2 * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
3 * Copyright (c) 1991-1994 by Xerox Corporation. All rights reserved.
4 * Copyright (c) 1996 by Silicon Graphics. All rights reserved.
5 * Copyright (c) 2000 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 #include "private/gc_priv.h"
18 #include "gc_inline.h" /* for GC_malloc_kind */
21 * These are extra allocation routines which are likely to be less
22 * frequently used than those in malloc.c. They are separate in the
23 * hope that the .o file will be excluded from statically linked
24 * executables. We should probably break this up further.
34 /* Some externally visible but unadvertised variables to allow access to */
35 /* free lists from inlined allocators without including gc_priv.h */
36 /* or introducing dependencies on internal data structure layouts. */
37 #include "gc_alloc_ptrs.h"
38 void ** const GC_objfreelist_ptr = GC_objfreelist;
39 void ** const GC_aobjfreelist_ptr = GC_aobjfreelist;
40 void ** const GC_uobjfreelist_ptr = GC_uobjfreelist;
41 # ifdef GC_ATOMIC_UNCOLLECTABLE
42 void ** const GC_auobjfreelist_ptr = GC_auobjfreelist;
45 GC_API int GC_CALL GC_get_kind_and_size(const void * p, size_t * psize)
50 *psize = (size_t)hhdr->hb_sz;
52 return hhdr -> hb_obj_kind;
55 GC_API GC_ATTR_MALLOC void * GC_CALL GC_generic_or_special_malloc(size_t lb,
61 return GC_malloc_kind(lb, knd);
63 # ifdef GC_ATOMIC_UNCOLLECTABLE
66 return GC_generic_malloc_uncollectable(lb, knd);
68 return GC_generic_malloc(lb, knd);
72 /* Change the size of the block pointed to by p to contain at least */
73 /* lb bytes. The object may be (and quite likely will be) moved. */
74 /* The kind (e.g. atomic) is the same as that of the old. */
75 /* Shrinking of large blocks is not implemented well. */
76 GC_API void * GC_CALL GC_realloc(void * p, size_t lb)
81 size_t sz; /* Current size in bytes */
82 size_t orig_sz; /* Original sz in bytes */
85 if (p == 0) return(GC_malloc(lb)); /* Required by ANSI */
86 if (0 == lb) /* and p != NULL */ {
94 sz = (size_t)hhdr->hb_sz;
95 obj_kind = hhdr -> hb_obj_kind;
98 if (sz > MAXOBJBYTES) {
99 /* Round it up to the next whole heap block */
100 word descr = GC_obj_kinds[obj_kind].ok_descriptor;
102 sz = (sz + HBLKSIZE-1) & ~HBLKMASK;
103 if (GC_obj_kinds[obj_kind].ok_relocate_descr)
105 /* GC_realloc might be changing the block size while */
106 /* GC_reclaim_block or GC_clear_hdr_marks is examining it. */
107 /* The change to the size field is benign, in that GC_reclaim */
108 /* (and GC_clear_hdr_marks) would work correctly with either */
109 /* value, since we are not changing the number of objects in */
110 /* the block. But seeing a half-updated value (though unlikely */
111 /* to occur in practice) could be probably bad. */
112 /* Using unordered atomic accesses on the size and hb_descr */
113 /* fields would solve the issue. (The alternate solution might */
114 /* be to initially overallocate large objects, so we do not */
115 /* have to adjust the size in GC_realloc, if they still fit. */
116 /* But that is probably more expensive, since we may end up */
117 /* scanning a bunch of zeros during GC.) */
118 # ifdef AO_HAVE_store
119 GC_STATIC_ASSERT(sizeof(hhdr->hb_sz) == sizeof(AO_t));
120 AO_store((volatile AO_t *)&hhdr->hb_sz, (AO_t)sz);
121 AO_store((volatile AO_t *)&hhdr->hb_descr, (AO_t)descr);
128 hhdr -> hb_descr = descr;
133 # ifdef MARK_BIT_PER_OBJ
134 GC_ASSERT(hhdr -> hb_inv_sz == LARGE_INV_SZ);
136 # ifdef MARK_BIT_PER_GRANULE
137 GC_ASSERT((hhdr -> hb_flags & LARGE_BLOCK) != 0
138 && hhdr -> hb_map[ANY_INDEX] == 1);
140 if (IS_UNCOLLECTABLE(obj_kind)) GC_non_gc_bytes += (sz - orig_sz);
141 /* Extra area is already cleared by GC_alloc_large_and_clear. */
143 if (ADD_SLOP(lb) <= sz) {
144 if (lb >= (sz >> 1)) {
146 /* Clear unneeded part of object to avoid bogus pointer */
148 BZERO(((ptr_t)p) + lb, orig_sz - lb);
155 result = GC_generic_or_special_malloc((word)lb, obj_kind);
156 if (result != NULL) {
157 /* In case of shrink, it could also return original object. */
158 /* But this gives the client warning of imminent disaster. */
159 BCOPY(p, result, sz);
167 # if defined(REDIRECT_MALLOC) && !defined(REDIRECT_REALLOC)
168 # define REDIRECT_REALLOC GC_realloc
171 # ifdef REDIRECT_REALLOC
173 /* As with malloc, avoid two levels of extra calls here. */
174 # define GC_debug_realloc_replacement(p, lb) \
175 GC_debug_realloc(p, lb, GC_DBG_EXTRAS)
177 # if !defined(REDIRECT_MALLOC_IN_HEADER)
178 void * realloc(void * p, size_t lb)
180 return(REDIRECT_REALLOC(p, lb));
184 # undef GC_debug_realloc_replacement
185 # endif /* REDIRECT_REALLOC */
187 /* Allocate memory such that only pointers to near the */
188 /* beginning of the object are considered. */
189 /* We avoid holding allocation lock while we clear the memory. */
190 GC_API GC_ATTR_MALLOC void * GC_CALL
191 GC_generic_malloc_ignore_off_page(size_t lb, int k)
201 return GC_generic_malloc(lb, k);
202 GC_ASSERT(k < MAXOBJKINDS);
203 lg = ROUNDED_UP_GRANULES(lb);
204 lb_rounded = GRANULES_TO_BYTES(lg);
205 n_blocks = OBJ_SZ_TO_BLOCKS(lb_rounded);
206 init = GC_obj_kinds[k].ok_init;
207 if (EXPECT(GC_have_errors, FALSE))
208 GC_print_all_errors();
209 GC_INVOKE_FINALIZERS();
210 GC_DBG_COLLECT_AT_MALLOC(lb);
212 result = (ptr_t)GC_alloc_large(ADD_SLOP(lb), k, IGNORE_OFF_PAGE);
213 if (NULL == result) {
214 GC_oom_func oom_fn = GC_oom_fn;
216 return (*oom_fn)(lb);
219 if (GC_debugging_started) {
220 BZERO(result, n_blocks * HBLKSIZE);
223 /* Clear any memory that might be used for GC descriptors */
224 /* before we release the lock. */
225 ((word *)result)[0] = 0;
226 ((word *)result)[1] = 0;
227 ((word *)result)[GRANULES_TO_WORDS(lg)-1] = 0;
228 ((word *)result)[GRANULES_TO_WORDS(lg)-2] = 0;
231 GC_bytes_allocd += lb_rounded;
233 if (init && !GC_debugging_started) {
234 BZERO(result, n_blocks * HBLKSIZE);
239 GC_API GC_ATTR_MALLOC void * GC_CALL GC_malloc_ignore_off_page(size_t lb)
241 return GC_generic_malloc_ignore_off_page(lb, NORMAL);
244 GC_API GC_ATTR_MALLOC void * GC_CALL
245 GC_malloc_atomic_ignore_off_page(size_t lb)
247 return GC_generic_malloc_ignore_off_page(lb, PTRFREE);
250 /* Increment GC_bytes_allocd from code that doesn't have direct access */
252 GC_API void GC_CALL GC_incr_bytes_allocd(size_t n)
254 GC_bytes_allocd += n;
257 /* The same for GC_bytes_freed. */
258 GC_API void GC_CALL GC_incr_bytes_freed(size_t n)
263 GC_API size_t GC_CALL GC_get_expl_freed_bytes_since_gc(void)
265 return (size_t)GC_bytes_freed;
268 # ifdef PARALLEL_MARK
269 STATIC volatile AO_t GC_bytes_allocd_tmp = 0;
270 /* Number of bytes of memory allocated since */
271 /* we released the GC lock. Instead of */
272 /* reacquiring the GC lock just to add this in, */
273 /* we add it in the next time we reacquire */
274 /* the lock. (Atomically adding it doesn't */
275 /* work, since we would have to atomically */
276 /* update it in GC_malloc, which is too */
278 # endif /* PARALLEL_MARK */
280 /* Return a list of 1 or more objects of the indicated size, linked */
281 /* through the first word in the object. This has the advantage that */
282 /* it acquires the allocation lock only once, and may greatly reduce */
283 /* time wasted contending for the allocation lock. Typical usage would */
284 /* be in a thread that requires many items of the same size. It would */
285 /* keep its own free list in thread-local storage, and call */
286 /* GC_malloc_many or friends to replenish it. (We do not round up */
287 /* object sizes, since a call indicates the intention to consume many */
288 /* objects of exactly this size.) */
289 /* We assume that the size is a multiple of GRANULE_BYTES. */
290 /* We return the free-list by assigning it to *result, since it is */
291 /* not safe to return, e.g. a linked list of pointer-free objects, */
292 /* since the collector would not retain the entire list if it were */
293 /* invoked just as we were returning. */
294 /* Note that the client should usually clear the link field. */
295 GC_API void GC_CALL GC_generic_malloc_many(size_t lb, int k, void **result)
300 size_t lw; /* Length in words. */
301 size_t lg; /* Length in granules. */
302 signed_word my_bytes_allocd = 0;
303 struct obj_kind * ok = &(GC_obj_kinds[k]);
307 GC_ASSERT(lb != 0 && (lb & (GRANULE_BYTES-1)) == 0);
308 /* Currently a single object is always allocated if manual VDB. */
309 /* TODO: GC_dirty should be called for each linked object (but */
310 /* the last one) to support multiple objects allocation. */
311 if (!SMALL_OBJ(lb) || GC_manual_vdb) {
312 op = GC_generic_malloc(lb, k);
313 if (EXPECT(0 != op, TRUE))
316 # ifndef GC_DISABLE_INCREMENTAL
317 if (GC_manual_vdb && GC_is_heap_ptr(result)) {
318 GC_dirty_inner(result);
319 REACHABLE_AFTER_DIRTY(op);
324 GC_ASSERT(k < MAXOBJKINDS);
325 lw = BYTES_TO_WORDS(lb);
326 lg = BYTES_TO_GRANULES(lb);
327 if (EXPECT(GC_have_errors, FALSE))
328 GC_print_all_errors();
329 GC_INVOKE_FINALIZERS();
330 GC_DBG_COLLECT_AT_MALLOC(lb);
331 if (!EXPECT(GC_is_initialized, TRUE)) GC_init();
333 /* Do our share of marking work */
334 if (GC_incremental && !GC_dont_gc) {
336 GC_collect_a_little_inner(1);
339 /* First see if we can reclaim a page of objects waiting to be */
341 rlh = ok -> ok_reclaim_list;
346 for (rlh += lg; (hbp = *rlh) != NULL; ) {
348 *rlh = hhdr -> hb_next;
349 GC_ASSERT(hhdr -> hb_sz == lb);
350 hhdr -> hb_last_reclaimed = (unsigned short) GC_gc_no;
351 # ifdef PARALLEL_MARK
353 signed_word my_bytes_allocd_tmp =
354 (signed_word)AO_load(&GC_bytes_allocd_tmp);
355 GC_ASSERT(my_bytes_allocd_tmp >= 0);
356 /* We only decrement it while holding the GC lock. */
357 /* Thus we can't accidentally adjust it down in more */
358 /* than one thread simultaneously. */
360 if (my_bytes_allocd_tmp != 0) {
361 (void)AO_fetch_and_add(&GC_bytes_allocd_tmp,
362 (AO_t)(-my_bytes_allocd_tmp));
363 GC_bytes_allocd += my_bytes_allocd_tmp;
365 GC_acquire_mark_lock();
366 ++ GC_fl_builder_count;
368 GC_release_mark_lock();
371 op = GC_reclaim_generic(hbp, hhdr, lb,
372 ok -> ok_init, 0, &my_bytes_allocd);
374 # ifdef PARALLEL_MARK
377 (void)AO_fetch_and_add(&GC_bytes_allocd_tmp,
378 (AO_t)my_bytes_allocd);
379 GC_acquire_mark_lock();
380 -- GC_fl_builder_count;
381 if (GC_fl_builder_count == 0) GC_notify_all_builder();
382 # ifdef THREAD_SANITIZER
383 GC_release_mark_lock();
385 GC_bytes_found += my_bytes_allocd;
388 GC_bytes_found += my_bytes_allocd;
389 /* The result may be inaccurate. */
390 GC_release_mark_lock();
392 (void) GC_clear_stack(0);
396 /* We also reclaimed memory, so we need to adjust */
398 GC_bytes_found += my_bytes_allocd;
399 GC_bytes_allocd += my_bytes_allocd;
402 # ifdef PARALLEL_MARK
404 GC_acquire_mark_lock();
405 -- GC_fl_builder_count;
406 if (GC_fl_builder_count == 0) GC_notify_all_builder();
407 GC_release_mark_lock();
409 /* GC lock is needed for reclaim list access. We */
410 /* must decrement fl_builder_count before reacquiring */
411 /* the lock. Hopefully this path is rare. */
416 /* Next try to use prefix of global free list if there is one. */
417 /* We don't refill it, but we need to use it up before allocating */
418 /* a new block ourselves. */
419 opp = &(GC_obj_kinds[k].ok_freelist[lg]);
420 if ( (op = *opp) != 0 ) {
423 for (p = op; p != 0; p = obj_link(p)) {
424 my_bytes_allocd += lb;
425 if ((word)my_bytes_allocd >= HBLKSIZE) {
431 GC_bytes_allocd += my_bytes_allocd;
434 /* Next try to allocate a new block worth of objects of this size. */
436 struct hblk *h = GC_allochblk(lb, k, 0);
437 if (h /* != NULL */) { /* CPPCHECK */
438 if (IS_UNCOLLECTABLE(k)) GC_set_hdr_marks(HDR(h));
439 GC_bytes_allocd += HBLKSIZE - HBLKSIZE % lb;
440 # ifdef PARALLEL_MARK
442 GC_acquire_mark_lock();
443 ++ GC_fl_builder_count;
445 GC_release_mark_lock();
447 op = GC_build_fl(h, lw,
448 (ok -> ok_init || GC_debugging_started), 0);
451 GC_acquire_mark_lock();
452 -- GC_fl_builder_count;
453 if (GC_fl_builder_count == 0) GC_notify_all_builder();
454 GC_release_mark_lock();
455 (void) GC_clear_stack(0);
459 op = GC_build_fl(h, lw, (ok -> ok_init || GC_debugging_started), 0);
464 /* As a last attempt, try allocating a single object. Note that */
465 /* this may trigger a collection or expand the heap. */
466 op = GC_generic_malloc_inner(lb, k);
467 if (0 != op) obj_link(op) = 0;
472 (void) GC_clear_stack(0);
475 /* Note that the "atomic" version of this would be unsafe, since the */
476 /* links would not be seen by the collector. */
477 GC_API GC_ATTR_MALLOC void * GC_CALL GC_malloc_many(size_t lb)
481 /* Add EXTRA_BYTES and round up to a multiple of a granule. */
482 lb = SIZET_SAT_ADD(lb, EXTRA_BYTES + GRANULE_BYTES - 1)
483 & ~(GRANULE_BYTES - 1);
485 GC_generic_malloc_many(lb, NORMAL, &result);
491 /* Debug version is tricky and currently missing. */
492 GC_API GC_ATTR_MALLOC void * GC_CALL GC_memalign(size_t align, size_t lb)
498 if (align <= GRANULE_BYTES) return GC_malloc(lb);
499 if (align >= HBLKSIZE/2 || lb >= HBLKSIZE/2) {
500 if (align > HBLKSIZE) {
501 return (*GC_get_oom_fn())(LONG_MAX-1024); /* Fail */
503 return GC_malloc(lb <= HBLKSIZE? HBLKSIZE : lb);
504 /* Will be HBLKSIZE aligned. */
506 /* We could also try to make sure that the real rounded-up object size */
507 /* is a multiple of align. That would be correct up to HBLKSIZE. */
508 new_lb = SIZET_SAT_ADD(lb, align - 1);
509 result = (ptr_t)GC_malloc(new_lb);
510 /* It is OK not to check result for NULL as in that case */
511 /* GC_memalign returns NULL too since (0 + 0 % align) is 0. */
512 offset = (word)result % align;
514 offset = align - offset;
515 if (!GC_all_interior_pointers) {
516 GC_STATIC_ASSERT(VALID_OFFSET_SZ <= HBLKSIZE);
517 GC_ASSERT(offset < VALID_OFFSET_SZ);
518 GC_register_displacement(offset);
522 GC_ASSERT((word)result % align == 0);
526 /* This one exists largely to redirect posix_memalign for leaks finding. */
527 GC_API int GC_CALL GC_posix_memalign(void **memptr, size_t align, size_t lb)
529 /* Check alignment properly. */
530 size_t align_minus_one = align - 1; /* to workaround a cppcheck warning */
531 if (align < sizeof(void *) || (align_minus_one & align) != 0) {
533 return ERROR_INVALID_PARAMETER;
539 if ((*memptr = GC_memalign(align, lb)) == NULL) {
541 return ERROR_NOT_ENOUGH_MEMORY;
549 /* provide a version of strdup() that uses the collector to allocate the
550 copy of the string */
551 GC_API GC_ATTR_MALLOC char * GC_CALL GC_strdup(const char *s)
555 if (s == NULL) return NULL;
557 copy = (char *)GC_malloc_atomic(lb);
568 GC_API GC_ATTR_MALLOC char * GC_CALL GC_strndup(const char *str, size_t size)
571 size_t len = strlen(str); /* str is expected to be non-NULL */
574 copy = (char *)GC_malloc_atomic(len + 1);
581 if (EXPECT(len > 0, TRUE))
582 BCOPY(str, copy, len);
587 #ifdef GC_REQUIRE_WCSDUP
588 # include <wchar.h> /* for wcslen() */
590 GC_API GC_ATTR_MALLOC wchar_t * GC_CALL GC_wcsdup(const wchar_t *str)
592 size_t lb = (wcslen(str) + 1) * sizeof(wchar_t);
593 wchar_t *copy = (wchar_t *)GC_malloc_atomic(lb);
601 BCOPY(str, copy, lb);
604 #endif /* GC_REQUIRE_WCSDUP */
607 GC_API void * GC_CALL GC_malloc_stubborn(size_t lb)
609 return GC_malloc(lb);
612 GC_API void GC_CALL GC_change_stubborn(const void *p GC_ATTR_UNUSED)
616 #endif /* !CPPCHECK */
618 GC_API void GC_CALL GC_end_stubborn_change(const void *p)
620 GC_dirty(p); /* entire object */
623 GC_API void GC_CALL GC_ptr_store_and_dirty(void *p, const void *q)
625 *(const void **)p = q;
627 REACHABLE_AFTER_DIRTY(q);