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 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
22 #include "mono/sgen/sgen-gc.h"
23 #include "mono/sgen/sgen-protocol.h"
24 #include "mono/sgen/sgen-cardtable.h"
25 #include "mono/sgen/sgen-memory-governor.h"
26 #include "mono/sgen/sgen-layout-stats.h"
27 #include "mono/sgen/sgen-pointer-queue.h"
28 #include "mono/sgen/sgen-array-list.h"
29 #include "mono/sgen/sgen-pinning.h"
30 #include "mono/sgen/sgen-workers.h"
31 #include "mono/sgen/sgen-thread-pool.h"
32 #include "mono/sgen/sgen-client.h"
33 #include "mono/utils/mono-memory-model.h"
35 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
36 #define MS_BLOCK_SIZE ARCH_MIN_MS_BLOCK_SIZE
37 #define MS_BLOCK_SIZE_SHIFT ARCH_MIN_MS_BLOCK_SIZE_SHIFT
39 #define MS_BLOCK_SIZE_SHIFT 14 /* INT FASTENABLE */
40 #define MS_BLOCK_SIZE (1 << MS_BLOCK_SIZE_SHIFT)
42 #define MAJOR_SECTION_SIZE MS_BLOCK_SIZE
43 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
46 * Don't allocate single blocks, but alloc a contingent of this many
47 * blocks in one swoop. This must be a power of two.
49 #define MS_BLOCK_ALLOC_NUM 32
52 * Number of bytes before the first object in a block. At the start
53 * of a block is the MSBlockHeader, then opional padding, then come
54 * the objects, so this must be >= sizeof (MSBlockHeader).
56 #define MS_BLOCK_SKIP ((sizeof (MSBlockHeader) + 15) & ~15)
58 #define MS_BLOCK_FREE (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
60 #define MS_NUM_MARK_WORDS ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
63 * Blocks progress from one state to the next:
65 * SWEPT The block is fully swept. It might or might not be in
68 * MARKING The block might or might not contain live objects. If
69 * we're in between an initial collection pause and the
70 * finishing pause, the block might or might not be in a
73 * CHECKING The sweep thread is investigating the block to determine
74 * whether or not it contains live objects. The block is
77 * NEED_SWEEPING The block contains live objects but has not yet been
78 * swept. It also contains free slots. It is in a block
81 * SWEEPING The block is being swept. It might be in a free list.
88 BLOCK_STATE_NEED_SWEEPING,
92 typedef struct _MSBlockInfo MSBlockInfo;
96 * FIXME: Do we even need this? It's only used during sweep and might be worth
97 * recalculating to save the space.
99 guint16 obj_size_index;
100 /* FIXME: Reduce this - it only needs a byte. */
101 volatile gint32 state;
103 unsigned int pinned : 1;
104 unsigned int has_references : 1;
105 unsigned int has_pinned : 1; /* means cannot evacuate */
106 unsigned int is_to_space : 1;
107 void ** volatile free_list;
108 MSBlockInfo * volatile next_free;
109 guint8 * volatile cardtable_mod_union;
110 mword mark_words [MS_NUM_MARK_WORDS];
113 #define MS_BLOCK_FOR_BLOCK_INFO(b) ((char*)(b))
115 #define MS_BLOCK_OBJ(b,i) ((GCObject *)(MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i)))
116 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size) (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
117 #define MS_BLOCK_DATA_FOR_OBJ(o) ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
123 #define MS_BLOCK_FOR_OBJ(o) (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
125 /* object index will always be small */
126 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
128 //casting to int is fine since blocks are 32k
129 #define MS_CALC_MARK_BIT(w,b,o) do { \
130 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
131 if (sizeof (mword) == 4) { \
140 #define MS_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] & (ONE_P << (b)))
141 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
143 #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))
145 #define MS_BLOCK_OBJ_SIZE_FACTOR (pow (2.0, 1.0 / 3))
148 * This way we can lookup block object size indexes for sizes up to
149 * 256 bytes with a single load.
151 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES 32
153 static int *block_obj_sizes;
154 static int num_block_obj_sizes;
155 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
157 #define MS_BLOCK_FLAG_PINNED 1
158 #define MS_BLOCK_FLAG_REFS 2
160 #define MS_BLOCK_TYPE_MAX 4
162 static gboolean *evacuate_block_obj_sizes;
163 static float evacuation_threshold = 0.666f;
165 static gboolean lazy_sweep = TRUE;
169 SWEEP_STATE_NEED_SWEEPING,
170 SWEEP_STATE_SWEEPING,
171 SWEEP_STATE_SWEEPING_AND_ITERATING,
172 SWEEP_STATE_COMPACTING
175 static volatile int sweep_state = SWEEP_STATE_SWEPT;
177 static gboolean concurrent_mark;
178 static gboolean concurrent_sweep = TRUE;
180 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl) SGEN_POINTER_IS_TAGGED_1 ((bl))
181 #define BLOCK_TAG_HAS_REFERENCES(bl) SGEN_POINTER_TAG_1 ((bl))
183 #define BLOCK_IS_TAGGED_CHECKING(bl) SGEN_POINTER_IS_TAGGED_2 ((bl))
184 #define BLOCK_TAG_CHECKING(bl) SGEN_POINTER_TAG_2 ((bl))
186 #define BLOCK_UNTAG(bl) ((MSBlockInfo *)SGEN_POINTER_UNTAG_12 ((bl)))
188 #define BLOCK_TAG(bl) ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
190 /* all allocated blocks in the system */
191 static SgenArrayList allocated_blocks = SGEN_ARRAY_LIST_INIT (NULL, NULL, NULL, INTERNAL_MEM_PIN_QUEUE);
193 /* non-allocated block free-list */
194 static void *empty_blocks = NULL;
195 static size_t num_empty_blocks = 0;
198 * We can iterate the block list also while sweep is in progress but we
199 * need to account for blocks that will be checked for sweeping and even
200 * freed in the process.
202 #define FOREACH_BLOCK_NO_LOCK(bl) { \
203 volatile gpointer *slot; \
204 SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) { \
205 (bl) = BLOCK_UNTAG (*slot); \
208 #define FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK(bl,hr) { \
209 volatile gpointer *slot; \
210 SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) { \
211 (bl) = (MSBlockInfo *) (*slot); \
214 (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl)); \
215 (bl) = BLOCK_UNTAG ((bl));
216 #define END_FOREACH_BLOCK_NO_LOCK } SGEN_ARRAY_LIST_END_FOREACH_SLOT; }
218 static volatile size_t num_major_sections = 0;
220 * One free block list for each block object size. We add and remove blocks from these
221 * lists lock-free via CAS.
223 * Blocks accessed/removed from `free_block_lists`:
224 * from the mutator (with GC lock held)
225 * in nursery collections
226 * in non-concurrent major collections
227 * in the finishing pause of concurrent major collections (whole list is cleared)
229 * Blocks added to `free_block_lists`:
230 * in the sweeping thread
231 * during nursery collections
232 * from domain clearing (with the world stopped and no sweeping happening)
234 * The only item of those that doesn't require the GC lock is the sweep thread. The sweep
235 * thread only ever adds blocks to the free list, so the ABA problem can't occur.
237 static MSBlockInfo * volatile *free_block_lists [MS_BLOCK_TYPE_MAX];
239 static guint64 stat_major_blocks_alloced = 0;
240 static guint64 stat_major_blocks_freed = 0;
241 static guint64 stat_major_blocks_lazy_swept = 0;
243 #if SIZEOF_VOID_P != 8
244 static guint64 stat_major_blocks_freed_ideal = 0;
245 static guint64 stat_major_blocks_freed_less_ideal = 0;
246 static guint64 stat_major_blocks_freed_individual = 0;
247 static guint64 stat_major_blocks_alloced_less_ideal = 0;
250 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
251 static guint64 num_major_objects_marked = 0;
252 #define INC_NUM_MAJOR_OBJECTS_MARKED() (++num_major_objects_marked)
254 #define INC_NUM_MAJOR_OBJECTS_MARKED()
257 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
258 static mono_mutex_t scanned_objects_list_lock;
259 static SgenPointerQueue scanned_objects_list;
262 add_scanned_object (void *ptr)
264 if (!binary_protocol_is_enabled ())
267 mono_os_mutex_lock (&scanned_objects_list_lock);
268 sgen_pointer_queue_add (&scanned_objects_list, ptr);
269 mono_os_mutex_unlock (&scanned_objects_list_lock);
273 static gboolean sweep_block (MSBlockInfo *block);
276 ms_find_block_obj_size_index (size_t size)
279 SGEN_ASSERT (9, size <= SGEN_MAX_SMALL_OBJ_SIZE, "size %zd is bigger than max small object size %d", size, SGEN_MAX_SMALL_OBJ_SIZE);
280 for (i = 0; i < num_block_obj_sizes; ++i)
281 if (block_obj_sizes [i] >= size)
283 g_error ("no object of size %zd\n", size);
287 #define FREE_BLOCKS_FROM(lists,p,r) (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
288 #define FREE_BLOCKS(p,r) (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
290 #define MS_BLOCK_OBJ_SIZE_INDEX(s) \
291 (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ? \
292 fast_block_obj_size_indexes [((s)+7)>>3] : \
293 ms_find_block_obj_size_index ((s)))
296 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
300 start = (char *)sgen_alloc_os_memory_aligned (nursery_size, nursery_align, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery");
302 start = (char *)sgen_alloc_os_memory (nursery_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery");
308 update_heap_boundaries_for_block (MSBlockInfo *block)
310 sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
317 ms_get_empty_block (void)
321 void *block, *empty, *next;
326 * We try allocating MS_BLOCK_ALLOC_NUM blocks first. If that's
327 * unsuccessful, we halve the number of blocks and try again, until we're at
328 * 1. If that doesn't work, either, we assert.
330 int alloc_num = MS_BLOCK_ALLOC_NUM;
332 p = (char *)sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE,
333 (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE),
334 alloc_num == 1 ? "major heap section" : NULL);
340 for (i = 0; i < alloc_num; ++i) {
343 * We do the free list update one after the
344 * other so that other threads can use the new
345 * blocks as quickly as possible.
348 empty = empty_blocks;
349 *(void**)block = empty;
350 } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
354 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
356 stat_major_blocks_alloced += alloc_num;
357 #if SIZEOF_VOID_P != 8
358 if (alloc_num != MS_BLOCK_ALLOC_NUM)
359 stat_major_blocks_alloced_less_ideal += alloc_num;
364 empty = empty_blocks;
368 next = *(void**)block;
369 } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
371 SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
373 *(void**)block = NULL;
375 g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
381 * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
382 * list, where it will either be freed later on, or reused in nursery collections.
385 ms_free_block (MSBlockInfo *info)
388 char *block = MS_BLOCK_FOR_BLOCK_INFO (info);
390 sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
391 if (info->cardtable_mod_union)
392 sgen_card_table_free_mod_union (info->cardtable_mod_union, block, MS_BLOCK_SIZE);
393 memset (block, 0, MS_BLOCK_SIZE);
396 empty = empty_blocks;
397 *(void**)block = empty;
398 } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
400 SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
402 binary_protocol_block_free (block, MS_BLOCK_SIZE);
406 sweep_in_progress (void)
408 int state = sweep_state;
409 return state == SWEEP_STATE_SWEEPING ||
410 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
411 state == SWEEP_STATE_COMPACTING;
414 static inline gboolean
415 block_is_swept_or_marking (MSBlockInfo *block)
417 gint32 state = block->state;
418 return state == BLOCK_STATE_SWEPT || state == BLOCK_STATE_MARKING;
421 //#define MARKSWEEP_CONSISTENCY_CHECK
423 #ifdef MARKSWEEP_CONSISTENCY_CHECK
425 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
427 SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
428 for (; block; block = block->next_free) {
429 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
430 g_assert (block->obj_size == size);
431 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
433 /* blocks in the free lists must have at least
435 g_assert (block->free_list);
437 /* the block must be in the allocated_blocks array */
438 g_assert (sgen_array_list_find (&allocated_blocks, BLOCK_TAG (block)) != (guint32)-1);
443 check_empty_blocks (void)
447 for (p = empty_blocks; p; p = *(void**)p)
449 g_assert (i == num_empty_blocks);
453 consistency_check (void)
458 /* check all blocks */
459 FOREACH_BLOCK_NO_LOCK (block) {
460 int count = MS_BLOCK_FREE / block->obj_size;
464 /* count number of free slots */
465 for (i = 0; i < count; ++i) {
466 void **obj = (void**) MS_BLOCK_OBJ (block, i);
467 if (!MS_OBJ_ALLOCED (obj, block))
471 /* check free list */
472 for (free = block->free_list; free; free = (void**)*free) {
473 g_assert (MS_BLOCK_FOR_OBJ (free) == block);
476 g_assert (num_free == 0);
478 /* check all mark words are zero */
479 if (!sgen_concurrent_collection_in_progress () && block_is_swept_or_marking (block)) {
480 for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
481 g_assert (block->mark_words [i] == 0);
483 } END_FOREACH_BLOCK_NO_LOCK;
485 /* check free blocks */
486 for (i = 0; i < num_block_obj_sizes; ++i) {
488 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
489 check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
492 check_empty_blocks ();
497 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
501 block->next_free = old = free_blocks [size_index];
502 } while (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], block, old) != old);
505 static void major_finish_sweep_checking (void);
508 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
510 int size = block_obj_sizes [size_index];
511 int count = MS_BLOCK_FREE / size;
513 MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
517 if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
520 info = (MSBlockInfo*)ms_get_empty_block ();
522 SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
524 info->obj_size = size;
525 info->obj_size_index = size_index;
526 info->pinned = pinned;
527 info->has_references = has_references;
528 info->has_pinned = pinned;
530 * Blocks that are to-space are not evacuated from. During an major collection
531 * blocks are allocated for two reasons: evacuating objects from the nursery and
532 * evacuating them from major blocks marked for evacuation. In both cases we don't
533 * want further evacuation. We also don't want to evacuate objects allocated during
534 * the concurrent mark since it would add pointless stress on the finishing pause.
536 info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD) || sgen_concurrent_collection_in_progress ();
537 info->state = info->is_to_space ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
538 SGEN_ASSERT (6, !sweep_in_progress () || info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
539 info->cardtable_mod_union = NULL;
541 update_heap_boundaries_for_block (info);
543 binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
545 /* build free list */
546 obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
547 info->free_list = (void**)obj_start;
548 /* we're skipping the last one - it must be nulled */
549 for (i = 0; i < count - 1; ++i) {
550 char *next_obj_start = obj_start + size;
551 *(void**)obj_start = next_obj_start;
552 obj_start = next_obj_start;
555 *(void**)obj_start = NULL;
557 add_free_block (free_blocks, size_index, info);
559 sgen_array_list_add (&allocated_blocks, BLOCK_TAG (info), 0, FALSE);
561 SGEN_ATOMIC_ADD_P (num_major_sections, 1);
566 ptr_is_from_pinned_alloc (char *ptr)
570 FOREACH_BLOCK_NO_LOCK (block) {
571 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE)
572 return block->pinned;
573 } END_FOREACH_BLOCK_NO_LOCK;
578 ensure_can_access_block_free_list (MSBlockInfo *block)
582 switch (block->state) {
583 case BLOCK_STATE_SWEPT:
584 case BLOCK_STATE_MARKING:
586 case BLOCK_STATE_CHECKING:
587 SGEN_ASSERT (0, FALSE, "How did we get a block that's being checked from a free list?");
589 case BLOCK_STATE_NEED_SWEEPING:
590 if (sweep_block (block))
591 ++stat_major_blocks_lazy_swept;
593 case BLOCK_STATE_SWEEPING:
594 /* FIXME: do this more elegantly */
598 SGEN_ASSERT (0, FALSE, "Illegal block state");
605 unlink_slot_from_free_list_uncontested (MSBlockInfo * volatile *free_blocks, int size_index)
607 MSBlockInfo *block, *next_free_block;
608 void *obj, *next_free_slot;
611 block = free_blocks [size_index];
612 SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
614 ensure_can_access_block_free_list (block);
616 obj = block->free_list;
617 SGEN_ASSERT (6, obj, "block %p in free list had no available object to alloc from", block);
619 next_free_slot = *(void**)obj;
620 if (next_free_slot) {
621 block->free_list = (gpointer *)next_free_slot;
625 next_free_block = block->next_free;
626 if (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], next_free_block, block) != block)
629 block->free_list = NULL;
630 block->next_free = NULL;
636 alloc_obj (GCVTable vtable, size_t size, gboolean pinned, gboolean has_references)
638 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
639 MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
642 if (!free_blocks [size_index]) {
643 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
647 obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
649 /* FIXME: assumes object layout */
650 *(GCVTable*)obj = vtable;
652 total_allocated_major += block_obj_sizes [size_index];
654 return (GCObject *)obj;
658 major_alloc_object (GCVTable vtable, size_t size, gboolean has_references)
660 return alloc_obj (vtable, size, FALSE, has_references);
664 * We're not freeing the block if it's empty. We leave that work for
665 * the next major collection.
667 * This is just called from the domain clearing code, which runs in a
668 * single thread and has the GC lock, so we don't need an extra lock.
671 free_object (GCObject *obj, size_t size, gboolean pinned)
673 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
675 gboolean in_free_list;
677 SGEN_ASSERT (9, sweep_state == SWEEP_STATE_SWEPT, "Should have waited for sweep to free objects.");
679 ensure_can_access_block_free_list (block);
680 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);
681 SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
682 MS_CALC_MARK_BIT (word, bit, obj);
683 SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set", obj);
685 memset (obj, 0, size);
687 in_free_list = !!block->free_list;
688 *(void**)obj = block->free_list;
689 block->free_list = (void**)obj;
692 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (pinned, block->has_references);
693 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
694 SGEN_ASSERT (9, !block->next_free, "block %p doesn't have a free-list of object but belongs to a free-list of blocks", block);
695 add_free_block (free_blocks, size_index, block);
700 major_free_non_pinned_object (GCObject *obj, size_t size)
702 free_object (obj, size, FALSE);
705 /* size is a multiple of SGEN_ALLOC_ALIGN */
707 major_alloc_small_pinned_obj (GCVTable vtable, size_t size, gboolean has_references)
711 res = alloc_obj (vtable, size, TRUE, has_references);
712 /*If we failed to alloc memory, we better try releasing memory
713 *as pinned alloc is requested by the runtime.
716 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE, TRUE);
717 res = alloc_obj (vtable, size, TRUE, has_references);
719 return (GCObject *)res;
723 free_pinned_object (GCObject *obj, size_t size)
725 free_object (obj, size, TRUE);
729 * size is already rounded up and we hold the GC lock.
732 major_alloc_degraded (GCVTable vtable, size_t size)
736 obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
737 if (G_LIKELY (obj)) {
738 HEAVY_STAT (++stat_objects_alloced_degraded);
739 HEAVY_STAT (stat_bytes_alloced_degraded += size);
745 * obj is some object. If it's not in the major heap (i.e. if it's in
746 * the nursery or LOS), return FALSE. Otherwise return whether it's
747 * been marked or copied.
750 major_is_object_live (GCObject *obj)
756 if (sgen_ptr_in_nursery (obj))
759 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
762 if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
765 /* now we know it's in a major block */
766 block = MS_BLOCK_FOR_OBJ (obj);
767 SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?", block);
768 MS_CALC_MARK_BIT (word, bit, obj);
769 return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
773 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
777 FOREACH_BLOCK_NO_LOCK (block) {
778 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
779 int count = MS_BLOCK_FREE / block->obj_size;
783 for (i = 0; i <= count; ++i) {
784 if (ptr >= (char*)MS_BLOCK_OBJ (block, i) && ptr < (char*)MS_BLOCK_OBJ (block, i + 1)) {
785 *start = (char *)MS_BLOCK_OBJ (block, i);
789 return !block->pinned;
791 } END_FOREACH_BLOCK_NO_LOCK;
796 try_set_sweep_state (int new_, int expected)
798 int old = SGEN_CAS (&sweep_state, new_, expected);
799 return old == expected;
803 set_sweep_state (int new_, int expected)
805 gboolean success = try_set_sweep_state (new_, expected);
806 SGEN_ASSERT (0, success, "Could not set sweep state.");
809 static gboolean ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked);
811 static SgenThreadPoolJob * volatile sweep_job;
812 static SgenThreadPoolJob * volatile sweep_blocks_job;
815 major_finish_sweep_checking (void)
818 SgenThreadPoolJob *job;
821 switch (sweep_state) {
822 case SWEEP_STATE_SWEPT:
823 case SWEEP_STATE_NEED_SWEEPING:
825 case SWEEP_STATE_SWEEPING:
826 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
829 case SWEEP_STATE_SWEEPING_AND_ITERATING:
830 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
832 case SWEEP_STATE_COMPACTING:
835 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
840 * We're running with the world stopped and the only other thread doing work is the
841 * sweep thread, which doesn't add blocks to the array, so we can safely access
844 for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index)
845 ensure_block_is_checked_for_sweeping (block_index, FALSE, NULL);
847 set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
852 sgen_thread_pool_job_wait (job);
853 SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
854 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
858 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
860 gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
861 gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
862 gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
865 major_finish_sweep_checking ();
866 FOREACH_BLOCK_NO_LOCK (block) {
867 int count = MS_BLOCK_FREE / block->obj_size;
870 if (block->pinned && !pinned)
872 if (!block->pinned && !non_pinned)
874 if (sweep && lazy_sweep) {
876 SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
879 for (i = 0; i < count; ++i) {
880 void **obj = (void**) MS_BLOCK_OBJ (block, i);
882 * We've finished sweep checking, but if we're sweeping lazily and
883 * the flags don't require us to sweep, the block might still need
884 * sweeping. In that case, we need to consult the mark bits to tell
885 * us whether an object slot is live.
887 if (!block_is_swept_or_marking (block)) {
889 SGEN_ASSERT (6, !sweep && block->state == BLOCK_STATE_NEED_SWEEPING, "Has sweeping not finished?");
890 MS_CALC_MARK_BIT (word, bit, obj);
891 if (!MS_MARK_BIT (block, word, bit))
894 if (MS_OBJ_ALLOCED (obj, block))
895 callback ((GCObject*)obj, block->obj_size, data);
897 } END_FOREACH_BLOCK_NO_LOCK;
901 major_is_valid_object (char *object)
905 FOREACH_BLOCK_NO_LOCK (block) {
909 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
912 idx = MS_BLOCK_OBJ_INDEX (object, block);
913 obj = (char*)MS_BLOCK_OBJ (block, idx);
916 return MS_OBJ_ALLOCED (obj, block);
917 } END_FOREACH_BLOCK_NO_LOCK;
924 major_describe_pointer (char *ptr)
928 FOREACH_BLOCK_NO_LOCK (block) {
936 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
939 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
940 MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
942 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
943 obj = (char*)MS_BLOCK_OBJ (block, idx);
944 live = MS_OBJ_ALLOCED (obj, block);
945 vtable = live ? SGEN_LOAD_VTABLE ((GCObject*)obj) : NULL;
947 MS_CALC_MARK_BIT (w, b, obj);
948 marked = MS_MARK_BIT (block, w, b);
953 SGEN_LOG (0, "object");
955 SGEN_LOG (0, "dead-object");
958 SGEN_LOG (0, "interior-ptr offset %zd", ptr - obj);
960 SGEN_LOG (0, "dead-interior-ptr offset %zd", ptr - obj);
963 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
966 } END_FOREACH_BLOCK_NO_LOCK;
972 major_check_scan_starts (void)
977 major_dump_heap (FILE *heap_dump_file)
980 int *slots_available = (int *)alloca (sizeof (int) * num_block_obj_sizes);
981 int *slots_used = (int *)alloca (sizeof (int) * num_block_obj_sizes);
984 for (i = 0; i < num_block_obj_sizes; ++i)
985 slots_available [i] = slots_used [i] = 0;
987 FOREACH_BLOCK_NO_LOCK (block) {
988 int index = ms_find_block_obj_size_index (block->obj_size);
989 int count = MS_BLOCK_FREE / block->obj_size;
991 slots_available [index] += count;
992 for (i = 0; i < count; ++i) {
993 if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
994 ++slots_used [index];
996 } END_FOREACH_BLOCK_NO_LOCK;
998 fprintf (heap_dump_file, "<occupancies>\n");
999 for (i = 0; i < num_block_obj_sizes; ++i) {
1000 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
1001 block_obj_sizes [i], slots_available [i], slots_used [i]);
1003 fprintf (heap_dump_file, "</occupancies>\n");
1005 FOREACH_BLOCK_NO_LOCK (block) {
1006 int count = MS_BLOCK_FREE / block->obj_size;
1010 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
1012 for (i = 0; i <= count; ++i) {
1013 if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
1018 sgen_dump_occupied ((char *)MS_BLOCK_OBJ (block, start), (char *)MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
1024 fprintf (heap_dump_file, "</section>\n");
1025 } END_FOREACH_BLOCK_NO_LOCK;
1029 get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate)
1031 guint8 *mod_union = block->cardtable_mod_union;
1037 mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1038 other = (guint8 *)SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
1040 SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
1043 sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1047 static inline guint8*
1048 major_get_cardtable_mod_union_for_reference (char *ptr)
1050 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (ptr);
1051 size_t offset = sgen_card_table_get_card_offset (ptr, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1052 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
1053 SGEN_ASSERT (0, mod_union, "FIXME: optionally allocate the mod union if it's not here and CAS it in.");
1054 return &mod_union [offset];
1058 * Mark the mod-union card for `ptr`, which must be a reference within the object `obj`.
1061 mark_mod_union_card (GCObject *obj, void **ptr, GCObject *value_obj)
1063 int type = sgen_obj_get_descriptor (obj) & DESC_TYPE_MASK;
1064 if (sgen_safe_object_is_small (obj, type)) {
1065 guint8 *card_byte = major_get_cardtable_mod_union_for_reference ((char*)ptr);
1066 SGEN_ASSERT (0, MS_BLOCK_FOR_OBJ (obj) == MS_BLOCK_FOR_OBJ (ptr), "How can an object and a reference inside it not be in the same block?");
1069 sgen_los_mark_mod_union_card (obj, ptr);
1071 binary_protocol_mod_union_remset (obj, ptr, value_obj, SGEN_LOAD_VTABLE (value_obj));
1074 static inline gboolean
1075 major_block_is_evacuating (MSBlockInfo *block)
1077 if (evacuate_block_obj_sizes [block->obj_size_index] &&
1078 !block->has_pinned &&
1079 !block->is_to_space)
1084 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do { \
1085 int __word, __bit; \
1086 MS_CALC_MARK_BIT (__word, __bit, (obj)); \
1087 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
1088 if (!MS_MARK_BIT ((block), __word, __bit)) { \
1089 MS_SET_MARK_BIT ((block), __word, __bit); \
1090 if (sgen_gc_descr_has_references (desc)) \
1091 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1092 binary_protocol_mark ((obj), (gpointer)SGEN_LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1093 INC_NUM_MAJOR_OBJECTS_MARKED (); \
1098 pin_major_object (GCObject *obj, SgenGrayQueue *queue)
1102 if (concurrent_mark)
1103 g_assert_not_reached ();
1105 block = MS_BLOCK_FOR_OBJ (obj);
1106 block->has_pinned = TRUE;
1107 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1110 #include "sgen-major-copy-object.h"
1113 major_get_and_reset_num_major_objects_marked (void)
1115 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1116 long long num = num_major_objects_marked;
1117 num_major_objects_marked = 0;
1124 #define PREFETCH_CARDS 1 /* BOOL FASTENABLE */
1126 #undef PREFETCH_CARDS
1129 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1130 #if defined(PLATFORM_MACOSX)
1131 #define GCC_VERSION (__GNUC__ * 10000 \
1132 + __GNUC_MINOR__ * 100 \
1133 + __GNUC_PATCHLEVEL__)
1134 #if GCC_VERSION <= 40300
1135 #undef PREFETCH_CARDS
1139 #ifdef HEAVY_STATISTICS
1140 static guint64 stat_optimized_copy;
1141 static guint64 stat_optimized_copy_nursery;
1142 static guint64 stat_optimized_copy_nursery_forwarded;
1143 static guint64 stat_optimized_copy_nursery_pinned;
1144 static guint64 stat_optimized_copy_major;
1145 static guint64 stat_optimized_copy_major_small_fast;
1146 static guint64 stat_optimized_copy_major_small_slow;
1147 static guint64 stat_optimized_copy_major_large;
1148 static guint64 stat_optimized_copy_major_forwarded;
1149 static guint64 stat_optimized_copy_major_small_evacuate;
1150 static guint64 stat_optimized_major_scan;
1151 static guint64 stat_optimized_major_scan_no_refs;
1153 static guint64 stat_drain_prefetch_fills;
1154 static guint64 stat_drain_prefetch_fill_failures;
1155 static guint64 stat_drain_loops;
1158 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_no_evacuation
1159 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_no_evacuation
1160 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_no_evacuation
1161 #include "sgen-marksweep-drain-gray-stack.h"
1163 #define COPY_OR_MARK_WITH_EVACUATION
1164 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_with_evacuation
1165 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_with_evacuation
1166 #define SCAN_VTYPE_FUNCTION_NAME major_scan_vtype_with_evacuation
1167 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_with_evacuation
1168 #define SCAN_PTR_FIELD_FUNCTION_NAME major_scan_ptr_field_with_evacuation
1169 #include "sgen-marksweep-drain-gray-stack.h"
1171 #define COPY_OR_MARK_CONCURRENT
1172 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_concurrent_no_evacuation
1173 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_concurrent_no_evacuation
1174 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_concurrent_no_evacuation
1175 #include "sgen-marksweep-drain-gray-stack.h"
1177 #define COPY_OR_MARK_CONCURRENT_WITH_EVACUATION
1178 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_concurrent_with_evacuation
1179 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_concurrent_with_evacuation
1180 #define SCAN_VTYPE_FUNCTION_NAME major_scan_vtype_concurrent_with_evacuation
1181 #define SCAN_PTR_FIELD_FUNCTION_NAME major_scan_ptr_field_concurrent_with_evacuation
1182 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_concurrent_with_evacuation
1183 #include "sgen-marksweep-drain-gray-stack.h"
1185 static inline gboolean
1186 major_is_evacuating (void)
1189 for (i = 0; i < num_block_obj_sizes; ++i) {
1190 if (evacuate_block_obj_sizes [i]) {
1199 drain_gray_stack (SgenGrayQueue *queue)
1201 if (major_is_evacuating ())
1202 return drain_gray_stack_with_evacuation (queue);
1204 return drain_gray_stack_no_evacuation (queue);
1208 drain_gray_stack_concurrent (SgenGrayQueue *queue)
1210 if (major_is_evacuating ())
1211 return drain_gray_stack_concurrent_with_evacuation (queue);
1213 return drain_gray_stack_concurrent_no_evacuation (queue);
1217 major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue)
1219 major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1223 major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *queue)
1225 major_copy_or_mark_object_concurrent_with_evacuation (ptr, *ptr, queue);
1229 major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue)
1231 major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1235 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1237 void **entry, **end;
1238 int last_index = -1;
1240 if (first_entry == last_entry)
1243 entry = sgen_pinning_get_entry (first_entry);
1244 end = sgen_pinning_get_entry (last_entry);
1246 for (; entry < end; ++entry) {
1247 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1249 SGEN_ASSERT (9, index >= 0 && index < MS_BLOCK_FREE / block->obj_size, "invalid object %p index %d max-index %d", *entry, index, (int)(MS_BLOCK_FREE / block->obj_size));
1250 if (index == last_index)
1252 obj = MS_BLOCK_OBJ (block, index);
1253 if (!MS_OBJ_ALLOCED (obj, block))
1255 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1256 sgen_pin_stats_register_object (obj, GENERATION_OLD);
1261 * There might have been potential pinning "pointers" into this block, but none of
1262 * them pointed to occupied slots, in which case we don't have to pin the block.
1264 if (last_index >= 0)
1265 block->has_pinned = TRUE;
1269 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1273 for (obj_index = 0; obj_index < count; ++obj_index) {
1275 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1277 MS_CALC_MARK_BIT (word, bit, obj);
1278 if (MS_MARK_BIT (block, word, bit)) {
1279 SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1281 /* an unmarked object */
1282 if (MS_OBJ_ALLOCED (obj, block)) {
1284 * FIXME: Merge consecutive
1285 * slots for lower reporting
1286 * overhead. Maybe memset
1287 * will also benefit?
1289 binary_protocol_empty (obj, obj_size);
1290 memset (obj, 0, obj_size);
1292 *(void**)obj = block->free_list;
1293 block->free_list = (void **)obj;
1298 static inline gboolean
1299 try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1301 gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
1302 gboolean success = old_state == expected_state;
1304 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
1309 set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1311 SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set");
1312 block->state = new_state;
1313 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, expected_state, new_state);
1317 * If `block` needs sweeping, sweep it and return TRUE. Otherwise return FALSE.
1319 * Sweeping means iterating through the block's slots and building the free-list from the
1320 * unmarked ones. They will also be zeroed. The mark bits will be reset.
1323 sweep_block (MSBlockInfo *block)
1326 void *reversed = NULL;
1329 switch (block->state) {
1330 case BLOCK_STATE_SWEPT:
1332 case BLOCK_STATE_MARKING:
1333 case BLOCK_STATE_CHECKING:
1334 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1336 case BLOCK_STATE_SWEEPING:
1337 /* FIXME: Do this more elegantly */
1340 case BLOCK_STATE_NEED_SWEEPING:
1341 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1345 SGEN_ASSERT (0, FALSE, "Illegal block state");
1348 SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1350 count = MS_BLOCK_FREE / block->obj_size;
1352 block->free_list = NULL;
1354 /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1355 // FIXME: Add more sizes
1356 switch (block->obj_size) {
1358 sweep_block_for_size (block, count, 16);
1361 sweep_block_for_size (block, count, block->obj_size);
1365 /* reset mark bits */
1366 memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1368 /* Reverse free list so that it's in address order */
1370 while (block->free_list) {
1371 void *next = *(void**)block->free_list;
1372 *(void**)block->free_list = reversed;
1373 reversed = block->free_list;
1374 block->free_list = (void **)next;
1376 block->free_list = (void **)reversed;
1378 mono_memory_write_barrier ();
1380 set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1391 if (sizeof (mword) == 8)
1392 count += __builtin_popcountll (d);
1394 count += __builtin_popcount (d);
1404 /* statistics for evacuation */
1405 static size_t *sweep_slots_available;
1406 static size_t *sweep_slots_used;
1407 static size_t *sweep_num_blocks;
1409 static volatile size_t num_major_sections_before_sweep;
1410 static volatile size_t num_major_sections_freed_in_sweep;
1417 for (i = 0; i < num_block_obj_sizes; ++i)
1418 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1420 /* clear all the free lists */
1421 for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1422 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1424 for (j = 0; j < num_block_obj_sizes; ++j)
1425 free_blocks [j] = NULL;
1428 sgen_array_list_remove_nulls (&allocated_blocks);
1431 static void sweep_finish (void);
1434 * If `wait` is TRUE and the block is currently being checked, this function will wait until
1435 * the checking has finished.
1437 * Returns whether the block is still there. If `wait` is FALSE, the return value will not
1438 * be correct, i.e. must not be used.
1441 ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked)
1444 gboolean have_live = FALSE;
1445 gboolean have_free = FALSE;
1451 volatile gpointer *block_slot = sgen_array_list_get_slot (&allocated_blocks, block_index);
1453 SGEN_ASSERT (6, sweep_in_progress (), "Why do we call this function if there's no sweep in progress?");
1456 *have_checked = FALSE;
1459 tagged_block = *(void * volatile *)block_slot;
1463 if (BLOCK_IS_TAGGED_CHECKING (tagged_block)) {
1466 /* FIXME: do this more elegantly */
1471 if (SGEN_CAS_PTR (block_slot, BLOCK_TAG_CHECKING (tagged_block), tagged_block) != tagged_block)
1474 block = BLOCK_UNTAG (tagged_block);
1475 block_state = block->state;
1477 if (!sweep_in_progress ()) {
1478 SGEN_ASSERT (6, block_state != BLOCK_STATE_SWEEPING && block_state != BLOCK_STATE_CHECKING, "Invalid block state.");
1480 SGEN_ASSERT (6, block_state != BLOCK_STATE_NEED_SWEEPING, "Invalid block state.");
1483 switch (block_state) {
1484 case BLOCK_STATE_SWEPT:
1485 case BLOCK_STATE_NEED_SWEEPING:
1486 case BLOCK_STATE_SWEEPING:
1488 case BLOCK_STATE_MARKING:
1490 case BLOCK_STATE_CHECKING:
1491 SGEN_ASSERT (0, FALSE, "We set the CHECKING bit - how can the stage be CHECKING?");
1494 SGEN_ASSERT (0, FALSE, "Illegal block state");
1498 SGEN_ASSERT (6, block->state == BLOCK_STATE_MARKING, "When we sweep all blocks must start out marking.");
1499 set_block_state (block, BLOCK_STATE_CHECKING, BLOCK_STATE_MARKING);
1502 *have_checked = TRUE;
1504 block->has_pinned = block->pinned;
1506 block->is_to_space = FALSE;
1508 count = MS_BLOCK_FREE / block->obj_size;
1510 if (block->cardtable_mod_union)
1511 memset (block->cardtable_mod_union, 0, CARDS_PER_BLOCK);
1513 /* Count marked objects in the block */
1514 for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1515 nused += bitcount (block->mark_words [i]);
1517 block->nused = nused;
1524 int obj_size_index = block->obj_size_index;
1525 gboolean has_pinned = block->has_pinned;
1527 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1530 * FIXME: Go straight to SWEPT if there are no free slots. We need
1531 * to set the free slot list to NULL, though, and maybe update some
1535 sweep_block (block);
1538 ++sweep_num_blocks [obj_size_index];
1539 sweep_slots_used [obj_size_index] += nused;
1540 sweep_slots_available [obj_size_index] += count;
1544 * If there are free slots in the block, add
1545 * the block to the corresponding free list.
1548 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1551 SGEN_ASSERT (6, block->free_list, "How do we not have a free list when there are free slots?");
1553 add_free_block (free_blocks, obj_size_index, block);
1556 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1557 update_heap_boundaries_for_block (block);
1560 * Blocks without live objects are removed from the
1561 * block list and freed.
1563 SGEN_ASSERT (6, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1564 SGEN_ASSERT (6, *block_slot == BLOCK_TAG_CHECKING (tagged_block), "How did the block move?");
1566 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1567 ms_free_block (block);
1569 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1571 tagged_block = NULL;
1575 *block_slot = tagged_block;
1576 return !!tagged_block;
1580 sweep_blocks_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1582 volatile gpointer *slot;
1585 SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) {
1586 bl = BLOCK_UNTAG (*slot);
1589 } SGEN_ARRAY_LIST_END_FOREACH_SLOT;
1591 mono_memory_write_barrier ();
1593 sweep_blocks_job = NULL;
1597 sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1599 guint32 block_index;
1600 guint32 num_blocks = num_major_sections_before_sweep;
1602 SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1603 SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1606 * We traverse the block array from high to low. Nursery collections will have to
1607 * cooperate with the sweep thread to finish sweeping, and they will traverse from
1608 * low to high, to avoid constantly colliding on the same blocks.
1610 for (block_index = num_blocks; block_index-- > 0;) {
1612 * The block might have been freed by another thread doing some checking
1615 if (!ensure_block_is_checked_for_sweeping (block_index, TRUE, NULL))
1616 ++num_major_sections_freed_in_sweep;
1619 while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) {
1621 * The main GC thread is currently iterating over the block array to help us
1622 * finish the sweep. We have already finished, but we don't want to mess up
1623 * that iteration, so we just wait for it.
1628 if (SGEN_MAX_ASSERT_LEVEL >= 6) {
1629 for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1630 MSBlockInfo *block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, block_index));
1631 SGEN_ASSERT (6, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1636 * Concurrently sweep all the blocks to reduce workload during minor
1637 * pauses where we need certain blocks to be swept. At the start of
1638 * the next major we need all blocks to be swept anyway.
1640 if (concurrent_sweep && lazy_sweep) {
1641 sweep_blocks_job = sgen_thread_pool_job_alloc ("sweep_blocks", sweep_blocks_job_func, sizeof (SgenThreadPoolJob));
1642 sgen_thread_pool_job_enqueue (sweep_blocks_job);
1653 mword used_slots_size = 0;
1656 for (i = 0; i < num_block_obj_sizes; ++i) {
1657 float usage = (float)sweep_slots_used [i] / (float)sweep_slots_available [i];
1658 if (sweep_num_blocks [i] > 5 && usage < evacuation_threshold) {
1659 evacuate_block_obj_sizes [i] = TRUE;
1661 g_print ("slot size %d - %d of %d used\n",
1662 block_obj_sizes [i], slots_used [i], slots_available [i]);
1665 evacuate_block_obj_sizes [i] = FALSE;
1668 used_slots_size += sweep_slots_used [i] * block_obj_sizes [i];
1671 sgen_memgov_major_post_sweep (used_slots_size);
1673 set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1674 if (concurrent_sweep)
1675 binary_protocol_concurrent_sweep_end (sgen_timestamp ());
1681 set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1685 SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?");
1687 num_major_sections_before_sweep = num_major_sections;
1688 num_major_sections_freed_in_sweep = 0;
1690 SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
1691 if (concurrent_sweep) {
1692 sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
1693 sgen_thread_pool_job_enqueue (sweep_job);
1695 sweep_job_func (NULL, NULL);
1700 major_have_swept (void)
1702 return sweep_state == SWEEP_STATE_SWEPT;
1705 static int count_pinned_ref;
1706 static int count_pinned_nonref;
1707 static int count_nonpinned_ref;
1708 static int count_nonpinned_nonref;
1711 count_nonpinned_callback (GCObject *obj, size_t size, void *data)
1713 GCVTable vtable = SGEN_LOAD_VTABLE (obj);
1715 if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1716 ++count_nonpinned_ref;
1718 ++count_nonpinned_nonref;
1722 count_pinned_callback (GCObject *obj, size_t size, void *data)
1724 GCVTable vtable = SGEN_LOAD_VTABLE (obj);
1726 if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1729 ++count_pinned_nonref;
1732 static G_GNUC_UNUSED void
1733 count_ref_nonref_objs (void)
1737 count_pinned_ref = 0;
1738 count_pinned_nonref = 0;
1739 count_nonpinned_ref = 0;
1740 count_nonpinned_nonref = 0;
1742 major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1743 major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1745 total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1747 g_print ("ref: %d pinned %d non-pinned non-ref: %d pinned %d non-pinned -- %.1f\n",
1748 count_pinned_ref, count_nonpinned_ref,
1749 count_pinned_nonref, count_nonpinned_nonref,
1750 (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1754 ms_calculate_block_obj_sizes (double factor, int *arr)
1761 * Have every possible slot size starting with the minimal
1762 * object size up to and including four times that size. Then
1763 * proceed by increasing geometrically with the given factor.
1766 for (int size = SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size <= 4 * SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size += SGEN_ALLOC_ALIGN) {
1768 arr [num_sizes] = size;
1772 target_size = (double)last_size;
1775 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1776 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1778 if (size != last_size) {
1780 arr [num_sizes] = size;
1785 target_size *= factor;
1786 } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1791 /* only valid during minor collections */
1792 static mword old_num_major_sections;
1795 major_start_nursery_collection (void)
1797 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1798 consistency_check ();
1801 old_num_major_sections = num_major_sections;
1805 major_finish_nursery_collection (void)
1807 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1808 consistency_check ();
1813 block_usage_comparer (const void *bl1, const void *bl2)
1815 const gint16 nused1 = (*(MSBlockInfo**)bl1)->nused;
1816 const gint16 nused2 = (*(MSBlockInfo**)bl2)->nused;
1818 return nused2 - nused1;
1822 sgen_evacuation_freelist_blocks (MSBlockInfo * volatile *block_list, int size_index)
1824 MSBlockInfo **evacuated_blocks;
1825 size_t index = 0, count, num_blocks = 0, num_used = 0;
1827 MSBlockInfo * volatile *prev;
1829 for (info = *block_list; info != NULL; info = info->next_free) {
1831 num_used += info->nused;
1835 * We have a set of blocks in the freelist which will be evacuated. Instead
1836 * of evacuating all of the blocks into new ones, we traverse the freelist
1837 * sorting it by the number of occupied slots, evacuating the objects from
1838 * blocks with fewer used slots into fuller blocks.
1840 * The number of used slots is set at the end of the previous sweep. Since
1841 * we sequentially unlink slots from blocks, except for the head of the
1842 * freelist, for blocks on the freelist, the number of used slots is the same
1843 * as at the end of the previous sweep.
1845 evacuated_blocks = (MSBlockInfo**)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_blocks, INTERNAL_MEM_TEMPORARY, TRUE);
1847 for (info = *block_list; info != NULL; info = info->next_free) {
1848 evacuated_blocks [index++] = info;
1851 SGEN_ASSERT (0, num_blocks == index, "Why did the freelist change ?");
1853 sgen_qsort (evacuated_blocks, num_blocks, sizeof (gpointer), block_usage_comparer);
1856 * Form a new freelist with the fullest blocks. These blocks will also be
1857 * marked as to_space so we don't evacuate from them.
1859 count = MS_BLOCK_FREE / block_obj_sizes [size_index];
1861 for (index = 0; index < (num_used + count - 1) / count; index++) {
1862 SGEN_ASSERT (0, index < num_blocks, "Why do we need more blocks for compaction than we already had ?");
1863 info = evacuated_blocks [index];
1864 info->is_to_space = TRUE;
1866 prev = &info->next_free;
1870 sgen_free_internal_dynamic (evacuated_blocks, sizeof (MSBlockInfo*) * num_blocks, INTERNAL_MEM_TEMPORARY);
1874 major_start_major_collection (void)
1879 major_finish_sweep_checking ();
1882 * Clear the free lists for block sizes where we do evacuation. For those block
1883 * sizes we will have to allocate new blocks.
1885 for (i = 0; i < num_block_obj_sizes; ++i) {
1886 if (!evacuate_block_obj_sizes [i])
1889 binary_protocol_evacuating_blocks (block_obj_sizes [i]);
1891 sgen_evacuation_freelist_blocks (&free_block_lists [0][i], i);
1892 sgen_evacuation_freelist_blocks (&free_block_lists [MS_BLOCK_FLAG_REFS][i], i);
1895 if (lazy_sweep && concurrent_sweep) {
1897 * sweep_blocks_job is created before sweep_finish, which we wait for above
1898 * (major_finish_sweep_checking). After the end of sweep, if we don't have
1899 * sweep_blocks_job set, it means that it has already been run.
1901 SgenThreadPoolJob *job = sweep_blocks_job;
1903 sgen_thread_pool_job_wait (job);
1906 if (lazy_sweep && !concurrent_sweep)
1907 binary_protocol_sweep_begin (GENERATION_OLD, TRUE);
1908 /* Sweep all unswept blocks and set them to MARKING */
1909 FOREACH_BLOCK_NO_LOCK (block) {
1910 if (lazy_sweep && !concurrent_sweep)
1911 sweep_block (block);
1912 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1913 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1915 * Swept blocks that have a null free_list are full. Evacuation is not
1916 * effective on these blocks since we expect them to have high usage anyway,
1917 * given that the survival rate for majors is relatively high.
1919 if (evacuate_block_obj_sizes [block->obj_size_index] && !block->free_list)
1920 block->is_to_space = TRUE;
1921 } END_FOREACH_BLOCK_NO_LOCK;
1922 if (lazy_sweep && !concurrent_sweep)
1923 binary_protocol_sweep_end (GENERATION_OLD, TRUE);
1925 set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1929 major_finish_major_collection (ScannedObjectCounts *counts)
1931 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1932 if (binary_protocol_is_enabled ()) {
1933 counts->num_scanned_objects = scanned_objects_list.next_slot;
1935 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1936 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1938 sgen_pointer_queue_clear (&scanned_objects_list);
1943 #if SIZEOF_VOID_P != 8
1945 compare_pointers (const void *va, const void *vb) {
1946 char *a = *(char**)va, *b = *(char**)vb;
1956 * This is called with sweep completed and the world stopped.
1959 major_free_swept_blocks (size_t allowance)
1961 /* FIXME: This is probably too much. It's assuming all objects are small. */
1962 size_t section_reserve = allowance / MS_BLOCK_SIZE;
1964 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1968 * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1969 * a VirtualAlloc ()-ed block.
1974 #if SIZEOF_VOID_P != 8
1976 int i, num_empty_blocks_orig, num_blocks, arr_length;
1978 void **empty_block_arr;
1979 void **rebuild_next;
1981 if (num_empty_blocks <= section_reserve)
1983 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1985 num_empty_blocks_orig = num_empty_blocks;
1986 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1987 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1988 if (!empty_block_arr)
1992 for (block = empty_blocks; block; block = *(void**)block)
1993 empty_block_arr [i++] = block;
1994 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1996 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1999 * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
2000 * contiguous ones. If we do, we free them. If that's not enough to get to
2001 * section_reserve, we halve the number of contiguous blocks we're looking
2002 * for and have another go, until we're done with looking for pairs of
2003 * blocks, at which point we give up and go to the fallback.
2005 arr_length = num_empty_blocks_orig;
2006 num_blocks = MS_BLOCK_ALLOC_NUM;
2007 while (num_empty_blocks > section_reserve && num_blocks > 1) {
2012 for (i = 0; i < arr_length; ++i) {
2014 void *block = empty_block_arr [i];
2015 SGEN_ASSERT (6, block, "we're not shifting correctly");
2017 empty_block_arr [dest] = block;
2019 * This is not strictly necessary, but we're
2022 empty_block_arr [i] = NULL;
2031 SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
2033 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
2038 if (d + 1 - first == num_blocks) {
2040 * We found num_blocks contiguous blocks. Free them
2041 * and null their array entries. As an optimization
2042 * we could, instead of nulling the entries, shift
2043 * the following entries over to the left, while
2047 sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
2048 for (j = first; j <= d; ++j)
2049 empty_block_arr [j] = NULL;
2053 num_empty_blocks -= num_blocks;
2055 stat_major_blocks_freed += num_blocks;
2056 if (num_blocks == MS_BLOCK_ALLOC_NUM)
2057 stat_major_blocks_freed_ideal += num_blocks;
2059 stat_major_blocks_freed_less_ideal += num_blocks;
2064 SGEN_ASSERT (6, dest <= i && dest <= arr_length, "array length is off");
2066 SGEN_ASSERT (6, arr_length == num_empty_blocks, "array length is off");
2071 /* rebuild empty_blocks free list */
2072 rebuild_next = (void**)&empty_blocks;
2073 for (i = 0; i < arr_length; ++i) {
2074 void *block = empty_block_arr [i];
2075 SGEN_ASSERT (6, block, "we're missing blocks");
2076 *rebuild_next = block;
2077 rebuild_next = (void**)block;
2079 *rebuild_next = NULL;
2082 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
2085 SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
2089 * This is our threshold. If there's not more empty than used blocks, we won't
2090 * release uncontiguous blocks, in fear of fragmenting the address space.
2092 if (num_empty_blocks <= num_major_sections)
2096 while (num_empty_blocks > section_reserve) {
2097 void *next = *(void**)empty_blocks;
2098 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
2099 empty_blocks = next;
2101 * Needs not be atomic because this is running
2106 ++stat_major_blocks_freed;
2107 #if SIZEOF_VOID_P != 8
2108 ++stat_major_blocks_freed_individual;
2114 major_pin_objects (SgenGrayQueue *queue)
2118 FOREACH_BLOCK_NO_LOCK (block) {
2119 size_t first_entry, last_entry;
2120 SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
2121 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
2122 &first_entry, &last_entry);
2123 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
2124 } END_FOREACH_BLOCK_NO_LOCK;
2128 major_init_to_space (void)
2133 major_report_pinned_memory_usage (void)
2135 g_assert_not_reached ();
2139 major_get_used_size (void)
2145 * We're holding the GC lock, but the sweep thread might be running. Make sure it's
2146 * finished, then we can iterate over the block array.
2148 major_finish_sweep_checking ();
2150 FOREACH_BLOCK_NO_LOCK (block) {
2151 int count = MS_BLOCK_FREE / block->obj_size;
2153 size += count * block->obj_size;
2154 for (iter = block->free_list; iter; iter = (void**)*iter)
2155 size -= block->obj_size;
2156 } END_FOREACH_BLOCK_NO_LOCK;
2161 /* FIXME: return number of bytes, not of sections */
2163 get_num_major_sections (void)
2165 return num_major_sections;
2169 * Returns the number of bytes in blocks that were present when the last sweep was
2170 * initiated, and were not freed during the sweep. They are the basis for calculating the
2174 get_bytes_survived_last_sweep (void)
2176 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2177 return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
2181 major_handle_gc_param (const char *opt)
2183 if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2184 const char *arg = strchr (opt, '=') + 1;
2185 int percentage = atoi (arg);
2186 if (percentage < 0 || percentage > 100) {
2187 fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2190 evacuation_threshold = (float)percentage / 100.0f;
2192 } else if (!strcmp (opt, "lazy-sweep")) {
2195 } else if (!strcmp (opt, "no-lazy-sweep")) {
2198 } else if (!strcmp (opt, "concurrent-sweep")) {
2199 concurrent_sweep = TRUE;
2201 } else if (!strcmp (opt, "no-concurrent-sweep")) {
2202 concurrent_sweep = FALSE;
2210 major_print_gc_param_usage (void)
2214 " evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2215 " (no-)lazy-sweep\n"
2216 " (no-)concurrent-sweep\n"
2221 * This callback is used to clear cards, move cards to the shadow table and do counting.
2224 major_iterate_block_ranges (sgen_cardtable_block_callback callback)
2227 gboolean has_references;
2229 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2231 callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2232 } END_FOREACH_BLOCK_NO_LOCK;
2236 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
2239 gboolean has_references;
2241 major_finish_sweep_checking ();
2242 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2244 callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2245 } END_FOREACH_BLOCK_NO_LOCK;
2248 #ifdef HEAVY_STATISTICS
2249 extern guint64 marked_cards;
2250 extern guint64 scanned_cards;
2251 extern guint64 scanned_objects;
2252 extern guint64 remarked_cards;
2255 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2257 * MS blocks are 16K aligned.
2258 * Cardtables are 4K aligned, at least.
2259 * This means that the cardtable of a given block is 32 bytes aligned.
2262 initial_skip_card (guint8 *card_data)
2264 mword *cards = (mword*)card_data;
2267 for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2273 if (i == CARD_WORDS_PER_BLOCK)
2274 return card_data + CARDS_PER_BLOCK;
2276 #if defined(__i386__) && defined(__GNUC__)
2277 return card_data + i * 4 + (__builtin_ffs (card) - 1) / 8;
2278 #elif defined(__x86_64__) && defined(__GNUC__)
2279 return card_data + i * 8 + (__builtin_ffsll (card) - 1) / 8;
2280 #elif defined(__s390x__) && defined(__GNUC__)
2281 return card_data + i * 8 + (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2283 for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2285 return &card_data [i];
2291 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2292 #define MS_BLOCK_OBJ_FAST(b,os,i) ((b) + MS_BLOCK_SKIP + (os) * (i))
2293 #define MS_OBJ_ALLOCED_FAST(o,b) (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2296 scan_card_table_for_block (MSBlockInfo *block, CardTableScanType scan_type, ScanCopyContext ctx)
2298 SgenGrayQueue *queue = ctx.queue;
2299 ScanObjectFunc scan_func = ctx.ops->scan_object;
2300 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2301 guint8 cards_copy [CARDS_PER_BLOCK];
2303 guint8 cards_preclean [CARDS_PER_BLOCK];
2304 gboolean small_objects;
2307 guint8 *card_data, *card_base;
2308 guint8 *card_data_end;
2309 char *scan_front = NULL;
2311 /* The concurrent mark doesn't enter evacuating blocks */
2312 if (scan_type == CARDTABLE_SCAN_MOD_UNION_PRECLEAN && major_block_is_evacuating (block))
2315 block_obj_size = block->obj_size;
2316 small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2318 block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2321 * This is safe in face of card aliasing for the following reason:
2323 * Major blocks are 16k aligned, or 32 cards aligned.
2324 * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2325 * sizes, they won't overflow the cardtable overlap modulus.
2327 if (scan_type & CARDTABLE_SCAN_MOD_UNION) {
2328 card_data = card_base = block->cardtable_mod_union;
2330 * This happens when the nursery collection that precedes finishing
2331 * the concurrent collection allocates new major blocks.
2336 if (scan_type == CARDTABLE_SCAN_MOD_UNION_PRECLEAN) {
2337 sgen_card_table_preclean_mod_union (card_data, cards_preclean, CARDS_PER_BLOCK);
2338 card_data = card_base = cards_preclean;
2341 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2342 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2344 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2346 card_data = card_base = cards_copy;
2349 card_data_end = card_data + CARDS_PER_BLOCK;
2351 card_data += MS_BLOCK_SKIP >> CARD_BITS;
2353 card_data = initial_skip_card (card_data);
2354 while (card_data < card_data_end) {
2355 size_t card_index, first_object_index;
2358 char *first_obj, *obj;
2360 HEAVY_STAT (++scanned_cards);
2367 card_index = card_data - card_base;
2368 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2369 end = start + CARD_SIZE_IN_BYTES;
2371 if (!block_is_swept_or_marking (block))
2372 sweep_block (block);
2374 HEAVY_STAT (++marked_cards);
2377 sgen_card_table_prepare_card_for_scanning (card_data);
2380 * If the card we're looking at starts at or in the block header, we
2381 * must start at the first object in the block, without calculating
2382 * the index of the object we're hypothetically starting at, because
2383 * it would be negative.
2385 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2386 first_object_index = 0;
2388 first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2390 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2392 binary_protocol_card_scan (first_obj, end - first_obj);
2395 if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2398 if (scan_type & CARDTABLE_SCAN_MOD_UNION) {
2399 /* FIXME: do this more efficiently */
2401 MS_CALC_MARK_BIT (w, b, obj);
2402 if (!MS_MARK_BIT (block, w, b))
2406 GCObject *object = (GCObject*)obj;
2408 if (small_objects) {
2409 HEAVY_STAT (++scanned_objects);
2410 scan_func (object, sgen_obj_get_descriptor (object), queue);
2412 size_t offset = sgen_card_table_get_card_offset (obj, block_start);
2413 sgen_cardtable_scan_object (object, block_obj_size, card_base + offset, ctx);
2416 obj += block_obj_size;
2417 g_assert (scan_front <= obj);
2421 HEAVY_STAT (if (*card_data) ++remarked_cards);
2426 card_data = card_base + sgen_card_table_get_card_offset (obj, block_start);
2431 major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx)
2434 gboolean has_references, was_sweeping, skip_scan;
2436 if (!concurrent_mark)
2437 g_assert (scan_type == CARDTABLE_SCAN_GLOBAL);
2439 if (scan_type != CARDTABLE_SCAN_GLOBAL)
2440 SGEN_ASSERT (0, !sweep_in_progress (), "Sweep should be finished when we scan mod union card table");
2441 was_sweeping = sweep_in_progress ();
2443 binary_protocol_major_card_table_scan_start (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
2444 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2445 #ifdef PREFETCH_CARDS
2446 int prefetch_index = __index + 6;
2447 if (prefetch_index < allocated_blocks.next_slot) {
2448 MSBlockInfo *prefetch_block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, prefetch_index));
2449 PREFETCH_READ (prefetch_block);
2450 if (scan_type == CARDTABLE_SCAN_GLOBAL) {
2451 guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2452 PREFETCH_WRITE (prefetch_cards);
2453 PREFETCH_WRITE (prefetch_cards + 32);
2458 if (!has_references)
2462 if (scan_type == CARDTABLE_SCAN_GLOBAL) {
2463 gpointer *card_start = (gpointer*) sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (block));
2464 gboolean has_dirty_cards = FALSE;
2466 for (i = 0; i < CARDS_PER_BLOCK / sizeof(gpointer); i++) {
2467 if (card_start [i]) {
2468 has_dirty_cards = TRUE;
2472 if (!has_dirty_cards) {
2476 * After the start of the concurrent collections, blocks change state
2477 * to marking. We should not sweep it in that case. We can't race with
2478 * sweep start since we are in a nursery collection. Also avoid CAS-ing
2480 if (sweep_in_progress ()) {
2481 skip_scan = !ensure_block_is_checked_for_sweeping (__index, TRUE, NULL);
2482 } else if (was_sweeping) {
2483 /* Recheck in case sweep finished after dereferencing the slot */
2484 skip_scan = *sgen_array_list_get_slot (&allocated_blocks, __index) == 0;
2489 scan_card_table_for_block (block, scan_type, ctx);
2490 } END_FOREACH_BLOCK_NO_LOCK;
2491 binary_protocol_major_card_table_scan_end (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
2495 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2498 gboolean has_references;
2499 long long total_cards = 0;
2500 long long marked_cards = 0;
2502 if (sweep_in_progress ()) {
2503 *num_total_cards = -1;
2504 *num_marked_cards = -1;
2508 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2509 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2512 if (!has_references)
2515 total_cards += CARDS_PER_BLOCK;
2516 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2520 } END_FOREACH_BLOCK_NO_LOCK;
2522 *num_total_cards = total_cards;
2523 *num_marked_cards = marked_cards;
2527 update_cardtable_mod_union (void)
2531 FOREACH_BLOCK_NO_LOCK (block) {
2532 gpointer *card_start = (gpointer*) sgen_card_table_get_card_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (block));
2533 gboolean has_dirty_cards = FALSE;
2535 for (i = 0; i < CARDS_PER_BLOCK / sizeof(gpointer); i++) {
2536 if (card_start [i]) {
2537 has_dirty_cards = TRUE;
2541 if (has_dirty_cards) {
2543 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
2544 sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2545 SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2547 } END_FOREACH_BLOCK_NO_LOCK;
2550 #undef pthread_create
2553 post_param_init (SgenMajorCollector *collector)
2555 collector->sweeps_lazily = lazy_sweep;
2556 collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
2560 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
2564 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2566 num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2567 block_obj_sizes = (int *)sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2568 ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2570 evacuate_block_obj_sizes = (gboolean *)sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2571 for (i = 0; i < num_block_obj_sizes; ++i)
2572 evacuate_block_obj_sizes [i] = FALSE;
2574 sweep_slots_available = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2575 sweep_slots_used = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2576 sweep_num_blocks = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2581 g_print ("block object sizes:\n");
2582 for (i = 0; i < num_block_obj_sizes; ++i)
2583 g_print ("%d\n", block_obj_sizes [i]);
2587 for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2588 free_block_lists [i] = (MSBlockInfo *volatile *)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2590 for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2591 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2592 for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2593 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2595 mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2596 mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2597 mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2598 #if SIZEOF_VOID_P != 8
2599 mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2600 mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2601 mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2602 mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2605 collector->section_size = MAJOR_SECTION_SIZE;
2607 concurrent_mark = is_concurrent;
2608 collector->is_concurrent = is_concurrent;
2609 collector->needs_thread_pool = is_concurrent || concurrent_sweep;
2610 collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2611 collector->supports_cardtable = TRUE;
2613 collector->alloc_heap = major_alloc_heap;
2614 collector->is_object_live = major_is_object_live;
2615 collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2616 collector->alloc_degraded = major_alloc_degraded;
2618 collector->alloc_object = major_alloc_object;
2619 collector->free_pinned_object = free_pinned_object;
2620 collector->iterate_objects = major_iterate_objects;
2621 collector->free_non_pinned_object = major_free_non_pinned_object;
2622 collector->pin_objects = major_pin_objects;
2623 collector->pin_major_object = pin_major_object;
2624 collector->scan_card_table = major_scan_card_table;
2625 collector->iterate_live_block_ranges = major_iterate_live_block_ranges;
2626 collector->iterate_block_ranges = major_iterate_block_ranges;
2627 if (is_concurrent) {
2628 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2629 collector->get_cardtable_mod_union_for_reference = major_get_cardtable_mod_union_for_reference;
2631 collector->init_to_space = major_init_to_space;
2632 collector->sweep = major_sweep;
2633 collector->have_swept = major_have_swept;
2634 collector->finish_sweeping = major_finish_sweep_checking;
2635 collector->free_swept_blocks = major_free_swept_blocks;
2636 collector->check_scan_starts = major_check_scan_starts;
2637 collector->dump_heap = major_dump_heap;
2638 collector->get_used_size = major_get_used_size;
2639 collector->start_nursery_collection = major_start_nursery_collection;
2640 collector->finish_nursery_collection = major_finish_nursery_collection;
2641 collector->start_major_collection = major_start_major_collection;
2642 collector->finish_major_collection = major_finish_major_collection;
2643 collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2644 collector->ptr_is_from_pinned_alloc = ptr_is_from_pinned_alloc;
2645 collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2646 collector->get_num_major_sections = get_num_major_sections;
2647 collector->get_bytes_survived_last_sweep = get_bytes_survived_last_sweep;
2648 collector->handle_gc_param = major_handle_gc_param;
2649 collector->print_gc_param_usage = major_print_gc_param_usage;
2650 collector->post_param_init = post_param_init;
2651 collector->is_valid_object = major_is_valid_object;
2652 collector->describe_pointer = major_describe_pointer;
2653 collector->count_cards = major_count_cards;
2655 collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
2656 collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
2657 collector->major_ops_serial.drain_gray_stack = drain_gray_stack;
2658 if (is_concurrent) {
2659 collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2660 collector->major_ops_concurrent_start.scan_object = major_scan_object_concurrent_with_evacuation;
2661 collector->major_ops_concurrent_start.scan_vtype = major_scan_vtype_concurrent_with_evacuation;
2662 collector->major_ops_concurrent_start.scan_ptr_field = major_scan_ptr_field_concurrent_with_evacuation;
2663 collector->major_ops_concurrent_start.drain_gray_stack = drain_gray_stack_concurrent;
2665 collector->major_ops_concurrent_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_finish_canonical;
2666 collector->major_ops_concurrent_finish.scan_object = major_scan_object_with_evacuation;
2667 collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_with_evacuation;
2668 collector->major_ops_concurrent_finish.scan_ptr_field = major_scan_ptr_field_with_evacuation;
2669 collector->major_ops_concurrent_finish.drain_gray_stack = drain_gray_stack;
2672 #ifdef HEAVY_STATISTICS
2673 mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2674 mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2675 mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2676 mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2677 mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2678 mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2679 mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2680 mono_counters_register ("Optimized copy major small evacuate", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_evacuate);
2681 mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2682 mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2683 mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2685 mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2686 mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2687 mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2690 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2691 mono_os_mutex_init (&scanned_objects_list_lock);
2694 SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2696 /*cardtable requires major pages to be 8 cards aligned*/
2697 g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2701 sgen_marksweep_init (SgenMajorCollector *collector)
2703 sgen_marksweep_init_internal (collector, FALSE);
2707 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2709 sgen_marksweep_init_internal (collector, TRUE);