2 * sgen-cardtable.c: Card table implementation for sgen
5 * Rodrigo Kumpera (rkumpera@novell.com)
7 * Copyright 2001-2003 Ximian, Inc
8 * Copyright 2003-2010 Novell, Inc.
9 * Copyright 2011 Xamarin Inc (http://www.xamarin.com)
10 * Copyright (C) 2012 Xamarin Inc
12 * This library is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Library General Public
14 * License 2.0 as published by the Free Software Foundation;
16 * This library is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Library General Public License for more details.
21 * You should have received a copy of the GNU Library General Public
22 * License 2.0 along with this library; if not, write to the Free
23 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 #include "metadata/sgen-gc.h"
30 #include "metadata/sgen-cardtable.h"
31 #include "metadata/sgen-memory-governor.h"
32 #include "metadata/sgen-protocol.h"
33 #include "metadata/sgen-layout-stats.h"
34 #include "utils/mono-counters.h"
35 #include "utils/mono-time.h"
36 #include "utils/mono-memory-model.h"
38 //#define CARDTABLE_STATS
43 #ifdef HAVE_SYS_MMAN_H
46 #include <sys/types.h>
48 #define ARRAY_OBJ_INDEX(ptr,array,elem_size) (((char*)(ptr) - ((char*)(array) + G_STRUCT_OFFSET (MonoArray, vector))) / (elem_size))
50 guint8 *sgen_cardtable;
52 static gboolean need_mod_union;
54 #ifdef HEAVY_STATISTICS
55 long long marked_cards;
56 long long scanned_cards;
57 long long scanned_objects;
58 long long remarked_cards;
60 static long long los_marked_cards;
61 static long long large_objects;
62 static long long bloby_objects;
63 static long long los_array_cards;
64 static long long los_array_remsets;
67 static long long major_card_scan_time;
68 static long long los_card_scan_time;
70 static long long last_major_scan_time;
71 static long long last_los_scan_time;
73 static void sgen_card_tables_collect_stats (gboolean begin);
76 /*WARNING: This function returns the number of cards regardless of overflow in case of overlapping cards.*/
78 cards_in_range (mword address, mword size)
80 mword end = address + MAX (1, size) - 1;
81 return (end >> CARD_BITS) - (address >> CARD_BITS) + 1;
85 sgen_card_table_wbarrier_set_field (MonoObject *obj, gpointer field_ptr, MonoObject* value)
87 *(void**)field_ptr = value;
88 if (need_mod_union || sgen_ptr_in_nursery (value))
89 sgen_card_table_mark_address ((mword)field_ptr);
90 sgen_dummy_use (value);
94 sgen_card_table_wbarrier_set_arrayref (MonoArray *arr, gpointer slot_ptr, MonoObject* value)
96 *(void**)slot_ptr = value;
97 if (need_mod_union || sgen_ptr_in_nursery (value))
98 sgen_card_table_mark_address ((mword)slot_ptr);
99 sgen_dummy_use (value);
103 sgen_card_table_wbarrier_arrayref_copy (gpointer dest_ptr, gpointer src_ptr, int count)
105 gpointer *dest = dest_ptr;
106 gpointer *src = src_ptr;
108 /*overlapping that required backward copying*/
109 if (src < dest && (src + count) > dest) {
110 gpointer *start = dest;
114 for (; dest >= start; --src, --dest) {
115 gpointer value = *src;
117 if (need_mod_union || sgen_ptr_in_nursery (value))
118 sgen_card_table_mark_address ((mword)dest);
119 sgen_dummy_use (value);
122 gpointer *end = dest + count;
123 for (; dest < end; ++src, ++dest) {
124 gpointer value = *src;
126 if (need_mod_union || sgen_ptr_in_nursery (value))
127 sgen_card_table_mark_address ((mword)dest);
128 sgen_dummy_use (value);
134 sgen_card_table_wbarrier_value_copy (gpointer dest, gpointer src, int count, MonoClass *klass)
136 size_t element_size = mono_class_value_size (klass, NULL);
137 size_t size = count * element_size;
139 #ifdef DISABLE_CRITICAL_REGION
143 ENTER_CRITICAL_REGION;
145 mono_gc_memmove_atomic (dest, src, size);
146 sgen_card_table_mark_range ((mword)dest, size);
147 #ifdef DISABLE_CRITICAL_REGION
150 EXIT_CRITICAL_REGION;
155 sgen_card_table_wbarrier_object_copy (MonoObject* obj, MonoObject *src)
157 int size = mono_object_class (obj)->instance_size;
159 #ifdef DISABLE_CRITICAL_REGION
163 ENTER_CRITICAL_REGION;
165 mono_gc_memmove_aligned ((char*)obj + sizeof (MonoObject), (char*)src + sizeof (MonoObject),
166 size - sizeof (MonoObject));
167 sgen_card_table_mark_range ((mword)obj, size);
168 #ifdef DISABLE_CRITICAL_REGION
171 EXIT_CRITICAL_REGION;
176 sgen_card_table_wbarrier_generic_nostore (gpointer ptr)
178 sgen_card_table_mark_address ((mword)ptr);
181 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
183 guint8 *sgen_shadow_cardtable;
185 #define SGEN_SHADOW_CARDTABLE_END (sgen_shadow_cardtable + CARD_COUNT_IN_BYTES)
186 #define SGEN_CARDTABLE_END (sgen_cardtable + CARD_COUNT_IN_BYTES)
189 sgen_card_table_region_begin_scanning (mword start, mword end)
191 /*XXX this can be improved to work on words and have a single loop induction var */
192 while (start <= end) {
193 if (sgen_card_table_card_begin_scanning (start))
195 start += CARD_SIZE_IN_BYTES;
203 sgen_card_table_region_begin_scanning (mword start, mword size)
205 gboolean res = FALSE;
206 guint8 *card = sgen_card_table_get_card_address (start);
207 guint8 *end = card + cards_in_range (start, size);
209 /*XXX this can be improved to work on words and have a branchless body */
210 while (card != end) {
217 memset (sgen_card_table_get_card_address (start), 0, size >> CARD_BITS);
224 /*FIXME this assumes that major blocks are multiple of 4K which is pretty reasonable */
226 sgen_card_table_get_card_data (guint8 *data_dest, mword address, mword cards)
228 mword *start = (mword*)sgen_card_table_get_card_scan_address (address);
229 mword *dest = (mword*)data_dest;
230 mword *end = (mword*)(data_dest + cards);
233 for (; dest < end; ++dest, ++start) {
238 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
247 sgen_card_table_align_pointer (void *ptr)
249 return (void*)((mword)ptr & ~(CARD_SIZE_IN_BYTES - 1));
253 sgen_card_table_mark_range (mword address, mword size)
255 memset (sgen_card_table_get_card_address (address), 1, cards_in_range (address, size));
259 sgen_card_table_is_range_marked (guint8 *cards, mword address, mword size)
261 guint8 *end = cards + cards_in_range (address, size);
263 /*This is safe since this function is only called by code that only passes continuous card blocks*/
264 while (cards != end) {
273 sgen_card_table_record_pointer (gpointer address)
275 *sgen_card_table_get_card_address ((mword)address) = 1;
279 sgen_card_table_find_address (char *addr)
281 return sgen_card_table_address_is_marked ((mword)addr);
285 sgen_card_table_find_address_with_cards (char *cards_start, guint8 *cards, char *addr)
287 cards_start = sgen_card_table_align_pointer (cards_start);
288 return cards [(addr - cards_start) >> CARD_BITS];
292 update_mod_union (guint8 *dest, gboolean init, guint8 *start_card, size_t num_cards)
295 memcpy (dest, start_card, num_cards);
298 for (i = 0; i < num_cards; ++i)
299 dest [i] |= start_card [i];
304 alloc_mod_union (size_t num_cards)
306 return sgen_alloc_internal_dynamic (num_cards, INTERNAL_MEM_CARDTABLE_MOD_UNION, TRUE);
310 sgen_card_table_update_mod_union_from_cards (guint8 *dest, guint8 *start_card, size_t num_cards)
312 gboolean init = dest == NULL;
315 dest = alloc_mod_union (num_cards);
317 update_mod_union (dest, init, start_card, num_cards);
323 sgen_card_table_update_mod_union (guint8 *dest, char *obj, mword obj_size, size_t *out_num_cards)
325 guint8 *start_card = sgen_card_table_get_card_address ((mword)obj);
326 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
327 guint8 *end_card = sgen_card_table_get_card_address ((mword)obj + obj_size - 1) + 1;
330 guint8 *result = NULL;
332 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
335 rest = num_cards = cards_in_range ((mword) obj, obj_size);
337 while (start_card + rest > SGEN_CARDTABLE_END) {
338 size_t count = SGEN_CARDTABLE_END - start_card;
339 dest = sgen_card_table_update_mod_union_from_cards (dest, start_card, count);
344 start_card = sgen_cardtable;
348 num_cards = end_card - start_card;
351 dest = sgen_card_table_update_mod_union_from_cards (dest, start_card, num_cards);
356 *out_num_cards = num_cards;
361 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
364 move_cards_to_shadow_table (mword start, mword size)
366 guint8 *from = sgen_card_table_get_card_address (start);
367 guint8 *to = sgen_card_table_get_shadow_card_address (start);
368 size_t bytes = cards_in_range (start, size);
370 if (bytes >= CARD_COUNT_IN_BYTES) {
371 memcpy (sgen_shadow_cardtable, sgen_cardtable, CARD_COUNT_IN_BYTES);
372 } else if (to + bytes > SGEN_SHADOW_CARDTABLE_END) {
373 size_t first_chunk = SGEN_SHADOW_CARDTABLE_END - to;
374 size_t second_chunk = MIN (CARD_COUNT_IN_BYTES, bytes) - first_chunk;
376 memcpy (to, from, first_chunk);
377 memcpy (sgen_shadow_cardtable, sgen_cardtable, second_chunk);
379 memcpy (to, from, bytes);
384 clear_cards (mword start, mword size)
386 guint8 *addr = sgen_card_table_get_card_address (start);
387 size_t bytes = cards_in_range (start, size);
389 if (bytes >= CARD_COUNT_IN_BYTES) {
390 memset (sgen_cardtable, 0, CARD_COUNT_IN_BYTES);
391 } else if (addr + bytes > SGEN_CARDTABLE_END) {
392 size_t first_chunk = SGEN_CARDTABLE_END - addr;
394 memset (addr, 0, first_chunk);
395 memset (sgen_cardtable, 0, bytes - first_chunk);
397 memset (addr, 0, bytes);
405 clear_cards (mword start, mword size)
407 memset (sgen_card_table_get_card_address (start), 0, cards_in_range (start, size));
414 sgen_card_table_prepare_for_major_collection (void)
416 /*XXX we could do this in 2 ways. using mincore or iterating over all sections/los objects */
417 sgen_major_collector_iterate_live_block_ranges (clear_cards);
418 sgen_los_iterate_live_block_ranges (clear_cards);
422 sgen_card_table_finish_minor_collection (void)
424 sgen_card_tables_collect_stats (FALSE);
428 sgen_card_table_finish_scan_remsets (void *start_nursery, void *end_nursery, SgenGrayQueue *queue)
430 SGEN_TV_DECLARE (atv);
431 SGEN_TV_DECLARE (btv);
433 sgen_card_tables_collect_stats (TRUE);
435 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
436 /*FIXME we should have a bit on each block/los object telling if the object have marked cards.*/
438 sgen_major_collector_iterate_live_block_ranges (move_cards_to_shadow_table);
439 sgen_los_iterate_live_block_ranges (move_cards_to_shadow_table);
442 sgen_card_table_prepare_for_major_collection ();
444 SGEN_TV_GETTIME (atv);
445 sgen_major_collector_scan_card_table (queue);
446 SGEN_TV_GETTIME (btv);
447 last_major_scan_time = SGEN_TV_ELAPSED (atv, btv);
448 major_card_scan_time += last_major_scan_time;
449 sgen_los_scan_card_table (FALSE, queue);
450 SGEN_TV_GETTIME (atv);
451 last_los_scan_time = SGEN_TV_ELAPSED (btv, atv);
452 los_card_scan_time += last_los_scan_time;
456 mono_gc_get_card_table (int *shift_bits, gpointer *mask)
461 *shift_bits = CARD_BITS;
462 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
463 *mask = (gpointer)CARD_MASK;
468 return sgen_cardtable;
472 mono_gc_card_table_nursery_check (void)
474 return !major_collector.is_concurrent;
479 collect_faulted_cards (void)
481 #define CARD_PAGES (CARD_COUNT_IN_BYTES / 4096)
483 unsigned char faulted [CARD_PAGES] = { 0 };
484 mincore (sgen_cardtable, CARD_COUNT_IN_BYTES, faulted);
486 for (i = 0; i < CARD_PAGES; ++i) {
491 printf ("TOTAL card pages %d faulted %d\n", CARD_PAGES, count);
495 sgen_card_table_dump_obj_card (char *object, size_t size, void *dummy)
497 guint8 *start = sgen_card_table_get_card_scan_address (object);
498 guint8 *end = start + cards_in_range (object, size);
500 printf ("--obj %p %d cards [%p %p]--", object, size, start, end);
501 for (; start < end; ++start) {
503 printf ("\n\t[%p] ", start);
504 printf ("%x ", *start);
513 #define MWORD_MASK (sizeof (mword) - 1)
516 find_card_offset (mword card)
518 /*XXX Use assembly as this generates some pretty bad code */
519 #if defined(__i386__) && defined(__GNUC__)
520 return (__builtin_ffs (card) - 1) / 8;
521 #elif defined(__x86_64__) && defined(__GNUC__)
522 return (__builtin_ffsll (card) - 1) / 8;
523 #elif defined(__s390x__)
524 return (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
527 guint8 *ptr = (guint8 *) &card;
528 for (i = 0; i < sizeof (mword); ++i) {
537 find_next_card (guint8 *card_data, guint8 *end)
539 mword *cards, *cards_end;
542 while ((((mword)card_data) & MWORD_MASK) && card_data < end) {
548 if (card_data == end)
551 cards = (mword*)card_data;
552 cards_end = (mword*)((mword)end & ~MWORD_MASK);
553 while (cards < cards_end) {
556 return (guint8*)cards + find_card_offset (card);
560 card_data = (guint8*)cards_end;
561 while (card_data < end) {
571 sgen_cardtable_scan_object (char *obj, mword block_obj_size, guint8 *cards, gboolean mod_union, SgenGrayQueue *queue)
573 MonoVTable *vt = (MonoVTable*)SGEN_LOAD_VTABLE (obj);
574 MonoClass *klass = vt->klass;
576 HEAVY_STAT (++large_objects);
578 if (!SGEN_VTABLE_HAS_REFERENCES (vt)) {
579 sgen_object_layout_scanned_bitmap (0);
584 guint8 *card_data, *card_base;
585 guint8 *card_data_end;
586 char *obj_start = sgen_card_table_align_pointer (obj);
587 mword obj_size = sgen_par_object_get_size (vt, (MonoObject*)obj);
588 char *obj_end = obj + obj_size;
590 size_t extra_idx = 0;
592 MonoArray *arr = (MonoArray*)obj;
593 mword desc = (mword)klass->element_class->gc_descr;
594 int elem_size = mono_array_element_size (klass);
596 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
597 guint8 *overflow_scan_end = NULL;
600 #ifdef SGEN_OBJECT_LAYOUT_STATISTICS
601 if (klass->element_class->valuetype)
602 sgen_object_layout_scanned_vtype_array ();
604 sgen_object_layout_scanned_ref_array ();
610 card_data = sgen_card_table_get_card_scan_address ((mword)obj);
612 card_base = card_data;
613 card_count = cards_in_range ((mword)obj, obj_size);
614 card_data_end = card_data + card_count;
617 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
618 /*Check for overflow and if so, setup to scan in two steps*/
619 if (!cards && card_data_end >= SGEN_SHADOW_CARDTABLE_END) {
620 overflow_scan_end = sgen_shadow_cardtable + (card_data_end - SGEN_SHADOW_CARDTABLE_END);
621 card_data_end = SGEN_SHADOW_CARDTABLE_END;
627 card_data = find_next_card (card_data, card_data_end);
628 for (; card_data < card_data_end; card_data = find_next_card (card_data + 1, card_data_end)) {
630 size_t idx = (card_data - card_base) + extra_idx;
631 char *start = (char*)(obj_start + idx * CARD_SIZE_IN_BYTES);
632 char *card_end = start + CARD_SIZE_IN_BYTES;
633 char *first_elem, *elem;
635 HEAVY_STAT (++los_marked_cards);
638 sgen_card_table_prepare_card_for_scanning (card_data);
640 card_end = MIN (card_end, obj_end);
642 if (start <= (char*)arr->vector)
645 index = ARRAY_OBJ_INDEX (start, obj, elem_size);
647 elem = first_elem = (char*)mono_array_addr_with_size_fast ((MonoArray*)obj, elem_size, index);
648 if (klass->element_class->valuetype) {
649 ScanVTypeFunc scan_vtype_func = sgen_get_current_object_ops ()->scan_vtype;
651 for (; elem < card_end; elem += elem_size)
652 scan_vtype_func (elem, desc, queue BINARY_PROTOCOL_ARG (elem_size));
654 CopyOrMarkObjectFunc copy_func = sgen_get_current_object_ops ()->copy_or_mark_object;
656 HEAVY_STAT (++los_array_cards);
657 for (; elem < card_end; elem += SIZEOF_VOID_P) {
658 gpointer new, old = *(gpointer*)elem;
659 if ((mod_union && old) || G_UNLIKELY (sgen_ptr_in_nursery (old))) {
660 HEAVY_STAT (++los_array_remsets);
661 copy_func ((void**)elem, queue);
662 new = *(gpointer*)elem;
663 if (G_UNLIKELY (sgen_ptr_in_nursery (new)))
664 sgen_add_to_global_remset (elem, new);
669 binary_protocol_card_scan (first_elem, elem - first_elem);
672 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
673 if (overflow_scan_end) {
674 extra_idx = card_data - card_base;
675 card_base = card_data = sgen_shadow_cardtable;
676 card_data_end = overflow_scan_end;
677 overflow_scan_end = NULL;
683 HEAVY_STAT (++bloby_objects);
685 if (sgen_card_table_is_range_marked (cards, (mword)obj, block_obj_size))
686 sgen_get_current_object_ops ()->scan_object (obj, queue);
687 } else if (sgen_card_table_region_begin_scanning ((mword)obj, block_obj_size)) {
688 sgen_get_current_object_ops ()->scan_object (obj, queue);
691 binary_protocol_card_scan (obj, sgen_safe_object_get_size ((MonoObject*)obj));
695 #ifdef CARDTABLE_STATS
698 int total, marked, remarked, gc_marked;
701 static card_stats major_stats, los_stats;
702 static card_stats *cur_stats;
705 count_marked_cards (mword start, mword size)
707 mword end = start + size;
708 while (start <= end) {
709 guint8 card = *sgen_card_table_get_card_address (start);
714 ++cur_stats->gc_marked;
715 start += CARD_SIZE_IN_BYTES;
720 count_remarked_cards (mword start, mword size)
722 mword end = start + size;
723 while (start <= end) {
724 if (sgen_card_table_address_is_marked (start)) {
725 ++cur_stats->remarked;
726 *sgen_card_table_get_card_address (start) = 2;
728 start += CARD_SIZE_IN_BYTES;
735 sgen_card_tables_collect_stats (gboolean begin)
737 #ifdef CARDTABLE_STATS
739 memset (&major_stats, 0, sizeof (card_stats));
740 memset (&los_stats, 0, sizeof (card_stats));
741 cur_stats = &major_stats;
742 sgen_major_collector_iterate_live_block_ranges (count_marked_cards);
743 cur_stats = &los_stats;
744 sgen_los_iterate_live_block_ranges (count_marked_cards);
746 cur_stats = &major_stats;
747 sgen_major_collector_iterate_live_block_ranges (count_remarked_cards);
748 cur_stats = &los_stats;
749 sgen_los_iterate_live_block_ranges (count_remarked_cards);
750 printf ("cards major (t %d m %d g %d r %d) los (t %d m %d g %d r %d) major_scan %.2fms los_scan %.2fms\n",
751 major_stats.total, major_stats.marked, major_stats.gc_marked, major_stats.remarked,
752 los_stats.total, los_stats.marked, los_stats.gc_marked, los_stats.remarked,
753 last_major_scan_time / 10000.0f, last_los_scan_time / 10000.0f);
759 sgen_card_table_init (SgenRemeberedSet *remset)
761 sgen_cardtable = sgen_alloc_os_memory (CARD_COUNT_IN_BYTES, SGEN_ALLOC_INTERNAL | SGEN_ALLOC_ACTIVATE, "card table");
763 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
764 sgen_shadow_cardtable = sgen_alloc_os_memory (CARD_COUNT_IN_BYTES, SGEN_ALLOC_INTERNAL | SGEN_ALLOC_ACTIVATE, "shadow card table");
767 #ifdef HEAVY_STATISTICS
768 mono_counters_register ("marked cards", MONO_COUNTER_GC | MONO_COUNTER_LONG, &marked_cards);
769 mono_counters_register ("scanned cards", MONO_COUNTER_GC | MONO_COUNTER_LONG, &scanned_cards);
770 mono_counters_register ("remarked cards", MONO_COUNTER_GC | MONO_COUNTER_LONG, &remarked_cards);
772 mono_counters_register ("los marked cards", MONO_COUNTER_GC | MONO_COUNTER_LONG, &los_marked_cards);
773 mono_counters_register ("los array cards scanned ", MONO_COUNTER_GC | MONO_COUNTER_LONG, &los_array_cards);
774 mono_counters_register ("los array remsets", MONO_COUNTER_GC | MONO_COUNTER_LONG, &los_array_remsets);
775 mono_counters_register ("cardtable scanned objects", MONO_COUNTER_GC | MONO_COUNTER_LONG, &scanned_objects);
776 mono_counters_register ("cardtable large objects", MONO_COUNTER_GC | MONO_COUNTER_LONG, &large_objects);
777 mono_counters_register ("cardtable bloby objects", MONO_COUNTER_GC | MONO_COUNTER_LONG, &bloby_objects);
779 mono_counters_register ("cardtable major scan time", MONO_COUNTER_GC | MONO_COUNTER_LONG | MONO_COUNTER_TIME, &major_card_scan_time);
780 mono_counters_register ("cardtable los scan time", MONO_COUNTER_GC | MONO_COUNTER_LONG | MONO_COUNTER_TIME, &los_card_scan_time);
783 remset->wbarrier_set_field = sgen_card_table_wbarrier_set_field;
784 remset->wbarrier_set_arrayref = sgen_card_table_wbarrier_set_arrayref;
785 remset->wbarrier_arrayref_copy = sgen_card_table_wbarrier_arrayref_copy;
786 remset->wbarrier_value_copy = sgen_card_table_wbarrier_value_copy;
787 remset->wbarrier_object_copy = sgen_card_table_wbarrier_object_copy;
788 remset->wbarrier_generic_nostore = sgen_card_table_wbarrier_generic_nostore;
789 remset->record_pointer = sgen_card_table_record_pointer;
791 remset->finish_scan_remsets = sgen_card_table_finish_scan_remsets;
793 remset->finish_minor_collection = sgen_card_table_finish_minor_collection;
794 remset->prepare_for_major_collection = sgen_card_table_prepare_for_major_collection;
796 remset->find_address = sgen_card_table_find_address;
797 remset->find_address_with_cards = sgen_card_table_find_address_with_cards;
799 need_mod_union = sgen_get_major_collector ()->is_concurrent;
802 #endif /*HAVE_SGEN_GC*/