2 * Copyright (c) 1994 by Xerox Corporation. All rights reserved.
3 * Copyright (c) 1996 by Silicon Graphics. All rights reserved.
4 * Copyright (c) 1998 by Fergus Henderson. All rights reserved.
5 * Copyright (c) 2000-2004 by Hewlett-Packard Company. All rights reserved.
7 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
8 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
10 * Permission is hereby granted to use or copy this program
11 * for any purpose, provided the above notices are retained on all copies.
12 * Permission to modify the code and to distribute modified code is granted,
13 * provided the above notices are retained, and a notice that the code was
14 * modified is included with the above copyright notice.
17 * Support code for LinuxThreads, the clone()-based kernel
18 * thread package for Linux which is included in libc6.
20 * This code relies on implementation details of LinuxThreads,
21 * (i.e. properties not guaranteed by the Pthread standard),
22 * though this version now does less of that than the other Pthreads
25 * Note that there is a lot of code duplication between linux_threads.c
26 * and thread support for some of the other Posix platforms; any changes
27 * made here may need to be reflected there too.
29 /* DG/UX ix86 support <takis@xfree86.org> */
31 * Linux_threads.c now also includes some code to support HPUX and
32 * OSF1 (Compaq Tru64 Unix, really). The OSF1 support is based on Eric Benson's
35 * Eric also suggested an alternate basis for a lock implementation in
37 * + #elif defined(OSF1)
38 * + unsigned long GC_allocate_lock = 0;
39 * + msemaphore GC_allocate_semaphore;
40 * + # define GC_TRY_LOCK() \
41 * + ((msem_lock(&GC_allocate_semaphore, MSEM_IF_NOWAIT) == 0) \
42 * + ? (GC_allocate_lock = 1) \
44 * + # define GC_LOCK_TAKEN GC_allocate_lock
47 /*#define DEBUG_THREADS 1*/
48 /*#define GC_ASSERTIONS*/
50 # include "private/pthread_support.h"
52 # if defined(GC_PTHREADS) && !defined(GC_SOLARIS_THREADS) \
53 && !defined(GC_WIN32_THREADS)
55 # if defined(GC_HPUX_THREADS) && !defined(USE_PTHREAD_SPECIFIC) \
56 && !defined(USE_COMPILER_TLS)
58 # define USE_PTHREAD_SPECIFIC
59 /* Empirically, as of gcc 3.3, USE_COMPILER_TLS doesn't work. */
61 # define USE_COMPILER_TLS
65 # if defined USE_HPUX_TLS
66 --> Macro replaced by USE_COMPILER_TLS
69 # if (defined(GC_DGUX386_THREADS) || defined(GC_OSF1_THREADS) || \
70 defined(GC_DARWIN_THREADS) || defined(GC_AIX_THREADS)) || \
71 defined(GC_NETBSD_THREADS) && !defined(USE_PTHREAD_SPECIFIC) || \
72 defined(GC_OPENBSD_THREADS)
73 # define USE_PTHREAD_SPECIFIC
76 # if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
77 # define _POSIX4A_DRAFT10_SOURCE 1
80 # if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
81 # define _USING_POSIX4A_DRAFT10 1
84 # ifdef THREAD_LOCAL_ALLOC
85 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
86 # include "private/specific.h"
89 /* Note that these macros should be used only to get/set the GC_thread pointer.
90 * We need to use both tls and pthread because we use the pthread_create function hook to
91 * free the data for foreign threads. When that doesn't happen, libgc could have old
92 * pthread_t that get reused...
94 # if defined(USE_PTHREAD_SPECIFIC)
95 # define GC_getspecific pthread_getspecific
96 # define GC_setspecific pthread_setspecific
97 # define GC_key_create pthread_key_create
98 typedef pthread_key_t GC_key_t
;
100 # if defined(USE_COMPILER_TLS)
101 /* Note sles9 gcc on powerpc gets confused by the define to set GC_thread_tls and pthread_setspecific
102 * so we actually use a static inline function decalred below that is equivalent to:
103 * define GC_setspecific(key, v) (GC_thread_tls = (v), pthread_setspecific ((key), (v)))
105 # define GC_getspecific(x) (GC_thread_tls)
106 # define GC_key_create pthread_key_create
107 typedef pthread_key_t GC_key_t
;
111 # include <pthread.h>
116 # include <sys/mman.h>
117 # include <sys/time.h>
118 # include <sys/types.h>
119 # include <sys/stat.h>
123 #if defined(GC_DARWIN_THREADS)
124 # include "private/darwin_semaphore.h"
126 # include <semaphore.h>
127 #endif /* !GC_DARWIN_THREADS */
129 #if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
130 # include <sys/sysctl.h>
131 #endif /* GC_DARWIN_THREADS */
133 #if defined(GC_NETBSD_THREADS) || defined(GC_OPENBSD_THREADS)
134 # include <sys/param.h>
135 # include <sys/sysctl.h>
140 #if defined(GC_DGUX386_THREADS)
141 # include <sys/dg_sys_info.h>
142 # include <sys/_int_psem.h>
143 /* sem_t is an uint in DG/UX */
144 typedef unsigned int sem_t
;
145 #endif /* GC_DGUX386_THREADS */
151 #ifdef GC_USE_LD_WRAP
152 # define WRAP_FUNC(f) __wrap_##f
153 # define REAL_FUNC(f) __real_##f
155 # define WRAP_FUNC(f) GC_##f
156 # if !defined(GC_DGUX386_THREADS)
157 # define REAL_FUNC(f) f
158 # else /* GC_DGUX386_THREADS */
159 # define REAL_FUNC(f) __d10_##f
160 # endif /* GC_DGUX386_THREADS */
161 # undef pthread_create
162 # if !defined(GC_DARWIN_THREADS)
163 # undef pthread_sigmask
166 # undef pthread_detach
170 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
171 && !defined(_PTHREAD_USE_PTDNAM_)
172 /* Restore the original mangled names on Tru64 UNIX. */
173 # define pthread_create __pthread_create
174 # define pthread_join __pthread_join
175 # define pthread_detach __pthread_detach
181 static GC_bool parallel_initialized
= FALSE
;
183 void GC_init_parallel();
185 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
187 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
189 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
190 #undef PACKAGE_BUGREPORT
192 #undef PACKAGE_STRING
193 #undef PACKAGE_TARNAME
194 #undef PACKAGE_VERSION
195 #include "mono/utils/mono-compiler.h"
198 GC_key_t GC_thread_key
;
200 #ifdef USE_COMPILER_TLS
201 __thread MONO_TLS_FAST
void* GC_thread_tls
;
204 * gcc errors out with /tmp/ccdPMFuq.s:2994: Error: symbol `.LTLS4' is already defined
205 * if the inline is added on powerpc
207 #if !defined(__ppc__) && !defined(__powerpc__)
210 static int GC_setspecific (GC_key_t key
, void *value
) {
211 GC_thread_tls
= value
;
212 return pthread_setspecific (key
, value
);
216 static GC_bool keys_initialized
;
218 #ifdef MONO_DEBUGGER_SUPPORTED
219 #include "include/libgc-mono-debugger.h"
222 /* Recover the contents of the freelist array fl into the global one gfl.*/
223 /* Note that the indexing scheme differs, in that gfl has finer size */
224 /* resolution, even if not all entries are used. */
225 /* We hold the allocator lock. */
226 static void return_freelists(ptr_t
*fl
, ptr_t
*gfl
)
232 for (i
= 1; i
< NFREELISTS
; ++i
) {
233 nwords
= i
* (GRANULARITY
/sizeof(word
));
236 if ((word
)q
>= HBLKSIZE
) {
237 if (gfl
[nwords
] == 0) {
241 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 void GC_delete_gc_thread(pthread_t id
, GC_thread gct
);
260 void GC_destroy_thread_local(GC_thread p
);
262 void GC_thread_deregister_foreign (void *data
)
264 GC_thread me
= (GC_thread
)data
;
265 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
266 if (me
-> flags
& FOREIGN_THREAD
) {
268 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
269 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
270 GC_destroy_thread_local (me
);
272 GC_delete_gc_thread(me
->id
, me
);
277 /* Each thread structure must be initialized. */
278 /* This call must be made from the new thread. */
279 /* Caller holds allocation lock. */
280 void GC_init_thread_local(GC_thread p
)
284 if (!keys_initialized
) {
285 if (0 != GC_key_create(&GC_thread_key
, GC_thread_deregister_foreign
)) {
286 ABORT("Failed to create key for local allocator");
288 keys_initialized
= TRUE
;
290 if (0 != GC_setspecific(GC_thread_key
, p
)) {
291 ABORT("Failed to set thread specific allocation pointers");
293 for (i
= 1; i
< NFREELISTS
; ++i
) {
294 p
-> ptrfree_freelists
[i
] = (ptr_t
)1;
295 p
-> normal_freelists
[i
] = (ptr_t
)1;
296 # ifdef GC_GCJ_SUPPORT
297 p
-> gcj_freelists
[i
] = (ptr_t
)1;
300 /* Set up the size 0 free lists. */
301 p
-> ptrfree_freelists
[0] = (ptr_t
)(&size_zero_object
);
302 p
-> normal_freelists
[0] = (ptr_t
)(&size_zero_object
);
303 # ifdef GC_GCJ_SUPPORT
304 p
-> gcj_freelists
[0] = (ptr_t
)(-1);
308 #ifdef GC_GCJ_SUPPORT
309 extern ptr_t
* GC_gcjobjfreelist
;
312 /* We hold the allocator lock. */
313 void GC_destroy_thread_local(GC_thread p
)
315 /* We currently only do this from the thread itself or from */
316 /* the fork handler for a child process. */
318 GC_ASSERT(GC_getspecific(GC_thread_key
) == (void *)p
);
320 return_freelists(p
-> ptrfree_freelists
, GC_aobjfreelist
);
321 return_freelists(p
-> normal_freelists
, GC_objfreelist
);
322 # ifdef GC_GCJ_SUPPORT
323 return_freelists(p
-> gcj_freelists
, GC_gcjobjfreelist
);
327 extern GC_PTR
GC_generic_malloc_many();
329 GC_PTR
GC_local_malloc(size_t bytes
)
331 if (EXPECT(!SMALL_ENOUGH(bytes
),0)) {
332 return(GC_malloc(bytes
));
334 int index
= INDEX_FROM_BYTES(bytes
);
337 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
338 GC_key_t k
= GC_thread_key
;
342 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
343 if (EXPECT(0 == k
, 0)) {
344 /* This can happen if we get called when the world is */
345 /* being initialized. Whether we can actually complete */
346 /* the initialization then is unclear. */
351 tsd
= GC_getspecific(GC_thread_key
);
352 # ifdef GC_ASSERTIONS
354 GC_ASSERT(tsd
== (void *)GC_lookup_thread(pthread_self()));
357 my_fl
= ((GC_thread
)tsd
) -> normal_freelists
+ index
;
359 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
360 ptr_t next
= obj_link(my_entry
);
361 GC_PTR result
= (GC_PTR
)my_entry
;
363 obj_link(my_entry
) = 0;
364 PREFETCH_FOR_WRITE(next
);
366 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
367 *my_fl
= my_entry
+ index
+ 1;
368 return GC_malloc(bytes
);
370 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), NORMAL
, my_fl
);
371 if (*my_fl
== 0) return GC_oom_fn(bytes
);
372 return GC_local_malloc(bytes
);
377 GC_PTR
GC_local_malloc_atomic(size_t bytes
)
379 if (EXPECT(!SMALL_ENOUGH(bytes
), 0)) {
380 return(GC_malloc_atomic(bytes
));
382 int index
= INDEX_FROM_BYTES(bytes
);
383 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
384 -> ptrfree_freelists
+ index
;
385 ptr_t my_entry
= *my_fl
;
387 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
388 GC_PTR result
= (GC_PTR
)my_entry
;
389 *my_fl
= obj_link(my_entry
);
391 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
392 *my_fl
= my_entry
+ index
+ 1;
393 return GC_malloc_atomic(bytes
);
395 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), PTRFREE
, my_fl
);
396 /* *my_fl is updated while the collector is excluded; */
397 /* the free list is always visible to the collector as */
399 if (*my_fl
== 0) return GC_oom_fn(bytes
);
400 return GC_local_malloc_atomic(bytes
);
405 #ifdef GC_GCJ_SUPPORT
407 #include "include/gc_gcj.h"
410 extern GC_bool GC_gcj_malloc_initialized
;
413 extern int GC_gcj_kind
;
415 GC_PTR
GC_local_gcj_malloc(size_t bytes
,
416 void * ptr_to_struct_containing_descr
)
418 GC_ASSERT(GC_gcj_malloc_initialized
);
419 if (EXPECT(!SMALL_ENOUGH(bytes
), 0)) {
420 return GC_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
422 int index
= INDEX_FROM_BYTES(bytes
);
423 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
424 -> gcj_freelists
+ index
;
425 ptr_t my_entry
= *my_fl
;
426 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
427 GC_PTR result
= (GC_PTR
)my_entry
;
428 GC_ASSERT(!GC_incremental
);
429 /* We assert that any concurrent marker will stop us. */
430 /* Thus it is impossible for a mark procedure to see the */
431 /* allocation of the next object, but to see this object */
432 /* still containing a free list pointer. Otherwise the */
433 /* marker might find a random "mark descriptor". */
434 *(volatile ptr_t
*)my_fl
= obj_link(my_entry
);
435 /* We must update the freelist before we store the pointer. */
436 /* Otherwise a GC at this point would see a corrupted */
438 /* A memory barrier is probably never needed, since the */
439 /* action of stopping this thread will cause prior writes */
441 GC_ASSERT(((void * volatile *)result
)[1] == 0);
442 *(void * volatile *)result
= ptr_to_struct_containing_descr
;
444 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
445 if (!GC_incremental
) *my_fl
= my_entry
+ index
+ 1;
446 /* In the incremental case, we always have to take this */
447 /* path. Thus we leave the counter alone. */
448 return GC_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
450 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), GC_gcj_kind
, my_fl
);
451 if (*my_fl
== 0) return GC_oom_fn(bytes
);
452 return GC_local_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
457 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
458 /* adjust it. The size is assumed to be such that it can be */
459 /* allocated as a small object. */
460 void * GC_local_gcj_fast_malloc(size_t lw
, void * ptr_to_struct_containing_descr
)
462 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
463 -> gcj_freelists
+ lw
;
464 ptr_t my_entry
= *my_fl
;
466 GC_ASSERT(GC_gcj_malloc_initialized
);
468 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
469 GC_PTR result
= (GC_PTR
)my_entry
;
470 GC_ASSERT(!GC_incremental
);
471 /* We assert that any concurrent marker will stop us. */
472 /* Thus it is impossible for a mark procedure to see the */
473 /* allocation of the next object, but to see this object */
474 /* still containing a free list pointer. Otherwise the */
475 /* marker might find a random "mark descriptor". */
476 *(volatile ptr_t
*)my_fl
= obj_link(my_entry
);
477 /* We must update the freelist before we store the pointer. */
478 /* Otherwise a GC at this point would see a corrupted */
480 /* A memory barrier is probably never needed, since the */
481 /* action of stopping this thread will cause prior writes */
483 GC_ASSERT(((void * volatile *)result
)[1] == 0);
484 *(void * volatile *)result
= ptr_to_struct_containing_descr
;
486 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
487 if (!GC_incremental
) *my_fl
= my_entry
+ lw
+ 1;
488 /* In the incremental case, we always have to take this */
489 /* path. Thus we leave the counter alone. */
490 return GC_gcj_fast_malloc(lw
, ptr_to_struct_containing_descr
);
492 GC_generic_malloc_many(BYTES_FROM_INDEX(lw
), GC_gcj_kind
, my_fl
);
493 if (*my_fl
== 0) return GC_oom_fn(BYTES_FROM_INDEX(lw
));
494 return GC_local_gcj_fast_malloc(lw
, ptr_to_struct_containing_descr
);
498 #endif /* GC_GCJ_SUPPORT */
500 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
502 # define GC_destroy_thread_local(t)
504 # endif /* !THREAD_LOCAL_ALLOC */
508 To make sure that we're using LinuxThreads and not some other thread
509 package, we generate a dummy reference to `pthread_kill_other_threads_np'
510 (was `__pthread_initial_thread_bos' but that disappeared),
511 which is a symbol defined in LinuxThreads, but (hopefully) not in other
514 We no longer do this, since this code is now portable enough that it might
515 actually work for something else.
517 void (*dummy_var_to_force_linux_threads
)() = pthread_kill_other_threads_np
;
520 long GC_nprocs
= 1; /* Number of processors. We may not have */
521 /* access to all of them, but this is as good */
522 /* a guess as any ... */
527 # define MAX_MARKERS 16
530 static ptr_t marker_sp
[MAX_MARKERS
] = {0};
532 void * GC_mark_thread(void * id
)
536 marker_sp
[(word
)id
] = GC_approx_sp();
537 for (;; ++my_mark_no
) {
538 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
539 /* promptly. This is important if it were called from the signal */
540 /* handler or from the GC lock acquisition code. Under Linux, it's */
541 /* not safe to call it from a signal handler, since it uses mutexes */
542 /* and condition variables. Since it is called only here, the */
543 /* argument is unnecessary. */
544 if (my_mark_no
< GC_mark_no
|| my_mark_no
> GC_mark_no
+ 2) {
545 /* resynchronize if we get far off, e.g. because GC_mark_no */
547 my_mark_no
= GC_mark_no
;
549 # ifdef DEBUG_THREADS
550 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no
);
552 GC_help_marker(my_mark_no
);
556 extern long GC_markers
; /* Number of mark threads we would */
557 /* like to have. Includes the */
558 /* initiating thread. */
560 pthread_t GC_mark_threads
[MAX_MARKERS
];
562 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
564 static void start_mark_threads()
569 if (GC_markers
> MAX_MARKERS
) {
570 WARN("Limiting number of mark threads\n", 0);
571 GC_markers
= MAX_MARKERS
;
573 if (0 != pthread_attr_init(&attr
)) ABORT("pthread_attr_init failed");
575 if (0 != pthread_attr_setdetachstate(&attr
, PTHREAD_CREATE_DETACHED
))
576 ABORT("pthread_attr_setdetachstate failed");
578 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
579 /* Default stack size is usually too small: fix it. */
580 /* Otherwise marker threads or GC may run out of */
582 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
587 if (pthread_attr_getstacksize(&attr
, &old_size
) != 0)
588 ABORT("pthread_attr_getstacksize failed\n");
589 if (old_size
< MIN_STACK_SIZE
) {
590 if (pthread_attr_setstacksize(&attr
, MIN_STACK_SIZE
) != 0)
591 ABORT("pthread_attr_setstacksize failed\n");
594 # endif /* HPUX || GC_DGUX386_THREADS */
596 if (GC_print_stats
) {
597 GC_printf1("Starting %ld marker threads\n", GC_markers
- 1);
600 for (i
= 0; i
< GC_markers
- 1; ++i
) {
601 if (0 != PTHREAD_CREATE(GC_mark_threads
+ i
, &attr
,
602 GC_mark_thread
, (void *)(word
)i
)) {
603 WARN("Marker thread creation failed, errno = %ld.\n", errno
);
608 #else /* !PARALLEL_MARK */
610 static __inline__
void start_mark_threads()
614 #endif /* !PARALLEL_MARK */
616 GC_bool GC_thr_initialized
= FALSE
;
618 volatile GC_thread GC_threads
[THREAD_TABLE_SZ
];
621 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
622 * put it into a separate function.
624 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
625 static __attribute__((noinline
)) unsigned char* get_gc_thread_key_addr
GC_PROTO((void))
627 return (unsigned char*)&GC_thread_key
;
630 void GC_push_thread_structures
GC_PROTO((void))
632 GC_push_all((ptr_t
)(GC_threads
), (ptr_t
)(GC_threads
)+sizeof(GC_threads
));
633 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
634 GC_push_all((ptr_t
)get_gc_thread_key_addr(),
635 (ptr_t
)(get_gc_thread_key_addr())+sizeof(&GC_thread_key
));
641 void GC_push_thread_structures
GC_PROTO((void))
643 GC_push_all((ptr_t
)(GC_threads
), (ptr_t
)(GC_threads
)+sizeof(GC_threads
));
644 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
645 GC_push_all((ptr_t
)(&GC_thread_key
),
646 (ptr_t
)(&GC_thread_key
)+sizeof(&GC_thread_key
));
652 #ifdef THREAD_LOCAL_ALLOC
653 /* We must explicitly mark ptrfree and gcj free lists, since the free */
654 /* list links wouldn't otherwise be found. We also set them in the */
655 /* normal free lists, since that involves touching less memory than if */
656 /* we scanned them normally. */
657 void GC_mark_thread_local_free_lists(void)
663 for (i
= 0; i
< THREAD_TABLE_SZ
; ++i
) {
664 for (p
= GC_threads
[i
]; 0 != p
; p
= p
-> next
) {
665 for (j
= 1; j
< NFREELISTS
; ++j
) {
666 q
= p
-> ptrfree_freelists
[j
];
667 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
668 q
= p
-> normal_freelists
[j
];
669 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
670 # ifdef GC_GCJ_SUPPORT
671 q
= p
-> gcj_freelists
[j
];
672 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
673 # endif /* GC_GCJ_SUPPORT */
678 #endif /* THREAD_LOCAL_ALLOC */
680 static struct GC_Thread_Rep first_thread
;
683 extern int nacl_thread_parked
[MAX_NACL_GC_THREADS
];
684 extern int nacl_thread_used
[MAX_NACL_GC_THREADS
];
685 extern int nacl_thread_parking_inited
;
686 extern int nacl_num_gc_threads
;
687 extern pthread_mutex_t nacl_thread_alloc_lock
;
688 extern __thread
int nacl_thread_idx
;
689 extern __thread GC_thread nacl_gc_thread_self
;
691 void nacl_initialize_gc_thread()
694 pthread_mutex_lock(&nacl_thread_alloc_lock
);
695 if (!nacl_thread_parking_inited
)
697 for (i
= 0; i
< MAX_NACL_GC_THREADS
; i
++) {
698 nacl_thread_used
[i
] = 0;
699 nacl_thread_parked
[i
] = 0;
701 nacl_thread_parking_inited
= 1;
703 GC_ASSERT(nacl_num_gc_threads
<= MAX_NACL_GC_THREADS
);
704 for (i
= 0; i
< MAX_NACL_GC_THREADS
; i
++) {
705 if (nacl_thread_used
[i
] == 0) {
706 nacl_thread_used
[i
] = 1;
708 nacl_num_gc_threads
++;
712 pthread_mutex_unlock(&nacl_thread_alloc_lock
);
715 void nacl_shutdown_gc_thread()
717 pthread_mutex_lock(&nacl_thread_alloc_lock
);
718 GC_ASSERT(nacl_thread_idx
>= 0 && nacl_thread_idx
< MAX_NACL_GC_THREADS
);
719 GC_ASSERT(nacl_thread_used
[nacl_thread_idx
] != 0);
720 nacl_thread_used
[nacl_thread_idx
] = 0;
721 nacl_thread_idx
= -1;
722 nacl_num_gc_threads
--;
723 pthread_mutex_unlock(&nacl_thread_alloc_lock
);
728 /* Add a thread to GC_threads. We assume it wasn't already there. */
729 /* Caller holds allocation lock. */
730 GC_thread
GC_new_thread(pthread_t id
)
732 int hv
= ((word
)id
) % THREAD_TABLE_SZ
;
734 static GC_bool first_thread_used
= FALSE
;
736 if (!first_thread_used
) {
737 result
= &first_thread
;
738 first_thread_used
= TRUE
;
740 result
= (struct GC_Thread_Rep
*)
741 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep
), NORMAL
);
743 if (result
== 0) return(0);
745 #ifdef PLATFORM_ANDROID
746 result
-> kernel_id
= gettid();
748 result
-> next
= GC_threads
[hv
];
749 GC_threads
[hv
] = result
;
751 nacl_gc_thread_self
= result
;
752 nacl_initialize_gc_thread();
754 GC_ASSERT(result
-> flags
== 0 && result
-> thread_blocked
== 0);
758 /* Delete a thread from GC_threads. We assume it is there. */
759 /* (The code intentionally traps if it wasn't.) */
760 /* Caller holds allocation lock. */
761 void GC_delete_thread(pthread_t id
)
763 int hv
= ((word
)id
) % THREAD_TABLE_SZ
;
764 register GC_thread p
= GC_threads
[hv
];
765 register GC_thread prev
= 0;
768 nacl_shutdown_gc_thread();
769 nacl_gc_thread_self
= NULL
;
772 while (!pthread_equal(p
-> id
, id
)) {
777 GC_threads
[hv
] = p
-> next
;
779 prev
-> next
= p
-> next
;
781 #ifdef MONO_DEBUGGER_SUPPORTED
782 if (gc_thread_vtable
&& gc_thread_vtable
->thread_exited
)
783 gc_thread_vtable
->thread_exited (id
, &p
->stop_info
.stack_ptr
);
786 #ifdef GC_DARWIN_THREADS
787 mach_port_deallocate(mach_task_self(), p
->stop_info
.mach_thread
);
793 /* If a thread has been joined, but we have not yet */
794 /* been notified, then there may be more than one thread */
795 /* in the table with the same pthread id. */
796 /* This is OK, but we need a way to delete a specific one. */
797 void GC_delete_gc_thread(pthread_t id
, GC_thread gc_id
)
799 int hv
= ((word
)id
) % THREAD_TABLE_SZ
;
800 register GC_thread p
= GC_threads
[hv
];
801 register GC_thread prev
= 0;
808 GC_threads
[hv
] = p
-> next
;
810 prev
-> next
= p
-> next
;
813 #ifdef GC_DARWIN_THREADS
814 mach_port_deallocate(mach_task_self(), p
->stop_info
.mach_thread
);
820 /* Return a GC_thread corresponding to a given pthread_t. */
821 /* Returns 0 if it's not there. */
822 /* Caller holds allocation lock or otherwise inhibits */
824 /* If there is more than one thread with the given id we */
825 /* return the most recent one. */
826 GC_thread
GC_lookup_thread(pthread_t id
)
828 int hv
= ((word
)id
) % THREAD_TABLE_SZ
;
829 register GC_thread p
= GC_threads
[hv
];
831 while (p
!= 0 && !pthread_equal(p
-> id
, id
)) p
= p
-> next
;
835 int GC_thread_is_registered (void)
840 ptr
= (void *)GC_lookup_thread(pthread_self());
847 /* Remove all entries from the GC_threads table, except the */
848 /* one for the current thread. We need to do this in the child */
849 /* process after a fork(), since only the current thread */
850 /* survives in the child. */
851 void GC_remove_all_threads_but_me(void)
853 pthread_t self
= pthread_self();
855 GC_thread p
, next
, me
;
857 for (hv
= 0; hv
< THREAD_TABLE_SZ
; ++hv
) {
859 for (p
= GC_threads
[hv
]; 0 != p
; p
= next
) {
861 if (p
-> id
== self
) {
865 # ifdef THREAD_LOCAL_ALLOC
866 if (!(p
-> flags
& FINISHED
)) {
867 GC_destroy_thread_local(p
);
869 # endif /* THREAD_LOCAL_ALLOC */
870 if (p
!= &first_thread
) GC_INTERNAL_FREE(p
);
877 #endif /* HANDLE_FORK */
879 #ifdef USE_PROC_FOR_LIBRARIES
880 int GC_segment_is_thread_stack(ptr_t lo
, ptr_t hi
)
885 # ifdef PARALLEL_MARK
886 for (i
= 0; i
< GC_markers
; ++i
) {
887 if (marker_sp
[i
] > lo
& marker_sp
[i
] < hi
) return 1;
890 for (i
= 0; i
< THREAD_TABLE_SZ
; i
++) {
891 for (p
= GC_threads
[i
]; p
!= 0; p
= p
-> next
) {
892 if (0 != p
-> stack_end
) {
893 # ifdef STACK_GROWS_UP
894 if (p
-> stack_end
>= lo
&& p
-> stack_end
< hi
) return 1;
895 # else /* STACK_GROWS_DOWN */
896 if (p
-> stack_end
> lo
&& p
-> stack_end
<= hi
) return 1;
903 #endif /* USE_PROC_FOR_LIBRARIES */
905 #ifdef GC_LINUX_THREADS
906 /* Return the number of processors, or i<= 0 if it can't be determined. */
909 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
910 /* appears to be buggy in many cases. */
911 /* We look for lines "cpu<n>" in /proc/stat. */
912 # define STAT_BUF_SIZE 4096
913 # define STAT_READ read
914 /* If read is wrapped, this may need to be redefined to call */
916 char stat_buf
[STAT_BUF_SIZE
];
919 /* Some old kernels only have a single "cpu nnnn ..." */
920 /* entry in /proc/stat. We identify those as */
924 f
= open("/proc/stat", O_RDONLY
);
925 if (f
< 0 || (len
= STAT_READ(f
, stat_buf
, STAT_BUF_SIZE
)) < 100) {
926 WARN("Couldn't read /proc/stat\n", 0);
929 for (i
= 0; i
< len
- 100; ++i
) {
930 if (stat_buf
[i
] == '\n' && stat_buf
[i
+1] == 'c'
931 && stat_buf
[i
+2] == 'p' && stat_buf
[i
+3] == 'u') {
932 int cpu_no
= atoi(stat_buf
+ i
+ 4);
933 if (cpu_no
>= result
) result
= cpu_no
+ 1;
939 #endif /* GC_LINUX_THREADS */
941 /* We hold the GC lock. Wait until an in-progress GC has finished. */
942 /* Repeatedly RELEASES GC LOCK in order to wait. */
943 /* If wait_for_all is true, then we exit with the GC lock held and no */
944 /* collection in progress; otherwise we just wait for the current GC */
946 extern GC_bool
GC_collection_in_progress();
947 void GC_wait_for_gc_completion(GC_bool wait_for_all
)
949 if (GC_incremental
&& GC_collection_in_progress()) {
950 int old_gc_no
= GC_gc_no
;
952 /* Make sure that no part of our stack is still on the mark stack, */
953 /* since it's about to be unmapped. */
954 while (GC_incremental
&& GC_collection_in_progress()
955 && (wait_for_all
|| old_gc_no
== GC_gc_no
)) {
957 GC_in_thread_creation
= TRUE
;
958 GC_collect_a_little_inner(1);
959 GC_in_thread_creation
= FALSE
;
969 /* Procedures called before and after a fork. The goal here is to make */
970 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
971 /* attainable, since the single UNIX spec seems to imply that one */
972 /* should only call async-signal-safe functions, and we probably can't */
973 /* quite guarantee that. But we give it our best shot. (That same */
974 /* spec also implies that it's not safe to call the system malloc */
975 /* between fork() and exec(). Thus we're doing no worse than it. */
977 /* Called before a fork() */
978 void GC_fork_prepare_proc(void)
980 /* Acquire all relevant locks, so that after releasing the locks */
981 /* the child will see a consistent state in which monitor */
982 /* invariants hold. Unfortunately, we can't acquire libc locks */
983 /* we might need, and there seems to be no guarantee that libc */
984 /* must install a suitable fork handler. */
985 /* Wait for an ongoing GC to finish, since we can't finish it in */
986 /* the (one remaining thread in) the child. */
988 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
989 GC_wait_for_reclaim();
991 GC_wait_for_gc_completion(TRUE
);
992 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
993 GC_acquire_mark_lock();
997 /* Called in parent after a fork() */
998 void GC_fork_parent_proc(void)
1000 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1001 GC_release_mark_lock();
1006 /* Called in child after a fork() */
1007 void GC_fork_child_proc(void)
1009 /* Clean up the thread table, so that just our thread is left. */
1010 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1011 GC_release_mark_lock();
1013 GC_remove_all_threads_but_me();
1014 # ifdef PARALLEL_MARK
1015 /* Turn off parallel marking in the child, since we are probably */
1016 /* just going to exec, and we would have to restart mark threads. */
1018 GC_parallel
= FALSE
;
1019 # endif /* PARALLEL_MARK */
1022 #endif /* HANDLE_FORK */
1024 #if defined(GC_DGUX386_THREADS)
1025 /* Return the number of processors, or i<= 0 if it can't be determined. */
1028 /* <takis@XFree86.Org> */
1030 struct dg_sys_info_pm_info pm_sysinfo
;
1033 status
= dg_sys_info((long int *) &pm_sysinfo
,
1034 DG_SYS_INFO_PM_INFO_TYPE
, DG_SYS_INFO_PM_CURRENT_VERSION
);
1036 /* set -1 for error */
1040 numCpus
= pm_sysinfo
.idle_vp_count
;
1042 # ifdef DEBUG_THREADS
1043 GC_printf1("Number of active CPUs in this system: %d\n", numCpus
);
1047 #endif /* GC_DGUX386_THREADS */
1049 /* We hold the allocation lock. */
1052 # ifndef GC_DARWIN_THREADS
1057 if (GC_thr_initialized
) return;
1058 GC_thr_initialized
= TRUE
;
1061 /* Prepare for a possible fork. */
1062 pthread_atfork(GC_fork_prepare_proc
, GC_fork_parent_proc
,
1063 GC_fork_child_proc
);
1064 # endif /* HANDLE_FORK */
1065 /* Add the initial thread, so we can stop it. */
1066 t
= GC_new_thread(pthread_self());
1067 # ifdef GC_DARWIN_THREADS
1068 t
-> stop_info
.mach_thread
= mach_thread_self();
1070 t
-> stop_info
.stack_ptr
= (ptr_t
)(&dummy
);
1072 t
-> flags
= DETACHED
| MAIN_THREAD
;
1073 #ifdef MONO_DEBUGGER_SUPPORTED
1074 if (gc_thread_vtable
&& gc_thread_vtable
->thread_created
)
1075 # ifdef GC_DARWIN_THREADS
1076 gc_thread_vtable
->thread_created (mach_thread_self (), &t
->stop_info
.stack_ptr
);
1078 gc_thread_vtable
->thread_created (pthread_self (), &t
->stop_info
.stack_ptr
);
1084 /* Set GC_nprocs. */
1086 char * nprocs_string
= GETENV("GC_NPROCS");
1088 if (nprocs_string
!= NULL
) GC_nprocs
= atoi(nprocs_string
);
1090 if (GC_nprocs
<= 0) {
1091 # if defined(GC_HPUX_THREADS)
1092 GC_nprocs
= pthread_num_processors_np();
1094 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
1095 GC_nprocs
= sysconf(_SC_NPROCESSORS_ONLN
);
1096 if (GC_nprocs
<= 0) GC_nprocs
= 1;
1098 # if defined(GC_IRIX_THREADS)
1099 GC_nprocs
= sysconf(_SC_NPROC_ONLN
);
1100 if (GC_nprocs
<= 0) GC_nprocs
= 1;
1102 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS) || defined(GC_NETBSD_THREADS) || defined(GC_OPENBSD_THREADS)
1104 size_t len
= sizeof(ncpus
);
1105 sysctl((int[2]) {CTL_HW
, HW_NCPU
}, 2, &ncpus
, &len
, NULL
, 0);
1108 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1109 GC_nprocs
= GC_get_nprocs();
1112 if (GC_nprocs
<= 0) {
1113 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs
);
1115 # ifdef PARALLEL_MARK
1119 # ifdef PARALLEL_MARK
1121 char * markers_string
= GETENV("GC_MARKERS");
1122 if (markers_string
!= NULL
) {
1123 GC_markers
= atoi(markers_string
);
1125 GC_markers
= GC_nprocs
;
1126 if (GC_markers
> MAX_MARKERS
)
1127 GC_markers
= MAX_MARKERS
;
1132 # ifdef PARALLEL_MARK
1134 if (GC_print_stats
) {
1135 GC_printf2("Number of processors = %ld, "
1136 "number of marker threads = %ld\n", GC_nprocs
, GC_markers
);
1139 if (GC_markers
== 1) {
1140 GC_parallel
= FALSE
;
1142 if (GC_print_stats
) {
1143 GC_printf0("Single marker thread, turning off parallel marking\n");
1148 /* Disable true incremental collection, but generational is OK. */
1149 GC_time_limit
= GC_TIME_UNLIMITED
;
1151 /* If we are using a parallel marker, actually start helper threads. */
1152 if (GC_parallel
) start_mark_threads();
1157 /* Perform all initializations, including those that */
1158 /* may require allocation. */
1159 /* Called without allocation lock. */
1160 /* Must be called before a second thread is created. */
1161 /* Called without allocation lock. */
1162 void GC_init_parallel()
1164 if (parallel_initialized
) return;
1165 parallel_initialized
= TRUE
;
1167 /* GC_init() calls us back, so set flag first. */
1168 if (!GC_is_initialized
) GC_init();
1169 /* Initialize thread local free lists if used. */
1170 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1172 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1178 #if !defined(GC_DARWIN_THREADS) && !defined(GC_OPENBSD_THREADS)
1180 int WRAP_FUNC(pthread_sigmask
)(int how
, const sigset_t
*set
, sigset_t
*oset
)
1182 sigset_t fudged_set
;
1184 if (set
!= NULL
&& (how
== SIG_BLOCK
|| how
== SIG_SETMASK
)) {
1186 sigdelset(&fudged_set
, SIG_SUSPEND
);
1189 return(REAL_FUNC(pthread_sigmask
)(how
, set
, oset
));
1192 #endif /* !GC_DARWIN_THREADS */
1194 /* Wrappers for functions that are likely to block for an appreciable */
1195 /* length of time. Must be called in pairs, if at all. */
1196 /* Nothing much beyond the system call itself should be executed */
1197 /* between these. */
1199 void GC_start_blocking(void) {
1200 # define SP_SLOP 128
1203 me
= GC_lookup_thread(pthread_self());
1204 GC_ASSERT(!(me
-> thread_blocked
));
1206 me
-> stop_info
.stack_ptr
= (ptr_t
)GC_save_regs_in_stack();
1208 # ifndef GC_DARWIN_THREADS
1209 me
-> stop_info
.stack_ptr
= (ptr_t
)GC_approx_sp();
1213 me
-> backing_store_ptr
= (ptr_t
)GC_save_regs_in_stack() + SP_SLOP
;
1215 /* Add some slop to the stack pointer, since the wrapped call may */
1216 /* end up pushing more callee-save registers. */
1217 # ifndef GC_DARWIN_THREADS
1218 # ifdef STACK_GROWS_UP
1219 me
-> stop_info
.stack_ptr
+= SP_SLOP
;
1221 me
-> stop_info
.stack_ptr
-= SP_SLOP
;
1224 me
-> thread_blocked
= TRUE
;
1228 void GC_end_blocking(void) {
1230 LOCK(); /* This will block if the world is stopped. */
1231 me
= GC_lookup_thread(pthread_self());
1232 GC_ASSERT(me
-> thread_blocked
);
1233 me
-> thread_blocked
= FALSE
;
1237 #if defined(GC_DGUX386_THREADS)
1238 #define __d10_sleep sleep
1239 #endif /* GC_DGUX386_THREADS */
1241 /* A wrapper for the standard C sleep function */
1242 int WRAP_FUNC(sleep
) (unsigned int seconds
)
1246 GC_start_blocking();
1247 result
= REAL_FUNC(sleep
)(seconds
);
1253 void *(*start_routine
)(void *);
1256 sem_t registered
; /* 1 ==> in our thread table, but */
1257 /* parent hasn't yet noticed. */
1260 /* Called at thread exit. */
1261 /* Never called for main thread. That's OK, since it */
1262 /* results in at most a tiny one-time leak. And */
1263 /* linuxthreads doesn't reclaim the main threads */
1264 /* resources or id anyway. */
1265 void GC_thread_exit_proc(void *arg
)
1270 me
= GC_lookup_thread(pthread_self());
1271 GC_destroy_thread_local(me
);
1272 if (me
-> flags
& DETACHED
) {
1273 # ifdef THREAD_LOCAL_ALLOC
1274 /* NULL out the tls key to prevent the dtor function from being called */
1275 if (0 != GC_setspecific(GC_thread_key
, NULL
))
1276 ABORT("Failed to set thread specific allocation pointers");
1278 GC_delete_thread(pthread_self());
1280 me
-> flags
|= FINISHED
;
1282 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1283 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1284 GC_remove_specific(GC_thread_key
);
1286 /* The following may run the GC from "nonexistent" thread. */
1287 GC_wait_for_gc_completion(FALSE
);
1291 int WRAP_FUNC(pthread_join
)(pthread_t thread
, void **retval
)
1294 GC_thread thread_gc_id
;
1297 thread_gc_id
= GC_lookup_thread(thread
);
1298 /* This is guaranteed to be the intended one, since the thread id */
1299 /* cant have been recycled by pthreads. */
1301 result
= REAL_FUNC(pthread_join
)(thread
, retval
);
1302 # if defined (GC_FREEBSD_THREADS)
1303 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1304 appears to be) a spurious EINTR which caused the test and real code
1305 to gratuitously fail. Having looked at system pthread library source
1306 code, I see how this return code may be generated. In one path of
1307 code, pthread_join() just returns the errno setting of the thread
1308 being joined. This does not match the POSIX specification or the
1309 local man pages thus I have taken the liberty to catch this one
1310 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1311 if (result
== EINTR
) result
= 0;
1315 /* Here the pthread thread id may have been recycled. */
1316 GC_delete_gc_thread(thread
, thread_gc_id
);
1323 /* Native Client doesn't support pthread cleanup functions, */
1324 /* so wrap pthread_exit and manually cleanup the thread. */
1326 WRAP_FUNC(pthread_exit
)(void *status
)
1328 GC_thread_exit_proc(0);
1329 REAL_FUNC(pthread_exit
)(status
);
1334 WRAP_FUNC(pthread_detach
)(pthread_t thread
)
1337 GC_thread thread_gc_id
;
1340 thread_gc_id
= GC_lookup_thread(thread
);
1342 result
= REAL_FUNC(pthread_detach
)(thread
);
1345 thread_gc_id
-> flags
|= DETACHED
;
1346 /* Here the pthread thread id may have been recycled. */
1347 if (thread_gc_id
-> flags
& FINISHED
) {
1348 GC_delete_gc_thread(thread
, thread_gc_id
);
1355 GC_bool GC_in_thread_creation
= FALSE
;
1357 typedef void *(*ThreadStartFn
)(void *);
1358 void * GC_start_routine_head(void * arg
, void *base_addr
,
1359 ThreadStartFn
*start
, void **start_arg
)
1361 struct start_info
* si
= arg
;
1364 pthread_t my_pthread
;
1366 my_pthread
= pthread_self();
1367 # ifdef DEBUG_THREADS
1368 GC_printf1("Starting thread 0x%lx\n", my_pthread
);
1369 GC_printf1("pid = %ld\n", (long) getpid());
1370 GC_printf1("sp = 0x%lx\n", (long) &arg
);
1373 GC_in_thread_creation
= TRUE
;
1374 me
= GC_new_thread(my_pthread
);
1375 GC_in_thread_creation
= FALSE
;
1376 #ifdef GC_DARWIN_THREADS
1377 me
-> stop_info
.mach_thread
= mach_thread_self();
1379 me
-> stop_info
.stack_ptr
= 0;
1381 me
-> flags
= si
-> flags
;
1382 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1383 /* doesn't work because the stack base in /proc/self/stat is the */
1384 /* one for the main thread. There is a strong argument that that's */
1385 /* a kernel bug, but a pervasive one. */
1386 # ifdef STACK_GROWS_DOWN
1387 me
-> stack_end
= (ptr_t
)(((word
)(base_addr
) + (GC_page_size
- 1))
1388 & ~(GC_page_size
- 1));
1389 # ifndef GC_DARWIN_THREADS
1390 me
-> stop_info
.stack_ptr
= me
-> stack_end
- 0x10;
1392 /* Needs to be plausible, since an asynchronous stack mark */
1393 /* should not crash. */
1395 me
-> stack_end
= (ptr_t
)((word
)(base_addr
) & ~(GC_page_size
- 1));
1396 me
-> stop_info
.stack_ptr
= me
-> stack_end
+ 0x10;
1398 /* This is dubious, since we may be more than a page into the stack, */
1399 /* and hence skip some of it, though it's not clear that matters. */
1401 me
-> backing_store_end
= (ptr_t
)
1402 (GC_save_regs_in_stack() & ~(GC_page_size
- 1));
1403 /* This is also < 100% convincing. We should also read this */
1404 /* from /proc, but the hook to do so isn't there yet. */
1406 #ifdef MONO_DEBUGGER_SUPPORTED
1407 if (gc_thread_vtable
&& gc_thread_vtable
->thread_created
)
1408 # ifdef GC_DARWIN_THREADS
1409 gc_thread_vtable
->thread_created (mach_thread_self(), &me
->stop_info
.stack_ptr
);
1411 gc_thread_vtable
->thread_created (my_pthread
, &me
->stop_info
.stack_ptr
);
1416 if (start
) *start
= si
-> start_routine
;
1417 if (start_arg
) *start_arg
= si
-> arg
;
1419 if (!(si
->flags
& FOREIGN_THREAD
))
1420 sem_post(&(si
-> registered
)); /* Last action on si. */
1421 /* OK to deallocate. */
1422 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1424 GC_init_thread_local(me
);
1431 int GC_thread_register_foreign (void *base_addr
)
1433 struct start_info si
= { 0, }; /* stacked for legibility & locking */
1436 # ifdef DEBUG_THREADS
1437 GC_printf1( "GC_thread_register_foreign %p\n", &si
);
1440 si
.flags
= FOREIGN_THREAD
;
1442 if (!parallel_initialized
) GC_init_parallel();
1444 if (!GC_thr_initialized
) GC_thr_init();
1448 me
= GC_start_routine_head(&si
, base_addr
, NULL
, NULL
);
1453 void * GC_start_routine(void * arg
)
1456 struct start_info
* si
= arg
;
1459 ThreadStartFn start
;
1462 me
= GC_start_routine_head (arg
, &dummy
, &start
, &start_arg
);
1464 pthread_cleanup_push(GC_thread_exit_proc
, 0);
1465 # ifdef DEBUG_THREADS
1466 GC_printf1("start_routine = 0x%lx\n", start
);
1468 result
= (*start
)(start_arg
);
1470 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1472 me
-> status
= result
;
1473 pthread_cleanup_pop(1);
1474 /* Cleanup acquires lock, ensuring that we can't exit */
1475 /* while a collection that thinks we're alive is trying to stop */
1481 WRAP_FUNC(pthread_create
)(pthread_t
*new_thread
,
1482 const pthread_attr_t
*attr
,
1483 void *(*start_routine
)(void *), void *arg
)
1488 struct start_info
* si
;
1489 /* This is otherwise saved only in an area mmapped by the thread */
1490 /* library, which isn't visible to the collector. */
1492 /* We resist the temptation to muck with the stack size here, */
1493 /* even if the default is unreasonably small. That's the client's */
1494 /* responsibility. */
1497 si
= (struct start_info
*)GC_INTERNAL_MALLOC(sizeof(struct start_info
),
1500 if (!parallel_initialized
) GC_init_parallel();
1501 if (0 == si
) return(ENOMEM
);
1502 sem_init(&(si
-> registered
), 0, 0);
1503 si
-> start_routine
= start_routine
;
1506 if (!GC_thr_initialized
) GC_thr_init();
1507 # ifdef GC_ASSERTIONS
1511 pthread_attr_t my_attr
;
1512 pthread_attr_init(&my_attr
);
1513 pthread_attr_getstacksize(&my_attr
, &stack_size
);
1515 pthread_attr_getstacksize(attr
, &stack_size
);
1517 # ifdef PARALLEL_MARK
1518 GC_ASSERT(stack_size
>= (8*HBLKSIZE
*sizeof(word
)));
1520 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1521 /* HBLKSIZE=8192, sizeof(word)=8 */
1522 GC_ASSERT(stack_size
>= 65536);
1524 /* Our threads may need to do some work for the GC. */
1525 /* Ridiculously small threads won't work, and they */
1526 /* probably wouldn't work anyway. */
1530 detachstate
= PTHREAD_CREATE_JOINABLE
;
1532 pthread_attr_getdetachstate(attr
, &detachstate
);
1534 if (PTHREAD_CREATE_DETACHED
== detachstate
) my_flags
|= DETACHED
;
1535 si
-> flags
= my_flags
;
1537 # ifdef DEBUG_THREADS
1538 GC_printf1("About to start new thread from thread 0x%X\n",
1542 result
= REAL_FUNC(pthread_create
)(new_thread
, attr
, GC_start_routine
, si
);
1544 # ifdef DEBUG_THREADS
1545 GC_printf1("Started thread 0x%X\n", *new_thread
);
1547 /* Wait until child has been added to the thread table. */
1548 /* This also ensures that we hold onto si until the child is done */
1549 /* with it. Thus it doesn't matter whether it is otherwise */
1550 /* visible to the collector. */
1552 while (0 != sem_wait(&(si
-> registered
))) {
1553 if (EINTR
!= errno
) ABORT("sem_wait failed");
1556 sem_destroy(&(si
-> registered
));
1558 GC_INTERNAL_FREE(si
);
1564 #ifdef GENERIC_COMPARE_AND_SWAP
1565 pthread_mutex_t GC_compare_and_swap_lock
= PTHREAD_MUTEX_INITIALIZER
;
1567 GC_bool
GC_compare_and_exchange(volatile GC_word
*addr
,
1568 GC_word old
, GC_word new_val
)
1571 pthread_mutex_lock(&GC_compare_and_swap_lock
);
1578 pthread_mutex_unlock(&GC_compare_and_swap_lock
);
1582 GC_word
GC_atomic_add(volatile GC_word
*addr
, GC_word how_much
)
1585 pthread_mutex_lock(&GC_compare_and_swap_lock
);
1587 *addr
= old
+ how_much
;
1588 pthread_mutex_unlock(&GC_compare_and_swap_lock
);
1592 #endif /* GENERIC_COMPARE_AND_SWAP */
1593 /* Spend a few cycles in a way that can't introduce contention with */
1594 /* othre threads. */
1598 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1599 volatile word dummy
= 0;
1602 for (i
= 0; i
< 10; ++i
) {
1603 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1604 __asm__
__volatile__ (" " : : : "memory");
1606 /* Something that's unlikely to be optimized away. */
1612 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1615 VOLATILE GC_bool GC_collecting
= 0;
1616 /* A hint that we're in the collector and */
1617 /* holding the allocation lock for an */
1618 /* extended period. */
1620 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1621 /* If we don't want to use the below spinlock implementation, either */
1622 /* because we don't have a GC_test_and_set implementation, or because */
1623 /* we don't want to risk sleeping, we can still try spinning on */
1624 /* pthread_mutex_trylock for a while. This appears to be very */
1625 /* beneficial in many cases. */
1626 /* I suspect that under high contention this is nearly always better */
1627 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1628 /* Hence we still default to the spin lock. */
1629 /* This is also used to acquire the mark lock for the parallel */
1632 /* Here we use a strict exponential backoff scheme. I don't know */
1633 /* whether that's better or worse than the above. We eventually */
1634 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1635 /* explicitly sleep. */
1639 unsigned long GC_spin_count
= 0;
1640 unsigned long GC_block_count
= 0;
1641 unsigned long GC_unlocked_count
= 0;
1644 void GC_generic_lock(pthread_mutex_t
* lock
)
1646 #ifndef NO_PTHREAD_TRYLOCK
1647 unsigned pause_length
= 1;
1650 if (0 == pthread_mutex_trylock(lock
)) {
1652 ++GC_unlocked_count
;
1656 for (; pause_length
<= SPIN_MAX
; pause_length
<<= 1) {
1657 for (i
= 0; i
< pause_length
; ++i
) {
1660 switch(pthread_mutex_trylock(lock
)) {
1669 ABORT("Unexpected error from pthread_mutex_trylock");
1672 #endif /* !NO_PTHREAD_TRYLOCK */
1676 pthread_mutex_lock(lock
);
1679 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1681 #if defined(USE_SPIN_LOCK)
1683 /* Reasonably fast spin locks. Basically the same implementation */
1684 /* as STL alloc.h. This isn't really the right way to do this. */
1685 /* but until the POSIX scheduling mess gets straightened out ... */
1687 volatile unsigned int GC_allocate_lock
= 0;
1692 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1693 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1694 static unsigned spin_max
= low_spin_max
;
1695 unsigned my_spin_max
;
1696 static unsigned last_spins
= 0;
1697 unsigned my_last_spins
;
1700 if (!GC_test_and_set(&GC_allocate_lock
)) {
1703 my_spin_max
= spin_max
;
1704 my_last_spins
= last_spins
;
1705 for (i
= 0; i
< my_spin_max
; i
++) {
1706 if (GC_collecting
|| GC_nprocs
== 1) goto yield
;
1707 if (i
< my_last_spins
/2 || GC_allocate_lock
) {
1711 if (!GC_test_and_set(&GC_allocate_lock
)) {
1714 * Spinning worked. Thus we're probably not being scheduled
1715 * against the other process with which we were contending.
1716 * Thus it makes sense to spin longer the next time.
1719 spin_max
= high_spin_max
;
1723 /* We are probably being scheduled against the other process. Sleep. */
1724 spin_max
= low_spin_max
;
1727 if (!GC_test_and_set(&GC_allocate_lock
)) {
1730 # define SLEEP_THRESHOLD 12
1731 /* Under Linux very short sleeps tend to wait until */
1732 /* the current time quantum expires. On old Linux */
1733 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1734 /* (Under 2.4, this happens only for real-time */
1735 /* processes.) We want to minimize both behaviors */
1737 if (i
< SLEEP_THRESHOLD
) {
1743 /* Don't wait for more than about 15msecs, even */
1744 /* under extreme contention. */
1746 ts
.tv_nsec
= 1 << i
;
1752 #else /* !USE_SPINLOCK */
1755 #ifndef NO_PTHREAD_TRYLOCK
1756 if (1 == GC_nprocs
|| GC_collecting
) {
1757 pthread_mutex_lock(&GC_allocate_ml
);
1759 GC_generic_lock(&GC_allocate_ml
);
1761 #else /* !NO_PTHREAD_TRYLOCK */
1762 pthread_mutex_lock(&GC_allocate_ml
);
1763 #endif /* !NO_PTHREAD_TRYLOCK */
1766 #endif /* !USE_SPINLOCK */
1768 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1770 #ifdef GC_ASSERTIONS
1771 pthread_t GC_mark_lock_holder
= NO_THREAD
;
1775 /* Ugly workaround for a linux threads bug in the final versions */
1776 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1777 /* field even when it fails to acquire the mutex. This causes */
1778 /* pthread_cond_wait to die. Remove for glibc2.2. */
1779 /* According to the man page, we should use */
1780 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1782 static pthread_mutex_t mark_mutex
=
1783 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP
, {0, 0}};
1785 static pthread_mutex_t mark_mutex
= PTHREAD_MUTEX_INITIALIZER
;
1788 static pthread_cond_t builder_cv
= PTHREAD_COND_INITIALIZER
;
1790 void GC_acquire_mark_lock()
1793 if (pthread_mutex_lock(&mark_mutex) != 0) {
1794 ABORT("pthread_mutex_lock failed");
1797 GC_generic_lock(&mark_mutex
);
1798 # ifdef GC_ASSERTIONS
1799 GC_mark_lock_holder
= pthread_self();
1803 void GC_release_mark_lock()
1805 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1806 # ifdef GC_ASSERTIONS
1807 GC_mark_lock_holder
= NO_THREAD
;
1809 if (pthread_mutex_unlock(&mark_mutex
) != 0) {
1810 ABORT("pthread_mutex_unlock failed");
1814 /* Collector must wait for a freelist builders for 2 reasons: */
1815 /* 1) Mark bits may still be getting examined without lock. */
1816 /* 2) Partial free lists referenced only by locals may not be scanned */
1817 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1818 /* free-list link may be ignored. */
1819 void GC_wait_builder()
1821 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1822 # ifdef GC_ASSERTIONS
1823 GC_mark_lock_holder
= NO_THREAD
;
1825 if (pthread_cond_wait(&builder_cv
, &mark_mutex
) != 0) {
1826 ABORT("pthread_cond_wait failed");
1828 GC_ASSERT(GC_mark_lock_holder
== NO_THREAD
);
1829 # ifdef GC_ASSERTIONS
1830 GC_mark_lock_holder
= pthread_self();
1834 void GC_wait_for_reclaim()
1836 GC_acquire_mark_lock();
1837 while (GC_fl_builder_count
> 0) {
1840 GC_release_mark_lock();
1843 void GC_notify_all_builder()
1845 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1846 if (pthread_cond_broadcast(&builder_cv
) != 0) {
1847 ABORT("pthread_cond_broadcast failed");
1851 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1853 #ifdef PARALLEL_MARK
1855 static pthread_cond_t mark_cv
= PTHREAD_COND_INITIALIZER
;
1857 void GC_wait_marker()
1859 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1860 # ifdef GC_ASSERTIONS
1861 GC_mark_lock_holder
= NO_THREAD
;
1863 if (pthread_cond_wait(&mark_cv
, &mark_mutex
) != 0) {
1864 ABORT("pthread_cond_wait failed");
1866 GC_ASSERT(GC_mark_lock_holder
== NO_THREAD
);
1867 # ifdef GC_ASSERTIONS
1868 GC_mark_lock_holder
= pthread_self();
1872 void GC_notify_all_marker()
1874 if (pthread_cond_broadcast(&mark_cv
) != 0) {
1875 ABORT("pthread_cond_broadcast failed");
1879 #endif /* PARALLEL_MARK */
1881 # endif /* GC_LINUX_THREADS and friends */