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 #ifndef USE_COMPILER_TLS
70 # if (defined(GC_DGUX386_THREADS) || defined(GC_OSF1_THREADS) || \
71 defined(GC_DARWIN_THREADS) || defined(GC_AIX_THREADS)) || \
72 defined(GC_NETBSD_THREADS) && !defined(USE_PTHREAD_SPECIFIC) || \
73 defined(GC_FREEBSD_THREADS) && !defined(USE_PTHREAD_SPECIFIC) || \
74 defined(GC_OPENBSD_THREADS)
75 # define USE_PTHREAD_SPECIFIC
79 # if defined(GC_DGUX386_THREADS) && !defined(_POSIX4A_DRAFT10_SOURCE)
80 # define _POSIX4A_DRAFT10_SOURCE 1
83 # if defined(GC_DGUX386_THREADS) && !defined(_USING_POSIX4A_DRAFT10)
84 # define _USING_POSIX4A_DRAFT10 1
87 # ifdef THREAD_LOCAL_ALLOC
88 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_COMPILER_TLS)
89 # include "private/specific.h"
92 /* Note that these macros should be used only to get/set the GC_thread pointer.
93 * We need to use both tls and pthread because we use the pthread_create function hook to
94 * free the data for foreign threads. When that doesn't happen, libgc could have old
95 * pthread_t that get reused...
97 # if defined(USE_PTHREAD_SPECIFIC)
98 # define GC_getspecific pthread_getspecific
99 # define GC_setspecific pthread_setspecific
100 # define GC_key_create pthread_key_create
101 typedef pthread_key_t GC_key_t
;
103 # if defined(USE_COMPILER_TLS)
104 /* Note sles9 gcc on powerpc gets confused by the define to set GC_thread_tls and pthread_setspecific
105 * so we actually use a static inline function decalred below that is equivalent to:
106 * define GC_setspecific(key, v) (GC_thread_tls = (v), pthread_setspecific ((key), (v)))
108 # define GC_getspecific(x) (GC_thread_tls)
109 # define GC_key_create pthread_key_create
110 typedef pthread_key_t GC_key_t
;
114 # include <pthread.h>
119 # include <sys/mman.h>
120 # include <sys/time.h>
121 # include <sys/types.h>
122 # include <sys/stat.h>
126 #if defined(GC_DARWIN_THREADS)
127 # include "private/darwin_semaphore.h"
129 # include <semaphore.h>
130 #endif /* !GC_DARWIN_THREADS */
132 #if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS)
133 # include <sys/sysctl.h>
134 #endif /* GC_DARWIN_THREADS */
136 #if defined(GC_NETBSD_THREADS) || defined(GC_OPENBSD_THREADS)
137 # include <sys/param.h>
138 # include <sys/sysctl.h>
143 #if defined(GC_DGUX386_THREADS)
144 # include <sys/dg_sys_info.h>
145 # include <sys/_int_psem.h>
146 /* sem_t is an uint in DG/UX */
147 typedef unsigned int sem_t
;
148 #endif /* GC_DGUX386_THREADS */
154 #ifdef GC_USE_LD_WRAP
155 # define WRAP_FUNC(f) __wrap_##f
156 # define REAL_FUNC(f) __real_##f
158 # define WRAP_FUNC(f) GC_##f
159 # if !defined(GC_DGUX386_THREADS)
160 # define REAL_FUNC(f) f
161 # else /* GC_DGUX386_THREADS */
162 # define REAL_FUNC(f) __d10_##f
163 # endif /* GC_DGUX386_THREADS */
164 # undef pthread_create
165 # if !defined(GC_DARWIN_THREADS)
166 # undef pthread_sigmask
169 # undef pthread_detach
173 # if defined(GC_OSF1_THREADS) && defined(_PTHREAD_USE_MANGLED_NAMES_) \
174 && !defined(_PTHREAD_USE_PTDNAM_)
175 /* Restore the original mangled names on Tru64 UNIX. */
176 # define pthread_create __pthread_create
177 # define pthread_join __pthread_join
178 # define pthread_detach __pthread_detach
184 static GC_bool parallel_initialized
= FALSE
;
186 void GC_init_parallel();
188 static pthread_t main_pthread_self
;
189 static void *main_stack
, *main_altstack
;
190 static int main_stack_size
, main_altstack_size
;
192 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
194 /* We don't really support thread-local allocation with DBG_HDRS_ALL */
196 /* work around a dlopen issue (bug #75390), undefs to avoid warnings with redefinitions */
197 #undef PACKAGE_BUGREPORT
199 #undef PACKAGE_STRING
200 #undef PACKAGE_TARNAME
201 #undef PACKAGE_VERSION
202 #include "mono/utils/mono-compiler.h"
205 GC_key_t GC_thread_key
;
207 #ifdef USE_COMPILER_TLS
208 __thread
void* GC_thread_tls
;
211 * gcc errors out with /tmp/ccdPMFuq.s:2994: Error: symbol `.LTLS4' is already defined
212 * if the inline is added on powerpc
214 #if !defined(__ppc__) && !defined(__powerpc__)
217 static int GC_setspecific (GC_key_t key
, void *value
) {
218 GC_thread_tls
= value
;
219 return pthread_setspecific (key
, value
);
223 static GC_bool keys_initialized
;
225 /* Recover the contents of the freelist array fl into the global one gfl.*/
226 /* Note that the indexing scheme differs, in that gfl has finer size */
227 /* resolution, even if not all entries are used. */
228 /* We hold the allocator lock. */
229 static void return_freelists(ptr_t
*fl
, ptr_t
*gfl
)
235 for (i
= 1; i
< NFREELISTS
; ++i
) {
236 nwords
= i
* (GRANULARITY
/sizeof(word
));
239 if ((word
)q
>= HBLKSIZE
) {
240 if (gfl
[nwords
] == 0) {
244 for (; (word
)q
>= HBLKSIZE
; qptr
= &(obj_link(q
)), q
= *qptr
);
250 /* Clear fl[i], since the thread structure may hang around. */
251 /* Do it in a way that is likely to trap if we access it. */
252 fl
[i
] = (ptr_t
)HBLKSIZE
;
256 /* We statically allocate a single "size 0" object. It is linked to */
257 /* itself, and is thus repeatedly reused for all size 0 allocation */
258 /* requests. (Size 0 gcj allocation requests are incorrect, and */
259 /* we arrange for those to fault asap.) */
260 static ptr_t size_zero_object
= (ptr_t
)(&size_zero_object
);
262 void GC_delete_gc_thread(pthread_t id
, GC_thread gct
);
263 void GC_destroy_thread_local(GC_thread p
);
265 void GC_thread_deregister_foreign (void *data
)
267 GC_thread me
= (GC_thread
)data
;
268 /* GC_fprintf1( "\n\n\n\n --- Deregister %x ---\n\n\n\n\n", me->flags ); */
269 if (me
-> flags
& FOREIGN_THREAD
) {
271 /* GC_fprintf0( "\n\n\n\n --- FOO ---\n\n\n\n\n" ); */
272 #if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
273 GC_destroy_thread_local (me
);
275 GC_delete_gc_thread(me
->id
, me
);
280 /* Each thread structure must be initialized. */
281 /* This call must be made from the new thread. */
282 /* Caller holds allocation lock. */
283 void GC_init_thread_local(GC_thread p
)
287 if (!keys_initialized
) {
288 if (0 != GC_key_create(&GC_thread_key
, GC_thread_deregister_foreign
)) {
289 ABORT("Failed to create key for local allocator");
291 keys_initialized
= TRUE
;
293 if (0 != GC_setspecific(GC_thread_key
, p
)) {
294 ABORT("Failed to set thread specific allocation pointers");
296 for (i
= 1; i
< NFREELISTS
; ++i
) {
297 p
-> tlfs
.ptrfree_freelists
[i
] = (ptr_t
)1;
298 p
-> tlfs
.normal_freelists
[i
] = (ptr_t
)1;
299 # ifdef GC_GCJ_SUPPORT
300 p
-> tlfs
.gcj_freelists
[i
] = (ptr_t
)1;
303 /* Set up the size 0 free lists. */
304 p
-> tlfs
.ptrfree_freelists
[0] = (ptr_t
)(&size_zero_object
);
305 p
-> tlfs
.normal_freelists
[0] = (ptr_t
)(&size_zero_object
);
306 # ifdef GC_GCJ_SUPPORT
307 p
-> tlfs
.gcj_freelists
[0] = (ptr_t
)(-1);
311 #ifdef GC_GCJ_SUPPORT
312 extern ptr_t
* GC_gcjobjfreelist
;
315 /* We hold the allocator lock. */
316 void GC_destroy_thread_local(GC_thread p
)
318 /* We currently only do this from the thread itself or from */
319 /* the fork handler for a child process. */
321 GC_ASSERT(GC_getspecific(GC_thread_key
) == (void *)p
);
323 return_freelists(p
-> tlfs
.ptrfree_freelists
, GC_aobjfreelist
);
324 return_freelists(p
-> tlfs
.normal_freelists
, GC_objfreelist
);
325 # ifdef GC_GCJ_SUPPORT
326 return_freelists(p
-> tlfs
.gcj_freelists
, GC_gcjobjfreelist
);
330 extern GC_PTR
GC_generic_malloc_many();
332 GC_PTR
GC_local_malloc(size_t bytes
)
334 if (EXPECT(!SMALL_ENOUGH(bytes
),0)) {
335 return(GC_malloc(bytes
));
337 int index
= INDEX_FROM_BYTES(bytes
);
340 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
341 GC_key_t k
= GC_thread_key
;
345 # if defined(REDIRECT_MALLOC) && !defined(USE_PTHREAD_SPECIFIC)
346 if (EXPECT(0 == k
, 0)) {
347 /* This can happen if we get called when the world is */
348 /* being initialized. Whether we can actually complete */
349 /* the initialization then is unclear. */
354 tsd
= GC_getspecific(GC_thread_key
);
355 # ifdef GC_ASSERTIONS
357 GC_ASSERT(tsd
== (void *)GC_lookup_thread(pthread_self()));
360 my_fl
= ((GC_thread
)tsd
) -> tlfs
.normal_freelists
+ index
;
362 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
363 ptr_t next
= obj_link(my_entry
);
364 GC_PTR result
= (GC_PTR
)my_entry
;
366 obj_link(my_entry
) = 0;
367 PREFETCH_FOR_WRITE(next
);
369 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
370 *my_fl
= my_entry
+ index
+ 1;
371 return GC_malloc(bytes
);
373 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), NORMAL
, my_fl
);
374 if (*my_fl
== 0) return GC_oom_fn(bytes
);
375 return GC_local_malloc(bytes
);
380 GC_PTR
GC_local_malloc_atomic(size_t bytes
)
382 if (EXPECT(!SMALL_ENOUGH(bytes
), 0)) {
383 return(GC_malloc_atomic(bytes
));
385 int index
= INDEX_FROM_BYTES(bytes
);
386 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
387 -> tlfs
.ptrfree_freelists
+ index
;
388 ptr_t my_entry
= *my_fl
;
390 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
391 GC_PTR result
= (GC_PTR
)my_entry
;
392 *my_fl
= obj_link(my_entry
);
394 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
395 *my_fl
= my_entry
+ index
+ 1;
396 return GC_malloc_atomic(bytes
);
398 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), PTRFREE
, my_fl
);
399 /* *my_fl is updated while the collector is excluded; */
400 /* the free list is always visible to the collector as */
402 if (*my_fl
== 0) return GC_oom_fn(bytes
);
403 return GC_local_malloc_atomic(bytes
);
408 #ifdef GC_GCJ_SUPPORT
410 #include "include/gc_gcj.h"
413 extern GC_bool GC_gcj_malloc_initialized
;
416 extern int GC_gcj_kind
;
418 GC_PTR
GC_local_gcj_malloc(size_t bytes
,
419 void * ptr_to_struct_containing_descr
)
421 GC_ASSERT(GC_gcj_malloc_initialized
);
422 if (EXPECT(!SMALL_ENOUGH(bytes
), 0)) {
423 return GC_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
425 int index
= INDEX_FROM_BYTES(bytes
);
426 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
427 -> tlfs
.gcj_freelists
+ index
;
428 ptr_t my_entry
= *my_fl
;
429 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
430 GC_PTR result
= (GC_PTR
)my_entry
;
431 GC_ASSERT(!GC_incremental
);
432 /* We assert that any concurrent marker will stop us. */
433 /* Thus it is impossible for a mark procedure to see the */
434 /* allocation of the next object, but to see this object */
435 /* still containing a free list pointer. Otherwise the */
436 /* marker might find a random "mark descriptor". */
437 *(volatile ptr_t
*)my_fl
= obj_link(my_entry
);
438 /* We must update the freelist before we store the pointer. */
439 /* Otherwise a GC at this point would see a corrupted */
441 /* A memory barrier is probably never needed, since the */
442 /* action of stopping this thread will cause prior writes */
444 GC_ASSERT(((void * volatile *)result
)[1] == 0);
445 *(void * volatile *)result
= ptr_to_struct_containing_descr
;
447 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
448 if (!GC_incremental
) *my_fl
= my_entry
+ index
+ 1;
449 /* In the incremental case, we always have to take this */
450 /* path. Thus we leave the counter alone. */
451 return GC_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
453 GC_generic_malloc_many(BYTES_FROM_INDEX(index
), GC_gcj_kind
, my_fl
);
454 if (*my_fl
== 0) return GC_oom_fn(bytes
);
455 return GC_local_gcj_malloc(bytes
, ptr_to_struct_containing_descr
);
460 /* Similar to GC_local_gcj_malloc, but the size is in words, and we don't */
461 /* adjust it. The size is assumed to be such that it can be */
462 /* allocated as a small object. */
463 void * GC_local_gcj_fast_malloc(size_t lw
, void * ptr_to_struct_containing_descr
)
465 ptr_t
* my_fl
= ((GC_thread
)GC_getspecific(GC_thread_key
))
466 -> tlfs
.gcj_freelists
+ lw
;
467 ptr_t my_entry
= *my_fl
;
469 GC_ASSERT(GC_gcj_malloc_initialized
);
471 if (EXPECT((word
)my_entry
>= HBLKSIZE
, 1)) {
472 GC_PTR result
= (GC_PTR
)my_entry
;
473 GC_ASSERT(!GC_incremental
);
474 /* We assert that any concurrent marker will stop us. */
475 /* Thus it is impossible for a mark procedure to see the */
476 /* allocation of the next object, but to see this object */
477 /* still containing a free list pointer. Otherwise the */
478 /* marker might find a random "mark descriptor". */
479 *(volatile ptr_t
*)my_fl
= obj_link(my_entry
);
480 /* We must update the freelist before we store the pointer. */
481 /* Otherwise a GC at this point would see a corrupted */
483 /* A memory barrier is probably never needed, since the */
484 /* action of stopping this thread will cause prior writes */
486 GC_ASSERT(((void * volatile *)result
)[1] == 0);
487 *(void * volatile *)result
= ptr_to_struct_containing_descr
;
489 } else if ((word
)my_entry
- 1 < DIRECT_GRANULES
) {
490 if (!GC_incremental
) *my_fl
= my_entry
+ lw
+ 1;
491 /* In the incremental case, we always have to take this */
492 /* path. Thus we leave the counter alone. */
493 return GC_gcj_fast_malloc(lw
, ptr_to_struct_containing_descr
);
495 GC_generic_malloc_many(BYTES_FROM_INDEX(lw
), GC_gcj_kind
, my_fl
);
496 if (*my_fl
== 0) return GC_oom_fn(BYTES_FROM_INDEX(lw
));
497 return GC_local_gcj_fast_malloc(lw
, ptr_to_struct_containing_descr
);
501 #endif /* GC_GCJ_SUPPORT */
503 # else /* !THREAD_LOCAL_ALLOC && !DBG_HDRS_ALL */
505 # define GC_destroy_thread_local(t)
507 # endif /* !THREAD_LOCAL_ALLOC */
511 To make sure that we're using LinuxThreads and not some other thread
512 package, we generate a dummy reference to `pthread_kill_other_threads_np'
513 (was `__pthread_initial_thread_bos' but that disappeared),
514 which is a symbol defined in LinuxThreads, but (hopefully) not in other
517 We no longer do this, since this code is now portable enough that it might
518 actually work for something else.
520 void (*dummy_var_to_force_linux_threads
)() = pthread_kill_other_threads_np
;
523 long GC_nprocs
= 1; /* Number of processors. We may not have */
524 /* access to all of them, but this is as good */
525 /* a guess as any ... */
530 # define MAX_MARKERS 16
533 static ptr_t marker_sp
[MAX_MARKERS
] = {0};
535 void * GC_mark_thread(void * id
)
539 marker_sp
[(word
)id
] = GC_approx_sp();
540 for (;; ++my_mark_no
) {
541 /* GC_mark_no is passed only to allow GC_help_marker to terminate */
542 /* promptly. This is important if it were called from the signal */
543 /* handler or from the GC lock acquisition code. Under Linux, it's */
544 /* not safe to call it from a signal handler, since it uses mutexes */
545 /* and condition variables. Since it is called only here, the */
546 /* argument is unnecessary. */
547 if (my_mark_no
< GC_mark_no
|| my_mark_no
> GC_mark_no
+ 2) {
548 /* resynchronize if we get far off, e.g. because GC_mark_no */
550 my_mark_no
= GC_mark_no
;
552 # ifdef DEBUG_THREADS
553 GC_printf1("Starting mark helper for mark number %ld\n", my_mark_no
);
555 GC_help_marker(my_mark_no
);
559 extern long GC_markers
; /* Number of mark threads we would */
560 /* like to have. Includes the */
561 /* initiating thread. */
563 pthread_t GC_mark_threads
[MAX_MARKERS
];
565 #define PTHREAD_CREATE REAL_FUNC(pthread_create)
567 static void start_mark_threads()
572 if (GC_markers
> MAX_MARKERS
) {
573 WARN("Limiting number of mark threads\n", 0);
574 GC_markers
= MAX_MARKERS
;
576 if (0 != pthread_attr_init(&attr
)) ABORT("pthread_attr_init failed");
578 if (0 != pthread_attr_setdetachstate(&attr
, PTHREAD_CREATE_DETACHED
))
579 ABORT("pthread_attr_setdetachstate failed");
581 # if defined(HPUX) || defined(GC_DGUX386_THREADS)
582 /* Default stack size is usually too small: fix it. */
583 /* Otherwise marker threads or GC may run out of */
585 # define MIN_STACK_SIZE (8*HBLKSIZE*sizeof(word))
590 if (pthread_attr_getstacksize(&attr
, &old_size
) != 0)
591 ABORT("pthread_attr_getstacksize failed\n");
592 if (old_size
< MIN_STACK_SIZE
) {
593 if (pthread_attr_setstacksize(&attr
, MIN_STACK_SIZE
) != 0)
594 ABORT("pthread_attr_setstacksize failed\n");
597 # endif /* HPUX || GC_DGUX386_THREADS */
599 if (GC_print_stats
) {
600 GC_printf1("Starting %ld marker threads\n", GC_markers
- 1);
603 for (i
= 0; i
< GC_markers
- 1; ++i
) {
604 if (0 != PTHREAD_CREATE(GC_mark_threads
+ i
, &attr
,
605 GC_mark_thread
, (void *)(word
)i
)) {
606 WARN("Marker thread creation failed, errno = %ld.\n", errno
);
611 #else /* !PARALLEL_MARK */
613 static __inline__
void start_mark_threads()
617 #endif /* !PARALLEL_MARK */
619 GC_bool GC_thr_initialized
= FALSE
;
621 volatile GC_thread GC_threads
[THREAD_TABLE_SZ
];
624 * gcc-3.3.6 miscompiles the &GC_thread_key+sizeof(&GC_thread_key) expression so
625 * put it into a separate function.
627 # if defined(__GNUC__) && defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
628 static __attribute__((noinline
)) unsigned char* get_gc_thread_key_addr
GC_PROTO((void))
630 return (unsigned char*)&GC_thread_key
;
633 void GC_push_thread_structures
GC_PROTO((void))
635 GC_push_all((ptr_t
)(GC_threads
), (ptr_t
)(GC_threads
)+sizeof(GC_threads
));
636 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
637 GC_push_all((ptr_t
)get_gc_thread_key_addr(),
638 (ptr_t
)(get_gc_thread_key_addr())+sizeof(&GC_thread_key
));
644 void GC_push_thread_structures
GC_PROTO((void))
646 GC_push_all((ptr_t
)(GC_threads
), (ptr_t
)(GC_threads
)+sizeof(GC_threads
));
647 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
648 GC_push_all((ptr_t
)(&GC_thread_key
),
649 (ptr_t
)(&GC_thread_key
)+sizeof(&GC_thread_key
));
655 #ifdef THREAD_LOCAL_ALLOC
656 /* We must explicitly mark ptrfree and gcj free lists, since the free */
657 /* list links wouldn't otherwise be found. We also set them in the */
658 /* normal free lists, since that involves touching less memory than if */
659 /* we scanned them normally. */
660 void GC_mark_thread_local_free_lists(void)
666 for (i
= 0; i
< THREAD_TABLE_SZ
; ++i
) {
667 for (p
= GC_threads
[i
]; 0 != p
; p
= p
-> next
) {
668 for (j
= 1; j
< NFREELISTS
; ++j
) {
669 q
= p
-> tlfs
.ptrfree_freelists
[j
];
670 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
671 q
= p
-> tlfs
.normal_freelists
[j
];
672 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
673 # ifdef GC_GCJ_SUPPORT
674 q
= p
-> tlfs
.gcj_freelists
[j
];
675 if ((word
)q
> HBLKSIZE
) GC_set_fl_marks(q
);
676 # endif /* GC_GCJ_SUPPORT */
681 #endif /* THREAD_LOCAL_ALLOC */
683 static struct GC_Thread_Rep first_thread
;
686 extern volatile int nacl_thread_parked
[MAX_NACL_GC_THREADS
];
687 extern volatile int nacl_thread_used
[MAX_NACL_GC_THREADS
];
688 extern volatile int nacl_thread_parking_inited
;
689 extern volatile int nacl_num_gc_threads
;
690 extern pthread_mutex_t nacl_thread_alloc_lock
;
691 extern __thread
int nacl_thread_idx
;
692 extern __thread GC_thread nacl_gc_thread_self
;
694 extern void nacl_pre_syscall_hook();
695 extern void nacl_post_syscall_hook();
696 extern void nacl_register_gc_hooks(void (*pre
)(), void (*post
)());
700 struct nacl_irt_blockhook
{
701 int (*register_block_hooks
)(void (*pre
)(void), void (*post
)(void));
704 extern size_t nacl_interface_query(const char *interface_ident
,
705 void *table
, size_t tablesize
);
707 void nacl_initialize_gc_thread()
710 static struct nacl_irt_blockhook gc_hook
;
712 pthread_mutex_lock(&nacl_thread_alloc_lock
);
713 if (!nacl_thread_parking_inited
)
715 for (i
= 0; i
< MAX_NACL_GC_THREADS
; i
++) {
716 nacl_thread_used
[i
] = 0;
717 nacl_thread_parked
[i
] = 0;
719 // TODO: replace with public 'register hook' function when
720 // available from glibc
721 nacl_interface_query("nacl-irt-blockhook-0.1", &gc_hook
, sizeof(gc_hook
));
722 gc_hook
.register_block_hooks(nacl_pre_syscall_hook
, nacl_post_syscall_hook
);
723 nacl_thread_parking_inited
= 1;
725 GC_ASSERT(nacl_num_gc_threads
<= MAX_NACL_GC_THREADS
);
726 for (i
= 0; i
< MAX_NACL_GC_THREADS
; i
++) {
727 if (nacl_thread_used
[i
] == 0) {
728 nacl_thread_used
[i
] = 1;
730 nacl_num_gc_threads
++;
734 pthread_mutex_unlock(&nacl_thread_alloc_lock
);
737 void nacl_shutdown_gc_thread()
739 pthread_mutex_lock(&nacl_thread_alloc_lock
);
740 GC_ASSERT(nacl_thread_idx
>= 0 && nacl_thread_idx
< MAX_NACL_GC_THREADS
);
741 GC_ASSERT(nacl_thread_used
[nacl_thread_idx
] != 0);
742 nacl_thread_used
[nacl_thread_idx
] = 0;
743 nacl_thread_idx
= -1;
744 nacl_num_gc_threads
--;
745 pthread_mutex_unlock(&nacl_thread_alloc_lock
);
750 /* Add a thread to GC_threads. We assume it wasn't already there. */
751 /* Caller holds allocation lock. */
752 GC_thread
GC_new_thread(pthread_t id
)
754 int hv
= ((unsigned long)id
) % THREAD_TABLE_SZ
;
756 static GC_bool first_thread_used
= FALSE
;
758 if (!first_thread_used
) {
759 result
= &first_thread
;
760 first_thread_used
= TRUE
;
762 result
= (struct GC_Thread_Rep
*)
763 GC_INTERNAL_MALLOC(sizeof(struct GC_Thread_Rep
), NORMAL
);
765 if (result
== 0) return(0);
768 result
-> kernel_id
= gettid();
770 result
-> next
= GC_threads
[hv
];
771 GC_threads
[hv
] = result
;
773 nacl_gc_thread_self
= result
;
774 nacl_initialize_gc_thread();
776 GC_ASSERT(result
-> flags
== 0 && result
-> thread_blocked
== 0);
780 /* Delete a thread from GC_threads. We assume it is there. */
781 /* (The code intentionally traps if it wasn't.) */
782 /* Caller holds allocation lock. */
783 void GC_delete_thread(pthread_t id
)
785 int hv
= ((unsigned long)id
) % THREAD_TABLE_SZ
;
786 register GC_thread p
= GC_threads
[hv
];
787 register GC_thread prev
= 0;
790 nacl_shutdown_gc_thread();
791 nacl_gc_thread_self
= NULL
;
794 while (!pthread_equal(p
-> id
, id
)) {
799 GC_threads
[hv
] = p
-> next
;
801 prev
-> next
= p
-> next
;
804 #ifdef GC_DARWIN_THREADS
805 mach_port_deallocate(mach_task_self(), p
->stop_info
.mach_thread
);
811 /* If a thread has been joined, but we have not yet */
812 /* been notified, then there may be more than one thread */
813 /* in the table with the same pthread id. */
814 /* This is OK, but we need a way to delete a specific one. */
815 void GC_delete_gc_thread(pthread_t id
, GC_thread gc_id
)
817 int hv
= ((unsigned long)id
) % THREAD_TABLE_SZ
;
818 register GC_thread p
= GC_threads
[hv
];
819 register GC_thread prev
= 0;
826 GC_threads
[hv
] = p
-> next
;
828 prev
-> next
= p
-> next
;
831 #ifdef GC_DARWIN_THREADS
832 mach_port_deallocate(mach_task_self(), p
->stop_info
.mach_thread
);
838 /* Return a GC_thread corresponding to a given pthread_t. */
839 /* Returns 0 if it's not there. */
840 /* Caller holds allocation lock or otherwise inhibits */
842 /* If there is more than one thread with the given id we */
843 /* return the most recent one. */
844 GC_thread
GC_lookup_thread(pthread_t id
)
846 int hv
= ((unsigned long)id
) % THREAD_TABLE_SZ
;
847 register GC_thread p
= GC_threads
[hv
];
849 while (p
!= 0 && !pthread_equal(p
-> id
, id
)) p
= p
-> next
;
853 int GC_thread_is_registered (void)
858 ptr
= (void *)GC_lookup_thread(pthread_self());
864 void GC_register_altstack (void *stack
, int stack_size
, void *altstack
, int altstack_size
)
869 thread
= (void *)GC_lookup_thread(pthread_self());
871 thread
->stack
= stack
;
872 thread
->stack_size
= stack_size
;
873 thread
->altstack
= altstack
;
874 thread
->altstack_size
= altstack_size
;
877 * This happens if we are called before GC_thr_init ().
879 main_pthread_self
= pthread_self ();
881 main_stack_size
= stack_size
;
882 main_altstack
= altstack
;
883 main_altstack_size
= altstack_size
;
889 /* Remove all entries from the GC_threads table, except the */
890 /* one for the current thread. We need to do this in the child */
891 /* process after a fork(), since only the current thread */
892 /* survives in the child. */
893 void GC_remove_all_threads_but_me(void)
895 pthread_t self
= pthread_self();
897 GC_thread p
, next
, me
;
899 for (hv
= 0; hv
< THREAD_TABLE_SZ
; ++hv
) {
901 for (p
= GC_threads
[hv
]; 0 != p
; p
= next
) {
903 if (p
-> id
== self
) {
907 # ifdef THREAD_LOCAL_ALLOC
908 if (!(p
-> flags
& FINISHED
)) {
909 GC_destroy_thread_local(p
);
911 # endif /* THREAD_LOCAL_ALLOC */
912 if (p
!= &first_thread
) GC_INTERNAL_FREE(p
);
919 #endif /* HANDLE_FORK */
921 #ifdef USE_PROC_FOR_LIBRARIES
922 int GC_segment_is_thread_stack(ptr_t lo
, ptr_t hi
)
927 # ifdef PARALLEL_MARK
928 for (i
= 0; i
< GC_markers
; ++i
) {
929 if (marker_sp
[i
] > lo
& marker_sp
[i
] < hi
) return 1;
932 for (i
= 0; i
< THREAD_TABLE_SZ
; i
++) {
933 for (p
= GC_threads
[i
]; p
!= 0; p
= p
-> next
) {
934 if (0 != p
-> stack_end
) {
935 # ifdef STACK_GROWS_UP
936 if (p
-> stack_end
>= lo
&& p
-> stack_end
< hi
) return 1;
937 # else /* STACK_GROWS_DOWN */
938 if (p
-> stack_end
> lo
&& p
-> stack_end
<= hi
) return 1;
945 #endif /* USE_PROC_FOR_LIBRARIES */
947 #ifdef GC_LINUX_THREADS
948 /* Return the number of processors, or i<= 0 if it can't be determined. */
952 /* Should be "return sysconf(_SC_NPROCESSORS_ONLN);" but that */
953 /* appears to be buggy in many cases. */
954 /* We look for lines "cpu<n>" in /proc/stat. */
955 # define STAT_BUF_SIZE 4096
956 # define STAT_READ read
957 /* If read is wrapped, this may need to be redefined to call */
959 char stat_buf
[STAT_BUF_SIZE
];
962 /* Some old kernels only have a single "cpu nnnn ..." */
963 /* entry in /proc/stat. We identify those as */
967 f
= open("/proc/stat", O_RDONLY
);
968 if (f
< 0 || (len
= STAT_READ(f
, stat_buf
, STAT_BUF_SIZE
)) < 100) {
969 WARN("Couldn't read /proc/stat\n", 0);
972 for (i
= 0; i
< len
- 100; ++i
) {
973 if (stat_buf
[i
] == '\n' && stat_buf
[i
+1] == 'c'
974 && stat_buf
[i
+2] == 'p' && stat_buf
[i
+3] == 'u') {
975 int cpu_no
= atoi(stat_buf
+ i
+ 4);
976 if (cpu_no
>= result
) result
= cpu_no
+ 1;
982 return sysconf(_SC_NPROCESSORS_ONLN
);
985 #endif /* GC_LINUX_THREADS */
987 /* We hold the GC lock. Wait until an in-progress GC has finished. */
988 /* Repeatedly RELEASES GC LOCK in order to wait. */
989 /* If wait_for_all is true, then we exit with the GC lock held and no */
990 /* collection in progress; otherwise we just wait for the current GC */
992 extern GC_bool
GC_collection_in_progress();
993 void GC_wait_for_gc_completion(GC_bool wait_for_all
)
995 if (GC_incremental
&& GC_collection_in_progress()) {
996 int old_gc_no
= GC_gc_no
;
998 /* Make sure that no part of our stack is still on the mark stack, */
999 /* since it's about to be unmapped. */
1000 while (GC_incremental
&& GC_collection_in_progress()
1001 && (wait_for_all
|| old_gc_no
== GC_gc_no
)) {
1003 GC_in_thread_creation
= TRUE
;
1004 GC_collect_a_little_inner(1);
1005 GC_in_thread_creation
= FALSE
;
1015 /* Procedures called before and after a fork. The goal here is to make */
1016 /* it safe to call GC_malloc() in a forked child. It's unclear that is */
1017 /* attainable, since the single UNIX spec seems to imply that one */
1018 /* should only call async-signal-safe functions, and we probably can't */
1019 /* quite guarantee that. But we give it our best shot. (That same */
1020 /* spec also implies that it's not safe to call the system malloc */
1021 /* between fork() and exec(). Thus we're doing no worse than it. */
1023 /* Called before a fork() */
1024 void GC_fork_prepare_proc(void)
1026 /* Acquire all relevant locks, so that after releasing the locks */
1027 /* the child will see a consistent state in which monitor */
1028 /* invariants hold. Unfortunately, we can't acquire libc locks */
1029 /* we might need, and there seems to be no guarantee that libc */
1030 /* must install a suitable fork handler. */
1031 /* Wait for an ongoing GC to finish, since we can't finish it in */
1032 /* the (one remaining thread in) the child. */
1034 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1035 GC_wait_for_reclaim();
1037 GC_wait_for_gc_completion(TRUE
);
1038 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1039 GC_acquire_mark_lock();
1043 /* Called in parent after a fork() */
1044 void GC_fork_parent_proc(void)
1046 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1047 GC_release_mark_lock();
1052 /* Called in child after a fork() */
1053 void GC_fork_child_proc(void)
1055 /* Clean up the thread table, so that just our thread is left. */
1056 # if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1057 GC_release_mark_lock();
1059 GC_remove_all_threads_but_me();
1060 # ifdef PARALLEL_MARK
1061 /* Turn off parallel marking in the child, since we are probably */
1062 /* just going to exec, and we would have to restart mark threads. */
1064 GC_parallel
= FALSE
;
1065 # endif /* PARALLEL_MARK */
1068 #endif /* HANDLE_FORK */
1070 #if defined(GC_DGUX386_THREADS)
1071 /* Return the number of processors, or i<= 0 if it can't be determined. */
1074 /* <takis@XFree86.Org> */
1076 struct dg_sys_info_pm_info pm_sysinfo
;
1079 status
= dg_sys_info((long int *) &pm_sysinfo
,
1080 DG_SYS_INFO_PM_INFO_TYPE
, DG_SYS_INFO_PM_CURRENT_VERSION
);
1082 /* set -1 for error */
1086 numCpus
= pm_sysinfo
.idle_vp_count
;
1088 # ifdef DEBUG_THREADS
1089 GC_printf1("Number of active CPUs in this system: %d\n", numCpus
);
1093 #endif /* GC_DGUX386_THREADS */
1095 /* We hold the allocation lock. */
1098 # ifndef GC_DARWIN_THREADS
1103 if (GC_thr_initialized
) return;
1104 GC_thr_initialized
= TRUE
;
1107 /* Prepare for a possible fork. */
1108 pthread_atfork(GC_fork_prepare_proc
, GC_fork_parent_proc
,
1109 GC_fork_child_proc
);
1110 # endif /* HANDLE_FORK */
1111 /* Add the initial thread, so we can stop it. */
1112 t
= GC_new_thread(pthread_self());
1113 # ifdef GC_DARWIN_THREADS
1114 t
-> stop_info
.mach_thread
= mach_thread_self();
1116 t
-> stop_info
.stack_ptr
= (ptr_t
)(&dummy
);
1118 t
-> flags
= DETACHED
| MAIN_THREAD
;
1119 if (pthread_self () == main_pthread_self
) {
1120 t
->stack
= main_stack
;
1121 t
->stack_size
= main_stack_size
;
1122 t
->altstack
= main_altstack
;
1123 t
->altstack_size
= main_altstack_size
;
1128 /* Set GC_nprocs. */
1130 char * nprocs_string
= GETENV("GC_NPROCS");
1132 if (nprocs_string
!= NULL
) GC_nprocs
= atoi(nprocs_string
);
1134 if (GC_nprocs
<= 0) {
1135 # if defined(GC_HPUX_THREADS)
1136 GC_nprocs
= pthread_num_processors_np();
1138 # if defined(GC_OSF1_THREADS) || defined(GC_AIX_THREADS)
1139 GC_nprocs
= sysconf(_SC_NPROCESSORS_ONLN
);
1140 if (GC_nprocs
<= 0) GC_nprocs
= 1;
1142 # if defined(GC_IRIX_THREADS)
1143 GC_nprocs
= sysconf(_SC_NPROC_ONLN
);
1144 if (GC_nprocs
<= 0) GC_nprocs
= 1;
1146 # if defined(GC_DARWIN_THREADS) || defined(GC_FREEBSD_THREADS) || defined(GC_NETBSD_THREADS) || defined(GC_OPENBSD_THREADS)
1148 size_t len
= sizeof(ncpus
);
1149 sysctl((int[2]) {CTL_HW
, HW_NCPU
}, 2, &ncpus
, &len
, NULL
, 0);
1152 # if defined(GC_LINUX_THREADS) || defined(GC_DGUX386_THREADS)
1153 GC_nprocs
= GC_get_nprocs();
1156 if (GC_nprocs
<= 0) {
1157 WARN("GC_get_nprocs() returned %ld\n", GC_nprocs
);
1159 # ifdef PARALLEL_MARK
1163 # ifdef PARALLEL_MARK
1165 char * markers_string
= GETENV("GC_MARKERS");
1166 if (markers_string
!= NULL
) {
1167 GC_markers
= atoi(markers_string
);
1169 GC_markers
= GC_nprocs
;
1170 if (GC_markers
> MAX_MARKERS
)
1171 GC_markers
= MAX_MARKERS
;
1176 # ifdef PARALLEL_MARK
1178 if (GC_print_stats
) {
1179 GC_printf2("Number of processors = %ld, "
1180 "number of marker threads = %ld\n", GC_nprocs
, GC_markers
);
1183 if (GC_markers
== 1) {
1184 GC_parallel
= FALSE
;
1186 if (GC_print_stats
) {
1187 GC_printf0("Single marker thread, turning off parallel marking\n");
1192 /* Disable true incremental collection, but generational is OK. */
1193 GC_time_limit
= GC_TIME_UNLIMITED
;
1195 /* If we are using a parallel marker, actually start helper threads. */
1196 if (GC_parallel
) start_mark_threads();
1201 /* Perform all initializations, including those that */
1202 /* may require allocation. */
1203 /* Called without allocation lock. */
1204 /* Must be called before a second thread is created. */
1205 /* Called without allocation lock. */
1206 void GC_init_parallel()
1208 if (parallel_initialized
) return;
1209 parallel_initialized
= TRUE
;
1211 /* GC_init() calls us back, so set flag first. */
1212 if (!GC_is_initialized
) GC_init();
1213 /* Initialize thread local free lists if used. */
1214 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1216 GC_init_thread_local(GC_lookup_thread(pthread_self()));
1222 #if !defined(GC_DARWIN_THREADS) && !defined(GC_OPENBSD_THREADS)
1224 int WRAP_FUNC(pthread_sigmask
)(int how
, const sigset_t
*set
, sigset_t
*oset
)
1226 sigset_t fudged_set
;
1228 if (set
!= NULL
&& (how
== SIG_BLOCK
|| how
== SIG_SETMASK
)) {
1230 sigdelset(&fudged_set
, SIG_SUSPEND
);
1233 return(REAL_FUNC(pthread_sigmask
)(how
, set
, oset
));
1236 #endif /* !GC_DARWIN_THREADS */
1238 /* Wrappers for functions that are likely to block for an appreciable */
1239 /* length of time. Must be called in pairs, if at all. */
1240 /* Nothing much beyond the system call itself should be executed */
1241 /* between these. */
1243 void GC_start_blocking(void) {
1244 # define SP_SLOP 128
1247 me
= GC_lookup_thread(pthread_self());
1248 GC_ASSERT(!(me
-> thread_blocked
));
1250 me
-> stop_info
.stack_ptr
= (ptr_t
)GC_save_regs_in_stack();
1252 # ifndef GC_DARWIN_THREADS
1253 me
-> stop_info
.stack_ptr
= (ptr_t
)GC_approx_sp();
1257 me
-> backing_store_ptr
= (ptr_t
)GC_save_regs_in_stack() + SP_SLOP
;
1259 /* Add some slop to the stack pointer, since the wrapped call may */
1260 /* end up pushing more callee-save registers. */
1261 # ifndef GC_DARWIN_THREADS
1262 # ifdef STACK_GROWS_UP
1263 me
-> stop_info
.stack_ptr
+= SP_SLOP
;
1265 me
-> stop_info
.stack_ptr
-= SP_SLOP
;
1268 me
-> thread_blocked
= TRUE
;
1272 void GC_end_blocking(void) {
1274 LOCK(); /* This will block if the world is stopped. */
1275 me
= GC_lookup_thread(pthread_self());
1276 GC_ASSERT(me
-> thread_blocked
);
1277 me
-> thread_blocked
= FALSE
;
1281 #if defined(GC_DGUX386_THREADS)
1282 #define __d10_sleep sleep
1283 #endif /* GC_DGUX386_THREADS */
1285 /* A wrapper for the standard C sleep function */
1286 int WRAP_FUNC(sleep
) (unsigned int seconds
)
1290 GC_start_blocking();
1291 result
= REAL_FUNC(sleep
)(seconds
);
1297 void *(*start_routine
)(void *);
1300 sem_t registered
; /* 1 ==> in our thread table, but */
1301 /* parent hasn't yet noticed. */
1304 /* Called at thread exit. */
1305 /* Never called for main thread. That's OK, since it */
1306 /* results in at most a tiny one-time leak. And */
1307 /* linuxthreads doesn't reclaim the main threads */
1308 /* resources or id anyway. */
1309 void GC_thread_exit_proc(void *arg
)
1314 me
= GC_lookup_thread(pthread_self());
1315 GC_destroy_thread_local(me
);
1316 if (me
-> flags
& DETACHED
) {
1317 # ifdef THREAD_LOCAL_ALLOC
1318 /* NULL out the tls key to prevent the dtor function from being called */
1319 if (0 != GC_setspecific(GC_thread_key
, NULL
))
1320 ABORT("Failed to set thread specific allocation pointers");
1322 GC_delete_thread(pthread_self());
1324 me
-> flags
|= FINISHED
;
1326 # if defined(THREAD_LOCAL_ALLOC) && !defined(USE_PTHREAD_SPECIFIC) \
1327 && !defined(USE_COMPILER_TLS) && !defined(DBG_HDRS_ALL)
1328 GC_remove_specific(GC_thread_key
);
1330 /* The following may run the GC from "nonexistent" thread. */
1331 GC_wait_for_gc_completion(FALSE
);
1335 int WRAP_FUNC(pthread_join
)(pthread_t thread
, void **retval
)
1338 GC_thread thread_gc_id
;
1341 thread_gc_id
= GC_lookup_thread(thread
);
1342 /* This is guaranteed to be the intended one, since the thread id */
1343 /* cant have been recycled by pthreads. */
1345 result
= REAL_FUNC(pthread_join
)(thread
, retval
);
1346 # if defined (GC_FREEBSD_THREADS)
1347 /* On FreeBSD, the wrapped pthread_join() sometimes returns (what
1348 appears to be) a spurious EINTR which caused the test and real code
1349 to gratuitously fail. Having looked at system pthread library source
1350 code, I see how this return code may be generated. In one path of
1351 code, pthread_join() just returns the errno setting of the thread
1352 being joined. This does not match the POSIX specification or the
1353 local man pages thus I have taken the liberty to catch this one
1354 spurious return value properly conditionalized on GC_FREEBSD_THREADS. */
1355 if (result
== EINTR
) result
= 0;
1359 /* Here the pthread thread id may have been recycled. */
1360 GC_delete_gc_thread(thread
, thread_gc_id
);
1367 /* TODO: remove, NaCl glibc now supports pthread cleanup functions. */
1369 WRAP_FUNC(pthread_exit
)(void *status
)
1371 REAL_FUNC(pthread_exit
)(status
);
1376 WRAP_FUNC(pthread_detach
)(pthread_t thread
)
1379 GC_thread thread_gc_id
;
1382 thread_gc_id
= GC_lookup_thread(thread
);
1384 result
= REAL_FUNC(pthread_detach
)(thread
);
1387 thread_gc_id
-> flags
|= DETACHED
;
1388 /* Here the pthread thread id may have been recycled. */
1389 if (thread_gc_id
-> flags
& FINISHED
) {
1390 GC_delete_gc_thread(thread
, thread_gc_id
);
1397 GC_bool GC_in_thread_creation
= FALSE
;
1399 typedef void *(*ThreadStartFn
)(void *);
1400 void * GC_start_routine_head(void * arg
, void *base_addr
,
1401 ThreadStartFn
*start
, void **start_arg
)
1403 struct start_info
* si
= arg
;
1406 pthread_t my_pthread
;
1408 my_pthread
= pthread_self();
1409 # ifdef DEBUG_THREADS
1410 GC_printf1("Starting thread 0x%lx\n", my_pthread
);
1411 GC_printf1("pid = %ld\n", (long) getpid());
1412 GC_printf1("sp = 0x%lx\n", (long) &arg
);
1415 GC_in_thread_creation
= TRUE
;
1416 me
= GC_new_thread(my_pthread
);
1417 GC_in_thread_creation
= FALSE
;
1418 #ifdef GC_DARWIN_THREADS
1419 me
-> stop_info
.mach_thread
= mach_thread_self();
1421 me
-> stop_info
.stack_ptr
= 0;
1423 me
-> flags
= si
-> flags
;
1424 /* me -> stack_end = GC_linux_stack_base(); -- currently (11/99) */
1425 /* doesn't work because the stack base in /proc/self/stat is the */
1426 /* one for the main thread. There is a strong argument that that's */
1427 /* a kernel bug, but a pervasive one. */
1428 # ifdef STACK_GROWS_DOWN
1429 me
-> stack_end
= (ptr_t
)(((word
)(base_addr
) + (GC_page_size
- 1))
1430 & ~(GC_page_size
- 1));
1431 # ifndef GC_DARWIN_THREADS
1432 me
-> stop_info
.stack_ptr
= me
-> stack_end
- 0x10;
1434 /* Needs to be plausible, since an asynchronous stack mark */
1435 /* should not crash. */
1437 me
-> stack_end
= (ptr_t
)((word
)(base_addr
) & ~(GC_page_size
- 1));
1438 me
-> stop_info
.stack_ptr
= me
-> stack_end
+ 0x10;
1440 /* This is dubious, since we may be more than a page into the stack, */
1441 /* and hence skip some of it, though it's not clear that matters. */
1443 me
-> backing_store_end
= (ptr_t
)
1444 (GC_save_regs_in_stack() & ~(GC_page_size
- 1));
1445 /* This is also < 100% convincing. We should also read this */
1446 /* from /proc, but the hook to do so isn't there yet. */
1450 if (start
) *start
= si
-> start_routine
;
1451 if (start_arg
) *start_arg
= si
-> arg
;
1453 if (!(si
->flags
& FOREIGN_THREAD
))
1454 sem_post(&(si
-> registered
)); /* Last action on si. */
1455 /* OK to deallocate. */
1456 # if defined(THREAD_LOCAL_ALLOC) && !defined(DBG_HDRS_ALL)
1458 GC_init_thread_local(me
);
1465 void GC_allow_register_threads (void)
1467 /* No-op for GC pre-v7. */
1470 int GC_register_my_thread (struct GC_stack_base
*sb
)
1472 struct start_info si
= { 0, }; /* stacked for legibility & locking */
1475 # ifdef DEBUG_THREADS
1476 GC_printf1( "GC_register_my_thread %p\n", &si
);
1479 si
.flags
= FOREIGN_THREAD
;
1481 if (!parallel_initialized
) GC_init_parallel();
1483 if (!GC_thr_initialized
) GC_thr_init();
1484 me
= GC_lookup_thread(pthread_self());
1487 return GC_DUPLICATE
;
1489 (void)GC_start_routine_head(&si
, sb
-> mem_base
, NULL
, NULL
);
1493 void * GC_start_routine(void * arg
)
1496 struct start_info
* si
= arg
;
1499 ThreadStartFn start
;
1502 me
= GC_start_routine_head (arg
, &dummy
, &start
, &start_arg
);
1504 pthread_cleanup_push(GC_thread_exit_proc
, 0);
1505 # ifdef DEBUG_THREADS
1506 GC_printf1("start_routine = 0x%lx\n", start
);
1508 result
= (*start
)(start_arg
);
1510 GC_printf1("Finishing thread 0x%x\n", pthread_self());
1512 me
-> status
= result
;
1513 pthread_cleanup_pop(1);
1514 /* Cleanup acquires lock, ensuring that we can't exit */
1515 /* while a collection that thinks we're alive is trying to stop */
1521 WRAP_FUNC(pthread_create
)(pthread_t
*new_thread
,
1522 const pthread_attr_t
*attr
,
1523 void *(*start_routine
)(void *), void *arg
)
1528 struct start_info
* si
;
1529 /* This is otherwise saved only in an area mmapped by the thread */
1530 /* library, which isn't visible to the collector. */
1532 /* We resist the temptation to muck with the stack size here, */
1533 /* even if the default is unreasonably small. That's the client's */
1534 /* responsibility. */
1537 si
= (struct start_info
*)GC_INTERNAL_MALLOC(sizeof(struct start_info
),
1540 if (!parallel_initialized
) GC_init_parallel();
1541 if (0 == si
) return(ENOMEM
);
1542 sem_init(&(si
-> registered
), 0, 0);
1543 si
-> start_routine
= start_routine
;
1546 if (!GC_thr_initialized
) GC_thr_init();
1547 # ifdef GC_ASSERTIONS
1551 pthread_attr_t my_attr
;
1552 pthread_attr_init(&my_attr
);
1553 pthread_attr_getstacksize(&my_attr
, &stack_size
);
1555 pthread_attr_getstacksize(attr
, &stack_size
);
1557 # ifdef PARALLEL_MARK
1558 GC_ASSERT(stack_size
>= (8*HBLKSIZE
*sizeof(word
)));
1560 /* FreeBSD-5.3/Alpha: default pthread stack is 64K, */
1561 /* HBLKSIZE=8192, sizeof(word)=8 */
1562 GC_ASSERT(stack_size
>= 65536);
1564 /* Our threads may need to do some work for the GC. */
1565 /* Ridiculously small threads won't work, and they */
1566 /* probably wouldn't work anyway. */
1570 detachstate
= PTHREAD_CREATE_JOINABLE
;
1572 pthread_attr_getdetachstate(attr
, &detachstate
);
1574 if (PTHREAD_CREATE_DETACHED
== detachstate
) my_flags
|= DETACHED
;
1575 si
-> flags
= my_flags
;
1577 # ifdef DEBUG_THREADS
1578 GC_printf1("About to start new thread from thread 0x%X\n",
1582 result
= REAL_FUNC(pthread_create
)(new_thread
, attr
, GC_start_routine
, si
);
1584 # ifdef DEBUG_THREADS
1585 GC_printf1("Started thread 0x%X\n", *new_thread
);
1587 /* Wait until child has been added to the thread table. */
1588 /* This also ensures that we hold onto si until the child is done */
1589 /* with it. Thus it doesn't matter whether it is otherwise */
1590 /* visible to the collector. */
1592 while (0 != sem_wait(&(si
-> registered
))) {
1593 if (EINTR
!= errno
) ABORT("sem_wait failed");
1596 sem_destroy(&(si
-> registered
));
1598 GC_INTERNAL_FREE(si
);
1604 #ifdef GENERIC_COMPARE_AND_SWAP
1605 pthread_mutex_t GC_compare_and_swap_lock
= PTHREAD_MUTEX_INITIALIZER
;
1607 GC_bool
GC_compare_and_exchange(volatile GC_word
*addr
,
1608 GC_word old
, GC_word new_val
)
1611 pthread_mutex_lock(&GC_compare_and_swap_lock
);
1618 pthread_mutex_unlock(&GC_compare_and_swap_lock
);
1622 GC_word
GC_atomic_add(volatile GC_word
*addr
, GC_word how_much
)
1625 pthread_mutex_lock(&GC_compare_and_swap_lock
);
1627 *addr
= old
+ how_much
;
1628 pthread_mutex_unlock(&GC_compare_and_swap_lock
);
1632 #endif /* GENERIC_COMPARE_AND_SWAP */
1633 /* Spend a few cycles in a way that can't introduce contention with */
1634 /* othre threads. */
1638 # if !defined(__GNUC__) || defined(__INTEL_COMPILER)
1639 volatile word dummy
= 0;
1642 for (i
= 0; i
< 10; ++i
) {
1643 # if defined(__GNUC__) && !defined(__INTEL_COMPILER)
1644 __asm__
__volatile__ (" " : : : "memory");
1646 /* Something that's unlikely to be optimized away. */
1652 #define SPIN_MAX 128 /* Maximum number of calls to GC_pause before */
1655 VOLATILE GC_bool GC_collecting
= 0;
1656 /* A hint that we're in the collector and */
1657 /* holding the allocation lock for an */
1658 /* extended period. */
1660 #if !defined(USE_SPIN_LOCK) || defined(PARALLEL_MARK)
1661 /* If we don't want to use the below spinlock implementation, either */
1662 /* because we don't have a GC_test_and_set implementation, or because */
1663 /* we don't want to risk sleeping, we can still try spinning on */
1664 /* pthread_mutex_trylock for a while. This appears to be very */
1665 /* beneficial in many cases. */
1666 /* I suspect that under high contention this is nearly always better */
1667 /* than the spin lock. But it's a bit slower on a uniprocessor. */
1668 /* Hence we still default to the spin lock. */
1669 /* This is also used to acquire the mark lock for the parallel */
1672 /* Here we use a strict exponential backoff scheme. I don't know */
1673 /* whether that's better or worse than the above. We eventually */
1674 /* yield by calling pthread_mutex_lock(); it never makes sense to */
1675 /* explicitly sleep. */
1679 unsigned long GC_spin_count
= 0;
1680 unsigned long GC_block_count
= 0;
1681 unsigned long GC_unlocked_count
= 0;
1684 void GC_generic_lock(pthread_mutex_t
* lock
)
1686 #ifndef NO_PTHREAD_TRYLOCK
1687 unsigned pause_length
= 1;
1690 if (0 == pthread_mutex_trylock(lock
)) {
1692 ++GC_unlocked_count
;
1696 for (; pause_length
<= SPIN_MAX
; pause_length
<<= 1) {
1697 for (i
= 0; i
< pause_length
; ++i
) {
1700 switch(pthread_mutex_trylock(lock
)) {
1709 ABORT("Unexpected error from pthread_mutex_trylock");
1712 #endif /* !NO_PTHREAD_TRYLOCK */
1716 pthread_mutex_lock(lock
);
1719 #endif /* !USE_SPIN_LOCK || PARALLEL_MARK */
1721 #if defined(USE_SPIN_LOCK)
1723 /* Reasonably fast spin locks. Basically the same implementation */
1724 /* as STL alloc.h. This isn't really the right way to do this. */
1725 /* but until the POSIX scheduling mess gets straightened out ... */
1727 volatile unsigned int GC_allocate_lock
= 0;
1732 # define low_spin_max 30 /* spin cycles if we suspect uniprocessor */
1733 # define high_spin_max SPIN_MAX /* spin cycles for multiprocessor */
1734 static unsigned spin_max
= low_spin_max
;
1735 unsigned my_spin_max
;
1736 static unsigned last_spins
= 0;
1737 unsigned my_last_spins
;
1740 if (!GC_test_and_set(&GC_allocate_lock
)) {
1743 my_spin_max
= spin_max
;
1744 my_last_spins
= last_spins
;
1745 for (i
= 0; i
< my_spin_max
; i
++) {
1746 if (GC_collecting
|| GC_nprocs
== 1) goto yield
;
1747 if (i
< my_last_spins
/2 || GC_allocate_lock
) {
1751 if (!GC_test_and_set(&GC_allocate_lock
)) {
1754 * Spinning worked. Thus we're probably not being scheduled
1755 * against the other process with which we were contending.
1756 * Thus it makes sense to spin longer the next time.
1759 spin_max
= high_spin_max
;
1763 /* We are probably being scheduled against the other process. Sleep. */
1764 spin_max
= low_spin_max
;
1767 if (!GC_test_and_set(&GC_allocate_lock
)) {
1770 # define SLEEP_THRESHOLD 12
1771 /* Under Linux very short sleeps tend to wait until */
1772 /* the current time quantum expires. On old Linux */
1773 /* kernels nanosleep(<= 2ms) just spins under Linux. */
1774 /* (Under 2.4, this happens only for real-time */
1775 /* processes.) We want to minimize both behaviors */
1777 if (i
< SLEEP_THRESHOLD
) {
1783 /* Don't wait for more than about 15msecs, even */
1784 /* under extreme contention. */
1786 ts
.tv_nsec
= 1 << i
;
1792 #else /* !USE_SPINLOCK */
1795 #ifndef NO_PTHREAD_TRYLOCK
1796 if (1 == GC_nprocs
|| GC_collecting
) {
1797 pthread_mutex_lock(&GC_allocate_ml
);
1799 GC_generic_lock(&GC_allocate_ml
);
1801 #else /* !NO_PTHREAD_TRYLOCK */
1802 pthread_mutex_lock(&GC_allocate_ml
);
1803 #endif /* !NO_PTHREAD_TRYLOCK */
1806 #endif /* !USE_SPINLOCK */
1808 #if defined(PARALLEL_MARK) || defined(THREAD_LOCAL_ALLOC)
1810 #ifdef GC_ASSERTIONS
1811 pthread_t GC_mark_lock_holder
= NO_THREAD
;
1815 /* Ugly workaround for a linux threads bug in the final versions */
1816 /* of glibc2.1. Pthread_mutex_trylock sets the mutex owner */
1817 /* field even when it fails to acquire the mutex. This causes */
1818 /* pthread_cond_wait to die. Remove for glibc2.2. */
1819 /* According to the man page, we should use */
1820 /* PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP, but that isn't actually */
1822 static pthread_mutex_t mark_mutex
=
1823 {0, 0, 0, PTHREAD_MUTEX_ERRORCHECK_NP
, {0, 0}};
1825 static pthread_mutex_t mark_mutex
= PTHREAD_MUTEX_INITIALIZER
;
1828 static pthread_cond_t builder_cv
= PTHREAD_COND_INITIALIZER
;
1830 void GC_acquire_mark_lock()
1833 if (pthread_mutex_lock(&mark_mutex) != 0) {
1834 ABORT("pthread_mutex_lock failed");
1837 GC_generic_lock(&mark_mutex
);
1838 # ifdef GC_ASSERTIONS
1839 GC_mark_lock_holder
= pthread_self();
1843 void GC_release_mark_lock()
1845 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1846 # ifdef GC_ASSERTIONS
1847 GC_mark_lock_holder
= NO_THREAD
;
1849 if (pthread_mutex_unlock(&mark_mutex
) != 0) {
1850 ABORT("pthread_mutex_unlock failed");
1854 /* Collector must wait for a freelist builders for 2 reasons: */
1855 /* 1) Mark bits may still be getting examined without lock. */
1856 /* 2) Partial free lists referenced only by locals may not be scanned */
1857 /* correctly, e.g. if they contain "pointer-free" objects, since the */
1858 /* free-list link may be ignored. */
1859 void GC_wait_builder()
1861 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1862 # ifdef GC_ASSERTIONS
1863 GC_mark_lock_holder
= NO_THREAD
;
1865 if (pthread_cond_wait(&builder_cv
, &mark_mutex
) != 0) {
1866 ABORT("pthread_cond_wait failed");
1868 GC_ASSERT(GC_mark_lock_holder
== NO_THREAD
);
1869 # ifdef GC_ASSERTIONS
1870 GC_mark_lock_holder
= pthread_self();
1874 void GC_wait_for_reclaim()
1876 GC_acquire_mark_lock();
1877 while (GC_fl_builder_count
> 0) {
1880 GC_release_mark_lock();
1883 void GC_notify_all_builder()
1885 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1886 if (pthread_cond_broadcast(&builder_cv
) != 0) {
1887 ABORT("pthread_cond_broadcast failed");
1891 #endif /* PARALLEL_MARK || THREAD_LOCAL_ALLOC */
1893 #ifdef PARALLEL_MARK
1895 static pthread_cond_t mark_cv
= PTHREAD_COND_INITIALIZER
;
1897 void GC_wait_marker()
1899 GC_ASSERT(GC_mark_lock_holder
== pthread_self());
1900 # ifdef GC_ASSERTIONS
1901 GC_mark_lock_holder
= NO_THREAD
;
1903 if (pthread_cond_wait(&mark_cv
, &mark_mutex
) != 0) {
1904 ABORT("pthread_cond_wait failed");
1906 GC_ASSERT(GC_mark_lock_holder
== NO_THREAD
);
1907 # ifdef GC_ASSERTIONS
1908 GC_mark_lock_holder
= pthread_self();
1912 void GC_notify_all_marker()
1914 if (pthread_cond_broadcast(&mark_cv
) != 0) {
1915 ABORT("pthread_cond_broadcast failed");
1919 #endif /* PARALLEL_MARK */
1921 # endif /* GC_LINUX_THREADS and friends */