2 * Copyright (c) 1994 by Xerox Corporation. All rights reserved.
3 * Copyright (c) 1996 by Silicon Graphics. All rights reserved.
4 * Copyright (c) 1998 by Fergus Henderson. All rights reserved.
5 * Copyright (c) 2000-2001 by Hewlett-Packard Company. All rights reserved.
7 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
8 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
10 * Permission is hereby granted to use or copy this program
11 * for any purpose, provided the above notices are retained on all copies.
12 * Permission to modify the code and to distribute modified code is granted,
13 * provided the above notices are retained, and a notice that the code was
14 * modified is included with the above copyright notice.
17 * Support code for LinuxThreads, the clone()-based kernel
18 * thread package for Linux which is included in libc6.
20 * This code relies on implementation details of LinuxThreads,
21 * (i.e. properties not guaranteed by the Pthread standard),
22 * though this version now does less of that than the other Pthreads
25 * Note that there is a lot of code duplication between linux_threads.c
26 * and thread support for some of the other Posix platforms; any changes
27 * made here may need to be reflected there too.
30 * Linux_threads.c now also includes some code to support HPUX and
31 * OSF1 (Compaq Tru64 Unix, really). The OSF1 support is not yet
32 * functional. The OSF1 code is based on Eric Benson's
33 * patch, though that was originally against hpux_irix_threads. The code
34 * here is completely untested. With 0.0000001% probability, it might
37 * Eric also suggested an alternate basis for a lock implementation in
39 * + #elif defined(OSF1)
40 * + unsigned long GC_allocate_lock = 0;
41 * + msemaphore GC_allocate_semaphore;
42 * + # define GC_TRY_LOCK() \
43 * + ((msem_lock(&GC_allocate_semaphore, MSEM_IF_NOWAIT) == 0) \
44 * + ? (GC_allocate_lock = 1) \
46 * + # define GC_LOCK_TAKEN GC_allocate_lock
49 /* #define DEBUG_THREADS 1 */
51 /* ANSI C requires that a compilation unit contains something */
53 # if defined(GC_LINUX_THREADS) || defined(LINUX_THREADS) \
54 || defined(GC_HPUX_THREADS) || defined(HPUX_THREADS) \
55 || defined(GC_OSF1_THREADS) || defined(OSF1_THREADS) \
57 # include "private/gc_priv.h"
59 # if defined(HPUX_THREADS) && !defined(USE_PTHREAD_SPECIFIC) \
60 && !defined(USE_HPUX_TLS)
64 # ifdef THREAD_LOCAL_ALLOC
65 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_HPUX_TLS)
66 # include "private/specific.h"
68 # if defined(USE_PTHREAD_SPECIFIC)
69 # define GC_getspecific pthread_getspecific
70 # define GC_setspecific pthread_setspecific
71 # define GC_key_create pthread_key_create
72 typedef pthread_key_t GC_key_t;
74 # if defined(USE_HPUX_TLS)
75 # define GC_getspecific(x) (x)
76 # define GC_setspecific(key, v) ((key) = (v), 0)
77 # define GC_key_create(key, d) 0
78 typedef void * GC_key_t;
87 # include <sys/mman.h>
88 # include <sys/time.h>
89 # include <semaphore.h>
91 # include <sys/types.h>
92 # include <sys/stat.h>
100 # define WRAP_FUNC(f) __wrap_##f
101 # define REAL_FUNC(f) __real_##f
103 # define WRAP_FUNC(f) GC_##f
104 # define REAL_FUNC(f) f
105 # undef pthread_create
106 # undef pthread_sigmask
108 # undef pthread_detach
115 void GC_print_sig_mask()
120 if (pthread_sigmask(SIG_BLOCK, NULL, &blocked) != 0)
121 ABORT("pthread_sigmask");
122 GC_printf0("Blocked: ");
123 for (i = 1; i <= MAXSIG; i++) {
124 if (sigismember(&blocked, i)) { GC_printf1("%ld ",(long) i); }
131 /* We use the allocation lock to protect thread-related data structures. */
133 /* The set of all known threads. We intercept thread creation and */
135 /* Protected by allocation/GC lock. */
136 /* Some of this should be declared volatile, but that's inconsistent */
137 /* with some library routine declarations. */
138 typedef struct GC_Thread_Rep {
139 struct GC_Thread_Rep * next; /* More recently allocated threads */
140 /* with a given pthread id come */
141 /* first. (All but the first are */
142 /* guaranteed to be dead, but we may */
143 /* not yet have registered the join.) */
146 # define FINISHED 1 /* Thread has exited. */
147 # define DETACHED 2 /* Thread is intended to be detached. */
148 # define MAIN_THREAD 4 /* True for the original thread only. */
149 short thread_blocked; /* Protected by GC lock. */
150 /* Treated as a boolean value. If set, */
151 /* thread will acquire GC lock before */
152 /* doing any pointer manipulations, and */
153 /* has set its sp value. Thus it does */
154 /* not need to be sent a signal to stop */
156 ptr_t stack_end; /* Cold end of the stack. */
157 ptr_t stack_ptr; /* Valid only when stopped. */
159 ptr_t backing_store_end;
160 ptr_t backing_store_ptr;
163 void * status; /* The value returned from the thread. */
164 /* Used only to avoid premature */
165 /* reclamation of any data it might */
167 # ifdef THREAD_LOCAL_ALLOC
168 # if CPP_WORDSZ == 64 && defined(ALIGN_DOUBLE)
169 # define GRANULARITY 16
170 # define NFREELISTS 49
172 # define GRANULARITY 8
173 # define NFREELISTS 65
175 /* The ith free list corresponds to size i*GRANULARITY */
176 # define INDEX_FROM_BYTES(n) ((ADD_SLOP(n) + GRANULARITY - 1)/GRANULARITY)
177 # define BYTES_FROM_INDEX(i) ((i) * GRANULARITY - EXTRA_BYTES)
178 # define SMALL_ENOUGH(bytes) (ADD_SLOP(bytes) <= \
179 (NFREELISTS-1)*GRANULARITY)
180 ptr_t ptrfree_freelists[NFREELISTS];
181 ptr_t normal_freelists[NFREELISTS];
182 # ifdef GC_GCJ_SUPPORT
183 ptr_t gcj_freelists[NFREELISTS];
185 /* Free lists contain either a pointer or a small count */
186 /* reflecting the number of granules allocated at that */
188 /* 0 ==> thread-local allocation in use, free list */
190 /* > 0, <= DIRECT_GRANULES ==> Using global allocation, */
191 /* too few objects of this size have been */
192 /* allocated by this thread. */
193 /* >= HBLKSIZE => pointer to nonempty free list. */
194 /* > DIRECT_GRANULES, < HBLKSIZE ==> transition to */
195 /* local alloc, equivalent to 0. */
196 # define DIRECT_GRANULES (HBLKSIZE/GRANULARITY)
197 /* Don't use local free lists for up to this much */
202 GC_thread GC_lookup_thread(pthread_t id);
204 static GC_bool parallel_initialized = FALSE;
206 # if defined(__GNUC__)
207 void GC_init_parallel() __attribute__ ((constructor));
209 void GC_init_parallel();
212 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
214 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
219 GC_key_t GC_thread_key;
221 static GC_bool keys_initialized;
223 /* Recover the contents of the freelist array fl into the global one gfl.*/
224 /* Note that the indexing scheme differs, in that gfl has finer size */
225 /* resolution, even if not all entries are used. */
226 /* We hold the allocator lock. */
227 static void return_freelists(ptr_t *fl, ptr_t *gfl)
233 for (i = 1; i < NFREELISTS; ++i) {
234 nwords = i * (GRANULARITY/sizeof(word));
237 if ((word)q < HBLKSIZE) continue;
238 if (gfl[nwords] == 0) {
242 for (; (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
247 /* Clear fl[i], since the thread structure may hang around. */
248 /* Do it in a way that is likely to trap if we access it. */
249 fl[i] = (ptr_t)HBLKSIZE;
253 /* We statically allocate a single "size 0" object. It is linked to */
254 /* itself, and is thus repeatedly reused for all size 0 allocation */
255 /* requests. (Size 0 gcj allocation requests are incorrect, and */
256 /* we arrange for those to fault asap.) */
257 static ptr_t size_zero_object = (ptr_t)(&size_zero_object);
259 /* Each thread structure must be initialized. */
260 /* This call must be made from the new thread. */
261 /* Caller holds allocation lock. */
262 void GC_init_thread_local(GC_thread p)
266 if (!keys_initialized) {
267 if (0 != GC_key_create(&GC_thread_key, 0)) {
268 ABORT("Failed to create key for local allocator");
270 keys_initialized = TRUE;
272 if (0 != GC_setspecific(GC_thread_key, p)) {
273 ABORT("Failed to set thread specific allocation pointers");
275 for (i = 1; i < NFREELISTS; ++i) {
276 p -> ptrfree_freelists[i] = (ptr_t)1;
277 p -> normal_freelists[i] = (ptr_t)1;
278 # ifdef GC_GCJ_SUPPORT
279 p -> gcj_freelists[i] = (ptr_t)1;
282 /* Set up the size 0 free lists. */
283 p -> ptrfree_freelists[0] = (ptr_t)(&size_zero_object);
284 p -> normal_freelists[0] = (ptr_t)(&size_zero_object);
285 # ifdef GC_GCJ_SUPPORT
286 p -> gcj_freelists[0] = (ptr_t)(-1);
290 #ifdef GC_GCJ_SUPPORT
291 extern ptr_t * GC_gcjobjfreelist;
294 /* We hold the allocator lock. */
295 void GC_destroy_thread_local(GC_thread p)
297 /* We currently only do this from the thread itself. */
298 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
299 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
300 return_freelists(p -> normal_freelists, GC_objfreelist);
301 # ifdef GC_GCJ_SUPPORT
302 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
306 extern GC_PTR GC_generic_malloc_many();
308 GC_PTR GC_local_malloc(size_t bytes)
310 if (EXPECT(!SMALL_ENOUGH(bytes),0)) {
311 return(GC_malloc(bytes));
313 int index = INDEX_FROM_BYTES(bytes);
316 GC_key_t k = GC_thread_key;
319 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
320 || !defined(__GNUC__)
321 if (EXPECT(0 == k, 0)) {
322 /* This can happen if we get called when the world is */
323 /* being initialized. Whether we can actually complete */
324 /* the initialization then is unclear. */
329 tsd = GC_getspecific(GC_thread_key);
330 # ifdef GC_ASSERTIONS
332 GC_ASSERT(tsd == (void *)GC_lookup_thread(pthread_self()));
335 my_fl = ((GC_thread)tsd) -> normal_freelists + index;
337 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
338 ptr_t next = obj_link(my_entry);
339 GC_PTR result = (GC_PTR)my_entry;
341 obj_link(my_entry) = 0;
342 PREFETCH_FOR_WRITE(next);
344 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
345 *my_fl = my_entry + index + 1;
346 return GC_malloc(bytes);
348 GC_generic_malloc_many(BYTES_FROM_INDEX(index), NORMAL, my_fl);
349 if (*my_fl == 0) return GC_oom_fn(bytes);
350 return GC_local_malloc(bytes);
355 GC_PTR GC_local_malloc_atomic(size_t bytes)
357 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
358 return(GC_malloc_atomic(bytes));
360 int index = INDEX_FROM_BYTES(bytes);
361 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
362 -> ptrfree_freelists + index;
363 ptr_t my_entry = *my_fl;
364 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
365 GC_PTR result = (GC_PTR)my_entry;
366 *my_fl = obj_link(my_entry);
368 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
369 *my_fl = my_entry + index + 1;
370 return GC_malloc_atomic(bytes);
372 GC_generic_malloc_many(BYTES_FROM_INDEX(index), PTRFREE, my_fl);
373 /* *my_fl is updated while the collector is excluded; */
374 /* the free list is always visible to the collector as */
376 if (*my_fl == 0) return GC_oom_fn(bytes);
377 return GC_local_malloc_atomic(bytes);
382 #ifdef GC_GCJ_SUPPORT
384 #include "include/gc_gcj.h"
387 extern GC_bool GC_gcj_malloc_initialized;
390 extern int GC_gcj_kind;
392 GC_PTR GC_local_gcj_malloc(size_t bytes,
393 void * ptr_to_struct_containing_descr)
395 GC_ASSERT(GC_gcj_malloc_initialized);
396 if (EXPECT(!SMALL_ENOUGH(bytes), 0)) {
397 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
399 int index = INDEX_FROM_BYTES(bytes);
400 ptr_t * my_fl = ((GC_thread)GC_getspecific(GC_thread_key))
401 -> gcj_freelists + index;
402 ptr_t my_entry = *my_fl;
403 if (EXPECT((word)my_entry >= HBLKSIZE, 1)) {
404 GC_PTR result = (GC_PTR)my_entry;
405 GC_ASSERT(!GC_incremental);
406 /* We assert that any concurrent marker will stop us. */
407 /* Thus it is impossible for a mark procedure to see the */
408 /* allocation of the next object, but to see this object */
409 /* still containing a free list pointer. Otherwise the */
410 /* marker might find a random "mark descriptor". */
411 *(volatile ptr_t *)my_fl = obj_link(my_entry);
412 /* We must update the freelist before we store the pointer. */
413 /* Otherwise a GC at this point would see a corrupted */
415 /* A memory barrier is probably never needed, since the */
416 /* action of stopping this thread will cause prior writes */
418 *(void * volatile *)result = ptr_to_struct_containing_descr;
420 } else if ((word)my_entry - 1 < DIRECT_GRANULES) {
421 *my_fl = my_entry + index + 1;
422 return GC_gcj_malloc(bytes, ptr_to_struct_containing_descr);
424 GC_generic_malloc_many(BYTES_FROM_INDEX(index), GC_gcj_kind, my_fl);
425 if (*my_fl == 0) return GC_oom_fn(bytes);
426 return GC_local_gcj_malloc(bytes, ptr_to_struct_containing_descr);
431 #endif /* GC_GCJ_SUPPORT */
433 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
435 # define GC_destroy_thread_local(t)
437 # endif /* !THREAD_LOCAL_ALLOC */
440 * We use signals to stop threads during GC.
442 * Suspended threads wait in signal handler for SIG_THR_RESTART.
443 * That's more portable than semaphores or condition variables.
444 * (We do use sem_post from a signal handler, but that should be portable.)
446 * The thread suspension signal SIG_SUSPEND is now defined in gc_priv.h.
447 * Note that we can't just stop a thread; we need it to save its stack
448 * pointer(s) and acknowledge.
451 #ifndef SIG_THR_RESTART
452 # if defined(HPUX_THREADS) || defined(GC_OSF1_THREADS)
453 # define SIG_THR_RESTART _SIGRTMIN + 5
455 # define SIG_THR_RESTART SIGXCPU
459 sem_t GC_suspend_ack_sem;
461 #if !defined(HPUX_THREADS) && !defined(GC_OSF1_THREADS)
463 To make sure that we're using LinuxThreads and not some other thread
464 package, we generate a dummy reference to `pthread_kill_other_threads_np'
465 (was `__pthread_initial_thread_bos' but that disappeared),
466 which is a symbol defined in LinuxThreads, but (hopefully) not in other
469 void (*dummy_var_to_force_linux_threads)() = pthread_kill_other_threads_np;
470 #endif /* !HPUX_THREADS */
472 #if defined(SPARC) || defined(IA64)
473 extern word GC_save_regs_in_stack();
476 long GC_nprocs = 1; /* Number of processors. We may not have */
477 /* access to all of them, but this is as good */
478 /* a guess as any ... */
483 # define MAX_MARKERS 16
486 static ptr_t marker_sp[MAX_MARKERS] = {0};
488 void * GC_mark_thread(void * id)
492 marker_sp[(word)id] = GC_approx_sp();
493 for (;; ++my_mark_no) {
494 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
495 /* promptly. This is important if it were called from the signal */
496 /* handler or from the GC lock acquisition code. Under Linux, it's */
497 /* not safe to call it from a signal handler, since it uses mutexes */
498 /* and condition variables. Since it is called only here, the */
499 /* argument is unnecessary. */
500 if (my_mark_no < GC_mark_no || my_mark_no > GC_mark_no + 2) {
501 /* resynchronize if we get far off, e.g. because GC_mark_no */
503 my_mark_no = GC_mark_no;
505 # ifdef DEBUG_THREADS
506 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no);
508 GC_help_marker(my_mark_no);
512 extern long GC_markers; /* Number of mark threads we would */
513 /* like to have. Includes the */
514 /* initiating thread. */
516 pthread_t GC_mark_threads[MAX_MARKERS];
518 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
520 static void start_mark_threads()
525 if (GC_markers > MAX_MARKERS) {
526 WARN("Limiting number of mark threads\n", 0);
527 GC_markers = MAX_MARKERS;
529 if (0 != pthread_attr_init(&attr)) ABORT("pthread_attr_init failed");
531 if (0 != pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED))
532 ABORT("pthread_attr_setdetachstate failed");
534 if (GC_print_stats) {
535 GC_printf1("Starting %ld marker threads\n", GC_markers - 1);
538 for (i = 0; i < GC_markers - 1; ++i) {
539 if (0 != PTHREAD_CREATE(GC_mark_threads + i, &attr,
540 GC_mark_thread, (void *)(word)i)) {
541 WARN("Marker thread creation failed, errno = %ld.\n", errno);
546 #else /* !PARALLEL_MARK */
548 static __inline__ void start_mark_threads()
552 #endif /* !PARALLEL_MARK */
554 void GC_suspend_handler(int sig)
557 pthread_t my_thread = pthread_self();
563 # ifdef PARALLEL_MARK
564 word my_mark_no = GC_mark_no;
565 /* Marker can't proceed until we acknowledge. Thus this is */
566 /* guaranteed to be the mark_no correspending to our */
567 /* suspension, i.e. the marker can't have incremented it yet. */
570 if (sig != SIG_SUSPEND) ABORT("Bad signal in suspend_handler");
573 GC_printf1("Suspending 0x%x\n", my_thread);
576 me = GC_lookup_thread(my_thread);
577 /* The lookup here is safe, since I'm doing this on behalf */
578 /* of a thread which holds the allocation lock in order */
579 /* to stop the world. Thus concurrent modification of the */
580 /* data structure is impossible. */
582 me -> stack_ptr = (ptr_t)GC_save_regs_in_stack();
584 me -> stack_ptr = (ptr_t)(&dummy);
587 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack();
590 /* Tell the thread that wants to stop the world that this */
591 /* thread has been stopped. Note that sem_post() is */
592 /* the only async-signal-safe primitive in LinuxThreads. */
593 sem_post(&GC_suspend_ack_sem);
595 /* Wait until that thread tells us to restart by sending */
596 /* this thread a SIG_THR_RESTART signal. */
597 /* SIG_THR_RESTART should be masked at this point. Thus there */
599 if (sigfillset(&mask) != 0) ABORT("sigfillset() failed");
600 if (sigdelset(&mask, SIG_THR_RESTART) != 0) ABORT("sigdelset() failed");
602 if (sigdelset(&mask, SIGINT) != 0) ABORT("sigdelset() failed");
603 if (sigdelset(&mask, SIGQUIT) != 0) ABORT("sigdelset() failed");
604 if (sigdelset(&mask, SIGTERM) != 0) ABORT("sigdelset() failed");
605 if (sigdelset(&mask, SIGABRT) != 0) ABORT("sigdelset() failed");
609 sigsuspend(&mask); /* Wait for signal */
610 } while (me->signal != SIG_THR_RESTART);
613 GC_printf1("Continuing 0x%x\n", my_thread);
617 void GC_restart_handler(int sig)
621 if (sig != SIG_THR_RESTART) ABORT("Bad signal in suspend_handler");
623 /* Let the GC_suspend_handler() know that we got a SIG_THR_RESTART. */
624 /* The lookup here is safe, since I'm doing this on behalf */
625 /* of a thread which holds the allocation lock in order */
626 /* to stop the world. Thus concurrent modification of the */
627 /* data structure is impossible. */
628 me = GC_lookup_thread(pthread_self());
629 me->signal = SIG_THR_RESTART;
632 ** Note: even if we didn't do anything useful here,
633 ** it would still be necessary to have a signal handler,
634 ** rather than ignoring the signals, otherwise
635 ** the signals will not be delivered at all, and
636 ** will thus not interrupt the sigsuspend() above.
640 GC_printf1("In GC_restart_handler for 0x%x\n", pthread_self());
644 /* Defining INSTALL_LOOPING_SEGV_HANDLER causes SIGSEGV and SIGBUS to */
645 /* result in an infinite loop in a signal handler. This can be very */
646 /* useful for debugging, since (as of RH7) gdb still seems to have */
647 /* serious problems with threads. */
648 #ifdef INSTALL_LOOPING_SEGV_HANDLER
649 void GC_looping_handler(int sig)
651 GC_printf3("Signal %ld in thread %lx, pid %ld\n",
652 sig, pthread_self(), getpid());
657 GC_bool GC_thr_initialized = FALSE;
659 # define THREAD_TABLE_SZ 128 /* Must be power of 2 */
660 volatile GC_thread GC_threads[THREAD_TABLE_SZ];
662 void GC_push_thread_structures GC_PROTO((void))
664 GC_push_all((ptr_t)(GC_threads), (ptr_t)(GC_threads)+sizeof(GC_threads));
667 #ifdef THREAD_LOCAL_ALLOC
668 /* We must explicitly mark ptrfree and gcj free lists, since the free */
669 /* list links wouldn't otherwise be found. We also set them in the */
670 /* normal free lists, since that involves touching less memory than if */
671 /* we scanned them normally. */
672 void GC_mark_thread_local_free_lists(void)
678 for (i = 0; i < THREAD_TABLE_SZ; ++i) {
679 for (p = GC_threads[i]; 0 != p; p = p -> next) {
680 for (j = 1; j < NFREELISTS; ++j) {
681 q = p -> ptrfree_freelists[j];
682 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
683 q = p -> normal_freelists[j];
684 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
685 # ifdef GC_GCJ_SUPPORT
686 q = p -> gcj_freelists[j];
687 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
688 # endif /* GC_GCJ_SUPPORT */
693 #endif /* THREAD_LOCAL_ALLOC */
695 /* Add a thread to GC_threads. We assume it wasn't already there. */
696 /* Caller holds allocation lock. */
697 GC_thread GC_new_thread(pthread_t id)
699 int hv = ((word)id) % THREAD_TABLE_SZ;
701 static struct GC_Thread_Rep first_thread;
702 static GC_bool first_thread_used = FALSE;
704 if (!first_thread_used) {
705 result = &first_thread;
706 first_thread_used = TRUE;
708 result = (struct GC_Thread_Rep *)
709 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep), NORMAL);
711 if (result == 0) return(0);
713 result -> next = GC_threads[hv];
714 GC_threads[hv] = result;
715 GC_ASSERT(result -> flags == 0 && result -> thread_blocked == 0);
719 /* Delete a thread from GC_threads. We assume it is there. */
720 /* (The code intentionally traps if it wasn't.) */
721 /* Caller holds allocation lock. */
722 void GC_delete_thread(pthread_t id)
724 int hv = ((word)id) % THREAD_TABLE_SZ;
725 register GC_thread p = GC_threads[hv];
726 register GC_thread prev = 0;
728 while (!pthread_equal(p -> id, id)) {
733 GC_threads[hv] = p -> next;
735 prev -> next = p -> next;
740 /* If a thread has been joined, but we have not yet */
741 /* been notified, then there may be more than one thread */
742 /* in the table with the same pthread id. */
743 /* This is OK, but we need a way to delete a specific one. */
744 void GC_delete_gc_thread(pthread_t id, GC_thread gc_id)
746 int hv = ((word)id) % THREAD_TABLE_SZ;
747 register GC_thread p = GC_threads[hv];
748 register GC_thread prev = 0;
755 GC_threads[hv] = p -> next;
757 prev -> next = p -> next;
762 /* Return a GC_thread corresponding to a given thread_t. */
763 /* Returns 0 if it's not there. */
764 /* Caller holds allocation lock or otherwise inhibits */
766 /* If there is more than one thread with the given id we */
767 /* return the most recent one. */
768 GC_thread GC_lookup_thread(pthread_t id)
770 int hv = ((word)id) % THREAD_TABLE_SZ;
771 register GC_thread p = GC_threads[hv];
773 while (p != 0 && !pthread_equal(p -> id, id)) p = p -> next;
777 /* There seems to be a very rare thread stopping problem. To help us */
778 /* debug that, we save the ids of the stopping thread. */
779 pthread_t GC_stopping_thread;
782 /* Caller holds allocation lock. */
785 pthread_t my_thread = pthread_self();
787 register GC_thread p;
788 register int n_live_threads = 0;
791 GC_stopping_thread = my_thread; /* debugging only. */
792 GC_stopping_pid = getpid(); /* debugging only. */
793 /* Make sure all free list construction has stopped before we start. */
794 /* No new construction can start, since free list construction is */
795 /* required to acquire and release the GC lock before it starts, */
796 /* and we have the lock. */
797 # ifdef PARALLEL_MARK
798 GC_acquire_mark_lock();
799 GC_ASSERT(GC_fl_builder_count == 0);
800 /* We should have previously waited for it to become zero. */
801 # endif /* PARALLEL_MARK */
802 for (i = 0; i < THREAD_TABLE_SZ; i++) {
803 for (p = GC_threads[i]; p != 0; p = p -> next) {
804 if (p -> id != my_thread) {
805 if (p -> flags & FINISHED) continue;
806 if (p -> thread_blocked) /* Will wait */ continue;
809 GC_printf1("Sending suspend signal to 0x%x\n", p -> id);
811 result = pthread_kill(p -> id, SIG_SUSPEND);
814 /* Not really there anymore. Possible? */
820 ABORT("pthread_kill failed");
825 for (i = 0; i < n_live_threads; i++) {
826 if (0 != sem_wait(&GC_suspend_ack_sem))
827 ABORT("sem_wait in handler failed");
829 # ifdef PARALLEL_MARK
830 GC_release_mark_lock();
833 GC_printf1("World stopped 0x%x\n", pthread_self());
835 GC_stopping_thread = 0; /* debugging only */
838 /* Caller holds allocation lock, and has held it continuously since */
839 /* the world stopped. */
840 void GC_start_world()
842 pthread_t my_thread = pthread_self();
844 register GC_thread p;
845 register int n_live_threads = 0;
849 GC_printf0("World starting\n");
852 for (i = 0; i < THREAD_TABLE_SZ; i++) {
853 for (p = GC_threads[i]; p != 0; p = p -> next) {
854 if (p -> id != my_thread) {
855 if (p -> flags & FINISHED) continue;
856 if (p -> thread_blocked) continue;
859 GC_printf1("Sending restart signal to 0x%x\n", p -> id);
861 result = pthread_kill(p -> id, SIG_THR_RESTART);
864 /* Not really there anymore. Possible? */
870 ABORT("pthread_kill failed");
876 GC_printf0("World started\n");
878 GC_stopping_thread = 0; /* debugging only */
882 # define IF_IA64(x) x
886 /* We hold allocation lock. Should do exactly the right thing if the */
887 /* world is stopped. Should not fail if it isn't. */
888 void GC_push_all_stacks()
892 ptr_t sp = GC_approx_sp();
894 /* On IA64, we also need to scan the register backing store. */
895 IF_IA64(ptr_t bs_lo; ptr_t bs_hi;)
896 pthread_t me = pthread_self();
898 if (!GC_thr_initialized) GC_thr_init();
900 GC_printf1("Pushing stacks from thread 0x%lx\n", (unsigned long) me);
902 for (i = 0; i < THREAD_TABLE_SZ; i++) {
903 for (p = GC_threads[i]; p != 0; p = p -> next) {
904 if (p -> flags & FINISHED) continue;
905 if (pthread_equal(p -> id, me)) {
907 lo = (ptr_t)GC_save_regs_in_stack();
911 IF_IA64(bs_hi = (ptr_t)GC_save_regs_in_stack();)
914 IF_IA64(bs_hi = p -> backing_store_ptr;)
916 if ((p -> flags & MAIN_THREAD) == 0) {
918 IF_IA64(bs_lo = p -> backing_store_end);
920 /* The original stack. */
922 IF_IA64(bs_lo = BACKING_STORE_BASE;)
925 GC_printf3("Stack for thread 0x%lx = [%lx,%lx)\n",
926 (unsigned long) p -> id,
927 (unsigned long) lo, (unsigned long) hi);
929 if (0 == lo) ABORT("GC_push_all_stacks: sp not set!\n");
930 # ifdef STACK_GROWS_UP
931 /* We got them backwards! */
932 GC_push_all_stack(hi, lo);
934 GC_push_all_stack(lo, hi);
937 if (pthread_equal(p -> id, me)) {
938 GC_push_all_eager(bs_lo, bs_hi);
940 GC_push_all_stack(bs_lo, bs_hi);
947 #ifdef USE_PROC_FOR_LIBRARIES
948 int GC_segment_is_thread_stack(ptr_t lo, ptr_t hi)
953 # ifdef PARALLEL_MARK
954 for (i = 0; i < GC_markers; ++i) {
955 if (marker_sp[i] > lo & marker_sp[i] < hi) return 1;
958 for (i = 0; i < THREAD_TABLE_SZ; i++) {
959 for (p = GC_threads[i]; p != 0; p = p -> next) {
960 if (0 != p -> stack_end) {
961 # ifdef STACK_GROWS_UP
962 if (p -> stack_end >= lo && p -> stack_end < hi) return 1;
963 # else /* STACK_GROWS_DOWN */
964 if (p -> stack_end > lo && p -> stack_end <= hi) return 1;
971 #endif /* USE_PROC_FOR_LIBRARIES */
974 /* Return the number of processors, or i<= 0 if it can't be determined. */
977 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
978 /* appears to be buggy in many cases. */
979 /* We look for lines "cpu<n>" in /proc/stat. */
980 # define STAT_BUF_SIZE 4096
981 # if defined(GC_USE_LD_WRAP)
982 # define STAT_READ __real_read
984 # define STAT_READ read
986 char stat_buf[STAT_BUF_SIZE];
990 /* Some old kernels only have a single "cpu nnnn ..." */
991 /* entry in /proc/stat. We identify those as */
995 f = open("/proc/stat", O_RDONLY);
996 if (f < 0 || (len = STAT_READ(f, stat_buf, STAT_BUF_SIZE)) < 100) {
997 WARN("Couldn't read /proc/stat\n", 0);
1000 for (i = 0; i < len - 100; ++i) {
1001 if (stat_buf[i] == '\n' && stat_buf[i+1] == 'c'
1002 && stat_buf[i+2] == 'p' && stat_buf[i+3] == 'u') {
1003 int cpu_no = atoi(stat_buf + i + 4);
1004 if (cpu_no >= result) result = cpu_no + 1;
1009 #endif /* LINUX_THREADS */
1011 /* We hold the allocation lock. */
1016 struct sigaction act;
1018 if (GC_thr_initialized) return;
1019 GC_thr_initialized = TRUE;
1021 if (sem_init(&GC_suspend_ack_sem, 0, 0) != 0)
1022 ABORT("sem_init failed");
1024 act.sa_flags = SA_RESTART;
1025 if (sigfillset(&act.sa_mask) != 0) {
1026 ABORT("sigfillset() failed");
1029 if (sigdelset(&act.sa_mask, SIGINT) != 0
1030 || sigdelset(&act.sa_mask, SIGQUIT != 0)
1031 || sigdelset(&act.sa_mask, SIGABRT != 0)
1032 || sigdelset(&act.sa_mask, SIGTERM != 0)) {
1033 ABORT("sigdelset() failed");
1037 /* SIG_THR_RESTART is unmasked by the handler when necessary. */
1038 act.sa_handler = GC_suspend_handler;
1039 if (sigaction(SIG_SUSPEND, &act, NULL) != 0) {
1040 ABORT("Cannot set SIG_SUSPEND handler");
1043 act.sa_handler = GC_restart_handler;
1044 if (sigaction(SIG_THR_RESTART, &act, NULL) != 0) {
1045 ABORT("Cannot set SIG_THR_RESTART handler");
1047 # ifdef INSTALL_LOOPING_SEGV_HANDLER
1048 act.sa_handler = GC_looping_handler;
1049 if (sigaction(SIGSEGV, &act, NULL) != 0
1050 || sigaction(SIGBUS, &act, NULL) != 0) {
1051 ABORT("Cannot set SIGSEGV or SIGBUS looping handler");
1053 # endif /* INSTALL_LOOPING_SEGV_HANDLER */
1055 /* Add the initial thread, so we can stop it. */
1056 t = GC_new_thread(pthread_self());
1057 t -> stack_ptr = (ptr_t)(&dummy);
1058 t -> flags = DETACHED | MAIN_THREAD;
1060 /* Set GC_nprocs. */
1062 char * nprocs_string = GETENV("GC_NPROCS");
1064 if (nprocs_string != NULL) GC_nprocs = atoi(nprocs_string);
1066 if (GC_nprocs <= 0) {
1067 # if defined(HPUX_THREADS)
1068 GC_nprocs = pthread_num_processors_np();
1070 # if defined(OSF1_THREADS)
1073 # ifdef LINUX_THREADS
1074 GC_nprocs = GC_get_nprocs();
1077 if (GC_nprocs <= 0) {
1078 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs);
1080 # ifdef PARALLEL_MARK
1084 # ifdef PARALLEL_MARK
1085 GC_markers = GC_nprocs;
1088 # ifdef PARALLEL_MARK
1090 if (GC_print_stats) {
1091 GC_printf2("Number of processors = %ld, "
1092 "number of marker threads = %ld\n", GC_nprocs, GC_markers);
1095 if (GC_markers == 1) {
1096 GC_parallel = FALSE;
1098 if (GC_print_stats) {
1099 GC_printf0("Single marker thread, turning off parallel marking\n");
1109 /* Perform all initializations, including those that */
1110 /* may require allocation. */
1111 /* Called as constructor without allocation lock. */
1112 /* Must be called before a second thread is created. */
1113 /* Called without allocation lock. */
1114 void GC_init_parallel()
1116 if (parallel_initialized) return;
1117 parallel_initialized = TRUE;
1118 /* GC_init() calls us back, so set flag first. */
1119 if (!GC_is_initialized) GC_init();
1120 /* If we are using a parallel marker, start the helper threads. */
1121 # ifdef PARALLEL_MARK
1122 if (GC_parallel) start_mark_threads();
1124 /* Initialize thread local free lists if used. */
1125 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1127 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1133 int WRAP_FUNC(pthread_sigmask)(int how, const sigset_t *set, sigset_t *oset)
1135 sigset_t fudged_set;
1137 if (set != NULL && (how == SIG_BLOCK || how == SIG_SETMASK)) {
1139 sigdelset(&fudged_set, SIG_SUSPEND);
1142 return(REAL_FUNC(pthread_sigmask)(how, set, oset));
1145 /* Wrappers for functions that are likely to block for an appreciable */
1146 /* length of time. Must be called in pairs, if at all. */
1147 /* Nothing much beyond the system call itself should be executed */
1148 /* between these. */
1150 void GC_start_blocking(void) {
1151 # define SP_SLOP 128
1154 me = GC_lookup_thread(pthread_self());
1155 GC_ASSERT(!(me -> thread_blocked));
1157 me -> stack_ptr = (ptr_t)GC_save_regs_in_stack();
1159 me -> stack_ptr = (ptr_t)GC_approx_sp();
1162 me -> backing_store_ptr = (ptr_t)GC_save_regs_in_stack() + SP_SLOP;
1164 /* Add some slop to the stack pointer, since the wrapped call may */
1165 /* end up pushing more callee-save registers. */
1166 # ifdef STACK_GROWS_UP
1167 me -> stack_ptr += SP_SLOP;
1169 me -> stack_ptr -= SP_SLOP;
1171 me -> thread_blocked = TRUE;
1175 GC_end_blocking(void) {
1177 LOCK(); /* This will block if the world is stopped. */
1178 me = GC_lookup_thread(pthread_self());
1179 GC_ASSERT(me -> thread_blocked);
1180 me -> thread_blocked = FALSE;
1184 /* A wrapper for the standard C sleep function */
1185 int WRAP_FUNC(sleep) (unsigned int seconds)
1189 GC_start_blocking();
1190 result = REAL_FUNC(sleep)(seconds);
1196 void *(*start_routine)(void *);
1199 sem_t registered; /* 1 ==> in our thread table, but */
1200 /* parent hasn't yet noticed. */
1203 /* Called at thread exit. */
1204 /* Never called for main thread. That's OK, since it */
1205 /* results in at most a tiny one-time leak. And */
1206 /* linuxthreads doesn't reclaim the main threads */
1207 /* resources or id anyway. */
1208 void GC_thread_exit_proc(void *arg)
1213 me = GC_lookup_thread(pthread_self());
1214 GC_destroy_thread_local(me);
1215 if (me -> flags & DETACHED) {
1216 GC_delete_thread(pthread_self());
1218 me -> flags |= FINISHED;
1220 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1221 && !defined(USE_HPUX_TLS) && !defined(DBG_HDRS_ALL)
1222 GC_remove_specific(GC_thread_key);
1224 if (GC_incremental && GC_collection_in_progress()) {
1225 int old_gc_no = GC_gc_no;
1227 /* Make sure that no part of our stack is still on the mark stack, */
1228 /* since it's about to be unmapped. */
1229 while (GC_incremental && GC_collection_in_progress()
1230 && old_gc_no == GC_gc_no) {
1232 GC_collect_a_little_inner(1);
1242 int WRAP_FUNC(pthread_join)(pthread_t thread, void **retval)
1245 GC_thread thread_gc_id;
1248 thread_gc_id = GC_lookup_thread(thread);
1249 /* This is guaranteed to be the intended one, since the thread id */
1250 /* cant have been recycled by pthreads. */
1252 result = REAL_FUNC(pthread_join)(thread, retval);
1255 /* Here the pthread thread id may have been recycled. */
1256 GC_delete_gc_thread(thread, thread_gc_id);
1263 WRAP_FUNC(pthread_detach)(pthread_t thread)
1266 GC_thread thread_gc_id;
1269 thread_gc_id = GC_lookup_thread(thread);
1271 result = REAL_FUNC(pthread_detach)(thread);
1274 thread_gc_id -> flags |= DETACHED;
1275 /* Here the pthread thread id may have been recycled. */
1276 if (thread_gc_id -> flags & FINISHED) {
1277 GC_delete_gc_thread(thread, thread_gc_id);
1284 void * GC_start_routine(void * arg)
1287 struct start_info * si = arg;
1290 pthread_t my_pthread;
1291 void *(*start)(void *);
1294 my_pthread = pthread_self();
1295 # ifdef DEBUG_THREADS
1296 GC_printf1("Starting thread 0x%lx\n", my_pthread);
1297 GC_printf1("pid = %ld\n", (long) getpid());
1298 GC_printf1("sp = 0x%lx\n", (long) &arg);
1301 me = GC_new_thread(my_pthread);
1302 me -> flags = si -> flags;
1303 me -> stack_ptr = 0;
1304 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1305 /* doesn't work because the stack base in /proc/self/stat is the */
1306 /* one for the main thread. There is a strong argument that that's */
1307 /* a kernel bug, but a pervasive one. */
1308 # ifdef STACK_GROWS_DOWN
1309 me -> stack_end = (ptr_t)(((word)(&dummy) + (GC_page_size - 1))
1310 & ~(GC_page_size - 1));
1311 me -> stack_ptr = me -> stack_end - 0x10;
1312 /* Needs to be plausible, since an asynchronous stack mark */
1313 /* should not crash. */
1315 me -> stack_end = (ptr_t)((word)(&dummy) & ~(GC_page_size - 1));
1316 me -> stack_ptr = me -> stack_end + 0x10;
1318 /* This is dubious, since we may be more than a page into the stack, */
1319 /* and hence skip some of it, though it's not clear that matters. */
1321 me -> backing_store_end = (ptr_t)
1322 (GC_save_regs_in_stack() & ~(GC_page_size - 1));
1323 /* This is also < 100% convincing. We should also read this */
1324 /* from /proc, but the hook to do so isn't there yet. */
1327 start = si -> start_routine;
1328 # ifdef DEBUG_THREADS
1329 GC_printf1("start_routine = 0x%lx\n", start);
1331 start_arg = si -> arg;
1332 sem_post(&(si -> registered)); /* Last action on si. */
1333 /* OK to deallocate. */
1334 pthread_cleanup_push(GC_thread_exit_proc, 0);
1335 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1337 GC_init_thread_local(me);
1340 result = (*start)(start_arg);
1342 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1344 me -> status = result;
1345 me -> flags |= FINISHED;
1346 pthread_cleanup_pop(1);
1347 /* Cleanup acquires lock, ensuring that we can't exit */
1348 /* while a collection that thinks we're alive is trying to stop */
1354 WRAP_FUNC(pthread_create)(pthread_t *new_thread,
1355 const pthread_attr_t *attr,
1356 void *(*start_routine)(void *), void *arg)
1360 pthread_t my_new_thread;
1363 struct start_info * si;
1364 /* This is otherwise saved only in an area mmapped by the thread */
1365 /* library, which isn't visible to the collector. */
1368 si = (struct start_info *)GC_INTERNAL_MALLOC(sizeof(struct start_info), NORMAL);
1370 if (!parallel_initialized) GC_init_parallel();
1371 if (0 == si) return(ENOMEM);
1372 sem_init(&(si -> registered), 0, 0);
1373 si -> start_routine = start_routine;
1376 if (!GC_thr_initialized) GC_thr_init();
1378 detachstate = PTHREAD_CREATE_JOINABLE;
1380 pthread_attr_getdetachstate(attr, &detachstate);
1382 if (PTHREAD_CREATE_DETACHED == detachstate) my_flags |= DETACHED;
1383 si -> flags = my_flags;
1385 # ifdef DEBUG_THREADS
1386 GC_printf1("About to start new thread from thread 0x%X\n",
1389 result = REAL_FUNC(pthread_create)(new_thread, attr, GC_start_routine, si);
1390 # ifdef DEBUG_THREADS
1391 GC_printf1("Started thread 0x%X\n", *new_thread);
1393 /* Wait until child has been added to the thread table. */
1394 /* This also ensures that we hold onto si until the child is done */
1395 /* with it. Thus it doesn't matter whether it is otherwise */
1396 /* visible to the collector. */
1397 while (0 != sem_wait(&(si -> registered))) {
1398 if (EINTR != errno) ABORT("sem_wait failed");
1400 sem_destroy(&(si -> registered));
1402 GC_INTERNAL_FREE(si);
1407 #ifdef GENERIC_COMPARE_AND_SWAP
1408 pthread_mutex_t GC_compare_and_swap_lock = PTHREAD_MUTEX_INITIALIZER;
1410 GC_bool GC_compare_and_exchange(volatile GC_word *addr,
1411 GC_word old, GC_word new_val)
1414 pthread_mutex_lock(&GC_compare_and_swap_lock);
1421 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1425 GC_word GC_atomic_add(volatile GC_word *addr, GC_word how_much)
1428 pthread_mutex_lock(&GC_compare_and_swap_lock);
1430 *addr = old + how_much;
1431 pthread_mutex_unlock(&GC_compare_and_swap_lock);
1435 #endif /* GENERIC_COMPARE_AND_SWAP */
1436 /* Spend a few cycles in a way that can't introduce contention with */
1437 /* othre threads. */
1441 volatile word dummy = 0;
1443 for (i = 0; i < 10; ++i) {
1445 __asm__ __volatile__ (" " : : : "memory");
1447 /* Something that's unlikely to be optimized away. */
1453 #define SPIN_MAX 1024 /* Maximum number of calls to GC_pause before */
1456 VOLATILE GC_bool GC_collecting = 0;
1457 /* A hint that we're in the collector and */
1458 /* holding the allocation lock for an */
1459 /* extended period. */
1461 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1462 /* If we don't want to use the below spinlock implementation, either */
1463 /* because we don't have a GC_test_and_set implementation, or because */
1464 /* we don't want to risk sleeping, we can still try spinning on */
1465 /* pthread_mutex_trylock for a while. This appears to be very */
1466 /* beneficial in many cases. */
1467 /* I suspect that under high contention this is nearly always better */
1468 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1469 /* Hence we still default to the spin lock. */
1470 /* This is also used to acquire the mark lock for the parallel */
1473 /* Here we use a strict exponential backoff scheme. I don't know */
1474 /* whether that's better or worse than the above. We eventually */
1475 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1476 /* explicitly sleep. */
1478 void GC_generic_lock(pthread_mutex_t * lock)
1480 unsigned pause_length = 1;
1483 if (0 == pthread_mutex_trylock(lock)) return;
1484 for (; pause_length <= SPIN_MAX; pause_length <<= 1) {
1485 for (i = 0; i < pause_length; ++i) {
1488 switch(pthread_mutex_trylock(lock)) {
1494 ABORT("Unexpected error from pthread_mutex_trylock");
1497 pthread_mutex_lock(lock);
1500 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1502 #if defined(USE_SPIN_LOCK)
1504 /* Reasonably fast spin locks. Basically the same implementation */
1505 /* as STL alloc.h. This isn't really the right way to do this. */
1506 /* but until the POSIX scheduling mess gets straightened out ... */
1508 volatile unsigned int GC_allocate_lock = 0;
1513 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1514 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1515 static unsigned spin_max = low_spin_max;
1516 unsigned my_spin_max;
1517 static unsigned last_spins = 0;
1518 unsigned my_last_spins;
1521 if (!GC_test_and_set(&GC_allocate_lock)) {
1524 my_spin_max = spin_max;
1525 my_last_spins = last_spins;
1526 for (i = 0; i < my_spin_max; i++) {
1527 if (GC_collecting || GC_nprocs == 1) goto yield;
1528 if (i < my_last_spins/2 || GC_allocate_lock) {
1532 if (!GC_test_and_set(&GC_allocate_lock)) {
1535 * Spinning worked. Thus we're probably not being scheduled
1536 * against the other process with which we were contending.
1537 * Thus it makes sense to spin longer the next time.
1540 spin_max = high_spin_max;
1544 /* We are probably being scheduled against the other process. Sleep. */
1545 spin_max = low_spin_max;
1548 if (!GC_test_and_set(&GC_allocate_lock)) {
1551 # define SLEEP_THRESHOLD 12
1552 /* nanosleep(<= 2ms) just spins under Linux. We */
1553 /* want to be careful to avoid that behavior. */
1554 if (i < SLEEP_THRESHOLD) {
1560 /* Don't wait for more than about 15msecs, even */
1561 /* under extreme contention. */
1563 ts.tv_nsec = 1 << i;
1569 #else /* !USE_SPINLOCK */
1573 if (1 == GC_nprocs || GC_collecting) {
1574 pthread_mutex_lock(&GC_allocate_ml);
1576 GC_generic_lock(&GC_allocate_ml);
1580 #endif /* !USE_SPINLOCK */
1582 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1584 #ifdef GC_ASSERTIONS
1585 pthread_t GC_mark_lock_holder = NO_THREAD;
1589 /* Ugly workaround for a linux threads bug in the final versions */
1590 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1591 /* field even when it fails to acquire the mutex. This causes */
1592 /* pthread_cond_wait to die. Remove for glibc2.2. */
1593 /* According to the man page, we should use */
1594 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1596 static pthread_mutex_t mark_mutex =
1597 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP, {0, 0}};
1599 static pthread_mutex_t mark_mutex = PTHREAD_MUTEX_INITIALIZER;
1602 static pthread_cond_t builder_cv = PTHREAD_COND_INITIALIZER;
1604 void GC_acquire_mark_lock()
1607 if (pthread_mutex_lock(&mark_mutex) != 0) {
1608 ABORT("pthread_mutex_lock failed");
1611 GC_generic_lock(&mark_mutex);
1612 # ifdef GC_ASSERTIONS
1613 GC_mark_lock_holder = pthread_self();
1617 void GC_release_mark_lock()
1619 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1620 # ifdef GC_ASSERTIONS
1621 GC_mark_lock_holder = NO_THREAD;
1623 if (pthread_mutex_unlock(&mark_mutex) != 0) {
1624 ABORT("pthread_mutex_unlock failed");
1628 /* Collector must wait for a freelist builders for 2 reasons: */
1629 /* 1) Mark bits may still be getting examined without lock. */
1630 /* 2) Partial free lists referenced only by locals may not be scanned */
1631 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1632 /* free-list link may be ignored. */
1633 void GC_wait_builder()
1635 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1636 # ifdef GC_ASSERTIONS
1637 GC_mark_lock_holder = NO_THREAD;
1639 if (pthread_cond_wait(&builder_cv, &mark_mutex) != 0) {
1640 ABORT("pthread_cond_wait failed");
1642 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1643 # ifdef GC_ASSERTIONS
1644 GC_mark_lock_holder = pthread_self();
1648 void GC_wait_for_reclaim()
1650 GC_acquire_mark_lock();
1651 while (GC_fl_builder_count > 0) {
1654 GC_release_mark_lock();
1657 void GC_notify_all_builder()
1659 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1660 if (pthread_cond_broadcast(&builder_cv) != 0) {
1661 ABORT("pthread_cond_broadcast failed");
1665 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1667 #ifdef PARALLEL_MARK
1669 static pthread_cond_t mark_cv = PTHREAD_COND_INITIALIZER;
1671 void GC_wait_marker()
1673 GC_ASSERT(GC_mark_lock_holder == pthread_self());
1674 # ifdef GC_ASSERTIONS
1675 GC_mark_lock_holder = NO_THREAD;
1677 if (pthread_cond_wait(&mark_cv, &mark_mutex) != 0) {
1678 ABORT("pthread_cond_wait failed");
1680 GC_ASSERT(GC_mark_lock_holder == NO_THREAD);
1681 # ifdef GC_ASSERTIONS
1682 GC_mark_lock_holder = pthread_self();
1686 void GC_notify_all_marker()
1688 if (pthread_cond_broadcast(&mark_cv) != 0) {
1689 ABORT("pthread_cond_broadcast failed");
1693 #endif /* PARALLEL_MARK */
1695 # endif /* LINUX_THREADS */