2 * Copyright (c) 2000-2005 by Hewlett-Packard Company. All rights reserved.
4 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
5 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
7 * Permission is hereby granted to use or copy this program
8 * for any purpose, provided the above notices are retained on all copies.
9 * Permission to modify the code and to distribute modified code is granted,
10 * provided the above notices are retained, and a notice that the code was
11 * modified is included with the above copyright notice.
13 #include "private/gc_priv.h"
15 # if defined(THREAD_LOCAL_ALLOC)
17 #include "private/thread_local_alloc.h"
18 #include "gc_inline.h"
22 #if defined(USE_COMPILER_TLS)
24 #elif defined(USE_WIN32_COMPILER_TLS)
27 GC_key_t GC_thread_key;
29 static GC_bool keys_initialized;
31 /* Return a single nonempty freelist fl to the global one pointed to */
34 static void return_single_freelist(void *fl, void **gfl)
41 GC_ASSERT(GC_size(fl) == GC_size(*gfl));
43 for (qptr = &(obj_link(fl)), q = *qptr;
44 (word)q >= HBLKSIZE; qptr = &(obj_link(q)), q = *qptr);
51 /* Recover the contents of the freelist array fl into the global one gfl.*/
52 /* We hold the allocator lock. */
53 static void return_freelists(void **fl, void **gfl)
57 for (i = 1; i < TINY_FREELISTS; ++i) {
58 if ((word)(fl[i]) >= HBLKSIZE) {
59 return_single_freelist(fl[i], gfl+i);
61 /* Clear fl[i], since the thread structure may hang around. */
62 /* Do it in a way that is likely to trap if we access it. */
63 fl[i] = (ptr_t)HBLKSIZE;
65 /* The 0 granule freelist really contains 1 granule objects. */
66 # ifdef GC_GCJ_SUPPORT
67 if (fl[0] == ERROR_FL) return;
69 if ((word)(fl[0]) >= HBLKSIZE) {
70 return_single_freelist(fl[0], gfl+1);
74 /* Each thread structure must be initialized. */
75 /* This call must be made from the new thread. */
76 void GC_init_thread_local(GC_tlfs p)
80 GC_ASSERT(I_HOLD_LOCK());
81 if (!keys_initialized) {
82 if (0 != GC_key_create(&GC_thread_key, 0)) {
83 ABORT("Failed to create key for local allocator");
85 keys_initialized = TRUE;
87 if (0 != GC_setspecific(GC_thread_key, p)) {
88 ABORT("Failed to set thread specific allocation pointers");
90 for (i = 1; i < TINY_FREELISTS; ++i) {
91 p -> ptrfree_freelists[i] = (void *)1;
92 p -> normal_freelists[i] = (void *)1;
93 # ifdef GC_GCJ_SUPPORT
94 p -> gcj_freelists[i] = (void *)1;
97 /* Set up the size 0 free lists. */
98 /* We now handle most of them like regular free lists, to ensure */
99 /* That explicit deallocation works. However, allocation of a */
100 /* size 0 "gcj" object is always an error. */
101 p -> ptrfree_freelists[0] = (void *)1;
102 p -> normal_freelists[0] = (void *)1;
103 # ifdef GC_GCJ_SUPPORT
104 p -> gcj_freelists[0] = ERROR_FL;
108 #ifdef GC_GCJ_SUPPORT
109 extern void ** GC_gcjobjfreelist;
112 /* We hold the allocator lock. */
113 void GC_destroy_thread_local(GC_tlfs p)
115 /* We currently only do this from the thread itself or from */
116 /* the fork handler for a child process. */
118 GC_ASSERT(GC_getspecific(GC_thread_key) == (void *)p);
120 return_freelists(p -> ptrfree_freelists, GC_aobjfreelist);
121 return_freelists(p -> normal_freelists, GC_objfreelist);
122 # ifdef GC_GCJ_SUPPORT
123 return_freelists(p -> gcj_freelists, GC_gcjobjfreelist);
127 #if defined(GC_ASSERTIONS) && defined(GC_PTHREADS) && !defined(CYGWIN32) \
128 && !defined(GC_WIN32_PTHREADS)
129 # include <pthread.h>
130 extern char * GC_lookup_thread(pthread_t id);
133 #if defined(GC_ASSERTIONS) && defined(GC_WIN32_THREADS)
134 extern char * GC_lookup_thread(int id);
137 void * GC_malloc(size_t bytes)
139 size_t granules = ROUNDED_UP_GRANULES(bytes);
144 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_WIN32_SPECIFIC)
145 GC_key_t k = GC_thread_key;
146 if (EXPECT(0 == k, 0)) {
147 /* We haven't yet run GC_init_parallel. That means */
148 /* we also aren't locking, so this is fairly cheap. */
149 return GC_core_malloc(bytes);
151 tsd = GC_getspecific(k);
153 tsd = GC_getspecific(GC_thread_key);
155 # if defined(USE_PTHREAD_SPECIFIC) || defined(USE_WIN32_SPECIFIC)
156 if (EXPECT(0 == tsd, 0)) {
157 return GC_core_malloc(bytes);
160 GC_ASSERT(GC_is_initialized);
161 # ifdef GC_ASSERTIONS
162 /* We can't check tsd correctly, since we don't have access to */
163 /* the right declarations. But we can check that it's close. */
166 # if defined(GC_WIN32_THREADS)
167 char * me = (char *)GC_lookup_thread_inner(GetCurrentThreadId());
169 char * me = GC_lookup_thread(pthread_self());
171 GC_ASSERT((char *)tsd > me && (char *)tsd < me + 1000);
175 tiny_fl = ((GC_tlfs)tsd) -> normal_freelists;
176 GC_FAST_MALLOC_GRANS(result, granules, tiny_fl, DIRECT_GRANULES,
177 NORMAL, GC_core_malloc(bytes), obj_link(result)=0);
179 GC_err_printf("GC_malloc(%d) = %p : %d\n", bytes, result, GC_gc_no);
184 void * GC_malloc_atomic(size_t bytes)
186 size_t granules = ROUNDED_UP_GRANULES(bytes);
191 # if !defined(USE_PTHREAD_SPECIFIC) && !defined(USE_WIN32_SPECIFIC)
192 GC_key_t k = GC_thread_key;
193 if (EXPECT(0 == k, 0)) {
194 /* We haven't yet run GC_init_parallel. That means */
195 /* we also aren't locking, so this is fairly cheap. */
196 return GC_core_malloc(bytes);
198 tsd = GC_getspecific(k);
200 tsd = GC_getspecific(GC_thread_key);
202 # if defined(USE_PTHREAD_SPECIFIC) || defined(USE_WIN32_SPECIFIC)
203 if (EXPECT(0 == tsd, 0)) {
204 return GC_core_malloc(bytes);
207 GC_ASSERT(GC_is_initialized);
208 tiny_fl = ((GC_tlfs)tsd) -> ptrfree_freelists;
209 GC_FAST_MALLOC_GRANS(result, granules, tiny_fl, DIRECT_GRANULES,
210 PTRFREE, GC_core_malloc_atomic(bytes), 0/* no init */);
214 #ifdef GC_GCJ_SUPPORT
216 #include "include/gc_gcj.h"
219 extern GC_bool GC_gcj_malloc_initialized;
222 extern int GC_gcj_kind;
224 /* Gcj-style allocation without locks is extremely tricky. The */
225 /* fundamental issue is that we may end up marking a free list, which */
226 /* has freelist links instead of "vtable" pointers. That is usually */
227 /* OK, since the next object on the free list will be cleared, and */
228 /* will thus be interpreted as containg a zero descriptor. That's fine */
229 /* if the object has not yet been initialized. But there are */
230 /* interesting potential races. */
231 /* In the case of incremental collection, this seems hopeless, since */
232 /* the marker may run asynchronously, and may pick up the pointer to */
233 /* the next freelist entry (which it thinks is a vtable pointer), get */
234 /* suspended for a while, and then see an allocated object instead */
235 /* of the vtable. This made be avoidable with either a handshake with */
236 /* the collector or, probably more easily, by moving the free list */
237 /* links to the second word of each object. The latter isn't a */
238 /* universal win, since on architecture like Itanium, nonzero offsets */
239 /* are not necessarily free. And there may be cache fill order issues. */
240 /* For now, we punt with incremental GC. This probably means that */
241 /* incremental GC should be enabled before we fork a second thread. */
242 /* Unlike the other thread local allocation calls, we assume that the */
243 /* collector has been explicitly initialized. */
244 void * GC_gcj_malloc(size_t bytes,
245 void * ptr_to_struct_containing_descr)
247 if (GC_EXPECT(GC_incremental, 0)) {
248 return GC_core_gcj_malloc(bytes, ptr_to_struct_containing_descr);
250 size_t granules = ROUNDED_UP_GRANULES(bytes);
252 void **tiny_fl = ((GC_tlfs)GC_getspecific(GC_thread_key))
254 GC_ASSERT(GC_gcj_malloc_initialized);
255 GC_FAST_MALLOC_GRANS(result, granules, tiny_fl, DIRECT_GRANULES,
257 GC_core_gcj_malloc(bytes,
258 ptr_to_struct_containing_descr),
259 {AO_compiler_barrier();
260 *(void **)result = ptr_to_struct_containing_descr;});
261 /* This forces the initialization of the "method ptr". */
262 /* This is necessary to ensure some very subtle properties */
263 /* required if a GC is run in the middle of such an allocation. */
264 /* Here we implicitly also assume atomicity for the free list. */
265 /* and method pointer assignments. */
266 /* We must update the freelist before we store the pointer. */
267 /* Otherwise a GC at this point would see a corrupted */
269 /* A real memory barrier is not needed, since the */
270 /* action of stopping this thread will cause prior writes */
272 /* We assert that any concurrent marker will stop us. */
273 /* Thus it is impossible for a mark procedure to see the */
274 /* allocation of the next object, but to see this object */
275 /* still containing a free list pointer. Otherwise the */
276 /* marker, by misinterpreting the freelist link as a vtable */
277 /* pointer, might find a random "mark descriptor" in the next */
283 #endif /* GC_GCJ_SUPPORT */
285 /* The thread support layer must arrange to mark thread-local */
286 /* free lists explicitly, since the link field is often */
287 /* invisible to the marker. It knows hoe to find all threads; */
288 /* we take care of an individual thread freelist structure. */
289 void GC_mark_thread_local_fls_for(GC_tlfs p)
294 for (j = 1; j < TINY_FREELISTS; ++j) {
295 q = p -> ptrfree_freelists[j];
296 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
297 q = p -> normal_freelists[j];
298 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
299 # ifdef GC_GCJ_SUPPORT
300 q = p -> gcj_freelists[j];
301 if ((word)q > HBLKSIZE) GC_set_fl_marks(q);
302 # endif /* GC_GCJ_SUPPORT */
306 #if defined(GC_ASSERTIONS)
307 /* Check that all thread-local free-lists in p are completely marked. */
308 void GC_check_tls_for(GC_tlfs p)
313 for (j = 1; j < TINY_FREELISTS; ++j) {
314 q = p -> ptrfree_freelists[j];
315 if ((word)q > HBLKSIZE) GC_check_fl_marks(q);
316 q = p -> normal_freelists[j];
317 if ((word)q > HBLKSIZE) GC_check_fl_marks(q);
318 # ifdef GC_GCJ_SUPPORT
319 q = p -> gcj_freelists[j];
320 if ((word)q > HBLKSIZE) GC_check_fl_marks(q);
321 # endif /* GC_GCJ_SUPPORT */
324 #endif /* GC_ASSERTIONS */
326 # else /* !THREAD_LOCAL_ALLOC */
328 # define GC_destroy_thread_local(t)
330 # endif /* !THREAD_LOCAL_ALLOC */