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 defined(GC_FREEBSD_THREADS) && !defined(USE_PTHREAD_SPECIFIC) || \
73 defined(GC_OPENBSD_THREADS)
74 # define USE_PTHREAD_SPECIFIC
77 # if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
78 # define _POSIX4A_DRAFT10_SOURCE 1
81 # if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
82 # define _USING_POSIX4A_DRAFT10 1
85 # ifdef THREAD_LOCAL_ALLOC
86 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
87 # include "private/specific.h"
90 /* Note that these macros should be used only to get/set the GC_thread pointer.
91 * We need to use both tls and pthread because we use the pthread_create function hook to
92 * free the data for foreign threads. When that doesn't happen, libgc could have old
93 * pthread_t that get reused...
95 # if defined(USE_PTHREAD_SPECIFIC)
96 # define GC_getspecific pthread_getspecific
97 # define GC_setspecific pthread_setspecific
98 # define GC_key_create pthread_key_create
99 typedef pthread_key_t GC_key_t;
101 # if defined(USE_COMPILER_TLS)
102 /* Note sles9 gcc on powerpc gets confused by the define to set GC_thread_tls and pthread_setspecific
103 * so we actually use a static inline function decalred below that is equivalent to:
104 * define GC_setspecific(key, v) (GC_thread_tls = (v), pthread_setspecific ((key), (v)))
106 # define GC_getspecific(x) (GC_thread_tls)
107 # define GC_key_create pthread_key_create
108 typedef pthread_key_t GC_key_t;
112 # include <pthread.h>
117 # include <sys/mman.h>
118 # include <sys/time.h>
119 # include <sys/types.h>
120 # include <sys/stat.h>
124 #if defined(GC_DARWIN_THREADS)
125 # include "private/darwin_semaphore.h"
127 # include <semaphore.h>
128 #endif /* !GC_DARWIN_THREADS */
130 #if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
131 # include <sys/sysctl.h>
132 #endif /* GC_DARWIN_THREADS */
134 #if defined(GC_NETBSD_THREADS) || defined(GC_OPENBSD_THREADS)
135 # include <sys/param.h>
136 # include <sys/sysctl.h>
141 #if defined(GC_DGUX386_THREADS)
142 # include <sys/dg_sys_info.h>
143 # include <sys/_int_psem.h>
144 /* sem_t is an uint in DG/UX */
145 typedef unsigned int sem_t;
146 #endif /* GC_DGUX386_THREADS */
152 #ifdef GC_USE_LD_WRAP
153 # define WRAP_FUNC(f) __wrap_##f
154 # define REAL_FUNC(f) __real_##f
156 # define WRAP_FUNC(f) GC_##f
157 # if !defined(GC_DGUX386_THREADS)
158 # define REAL_FUNC(f) f
159 # else /* GC_DGUX386_THREADS */
160 # define REAL_FUNC(f) __d10_##f
161 # endif /* GC_DGUX386_THREADS */
162 # undef pthread_create
163 # if !defined(GC_DARWIN_THREADS)
164 # undef pthread_sigmask
167 # undef pthread_detach
171 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
172 && !defined(_PTHREAD_USE_PTDNAM_)
173 /* Restore the original mangled names on Tru64 UNIX. */
174 # define pthread_create __pthread_create
175 # define pthread_join __pthread_join
176 # define pthread_detach __pthread_detach
182 static GC_bool parallel_initialized = FALSE;
184 void GC_init_parallel();
186 static pthread_t main_pthread_self;
187 static void *main_stack, *main_altstack;
188 static int main_stack_size, main_altstack_size;
190 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
192 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
194 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
195 #undef PACKAGE_BUGREPORT
197 #undef PACKAGE_STRING
198 #undef PACKAGE_TARNAME
199 #undef PACKAGE_VERSION
200 #include "mono/utils/mono-compiler.h"
203 GC_key_t GC_thread_key;
205 #ifdef USE_COMPILER_TLS
206 __thread MONO_TLS_FAST void* GC_thread_tls;
209 * gcc errors out with /tmp/ccdPMFuq.s:2994: Error: symbol `.LTLS4' is already defined
210 * if the inline is added on powerpc
212 #if !defined(__ppc__) && !defined(__powerpc__)
215 static int GC_setspecific (GC_key_t key, void *value) {
216 GC_thread_tls = value;
217 return pthread_setspecific (key, value);
221 static GC_bool keys_initialized;
223 #ifdef MONO_DEBUGGER_SUPPORTED
224 #include "include/libgc-mono-debugger.h"
227 /* Recover the contents of the freelist array fl into the global one gfl.*/
228 /* Note that the indexing scheme differs, in that gfl has finer size */
229 /* resolution, even if not all entries are used. */
230 /* We hold the allocator lock. */
231 static void return_freelists(ptr_t *fl, ptr_t *gfl)
237 for (i = 1; i < NFREELISTS; ++i) {
238 nwords = i * (GRANULARITY/sizeof(word));
241 if ((word)q >= HBLKSIZE) {
242 if (gfl[nwords] == 0) {
246 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
252 /* Clear fl[i], since the thread structure may hang around. */
253 /* Do it in a way that is likely to trap if we access it. */
254 fl[i] = (ptr_t)HBLKSIZE;
258 /* We statically allocate a single "size 0" object. It is linked to */
259 /* itself, and is thus repeatedly reused for all size 0 allocation */
260 /* requests. (Size 0 gcj allocation requests are incorrect, and */
261 /* we arrange for those to fault asap.) */
262 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
264 void GC_delete_gc_thread(pthread_t id, GC_thread gct);
265 void GC_destroy_thread_local(GC_thread p);
267 void GC_thread_deregister_foreign (void *data)
269 GC_thread me = (GC_thread)data;
270 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
271 if (me -> flags & FOREIGN_THREAD) {
273 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
274 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
275 GC_destroy_thread_local (me);
277 GC_delete_gc_thread(me->id, me);
282 /* Each thread structure must be initialized. */
283 /* This call must be made from the new thread. */
284 /* Caller holds allocation lock. */
285 void GC_init_thread_local(GC_thread p)
289 if (!keys_initialized) {
290 if (0 != GC_key_create(&GC_thread_key, GC_thread_deregister_foreign)) {
291 ABORT("Failed to create key for local allocator");
293 keys_initialized = TRUE;
295 if (0 != GC_setspecific(GC_thread_key, p)) {
296 ABORT("Failed to set thread specific allocation pointers");
298 for (i = 1; i < NFREELISTS; ++i) {
299 p -> ptrfree_freelists[i] = (ptr_t)1;
300 p -> normal_freelists[i] = (ptr_t)1;
301 # ifdef GC_GCJ_SUPPORT
302 p -> gcj_freelists[i] = (ptr_t)1;
305 /* Set up the size 0 free lists. */
306 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
307 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
308 # ifdef GC_GCJ_SUPPORT
309 p -> gcj_freelists[0] = (ptr_t)(-1);
313 #ifdef GC_GCJ_SUPPORT
314 extern ptr_t * GC_gcjobjfreelist;
317 /* We hold the allocator lock. */
318 void GC_destroy_thread_local(GC_thread p)
320 /* We currently only do this from the thread itself or from */
321 /* the fork handler for a child process. */
323 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
325 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
326 return_freelists(p -> normal_freelists, GC_objfreelist);
327 # ifdef GC_GCJ_SUPPORT
328 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
332 extern GC_PTR GC_generic_malloc_many();
334 GC_PTR GC_local_malloc(size_t bytes)
336 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
337 return(GC_malloc(bytes));
339 int index = INDEX_FROM_BYTES(bytes);
342 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
343 GC_key_t k = GC_thread_key;
347 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
348 if (EXPECT(0 == k, 0)) {
349 /* This can happen if we get called when the world is */
350 /* being initialized. Whether we can actually complete */
351 /* the initialization then is unclear. */
356 tsd = GC_getspecific(GC_thread_key);
357 # ifdef GC_ASSERTIONS
359 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
362 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
364 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
365 ptr_t next = obj_link(my_entry);
366 GC_PTR result = (GC_PTR)my_entry;
368 obj_link(my_entry) = 0;
369 PREFETCH_FOR_WRITE(next);
371 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
372 *my_fl = my_entry + index + 1;
373 return GC_malloc(bytes);
375 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
376 if (*my_fl == 0) return GC_oom_fn(bytes);
377 return GC_local_malloc(bytes);
382 GC_PTR GC_local_malloc_atomic(size_t bytes)
384 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
385 return(GC_malloc_atomic(bytes));
387 int index = INDEX_FROM_BYTES(bytes);
388 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
389 -> ptrfree_freelists + index;
390 ptr_t my_entry = *my_fl;
392 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
393 GC_PTR result = (GC_PTR)my_entry;
394 *my_fl = obj_link(my_entry);
396 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
397 *my_fl = my_entry + index + 1;
398 return GC_malloc_atomic(bytes);
400 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
401 /* *my_fl is updated while the collector is excluded; */
402 /* the free list is always visible to the collector as */
404 if (*my_fl == 0) return GC_oom_fn(bytes);
405 return GC_local_malloc_atomic(bytes);
410 #ifdef GC_GCJ_SUPPORT
412 #include "include/gc_gcj.h"
415 extern GC_bool GC_gcj_malloc_initialized;
418 extern int GC_gcj_kind;
420 GC_PTR GC_local_gcj_malloc(size_t bytes,
421 void * ptr_to_struct_containing_descr)
423 GC_ASSERT(GC_gcj_malloc_initialized);
424 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
425 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
427 int index = INDEX_FROM_BYTES(bytes);
428 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
429 -> gcj_freelists + index;
430 ptr_t my_entry = *my_fl;
431 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
432 GC_PTR result = (GC_PTR)my_entry;
433 GC_ASSERT(!GC_incremental);
434 /* We assert that any concurrent marker will stop us. */
435 /* Thus it is impossible for a mark procedure to see the */
436 /* allocation of the next object, but to see this object */
437 /* still containing a free list pointer. Otherwise the */
438 /* marker might find a random "mark descriptor". */
439 *(volatile ptr_t *)my_fl = obj_link(my_entry);
440 /* We must update the freelist before we store the pointer. */
441 /* Otherwise a GC at this point would see a corrupted */
443 /* A memory barrier is probably never needed, since the */
444 /* action of stopping this thread will cause prior writes */
446 GC_ASSERT(((void * volatile *)result)[1] == 0);
447 *(void * volatile *)result = ptr_to_struct_containing_descr;
449 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
450 if (!GC_incremental) *my_fl = my_entry + index + 1;
451 /* In the incremental case, we always have to take this */
452 /* path. Thus we leave the counter alone. */
453 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
455 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
456 if (*my_fl == 0) return GC_oom_fn(bytes);
457 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
462 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
463 /* adjust it. The size is assumed to be such that it can be */
464 /* allocated as a small object. */
465 void * GC_local_gcj_fast_malloc(size_t lw, void * ptr_to_struct_containing_descr)
467 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
468 -> gcj_freelists + lw;
469 ptr_t my_entry = *my_fl;
471 GC_ASSERT(GC_gcj_malloc_initialized);
473 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
474 GC_PTR result = (GC_PTR)my_entry;
475 GC_ASSERT(!GC_incremental);
476 /* We assert that any concurrent marker will stop us. */
477 /* Thus it is impossible for a mark procedure to see the */
478 /* allocation of the next object, but to see this object */
479 /* still containing a free list pointer. Otherwise the */
480 /* marker might find a random "mark descriptor". */
481 *(volatile ptr_t *)my_fl = obj_link(my_entry);
482 /* We must update the freelist before we store the pointer. */
483 /* Otherwise a GC at this point would see a corrupted */
485 /* A memory barrier is probably never needed, since the */
486 /* action of stopping this thread will cause prior writes */
488 GC_ASSERT(((void * volatile *)result)[1] == 0);
489 *(void * volatile *)result = ptr_to_struct_containing_descr;
491 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
492 if (!GC_incremental) *my_fl = my_entry + lw + 1;
493 /* In the incremental case, we always have to take this */
494 /* path. Thus we leave the counter alone. */
495 return GC_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
497 GC_generic_malloc_many(BYTES_FROM_INDEX(lw), GC_gcj_kind, my_fl);
498 if (*my_fl == 0) return GC_oom_fn(BYTES_FROM_INDEX(lw));
499 return GC_local_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
503 #endif /* GC_GCJ_SUPPORT */
505 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
507 # define GC_destroy_thread_local(t)
509 # endif /* !THREAD_LOCAL_ALLOC */
513 To make sure that we're using LinuxThreads and not some other thread
514 package, we generate a dummy reference to `pthread_kill_other_threads_np'
515 (was `__pthread_initial_thread_bos' but that disappeared),
516 which is a symbol defined in LinuxThreads, but (hopefully) not in other
519 We no longer do this, since this code is now portable enough that it might
520 actually work for something else.
522 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
525 long GC_nprocs = 1; /* Number of processors. We may not have */
526 /* access to all of them, but this is as good */
527 /* a guess as any ... */
532 # define MAX_MARKERS 16
535 static ptr_t marker_sp[MAX_MARKERS] = {0};
537 void * GC_mark_thread(void * id)
541 marker_sp[(word)id] = GC_approx_sp();
542 for (;; ++my_mark_no) {
543 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
544 /* promptly. This is important if it were called from the signal */
545 /* handler or from the GC lock acquisition code. Under Linux, it's */
546 /* not safe to call it from a signal handler, since it uses mutexes */
547 /* and condition variables. Since it is called only here, the */
548 /* argument is unnecessary. */
549 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
550 /* resynchronize if we get far off, e.g. because GC_mark_no */
552 my_mark_no = GC_mark_no;
554 # ifdef DEBUG_THREADS
555 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
557 GC_help_marker(my_mark_no);
561 extern long GC_markers; /* Number of mark threads we would */
562 /* like to have. Includes the */
563 /* initiating thread. */
565 pthread_t GC_mark_threads[MAX_MARKERS];
567 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
569 static void start_mark_threads()
574 if (GC_markers > MAX_MARKERS) {
575 WARN("Limiting number of mark threads\n", 0);
576 GC_markers = MAX_MARKERS;
578 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
580 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
581 ABORT("pthread_attr_setdetachstate failed");
583 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
584 /* Default stack size is usually too small: fix it. */
585 /* Otherwise marker threads or GC may run out of */
587 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
592 if (pthread_attr_getstacksize(&attr, &old_size) != 0)
593 ABORT("pthread_attr_getstacksize failed\n");
594 if (old_size < MIN_STACK_SIZE) {
595 if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
596 ABORT("pthread_attr_setstacksize failed\n");
599 # endif /* HPUX || GC_DGUX386_THREADS */
601 if (GC_print_stats) {
602 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
605 for (i = 0; i < GC_markers - 1; ++i) {
606 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
607 GC_mark_thread, (void *)(word)i)) {
608 WARN("Marker thread creation failed, errno = %ld.\n", errno);
613 #else /* !PARALLEL_MARK */
615 static __inline__ void start_mark_threads()
619 #endif /* !PARALLEL_MARK */
621 GC_bool GC_thr_initialized = FALSE;
623 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
626 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
627 * put it into a separate function.
629 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
630 static __attribute__((noinline)) unsigned char* get_gc_thread_key_addr GC_PROTO((void))
632 return (unsigned char*)&GC_thread_key;
635 void GC_push_thread_structures GC_PROTO((void))
637 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
638 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
639 GC_push_all((ptr_t)get_gc_thread_key_addr(),
640 (ptr_t)(get_gc_thread_key_addr())+sizeof(&GC_thread_key));
646 void GC_push_thread_structures GC_PROTO((void))
648 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
649 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
650 GC_push_all((ptr_t)(&GC_thread_key),
651 (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
657 #ifdef THREAD_LOCAL_ALLOC
658 /* We must explicitly mark ptrfree and gcj free lists, since the free */
659 /* list links wouldn't otherwise be found. We also set them in the */
660 /* normal free lists, since that involves touching less memory than if */
661 /* we scanned them normally. */
662 void GC_mark_thread_local_free_lists(void)
668 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
669 for (p = GC_threads[i]; 0 != p; p = p -> next) {
670 for (j = 1; j < NFREELISTS; ++j) {
671 q = p -> ptrfree_freelists[j];
672 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
673 q = p -> normal_freelists[j];
674 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
675 # ifdef GC_GCJ_SUPPORT
676 q = p -> gcj_freelists[j];
677 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
678 # endif /* GC_GCJ_SUPPORT */
683 #endif /* THREAD_LOCAL_ALLOC */
685 static struct GC_Thread_Rep first_thread;
688 extern int nacl_thread_parked[MAX_NACL_GC_THREADS];
689 extern int nacl_thread_used[MAX_NACL_GC_THREADS];
690 extern int nacl_thread_parking_inited;
691 extern int nacl_num_gc_threads;
692 extern pthread_mutex_t nacl_thread_alloc_lock;
693 extern __thread int nacl_thread_idx;
694 extern __thread GC_thread nacl_gc_thread_self;
696 extern void nacl_pre_syscall_hook();
697 extern void nacl_post_syscall_hook();
698 extern void nacl_register_gc_hooks(void (*pre)(), void (*post)());
700 void nacl_initialize_gc_thread()
703 nacl_register_gc_hooks(nacl_pre_syscall_hook, nacl_post_syscall_hook);
704 pthread_mutex_lock(&nacl_thread_alloc_lock);
705 if (!nacl_thread_parking_inited)
707 for (i = 0; i < MAX_NACL_GC_THREADS; i++) {
708 nacl_thread_used[i] = 0;
709 nacl_thread_parked[i] = 0;
711 nacl_thread_parking_inited = 1;
713 GC_ASSERT(nacl_num_gc_threads <= MAX_NACL_GC_THREADS);
714 for (i = 0; i < MAX_NACL_GC_THREADS; i++) {
715 if (nacl_thread_used[i] == 0) {
716 nacl_thread_used[i] = 1;
718 nacl_num_gc_threads++;
722 pthread_mutex_unlock(&nacl_thread_alloc_lock);
725 void nacl_shutdown_gc_thread()
727 pthread_mutex_lock(&nacl_thread_alloc_lock);
728 GC_ASSERT(nacl_thread_idx >= 0 && nacl_thread_idx < MAX_NACL_GC_THREADS);
729 GC_ASSERT(nacl_thread_used[nacl_thread_idx] != 0);
730 nacl_thread_used[nacl_thread_idx] = 0;
731 nacl_thread_idx = -1;
732 nacl_num_gc_threads--;
733 pthread_mutex_unlock(&nacl_thread_alloc_lock);
738 /* Add a thread to GC_threads. We assume it wasn't already there. */
739 /* Caller holds allocation lock. */
740 GC_thread GC_new_thread(pthread_t id)
742 int hv = ((unsigned long)id) % THREAD_TABLE_SZ;
744 static GC_bool first_thread_used = FALSE;
746 if (!first_thread_used) {
747 result = &first_thread;
748 first_thread_used = TRUE;
750 result = (struct GC_Thread_Rep *)
751 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
753 if (result == 0) return(0);
755 #ifdef PLATFORM_ANDROID
756 result -> kernel_id = gettid();
758 result -> next = GC_threads[hv];
759 GC_threads[hv] = result;
761 nacl_gc_thread_self = result;
762 nacl_initialize_gc_thread();
764 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
768 /* Delete a thread from GC_threads. We assume it is there. */
769 /* (The code intentionally traps if it wasn't.) */
770 /* Caller holds allocation lock. */
771 void GC_delete_thread(pthread_t id)
773 int hv = ((unsigned long)id) % THREAD_TABLE_SZ;
774 register GC_thread p = GC_threads[hv];
775 register GC_thread prev = 0;
778 nacl_shutdown_gc_thread();
779 nacl_gc_thread_self = NULL;
782 while (!pthread_equal(p -> id, id)) {
787 GC_threads[hv] = p -> next;
789 prev -> next = p -> next;
791 #ifdef MONO_DEBUGGER_SUPPORTED
792 if (gc_thread_vtable && gc_thread_vtable->thread_exited)
793 gc_thread_vtable->thread_exited (id, &p->stop_info.stack_ptr);
796 #ifdef GC_DARWIN_THREADS
797 mach_port_deallocate(mach_task_self(), p->stop_info.mach_thread);
803 /* If a thread has been joined, but we have not yet */
804 /* been notified, then there may be more than one thread */
805 /* in the table with the same pthread id. */
806 /* This is OK, but we need a way to delete a specific one. */
807 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
809 int hv = ((unsigned long)id) % THREAD_TABLE_SZ;
810 register GC_thread p = GC_threads[hv];
811 register GC_thread prev = 0;
818 GC_threads[hv] = p -> next;
820 prev -> next = p -> next;
823 #ifdef GC_DARWIN_THREADS
824 mach_port_deallocate(mach_task_self(), p->stop_info.mach_thread);
830 /* Return a GC_thread corresponding to a given pthread_t. */
831 /* Returns 0 if it's not there. */
832 /* Caller holds allocation lock or otherwise inhibits */
834 /* If there is more than one thread with the given id we */
835 /* return the most recent one. */
836 GC_thread GC_lookup_thread(pthread_t id)
838 int hv = ((unsigned long)id) % THREAD_TABLE_SZ;
839 register GC_thread p = GC_threads[hv];
841 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
845 int GC_thread_is_registered (void)
850 ptr = (void *)GC_lookup_thread(pthread_self());
856 void GC_register_altstack (void *stack, int stack_size, void *altstack, int altstack_size)
861 thread = (void *)GC_lookup_thread(pthread_self());
863 thread->stack = stack;
864 thread->stack_size = stack_size;
865 thread->altstack = altstack;
866 thread->altstack_size = altstack_size;
869 * This happens if we are called before GC_thr_init ().
871 main_pthread_self = pthread_self ();
873 main_stack_size = stack_size;
874 main_altstack = altstack;
875 main_altstack_size = altstack_size;
881 /* Remove all entries from the GC_threads table, except the */
882 /* one for the current thread. We need to do this in the child */
883 /* process after a fork(), since only the current thread */
884 /* survives in the child. */
885 void GC_remove_all_threads_but_me(void)
887 pthread_t self = pthread_self();
889 GC_thread p, next, me;
891 for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
893 for (p = GC_threads[hv]; 0 != p; p = next) {
895 if (p -> id == self) {
899 # ifdef THREAD_LOCAL_ALLOC
900 if (!(p -> flags & FINISHED)) {
901 GC_destroy_thread_local(p);
903 # endif /* THREAD_LOCAL_ALLOC */
904 if (p != &first_thread) GC_INTERNAL_FREE(p);
911 #endif /* HANDLE_FORK */
913 #ifdef USE_PROC_FOR_LIBRARIES
914 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
919 # ifdef PARALLEL_MARK
920 for (i = 0; i < GC_markers; ++i) {
921 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
924 for (i = 0; i < THREAD_TABLE_SZ; i++) {
925 for (p = GC_threads[i]; p != 0; p = p -> next) {
926 if (0 != p -> stack_end) {
927 # ifdef STACK_GROWS_UP
928 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
929 # else /* STACK_GROWS_DOWN */
930 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
937 #endif /* USE_PROC_FOR_LIBRARIES */
939 #ifdef GC_LINUX_THREADS
940 /* Return the number of processors, or i<= 0 if it can't be determined. */
943 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
944 /* appears to be buggy in many cases. */
945 /* We look for lines "cpu<n>" in /proc/stat. */
946 # define STAT_BUF_SIZE 4096
947 # define STAT_READ read
948 /* If read is wrapped, this may need to be redefined to call */
950 char stat_buf[STAT_BUF_SIZE];
953 /* Some old kernels only have a single "cpu nnnn ..." */
954 /* entry in /proc/stat. We identify those as */
958 f = open("/proc/stat", O_RDONLY);
959 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
960 WARN("Couldn't read /proc/stat\n", 0);
963 for (i = 0; i < len - 100; ++i) {
964 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
965 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
966 int cpu_no = atoi(stat_buf + i + 4);
967 if (cpu_no >= result) result = cpu_no + 1;
973 #endif /* GC_LINUX_THREADS */
975 /* We hold the GC lock. Wait until an in-progress GC has finished. */
976 /* Repeatedly RELEASES GC LOCK in order to wait. */
977 /* If wait_for_all is true, then we exit with the GC lock held and no */
978 /* collection in progress; otherwise we just wait for the current GC */
980 extern GC_bool GC_collection_in_progress();
981 void GC_wait_for_gc_completion(GC_bool wait_for_all)
983 if (GC_incremental && GC_collection_in_progress()) {
984 int old_gc_no = GC_gc_no;
986 /* Make sure that no part of our stack is still on the mark stack, */
987 /* since it's about to be unmapped. */
988 while (GC_incremental && GC_collection_in_progress()
989 && (wait_for_all || old_gc_no == GC_gc_no)) {
991 GC_in_thread_creation = TRUE;
992 GC_collect_a_little_inner(1);
993 GC_in_thread_creation = FALSE;
1003 /* Procedures called before and after a fork. The goal here is to make */
1004 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
1005 /* attainable, since the single UNIX spec seems to imply that one */
1006 /* should only call async-signal-safe functions, and we probably can't */
1007 /* quite guarantee that. But we give it our best shot. (That same */
1008 /* spec also implies that it's not safe to call the system malloc */
1009 /* between fork() and exec(). Thus we're doing no worse than it. */
1011 /* Called before a fork() */
1012 void GC_fork_prepare_proc(void)
1014 /* Acquire all relevant locks, so that after releasing the locks */
1015 /* the child will see a consistent state in which monitor */
1016 /* invariants hold. Unfortunately, we can't acquire libc locks */
1017 /* we might need, and there seems to be no guarantee that libc */
1018 /* must install a suitable fork handler. */
1019 /* Wait for an ongoing GC to finish, since we can't finish it in */
1020 /* the (one remaining thread in) the child. */
1022 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1023 GC_wait_for_reclaim();
1025 GC_wait_for_gc_completion(TRUE);
1026 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1027 GC_acquire_mark_lock();
1031 /* Called in parent after a fork() */
1032 void GC_fork_parent_proc(void)
1034 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1035 GC_release_mark_lock();
1040 /* Called in child after a fork() */
1041 void GC_fork_child_proc(void)
1043 /* Clean up the thread table, so that just our thread is left. */
1044 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1045 GC_release_mark_lock();
1047 GC_remove_all_threads_but_me();
1048 # ifdef PARALLEL_MARK
1049 /* Turn off parallel marking in the child, since we are probably */
1050 /* just going to exec, and we would have to restart mark threads. */
1052 GC_parallel = FALSE;
1053 # endif /* PARALLEL_MARK */
1056 #endif /* HANDLE_FORK */
1058 #if defined(GC_DGUX386_THREADS)
1059 /* Return the number of processors, or i<= 0 if it can't be determined. */
1062 /* <takis@XFree86.Org> */
1064 struct dg_sys_info_pm_info pm_sysinfo;
1067 status = dg_sys_info((long int *) &pm_sysinfo,
1068 DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
1070 /* set -1 for error */
1074 numCpus = pm_sysinfo.idle_vp_count;
1076 # ifdef DEBUG_THREADS
1077 GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
1081 #endif /* GC_DGUX386_THREADS */
1083 /* We hold the allocation lock. */
1086 # ifndef GC_DARWIN_THREADS
1091 if (GC_thr_initialized) return;
1092 GC_thr_initialized = TRUE;
1095 /* Prepare for a possible fork. */
1096 pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
1097 GC_fork_child_proc);
1098 # endif /* HANDLE_FORK */
1099 /* Add the initial thread, so we can stop it. */
1100 t = GC_new_thread(pthread_self());
1101 # ifdef GC_DARWIN_THREADS
1102 t -> stop_info.mach_thread = mach_thread_self();
1104 t -> stop_info.stack_ptr = (ptr_t)(&dummy);
1106 t -> flags = DETACHED | MAIN_THREAD;
1107 #ifdef MONO_DEBUGGER_SUPPORTED
1108 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1109 # ifdef GC_DARWIN_THREADS
1110 gc_thread_vtable->thread_created (mach_thread_self (), &t->stop_info.stack_ptr);
1112 gc_thread_vtable->thread_created (pthread_self (), &t->stop_info.stack_ptr);
1115 if (pthread_self () == main_pthread_self) {
1116 t->stack = main_stack;
1117 t->stack_size = main_stack_size;
1118 t->altstack = main_altstack;
1119 t->altstack_size = main_altstack_size;
1124 /* Set GC_nprocs. */
1126 char * nprocs_string = GETENV("GC_NPROCS");
1128 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
1130 if (GC_nprocs <= 0) {
1131 # if defined(GC_HPUX_THREADS)
1132 GC_nprocs = pthread_num_processors_np();
1134 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
1135 GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
1136 if (GC_nprocs <= 0) GC_nprocs = 1;
1138 # if defined(GC_IRIX_THREADS)
1139 GC_nprocs = sysconf(_SC_NPROC_ONLN);
1140 if (GC_nprocs <= 0) GC_nprocs = 1;
1142 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS) || defined(GC_NETBSD_THREADS) || defined(GC_OPENBSD_THREADS)
1144 size_t len = sizeof(ncpus);
1145 sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
1148 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1149 GC_nprocs = GC_get_nprocs();
1152 if (GC_nprocs <= 0) {
1153 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
1155 # ifdef PARALLEL_MARK
1159 # ifdef PARALLEL_MARK
1161 char * markers_string = GETENV("GC_MARKERS");
1162 if (markers_string != NULL) {
1163 GC_markers = atoi(markers_string);
1165 GC_markers = GC_nprocs;
1166 if (GC_markers > MAX_MARKERS)
1167 GC_markers = MAX_MARKERS;
1172 # ifdef PARALLEL_MARK
1174 if (GC_print_stats) {
1175 GC_printf2("Number of processors = %ld, "
1176 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
1179 if (GC_markers == 1) {
1180 GC_parallel = FALSE;
1182 if (GC_print_stats) {
1183 GC_printf0("Single marker thread, turning off parallel marking\n");
1188 /* Disable true incremental collection, but generational is OK. */
1189 GC_time_limit = GC_TIME_UNLIMITED;
1191 /* If we are using a parallel marker, actually start helper threads. */
1192 if (GC_parallel) start_mark_threads();
1197 /* Perform all initializations, including those that */
1198 /* may require allocation. */
1199 /* Called without allocation lock. */
1200 /* Must be called before a second thread is created. */
1201 /* Called without allocation lock. */
1202 void GC_init_parallel()
1204 if (parallel_initialized) return;
1205 parallel_initialized = TRUE;
1207 /* GC_init() calls us back, so set flag first. */
1208 if (!GC_is_initialized) GC_init();
1209 /* Initialize thread local free lists if used. */
1210 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1212 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1218 #if !defined(GC_DARWIN_THREADS) && !defined(GC_OPENBSD_THREADS)
1220 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
1222 sigset_t fudged_set;
1224 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
1226 sigdelset(&fudged_set, SIG_SUSPEND);
1229 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
1232 #endif /* !GC_DARWIN_THREADS */
1234 /* Wrappers for functions that are likely to block for an appreciable */
1235 /* length of time. Must be called in pairs, if at all. */
1236 /* Nothing much beyond the system call itself should be executed */
1237 /* between these. */
1239 void GC_start_blocking(void) {
1240 # define SP_SLOP 128
1243 me = GC_lookup_thread(pthread_self());
1244 GC_ASSERT(!(me -> thread_blocked));
1246 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
1248 # ifndef GC_DARWIN_THREADS
1249 me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
1253 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1255 /* Add some slop to the stack pointer, since the wrapped call may */
1256 /* end up pushing more callee-save registers. */
1257 # ifndef GC_DARWIN_THREADS
1258 # ifdef STACK_GROWS_UP
1259 me -> stop_info.stack_ptr += SP_SLOP;
1261 me -> stop_info.stack_ptr -= SP_SLOP;
1264 me -> thread_blocked = TRUE;
1268 void GC_end_blocking(void) {
1270 LOCK(); /* This will block if the world is stopped. */
1271 me = GC_lookup_thread(pthread_self());
1272 GC_ASSERT(me -> thread_blocked);
1273 me -> thread_blocked = FALSE;
1277 #if defined(GC_DGUX386_THREADS)
1278 #define __d10_sleep sleep
1279 #endif /* GC_DGUX386_THREADS */
1281 /* A wrapper for the standard C sleep function */
1282 int WRAP_FUNC(sleep) (unsigned int seconds)
1286 GC_start_blocking();
1287 result = REAL_FUNC(sleep)(seconds);
1293 void *(*start_routine)(void *);
1296 sem_t registered; /* 1 ==> in our thread table, but */
1297 /* parent hasn't yet noticed. */
1300 /* Called at thread exit. */
1301 /* Never called for main thread. That's OK, since it */
1302 /* results in at most a tiny one-time leak. And */
1303 /* linuxthreads doesn't reclaim the main threads */
1304 /* resources or id anyway. */
1305 void GC_thread_exit_proc(void *arg)
1310 me = GC_lookup_thread(pthread_self());
1311 GC_destroy_thread_local(me);
1312 if (me -> flags & DETACHED) {
1313 # ifdef THREAD_LOCAL_ALLOC
1314 /* NULL out the tls key to prevent the dtor function from being called */
1315 if (0 != GC_setspecific(GC_thread_key, NULL))
1316 ABORT("Failed to set thread specific allocation pointers");
1318 GC_delete_thread(pthread_self());
1320 me -> flags |= FINISHED;
1322 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1323 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1324 GC_remove_specific(GC_thread_key);
1326 /* The following may run the GC from "nonexistent" thread. */
1327 GC_wait_for_gc_completion(FALSE);
1331 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1334 GC_thread thread_gc_id;
1337 thread_gc_id = GC_lookup_thread(thread);
1338 /* This is guaranteed to be the intended one, since the thread id */
1339 /* cant have been recycled by pthreads. */
1341 result = REAL_FUNC(pthread_join)(thread, retval);
1342 # if defined (GC_FREEBSD_THREADS)
1343 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1344 appears to be) a spurious EINTR which caused the test and real code
1345 to gratuitously fail. Having looked at system pthread library source
1346 code, I see how this return code may be generated. In one path of
1347 code, pthread_join() just returns the errno setting of the thread
1348 being joined. This does not match the POSIX specification or the
1349 local man pages thus I have taken the liberty to catch this one
1350 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1351 if (result == EINTR) result = 0;
1355 /* Here the pthread thread id may have been recycled. */
1356 GC_delete_gc_thread(thread, thread_gc_id);
1363 /* Native Client doesn't support pthread cleanup functions, */
1364 /* so wrap pthread_exit and manually cleanup the thread. */
1366 WRAP_FUNC(pthread_exit)(void *status)
1368 GC_thread_exit_proc(0);
1369 REAL_FUNC(pthread_exit)(status);
1374 WRAP_FUNC(pthread_detach)(pthread_t thread)
1377 GC_thread thread_gc_id;
1380 thread_gc_id = GC_lookup_thread(thread);
1382 result = REAL_FUNC(pthread_detach)(thread);
1385 thread_gc_id -> flags |= DETACHED;
1386 /* Here the pthread thread id may have been recycled. */
1387 if (thread_gc_id -> flags & FINISHED) {
1388 GC_delete_gc_thread(thread, thread_gc_id);
1395 GC_bool GC_in_thread_creation = FALSE;
1397 typedef void *(*ThreadStartFn)(void *);
1398 void * GC_start_routine_head(void * arg, void *base_addr,
1399 ThreadStartFn *start, void **start_arg )
1401 struct start_info * si = arg;
1404 pthread_t my_pthread;
1406 my_pthread = pthread_self();
1407 # ifdef DEBUG_THREADS
1408 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1409 GC_printf1("pid = %ld\n", (long) getpid());
1410 GC_printf1("sp = 0x%lx\n", (long) &arg);
1413 GC_in_thread_creation = TRUE;
1414 me = GC_new_thread(my_pthread);
1415 GC_in_thread_creation = FALSE;
1416 #ifdef GC_DARWIN_THREADS
1417 me -> stop_info.mach_thread = mach_thread_self();
1419 me -> stop_info.stack_ptr = 0;
1421 me -> flags = si -> flags;
1422 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1423 /* doesn't work because the stack base in /proc/self/stat is the */
1424 /* one for the main thread. There is a strong argument that that's */
1425 /* a kernel bug, but a pervasive one. */
1426 # ifdef STACK_GROWS_DOWN
1427 me -> stack_end = (ptr_t)(((word)(base_addr) + (GC_page_size - 1))
1428 & ~(GC_page_size - 1));
1429 # ifndef GC_DARWIN_THREADS
1430 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1432 /* Needs to be plausible, since an asynchronous stack mark */
1433 /* should not crash. */
1435 me -> stack_end = (ptr_t)((word)(base_addr) & ~(GC_page_size - 1));
1436 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1438 /* This is dubious, since we may be more than a page into the stack, */
1439 /* and hence skip some of it, though it's not clear that matters. */
1441 me -> backing_store_end = (ptr_t)
1442 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1443 /* This is also < 100% convincing. We should also read this */
1444 /* from /proc, but the hook to do so isn't there yet. */
1446 #ifdef MONO_DEBUGGER_SUPPORTED
1447 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1448 # ifdef GC_DARWIN_THREADS
1449 gc_thread_vtable->thread_created (mach_thread_self(), &me->stop_info.stack_ptr);
1451 gc_thread_vtable->thread_created (my_pthread, &me->stop_info.stack_ptr);
1456 if (start) *start = si -> start_routine;
1457 if (start_arg) *start_arg = si -> arg;
1459 if (!(si->flags & FOREIGN_THREAD))
1460 sem_post(&(si -> registered)); /* Last action on si. */
1461 /* OK to deallocate. */
1462 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1464 GC_init_thread_local(me);
1471 int GC_thread_register_foreign (void *base_addr)
1473 struct start_info si = { 0, }; /* stacked for legibility & locking */
1476 # ifdef DEBUG_THREADS
1477 GC_printf1( "GC_thread_register_foreign %p\n", &si );
1480 si.flags = FOREIGN_THREAD;
1482 if (!parallel_initialized) GC_init_parallel();
1484 if (!GC_thr_initialized) GC_thr_init();
1488 me = GC_start_routine_head(&si, base_addr, NULL, NULL);
1493 void * GC_start_routine(void * arg)
1496 struct start_info * si = arg;
1499 ThreadStartFn start;
1502 me = GC_start_routine_head (arg, &dummy, &start, &start_arg);
1504 pthread_cleanup_push(GC_thread_exit_proc, 0);
1505 # ifdef DEBUG_THREADS
1506 GC_printf1("start_routine = 0x%lx\n", start);
1508 result = (*start)(start_arg);
1510 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1512 me -> status = result;
1513 pthread_cleanup_pop(1);
1514 /* Cleanup acquires lock, ensuring that we can't exit */
1515 /* while a collection that thinks we're alive is trying to stop */
1521 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1522 const pthread_attr_t *attr,
1523 void *(*start_routine)(void *), void *arg)
1528 struct start_info * si;
1529 /* This is otherwise saved only in an area mmapped by the thread */
1530 /* library, which isn't visible to the collector. */
1532 /* We resist the temptation to muck with the stack size here, */
1533 /* even if the default is unreasonably small. That's the client's */
1534 /* responsibility. */
1537 si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
1540 if (!parallel_initialized) GC_init_parallel();
1541 if (0 == si) return(ENOMEM);
1542 sem_init(&(si -> registered), 0, 0);
1543 si -> start_routine = start_routine;
1546 if (!GC_thr_initialized) GC_thr_init();
1547 # ifdef GC_ASSERTIONS
1551 pthread_attr_t my_attr;
1552 pthread_attr_init(&my_attr);
1553 pthread_attr_getstacksize(&my_attr, &stack_size);
1555 pthread_attr_getstacksize(attr, &stack_size);
1557 # ifdef PARALLEL_MARK
1558 GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1560 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1561 /* HBLKSIZE=8192, sizeof(word)=8 */
1562 GC_ASSERT(stack_size >= 65536);
1564 /* Our threads may need to do some work for the GC. */
1565 /* Ridiculously small threads won't work, and they */
1566 /* probably wouldn't work anyway. */
1570 detachstate = PTHREAD_CREATE_JOINABLE;
1572 pthread_attr_getdetachstate(attr, &detachstate);
1574 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1575 si -> flags = my_flags;
1577 # ifdef DEBUG_THREADS
1578 GC_printf1("About to start new thread from thread 0x%X\n",
1582 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1584 # ifdef DEBUG_THREADS
1585 GC_printf1("Started thread 0x%X\n", *new_thread);
1587 /* Wait until child has been added to the thread table. */
1588 /* This also ensures that we hold onto si until the child is done */
1589 /* with it. Thus it doesn't matter whether it is otherwise */
1590 /* visible to the collector. */
1592 while (0 != sem_wait(&(si -> registered))) {
1593 if (EINTR != errno) ABORT("sem_wait failed");
1596 sem_destroy(&(si -> registered));
1598 GC_INTERNAL_FREE(si);
1604 #ifdef GENERIC_COMPARE_AND_SWAP
1605 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1607 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1608 GC_word old, GC_word new_val)
1611 pthread_mutex_lock(&GC_compare_and_swap_lock);
1618 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1622 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1625 pthread_mutex_lock(&GC_compare_and_swap_lock);
1627 *addr = old + how_much;
1628 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1632 #endif /* GENERIC_COMPARE_AND_SWAP */
1633 /* Spend a few cycles in a way that can't introduce contention with */
1634 /* othre threads. */
1638 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1639 volatile word dummy = 0;
1642 for (i = 0; i < 10; ++i) {
1643 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1644 __asm__ __volatile__ (" " : : : "memory");
1646 /* Something that's unlikely to be optimized away. */
1652 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1655 VOLATILE GC_bool GC_collecting = 0;
1656 /* A hint that we're in the collector and */
1657 /* holding the allocation lock for an */
1658 /* extended period. */
1660 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1661 /* If we don't want to use the below spinlock implementation, either */
1662 /* because we don't have a GC_test_and_set implementation, or because */
1663 /* we don't want to risk sleeping, we can still try spinning on */
1664 /* pthread_mutex_trylock for a while. This appears to be very */
1665 /* beneficial in many cases. */
1666 /* I suspect that under high contention this is nearly always better */
1667 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1668 /* Hence we still default to the spin lock. */
1669 /* This is also used to acquire the mark lock for the parallel */
1672 /* Here we use a strict exponential backoff scheme. I don't know */
1673 /* whether that's better or worse than the above. We eventually */
1674 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1675 /* explicitly sleep. */
1679 unsigned long GC_spin_count = 0;
1680 unsigned long GC_block_count = 0;
1681 unsigned long GC_unlocked_count = 0;
1684 void GC_generic_lock(pthread_mutex_t * lock)
1686 #ifndef NO_PTHREAD_TRYLOCK
1687 unsigned pause_length = 1;
1690 if (0 == pthread_mutex_trylock(lock)) {
1692 ++GC_unlocked_count;
1696 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1697 for (i = 0; i < pause_length; ++i) {
1700 switch(pthread_mutex_trylock(lock)) {
1709 ABORT("Unexpected error from pthread_mutex_trylock");
1712 #endif /* !NO_PTHREAD_TRYLOCK */
1716 pthread_mutex_lock(lock);
1719 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1721 #if defined(USE_SPIN_LOCK)
1723 /* Reasonably fast spin locks. Basically the same implementation */
1724 /* as STL alloc.h. This isn't really the right way to do this. */
1725 /* but until the POSIX scheduling mess gets straightened out ... */
1727 volatile unsigned int GC_allocate_lock = 0;
1732 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1733 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1734 static unsigned spin_max = low_spin_max;
1735 unsigned my_spin_max;
1736 static unsigned last_spins = 0;
1737 unsigned my_last_spins;
1740 if (!GC_test_and_set(&GC_allocate_lock)) {
1743 my_spin_max = spin_max;
1744 my_last_spins = last_spins;
1745 for (i = 0; i < my_spin_max; i++) {
1746 if (GC_collecting || GC_nprocs == 1) goto yield;
1747 if (i < my_last_spins/2 || GC_allocate_lock) {
1751 if (!GC_test_and_set(&GC_allocate_lock)) {
1754 * Spinning worked. Thus we're probably not being scheduled
1755 * against the other process with which we were contending.
1756 * Thus it makes sense to spin longer the next time.
1759 spin_max = high_spin_max;
1763 /* We are probably being scheduled against the other process. Sleep. */
1764 spin_max = low_spin_max;
1767 if (!GC_test_and_set(&GC_allocate_lock)) {
1770 # define SLEEP_THRESHOLD 12
1771 /* Under Linux very short sleeps tend to wait until */
1772 /* the current time quantum expires. On old Linux */
1773 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1774 /* (Under 2.4, this happens only for real-time */
1775 /* processes.) We want to minimize both behaviors */
1777 if (i < SLEEP_THRESHOLD) {
1783 /* Don't wait for more than about 15msecs, even */
1784 /* under extreme contention. */
1786 ts.tv_nsec = 1 << i;
1792 #else /* !USE_SPINLOCK */
1795 #ifndef NO_PTHREAD_TRYLOCK
1796 if (1 == GC_nprocs || GC_collecting) {
1797 pthread_mutex_lock(&GC_allocate_ml);
1799 GC_generic_lock(&GC_allocate_ml);
1801 #else /* !NO_PTHREAD_TRYLOCK */
1802 pthread_mutex_lock(&GC_allocate_ml);
1803 #endif /* !NO_PTHREAD_TRYLOCK */
1806 #endif /* !USE_SPINLOCK */
1808 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1810 #ifdef GC_ASSERTIONS
1811 pthread_t GC_mark_lock_holder = NO_THREAD;
1815 /* Ugly workaround for a linux threads bug in the final versions */
1816 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1817 /* field even when it fails to acquire the mutex. This causes */
1818 /* pthread_cond_wait to die. Remove for glibc2.2. */
1819 /* According to the man page, we should use */
1820 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1822 static pthread_mutex_t mark_mutex =
1823 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1825 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1828 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1830 void GC_acquire_mark_lock()
1833 if (pthread_mutex_lock(&mark_mutex) != 0) {
1834 ABORT("pthread_mutex_lock failed");
1837 GC_generic_lock(&mark_mutex);
1838 # ifdef GC_ASSERTIONS
1839 GC_mark_lock_holder = pthread_self();
1843 void GC_release_mark_lock()
1845 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1846 # ifdef GC_ASSERTIONS
1847 GC_mark_lock_holder = NO_THREAD;
1849 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1850 ABORT("pthread_mutex_unlock failed");
1854 /* Collector must wait for a freelist builders for 2 reasons: */
1855 /* 1) Mark bits may still be getting examined without lock. */
1856 /* 2) Partial free lists referenced only by locals may not be scanned */
1857 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1858 /* free-list link may be ignored. */
1859 void GC_wait_builder()
1861 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1862 # ifdef GC_ASSERTIONS
1863 GC_mark_lock_holder = NO_THREAD;
1865 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1866 ABORT("pthread_cond_wait failed");
1868 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1869 # ifdef GC_ASSERTIONS
1870 GC_mark_lock_holder = pthread_self();
1874 void GC_wait_for_reclaim()
1876 GC_acquire_mark_lock();
1877 while (GC_fl_builder_count > 0) {
1880 GC_release_mark_lock();
1883 void GC_notify_all_builder()
1885 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1886 if (pthread_cond_broadcast(&builder_cv) != 0) {
1887 ABORT("pthread_cond_broadcast failed");
1891 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1893 #ifdef PARALLEL_MARK
1895 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1897 void GC_wait_marker()
1899 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1900 # ifdef GC_ASSERTIONS
1901 GC_mark_lock_holder = NO_THREAD;
1903 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1904 ABORT("pthread_cond_wait failed");
1906 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1907 # ifdef GC_ASSERTIONS
1908 GC_mark_lock_holder = pthread_self();
1912 void GC_notify_all_marker()
1914 if (pthread_cond_broadcast(&mark_cv) != 0) {
1915 ABORT("pthread_cond_broadcast failed");
1919 #endif /* PARALLEL_MARK */
1921 # endif /* GC_LINUX_THREADS and friends */