2 * Copyright (c) 1994 by Xerox Corporation. All rights reserved.
3 * Copyright (c) 1996 by Silicon Graphics. All rights reserved.
4 * Copyright (c) 1998 by Fergus Henderson. All rights reserved.
5 * Copyright (c) 2000-2004 by Hewlett-Packard Company. All rights reserved.
7 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
8 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
10 * Permission is hereby granted to use or copy this program
11 * for any purpose, provided the above notices are retained on all copies.
12 * Permission to modify the code and to distribute modified code is granted,
13 * provided the above notices are retained, and a notice that the code was
14 * modified is included with the above copyright notice.
17 * Support code for LinuxThreads, the clone()-based kernel
18 * thread package for Linux which is included in libc6.
20 * This code relies on implementation details of LinuxThreads,
21 * (i.e. properties not guaranteed by the Pthread standard),
22 * though this version now does less of that than the other Pthreads
25 * Note that there is a lot of code duplication between linux_threads.c
26 * and thread support for some of the other Posix platforms; any changes
27 * made here may need to be reflected there too.
29 /* DG/UX ix86 support <takis@xfree86.org> */
31 * Linux_threads.c now also includes some code to support HPUX and
32 * OSF1 (Compaq Tru64 Unix, really). The OSF1 support is based on Eric Benson's
35 * Eric also suggested an alternate basis for a lock implementation in
37 * + #elif defined(OSF1)
38 * + unsigned long GC_allocate_lock = 0;
39 * + msemaphore GC_allocate_semaphore;
40 * + # define GC_TRY_LOCK() \
41 * + ((msem_lock(&GC_allocate_semaphore, MSEM_IF_NOWAIT) == 0) \
42 * + ? (GC_allocate_lock = 1) \
44 * + # define GC_LOCK_TAKEN GC_allocate_lock
47 /*#define DEBUG_THREADS 1*/
48 /*#define GC_ASSERTIONS*/
50 # include "private/pthread_support.h"
52 # if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
53 && !defined(GC_WIN32_THREADS)
55 # if defined(GC_HPUX_THREADS) && !defined(USE_PTHREAD_SPECIFIC) \
56 && !defined(USE_COMPILER_TLS)
58 # define USE_PTHREAD_SPECIFIC
59 /* Empirically, as of gcc 3.3, USE_COMPILER_TLS doesn't work. */
61 # define USE_COMPILER_TLS
65 # if defined USE_HPUX_TLS
66 --> Macro replaced by USE_COMPILER_TLS
69 # if (defined(GC_DGUX386_THREADS) || defined(GC_OSF1_THREADS) || \
70 defined(GC_DARWIN_THREADS) || defined(GC_AIX_THREADS)) || \
71 defined(GC_NETBSD_THREADS) && !defined(USE_PTHREAD_SPECIFIC)
72 # define USE_PTHREAD_SPECIFIC
75 # if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
76 # define _POSIX4A_DRAFT10_SOURCE 1
79 # if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
80 # define _USING_POSIX4A_DRAFT10 1
83 # ifdef THREAD_LOCAL_ALLOC
84 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
85 # include "private/specific.h"
88 /* Note that these macros should be used only to get/set the GC_thread pointer.
89 * We need to use both tls and pthread because we use the pthread_create function hook to
90 * free the data for foreign threads. When that doesn't happen, libgc could have old
91 * pthread_t that get reused...
93 # if defined(USE_PTHREAD_SPECIFIC)
94 # define GC_getspecific pthread_getspecific
95 # define GC_setspecific pthread_setspecific
96 # define GC_key_create pthread_key_create
97 typedef pthread_key_t GC_key_t
;
99 # if defined(USE_COMPILER_TLS)
100 /* Note sles9 gcc on powerpc gets confused by the define to set GC_thread_tls and pthread_setspecific
101 * so we actually use a static inline function decalred below that is equivalent to:
102 * define GC_setspecific(key, v) (GC_thread_tls = (v), pthread_setspecific ((key), (v)))
104 # define GC_getspecific(x) (GC_thread_tls)
105 # define GC_key_create pthread_key_create
106 typedef pthread_key_t GC_key_t
;
110 # include <pthread.h>
115 # include <sys/mman.h>
116 # include <sys/time.h>
117 # include <sys/types.h>
118 # include <sys/stat.h>
122 #if defined(GC_DARWIN_THREADS)
123 # include "private/darwin_semaphore.h"
125 # include <semaphore.h>
126 #endif /* !GC_DARWIN_THREADS */
128 #if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
129 # include <sys/sysctl.h>
130 #endif /* GC_DARWIN_THREADS */
132 #if defined(GC_NETBSD_THREADS)
133 # include <sys/param.h>
134 # include <sys/sysctl.h>
139 #if defined(GC_DGUX386_THREADS)
140 # include <sys/dg_sys_info.h>
141 # include <sys/_int_psem.h>
142 /* sem_t is an uint in DG/UX */
143 typedef unsigned int sem_t
;
144 #endif /* GC_DGUX386_THREADS */
150 #ifdef GC_USE_LD_WRAP
151 # define WRAP_FUNC(f) __wrap_##f
152 # define REAL_FUNC(f) __real_##f
154 # define WRAP_FUNC(f) GC_##f
155 # if !defined(GC_DGUX386_THREADS)
156 # define REAL_FUNC(f) f
157 # else /* GC_DGUX386_THREADS */
158 # define REAL_FUNC(f) __d10_##f
159 # endif /* GC_DGUX386_THREADS */
160 # undef pthread_create
161 # if !defined(GC_DARWIN_THREADS)
162 # undef pthread_sigmask
165 # undef pthread_detach
166 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
167 && !defined(_PTHREAD_USE_PTDNAM_)
168 /* Restore the original mangled names on Tru64 UNIX. */
169 # define pthread_create __pthread_create
170 # define pthread_join __pthread_join
171 # define pthread_detach __pthread_detach
177 static GC_bool parallel_initialized
= FALSE
;
179 void GC_init_parallel();
181 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
183 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
185 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
186 #undef PACKAGE_BUGREPORT
188 #undef PACKAGE_STRING
189 #undef PACKAGE_TARNAME
190 #undef PACKAGE_VERSION
191 #include "mono/utils/mono-compiler.h"
194 GC_key_t GC_thread_key
;
196 #ifdef USE_COMPILER_TLS
197 __thread MONO_TLS_FAST
void* GC_thread_tls
;
200 * gcc errors out with /tmp/ccdPMFuq.s:2994: Error: symbol `.LTLS4' is already defined
201 * if the inline is added on powerpc
203 #if !defined(__ppc__) && !defined(__powerpc__)
206 static int GC_setspecific (GC_key_t key
, void *value
) {
207 GC_thread_tls
= value
;
208 return pthread_setspecific (key
, value
);
212 static GC_bool keys_initialized
;
214 #ifdef MONO_DEBUGGER_SUPPORTED
215 #include "include/libgc-mono-debugger.h"
218 /* Recover the contents of the freelist array fl into the global one gfl.*/
219 /* Note that the indexing scheme differs, in that gfl has finer size */
220 /* resolution, even if not all entries are used. */
221 /* We hold the allocator lock. */
222 static void return_freelists(ptr_t
*fl
, ptr_t
*gfl
)
228 for (i
= 1; i
< NFREELISTS
; ++i
) {
229 nwords
= i
* (GRANULARITY
/sizeof(word
));
232 if ((word
)q
>= HBLKSIZE
) {
233 if (gfl
[nwords
] == 0) {
237 for (; (word
)q
>= HBLKSIZE
; qptr
= &(obj_link(q
)), q
= *qptr
);
243 /* Clear fl[i], since the thread structure may hang around. */
244 /* Do it in a way that is likely to trap if we access it. */
245 fl
[i
] = (ptr_t
)HBLKSIZE
;
249 /* We statically allocate a single "size 0" object. It is linked to */
250 /* itself, and is thus repeatedly reused for all size 0 allocation */
251 /* requests. (Size 0 gcj allocation requests are incorrect, and */
252 /* we arrange for those to fault asap.) */
253 static ptr_t size_zero_object
= (ptr_t
)(&size_zero_object
);
255 void GC_delete_gc_thread(pthread_t id
, GC_thread gct
);
256 void GC_destroy_thread_local(GC_thread p
);
258 void GC_thread_deregister_foreign (void *data
)
260 GC_thread me
= (GC_thread
)data
;
261 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
262 if (me
-> flags
& FOREIGN_THREAD
) {
264 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
265 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
266 GC_destroy_thread_local (me
);
268 GC_delete_gc_thread(me
->id
, me
);
273 /* Each thread structure must be initialized. */
274 /* This call must be made from the new thread. */
275 /* Caller holds allocation lock. */
276 void GC_init_thread_local(GC_thread p
)
280 if (!keys_initialized
) {
281 if (0 != GC_key_create(&GC_thread_key
, GC_thread_deregister_foreign
)) {
282 ABORT("Failed to create key for local allocator");
284 keys_initialized
= TRUE
;
286 if (0 != GC_setspecific(GC_thread_key
, p
)) {
287 ABORT("Failed to set thread specific allocation pointers");
289 for (i
= 1; i
< NFREELISTS
; ++i
) {
290 p
-> ptrfree_freelists
[i
] = (ptr_t
)1;
291 p
-> normal_freelists
[i
] = (ptr_t
)1;
292 # ifdef GC_GCJ_SUPPORT
293 p
-> gcj_freelists
[i
] = (ptr_t
)1;
296 /* Set up the size 0 free lists. */
297 p
-> ptrfree_freelists
[0] = (ptr_t
)(&size_zero_object
);
298 p
-> normal_freelists
[0] = (ptr_t
)(&size_zero_object
);
299 # ifdef GC_GCJ_SUPPORT
300 p
-> gcj_freelists
[0] = (ptr_t
)(-1);
304 #ifdef GC_GCJ_SUPPORT
305 extern ptr_t
* GC_gcjobjfreelist
;
308 /* We hold the allocator lock. */
309 void GC_destroy_thread_local(GC_thread p
)
311 /* We currently only do this from the thread itself or from */
312 /* the fork handler for a child process. */
314 GC_ASSERT(GC_getspecific(GC_thread_key
) == (void *)p
);
316 return_freelists(p
-> ptrfree_freelists
, GC_aobjfreelist
);
317 return_freelists(p
-> normal_freelists
, GC_objfreelist
);
318 # ifdef GC_GCJ_SUPPORT
319 return_freelists(p
-> gcj_freelists
, GC_gcjobjfreelist
);
323 extern GC_PTR
GC_generic_malloc_many();
325 GC_PTR
GC_local_malloc(size_t bytes
)
327 if (EXPECT(!SMALL_ENOUGH(bytes
),0)) {
328 return(GC_malloc(bytes
));
330 int index
= INDEX_FROM_BYTES(bytes
);
333 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
334 GC_key_t k
= GC_thread_key
;
338 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
339 if (EXPECT(0 == k
, 0)) {
340 /* This can happen if we get called when the world is */
341 /* being initialized. Whether we can actually complete */
342 /* the initialization then is unclear. */
347 tsd
= GC_getspecific(GC_thread_key
);
348 # ifdef GC_ASSERTIONS
350 GC_ASSERT(tsd
== (void *)GC_lookup_thread(pthread_self()));
353 my_fl
= ((GC_thread
)tsd
) -> normal_freelists
+ index
;
355 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
356 ptr_t next
= obj_link(my_entry
);
357 GC_PTR result
= (GC_PTR
)my_entry
;
359 obj_link(my_entry
) = 0;
360 PREFETCH_FOR_WRITE(next
);
362 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
363 *my_fl
= my_entry
+ index
+ 1;
364 return GC_malloc(bytes
);
366 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), NORMAL
, my_fl
);
367 if (*my_fl
== 0) return GC_oom_fn(bytes
);
368 return GC_local_malloc(bytes
);
373 GC_PTR
GC_local_malloc_atomic(size_t bytes
)
375 if (EXPECT(!SMALL_ENOUGH(bytes
), 0)) {
376 return(GC_malloc_atomic(bytes
));
378 int index
= INDEX_FROM_BYTES(bytes
);
379 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
380 -> ptrfree_freelists
+ index
;
381 ptr_t my_entry
= *my_fl
;
383 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
384 GC_PTR result
= (GC_PTR
)my_entry
;
385 *my_fl
= obj_link(my_entry
);
387 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
388 *my_fl
= my_entry
+ index
+ 1;
389 return GC_malloc_atomic(bytes
);
391 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), PTRFREE
, my_fl
);
392 /* *my_fl is updated while the collector is excluded; */
393 /* the free list is always visible to the collector as */
395 if (*my_fl
== 0) return GC_oom_fn(bytes
);
396 return GC_local_malloc_atomic(bytes
);
401 #ifdef GC_GCJ_SUPPORT
403 #include "include/gc_gcj.h"
406 extern GC_bool GC_gcj_malloc_initialized
;
409 extern int GC_gcj_kind
;
411 GC_PTR
GC_local_gcj_malloc(size_t bytes
,
412 void * ptr_to_struct_containing_descr
)
414 GC_ASSERT(GC_gcj_malloc_initialized
);
415 if (EXPECT(!SMALL_ENOUGH(bytes
), 0)) {
416 return GC_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
418 int index
= INDEX_FROM_BYTES(bytes
);
419 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
420 -> gcj_freelists
+ index
;
421 ptr_t my_entry
= *my_fl
;
422 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
423 GC_PTR result
= (GC_PTR
)my_entry
;
424 GC_ASSERT(!GC_incremental
);
425 /* We assert that any concurrent marker will stop us. */
426 /* Thus it is impossible for a mark procedure to see the */
427 /* allocation of the next object, but to see this object */
428 /* still containing a free list pointer. Otherwise the */
429 /* marker might find a random "mark descriptor". */
430 *(volatile ptr_t
*)my_fl
= obj_link(my_entry
);
431 /* We must update the freelist before we store the pointer. */
432 /* Otherwise a GC at this point would see a corrupted */
434 /* A memory barrier is probably never needed, since the */
435 /* action of stopping this thread will cause prior writes */
437 GC_ASSERT(((void * volatile *)result
)[1] == 0);
438 *(void * volatile *)result
= ptr_to_struct_containing_descr
;
440 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
441 if (!GC_incremental
) *my_fl
= my_entry
+ index
+ 1;
442 /* In the incremental case, we always have to take this */
443 /* path. Thus we leave the counter alone. */
444 return GC_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
446 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), GC_gcj_kind
, my_fl
);
447 if (*my_fl
== 0) return GC_oom_fn(bytes
);
448 return GC_local_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
453 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
454 /* adjust it. The size is assumed to be such that it can be */
455 /* allocated as a small object. */
456 void * GC_local_gcj_fast_malloc(size_t lw
, void * ptr_to_struct_containing_descr
)
458 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
459 -> gcj_freelists
+ lw
;
460 ptr_t my_entry
= *my_fl
;
462 GC_ASSERT(GC_gcj_malloc_initialized
);
464 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
465 GC_PTR result
= (GC_PTR
)my_entry
;
466 GC_ASSERT(!GC_incremental
);
467 /* We assert that any concurrent marker will stop us. */
468 /* Thus it is impossible for a mark procedure to see the */
469 /* allocation of the next object, but to see this object */
470 /* still containing a free list pointer. Otherwise the */
471 /* marker might find a random "mark descriptor". */
472 *(volatile ptr_t
*)my_fl
= obj_link(my_entry
);
473 /* We must update the freelist before we store the pointer. */
474 /* Otherwise a GC at this point would see a corrupted */
476 /* A memory barrier is probably never needed, since the */
477 /* action of stopping this thread will cause prior writes */
479 GC_ASSERT(((void * volatile *)result
)[1] == 0);
480 *(void * volatile *)result
= ptr_to_struct_containing_descr
;
482 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
483 if (!GC_incremental
) *my_fl
= my_entry
+ lw
+ 1;
484 /* In the incremental case, we always have to take this */
485 /* path. Thus we leave the counter alone. */
486 return GC_gcj_fast_malloc(lw
, ptr_to_struct_containing_descr
);
488 GC_generic_malloc_many(BYTES_FROM_INDEX(lw
), GC_gcj_kind
, my_fl
);
489 if (*my_fl
== 0) return GC_oom_fn(BYTES_FROM_INDEX(lw
));
490 return GC_local_gcj_fast_malloc(lw
, ptr_to_struct_containing_descr
);
494 #endif /* GC_GCJ_SUPPORT */
496 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
498 # define GC_destroy_thread_local(t)
500 # endif /* !THREAD_LOCAL_ALLOC */
504 To make sure that we're using LinuxThreads and not some other thread
505 package, we generate a dummy reference to `pthread_kill_other_threads_np'
506 (was `__pthread_initial_thread_bos' but that disappeared),
507 which is a symbol defined in LinuxThreads, but (hopefully) not in other
510 We no longer do this, since this code is now portable enough that it might
511 actually work for something else.
513 void (*dummy_var_to_force_linux_threads
)() = pthread_kill_other_threads_np
;
516 long GC_nprocs
= 1; /* Number of processors. We may not have */
517 /* access to all of them, but this is as good */
518 /* a guess as any ... */
523 # define MAX_MARKERS 16
526 static ptr_t marker_sp
[MAX_MARKERS
] = {0};
528 void * GC_mark_thread(void * id
)
532 marker_sp
[(word
)id
] = GC_approx_sp();
533 for (;; ++my_mark_no
) {
534 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
535 /* promptly. This is important if it were called from the signal */
536 /* handler or from the GC lock acquisition code. Under Linux, it's */
537 /* not safe to call it from a signal handler, since it uses mutexes */
538 /* and condition variables. Since it is called only here, the */
539 /* argument is unnecessary. */
540 if (my_mark_no
< GC_mark_no
|| my_mark_no
> GC_mark_no
+ 2) {
541 /* resynchronize if we get far off, e.g. because GC_mark_no */
543 my_mark_no
= GC_mark_no
;
545 # ifdef DEBUG_THREADS
546 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no
);
548 GC_help_marker(my_mark_no
);
552 extern long GC_markers
; /* Number of mark threads we would */
553 /* like to have. Includes the */
554 /* initiating thread. */
556 pthread_t GC_mark_threads
[MAX_MARKERS
];
558 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
560 static void start_mark_threads()
565 if (GC_markers
> MAX_MARKERS
) {
566 WARN("Limiting number of mark threads\n", 0);
567 GC_markers
= MAX_MARKERS
;
569 if (0 != pthread_attr_init(&attr
)) ABORT("pthread_attr_init failed");
571 if (0 != pthread_attr_setdetachstate(&attr
, PTHREAD_CREATE_DETACHED
))
572 ABORT("pthread_attr_setdetachstate failed");
574 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
575 /* Default stack size is usually too small: fix it. */
576 /* Otherwise marker threads or GC may run out of */
578 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
583 if (pthread_attr_getstacksize(&attr
, &old_size
) != 0)
584 ABORT("pthread_attr_getstacksize failed\n");
585 if (old_size
< MIN_STACK_SIZE
) {
586 if (pthread_attr_setstacksize(&attr
, MIN_STACK_SIZE
) != 0)
587 ABORT("pthread_attr_setstacksize failed\n");
590 # endif /* HPUX || GC_DGUX386_THREADS */
592 if (GC_print_stats
) {
593 GC_printf1("Starting %ld marker threads\n", GC_markers
- 1);
596 for (i
= 0; i
< GC_markers
- 1; ++i
) {
597 if (0 != PTHREAD_CREATE(GC_mark_threads
+ i
, &attr
,
598 GC_mark_thread
, (void *)(word
)i
)) {
599 WARN("Marker thread creation failed, errno = %ld.\n", errno
);
604 #else /* !PARALLEL_MARK */
606 static __inline__
void start_mark_threads()
610 #endif /* !PARALLEL_MARK */
612 GC_bool GC_thr_initialized
= FALSE
;
614 volatile GC_thread GC_threads
[THREAD_TABLE_SZ
];
617 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
618 * put it into a separate function.
620 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
621 static __attribute__((noinline
)) unsigned char* get_gc_thread_key_addr
GC_PROTO((void))
623 return (unsigned char*)&GC_thread_key
;
626 void GC_push_thread_structures
GC_PROTO((void))
628 GC_push_all((ptr_t
)(GC_threads
), (ptr_t
)(GC_threads
)+sizeof(GC_threads
));
629 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
630 GC_push_all((ptr_t
)get_gc_thread_key_addr(),
631 (ptr_t
)(get_gc_thread_key_addr())+sizeof(&GC_thread_key
));
637 void GC_push_thread_structures
GC_PROTO((void))
639 GC_push_all((ptr_t
)(GC_threads
), (ptr_t
)(GC_threads
)+sizeof(GC_threads
));
640 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
641 GC_push_all((ptr_t
)(&GC_thread_key
),
642 (ptr_t
)(&GC_thread_key
)+sizeof(&GC_thread_key
));
648 #ifdef THREAD_LOCAL_ALLOC
649 /* We must explicitly mark ptrfree and gcj free lists, since the free */
650 /* list links wouldn't otherwise be found. We also set them in the */
651 /* normal free lists, since that involves touching less memory than if */
652 /* we scanned them normally. */
653 void GC_mark_thread_local_free_lists(void)
659 for (i
= 0; i
< THREAD_TABLE_SZ
; ++i
) {
660 for (p
= GC_threads
[i
]; 0 != p
; p
= p
-> next
) {
661 for (j
= 1; j
< NFREELISTS
; ++j
) {
662 q
= p
-> ptrfree_freelists
[j
];
663 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
664 q
= p
-> normal_freelists
[j
];
665 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
666 # ifdef GC_GCJ_SUPPORT
667 q
= p
-> gcj_freelists
[j
];
668 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
669 # endif /* GC_GCJ_SUPPORT */
674 #endif /* THREAD_LOCAL_ALLOC */
676 static struct GC_Thread_Rep first_thread
;
678 /* Add a thread to GC_threads. We assume it wasn't already there. */
679 /* Caller holds allocation lock. */
680 GC_thread
GC_new_thread(pthread_t id
)
682 int hv
= ((word
)id
) % THREAD_TABLE_SZ
;
684 static GC_bool first_thread_used
= FALSE
;
686 if (!first_thread_used
) {
687 result
= &first_thread
;
688 first_thread_used
= TRUE
;
690 result
= (struct GC_Thread_Rep
*)
691 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep
), NORMAL
);
693 if (result
== 0) return(0);
695 result
-> next
= GC_threads
[hv
];
696 GC_threads
[hv
] = result
;
697 GC_ASSERT(result
-> flags
== 0 && result
-> thread_blocked
== 0);
701 /* Delete a thread from GC_threads. We assume it is there. */
702 /* (The code intentionally traps if it wasn't.) */
703 /* Caller holds allocation lock. */
704 void GC_delete_thread(pthread_t id
)
706 int hv
= ((word
)id
) % THREAD_TABLE_SZ
;
707 register GC_thread p
= GC_threads
[hv
];
708 register GC_thread prev
= 0;
710 while (!pthread_equal(p
-> id
, id
)) {
715 GC_threads
[hv
] = p
-> next
;
717 prev
-> next
= p
-> next
;
719 #ifdef MONO_DEBUGGER_SUPPORTED
720 if (gc_thread_vtable
&& gc_thread_vtable
->thread_exited
)
721 gc_thread_vtable
->thread_exited (id
, &p
->stop_info
.stack_ptr
);
724 #ifdef GC_DARWIN_THREADS
725 mach_port_deallocate(mach_task_self(), p
->stop_info
.mach_thread
);
731 /* If a thread has been joined, but we have not yet */
732 /* been notified, then there may be more than one thread */
733 /* in the table with the same pthread id. */
734 /* This is OK, but we need a way to delete a specific one. */
735 void GC_delete_gc_thread(pthread_t id
, GC_thread gc_id
)
737 int hv
= ((word
)id
) % THREAD_TABLE_SZ
;
738 register GC_thread p
= GC_threads
[hv
];
739 register GC_thread prev
= 0;
746 GC_threads
[hv
] = p
-> next
;
748 prev
-> next
= p
-> next
;
751 #ifdef GC_DARWIN_THREADS
752 mach_port_deallocate(mach_task_self(), p
->stop_info
.mach_thread
);
758 /* Return a GC_thread corresponding to a given pthread_t. */
759 /* Returns 0 if it's not there. */
760 /* Caller holds allocation lock or otherwise inhibits */
762 /* If there is more than one thread with the given id we */
763 /* return the most recent one. */
764 GC_thread
GC_lookup_thread(pthread_t id
)
766 int hv
= ((word
)id
) % THREAD_TABLE_SZ
;
767 register GC_thread p
= GC_threads
[hv
];
769 while (p
!= 0 && !pthread_equal(p
-> id
, id
)) p
= p
-> next
;
773 int GC_thread_is_registered (void)
778 ptr
= (void *)GC_lookup_thread(pthread_self());
785 /* Remove all entries from the GC_threads table, except the */
786 /* one for the current thread. We need to do this in the child */
787 /* process after a fork(), since only the current thread */
788 /* survives in the child. */
789 void GC_remove_all_threads_but_me(void)
791 pthread_t self
= pthread_self();
793 GC_thread p
, next
, me
;
795 for (hv
= 0; hv
< THREAD_TABLE_SZ
; ++hv
) {
797 for (p
= GC_threads
[hv
]; 0 != p
; p
= next
) {
799 if (p
-> id
== self
) {
803 # ifdef THREAD_LOCAL_ALLOC
804 if (!(p
-> flags
& FINISHED
)) {
805 GC_destroy_thread_local(p
);
807 # endif /* THREAD_LOCAL_ALLOC */
808 if (p
!= &first_thread
) GC_INTERNAL_FREE(p
);
815 #endif /* HANDLE_FORK */
817 #ifdef USE_PROC_FOR_LIBRARIES
818 int GC_segment_is_thread_stack(ptr_t lo
, ptr_t hi
)
823 # ifdef PARALLEL_MARK
824 for (i
= 0; i
< GC_markers
; ++i
) {
825 if (marker_sp
[i
] > lo
& marker_sp
[i
] < hi
) return 1;
828 for (i
= 0; i
< THREAD_TABLE_SZ
; i
++) {
829 for (p
= GC_threads
[i
]; p
!= 0; p
= p
-> next
) {
830 if (0 != p
-> stack_end
) {
831 # ifdef STACK_GROWS_UP
832 if (p
-> stack_end
>= lo
&& p
-> stack_end
< hi
) return 1;
833 # else /* STACK_GROWS_DOWN */
834 if (p
-> stack_end
> lo
&& p
-> stack_end
<= hi
) return 1;
841 #endif /* USE_PROC_FOR_LIBRARIES */
843 #ifdef GC_LINUX_THREADS
844 /* Return the number of processors, or i<= 0 if it can't be determined. */
847 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
848 /* appears to be buggy in many cases. */
849 /* We look for lines "cpu<n>" in /proc/stat. */
850 # define STAT_BUF_SIZE 4096
851 # define STAT_READ read
852 /* If read is wrapped, this may need to be redefined to call */
854 char stat_buf
[STAT_BUF_SIZE
];
857 /* Some old kernels only have a single "cpu nnnn ..." */
858 /* entry in /proc/stat. We identify those as */
862 f
= open("/proc/stat", O_RDONLY
);
863 if (f
< 0 || (len
= STAT_READ(f
, stat_buf
, STAT_BUF_SIZE
)) < 100) {
864 WARN("Couldn't read /proc/stat\n", 0);
867 for (i
= 0; i
< len
- 100; ++i
) {
868 if (stat_buf
[i
] == '\n' && stat_buf
[i
+1] == 'c'
869 && stat_buf
[i
+2] == 'p' && stat_buf
[i
+3] == 'u') {
870 int cpu_no
= atoi(stat_buf
+ i
+ 4);
871 if (cpu_no
>= result
) result
= cpu_no
+ 1;
877 #endif /* GC_LINUX_THREADS */
879 /* We hold the GC lock. Wait until an in-progress GC has finished. */
880 /* Repeatedly RELEASES GC LOCK in order to wait. */
881 /* If wait_for_all is true, then we exit with the GC lock held and no */
882 /* collection in progress; otherwise we just wait for the current GC */
884 extern GC_bool
GC_collection_in_progress();
885 void GC_wait_for_gc_completion(GC_bool wait_for_all
)
887 if (GC_incremental
&& GC_collection_in_progress()) {
888 int old_gc_no
= GC_gc_no
;
890 /* Make sure that no part of our stack is still on the mark stack, */
891 /* since it's about to be unmapped. */
892 while (GC_incremental
&& GC_collection_in_progress()
893 && (wait_for_all
|| old_gc_no
== GC_gc_no
)) {
895 GC_in_thread_creation
= TRUE
;
896 GC_collect_a_little_inner(1);
897 GC_in_thread_creation
= FALSE
;
907 /* Procedures called before and after a fork. The goal here is to make */
908 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
909 /* attainable, since the single UNIX spec seems to imply that one */
910 /* should only call async-signal-safe functions, and we probably can't */
911 /* quite guarantee that. But we give it our best shot. (That same */
912 /* spec also implies that it's not safe to call the system malloc */
913 /* between fork() and exec(). Thus we're doing no worse than it. */
915 /* Called before a fork() */
916 void GC_fork_prepare_proc(void)
918 /* Acquire all relevant locks, so that after releasing the locks */
919 /* the child will see a consistent state in which monitor */
920 /* invariants hold. Unfortunately, we can't acquire libc locks */
921 /* we might need, and there seems to be no guarantee that libc */
922 /* must install a suitable fork handler. */
923 /* Wait for an ongoing GC to finish, since we can't finish it in */
924 /* the (one remaining thread in) the child. */
926 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
927 GC_wait_for_reclaim();
929 GC_wait_for_gc_completion(TRUE
);
930 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
931 GC_acquire_mark_lock();
935 /* Called in parent after a fork() */
936 void GC_fork_parent_proc(void)
938 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
939 GC_release_mark_lock();
944 /* Called in child after a fork() */
945 void GC_fork_child_proc(void)
947 /* Clean up the thread table, so that just our thread is left. */
948 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
949 GC_release_mark_lock();
951 GC_remove_all_threads_but_me();
952 # ifdef PARALLEL_MARK
953 /* Turn off parallel marking in the child, since we are probably */
954 /* just going to exec, and we would have to restart mark threads. */
957 # endif /* PARALLEL_MARK */
960 #endif /* HANDLE_FORK */
962 #if defined(GC_DGUX386_THREADS)
963 /* Return the number of processors, or i<= 0 if it can't be determined. */
966 /* <takis@XFree86.Org> */
968 struct dg_sys_info_pm_info pm_sysinfo
;
971 status
= dg_sys_info((long int *) &pm_sysinfo
,
972 DG_SYS_INFO_PM_INFO_TYPE
, DG_SYS_INFO_PM_CURRENT_VERSION
);
974 /* set -1 for error */
978 numCpus
= pm_sysinfo
.idle_vp_count
;
980 # ifdef DEBUG_THREADS
981 GC_printf1("Number of active CPUs in this system: %d\n", numCpus
);
985 #endif /* GC_DGUX386_THREADS */
987 /* We hold the allocation lock. */
990 # ifndef GC_DARWIN_THREADS
995 if (GC_thr_initialized
) return;
996 GC_thr_initialized
= TRUE
;
999 /* Prepare for a possible fork. */
1000 pthread_atfork(GC_fork_prepare_proc
, GC_fork_parent_proc
,
1001 GC_fork_child_proc
);
1002 # endif /* HANDLE_FORK */
1003 /* Add the initial thread, so we can stop it. */
1004 t
= GC_new_thread(pthread_self());
1005 # ifdef GC_DARWIN_THREADS
1006 t
-> stop_info
.mach_thread
= mach_thread_self();
1008 t
-> stop_info
.stack_ptr
= (ptr_t
)(&dummy
);
1010 t
-> flags
= DETACHED
| MAIN_THREAD
;
1011 #ifdef MONO_DEBUGGER_SUPPORTED
1012 if (gc_thread_vtable
&& gc_thread_vtable
->thread_created
)
1013 # ifdef GC_DARWIN_THREADS
1014 gc_thread_vtable
->thread_created (mach_thread_self (), &t
->stop_info
.stack_ptr
);
1016 gc_thread_vtable
->thread_created (pthread_self (), &t
->stop_info
.stack_ptr
);
1022 /* Set GC_nprocs. */
1024 char * nprocs_string
= GETENV("GC_NPROCS");
1026 if (nprocs_string
!= NULL
) GC_nprocs
= atoi(nprocs_string
);
1028 if (GC_nprocs
<= 0) {
1029 # if defined(GC_HPUX_THREADS)
1030 GC_nprocs
= pthread_num_processors_np();
1032 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
1033 GC_nprocs
= sysconf(_SC_NPROCESSORS_ONLN
);
1034 if (GC_nprocs
<= 0) GC_nprocs
= 1;
1036 # if defined(GC_IRIX_THREADS)
1037 GC_nprocs
= sysconf(_SC_NPROC_ONLN
);
1038 if (GC_nprocs
<= 0) GC_nprocs
= 1;
1040 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS) || defined(GC_NETBSD_THREADS)
1042 size_t len
= sizeof(ncpus
);
1043 sysctl((int[2]) {CTL_HW
, HW_NCPU
}, 2, &ncpus
, &len
, NULL
, 0);
1046 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1047 GC_nprocs
= GC_get_nprocs();
1050 if (GC_nprocs
<= 0) {
1051 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs
);
1053 # ifdef PARALLEL_MARK
1057 # ifdef PARALLEL_MARK
1059 char * markers_string
= GETENV("GC_MARKERS");
1060 if (markers_string
!= NULL
) {
1061 GC_markers
= atoi(markers_string
);
1063 GC_markers
= GC_nprocs
;
1068 # ifdef PARALLEL_MARK
1070 if (GC_print_stats
) {
1071 GC_printf2("Number of processors = %ld, "
1072 "number of marker threads = %ld\n", GC_nprocs
, GC_markers
);
1075 if (GC_markers
== 1) {
1076 GC_parallel
= FALSE
;
1078 if (GC_print_stats
) {
1079 GC_printf0("Single marker thread, turning off parallel marking\n");
1084 /* Disable true incremental collection, but generational is OK. */
1085 GC_time_limit
= GC_TIME_UNLIMITED
;
1087 /* If we are using a parallel marker, actually start helper threads. */
1088 if (GC_parallel
) start_mark_threads();
1093 /* Perform all initializations, including those that */
1094 /* may require allocation. */
1095 /* Called without allocation lock. */
1096 /* Must be called before a second thread is created. */
1097 /* Called without allocation lock. */
1098 void GC_init_parallel()
1100 if (parallel_initialized
) return;
1101 parallel_initialized
= TRUE
;
1103 /* GC_init() calls us back, so set flag first. */
1104 if (!GC_is_initialized
) GC_init();
1105 /* Initialize thread local free lists if used. */
1106 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1108 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1114 #if !defined(GC_DARWIN_THREADS)
1115 int WRAP_FUNC(pthread_sigmask
)(int how
, const sigset_t
*set
, sigset_t
*oset
)
1117 sigset_t fudged_set
;
1119 if (set
!= NULL
&& (how
== SIG_BLOCK
|| how
== SIG_SETMASK
)) {
1121 sigdelset(&fudged_set
, SIG_SUSPEND
);
1124 return(REAL_FUNC(pthread_sigmask
)(how
, set
, oset
));
1126 #endif /* !GC_DARWIN_THREADS */
1128 /* Wrappers for functions that are likely to block for an appreciable */
1129 /* length of time. Must be called in pairs, if at all. */
1130 /* Nothing much beyond the system call itself should be executed */
1131 /* between these. */
1133 void GC_start_blocking(void) {
1134 # define SP_SLOP 128
1137 me
= GC_lookup_thread(pthread_self());
1138 GC_ASSERT(!(me
-> thread_blocked
));
1140 me
-> stop_info
.stack_ptr
= (ptr_t
)GC_save_regs_in_stack();
1142 # ifndef GC_DARWIN_THREADS
1143 me
-> stop_info
.stack_ptr
= (ptr_t
)GC_approx_sp();
1147 me
-> backing_store_ptr
= (ptr_t
)GC_save_regs_in_stack() + SP_SLOP
;
1149 /* Add some slop to the stack pointer, since the wrapped call may */
1150 /* end up pushing more callee-save registers. */
1151 # ifndef GC_DARWIN_THREADS
1152 # ifdef STACK_GROWS_UP
1153 me
-> stop_info
.stack_ptr
+= SP_SLOP
;
1155 me
-> stop_info
.stack_ptr
-= SP_SLOP
;
1158 me
-> thread_blocked
= TRUE
;
1162 void GC_end_blocking(void) {
1164 LOCK(); /* This will block if the world is stopped. */
1165 me
= GC_lookup_thread(pthread_self());
1166 GC_ASSERT(me
-> thread_blocked
);
1167 me
-> thread_blocked
= FALSE
;
1171 #if defined(GC_DGUX386_THREADS)
1172 #define __d10_sleep sleep
1173 #endif /* GC_DGUX386_THREADS */
1175 /* A wrapper for the standard C sleep function */
1176 int WRAP_FUNC(sleep
) (unsigned int seconds
)
1180 GC_start_blocking();
1181 result
= REAL_FUNC(sleep
)(seconds
);
1187 void *(*start_routine
)(void *);
1190 sem_t registered
; /* 1 ==> in our thread table, but */
1191 /* parent hasn't yet noticed. */
1194 /* Called at thread exit. */
1195 /* Never called for main thread. That's OK, since it */
1196 /* results in at most a tiny one-time leak. And */
1197 /* linuxthreads doesn't reclaim the main threads */
1198 /* resources or id anyway. */
1199 void GC_thread_exit_proc(void *arg
)
1204 me
= GC_lookup_thread(pthread_self());
1205 GC_destroy_thread_local(me
);
1206 if (me
-> flags
& DETACHED
) {
1207 # ifdef THREAD_LOCAL_ALLOC
1208 /* NULL out the tls key to prevent the dtor function from being called */
1209 if (0 != GC_setspecific(GC_thread_key
, NULL
))
1210 ABORT("Failed to set thread specific allocation pointers");
1212 GC_delete_thread(pthread_self());
1214 me
-> flags
|= FINISHED
;
1216 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1217 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1218 GC_remove_specific(GC_thread_key
);
1220 /* The following may run the GC from "nonexistent" thread. */
1221 GC_wait_for_gc_completion(FALSE
);
1225 int WRAP_FUNC(pthread_join
)(pthread_t thread
, void **retval
)
1228 GC_thread thread_gc_id
;
1231 thread_gc_id
= GC_lookup_thread(thread
);
1232 /* This is guaranteed to be the intended one, since the thread id */
1233 /* cant have been recycled by pthreads. */
1235 result
= REAL_FUNC(pthread_join
)(thread
, retval
);
1236 # if defined (GC_FREEBSD_THREADS)
1237 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1238 appears to be) a spurious EINTR which caused the test and real code
1239 to gratuitously fail. Having looked at system pthread library source
1240 code, I see how this return code may be generated. In one path of
1241 code, pthread_join() just returns the errno setting of the thread
1242 being joined. This does not match the POSIX specification or the
1243 local man pages thus I have taken the liberty to catch this one
1244 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1245 if (result
== EINTR
) result
= 0;
1249 /* Here the pthread thread id may have been recycled. */
1250 GC_delete_gc_thread(thread
, thread_gc_id
);
1257 WRAP_FUNC(pthread_detach
)(pthread_t thread
)
1260 GC_thread thread_gc_id
;
1263 thread_gc_id
= GC_lookup_thread(thread
);
1265 result
= REAL_FUNC(pthread_detach
)(thread
);
1268 thread_gc_id
-> flags
|= DETACHED
;
1269 /* Here the pthread thread id may have been recycled. */
1270 if (thread_gc_id
-> flags
& FINISHED
) {
1271 GC_delete_gc_thread(thread
, thread_gc_id
);
1278 GC_bool GC_in_thread_creation
= FALSE
;
1280 typedef void *(*ThreadStartFn
)(void *);
1281 void * GC_start_routine_head(void * arg
, void *base_addr
,
1282 ThreadStartFn
*start
, void **start_arg
)
1284 struct start_info
* si
= arg
;
1287 pthread_t my_pthread
;
1289 my_pthread
= pthread_self();
1290 # ifdef DEBUG_THREADS
1291 GC_printf1("Starting thread 0x%lx\n", my_pthread
);
1292 GC_printf1("pid = %ld\n", (long) getpid());
1293 GC_printf1("sp = 0x%lx\n", (long) &arg
);
1296 GC_in_thread_creation
= TRUE
;
1297 me
= GC_new_thread(my_pthread
);
1298 GC_in_thread_creation
= FALSE
;
1299 #ifdef GC_DARWIN_THREADS
1300 me
-> stop_info
.mach_thread
= mach_thread_self();
1302 me
-> stop_info
.stack_ptr
= 0;
1304 me
-> flags
= si
-> flags
;
1305 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1306 /* doesn't work because the stack base in /proc/self/stat is the */
1307 /* one for the main thread. There is a strong argument that that's */
1308 /* a kernel bug, but a pervasive one. */
1309 # ifdef STACK_GROWS_DOWN
1310 me
-> stack_end
= (ptr_t
)(((word
)(base_addr
) + (GC_page_size
- 1))
1311 & ~(GC_page_size
- 1));
1312 # ifndef GC_DARWIN_THREADS
1313 me
-> stop_info
.stack_ptr
= me
-> stack_end
- 0x10;
1315 /* Needs to be plausible, since an asynchronous stack mark */
1316 /* should not crash. */
1318 me
-> stack_end
= (ptr_t
)((word
)(base_addr
) & ~(GC_page_size
- 1));
1319 me
-> stop_info
.stack_ptr
= me
-> stack_end
+ 0x10;
1321 /* This is dubious, since we may be more than a page into the stack, */
1322 /* and hence skip some of it, though it's not clear that matters. */
1324 me
-> backing_store_end
= (ptr_t
)
1325 (GC_save_regs_in_stack() & ~(GC_page_size
- 1));
1326 /* This is also < 100% convincing. We should also read this */
1327 /* from /proc, but the hook to do so isn't there yet. */
1329 #ifdef MONO_DEBUGGER_SUPPORTED
1330 if (gc_thread_vtable
&& gc_thread_vtable
->thread_created
)
1331 # ifdef GC_DARWIN_THREADS
1332 gc_thread_vtable
->thread_created (mach_thread_self(), &me
->stop_info
.stack_ptr
);
1334 gc_thread_vtable
->thread_created (my_pthread
, &me
->stop_info
.stack_ptr
);
1339 if (start
) *start
= si
-> start_routine
;
1340 if (start_arg
) *start_arg
= si
-> arg
;
1342 sem_post(&(si
-> registered
)); /* Last action on si. */
1343 /* OK to deallocate. */
1344 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1346 GC_init_thread_local(me
);
1353 int GC_thread_register_foreign (void *base_addr
)
1355 struct start_info si
= { 0, }; /* stacked for legibility & locking */
1358 # ifdef DEBUG_THREADS
1359 GC_printf1( "GC_thread_register_foreign %p\n", &si
);
1362 si
.flags
= FOREIGN_THREAD
;
1364 if (!parallel_initialized
) GC_init_parallel();
1366 if (!GC_thr_initialized
) GC_thr_init();
1370 me
= GC_start_routine_head(&si
, base_addr
, NULL
, NULL
);
1375 void * GC_start_routine(void * arg
)
1378 struct start_info
* si
= arg
;
1381 ThreadStartFn start
;
1384 me
= GC_start_routine_head (arg
, &dummy
, &start
, &start_arg
);
1386 pthread_cleanup_push(GC_thread_exit_proc
, 0);
1387 # ifdef DEBUG_THREADS
1388 GC_printf1("start_routine = 0x%lx\n", start
);
1390 result
= (*start
)(start_arg
);
1392 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1394 me
-> status
= result
;
1395 pthread_cleanup_pop(1);
1396 /* Cleanup acquires lock, ensuring that we can't exit */
1397 /* while a collection that thinks we're alive is trying to stop */
1403 WRAP_FUNC(pthread_create
)(pthread_t
*new_thread
,
1404 const pthread_attr_t
*attr
,
1405 void *(*start_routine
)(void *), void *arg
)
1410 struct start_info
* si
;
1411 /* This is otherwise saved only in an area mmapped by the thread */
1412 /* library, which isn't visible to the collector. */
1414 /* We resist the temptation to muck with the stack size here, */
1415 /* even if the default is unreasonably small. That's the client's */
1416 /* responsibility. */
1419 si
= (struct start_info
*)GC_INTERNAL_MALLOC(sizeof(struct start_info
),
1422 if (!parallel_initialized
) GC_init_parallel();
1423 if (0 == si
) return(ENOMEM
);
1424 sem_init(&(si
-> registered
), 0, 0);
1425 si
-> start_routine
= start_routine
;
1428 if (!GC_thr_initialized
) GC_thr_init();
1429 # ifdef GC_ASSERTIONS
1433 pthread_attr_t my_attr
;
1434 pthread_attr_init(&my_attr
);
1435 pthread_attr_getstacksize(&my_attr
, &stack_size
);
1437 pthread_attr_getstacksize(attr
, &stack_size
);
1439 # ifdef PARALLEL_MARK
1440 GC_ASSERT(stack_size
>= (8*HBLKSIZE
*sizeof(word
)));
1442 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1443 /* HBLKSIZE=8192, sizeof(word)=8 */
1444 GC_ASSERT(stack_size
>= 65536);
1446 /* Our threads may need to do some work for the GC. */
1447 /* Ridiculously small threads won't work, and they */
1448 /* probably wouldn't work anyway. */
1452 detachstate
= PTHREAD_CREATE_JOINABLE
;
1454 pthread_attr_getdetachstate(attr
, &detachstate
);
1456 if (PTHREAD_CREATE_DETACHED
== detachstate
) my_flags
|= DETACHED
;
1457 si
-> flags
= my_flags
;
1459 # ifdef DEBUG_THREADS
1460 GC_printf1("About to start new thread from thread 0x%X\n",
1464 result
= REAL_FUNC(pthread_create
)(new_thread
, attr
, GC_start_routine
, si
);
1466 # ifdef DEBUG_THREADS
1467 GC_printf1("Started thread 0x%X\n", *new_thread
);
1469 /* Wait until child has been added to the thread table. */
1470 /* This also ensures that we hold onto si until the child is done */
1471 /* with it. Thus it doesn't matter whether it is otherwise */
1472 /* visible to the collector. */
1474 while (0 != sem_wait(&(si
-> registered
))) {
1475 if (EINTR
!= errno
) ABORT("sem_wait failed");
1478 sem_destroy(&(si
-> registered
));
1480 GC_INTERNAL_FREE(si
);
1486 #ifdef GENERIC_COMPARE_AND_SWAP
1487 pthread_mutex_t GC_compare_and_swap_lock
= PTHREAD_MUTEX_INITIALIZER
;
1489 GC_bool
GC_compare_and_exchange(volatile GC_word
*addr
,
1490 GC_word old
, GC_word new_val
)
1493 pthread_mutex_lock(&GC_compare_and_swap_lock
);
1500 pthread_mutex_unlock(&GC_compare_and_swap_lock
);
1504 GC_word
GC_atomic_add(volatile GC_word
*addr
, GC_word how_much
)
1507 pthread_mutex_lock(&GC_compare_and_swap_lock
);
1509 *addr
= old
+ how_much
;
1510 pthread_mutex_unlock(&GC_compare_and_swap_lock
);
1514 #endif /* GENERIC_COMPARE_AND_SWAP */
1515 /* Spend a few cycles in a way that can't introduce contention with */
1516 /* othre threads. */
1520 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1521 volatile word dummy
= 0;
1524 for (i
= 0; i
< 10; ++i
) {
1525 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1526 __asm__
__volatile__ (" " : : : "memory");
1528 /* Something that's unlikely to be optimized away. */
1534 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1537 VOLATILE GC_bool GC_collecting
= 0;
1538 /* A hint that we're in the collector and */
1539 /* holding the allocation lock for an */
1540 /* extended period. */
1542 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1543 /* If we don't want to use the below spinlock implementation, either */
1544 /* because we don't have a GC_test_and_set implementation, or because */
1545 /* we don't want to risk sleeping, we can still try spinning on */
1546 /* pthread_mutex_trylock for a while. This appears to be very */
1547 /* beneficial in many cases. */
1548 /* I suspect that under high contention this is nearly always better */
1549 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1550 /* Hence we still default to the spin lock. */
1551 /* This is also used to acquire the mark lock for the parallel */
1554 /* Here we use a strict exponential backoff scheme. I don't know */
1555 /* whether that's better or worse than the above. We eventually */
1556 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1557 /* explicitly sleep. */
1561 unsigned long GC_spin_count
= 0;
1562 unsigned long GC_block_count
= 0;
1563 unsigned long GC_unlocked_count
= 0;
1566 void GC_generic_lock(pthread_mutex_t
* lock
)
1568 #ifndef NO_PTHREAD_TRYLOCK
1569 unsigned pause_length
= 1;
1572 if (0 == pthread_mutex_trylock(lock
)) {
1574 ++GC_unlocked_count
;
1578 for (; pause_length
<= SPIN_MAX
; pause_length
<<= 1) {
1579 for (i
= 0; i
< pause_length
; ++i
) {
1582 switch(pthread_mutex_trylock(lock
)) {
1591 ABORT("Unexpected error from pthread_mutex_trylock");
1594 #endif /* !NO_PTHREAD_TRYLOCK */
1598 pthread_mutex_lock(lock
);
1601 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1603 #if defined(USE_SPIN_LOCK)
1605 /* Reasonably fast spin locks. Basically the same implementation */
1606 /* as STL alloc.h. This isn't really the right way to do this. */
1607 /* but until the POSIX scheduling mess gets straightened out ... */
1609 volatile unsigned int GC_allocate_lock
= 0;
1614 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1615 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1616 static unsigned spin_max
= low_spin_max
;
1617 unsigned my_spin_max
;
1618 static unsigned last_spins
= 0;
1619 unsigned my_last_spins
;
1622 if (!GC_test_and_set(&GC_allocate_lock
)) {
1625 my_spin_max
= spin_max
;
1626 my_last_spins
= last_spins
;
1627 for (i
= 0; i
< my_spin_max
; i
++) {
1628 if (GC_collecting
|| GC_nprocs
== 1) goto yield
;
1629 if (i
< my_last_spins
/2 || GC_allocate_lock
) {
1633 if (!GC_test_and_set(&GC_allocate_lock
)) {
1636 * Spinning worked. Thus we're probably not being scheduled
1637 * against the other process with which we were contending.
1638 * Thus it makes sense to spin longer the next time.
1641 spin_max
= high_spin_max
;
1645 /* We are probably being scheduled against the other process. Sleep. */
1646 spin_max
= low_spin_max
;
1649 if (!GC_test_and_set(&GC_allocate_lock
)) {
1652 # define SLEEP_THRESHOLD 12
1653 /* Under Linux very short sleeps tend to wait until */
1654 /* the current time quantum expires. On old Linux */
1655 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1656 /* (Under 2.4, this happens only for real-time */
1657 /* processes.) We want to minimize both behaviors */
1659 if (i
< SLEEP_THRESHOLD
) {
1665 /* Don't wait for more than about 15msecs, even */
1666 /* under extreme contention. */
1668 ts
.tv_nsec
= 1 << i
;
1674 #else /* !USE_SPINLOCK */
1677 #ifndef NO_PTHREAD_TRYLOCK
1678 if (1 == GC_nprocs
|| GC_collecting
) {
1679 pthread_mutex_lock(&GC_allocate_ml
);
1681 GC_generic_lock(&GC_allocate_ml
);
1683 #else /* !NO_PTHREAD_TRYLOCK */
1684 pthread_mutex_lock(&GC_allocate_ml
);
1685 #endif /* !NO_PTHREAD_TRYLOCK */
1688 #endif /* !USE_SPINLOCK */
1690 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1692 #ifdef GC_ASSERTIONS
1693 pthread_t GC_mark_lock_holder
= NO_THREAD
;
1697 /* Ugly workaround for a linux threads bug in the final versions */
1698 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1699 /* field even when it fails to acquire the mutex. This causes */
1700 /* pthread_cond_wait to die. Remove for glibc2.2. */
1701 /* According to the man page, we should use */
1702 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1704 static pthread_mutex_t mark_mutex
=
1705 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP
, {0, 0}};
1707 static pthread_mutex_t mark_mutex
= PTHREAD_MUTEX_INITIALIZER
;
1710 static pthread_cond_t builder_cv
= PTHREAD_COND_INITIALIZER
;
1712 void GC_acquire_mark_lock()
1715 if (pthread_mutex_lock(&mark_mutex) != 0) {
1716 ABORT("pthread_mutex_lock failed");
1719 GC_generic_lock(&mark_mutex
);
1720 # ifdef GC_ASSERTIONS
1721 GC_mark_lock_holder
= pthread_self();
1725 void GC_release_mark_lock()
1727 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1728 # ifdef GC_ASSERTIONS
1729 GC_mark_lock_holder
= NO_THREAD
;
1731 if (pthread_mutex_unlock(&mark_mutex
) != 0) {
1732 ABORT("pthread_mutex_unlock failed");
1736 /* Collector must wait for a freelist builders for 2 reasons: */
1737 /* 1) Mark bits may still be getting examined without lock. */
1738 /* 2) Partial free lists referenced only by locals may not be scanned */
1739 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1740 /* free-list link may be ignored. */
1741 void GC_wait_builder()
1743 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1744 # ifdef GC_ASSERTIONS
1745 GC_mark_lock_holder
= NO_THREAD
;
1747 if (pthread_cond_wait(&builder_cv
, &mark_mutex
) != 0) {
1748 ABORT("pthread_cond_wait failed");
1750 GC_ASSERT(GC_mark_lock_holder
== NO_THREAD
);
1751 # ifdef GC_ASSERTIONS
1752 GC_mark_lock_holder
= pthread_self();
1756 void GC_wait_for_reclaim()
1758 GC_acquire_mark_lock();
1759 while (GC_fl_builder_count
> 0) {
1762 GC_release_mark_lock();
1765 void GC_notify_all_builder()
1767 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1768 if (pthread_cond_broadcast(&builder_cv
) != 0) {
1769 ABORT("pthread_cond_broadcast failed");
1773 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1775 #ifdef PARALLEL_MARK
1777 static pthread_cond_t mark_cv
= PTHREAD_COND_INITIALIZER
;
1779 void GC_wait_marker()
1781 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1782 # ifdef GC_ASSERTIONS
1783 GC_mark_lock_holder
= NO_THREAD
;
1785 if (pthread_cond_wait(&mark_cv
, &mark_mutex
) != 0) {
1786 ABORT("pthread_cond_wait failed");
1788 GC_ASSERT(GC_mark_lock_holder
== NO_THREAD
);
1789 # ifdef GC_ASSERTIONS
1790 GC_mark_lock_holder
= pthread_self();
1794 void GC_notify_all_marker()
1796 if (pthread_cond_broadcast(&mark_cv
) != 0) {
1797 ABORT("pthread_cond_broadcast failed");
1801 #endif /* PARALLEL_MARK */
1803 # endif /* GC_LINUX_THREADS and friends */