2 * sgen-marksweep.c: The Mark & Sweep major collector.
5 * Mark Probst <mark.probst@gmail.com>
7 * Copyright 2009-2010 Novell, Inc.
8 * Copyright (C) 2012 Xamarin Inc
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Library General Public
12 * License 2.0 as published by the Free Software Foundation;
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Library General Public License for more details.
19 * You should have received a copy of the GNU Library General Public
20 * License 2.0 along with this library; if not, write to the Free
21 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
31 #include "utils/mono-counters.h"
32 #include "utils/mono-semaphore.h"
33 #include "utils/mono-time.h"
34 #include "metadata/object-internals.h"
35 #include "metadata/profiler-private.h"
37 #include "metadata/sgen-gc.h"
38 #include "metadata/sgen-protocol.h"
39 #include "metadata/sgen-cardtable.h"
40 #include "metadata/sgen-memory-governor.h"
41 #include "metadata/sgen-layout-stats.h"
42 #include "metadata/gc-internal.h"
43 #include "metadata/sgen-pointer-queue.h"
44 #include "metadata/sgen-pinning.h"
45 #include "metadata/sgen-workers.h"
47 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
48 #define MS_BLOCK_SIZE ARCH_MIN_MS_BLOCK_SIZE
49 #define MS_BLOCK_SIZE_SHIFT ARCH_MIN_MS_BLOCK_SIZE_SHIFT
51 #define MS_BLOCK_SIZE_SHIFT 14 /* INT FASTENABLE */
52 #define MS_BLOCK_SIZE (1 << MS_BLOCK_SIZE_SHIFT)
54 #define MAJOR_SECTION_SIZE MS_BLOCK_SIZE
55 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
58 * Don't allocate single blocks, but alloc a contingent of this many
59 * blocks in one swoop. This must be a power of two.
61 #define MS_BLOCK_ALLOC_NUM 32
64 * Number of bytes before the first object in a block. At the start
65 * of a block is the MSBlockHeader, then opional padding, then come
66 * the objects, so this must be >= sizeof (MSBlockHeader).
68 #define MS_BLOCK_SKIP ((sizeof (MSBlockHeader) + 15) & ~15)
70 #define MS_BLOCK_FREE (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
72 #define MS_NUM_MARK_WORDS ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
74 typedef struct _MSBlockInfo MSBlockInfo;
78 unsigned int pinned : 1;
79 unsigned int has_references : 1;
80 unsigned int has_pinned : 1; /* means cannot evacuate */
81 unsigned int is_to_space : 1;
82 unsigned int swept : 1;
84 MSBlockInfo *next_free;
85 size_t pin_queue_first_entry;
86 size_t pin_queue_last_entry;
87 guint8 *cardtable_mod_union;
88 mword mark_words [MS_NUM_MARK_WORDS];
91 #define MS_BLOCK_FOR_BLOCK_INFO(b) ((char*)(b))
93 #define MS_BLOCK_OBJ(b,i) (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i))
94 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size) (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
95 #define MS_BLOCK_DATA_FOR_OBJ(o) ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
101 #define MS_BLOCK_FOR_OBJ(o) (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
103 /* object index will always be small */
104 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
106 //casting to int is fine since blocks are 32k
107 #define MS_CALC_MARK_BIT(w,b,o) do { \
108 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
109 if (sizeof (mword) == 4) { \
118 #define MS_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] & (ONE_P << (b)))
119 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
121 #define MS_OBJ_ALLOCED(o,b) (*(void**)(o) && (*(char**)(o) < MS_BLOCK_FOR_BLOCK_INFO (b) || *(char**)(o) >= MS_BLOCK_FOR_BLOCK_INFO (b) + MS_BLOCK_SIZE))
123 #define MS_BLOCK_OBJ_SIZE_FACTOR (pow (2.0, 1.0 / 3))
126 * This way we can lookup block object size indexes for sizes up to
127 * 256 bytes with a single load.
129 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES 32
131 static int *block_obj_sizes;
132 static int num_block_obj_sizes;
133 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
135 #define MS_BLOCK_FLAG_PINNED 1
136 #define MS_BLOCK_FLAG_REFS 2
138 #define MS_BLOCK_TYPE_MAX 4
140 static gboolean *evacuate_block_obj_sizes;
141 static float evacuation_threshold = 0.666f;
142 static float concurrent_evacuation_threshold = 0.666f;
143 static gboolean want_evacuation = FALSE;
145 static gboolean lazy_sweep = TRUE;
146 static gboolean have_swept = TRUE;
148 static gboolean concurrent_mark;
150 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl) SGEN_POINTER_IS_TAGGED_1 ((bl))
151 #define BLOCK_TAG_HAS_REFERENCES(bl) SGEN_POINTER_TAG_1 ((bl))
152 #define BLOCK_UNTAG_HAS_REFERENCES(bl) SGEN_POINTER_UNTAG_1 ((bl))
154 #define BLOCK_TAG(bl) ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
156 /* all allocated blocks in the system */
157 static SgenPointerQueue allocated_blocks;
159 /* non-allocated block free-list */
160 static void *empty_blocks = NULL;
161 static size_t num_empty_blocks = 0;
163 #define FOREACH_BLOCK(bl) { size_t __index; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [__index]);
164 #define FOREACH_BLOCK_HAS_REFERENCES(bl,hr) { size_t __index; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = allocated_blocks.data [__index]; (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl)); (bl) = BLOCK_UNTAG_HAS_REFERENCES ((bl));
165 #define END_FOREACH_BLOCK } }
166 #define DELETE_BLOCK_IN_FOREACH() (allocated_blocks.data [__index] = NULL)
168 static size_t num_major_sections = 0;
169 /* one free block list for each block object size */
170 static MSBlockInfo **free_block_lists [MS_BLOCK_TYPE_MAX];
172 static guint64 stat_major_blocks_alloced = 0;
173 static guint64 stat_major_blocks_freed = 0;
174 static guint64 stat_major_blocks_lazy_swept = 0;
175 static guint64 stat_major_objects_evacuated = 0;
177 #if SIZEOF_VOID_P != 8
178 static guint64 stat_major_blocks_freed_ideal = 0;
179 static guint64 stat_major_blocks_freed_less_ideal = 0;
180 static guint64 stat_major_blocks_freed_individual = 0;
181 static guint64 stat_major_blocks_alloced_less_ideal = 0;
184 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
185 static guint64 num_major_objects_marked = 0;
186 #define INC_NUM_MAJOR_OBJECTS_MARKED() (++num_major_objects_marked)
188 #define INC_NUM_MAJOR_OBJECTS_MARKED()
191 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
192 static mono_mutex_t scanned_objects_list_lock;
193 static SgenPointerQueue scanned_objects_list;
196 add_scanned_object (void *ptr)
198 if (!binary_protocol_is_enabled ())
201 mono_mutex_lock (&scanned_objects_list_lock);
202 sgen_pointer_queue_add (&scanned_objects_list, ptr);
203 mono_mutex_unlock (&scanned_objects_list_lock);
208 sweep_block (MSBlockInfo *block, gboolean during_major_collection);
211 ms_find_block_obj_size_index (size_t size)
214 SGEN_ASSERT (9, size <= SGEN_MAX_SMALL_OBJ_SIZE, "size %d is bigger than max small object size %d", size, SGEN_MAX_SMALL_OBJ_SIZE);
215 for (i = 0; i < num_block_obj_sizes; ++i)
216 if (block_obj_sizes [i] >= size)
218 g_error ("no object of size %d\n", size);
221 #define FREE_BLOCKS_FROM(lists,p,r) (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
222 #define FREE_BLOCKS(p,r) (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
224 #define MS_BLOCK_OBJ_SIZE_INDEX(s) \
225 (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ? \
226 fast_block_obj_size_indexes [((s)+7)>>3] : \
227 ms_find_block_obj_size_index ((s)))
230 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
234 start = sgen_alloc_os_memory_aligned (nursery_size, nursery_align, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
236 start = sgen_alloc_os_memory (nursery_size, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
242 update_heap_boundaries_for_block (MSBlockInfo *block)
244 sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
248 ms_get_empty_block (void)
252 void *block, *empty, *next;
257 * We try allocating MS_BLOCK_ALLOC_NUM blocks first. If that's
258 * unsuccessful, we halve the number of blocks and try again, until we're at
259 * 1. If that doesn't work, either, we assert.
261 int alloc_num = MS_BLOCK_ALLOC_NUM;
263 p = sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE,
264 alloc_num == 1 ? "major heap section" : NULL);
270 for (i = 0; i < alloc_num; ++i) {
273 * We do the free list update one after the
274 * other so that other threads can use the new
275 * blocks as quickly as possible.
278 empty = empty_blocks;
279 *(void**)block = empty;
280 } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
284 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
286 stat_major_blocks_alloced += alloc_num;
287 #if SIZEOF_VOID_P != 8
288 if (alloc_num != MS_BLOCK_ALLOC_NUM)
289 stat_major_blocks_alloced_less_ideal += alloc_num;
294 empty = empty_blocks;
298 next = *(void**)block;
299 } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
301 SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
303 *(void**)block = NULL;
305 g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
311 * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
312 * list, where it will either be freed later on, or reused in nursery collections.
315 ms_free_block (void *block)
319 sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
320 memset (block, 0, MS_BLOCK_SIZE);
323 empty = empty_blocks;
324 *(void**)block = empty;
325 } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
327 SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
330 //#define MARKSWEEP_CONSISTENCY_CHECK
332 #ifdef MARKSWEEP_CONSISTENCY_CHECK
334 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
338 for (; block; block = block->next_free) {
339 g_assert (block->obj_size == size);
340 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
342 /* blocks in the free lists must have at least
345 g_assert (block->free_list);
347 /* the block must be in the allocated_blocks array */
348 g_assert (sgen_pointer_queue_find (&allocated_blocks, BLOCK_TAG (block)) != (size_t)-1);
353 check_empty_blocks (void)
357 for (p = empty_blocks; p; p = *(void**)p)
359 g_assert (i == num_empty_blocks);
363 consistency_check (void)
368 /* check all blocks */
369 FOREACH_BLOCK (block) {
370 int count = MS_BLOCK_FREE / block->obj_size;
374 /* check block header */
375 g_assert (((MSBlockHeader*)block->block)->info == block);
377 /* count number of free slots */
378 for (i = 0; i < count; ++i) {
379 void **obj = (void**) MS_BLOCK_OBJ (block, i);
380 if (!MS_OBJ_ALLOCED (obj, block))
384 /* check free list */
385 for (free = block->free_list; free; free = (void**)*free) {
386 g_assert (MS_BLOCK_FOR_OBJ (free) == block);
389 g_assert (num_free == 0);
391 /* check all mark words are zero */
393 for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
394 g_assert (block->mark_words [i] == 0);
398 /* check free blocks */
399 for (i = 0; i < num_block_obj_sizes; ++i) {
401 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
402 check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
405 check_empty_blocks ();
410 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
412 int size = block_obj_sizes [size_index];
413 int count = MS_BLOCK_FREE / size;
415 MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, has_references);
419 if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
422 info = (MSBlockInfo*)ms_get_empty_block ();
424 SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
426 info->obj_size = size;
427 info->obj_size_index = size_index;
428 info->pinned = pinned;
429 info->has_references = has_references;
430 info->has_pinned = pinned;
432 * Blocks that are to-space are not evacuated from. During an major collection
433 * blocks are allocated for two reasons: evacuating objects from the nursery and
434 * evacuating them from major blocks marked for evacuation. In both cases we don't
435 * want further evacuation.
437 info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD);
439 info->cardtable_mod_union = NULL;
441 update_heap_boundaries_for_block (info);
443 /* build free list */
444 obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
445 info->free_list = (void**)obj_start;
446 /* we're skipping the last one - it must be nulled */
447 for (i = 0; i < count - 1; ++i) {
448 char *next_obj_start = obj_start + size;
449 *(void**)obj_start = next_obj_start;
450 obj_start = next_obj_start;
453 *(void**)obj_start = NULL;
455 info->next_free = free_blocks [size_index];
456 free_blocks [size_index] = info;
458 sgen_pointer_queue_add (&allocated_blocks, BLOCK_TAG (info));
460 ++num_major_sections;
465 obj_is_from_pinned_alloc (char *ptr)
469 FOREACH_BLOCK (block) {
470 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE)
471 return block->pinned;
477 unlink_slot_from_free_list_uncontested (MSBlockInfo **free_blocks, int size_index)
482 block = free_blocks [size_index];
483 SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
485 if (G_UNLIKELY (!block->swept)) {
486 stat_major_blocks_lazy_swept ++;
487 sweep_block (block, FALSE);
490 obj = block->free_list;
491 SGEN_ASSERT (9, obj, "block %p in free list had no available object to alloc from", block);
493 block->free_list = *(void**)obj;
494 if (!block->free_list) {
495 free_blocks [size_index] = block->next_free;
496 block->next_free = NULL;
503 alloc_obj (MonoVTable *vtable, size_t size, gboolean pinned, gboolean has_references)
505 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
506 MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, has_references);
509 if (!free_blocks [size_index]) {
510 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
514 obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
516 *(MonoVTable**)obj = vtable;
522 major_alloc_object (MonoVTable *vtable, size_t size, gboolean has_references)
524 return alloc_obj (vtable, size, FALSE, has_references);
528 * We're not freeing the block if it's empty. We leave that work for
529 * the next major collection.
531 * This is just called from the domain clearing code, which runs in a
532 * single thread and has the GC lock, so we don't need an extra lock.
535 free_object (char *obj, size_t size, gboolean pinned)
537 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
541 sweep_block (block, FALSE);
542 SGEN_ASSERT (9, (pinned && block->pinned) || (!pinned && !block->pinned), "free-object pinning mixup object %p pinned %d block %p pinned %d", obj, pinned, block, block->pinned);
543 SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
544 MS_CALC_MARK_BIT (word, bit, obj);
545 SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set");
546 if (!block->free_list) {
547 MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, block->has_references);
548 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
549 SGEN_ASSERT (9, !block->next_free, "block %p doesn't have a free-list of object but belongs to a free-list of blocks");
550 block->next_free = free_blocks [size_index];
551 free_blocks [size_index] = block;
553 memset (obj, 0, size);
554 *(void**)obj = block->free_list;
555 block->free_list = (void**)obj;
559 major_free_non_pinned_object (char *obj, size_t size)
561 free_object (obj, size, FALSE);
564 /* size is a multiple of SGEN_ALLOC_ALIGN */
566 major_alloc_small_pinned_obj (MonoVTable *vtable, size_t size, gboolean has_references)
570 res = alloc_obj (vtable, size, TRUE, has_references);
571 /*If we failed to alloc memory, we better try releasing memory
572 *as pinned alloc is requested by the runtime.
575 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE);
576 res = alloc_obj (vtable, size, TRUE, has_references);
582 free_pinned_object (char *obj, size_t size)
584 free_object (obj, size, TRUE);
588 * size is already rounded up and we hold the GC lock.
591 major_alloc_degraded (MonoVTable *vtable, size_t size)
594 size_t old_num_sections;
596 old_num_sections = num_major_sections;
598 obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
599 if (G_LIKELY (obj)) {
600 HEAVY_STAT (++stat_objects_alloced_degraded);
601 HEAVY_STAT (stat_bytes_alloced_degraded += size);
602 g_assert (num_major_sections >= old_num_sections);
603 sgen_register_major_sections_alloced (num_major_sections - old_num_sections);
609 * obj is some object. If it's not in the major heap (i.e. if it's in
610 * the nursery or LOS), return FALSE. Otherwise return whether it's
611 * been marked or copied.
614 major_is_object_live (char *obj)
620 if (sgen_ptr_in_nursery (obj))
623 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
626 if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
629 /* now we know it's in a major block */
630 block = MS_BLOCK_FOR_OBJ (obj);
631 SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?");
632 MS_CALC_MARK_BIT (word, bit, obj);
633 return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
637 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
641 FOREACH_BLOCK (block) {
642 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
643 int count = MS_BLOCK_FREE / block->obj_size;
647 for (i = 0; i <= count; ++i) {
648 if (ptr >= MS_BLOCK_OBJ (block, i) && ptr < MS_BLOCK_OBJ (block, i + 1)) {
649 *start = MS_BLOCK_OBJ (block, i);
653 return !block->pinned;
660 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
662 gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
663 gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
664 gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
667 FOREACH_BLOCK (block) {
668 int count = MS_BLOCK_FREE / block->obj_size;
671 if (block->pinned && !pinned)
673 if (!block->pinned && !non_pinned)
675 if (sweep && lazy_sweep) {
676 sweep_block (block, FALSE);
677 SGEN_ASSERT (0, block->swept, "Block must be swept after sweeping");
680 for (i = 0; i < count; ++i) {
681 void **obj = (void**) MS_BLOCK_OBJ (block, i);
684 MS_CALC_MARK_BIT (word, bit, obj);
685 if (!MS_MARK_BIT (block, word, bit))
688 if (MS_OBJ_ALLOCED (obj, block))
689 callback ((char*)obj, block->obj_size, data);
695 major_is_valid_object (char *object)
699 FOREACH_BLOCK (block) {
703 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
706 idx = MS_BLOCK_OBJ_INDEX (object, block);
707 obj = (char*)MS_BLOCK_OBJ (block, idx);
710 return MS_OBJ_ALLOCED (obj, block);
718 major_describe_pointer (char *ptr)
722 FOREACH_BLOCK (block) {
730 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
733 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
734 MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
736 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
737 obj = (char*)MS_BLOCK_OBJ (block, idx);
738 live = MS_OBJ_ALLOCED (obj, block);
739 vtable = live ? (MonoVTable*)SGEN_LOAD_VTABLE (obj) : NULL;
741 MS_CALC_MARK_BIT (w, b, obj);
742 marked = MS_MARK_BIT (block, w, b);
747 SGEN_LOG (0, "object");
749 SGEN_LOG (0, "dead-object");
752 SGEN_LOG (0, "interior-ptr offset %td", ptr - obj);
754 SGEN_LOG (0, "dead-interior-ptr offset %td", ptr - obj);
757 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
766 major_check_scan_starts (void)
771 major_dump_heap (FILE *heap_dump_file)
774 int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
775 int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
778 for (i = 0; i < num_block_obj_sizes; ++i)
779 slots_available [i] = slots_used [i] = 0;
781 FOREACH_BLOCK (block) {
782 int index = ms_find_block_obj_size_index (block->obj_size);
783 int count = MS_BLOCK_FREE / block->obj_size;
785 slots_available [index] += count;
786 for (i = 0; i < count; ++i) {
787 if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
788 ++slots_used [index];
792 fprintf (heap_dump_file, "<occupancies>\n");
793 for (i = 0; i < num_block_obj_sizes; ++i) {
794 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
795 block_obj_sizes [i], slots_available [i], slots_used [i]);
797 fprintf (heap_dump_file, "</occupancies>\n");
799 FOREACH_BLOCK (block) {
800 int count = MS_BLOCK_FREE / block->obj_size;
804 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
806 for (i = 0; i <= count; ++i) {
807 if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
812 sgen_dump_occupied (MS_BLOCK_OBJ (block, start), MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
818 fprintf (heap_dump_file, "</section>\n");
822 #define LOAD_VTABLE SGEN_LOAD_VTABLE
824 #define MS_MARK_OBJECT_AND_ENQUEUE_CHECKED(obj,desc,block,queue) do { \
826 MS_CALC_MARK_BIT (__word, __bit, (obj)); \
827 if (!MS_MARK_BIT ((block), __word, __bit) && MS_OBJ_ALLOCED ((obj), (block))) { \
828 MS_SET_MARK_BIT ((block), __word, __bit); \
829 if (sgen_gc_descr_has_references (desc)) \
830 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
831 binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((MonoObject*)(obj))); \
832 INC_NUM_MAJOR_OBJECTS_MARKED (); \
835 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do { \
837 MS_CALC_MARK_BIT (__word, __bit, (obj)); \
838 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
839 if (!MS_MARK_BIT ((block), __word, __bit)) { \
840 MS_SET_MARK_BIT ((block), __word, __bit); \
841 if (sgen_gc_descr_has_references (desc)) \
842 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
843 binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((MonoObject*)(obj))); \
844 INC_NUM_MAJOR_OBJECTS_MARKED (); \
849 pin_major_object (char *obj, SgenGrayQueue *queue)
854 g_assert_not_reached ();
856 block = MS_BLOCK_FOR_OBJ (obj);
857 block->has_pinned = TRUE;
858 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
861 #include "sgen-major-copy-object.h"
864 major_copy_or_mark_object_with_evacuation_concurrent (void **ptr, void *obj, SgenGrayQueue *queue)
866 SGEN_ASSERT (9, sgen_concurrent_collection_in_progress (), "Why are we scanning concurrently when there's no concurrent collection on?");
867 SGEN_ASSERT (9, !sgen_workers_are_working () || sgen_is_worker_thread (mono_native_thread_id_get ()), "We must not scan from two threads at the same time!");
869 g_assert (!SGEN_OBJECT_IS_FORWARDED (obj));
871 if (!sgen_ptr_in_nursery (obj)) {
874 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
876 if (objsize <= SGEN_MAX_SMALL_OBJ_SIZE) {
877 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
878 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
880 if (sgen_los_object_is_pinned (obj))
884 if (G_UNLIKELY (MONO_GC_OBJ_PINNED_ENABLED ())) {
885 MonoVTable *vt = (MonoVTable*)SGEN_LOAD_VTABLE (obj);
886 MONO_GC_OBJ_PINNED ((mword)obj, sgen_safe_object_get_size (obj), vt->klass->name_space, vt->klass->name, GENERATION_OLD);
890 sgen_los_pin_object (obj);
891 if (SGEN_OBJECT_HAS_REFERENCES (obj))
892 GRAY_OBJECT_ENQUEUE (queue, obj, sgen_obj_get_descriptor (obj));
893 INC_NUM_MAJOR_OBJECTS_MARKED ();
899 major_get_and_reset_num_major_objects_marked (void)
901 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
902 long long num = num_major_objects_marked;
903 num_major_objects_marked = 0;
910 #define PREFETCH_CARDS 1 /* BOOL FASTENABLE */
912 #undef PREFETCH_CARDS
915 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
916 #if defined(PLATFORM_MACOSX)
917 #define GCC_VERSION (__GNUC__ * 10000 \
918 + __GNUC_MINOR__ * 100 \
919 + __GNUC_PATCHLEVEL__)
920 #if GCC_VERSION <= 40300
921 #undef PREFETCH_CARDS
925 #ifdef HEAVY_STATISTICS
926 static guint64 stat_optimized_copy;
927 static guint64 stat_optimized_copy_nursery;
928 static guint64 stat_optimized_copy_nursery_forwarded;
929 static guint64 stat_optimized_copy_nursery_pinned;
930 static guint64 stat_optimized_copy_major;
931 static guint64 stat_optimized_copy_major_small_fast;
932 static guint64 stat_optimized_copy_major_small_slow;
933 static guint64 stat_optimized_copy_major_large;
934 static guint64 stat_optimized_copy_major_forwarded;
935 static guint64 stat_optimized_copy_major_small_evacuate;
936 static guint64 stat_optimized_major_scan;
937 static guint64 stat_optimized_major_scan_no_refs;
939 static guint64 stat_drain_prefetch_fills;
940 static guint64 stat_drain_prefetch_fill_failures;
941 static guint64 stat_drain_loops;
944 static void major_scan_object_with_evacuation (char *start, mword desc, SgenGrayQueue *queue);
946 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_no_evacuation
947 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_no_evacuation
948 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_no_evacuation
949 #include "sgen-marksweep-drain-gray-stack.h"
951 #define COPY_OR_MARK_WITH_EVACUATION
952 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_with_evacuation
953 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_with_evacuation
954 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_with_evacuation
955 #include "sgen-marksweep-drain-gray-stack.h"
958 drain_gray_stack (ScanCopyContext ctx)
960 gboolean evacuation = FALSE;
962 for (i = 0; i < num_block_obj_sizes; ++i) {
963 if (evacuate_block_obj_sizes [i]) {
970 return drain_gray_stack_with_evacuation (ctx);
972 return drain_gray_stack_no_evacuation (ctx);
975 #include "sgen-marksweep-scan-object-concurrent.h"
978 major_copy_or_mark_object_canonical (void **ptr, SgenGrayQueue *queue)
980 major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
984 major_copy_or_mark_object_concurrent_canonical (void **ptr, SgenGrayQueue *queue)
986 major_copy_or_mark_object_with_evacuation_concurrent (ptr, *ptr, queue);
990 mark_pinned_objects_in_block (MSBlockInfo *block, SgenGrayQueue *queue)
995 if (block->pin_queue_first_entry == block->pin_queue_last_entry)
998 block->has_pinned = TRUE;
1000 entry = sgen_pinning_get_entry (block->pin_queue_first_entry);
1001 end = sgen_pinning_get_entry (block->pin_queue_last_entry);
1003 for (; entry < end; ++entry) {
1004 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1006 SGEN_ASSERT (9, index >= 0 && index < MS_BLOCK_FREE / block->obj_size, "invalid object %p index %d max-index %d", *entry, index, MS_BLOCK_FREE / block->obj_size);
1007 if (index == last_index)
1009 obj = MS_BLOCK_OBJ (block, index);
1010 MS_MARK_OBJECT_AND_ENQUEUE_CHECKED (obj, sgen_obj_get_descriptor (obj), block, queue);
1016 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1020 for (obj_index = 0; obj_index < count; ++obj_index) {
1022 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1024 MS_CALC_MARK_BIT (word, bit, obj);
1025 if (MS_MARK_BIT (block, word, bit)) {
1026 SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1028 /* an unmarked object */
1029 if (MS_OBJ_ALLOCED (obj, block)) {
1031 * FIXME: Merge consecutive
1032 * slots for lower reporting
1033 * overhead. Maybe memset
1034 * will also benefit?
1036 binary_protocol_empty (obj, obj_size);
1037 MONO_GC_MAJOR_SWEPT ((mword)obj, obj_size);
1038 memset (obj, 0, obj_size);
1040 *(void**)obj = block->free_list;
1041 block->free_list = obj;
1049 * Traverse BLOCK, freeing and zeroing unused objects.
1052 sweep_block (MSBlockInfo *block, gboolean during_major_collection)
1055 void *reversed = NULL;
1057 if (!during_major_collection)
1058 g_assert (!sgen_concurrent_collection_in_progress ());
1063 count = MS_BLOCK_FREE / block->obj_size;
1065 block->free_list = NULL;
1067 /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1068 // FIXME: Add more sizes
1069 switch (block->obj_size) {
1071 sweep_block_for_size (block, count, 16);
1074 sweep_block_for_size (block, count, block->obj_size);
1078 /* reset mark bits */
1079 memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1081 /* Reverse free list so that it's in address order */
1083 while (block->free_list) {
1084 void *next = *(void**)block->free_list;
1085 *(void**)block->free_list = reversed;
1086 reversed = block->free_list;
1087 block->free_list = next;
1089 block->free_list = reversed;
1100 if (sizeof (mword) == sizeof (unsigned long))
1101 count += __builtin_popcountl (d);
1103 count += __builtin_popcount (d);
1119 /* statistics for evacuation */
1120 int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
1121 int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
1122 int *num_blocks = alloca (sizeof (int) * num_block_obj_sizes);
1124 mword total_evacuate_heap = 0;
1125 mword total_evacuate_saved = 0;
1127 for (i = 0; i < num_block_obj_sizes; ++i)
1128 slots_available [i] = slots_used [i] = num_blocks [i] = 0;
1130 /* clear all the free lists */
1131 for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1132 MSBlockInfo **free_blocks = free_block_lists [i];
1134 for (j = 0; j < num_block_obj_sizes; ++j)
1135 free_blocks [j] = NULL;
1138 /* traverse all blocks, free and zero unmarked objects */
1139 FOREACH_BLOCK (block) {
1141 gboolean have_live = FALSE;
1142 gboolean has_pinned;
1143 gboolean have_free = FALSE;
1147 obj_size_index = block->obj_size_index;
1149 has_pinned = block->has_pinned;
1150 block->has_pinned = block->pinned;
1152 block->is_to_space = FALSE;
1155 count = MS_BLOCK_FREE / block->obj_size;
1157 if (block->cardtable_mod_union) {
1158 sgen_free_internal_dynamic (block->cardtable_mod_union, CARDS_PER_BLOCK, INTERNAL_MEM_CARDTABLE_MOD_UNION);
1159 block->cardtable_mod_union = NULL;
1162 /* Count marked objects in the block */
1163 for (i = 0; i < MS_NUM_MARK_WORDS; ++i) {
1164 nused += bitcount (block->mark_words [i]);
1173 sweep_block (block, TRUE);
1177 ++num_blocks [obj_size_index];
1178 slots_used [obj_size_index] += nused;
1179 slots_available [obj_size_index] += count;
1183 * If there are free slots in the block, add
1184 * the block to the corresponding free list.
1187 MSBlockInfo **free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1188 int index = MS_BLOCK_OBJ_SIZE_INDEX (block->obj_size);
1189 block->next_free = free_blocks [index];
1190 free_blocks [index] = block;
1193 update_heap_boundaries_for_block (block);
1196 * Blocks without live objects are removed from the
1197 * block list and freed.
1199 DELETE_BLOCK_IN_FOREACH ();
1201 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1202 ms_free_block (block);
1204 --num_major_sections;
1206 } END_FOREACH_BLOCK;
1207 sgen_pointer_queue_remove_nulls (&allocated_blocks);
1209 for (i = 0; i < num_block_obj_sizes; ++i) {
1210 float usage = (float)slots_used [i] / (float)slots_available [i];
1211 if (num_blocks [i] > 5 && usage < evacuation_threshold) {
1212 evacuate_block_obj_sizes [i] = TRUE;
1214 g_print ("slot size %d - %d of %d used\n",
1215 block_obj_sizes [i], slots_used [i], slots_available [i]);
1218 evacuate_block_obj_sizes [i] = FALSE;
1221 mword total_bytes = block_obj_sizes [i] * slots_available [i];
1222 total_evacuate_heap += total_bytes;
1223 if (evacuate_block_obj_sizes [i])
1224 total_evacuate_saved += total_bytes - block_obj_sizes [i] * slots_used [i];
1228 want_evacuation = (float)total_evacuate_saved / (float)total_evacuate_heap > (1 - concurrent_evacuation_threshold);
1234 major_have_finished_sweeping (void)
1239 static int count_pinned_ref;
1240 static int count_pinned_nonref;
1241 static int count_nonpinned_ref;
1242 static int count_nonpinned_nonref;
1245 count_nonpinned_callback (char *obj, size_t size, void *data)
1247 MonoVTable *vtable = (MonoVTable*)LOAD_VTABLE (obj);
1249 if (vtable->klass->has_references)
1250 ++count_nonpinned_ref;
1252 ++count_nonpinned_nonref;
1256 count_pinned_callback (char *obj, size_t size, void *data)
1258 MonoVTable *vtable = (MonoVTable*)LOAD_VTABLE (obj);
1260 if (vtable->klass->has_references)
1263 ++count_pinned_nonref;
1266 static G_GNUC_UNUSED void
1267 count_ref_nonref_objs (void)
1271 count_pinned_ref = 0;
1272 count_pinned_nonref = 0;
1273 count_nonpinned_ref = 0;
1274 count_nonpinned_nonref = 0;
1276 major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1277 major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1279 total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1281 g_print ("ref: %d pinned %d non-pinned non-ref: %d pinned %d non-pinned -- %.1f\n",
1282 count_pinned_ref, count_nonpinned_ref,
1283 count_pinned_nonref, count_nonpinned_nonref,
1284 (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1288 ms_calculate_block_obj_sizes (double factor, int *arr)
1290 double target_size = sizeof (MonoObject);
1295 int target_count = (int)ceil (MS_BLOCK_FREE / target_size);
1296 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1298 if (size != last_size) {
1300 arr [num_sizes] = size;
1305 target_size *= factor;
1306 } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1311 /* only valid during minor collections */
1312 static mword old_num_major_sections;
1315 major_start_nursery_collection (void)
1317 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1318 consistency_check ();
1321 old_num_major_sections = num_major_sections;
1325 major_finish_nursery_collection (void)
1327 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1328 consistency_check ();
1330 sgen_register_major_sections_alloced (num_major_sections - old_num_major_sections);
1334 major_start_major_collection (void)
1338 /* clear the free lists */
1339 for (i = 0; i < num_block_obj_sizes; ++i) {
1340 if (!evacuate_block_obj_sizes [i])
1343 free_block_lists [0][i] = NULL;
1344 free_block_lists [MS_BLOCK_FLAG_REFS][i] = NULL;
1347 // Sweep all unswept blocks
1351 MONO_GC_SWEEP_BEGIN (GENERATION_OLD, TRUE);
1353 FOREACH_BLOCK (block) {
1354 sweep_block (block, TRUE);
1355 } END_FOREACH_BLOCK;
1357 MONO_GC_SWEEP_END (GENERATION_OLD, TRUE);
1360 SGEN_ASSERT (0, have_swept, "Cannot start major collection without having finished sweeping");
1365 major_finish_major_collection (ScannedObjectCounts *counts)
1367 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1368 if (binary_protocol_is_enabled ()) {
1369 counts->num_scanned_objects = scanned_objects_list.next_slot;
1371 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1372 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1374 sgen_pointer_queue_clear (&scanned_objects_list);
1379 #if SIZEOF_VOID_P != 8
1381 compare_pointers (const void *va, const void *vb) {
1382 char *a = *(char**)va, *b = *(char**)vb;
1392 major_free_swept_blocks (void)
1394 size_t section_reserve = sgen_get_minor_collection_allowance () / MS_BLOCK_SIZE;
1396 g_assert (have_swept);
1398 #if SIZEOF_VOID_P != 8
1400 int i, num_empty_blocks_orig, num_blocks, arr_length;
1402 void **empty_block_arr;
1403 void **rebuild_next;
1407 * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1408 * a VirtualAlloc ()-ed block.
1413 if (num_empty_blocks <= section_reserve)
1415 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1417 num_empty_blocks_orig = num_empty_blocks;
1418 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1419 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1420 if (!empty_block_arr)
1424 for (block = empty_blocks; block; block = *(void**)block)
1425 empty_block_arr [i++] = block;
1426 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1428 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1431 * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1432 * contiguous ones. If we do, we free them. If that's not enough to get to
1433 * section_reserve, we halve the number of contiguous blocks we're looking
1434 * for and have another go, until we're done with looking for pairs of
1435 * blocks, at which point we give up and go to the fallback.
1437 arr_length = num_empty_blocks_orig;
1438 num_blocks = MS_BLOCK_ALLOC_NUM;
1439 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1444 for (i = 0; i < arr_length; ++i) {
1446 void *block = empty_block_arr [i];
1447 SGEN_ASSERT (0, block, "we're not shifting correctly");
1449 empty_block_arr [dest] = block;
1451 * This is not strictly necessary, but we're
1454 empty_block_arr [i] = NULL;
1463 SGEN_ASSERT (0, first >= 0 && d > first, "algorithm is wrong");
1465 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
1470 if (d + 1 - first == num_blocks) {
1472 * We found num_blocks contiguous blocks. Free them
1473 * and null their array entries. As an optimization
1474 * we could, instead of nulling the entries, shift
1475 * the following entries over to the left, while
1479 sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
1480 for (j = first; j <= d; ++j)
1481 empty_block_arr [j] = NULL;
1485 num_empty_blocks -= num_blocks;
1487 stat_major_blocks_freed += num_blocks;
1488 if (num_blocks == MS_BLOCK_ALLOC_NUM)
1489 stat_major_blocks_freed_ideal += num_blocks;
1491 stat_major_blocks_freed_less_ideal += num_blocks;
1496 SGEN_ASSERT (0, dest <= i && dest <= arr_length, "array length is off");
1498 SGEN_ASSERT (0, arr_length == num_empty_blocks, "array length is off");
1503 /* rebuild empty_blocks free list */
1504 rebuild_next = (void**)&empty_blocks;
1505 for (i = 0; i < arr_length; ++i) {
1506 void *block = empty_block_arr [i];
1507 SGEN_ASSERT (0, block, "we're missing blocks");
1508 *rebuild_next = block;
1509 rebuild_next = (void**)block;
1511 *rebuild_next = NULL;
1514 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
1517 SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
1521 * This is our threshold. If there's not more empty than used blocks, we won't
1522 * release uncontiguous blocks, in fear of fragmenting the address space.
1524 if (num_empty_blocks <= num_major_sections)
1528 while (num_empty_blocks > section_reserve) {
1529 void *next = *(void**)empty_blocks;
1530 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
1531 empty_blocks = next;
1533 * Needs not be atomic because this is running
1538 ++stat_major_blocks_freed;
1539 #if SIZEOF_VOID_P != 8
1540 ++stat_major_blocks_freed_individual;
1546 major_find_pin_queue_start_ends (SgenGrayQueue *queue)
1550 FOREACH_BLOCK (block) {
1551 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
1552 &block->pin_queue_first_entry, &block->pin_queue_last_entry);
1553 } END_FOREACH_BLOCK;
1557 major_pin_objects (SgenGrayQueue *queue)
1561 FOREACH_BLOCK (block) {
1562 mark_pinned_objects_in_block (block, queue);
1563 } END_FOREACH_BLOCK;
1567 major_init_to_space (void)
1572 major_report_pinned_memory_usage (void)
1574 g_assert_not_reached ();
1578 major_get_used_size (void)
1583 FOREACH_BLOCK (block) {
1584 int count = MS_BLOCK_FREE / block->obj_size;
1586 size += count * block->obj_size;
1587 for (iter = block->free_list; iter; iter = (void**)*iter)
1588 size -= block->obj_size;
1589 } END_FOREACH_BLOCK;
1595 get_num_major_sections (void)
1597 return num_major_sections;
1601 major_handle_gc_param (const char *opt)
1603 if (g_str_has_prefix (opt, "evacuation-threshold=")) {
1604 const char *arg = strchr (opt, '=') + 1;
1605 int percentage = atoi (arg);
1606 if (percentage < 0 || percentage > 100) {
1607 fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
1610 evacuation_threshold = (float)percentage / 100.0f;
1612 } else if (!strcmp (opt, "lazy-sweep")) {
1615 } else if (!strcmp (opt, "no-lazy-sweep")) {
1624 major_print_gc_param_usage (void)
1628 " evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
1629 " (no-)lazy-sweep\n"
1634 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
1637 gboolean has_references;
1639 FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1641 callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1642 } END_FOREACH_BLOCK;
1645 #ifdef HEAVY_STATISTICS
1646 extern guint64 marked_cards;
1647 extern guint64 scanned_cards;
1648 extern guint64 scanned_objects;
1649 extern guint64 remarked_cards;
1652 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
1654 * MS blocks are 16K aligned.
1655 * Cardtables are 4K aligned, at least.
1656 * This means that the cardtable of a given block is 32 bytes aligned.
1659 initial_skip_card (guint8 *card_data)
1661 mword *cards = (mword*)card_data;
1664 for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
1670 if (i == CARD_WORDS_PER_BLOCK)
1671 return card_data + CARDS_PER_BLOCK;
1673 #if defined(__i386__) && defined(__GNUC__)
1674 return card_data + i * 4 + (__builtin_ffs (card) - 1) / 8;
1675 #elif defined(__x86_64__) && defined(__GNUC__)
1676 return card_data + i * 8 + (__builtin_ffsll (card) - 1) / 8;
1677 #elif defined(__s390x__) && defined(__GNUC__)
1678 return card_data + i * 8 + (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
1680 for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
1682 return &card_data [i];
1688 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
1689 #define MS_BLOCK_OBJ_FAST(b,os,i) ((b) + MS_BLOCK_SKIP + (os) * (i))
1690 #define MS_OBJ_ALLOCED_FAST(o,b) (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
1693 card_offset (char *obj, char *base)
1695 return (obj - base) >> CARD_BITS;
1699 major_scan_card_table (gboolean mod_union, SgenGrayQueue *queue)
1702 gboolean has_references;
1703 ScanObjectFunc scan_func = sgen_get_current_object_ops ()->scan_object;
1705 if (!concurrent_mark)
1706 g_assert (!mod_union);
1708 FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1709 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
1710 guint8 cards_copy [CARDS_PER_BLOCK];
1712 gboolean small_objects;
1715 guint8 *card_data, *card_base;
1716 guint8 *card_data_end;
1717 char *scan_front = NULL;
1719 #ifdef PREFETCH_CARDS
1720 int prefetch_index = __index + 6;
1721 if (prefetch_index < allocated_blocks.next_slot) {
1722 MSBlockInfo *prefetch_block = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [prefetch_index]);
1723 guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
1724 PREFETCH_READ (prefetch_block);
1725 PREFETCH_WRITE (prefetch_cards);
1726 PREFETCH_WRITE (prefetch_cards + 32);
1730 if (!has_references)
1733 block_obj_size = block->obj_size;
1734 small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
1736 block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
1739 * This is safe in face of card aliasing for the following reason:
1741 * Major blocks are 16k aligned, or 32 cards aligned.
1742 * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
1743 * sizes, they won't overflow the cardtable overlap modulus.
1746 card_data = card_base = block->cardtable_mod_union;
1748 * This happens when the nursery collection that precedes finishing
1749 * the concurrent collection allocates new major blocks.
1754 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
1755 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
1757 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
1759 card_data = card_base = cards_copy;
1762 card_data_end = card_data + CARDS_PER_BLOCK;
1764 card_data += MS_BLOCK_SKIP >> CARD_BITS;
1766 card_data = initial_skip_card (card_data);
1767 while (card_data < card_data_end) {
1768 size_t card_index, first_object_index;
1771 char *first_obj, *obj;
1773 HEAVY_STAT (++scanned_cards);
1780 card_index = card_data - card_base;
1781 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
1782 end = start + CARD_SIZE_IN_BYTES;
1785 sweep_block (block, FALSE);
1787 HEAVY_STAT (++marked_cards);
1790 sgen_card_table_prepare_card_for_scanning (card_data);
1793 * If the card we're looking at starts at or in the block header, we
1794 * must start at the first object in the block, without calculating
1795 * the index of the object we're hypothetically starting at, because
1796 * it would be negative.
1798 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
1799 first_object_index = 0;
1801 first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
1803 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
1806 if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
1810 /* FIXME: do this more efficiently */
1812 MS_CALC_MARK_BIT (w, b, obj);
1813 if (!MS_MARK_BIT (block, w, b))
1817 if (small_objects) {
1818 HEAVY_STAT (++scanned_objects);
1819 scan_func (obj, sgen_obj_get_descriptor (obj), queue);
1821 size_t offset = card_offset (obj, block_start);
1822 sgen_cardtable_scan_object (obj, block_obj_size, card_base + offset, mod_union, queue);
1825 obj += block_obj_size;
1826 g_assert (scan_front <= obj);
1830 HEAVY_STAT (if (*card_data) ++remarked_cards);
1831 binary_protocol_card_scan (first_obj, obj - first_obj);
1836 card_data = card_base + card_offset (obj, block_start);
1838 } END_FOREACH_BLOCK;
1842 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
1845 gboolean has_references;
1846 long long total_cards = 0;
1847 long long marked_cards = 0;
1849 FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1850 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
1853 if (!has_references)
1856 total_cards += CARDS_PER_BLOCK;
1857 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
1861 } END_FOREACH_BLOCK;
1863 *num_total_cards = total_cards;
1864 *num_marked_cards = marked_cards;
1868 update_cardtable_mod_union (void)
1872 FOREACH_BLOCK (block) {
1875 block->cardtable_mod_union = sgen_card_table_update_mod_union (block->cardtable_mod_union,
1876 MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
1878 SGEN_ASSERT (0, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
1879 } END_FOREACH_BLOCK;
1883 major_get_cardtable_mod_union_for_object (char *obj)
1885 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
1886 size_t offset = card_offset (obj, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1887 return &block->cardtable_mod_union [offset];
1891 alloc_free_block_lists (MSBlockInfo ***lists)
1894 for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
1895 lists [i] = sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
1898 #undef pthread_create
1901 post_param_init (SgenMajorCollector *collector)
1903 collector->sweeps_lazily = lazy_sweep;
1907 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
1911 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
1913 num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
1914 block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
1915 ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
1917 evacuate_block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
1918 for (i = 0; i < num_block_obj_sizes; ++i)
1919 evacuate_block_obj_sizes [i] = FALSE;
1924 g_print ("block object sizes:\n");
1925 for (i = 0; i < num_block_obj_sizes; ++i)
1926 g_print ("%d\n", block_obj_sizes [i]);
1930 alloc_free_block_lists (free_block_lists);
1932 for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
1933 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
1934 for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
1935 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
1937 mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
1938 mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
1939 mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
1940 mono_counters_register ("# major objects evacuated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_objects_evacuated);
1941 #if SIZEOF_VOID_P != 8
1942 mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
1943 mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
1944 mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
1945 mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
1948 collector->section_size = MAJOR_SECTION_SIZE;
1950 concurrent_mark = is_concurrent;
1951 if (is_concurrent) {
1952 collector->is_concurrent = TRUE;
1953 collector->want_synchronous_collection = &want_evacuation;
1955 collector->is_concurrent = FALSE;
1956 collector->want_synchronous_collection = NULL;
1958 collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
1959 collector->supports_cardtable = TRUE;
1961 collector->alloc_heap = major_alloc_heap;
1962 collector->is_object_live = major_is_object_live;
1963 collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
1964 collector->alloc_degraded = major_alloc_degraded;
1966 collector->alloc_object = major_alloc_object;
1967 collector->free_pinned_object = free_pinned_object;
1968 collector->iterate_objects = major_iterate_objects;
1969 collector->free_non_pinned_object = major_free_non_pinned_object;
1970 collector->find_pin_queue_start_ends = major_find_pin_queue_start_ends;
1971 collector->pin_objects = major_pin_objects;
1972 collector->pin_major_object = pin_major_object;
1973 collector->scan_card_table = major_scan_card_table;
1974 collector->iterate_live_block_ranges = (void*)(void*) major_iterate_live_block_ranges;
1975 if (is_concurrent) {
1976 collector->update_cardtable_mod_union = update_cardtable_mod_union;
1977 collector->get_cardtable_mod_union_for_object = major_get_cardtable_mod_union_for_object;
1979 collector->init_to_space = major_init_to_space;
1980 collector->sweep = major_sweep;
1981 collector->have_finished_sweeping = major_have_finished_sweeping;
1982 collector->free_swept_blocks = major_free_swept_blocks;
1983 collector->check_scan_starts = major_check_scan_starts;
1984 collector->dump_heap = major_dump_heap;
1985 collector->get_used_size = major_get_used_size;
1986 collector->start_nursery_collection = major_start_nursery_collection;
1987 collector->finish_nursery_collection = major_finish_nursery_collection;
1988 collector->start_major_collection = major_start_major_collection;
1989 collector->finish_major_collection = major_finish_major_collection;
1990 collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
1991 collector->obj_is_from_pinned_alloc = obj_is_from_pinned_alloc;
1992 collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
1993 collector->get_num_major_sections = get_num_major_sections;
1994 collector->handle_gc_param = major_handle_gc_param;
1995 collector->print_gc_param_usage = major_print_gc_param_usage;
1996 collector->post_param_init = post_param_init;
1997 collector->is_valid_object = major_is_valid_object;
1998 collector->describe_pointer = major_describe_pointer;
1999 collector->count_cards = major_count_cards;
2001 collector->major_ops.copy_or_mark_object = major_copy_or_mark_object_canonical;
2002 collector->major_ops.scan_object = major_scan_object_with_evacuation;
2003 if (is_concurrent) {
2004 collector->major_concurrent_ops.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2005 collector->major_concurrent_ops.scan_object = major_scan_object_no_mark_concurrent;
2006 collector->major_concurrent_ops.scan_vtype = major_scan_vtype_concurrent;
2009 #if !defined (FIXED_HEAP) && !defined (SGEN_PARALLEL_MARK)
2010 /* FIXME: this will not work with evacuation or the split nursery. */
2012 collector->drain_gray_stack = drain_gray_stack;
2014 #ifdef HEAVY_STATISTICS
2015 mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2016 mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2017 mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2018 mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2019 mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2020 mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2021 mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2022 mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2023 mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2024 mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2026 mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2027 mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2028 mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2032 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2033 mono_mutex_init (&scanned_objects_list_lock);
2036 SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2038 /*cardtable requires major pages to be 8 cards aligned*/
2039 g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2043 sgen_marksweep_init (SgenMajorCollector *collector)
2045 sgen_marksweep_init_internal (collector, FALSE);
2049 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2051 sgen_marksweep_init_internal (collector, TRUE);