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_OPENBSD_THREADS)
73 # define USE_PTHREAD_SPECIFIC
76 # if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
77 # define _POSIX4A_DRAFT10_SOURCE 1
80 # if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
81 # define _USING_POSIX4A_DRAFT10 1
84 # ifdef THREAD_LOCAL_ALLOC
85 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
86 # include "private/specific.h"
89 /* Note that these macros should be used only to get/set the GC_thread pointer.
90 * We need to use both tls and pthread because we use the pthread_create function hook to
91 * free the data for foreign threads. When that doesn't happen, libgc could have old
92 * pthread_t that get reused...
94 # if defined(USE_PTHREAD_SPECIFIC)
95 # define GC_getspecific pthread_getspecific
96 # define GC_setspecific pthread_setspecific
97 # define GC_key_create pthread_key_create
98 typedef pthread_key_t GC_key_t;
100 # if defined(USE_COMPILER_TLS)
101 /* Note sles9 gcc on powerpc gets confused by the define to set GC_thread_tls and pthread_setspecific
102 * so we actually use a static inline function decalred below that is equivalent to:
103 * define GC_setspecific(key, v) (GC_thread_tls = (v), pthread_setspecific ((key), (v)))
105 # define GC_getspecific(x) (GC_thread_tls)
106 # define GC_key_create pthread_key_create
107 typedef pthread_key_t GC_key_t;
111 # include <pthread.h>
116 # include <sys/mman.h>
117 # include <sys/time.h>
118 # include <sys/types.h>
119 # include <sys/stat.h>
123 #if defined(GC_DARWIN_THREADS)
124 # include "private/darwin_semaphore.h"
126 # include <semaphore.h>
127 #endif /* !GC_DARWIN_THREADS */
129 #if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
130 # include <sys/sysctl.h>
131 #endif /* GC_DARWIN_THREADS */
133 #if defined(GC_NETBSD_THREADS) || defined(GC_OPENBSD_THREADS)
134 # include <sys/param.h>
135 # include <sys/sysctl.h>
140 #if defined(GC_DGUX386_THREADS)
141 # include <sys/dg_sys_info.h>
142 # include <sys/_int_psem.h>
143 /* sem_t is an uint in DG/UX */
144 typedef unsigned int sem_t;
145 #endif /* GC_DGUX386_THREADS */
151 #ifdef GC_USE_LD_WRAP
152 # define WRAP_FUNC(f) __wrap_##f
153 # define REAL_FUNC(f) __real_##f
155 # define WRAP_FUNC(f) GC_##f
156 # if !defined(GC_DGUX386_THREADS)
157 # define REAL_FUNC(f) f
158 # else /* GC_DGUX386_THREADS */
159 # define REAL_FUNC(f) __d10_##f
160 # endif /* GC_DGUX386_THREADS */
161 # undef pthread_create
162 # if !defined(GC_DARWIN_THREADS)
163 # undef pthread_sigmask
166 # undef pthread_detach
170 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
171 && !defined(_PTHREAD_USE_PTDNAM_)
172 /* Restore the original mangled names on Tru64 UNIX. */
173 # define pthread_create __pthread_create
174 # define pthread_join __pthread_join
175 # define pthread_detach __pthread_detach
181 static GC_bool parallel_initialized = FALSE;
183 void GC_init_parallel();
185 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
187 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
189 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
190 #undef PACKAGE_BUGREPORT
192 #undef PACKAGE_STRING
193 #undef PACKAGE_TARNAME
194 #undef PACKAGE_VERSION
195 #include "mono/utils/mono-compiler.h"
198 GC_key_t GC_thread_key;
200 #ifdef USE_COMPILER_TLS
201 __thread MONO_TLS_FAST void* GC_thread_tls;
204 * gcc errors out with /tmp/ccdPMFuq.s:2994: Error: symbol `.LTLS4' is already defined
205 * if the inline is added on powerpc
207 #if !defined(__ppc__) && !defined(__powerpc__)
210 static int GC_setspecific (GC_key_t key, void *value) {
211 GC_thread_tls = value;
212 return pthread_setspecific (key, value);
216 static GC_bool keys_initialized;
218 static pthread_t main_pthread_self;
219 static void *main_stack, *main_altstack;
220 static int main_stack_size, main_altstack_size;
222 #ifdef MONO_DEBUGGER_SUPPORTED
223 #include "include/libgc-mono-debugger.h"
226 /* Recover the contents of the freelist array fl into the global one gfl.*/
227 /* Note that the indexing scheme differs, in that gfl has finer size */
228 /* resolution, even if not all entries are used. */
229 /* We hold the allocator lock. */
230 static void return_freelists(ptr_t *fl, ptr_t *gfl)
236 for (i = 1; i < NFREELISTS; ++i) {
237 nwords = i * (GRANULARITY/sizeof(word));
240 if ((word)q >= HBLKSIZE) {
241 if (gfl[nwords] == 0) {
245 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
251 /* Clear fl[i], since the thread structure may hang around. */
252 /* Do it in a way that is likely to trap if we access it. */
253 fl[i] = (ptr_t)HBLKSIZE;
257 /* We statically allocate a single "size 0" object. It is linked to */
258 /* itself, and is thus repeatedly reused for all size 0 allocation */
259 /* requests. (Size 0 gcj allocation requests are incorrect, and */
260 /* we arrange for those to fault asap.) */
261 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
263 void GC_delete_gc_thread(pthread_t id, GC_thread gct);
264 void GC_destroy_thread_local(GC_thread p);
266 void GC_thread_deregister_foreign (void *data)
268 GC_thread me = (GC_thread)data;
269 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
270 if (me -> flags & FOREIGN_THREAD) {
272 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
273 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
274 GC_destroy_thread_local (me);
276 GC_delete_gc_thread(me->id, me);
281 /* Each thread structure must be initialized. */
282 /* This call must be made from the new thread. */
283 /* Caller holds allocation lock. */
284 void GC_init_thread_local(GC_thread p)
288 if (!keys_initialized) {
289 if (0 != GC_key_create(&GC_thread_key, GC_thread_deregister_foreign)) {
290 ABORT("Failed to create key for local allocator");
292 keys_initialized = TRUE;
294 if (0 != GC_setspecific(GC_thread_key, p)) {
295 ABORT("Failed to set thread specific allocation pointers");
297 for (i = 1; i < NFREELISTS; ++i) {
298 p -> ptrfree_freelists[i] = (ptr_t)1;
299 p -> normal_freelists[i] = (ptr_t)1;
300 # ifdef GC_GCJ_SUPPORT
301 p -> gcj_freelists[i] = (ptr_t)1;
304 /* Set up the size 0 free lists. */
305 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
306 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
307 # ifdef GC_GCJ_SUPPORT
308 p -> gcj_freelists[0] = (ptr_t)(-1);
312 #ifdef GC_GCJ_SUPPORT
313 extern ptr_t * GC_gcjobjfreelist;
316 /* We hold the allocator lock. */
317 void GC_destroy_thread_local(GC_thread p)
319 /* We currently only do this from the thread itself or from */
320 /* the fork handler for a child process. */
322 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
324 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
325 return_freelists(p -> normal_freelists, GC_objfreelist);
326 # ifdef GC_GCJ_SUPPORT
327 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
331 extern GC_PTR GC_generic_malloc_many();
333 GC_PTR GC_local_malloc(size_t bytes)
335 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
336 return(GC_malloc(bytes));
338 int index = INDEX_FROM_BYTES(bytes);
341 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
342 GC_key_t k = GC_thread_key;
346 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
347 if (EXPECT(0 == k, 0)) {
348 /* This can happen if we get called when the world is */
349 /* being initialized. Whether we can actually complete */
350 /* the initialization then is unclear. */
355 tsd = GC_getspecific(GC_thread_key);
356 # ifdef GC_ASSERTIONS
358 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
361 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
363 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
364 ptr_t next = obj_link(my_entry);
365 GC_PTR result = (GC_PTR)my_entry;
367 obj_link(my_entry) = 0;
368 PREFETCH_FOR_WRITE(next);
370 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
371 *my_fl = my_entry + index + 1;
372 return GC_malloc(bytes);
374 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
375 if (*my_fl == 0) return GC_oom_fn(bytes);
376 return GC_local_malloc(bytes);
381 GC_PTR GC_local_malloc_atomic(size_t bytes)
383 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
384 return(GC_malloc_atomic(bytes));
386 int index = INDEX_FROM_BYTES(bytes);
387 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
388 -> ptrfree_freelists + index;
389 ptr_t my_entry = *my_fl;
391 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
392 GC_PTR result = (GC_PTR)my_entry;
393 *my_fl = obj_link(my_entry);
395 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
396 *my_fl = my_entry + index + 1;
397 return GC_malloc_atomic(bytes);
399 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
400 /* *my_fl is updated while the collector is excluded; */
401 /* the free list is always visible to the collector as */
403 if (*my_fl == 0) return GC_oom_fn(bytes);
404 return GC_local_malloc_atomic(bytes);
409 #ifdef GC_GCJ_SUPPORT
411 #include "include/gc_gcj.h"
414 extern GC_bool GC_gcj_malloc_initialized;
417 extern int GC_gcj_kind;
419 GC_PTR GC_local_gcj_malloc(size_t bytes,
420 void * ptr_to_struct_containing_descr)
422 GC_ASSERT(GC_gcj_malloc_initialized);
423 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
424 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
426 int index = INDEX_FROM_BYTES(bytes);
427 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
428 -> gcj_freelists + index;
429 ptr_t my_entry = *my_fl;
430 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
431 GC_PTR result = (GC_PTR)my_entry;
432 GC_ASSERT(!GC_incremental);
433 /* We assert that any concurrent marker will stop us. */
434 /* Thus it is impossible for a mark procedure to see the */
435 /* allocation of the next object, but to see this object */
436 /* still containing a free list pointer. Otherwise the */
437 /* marker might find a random "mark descriptor". */
438 *(volatile ptr_t *)my_fl = obj_link(my_entry);
439 /* We must update the freelist before we store the pointer. */
440 /* Otherwise a GC at this point would see a corrupted */
442 /* A memory barrier is probably never needed, since the */
443 /* action of stopping this thread will cause prior writes */
445 GC_ASSERT(((void * volatile *)result)[1] == 0);
446 *(void * volatile *)result = ptr_to_struct_containing_descr;
448 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
449 if (!GC_incremental) *my_fl = my_entry + index + 1;
450 /* In the incremental case, we always have to take this */
451 /* path. Thus we leave the counter alone. */
452 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
454 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
455 if (*my_fl == 0) return GC_oom_fn(bytes);
456 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
461 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
462 /* adjust it. The size is assumed to be such that it can be */
463 /* allocated as a small object. */
464 void * GC_local_gcj_fast_malloc(size_t lw, void * ptr_to_struct_containing_descr)
466 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
467 -> gcj_freelists + lw;
468 ptr_t my_entry = *my_fl;
470 GC_ASSERT(GC_gcj_malloc_initialized);
472 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
473 GC_PTR result = (GC_PTR)my_entry;
474 GC_ASSERT(!GC_incremental);
475 /* We assert that any concurrent marker will stop us. */
476 /* Thus it is impossible for a mark procedure to see the */
477 /* allocation of the next object, but to see this object */
478 /* still containing a free list pointer. Otherwise the */
479 /* marker might find a random "mark descriptor". */
480 *(volatile ptr_t *)my_fl = obj_link(my_entry);
481 /* We must update the freelist before we store the pointer. */
482 /* Otherwise a GC at this point would see a corrupted */
484 /* A memory barrier is probably never needed, since the */
485 /* action of stopping this thread will cause prior writes */
487 GC_ASSERT(((void * volatile *)result)[1] == 0);
488 *(void * volatile *)result = ptr_to_struct_containing_descr;
490 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
491 if (!GC_incremental) *my_fl = my_entry + lw + 1;
492 /* In the incremental case, we always have to take this */
493 /* path. Thus we leave the counter alone. */
494 return GC_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
496 GC_generic_malloc_many(BYTES_FROM_INDEX(lw), GC_gcj_kind, my_fl);
497 if (*my_fl == 0) return GC_oom_fn(BYTES_FROM_INDEX(lw));
498 return GC_local_gcj_fast_malloc(lw, ptr_to_struct_containing_descr);
502 #endif /* GC_GCJ_SUPPORT */
504 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
506 # define GC_destroy_thread_local(t)
508 # endif /* !THREAD_LOCAL_ALLOC */
512 To make sure that we're using LinuxThreads and not some other thread
513 package, we generate a dummy reference to `pthread_kill_other_threads_np'
514 (was `__pthread_initial_thread_bos' but that disappeared),
515 which is a symbol defined in LinuxThreads, but (hopefully) not in other
518 We no longer do this, since this code is now portable enough that it might
519 actually work for something else.
521 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
524 long GC_nprocs = 1; /* Number of processors. We may not have */
525 /* access to all of them, but this is as good */
526 /* a guess as any ... */
531 # define MAX_MARKERS 16
534 static ptr_t marker_sp[MAX_MARKERS] = {0};
536 void * GC_mark_thread(void * id)
540 marker_sp[(word)id] = GC_approx_sp();
541 for (;; ++my_mark_no) {
542 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
543 /* promptly. This is important if it were called from the signal */
544 /* handler or from the GC lock acquisition code. Under Linux, it's */
545 /* not safe to call it from a signal handler, since it uses mutexes */
546 /* and condition variables. Since it is called only here, the */
547 /* argument is unnecessary. */
548 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
549 /* resynchronize if we get far off, e.g. because GC_mark_no */
551 my_mark_no = GC_mark_no;
553 # ifdef DEBUG_THREADS
554 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
556 GC_help_marker(my_mark_no);
560 extern long GC_markers; /* Number of mark threads we would */
561 /* like to have. Includes the */
562 /* initiating thread. */
564 pthread_t GC_mark_threads[MAX_MARKERS];
566 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
568 static void start_mark_threads()
573 if (GC_markers > MAX_MARKERS) {
574 WARN("Limiting number of mark threads\n", 0);
575 GC_markers = MAX_MARKERS;
577 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
579 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
580 ABORT("pthread_attr_setdetachstate failed");
582 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
583 /* Default stack size is usually too small: fix it. */
584 /* Otherwise marker threads or GC may run out of */
586 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
591 if (pthread_attr_getstacksize(&attr, &old_size) != 0)
592 ABORT("pthread_attr_getstacksize failed\n");
593 if (old_size < MIN_STACK_SIZE) {
594 if (pthread_attr_setstacksize(&attr, MIN_STACK_SIZE) != 0)
595 ABORT("pthread_attr_setstacksize failed\n");
598 # endif /* HPUX || GC_DGUX386_THREADS */
600 if (GC_print_stats) {
601 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
604 for (i = 0; i < GC_markers - 1; ++i) {
605 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
606 GC_mark_thread, (void *)(word)i)) {
607 WARN("Marker thread creation failed, errno = %ld.\n", errno);
612 #else /* !PARALLEL_MARK */
614 static __inline__ void start_mark_threads()
618 #endif /* !PARALLEL_MARK */
620 GC_bool GC_thr_initialized = FALSE;
622 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
625 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
626 * put it into a separate function.
628 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
629 static __attribute__((noinline)) unsigned char* get_gc_thread_key_addr GC_PROTO((void))
631 return (unsigned char*)&GC_thread_key;
634 void GC_push_thread_structures GC_PROTO((void))
636 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
637 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
638 GC_push_all((ptr_t)get_gc_thread_key_addr(),
639 (ptr_t)(get_gc_thread_key_addr())+sizeof(&GC_thread_key));
645 void GC_push_thread_structures GC_PROTO((void))
647 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
648 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
649 GC_push_all((ptr_t)(&GC_thread_key),
650 (ptr_t)(&GC_thread_key)+sizeof(&GC_thread_key));
656 #ifdef THREAD_LOCAL_ALLOC
657 /* We must explicitly mark ptrfree and gcj free lists, since the free */
658 /* list links wouldn't otherwise be found. We also set them in the */
659 /* normal free lists, since that involves touching less memory than if */
660 /* we scanned them normally. */
661 void GC_mark_thread_local_free_lists(void)
667 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
668 for (p = GC_threads[i]; 0 != p; p = p -> next) {
669 for (j = 1; j < NFREELISTS; ++j) {
670 q = p -> ptrfree_freelists[j];
671 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
672 q = p -> normal_freelists[j];
673 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
674 # ifdef GC_GCJ_SUPPORT
675 q = p -> gcj_freelists[j];
676 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
677 # endif /* GC_GCJ_SUPPORT */
682 #endif /* THREAD_LOCAL_ALLOC */
684 static struct GC_Thread_Rep first_thread;
687 extern int nacl_thread_parked[MAX_NACL_GC_THREADS];
688 extern int nacl_thread_used[MAX_NACL_GC_THREADS];
689 extern int nacl_thread_parking_inited;
690 extern int nacl_num_gc_threads;
691 extern pthread_mutex_t nacl_thread_alloc_lock;
692 extern __thread int nacl_thread_idx;
693 extern __thread GC_thread nacl_gc_thread_self;
695 extern void nacl_pre_syscall_hook();
696 extern void nacl_post_syscall_hook();
697 extern void nacl_register_gc_hooks(void (*pre)(), void (*post)());
699 void nacl_initialize_gc_thread()
702 nacl_register_gc_hooks(nacl_pre_syscall_hook, nacl_post_syscall_hook);
703 pthread_mutex_lock(&nacl_thread_alloc_lock);
704 if (!nacl_thread_parking_inited)
706 for (i = 0; i < MAX_NACL_GC_THREADS; i++) {
707 nacl_thread_used[i] = 0;
708 nacl_thread_parked[i] = 0;
710 nacl_thread_parking_inited = 1;
712 GC_ASSERT(nacl_num_gc_threads <= MAX_NACL_GC_THREADS);
713 for (i = 0; i < MAX_NACL_GC_THREADS; i++) {
714 if (nacl_thread_used[i] == 0) {
715 nacl_thread_used[i] = 1;
717 nacl_num_gc_threads++;
721 pthread_mutex_unlock(&nacl_thread_alloc_lock);
724 void nacl_shutdown_gc_thread()
726 pthread_mutex_lock(&nacl_thread_alloc_lock);
727 GC_ASSERT(nacl_thread_idx >= 0 && nacl_thread_idx < MAX_NACL_GC_THREADS);
728 GC_ASSERT(nacl_thread_used[nacl_thread_idx] != 0);
729 nacl_thread_used[nacl_thread_idx] = 0;
730 nacl_thread_idx = -1;
731 nacl_num_gc_threads--;
732 pthread_mutex_unlock(&nacl_thread_alloc_lock);
737 /* Add a thread to GC_threads. We assume it wasn't already there. */
738 /* Caller holds allocation lock. */
739 GC_thread GC_new_thread(pthread_t id)
741 int hv = ((unsigned long)id) % THREAD_TABLE_SZ;
743 static GC_bool first_thread_used = FALSE;
745 if (!first_thread_used) {
746 result = &first_thread;
747 first_thread_used = TRUE;
749 result = (struct GC_Thread_Rep *)
750 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
752 if (result == 0) return(0);
754 #ifdef PLATFORM_ANDROID
755 result -> kernel_id = gettid();
757 result -> next = GC_threads[hv];
758 GC_threads[hv] = result;
760 nacl_gc_thread_self = result;
761 nacl_initialize_gc_thread();
763 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
767 /* Delete a thread from GC_threads. We assume it is there. */
768 /* (The code intentionally traps if it wasn't.) */
769 /* Caller holds allocation lock. */
770 void GC_delete_thread(pthread_t id)
772 int hv = ((unsigned long)id) % THREAD_TABLE_SZ;
773 register GC_thread p = GC_threads[hv];
774 register GC_thread prev = 0;
777 nacl_shutdown_gc_thread();
778 nacl_gc_thread_self = NULL;
781 while (!pthread_equal(p -> id, id)) {
786 GC_threads[hv] = p -> next;
788 prev -> next = p -> next;
790 #ifdef MONO_DEBUGGER_SUPPORTED
791 if (gc_thread_vtable && gc_thread_vtable->thread_exited)
792 gc_thread_vtable->thread_exited (id, &p->stop_info.stack_ptr);
795 #ifdef GC_DARWIN_THREADS
796 mach_port_deallocate(mach_task_self(), p->stop_info.mach_thread);
802 /* If a thread has been joined, but we have not yet */
803 /* been notified, then there may be more than one thread */
804 /* in the table with the same pthread id. */
805 /* This is OK, but we need a way to delete a specific one. */
806 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
808 int hv = ((unsigned long)id) % THREAD_TABLE_SZ;
809 register GC_thread p = GC_threads[hv];
810 register GC_thread prev = 0;
817 GC_threads[hv] = p -> next;
819 prev -> next = p -> next;
822 #ifdef GC_DARWIN_THREADS
823 mach_port_deallocate(mach_task_self(), p->stop_info.mach_thread);
829 /* Return a GC_thread corresponding to a given pthread_t. */
830 /* Returns 0 if it's not there. */
831 /* Caller holds allocation lock or otherwise inhibits */
833 /* If there is more than one thread with the given id we */
834 /* return the most recent one. */
835 GC_thread GC_lookup_thread(pthread_t id)
837 int hv = ((unsigned long)id) % THREAD_TABLE_SZ;
838 register GC_thread p = GC_threads[hv];
840 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
844 int GC_thread_is_registered (void)
849 ptr = (void *)GC_lookup_thread(pthread_self());
855 void GC_register_altstack (void *stack, int stack_size, void *altstack, int altstack_size)
860 thread = (void *)GC_lookup_thread(pthread_self());
862 thread->stack = stack;
863 thread->stack_size = stack_size;
864 thread->altstack = altstack;
865 thread->altstack_size = altstack_size;
868 * This happens if we are called before GC_thr_init ().
870 main_pthread_self = pthread_self ();
872 main_stack_size = stack_size;
873 main_altstack = altstack;
874 main_altstack_size = altstack_size;
880 /* Remove all entries from the GC_threads table, except the */
881 /* one for the current thread. We need to do this in the child */
882 /* process after a fork(), since only the current thread */
883 /* survives in the child. */
884 void GC_remove_all_threads_but_me(void)
886 pthread_t self = pthread_self();
888 GC_thread p, next, me;
890 for (hv = 0; hv < THREAD_TABLE_SZ; ++hv) {
892 for (p = GC_threads[hv]; 0 != p; p = next) {
894 if (p -> id == self) {
898 # ifdef THREAD_LOCAL_ALLOC
899 if (!(p -> flags & FINISHED)) {
900 GC_destroy_thread_local(p);
902 # endif /* THREAD_LOCAL_ALLOC */
903 if (p != &first_thread) GC_INTERNAL_FREE(p);
910 #endif /* HANDLE_FORK */
912 #ifdef USE_PROC_FOR_LIBRARIES
913 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
918 # ifdef PARALLEL_MARK
919 for (i = 0; i < GC_markers; ++i) {
920 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
923 for (i = 0; i < THREAD_TABLE_SZ; i++) {
924 for (p = GC_threads[i]; p != 0; p = p -> next) {
925 if (0 != p -> stack_end) {
926 # ifdef STACK_GROWS_UP
927 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
928 # else /* STACK_GROWS_DOWN */
929 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
936 #endif /* USE_PROC_FOR_LIBRARIES */
938 #ifdef GC_LINUX_THREADS
939 /* Return the number of processors, or i<= 0 if it can't be determined. */
942 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
943 /* appears to be buggy in many cases. */
944 /* We look for lines "cpu<n>" in /proc/stat. */
945 # define STAT_BUF_SIZE 4096
946 # define STAT_READ read
947 /* If read is wrapped, this may need to be redefined to call */
949 char stat_buf[STAT_BUF_SIZE];
952 /* Some old kernels only have a single "cpu nnnn ..." */
953 /* entry in /proc/stat. We identify those as */
957 f = open("/proc/stat", O_RDONLY);
958 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
959 WARN("Couldn't read /proc/stat\n", 0);
962 for (i = 0; i < len - 100; ++i) {
963 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
964 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
965 int cpu_no = atoi(stat_buf + i + 4);
966 if (cpu_no >= result) result = cpu_no + 1;
972 #endif /* GC_LINUX_THREADS */
974 /* We hold the GC lock. Wait until an in-progress GC has finished. */
975 /* Repeatedly RELEASES GC LOCK in order to wait. */
976 /* If wait_for_all is true, then we exit with the GC lock held and no */
977 /* collection in progress; otherwise we just wait for the current GC */
979 extern GC_bool GC_collection_in_progress();
980 void GC_wait_for_gc_completion(GC_bool wait_for_all)
982 if (GC_incremental && GC_collection_in_progress()) {
983 int old_gc_no = GC_gc_no;
985 /* Make sure that no part of our stack is still on the mark stack, */
986 /* since it's about to be unmapped. */
987 while (GC_incremental && GC_collection_in_progress()
988 && (wait_for_all || old_gc_no == GC_gc_no)) {
990 GC_in_thread_creation = TRUE;
991 GC_collect_a_little_inner(1);
992 GC_in_thread_creation = FALSE;
1002 /* Procedures called before and after a fork. The goal here is to make */
1003 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
1004 /* attainable, since the single UNIX spec seems to imply that one */
1005 /* should only call async-signal-safe functions, and we probably can't */
1006 /* quite guarantee that. But we give it our best shot. (That same */
1007 /* spec also implies that it's not safe to call the system malloc */
1008 /* between fork() and exec(). Thus we're doing no worse than it. */
1010 /* Called before a fork() */
1011 void GC_fork_prepare_proc(void)
1013 /* Acquire all relevant locks, so that after releasing the locks */
1014 /* the child will see a consistent state in which monitor */
1015 /* invariants hold. Unfortunately, we can't acquire libc locks */
1016 /* we might need, and there seems to be no guarantee that libc */
1017 /* must install a suitable fork handler. */
1018 /* Wait for an ongoing GC to finish, since we can't finish it in */
1019 /* the (one remaining thread in) the child. */
1021 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1022 GC_wait_for_reclaim();
1024 GC_wait_for_gc_completion(TRUE);
1025 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1026 GC_acquire_mark_lock();
1030 /* Called in parent after a fork() */
1031 void GC_fork_parent_proc(void)
1033 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1034 GC_release_mark_lock();
1039 /* Called in child after a fork() */
1040 void GC_fork_child_proc(void)
1042 /* Clean up the thread table, so that just our thread is left. */
1043 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1044 GC_release_mark_lock();
1046 GC_remove_all_threads_but_me();
1047 # ifdef PARALLEL_MARK
1048 /* Turn off parallel marking in the child, since we are probably */
1049 /* just going to exec, and we would have to restart mark threads. */
1051 GC_parallel = FALSE;
1052 # endif /* PARALLEL_MARK */
1055 #endif /* HANDLE_FORK */
1057 #if defined(GC_DGUX386_THREADS)
1058 /* Return the number of processors, or i<= 0 if it can't be determined. */
1061 /* <takis@XFree86.Org> */
1063 struct dg_sys_info_pm_info pm_sysinfo;
1066 status = dg_sys_info((long int *) &pm_sysinfo,
1067 DG_SYS_INFO_PM_INFO_TYPE, DG_SYS_INFO_PM_CURRENT_VERSION);
1069 /* set -1 for error */
1073 numCpus = pm_sysinfo.idle_vp_count;
1075 # ifdef DEBUG_THREADS
1076 GC_printf1("Number of active CPUs in this system: %d\n", numCpus);
1080 #endif /* GC_DGUX386_THREADS */
1082 /* We hold the allocation lock. */
1085 # ifndef GC_DARWIN_THREADS
1090 if (GC_thr_initialized) return;
1091 GC_thr_initialized = TRUE;
1094 /* Prepare for a possible fork. */
1095 pthread_atfork(GC_fork_prepare_proc, GC_fork_parent_proc,
1096 GC_fork_child_proc);
1097 # endif /* HANDLE_FORK */
1098 /* Add the initial thread, so we can stop it. */
1099 t = GC_new_thread(pthread_self());
1100 # ifdef GC_DARWIN_THREADS
1101 t -> stop_info.mach_thread = mach_thread_self();
1103 t -> stop_info.stack_ptr = (ptr_t)(&dummy);
1105 t -> flags = DETACHED | MAIN_THREAD;
1106 #ifdef MONO_DEBUGGER_SUPPORTED
1107 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1108 # ifdef GC_DARWIN_THREADS
1109 gc_thread_vtable->thread_created (mach_thread_self (), &t->stop_info.stack_ptr);
1111 gc_thread_vtable->thread_created (pthread_self (), &t->stop_info.stack_ptr);
1114 if (pthread_self () == main_pthread_self) {
1115 t->stack = main_stack;
1116 t->stack_size = main_stack_size;
1117 t->altstack = main_altstack;
1118 t->altstack_size = main_altstack_size;
1123 /* Set GC_nprocs. */
1125 char * nprocs_string = GETENV("GC_NPROCS");
1127 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
1129 if (GC_nprocs <= 0) {
1130 # if defined(GC_HPUX_THREADS)
1131 GC_nprocs = pthread_num_processors_np();
1133 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
1134 GC_nprocs = sysconf(_SC_NPROCESSORS_ONLN);
1135 if (GC_nprocs <= 0) GC_nprocs = 1;
1137 # if defined(GC_IRIX_THREADS)
1138 GC_nprocs = sysconf(_SC_NPROC_ONLN);
1139 if (GC_nprocs <= 0) GC_nprocs = 1;
1141 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS) || defined(GC_NETBSD_THREADS) || defined(GC_OPENBSD_THREADS)
1143 size_t len = sizeof(ncpus);
1144 sysctl((int[2]) {CTL_HW, HW_NCPU}, 2, &ncpus, &len, NULL, 0);
1147 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1148 GC_nprocs = GC_get_nprocs();
1151 if (GC_nprocs <= 0) {
1152 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
1154 # ifdef PARALLEL_MARK
1158 # ifdef PARALLEL_MARK
1160 char * markers_string = GETENV("GC_MARKERS");
1161 if (markers_string != NULL) {
1162 GC_markers = atoi(markers_string);
1164 GC_markers = GC_nprocs;
1165 if (GC_markers > MAX_MARKERS)
1166 GC_markers = MAX_MARKERS;
1171 # ifdef PARALLEL_MARK
1173 if (GC_print_stats) {
1174 GC_printf2("Number of processors = %ld, "
1175 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
1178 if (GC_markers == 1) {
1179 GC_parallel = FALSE;
1181 if (GC_print_stats) {
1182 GC_printf0("Single marker thread, turning off parallel marking\n");
1187 /* Disable true incremental collection, but generational is OK. */
1188 GC_time_limit = GC_TIME_UNLIMITED;
1190 /* If we are using a parallel marker, actually start helper threads. */
1191 if (GC_parallel) start_mark_threads();
1196 /* Perform all initializations, including those that */
1197 /* may require allocation. */
1198 /* Called without allocation lock. */
1199 /* Must be called before a second thread is created. */
1200 /* Called without allocation lock. */
1201 void GC_init_parallel()
1203 if (parallel_initialized) return;
1204 parallel_initialized = TRUE;
1206 /* GC_init() calls us back, so set flag first. */
1207 if (!GC_is_initialized) GC_init();
1208 /* Initialize thread local free lists if used. */
1209 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1211 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1217 #if !defined(GC_DARWIN_THREADS) && !defined(GC_OPENBSD_THREADS)
1219 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
1221 sigset_t fudged_set;
1223 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
1225 sigdelset(&fudged_set, SIG_SUSPEND);
1228 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
1231 #endif /* !GC_DARWIN_THREADS */
1233 /* Wrappers for functions that are likely to block for an appreciable */
1234 /* length of time. Must be called in pairs, if at all. */
1235 /* Nothing much beyond the system call itself should be executed */
1236 /* between these. */
1238 void GC_start_blocking(void) {
1239 # define SP_SLOP 128
1242 me = GC_lookup_thread(pthread_self());
1243 GC_ASSERT(!(me -> thread_blocked));
1245 me -> stop_info.stack_ptr = (ptr_t)GC_save_regs_in_stack();
1247 # ifndef GC_DARWIN_THREADS
1248 me -> stop_info.stack_ptr = (ptr_t)GC_approx_sp();
1252 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1254 /* Add some slop to the stack pointer, since the wrapped call may */
1255 /* end up pushing more callee-save registers. */
1256 # ifndef GC_DARWIN_THREADS
1257 # ifdef STACK_GROWS_UP
1258 me -> stop_info.stack_ptr += SP_SLOP;
1260 me -> stop_info.stack_ptr -= SP_SLOP;
1263 me -> thread_blocked = TRUE;
1267 void GC_end_blocking(void) {
1269 LOCK(); /* This will block if the world is stopped. */
1270 me = GC_lookup_thread(pthread_self());
1271 GC_ASSERT(me -> thread_blocked);
1272 me -> thread_blocked = FALSE;
1276 #if defined(GC_DGUX386_THREADS)
1277 #define __d10_sleep sleep
1278 #endif /* GC_DGUX386_THREADS */
1280 /* A wrapper for the standard C sleep function */
1281 int WRAP_FUNC(sleep) (unsigned int seconds)
1285 GC_start_blocking();
1286 result = REAL_FUNC(sleep)(seconds);
1292 void *(*start_routine)(void *);
1295 sem_t registered; /* 1 ==> in our thread table, but */
1296 /* parent hasn't yet noticed. */
1299 /* Called at thread exit. */
1300 /* Never called for main thread. That's OK, since it */
1301 /* results in at most a tiny one-time leak. And */
1302 /* linuxthreads doesn't reclaim the main threads */
1303 /* resources or id anyway. */
1304 void GC_thread_exit_proc(void *arg)
1309 me = GC_lookup_thread(pthread_self());
1310 GC_destroy_thread_local(me);
1311 if (me -> flags & DETACHED) {
1312 # ifdef THREAD_LOCAL_ALLOC
1313 /* NULL out the tls key to prevent the dtor function from being called */
1314 if (0 != GC_setspecific(GC_thread_key, NULL))
1315 ABORT("Failed to set thread specific allocation pointers");
1317 GC_delete_thread(pthread_self());
1319 me -> flags |= FINISHED;
1321 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1322 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1323 GC_remove_specific(GC_thread_key);
1325 /* The following may run the GC from "nonexistent" thread. */
1326 GC_wait_for_gc_completion(FALSE);
1330 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1333 GC_thread thread_gc_id;
1336 thread_gc_id = GC_lookup_thread(thread);
1337 /* This is guaranteed to be the intended one, since the thread id */
1338 /* cant have been recycled by pthreads. */
1340 result = REAL_FUNC(pthread_join)(thread, retval);
1341 # if defined (GC_FREEBSD_THREADS)
1342 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1343 appears to be) a spurious EINTR which caused the test and real code
1344 to gratuitously fail. Having looked at system pthread library source
1345 code, I see how this return code may be generated. In one path of
1346 code, pthread_join() just returns the errno setting of the thread
1347 being joined. This does not match the POSIX specification or the
1348 local man pages thus I have taken the liberty to catch this one
1349 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1350 if (result == EINTR) result = 0;
1354 /* Here the pthread thread id may have been recycled. */
1355 GC_delete_gc_thread(thread, thread_gc_id);
1362 /* Native Client doesn't support pthread cleanup functions, */
1363 /* so wrap pthread_exit and manually cleanup the thread. */
1365 WRAP_FUNC(pthread_exit)(void *status)
1367 GC_thread_exit_proc(0);
1368 REAL_FUNC(pthread_exit)(status);
1373 WRAP_FUNC(pthread_detach)(pthread_t thread)
1376 GC_thread thread_gc_id;
1379 thread_gc_id = GC_lookup_thread(thread);
1381 result = REAL_FUNC(pthread_detach)(thread);
1384 thread_gc_id -> flags |= DETACHED;
1385 /* Here the pthread thread id may have been recycled. */
1386 if (thread_gc_id -> flags & FINISHED) {
1387 GC_delete_gc_thread(thread, thread_gc_id);
1394 GC_bool GC_in_thread_creation = FALSE;
1396 typedef void *(*ThreadStartFn)(void *);
1397 void * GC_start_routine_head(void * arg, void *base_addr,
1398 ThreadStartFn *start, void **start_arg )
1400 struct start_info * si = arg;
1403 pthread_t my_pthread;
1405 my_pthread = pthread_self();
1406 # ifdef DEBUG_THREADS
1407 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1408 GC_printf1("pid = %ld\n", (long) getpid());
1409 GC_printf1("sp = 0x%lx\n", (long) &arg);
1412 GC_in_thread_creation = TRUE;
1413 me = GC_new_thread(my_pthread);
1414 GC_in_thread_creation = FALSE;
1415 #ifdef GC_DARWIN_THREADS
1416 me -> stop_info.mach_thread = mach_thread_self();
1418 me -> stop_info.stack_ptr = 0;
1420 me -> flags = si -> flags;
1421 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1422 /* doesn't work because the stack base in /proc/self/stat is the */
1423 /* one for the main thread. There is a strong argument that that's */
1424 /* a kernel bug, but a pervasive one. */
1425 # ifdef STACK_GROWS_DOWN
1426 me -> stack_end = (ptr_t)(((word)(base_addr) + (GC_page_size - 1))
1427 & ~(GC_page_size - 1));
1428 # ifndef GC_DARWIN_THREADS
1429 me -> stop_info.stack_ptr = me -> stack_end - 0x10;
1431 /* Needs to be plausible, since an asynchronous stack mark */
1432 /* should not crash. */
1434 me -> stack_end = (ptr_t)((word)(base_addr) & ~(GC_page_size - 1));
1435 me -> stop_info.stack_ptr = me -> stack_end + 0x10;
1437 /* This is dubious, since we may be more than a page into the stack, */
1438 /* and hence skip some of it, though it's not clear that matters. */
1440 me -> backing_store_end = (ptr_t)
1441 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1442 /* This is also < 100% convincing. We should also read this */
1443 /* from /proc, but the hook to do so isn't there yet. */
1445 #ifdef MONO_DEBUGGER_SUPPORTED
1446 if (gc_thread_vtable && gc_thread_vtable->thread_created)
1447 # ifdef GC_DARWIN_THREADS
1448 gc_thread_vtable->thread_created (mach_thread_self(), &me->stop_info.stack_ptr);
1450 gc_thread_vtable->thread_created (my_pthread, &me->stop_info.stack_ptr);
1455 if (start) *start = si -> start_routine;
1456 if (start_arg) *start_arg = si -> arg;
1458 if (!(si->flags & FOREIGN_THREAD))
1459 sem_post(&(si -> registered)); /* Last action on si. */
1460 /* OK to deallocate. */
1461 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1463 GC_init_thread_local(me);
1470 int GC_thread_register_foreign (void *base_addr)
1472 struct start_info si = { 0, }; /* stacked for legibility & locking */
1475 # ifdef DEBUG_THREADS
1476 GC_printf1( "GC_thread_register_foreign %p\n", &si );
1479 si.flags = FOREIGN_THREAD;
1481 if (!parallel_initialized) GC_init_parallel();
1483 if (!GC_thr_initialized) GC_thr_init();
1487 me = GC_start_routine_head(&si, base_addr, NULL, NULL);
1492 void * GC_start_routine(void * arg)
1495 struct start_info * si = arg;
1498 ThreadStartFn start;
1501 me = GC_start_routine_head (arg, &dummy, &start, &start_arg);
1503 pthread_cleanup_push(GC_thread_exit_proc, 0);
1504 # ifdef DEBUG_THREADS
1505 GC_printf1("start_routine = 0x%lx\n", start);
1507 result = (*start)(start_arg);
1509 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1511 me -> status = result;
1512 pthread_cleanup_pop(1);
1513 /* Cleanup acquires lock, ensuring that we can't exit */
1514 /* while a collection that thinks we're alive is trying to stop */
1520 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1521 const pthread_attr_t *attr,
1522 void *(*start_routine)(void *), void *arg)
1527 struct start_info * si;
1528 /* This is otherwise saved only in an area mmapped by the thread */
1529 /* library, which isn't visible to the collector. */
1531 /* We resist the temptation to muck with the stack size here, */
1532 /* even if the default is unreasonably small. That's the client's */
1533 /* responsibility. */
1536 si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info),
1539 if (!parallel_initialized) GC_init_parallel();
1540 if (0 == si) return(ENOMEM);
1541 sem_init(&(si -> registered), 0, 0);
1542 si -> start_routine = start_routine;
1545 if (!GC_thr_initialized) GC_thr_init();
1546 # ifdef GC_ASSERTIONS
1550 pthread_attr_t my_attr;
1551 pthread_attr_init(&my_attr);
1552 pthread_attr_getstacksize(&my_attr, &stack_size);
1554 pthread_attr_getstacksize(attr, &stack_size);
1556 # ifdef PARALLEL_MARK
1557 GC_ASSERT(stack_size >= (8*HBLKSIZE*sizeof(word)));
1559 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1560 /* HBLKSIZE=8192, sizeof(word)=8 */
1561 GC_ASSERT(stack_size >= 65536);
1563 /* Our threads may need to do some work for the GC. */
1564 /* Ridiculously small threads won't work, and they */
1565 /* probably wouldn't work anyway. */
1569 detachstate = PTHREAD_CREATE_JOINABLE;
1571 pthread_attr_getdetachstate(attr, &detachstate);
1573 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1574 si -> flags = my_flags;
1576 # ifdef DEBUG_THREADS
1577 GC_printf1("About to start new thread from thread 0x%X\n",
1581 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1583 # ifdef DEBUG_THREADS
1584 GC_printf1("Started thread 0x%X\n", *new_thread);
1586 /* Wait until child has been added to the thread table. */
1587 /* This also ensures that we hold onto si until the child is done */
1588 /* with it. Thus it doesn't matter whether it is otherwise */
1589 /* visible to the collector. */
1591 while (0 != sem_wait(&(si -> registered))) {
1592 if (EINTR != errno) ABORT("sem_wait failed");
1595 sem_destroy(&(si -> registered));
1597 GC_INTERNAL_FREE(si);
1603 #ifdef GENERIC_COMPARE_AND_SWAP
1604 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1606 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1607 GC_word old, GC_word new_val)
1610 pthread_mutex_lock(&GC_compare_and_swap_lock);
1617 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1621 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1624 pthread_mutex_lock(&GC_compare_and_swap_lock);
1626 *addr = old + how_much;
1627 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1631 #endif /* GENERIC_COMPARE_AND_SWAP */
1632 /* Spend a few cycles in a way that can't introduce contention with */
1633 /* othre threads. */
1637 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1638 volatile word dummy = 0;
1641 for (i = 0; i < 10; ++i) {
1642 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1643 __asm__ __volatile__ (" " : : : "memory");
1645 /* Something that's unlikely to be optimized away. */
1651 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1654 VOLATILE GC_bool GC_collecting = 0;
1655 /* A hint that we're in the collector and */
1656 /* holding the allocation lock for an */
1657 /* extended period. */
1659 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1660 /* If we don't want to use the below spinlock implementation, either */
1661 /* because we don't have a GC_test_and_set implementation, or because */
1662 /* we don't want to risk sleeping, we can still try spinning on */
1663 /* pthread_mutex_trylock for a while. This appears to be very */
1664 /* beneficial in many cases. */
1665 /* I suspect that under high contention this is nearly always better */
1666 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1667 /* Hence we still default to the spin lock. */
1668 /* This is also used to acquire the mark lock for the parallel */
1671 /* Here we use a strict exponential backoff scheme. I don't know */
1672 /* whether that's better or worse than the above. We eventually */
1673 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1674 /* explicitly sleep. */
1678 unsigned long GC_spin_count = 0;
1679 unsigned long GC_block_count = 0;
1680 unsigned long GC_unlocked_count = 0;
1683 void GC_generic_lock(pthread_mutex_t * lock)
1685 #ifndef NO_PTHREAD_TRYLOCK
1686 unsigned pause_length = 1;
1689 if (0 == pthread_mutex_trylock(lock)) {
1691 ++GC_unlocked_count;
1695 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1696 for (i = 0; i < pause_length; ++i) {
1699 switch(pthread_mutex_trylock(lock)) {
1708 ABORT("Unexpected error from pthread_mutex_trylock");
1711 #endif /* !NO_PTHREAD_TRYLOCK */
1715 pthread_mutex_lock(lock);
1718 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1720 #if defined(USE_SPIN_LOCK)
1722 /* Reasonably fast spin locks. Basically the same implementation */
1723 /* as STL alloc.h. This isn't really the right way to do this. */
1724 /* but until the POSIX scheduling mess gets straightened out ... */
1726 volatile unsigned int GC_allocate_lock = 0;
1731 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1732 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1733 static unsigned spin_max = low_spin_max;
1734 unsigned my_spin_max;
1735 static unsigned last_spins = 0;
1736 unsigned my_last_spins;
1739 if (!GC_test_and_set(&GC_allocate_lock)) {
1742 my_spin_max = spin_max;
1743 my_last_spins = last_spins;
1744 for (i = 0; i < my_spin_max; i++) {
1745 if (GC_collecting || GC_nprocs == 1) goto yield;
1746 if (i < my_last_spins/2 || GC_allocate_lock) {
1750 if (!GC_test_and_set(&GC_allocate_lock)) {
1753 * Spinning worked. Thus we're probably not being scheduled
1754 * against the other process with which we were contending.
1755 * Thus it makes sense to spin longer the next time.
1758 spin_max = high_spin_max;
1762 /* We are probably being scheduled against the other process. Sleep. */
1763 spin_max = low_spin_max;
1766 if (!GC_test_and_set(&GC_allocate_lock)) {
1769 # define SLEEP_THRESHOLD 12
1770 /* Under Linux very short sleeps tend to wait until */
1771 /* the current time quantum expires. On old Linux */
1772 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1773 /* (Under 2.4, this happens only for real-time */
1774 /* processes.) We want to minimize both behaviors */
1776 if (i < SLEEP_THRESHOLD) {
1782 /* Don't wait for more than about 15msecs, even */
1783 /* under extreme contention. */
1785 ts.tv_nsec = 1 << i;
1791 #else /* !USE_SPINLOCK */
1794 #ifndef NO_PTHREAD_TRYLOCK
1795 if (1 == GC_nprocs || GC_collecting) {
1796 pthread_mutex_lock(&GC_allocate_ml);
1798 GC_generic_lock(&GC_allocate_ml);
1800 #else /* !NO_PTHREAD_TRYLOCK */
1801 pthread_mutex_lock(&GC_allocate_ml);
1802 #endif /* !NO_PTHREAD_TRYLOCK */
1805 #endif /* !USE_SPINLOCK */
1807 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1809 #ifdef GC_ASSERTIONS
1810 pthread_t GC_mark_lock_holder = NO_THREAD;
1814 /* Ugly workaround for a linux threads bug in the final versions */
1815 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1816 /* field even when it fails to acquire the mutex. This causes */
1817 /* pthread_cond_wait to die. Remove for glibc2.2. */
1818 /* According to the man page, we should use */
1819 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1821 static pthread_mutex_t mark_mutex =
1822 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1824 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1827 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1829 void GC_acquire_mark_lock()
1832 if (pthread_mutex_lock(&mark_mutex) != 0) {
1833 ABORT("pthread_mutex_lock failed");
1836 GC_generic_lock(&mark_mutex);
1837 # ifdef GC_ASSERTIONS
1838 GC_mark_lock_holder = pthread_self();
1842 void GC_release_mark_lock()
1844 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1845 # ifdef GC_ASSERTIONS
1846 GC_mark_lock_holder = NO_THREAD;
1848 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1849 ABORT("pthread_mutex_unlock failed");
1853 /* Collector must wait for a freelist builders for 2 reasons: */
1854 /* 1) Mark bits may still be getting examined without lock. */
1855 /* 2) Partial free lists referenced only by locals may not be scanned */
1856 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1857 /* free-list link may be ignored. */
1858 void GC_wait_builder()
1860 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1861 # ifdef GC_ASSERTIONS
1862 GC_mark_lock_holder = NO_THREAD;
1864 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1865 ABORT("pthread_cond_wait failed");
1867 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1868 # ifdef GC_ASSERTIONS
1869 GC_mark_lock_holder = pthread_self();
1873 void GC_wait_for_reclaim()
1875 GC_acquire_mark_lock();
1876 while (GC_fl_builder_count > 0) {
1879 GC_release_mark_lock();
1882 void GC_notify_all_builder()
1884 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1885 if (pthread_cond_broadcast(&builder_cv) != 0) {
1886 ABORT("pthread_cond_broadcast failed");
1890 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1892 #ifdef PARALLEL_MARK
1894 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1896 void GC_wait_marker()
1898 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1899 # ifdef GC_ASSERTIONS
1900 GC_mark_lock_holder = NO_THREAD;
1902 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1903 ABORT("pthread_cond_wait failed");
1905 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1906 # ifdef GC_ASSERTIONS
1907 GC_mark_lock_holder = pthread_self();
1911 void GC_notify_all_marker()
1913 if (pthread_cond_broadcast(&mark_cv) != 0) {
1914 ABORT("pthread_cond_broadcast failed");
1918 #endif /* PARALLEL_MARK */
1920 # endif /* GC_LINUX_THREADS and friends */