEBUGGER_SUPPORTED
1015 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1016 # ifdef GC_DARWIN_THREADS
1017 gc_thread_vtable->thread_created (mach_thread_self (), &t->stop_info.stack_ptr);
1019 gc_thread_vtable->thread_created (pthread_self (), &t->stop_info.stack_ptr);
1025 /* Set GC_nprocs. */
1027 char * nprocs_string = GETENV("GC_NPROCS");
1029 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
1031 if (GC_nprocs <= 0) {
1032 # if defined(GC_HPUX_THREADS)
1033 GC_nprocs = pthread_num_processors_np();
1035 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
1036 GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
1037 if (GC_nprocs <= 0) GC_nprocs = 1;
1039 # if defined(GC_IRIX_THREADS)
1040 GC_nprocs = sysconf(_SC_NPROC_ONLN);
1041 if (GC_nprocs <= 0) GC_nprocs = 1;
1043 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS) || defined(GC_NETBSD_THREADS)
1045 size_t len = sizeof(ncpus);
1046 sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
1049 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1050 GC_nprocs = GC_get_nprocs();
1053 if (GC_nprocs <= 0) {
1054 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
1056 # ifdef PARALLEL_MARK
1060 # ifdef PARALLEL_MARK
1062 char * markers_string = GETENV("GC_MARKERS");
1063 if (markers_string != NULL) {
1064 GC_markers = atoi(markers_string);
1066 GC_markers = GC_nprocs;
1067 if (GC_markers > MAX_MARKERS)
1068 GC_markers = MAX_MARKERS;
1073 # ifdef PARALLEL_MARK
1075 if (GC_print_stats) {
1076 GC_printf2("Number of processors = %ld, "
1077 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
1080 if (GC_markers == 1) {
1081 GC_parallel = FALSE;
1083 if (GC_print_stats) {
1084 GC_printf0("Single marker thread, turning off parallel marking\n");
1089 /* Disable true incremental collection, but generational is OK. */
1090 GC_time_limit = GC_TIME_UNLIMITED;
1092 /* If we are using a parallel marker, actually start helper threads. */
1093 if (GC_parallel) start_mark_threads();
1098 /* Perform all initializations, including those that */
1099 /* may require allocation. */
1100 /* Called without allocation lock. */
1101 /* Must be called before a second thread is created. */
1102 /* Called without allocation lock. */
1103 void GC_init_parallel()
1105 if (parallel_initialized) return;
1106 parallel_initialized = TRUE;
1108 /* GC_init() calls us back, so set flag first. */
1109 if (!GC_is_initialized) GC_init();
1110 /* Initialize thread local free lists if used. */
1111 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1113 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1119 #if !defined(GC_DARWIN_THREADS)
1120 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
1122 sigset_t fudged_set;
1124 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
1126 sigdelset(&fudged_set, SIG_SUSPEND);
1129 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
1131 #endif /* !GC_DARWIN_THREADS */
1133 /* Wrappers for functions that are likely to block for an appreciable */
1134 /* length of time. Must be called in pairs, if at all. */
1135 /* Nothing much beyond the system call itself should be executed */
1136 /* between these. */
1138 void GC_start_blocking(void) {
1139 # define SP_SLOP 128
1142 me = GC_lookup_thread(pthread_self());
1143 GC_ASSERT(!(me -> thread_blocked));
1145 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
1147 # ifndef GC_DARWIN_THREADS
1148 me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
1152 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1154 /* Add some slop to the stack pointer, since the wrapped call may */
1155 /* end up pushing more callee-save registers. */
1156 # ifndef GC_DARWIN_THREADS
1157 # ifdef STACK_GROWS_UP
1158 me -> stop_info.stack_ptr += SP_SLOP;
1160 me -> stop_info.stack_ptr -= SP_SLOP;
1163 me -> thread_blocked = TRUE;
1167 void GC_end_blocking(void) {
1169 LOCK(); /* This will block if the world is stopped. */
1170 me = GC_lookup_thread(pthread_self());
1171 GC_ASSERT(me -> thread_blocked);
1172 me -> thread_blocked = FALSE;
1176 #if defined(GC_DGUX386_THREADS)
1177 #define __d10_sleep sleep
1178 #endif /* GC_DGUX386_THREADS */
1180 /* A wrapper for the standard C sleep function */
1181 int WRAP_FUNC(sleep) (unsigned int seconds)
1185 GC_start_blocking();
1186 result = REAL_FUNC(sleep)(seconds);
1192 void *(*start_routine)(void *);
1195 sem_t registered; /* 1 ==> in our thread table, but */
1196 /* parent hasn't yet noticed. */
1199 /* Called at thread exit. */
1200 /* Never called for main thread. That's OK, since it */
1201 /* results in at most a tiny one-time leak. And */
1202 /* linuxthreads doesn't reclaim the main threads */
1203 /* resources or id anyway. */
1204 void GC_thread_exit_proc(void *arg)
1209 me = GC_lookup_thread(pthread_self());
1210 GC_destroy_thread_local(me);
1211 if (me -> flags & DETACHED) {
1212 # ifdef THREAD_LOCAL_ALLOC
1213 /* NULL out the tls key to prevent the dtor function from being called */
1214 if (0 != GC_setspecific(GC_thread_key, NULL))
1215 ABORT("Failed to set thread specific allocation pointers");
1217 GC_delete_thread(pthread_self());
1219 me -> flags |= FINISHED;
1221 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1222 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1223 GC_remove_specific(GC_thread_key);
1225 /* The following may run the GC from "nonexistent" thread. */
1226 GC_wait_for_gc_completion(FALSE);
1230 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1233 GC_thread thread_gc_id;
1236 thread_gc_id = GC_lookup_thread(thread);
1237 /* This is guaranteed to be the intended one, since the thread id */
1238 /* cant have been recycled by pthreads. */
1240 result = REAL_FUNC(pthread_join)(thread, retval);
1241 # if defined (GC_FREEBSD_THREADS)
1242 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1243 appears to be) a spurious EINTR which caused the test and real code
1244 to gratuitously fail. Having looked at system pthread library source
1245 code, I see how this return code may be generated. In one path of
1246 code, pthread_join() just returns the errno setting of the thread
1247 being joined. This does not match the POSIX specification or the
1248 local man pages thus I have taken the liberty to catch this one
1249 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1250 if (result == EINTR) result = 0;
1254 /* Here the pthread thread id may have been recycled. */
1255 GC_delete_gc_thread(thread, thread_gc_id);
1262 WRAP_FUNC(pthread_detach)(pthread_t thread)
1265 GC_thread thread_gc_id;
1268 thread_gc_id = GC_lookup_thread(thread);
1270 result = REAL_FUNC(pthread_detach)(thread);
1273 thread_gc_id -> flags |= DETACHED;
1274 /* Here the pthread thread id may have been recycled. */
1275 if (thread_gc_id -> flags & FINISHED) {
1276 GC_delete_gc_thread(thread, thread_gc_id);
1283 GC_bool GC_in_thread_creation = FALSE;
1285 typedef void *(*ThreadStartFn)(void *);
1286 void * GC_start_routine_head(void * arg, void *base_addr,
1287 ThreadStartFn *start, void **start_arg )
1289 struct start_info * si = arg;
1292 pthread_t my_pthread;
1294 my_pthread = pthread_self();
1295 # ifdef DEBUG_THREADS
1296 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1297 GC_printf1("pid = %ld\n", (long) getpid());
1298 GC_printf1("sp = 0x%lx\n", (long) &arg);
1301 GC_in_thread_creation = TRUE;
1302 me = GC_new_thread(my_pthread);
1303 GC_in_thread_creation = FALSE;
1304 #ifdef GC_DARWIN_THREADS
1305 me -> stop_info.mach_thread = mach_thread_self();
1307 me -> stop_info.stack_ptr = 0;
1309 me -> flags = si -> flags;
1310 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1311 /* doesn't work because the stack base in /proc/self/stat is the */
1312 /* one for the main thread. There is a strong argument that that's */
1313 /* a kernel bug, but a pervasive one. */
1314 # ifdef STACK_GROWS_DOWN
1315 me -> stack_end = (ptr_t)(((word)(base_addr) + (GC_page_size - 1))
1316 & ~(GC_page_size - 1));
1317 # ifndef GC_DARWIN_THREADS
1318 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1320 /* Needs to be plausible, since an asynchronous stack mark */
1321 /* should not crash. */
1323 me -> stack_end = (ptr_t)((word)(base_addr) & ~(GC_page_size - 1));
1324 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1326 /* This is dubious, since we may be more than a page into the stack, */
1327 /* and hence skip some of it, though it's not clear that matters. */
1329 me -> backing_store_end = (ptr_t)
1330 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1331 /* This is also < 100% convincing. We should also read this */
1332 /* from /proc, but the hook to do so isn't there yet. */
1334 #ifdef MONO_DEBUGGER_SUPPORTED
1335 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1336 # ifdef GC_DARWIN_THREADS
1337 gc_thread_vtable->thread_created (mach_thread_self(), &me->stop_info.stack_ptr);
1339 gc_thread_vtable->thread_created (my_pthread, &me->stop_info.stack_ptr);
1344 if (start) *start = si -> start_routine;
1345 if (start_arg) *start_arg = si -> arg;
1347 sem_post(&(si -> registered)); /* Last action on si. */
1348 /* OK to deallocate. */
1349 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1351 GC_init_thread_local(me);
1358 int GC_thread_register_foreign (void *base_addr)
1360 struct start_info si = { 0, }; /* stacked for legibility & locking */
1363 # ifdef DEBUG_THREADS
1364 GC_printf1( "GC_thread_register_foreign %p\n", &si );
1367 si.flags = FOREIGN_THREAD;
1369 if (!parallel_initialized) GC_init_parallel();
1371 if (!GC_thr_initialized) GC_thr_init();
1375 me = GC_start_routine_head(&si, base_addr, NULL, NULL);
1380 void * GC_start_routine(void * arg)
1383 struct start_info * si = arg;
1386 ThreadStartFn start;
1389 me = GC_start_routine_head (arg, &dummy, &start, &start_arg);
1391 pthread_cleanup_push(GC_thread_exit_proc, 0);
1392 # ifdef DEBUG_THREADS
1393 GC_printf1("start_routine = 0x%lx\n", start);
1395 result = (*start)(start_arg);
1397 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1399 me -> status = result;
1400 pthread_cleanup_pop(1);
1401 /* Cleanup acquires lock, ensuring that we can't exit */
1402 /* while a collection that thinks we're alive is trying to stop */
1408 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1409 const pthread_attr_t *attr,
1410 void *(*start_routine)(void *), void *arg)
1415 struct start_info * si;
1416 /* This is otherwise saved only in an area mmapped by the thread */
1417 /* library, which isn't visible to the collector. */
1419 /* We resist the temptation to muck with the stack size here, */
1420 /* even if the default is unreasonably small. That's the client's */
1421 /* responsibility. */
1424 si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
1427 if (!parallel_initialized) GC_init_parallel();
1428 if (0 == si) return(ENOMEM);
1429 sem_init(&(si -> registered), 0, 0);
1430 si -> start_routine = start_routine;
1433 if (!GC_thr_initialized) GC_thr_init();
1434 # ifdef GC_ASSERTIONS
1438 pthread_attr_t my_attr;
1439 pthread_attr_init(&my_attr);
1440 pthread_attr_getstacksize(&my_attr, &stack_size);
1442 pthread_attr_getstacksize(attr, &stack_size);
1444 # ifdef PARALLEL_MARK
1445 GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1447 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1448 /* HBLKSIZE=8192, sizeof(word)=8 */
1449 GC_ASSERT(stack_size >= 65536);
1451 /* Our threads may need to do some work for the GC. */
1452 /* Ridiculously small threads won't work, and they */
1453 /* probably wouldn't work anyway. */
1457 detachstate = PTHREAD_CREATE_JOINABLE;
1459 pthread_attr_getdetachstate(attr, &detachstate);
1461 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1462 si -> flags = my_flags;
1464 # ifdef DEBUG_THREADS
1465 GC_printf1("About to start new thread from thread 0x%X\n",
1469 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1471 # ifdef DEBUG_THREADS
1472 GC_printf1("Started thread 0x%X\n", *new_thread);
1474 /* Wait until child has been added to the thread table. */
1475 /* This also ensures that we hold onto si until the child is done */
1476 /* with it. Thus it doesn't matter whether it is otherwise */
1477 /* visible to the collector. */
1479 while (0 != sem_wait(&(si -> registered))) {
1480 if (EINTR != errno) ABORT("sem_wait failed");
1483 sem_destroy(&(si -> registered));
1485 GC_INTERNAL_FREE(si);
1491 #ifdef GENERIC_COMPARE_AND_SWAP
1492 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1494 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1495 GC_word old, GC_word new_val)
1498 pthread_mutex_lock(&GC_compare_and_swap_lock);
1505 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1509 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1512 pthread_mutex_lock(&GC_compare_and_swap_lock);
1514 *addr = old + how_much;
1515 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1519 #endif /* GENERIC_COMPARE_AND_SWAP */
1520 /* Spend a few cycles in a way that can't introduce contention with */
1521 /* othre threads. */
1525 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1526 volatile word dummy = 0;
1529 for (i = 0; i < 10; ++i) {
1530 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1531 __asm__ __volatile__ (" " : : : "memory");
1533 /* Something that's unlikely to be optimized away. */
1539 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1542 VOLATILE GC_bool GC_collecting = 0;
1543 /* A hint that we're in the collector and */
1544 /* holding the allocation lock for an */
1545 /* extended period. */
1547 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1548 /* If we don't want to use the below spinlock implementation, either */
1549 /* because we don't have a GC_test_and_set implementation, or because */
1550 /* we don't want to risk sleeping, we can still try spinning on */
1551 /* pthread_mutex_trylock for a while. This appears to be very */
1552 /* beneficial in many cases. */
1553 /* I suspect that under high contention this is nearly always better */
1554 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1555 /* Hence we still default to the spin lock. */
1556 /* This is also used to acquire the mark lock for the parallel */
1559 /* Here we use a strict exponential backoff scheme. I don't know */
1560 /* whether that's better or worse than the above. We eventually */
1561 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1562 /* explicitly sleep. */
1566 unsigned long GC_spin_count = 0;
1567 unsigned long GC_block_count = 0;
1568 unsigned long GC_unlocked_count = 0;
1571 void GC_generic_lock(pthread_mutex_t * lock)
1573 #ifndef NO_PTHREAD_TRYLOCK
1574 unsigned pause_length = 1;
1577 if (0 == pthread_mutex_trylock(lock)) {
1579 ++GC_unlocked_count;
1583 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1584 for (i = 0; i < pause_length; ++i) {
1587 switch(pthread_mutex_trylock(lock)) {
1596 ABORT("Unexpected error from pthread_mutex_trylock");
1599 #endif /* !NO_PTHREAD_TRYLOCK */
1603 pthread_mutex_lock(lock);
1606 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1608 #if defined(USE_SPIN_LOCK)
1610 /* Reasonably fast spin locks. Basically the same implementation */
1611 /* as STL alloc.h. This isn't really the right way to do this. */
1612 /* but until the POSIX scheduling mess gets straightened out ... */
1614 volatile unsigned int GC_allocate_lock = 0;
1619 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1620 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1621 static unsigned spin_max = low_spin_max;
1622 unsigned my_spin_max;
1623 static unsigned last_spins = 0;
1624 unsigned my_last_spins;
1627 if (!GC_test_and_set(&GC_allocate_lock)) {
1630 my_spin_max = spin_max;
1631 my_last_spins = last_spins;
1632 for (i = 0; i < my_spin_max; i++) {
1633 if (GC_collecting || GC_nprocs == 1) goto yield;
1634 if (i < my_last_spins/2 || GC_allocate_lock) {
1638 if (!GC_test_and_set(&GC_allocate_lock)) {
1641 * Spinning worked. Thus we're probably not being scheduled
1642 * against the other process with which we were contending.
1643 * Thus it makes sense to spin longer the next time.
1646 spin_max = high_spin_max;
1650 /* We are probably being scheduled against the other process. Sleep. */
1651 spin_max = low_spin_max;
1654 if (!GC_test_and_set(&GC_allocate_lock)) {
1657 # define SLEEP_THRESHOLD 12
1658 /* Under Linux very short sleeps tend to wait until */
1659 /* the current time quantum expires. On old Linux */
1660 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1661 /* (Under 2.4, this happens only for real-time */
1662 /* processes.) We want to minimize both behaviors */
1664 if (i < SLEEP_THRESHOLD) {
1670 /* Don't wait for more than about 15msecs, even */
1671 /* under extreme contention. */
1673 ts.tv_nsec = 1 << i;
1679 #else /* !USE_SPINLOCK */
1682 #ifndef NO_PTHREAD_TRYLOCK
1683 if (1 == GC_nprocs || GC_collecting) {
1684 pthread_mutex_lock(&GC_allocate_ml);
1686 GC_generic_lock(&GC_allocate_ml);
1688 #else /* !NO_PTHREAD_TRYLOCK */
1689 pthread_mutex_lock(&GC_allocate_ml);
1690 #endif /* !NO_PTHREAD_TRYLOCK */
1693 #endif /* !USE_SPINLOCK */
1695 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1697 #ifdef GC_ASSERTIONS
1698 pthread_t GC_mark_lock_holder = NO_THREAD;
1702 /* Ugly workaround for a linux threads bug in the final versions */
1703 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1704 /* field even when it fails to acquire the mutex. This causes */
1705 /* pthread_cond_wait to die. Remove for glibc2.2. */
1706 /* According to the man page, we should use */
1707 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1709 static pthread_mutex_t mark_mutex =
1710 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1712 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1715 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1717 void GC_acquire_mark_lock()
1720 if (pthread_mutex_lock(&mark_mutex) != 0) {
1721 ABORT("pthread_mutex_lock failed");
1724 GC_generic_lock(&mark_mutex);
1725 # ifdef GC_ASSERTIONS
1726 GC_mark_lock_holder = pthread_self();
1730 void GC_release_mark_lock()
1732 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1733 # ifdef GC_ASSERTIONS
1734 GC_mark_lock_holder = NO_THREAD;
1736 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1737 ABORT("pthread_mutex_unlock failed");
1741 /* Collector must wait for a freelist builders for 2 reasons: */
1742 /* 1) Mark bits may still be getting examined without lock. */
1743 /* 2) Partial free lists referenced only by locals may not be scanned */
1744 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1745 /* free-list link may be ignored. */
1746 void GC_wait_builder()
1748 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1749 # ifdef GC_ASSERTIONS
1750 GC_mark_lock_holder = NO_THREAD;
1752 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1753 ABORT("pthread_cond_wait failed");
1755 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1756 # ifdef GC_ASSERTIONS
1757 GC_mark_lock_holder = pthread_self();
1761 void GC_wait_for_reclaim()
1763 GC_acquire_mark_lock();
1764 while (GC_fl_builder_count > 0) {
1767 GC_release_mark_lock();
1770 void GC_notify_all_builder()
1772 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1773 if (pthread_cond_broadcast(&builder_cv) != 0) {
1774 ABORT("pthread_cond_broadcast failed");
1778 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1780 #ifdef PARALLEL_MARK
1782 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1784 void GC_wait_marker()
1786 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1787 # ifdef GC_ASSERTIONS
1788 GC_mark_lock_holder = NO_THREAD;
1790 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1791 ABORT("pthread_cond_wait failed");
1793 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1794 # ifdef GC_ASSERTIONS
1795 GC_mark_lock_holder = pthread_self();
1799 void GC_notify_all_marker()
1801 if (pthread_cond_broadcast(&mark_cv) != 0) {
1802 ABORT("pthread_cond_broadcast failed");
1806 #endif /* PARALLEL_MARK */
1808 # endif /* GC_LINUX_THREADS and friends */