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, sgen_array_list_default_is_slot_set, sgen_array_list_default_cas_setter, 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 static guint64 stat_major_blocks_freed_ideal = 0;
244 static guint64 stat_major_blocks_freed_less_ideal = 0;
245 static guint64 stat_major_blocks_freed_individual = 0;
246 static guint64 stat_major_blocks_alloced_less_ideal = 0;
248 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
249 static guint64 num_major_objects_marked = 0;
250 #define INC_NUM_MAJOR_OBJECTS_MARKED() (++num_major_objects_marked)
252 #define INC_NUM_MAJOR_OBJECTS_MARKED()
255 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
256 static mono_mutex_t scanned_objects_list_lock;
257 static SgenPointerQueue scanned_objects_list;
260 add_scanned_object (void *ptr)
262 if (!binary_protocol_is_enabled ())
265 mono_os_mutex_lock (&scanned_objects_list_lock);
266 sgen_pointer_queue_add (&scanned_objects_list, ptr);
267 mono_os_mutex_unlock (&scanned_objects_list_lock);
271 static gboolean sweep_block (MSBlockInfo *block);
274 ms_find_block_obj_size_index (size_t size)
277 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);
278 for (i = 0; i < num_block_obj_sizes; ++i)
279 if (block_obj_sizes [i] >= size)
281 g_error ("no object of size %zd\n", size);
285 #define FREE_BLOCKS_FROM(lists,p,r) (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
286 #define FREE_BLOCKS(p,r) (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
288 #define MS_BLOCK_OBJ_SIZE_INDEX(s) \
289 (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ? \
290 fast_block_obj_size_indexes [((s)+7)>>3] : \
291 ms_find_block_obj_size_index ((s)))
294 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
298 start = (char *)sgen_alloc_os_memory_aligned (nursery_size, nursery_align, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery", MONO_MEM_ACCOUNT_SGEN_NURSERY);
300 start = (char *)sgen_alloc_os_memory (nursery_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery", MONO_MEM_ACCOUNT_SGEN_NURSERY);
306 update_heap_boundaries_for_block (MSBlockInfo *block)
308 sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
315 ms_get_empty_block (void)
319 void *block, *empty, *next;
324 * We try allocating MS_BLOCK_ALLOC_NUM blocks first. If that's
325 * unsuccessful, we halve the number of blocks and try again, until we're at
326 * 1. If that doesn't work, either, we assert.
328 int alloc_num = MS_BLOCK_ALLOC_NUM;
330 p = (char *)sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE,
331 (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE),
332 alloc_num == 1 ? "major heap section" : NULL, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
338 for (i = 0; i < alloc_num; ++i) {
341 * We do the free list update one after the
342 * other so that other threads can use the new
343 * blocks as quickly as possible.
346 empty = empty_blocks;
347 *(void**)block = empty;
348 } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
352 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
354 stat_major_blocks_alloced += alloc_num;
355 #if SIZEOF_VOID_P != 8
356 if (alloc_num != MS_BLOCK_ALLOC_NUM)
357 stat_major_blocks_alloced_less_ideal += alloc_num;
362 empty = empty_blocks;
366 next = *(void**)block;
367 } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
369 SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
371 *(void**)block = NULL;
373 g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
379 * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
380 * list, where it will either be freed later on, or reused in nursery collections.
383 ms_free_block (MSBlockInfo *info)
386 char *block = MS_BLOCK_FOR_BLOCK_INFO (info);
388 sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
389 if (info->cardtable_mod_union)
390 sgen_card_table_free_mod_union (info->cardtable_mod_union, block, MS_BLOCK_SIZE);
391 memset (block, 0, MS_BLOCK_SIZE);
394 empty = empty_blocks;
395 *(void**)block = empty;
396 } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
398 SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
400 binary_protocol_block_free (block, MS_BLOCK_SIZE);
404 sweep_in_progress (void)
406 int state = sweep_state;
407 return state == SWEEP_STATE_SWEEPING ||
408 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
409 state == SWEEP_STATE_COMPACTING;
412 static inline gboolean
413 block_is_swept_or_marking (MSBlockInfo *block)
415 gint32 state = block->state;
416 return state == BLOCK_STATE_SWEPT || state == BLOCK_STATE_MARKING;
419 //#define MARKSWEEP_CONSISTENCY_CHECK
421 #ifdef MARKSWEEP_CONSISTENCY_CHECK
423 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
425 SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
426 for (; block; block = block->next_free) {
427 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
428 g_assert (block->obj_size == size);
429 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
431 /* blocks in the free lists must have at least
433 g_assert (block->free_list);
435 /* the block must be in the allocated_blocks array */
436 g_assert (sgen_array_list_find (&allocated_blocks, BLOCK_TAG (block)) != (guint32)-1);
441 check_empty_blocks (void)
445 for (p = empty_blocks; p; p = *(void**)p)
447 g_assert (i == num_empty_blocks);
451 consistency_check (void)
456 /* check all blocks */
457 FOREACH_BLOCK_NO_LOCK (block) {
458 int count = MS_BLOCK_FREE / block->obj_size;
462 /* count number of free slots */
463 for (i = 0; i < count; ++i) {
464 void **obj = (void**) MS_BLOCK_OBJ (block, i);
465 if (!MS_OBJ_ALLOCED (obj, block))
469 /* check free list */
470 for (free = block->free_list; free; free = (void**)*free) {
471 g_assert (MS_BLOCK_FOR_OBJ (free) == block);
474 g_assert (num_free == 0);
476 /* check all mark words are zero */
477 if (!sgen_concurrent_collection_in_progress () && block_is_swept_or_marking (block)) {
478 for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
479 g_assert (block->mark_words [i] == 0);
481 } END_FOREACH_BLOCK_NO_LOCK;
483 /* check free blocks */
484 for (i = 0; i < num_block_obj_sizes; ++i) {
486 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
487 check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
490 check_empty_blocks ();
495 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
499 block->next_free = old = free_blocks [size_index];
500 } while (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], block, old) != old);
503 static void major_finish_sweep_checking (void);
506 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
508 int size = block_obj_sizes [size_index];
509 int count = MS_BLOCK_FREE / size;
511 MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
515 if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
518 info = (MSBlockInfo*)ms_get_empty_block ();
520 SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
522 info->obj_size = size;
523 info->obj_size_index = size_index;
524 info->pinned = pinned;
525 info->has_references = has_references;
526 info->has_pinned = pinned;
528 * Blocks that are to-space are not evacuated from. During an major collection
529 * blocks are allocated for two reasons: evacuating objects from the nursery and
530 * evacuating them from major blocks marked for evacuation. In both cases we don't
531 * want further evacuation. We also don't want to evacuate objects allocated during
532 * the concurrent mark since it would add pointless stress on the finishing pause.
534 info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD) || sgen_concurrent_collection_in_progress ();
535 info->state = info->is_to_space ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
536 SGEN_ASSERT (6, !sweep_in_progress () || info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
537 info->cardtable_mod_union = NULL;
539 update_heap_boundaries_for_block (info);
541 binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
543 /* build free list */
544 obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
545 info->free_list = (void**)obj_start;
546 /* we're skipping the last one - it must be nulled */
547 for (i = 0; i < count - 1; ++i) {
548 char *next_obj_start = obj_start + size;
549 *(void**)obj_start = next_obj_start;
550 obj_start = next_obj_start;
553 *(void**)obj_start = NULL;
555 add_free_block (free_blocks, size_index, info);
557 sgen_array_list_add (&allocated_blocks, BLOCK_TAG (info), 0, FALSE);
559 SGEN_ATOMIC_ADD_P (num_major_sections, 1);
564 ptr_is_in_major_block (char *ptr, char **start, gboolean *pinned)
568 FOREACH_BLOCK_NO_LOCK (block) {
569 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
570 int count = MS_BLOCK_FREE / block->obj_size;
575 for (i = 0; i <= count; ++i) {
576 if (ptr >= (char*)MS_BLOCK_OBJ (block, i) && ptr < (char*)MS_BLOCK_OBJ (block, i + 1)) {
578 *start = (char *)MS_BLOCK_OBJ (block, i);
583 *pinned = block->pinned;
586 } END_FOREACH_BLOCK_NO_LOCK;
591 ptr_is_from_pinned_alloc (char *ptr)
594 if (ptr_is_in_major_block (ptr, NULL, &pinned))
600 ensure_can_access_block_free_list (MSBlockInfo *block)
604 switch (block->state) {
605 case BLOCK_STATE_SWEPT:
606 case BLOCK_STATE_MARKING:
608 case BLOCK_STATE_CHECKING:
609 SGEN_ASSERT (0, FALSE, "How did we get a block that's being checked from a free list?");
611 case BLOCK_STATE_NEED_SWEEPING:
612 if (sweep_block (block))
613 ++stat_major_blocks_lazy_swept;
615 case BLOCK_STATE_SWEEPING:
616 /* FIXME: do this more elegantly */
620 SGEN_ASSERT (0, FALSE, "Illegal block state");
627 unlink_slot_from_free_list_uncontested (MSBlockInfo * volatile *free_blocks, int size_index)
629 MSBlockInfo *block, *next_free_block;
630 void *obj, *next_free_slot;
633 block = free_blocks [size_index];
634 SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
636 ensure_can_access_block_free_list (block);
638 obj = block->free_list;
639 SGEN_ASSERT (6, obj, "block %p in free list had no available object to alloc from", block);
641 next_free_slot = *(void**)obj;
642 if (next_free_slot) {
643 block->free_list = (gpointer *)next_free_slot;
647 next_free_block = block->next_free;
648 if (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], next_free_block, block) != block)
651 block->free_list = NULL;
652 block->next_free = NULL;
658 alloc_obj (GCVTable vtable, size_t size, gboolean pinned, gboolean has_references)
660 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
661 MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
664 if (!free_blocks [size_index]) {
665 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
669 obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
671 /* FIXME: assumes object layout */
672 *(GCVTable*)obj = vtable;
674 total_allocated_major += block_obj_sizes [size_index];
676 return (GCObject *)obj;
680 major_alloc_object (GCVTable vtable, size_t size, gboolean has_references)
682 return alloc_obj (vtable, size, FALSE, has_references);
686 * We're not freeing the block if it's empty. We leave that work for
687 * the next major collection.
689 * This is just called from the domain clearing code, which runs in a
690 * single thread and has the GC lock, so we don't need an extra lock.
693 free_object (GCObject *obj, size_t size, gboolean pinned)
695 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
697 gboolean in_free_list;
699 SGEN_ASSERT (9, sweep_state == SWEEP_STATE_SWEPT, "Should have waited for sweep to free objects.");
701 ensure_can_access_block_free_list (block);
702 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);
703 SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
704 MS_CALC_MARK_BIT (word, bit, obj);
705 SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set", obj);
707 memset (obj, 0, size);
709 in_free_list = !!block->free_list;
710 *(void**)obj = block->free_list;
711 block->free_list = (void**)obj;
714 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (pinned, block->has_references);
715 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
716 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);
717 add_free_block (free_blocks, size_index, block);
722 major_free_non_pinned_object (GCObject *obj, size_t size)
724 free_object (obj, size, FALSE);
727 /* size is a multiple of SGEN_ALLOC_ALIGN */
729 major_alloc_small_pinned_obj (GCVTable vtable, size_t size, gboolean has_references)
733 res = alloc_obj (vtable, size, TRUE, has_references);
734 /*If we failed to alloc memory, we better try releasing memory
735 *as pinned alloc is requested by the runtime.
738 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE, TRUE);
739 res = alloc_obj (vtable, size, TRUE, has_references);
741 return (GCObject *)res;
745 free_pinned_object (GCObject *obj, size_t size)
747 free_object (obj, size, TRUE);
751 * size is already rounded up and we hold the GC lock.
754 major_alloc_degraded (GCVTable vtable, size_t size)
758 obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
759 if (G_LIKELY (obj)) {
760 HEAVY_STAT (++stat_objects_alloced_degraded);
761 HEAVY_STAT (stat_bytes_alloced_degraded += size);
767 * obj is some object. If it's not in the major heap (i.e. if it's in
768 * the nursery or LOS), return FALSE. Otherwise return whether it's
769 * been marked or copied.
772 major_is_object_live (GCObject *obj)
778 if (sgen_ptr_in_nursery (obj))
781 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
784 if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
787 /* now we know it's in a major block */
788 block = MS_BLOCK_FOR_OBJ (obj);
789 SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?", block);
790 MS_CALC_MARK_BIT (word, bit, obj);
791 return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
795 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
798 if (ptr_is_in_major_block (ptr, start, &pinned))
804 try_set_sweep_state (int new_, int expected)
806 int old = SGEN_CAS (&sweep_state, new_, expected);
807 return old == expected;
811 set_sweep_state (int new_, int expected)
813 gboolean success = try_set_sweep_state (new_, expected);
814 SGEN_ASSERT (0, success, "Could not set sweep state.");
817 static gboolean ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked);
819 static SgenThreadPoolJob * volatile sweep_job;
820 static SgenThreadPoolJob * volatile sweep_blocks_job;
823 major_finish_sweep_checking (void)
826 SgenThreadPoolJob *job;
829 switch (sweep_state) {
830 case SWEEP_STATE_SWEPT:
831 case SWEEP_STATE_NEED_SWEEPING:
833 case SWEEP_STATE_SWEEPING:
834 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
837 case SWEEP_STATE_SWEEPING_AND_ITERATING:
838 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
840 case SWEEP_STATE_COMPACTING:
843 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
848 * We're running with the world stopped and the only other thread doing work is the
849 * sweep thread, which doesn't add blocks to the array, so we can safely access
852 for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index)
853 ensure_block_is_checked_for_sweeping (block_index, FALSE, NULL);
855 set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
860 sgen_thread_pool_job_wait (job);
861 SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
862 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
866 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
868 gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
869 gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
870 gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
873 /* No actual sweeping will take place if we are in the middle of a major collection. */
874 major_finish_sweep_checking ();
875 FOREACH_BLOCK_NO_LOCK (block) {
876 int count = MS_BLOCK_FREE / block->obj_size;
879 if (block->pinned && !pinned)
881 if (!block->pinned && !non_pinned)
883 if (sweep && lazy_sweep && !block_is_swept_or_marking (block)) {
885 SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
888 for (i = 0; i < count; ++i) {
889 void **obj = (void**) MS_BLOCK_OBJ (block, i);
890 if (MS_OBJ_ALLOCED (obj, block))
891 callback ((GCObject*)obj, block->obj_size, data);
893 } END_FOREACH_BLOCK_NO_LOCK;
897 major_is_valid_object (char *object)
901 FOREACH_BLOCK_NO_LOCK (block) {
905 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
908 idx = MS_BLOCK_OBJ_INDEX (object, block);
909 obj = (char*)MS_BLOCK_OBJ (block, idx);
912 return MS_OBJ_ALLOCED (obj, block);
913 } END_FOREACH_BLOCK_NO_LOCK;
920 major_describe_pointer (char *ptr)
924 FOREACH_BLOCK_NO_LOCK (block) {
932 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
935 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
936 MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
938 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
939 obj = (char*)MS_BLOCK_OBJ (block, idx);
940 live = MS_OBJ_ALLOCED (obj, block);
941 vtable = live ? SGEN_LOAD_VTABLE ((GCObject*)obj) : NULL;
943 MS_CALC_MARK_BIT (w, b, obj);
944 marked = MS_MARK_BIT (block, w, b);
949 SGEN_LOG (0, "object");
951 SGEN_LOG (0, "dead-object");
954 SGEN_LOG (0, "interior-ptr offset %zd", ptr - obj);
956 SGEN_LOG (0, "dead-interior-ptr offset %zd", ptr - obj);
959 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
962 } END_FOREACH_BLOCK_NO_LOCK;
968 major_check_scan_starts (void)
973 major_dump_heap (FILE *heap_dump_file)
976 int *slots_available = (int *)alloca (sizeof (int) * num_block_obj_sizes);
977 int *slots_used = (int *)alloca (sizeof (int) * num_block_obj_sizes);
980 for (i = 0; i < num_block_obj_sizes; ++i)
981 slots_available [i] = slots_used [i] = 0;
983 FOREACH_BLOCK_NO_LOCK (block) {
984 int index = ms_find_block_obj_size_index (block->obj_size);
985 int count = MS_BLOCK_FREE / block->obj_size;
987 slots_available [index] += count;
988 for (i = 0; i < count; ++i) {
989 if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
990 ++slots_used [index];
992 } END_FOREACH_BLOCK_NO_LOCK;
994 fprintf (heap_dump_file, "<occupancies>\n");
995 for (i = 0; i < num_block_obj_sizes; ++i) {
996 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
997 block_obj_sizes [i], slots_available [i], slots_used [i]);
999 fprintf (heap_dump_file, "</occupancies>\n");
1001 FOREACH_BLOCK_NO_LOCK (block) {
1002 int count = MS_BLOCK_FREE / block->obj_size;
1006 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
1008 for (i = 0; i <= count; ++i) {
1009 if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
1014 sgen_dump_occupied ((char *)MS_BLOCK_OBJ (block, start), (char *)MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
1020 fprintf (heap_dump_file, "</section>\n");
1021 } END_FOREACH_BLOCK_NO_LOCK;
1025 get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate)
1027 guint8 *mod_union = block->cardtable_mod_union;
1033 mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1034 other = (guint8 *)SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
1036 SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
1039 sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1043 static inline guint8*
1044 major_get_cardtable_mod_union_for_reference (char *ptr)
1046 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (ptr);
1047 size_t offset = sgen_card_table_get_card_offset (ptr, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1048 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
1049 SGEN_ASSERT (0, mod_union, "FIXME: optionally allocate the mod union if it's not here and CAS it in.");
1050 return &mod_union [offset];
1054 * Mark the mod-union card for `ptr`, which must be a reference within the object `obj`.
1057 mark_mod_union_card (GCObject *obj, void **ptr, GCObject *value_obj)
1059 int type = sgen_obj_get_descriptor (obj) & DESC_TYPE_MASK;
1060 if (sgen_safe_object_is_small (obj, type)) {
1061 guint8 *card_byte = major_get_cardtable_mod_union_for_reference ((char*)ptr);
1062 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?");
1065 sgen_los_mark_mod_union_card (obj, ptr);
1067 binary_protocol_mod_union_remset (obj, ptr, value_obj, SGEN_LOAD_VTABLE (value_obj));
1070 static inline gboolean
1071 major_block_is_evacuating (MSBlockInfo *block)
1073 if (evacuate_block_obj_sizes [block->obj_size_index] &&
1074 !block->has_pinned &&
1075 !block->is_to_space)
1080 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do { \
1081 int __word, __bit; \
1082 MS_CALC_MARK_BIT (__word, __bit, (obj)); \
1083 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
1084 if (!MS_MARK_BIT ((block), __word, __bit)) { \
1085 MS_SET_MARK_BIT ((block), __word, __bit); \
1086 if (sgen_gc_descr_has_references (desc)) \
1087 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1088 binary_protocol_mark ((obj), (gpointer)SGEN_LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1089 INC_NUM_MAJOR_OBJECTS_MARKED (); \
1094 pin_major_object (GCObject *obj, SgenGrayQueue *queue)
1098 if (concurrent_mark)
1099 g_assert_not_reached ();
1101 block = MS_BLOCK_FOR_OBJ (obj);
1102 block->has_pinned = TRUE;
1103 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1106 #include "sgen-major-copy-object.h"
1109 major_get_and_reset_num_major_objects_marked (void)
1111 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1112 long long num = num_major_objects_marked;
1113 num_major_objects_marked = 0;
1120 #define PREFETCH_CARDS 1 /* BOOL FASTENABLE */
1122 #undef PREFETCH_CARDS
1125 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1126 #if defined(PLATFORM_MACOSX)
1127 #if MONO_GNUC_VERSION <= 40300
1128 #undef PREFETCH_CARDS
1132 #ifdef HEAVY_STATISTICS
1133 static guint64 stat_optimized_copy;
1134 static guint64 stat_optimized_copy_nursery;
1135 static guint64 stat_optimized_copy_nursery_forwarded;
1136 static guint64 stat_optimized_copy_nursery_pinned;
1137 static guint64 stat_optimized_copy_major;
1138 static guint64 stat_optimized_copy_major_small_fast;
1139 static guint64 stat_optimized_copy_major_small_slow;
1140 static guint64 stat_optimized_copy_major_large;
1141 static guint64 stat_optimized_copy_major_forwarded;
1142 static guint64 stat_optimized_copy_major_small_evacuate;
1143 static guint64 stat_optimized_major_scan;
1144 static guint64 stat_optimized_major_scan_no_refs;
1146 static guint64 stat_drain_prefetch_fills;
1147 static guint64 stat_drain_prefetch_fill_failures;
1148 static guint64 stat_drain_loops;
1151 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_no_evacuation
1152 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_no_evacuation
1153 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_no_evacuation
1154 #include "sgen-marksweep-drain-gray-stack.h"
1156 #define COPY_OR_MARK_WITH_EVACUATION
1157 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_with_evacuation
1158 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_with_evacuation
1159 #define SCAN_VTYPE_FUNCTION_NAME major_scan_vtype_with_evacuation
1160 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_with_evacuation
1161 #define SCAN_PTR_FIELD_FUNCTION_NAME major_scan_ptr_field_with_evacuation
1162 #include "sgen-marksweep-drain-gray-stack.h"
1164 #define COPY_OR_MARK_CONCURRENT
1165 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_concurrent_no_evacuation
1166 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_concurrent_no_evacuation
1167 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_concurrent_no_evacuation
1168 #include "sgen-marksweep-drain-gray-stack.h"
1170 #define COPY_OR_MARK_CONCURRENT_WITH_EVACUATION
1171 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_concurrent_with_evacuation
1172 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_concurrent_with_evacuation
1173 #define SCAN_VTYPE_FUNCTION_NAME major_scan_vtype_concurrent_with_evacuation
1174 #define SCAN_PTR_FIELD_FUNCTION_NAME major_scan_ptr_field_concurrent_with_evacuation
1175 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_concurrent_with_evacuation
1176 #include "sgen-marksweep-drain-gray-stack.h"
1178 static inline gboolean
1179 major_is_evacuating (void)
1182 for (i = 0; i < num_block_obj_sizes; ++i) {
1183 if (evacuate_block_obj_sizes [i]) {
1192 drain_gray_stack (SgenGrayQueue *queue)
1194 if (major_is_evacuating ())
1195 return drain_gray_stack_with_evacuation (queue);
1197 return drain_gray_stack_no_evacuation (queue);
1201 drain_gray_stack_concurrent (SgenGrayQueue *queue)
1203 if (major_is_evacuating ())
1204 return drain_gray_stack_concurrent_with_evacuation (queue);
1206 return drain_gray_stack_concurrent_no_evacuation (queue);
1210 major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue)
1212 major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1216 major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *queue)
1218 major_copy_or_mark_object_concurrent_with_evacuation (ptr, *ptr, queue);
1222 major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue)
1224 major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1228 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1230 void **entry, **end;
1231 int last_index = -1;
1233 if (first_entry == last_entry)
1236 entry = sgen_pinning_get_entry (first_entry);
1237 end = sgen_pinning_get_entry (last_entry);
1239 for (; entry < end; ++entry) {
1240 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1242 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));
1243 if (index == last_index)
1245 obj = MS_BLOCK_OBJ (block, index);
1246 if (!MS_OBJ_ALLOCED (obj, block))
1248 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1249 sgen_pin_stats_register_object (obj, GENERATION_OLD);
1254 * There might have been potential pinning "pointers" into this block, but none of
1255 * them pointed to occupied slots, in which case we don't have to pin the block.
1257 if (last_index >= 0)
1258 block->has_pinned = TRUE;
1262 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1266 for (obj_index = 0; obj_index < count; ++obj_index) {
1268 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1270 MS_CALC_MARK_BIT (word, bit, obj);
1271 if (MS_MARK_BIT (block, word, bit)) {
1272 SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1274 /* an unmarked object */
1275 if (MS_OBJ_ALLOCED (obj, block)) {
1277 * FIXME: Merge consecutive
1278 * slots for lower reporting
1279 * overhead. Maybe memset
1280 * will also benefit?
1282 binary_protocol_empty (obj, obj_size);
1283 memset (obj, 0, obj_size);
1285 *(void**)obj = block->free_list;
1286 block->free_list = (void **)obj;
1291 static inline gboolean
1292 try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1294 gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
1295 gboolean success = old_state == expected_state;
1297 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
1302 set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1304 SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set");
1305 block->state = new_state;
1306 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, expected_state, new_state);
1310 * If `block` needs sweeping, sweep it and return TRUE. Otherwise return FALSE.
1312 * Sweeping means iterating through the block's slots and building the free-list from the
1313 * unmarked ones. They will also be zeroed. The mark bits will be reset.
1316 sweep_block (MSBlockInfo *block)
1319 void *reversed = NULL;
1322 switch (block->state) {
1323 case BLOCK_STATE_SWEPT:
1325 case BLOCK_STATE_MARKING:
1326 case BLOCK_STATE_CHECKING:
1327 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1329 case BLOCK_STATE_SWEEPING:
1330 /* FIXME: Do this more elegantly */
1333 case BLOCK_STATE_NEED_SWEEPING:
1334 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1338 SGEN_ASSERT (0, FALSE, "Illegal block state");
1341 SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1343 count = MS_BLOCK_FREE / block->obj_size;
1345 block->free_list = NULL;
1347 /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1348 // FIXME: Add more sizes
1349 switch (block->obj_size) {
1351 sweep_block_for_size (block, count, 16);
1354 sweep_block_for_size (block, count, block->obj_size);
1358 /* reset mark bits */
1359 memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1361 /* Reverse free list so that it's in address order */
1363 while (block->free_list) {
1364 void *next = *(void**)block->free_list;
1365 *(void**)block->free_list = reversed;
1366 reversed = block->free_list;
1367 block->free_list = (void **)next;
1369 block->free_list = (void **)reversed;
1371 mono_memory_write_barrier ();
1373 set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1384 if (sizeof (mword) == 8)
1385 count += __builtin_popcountll (d);
1387 count += __builtin_popcount (d);
1397 /* statistics for evacuation */
1398 static size_t *sweep_slots_available;
1399 static size_t *sweep_slots_used;
1400 static size_t *sweep_num_blocks;
1402 static volatile size_t num_major_sections_before_sweep;
1403 static volatile size_t num_major_sections_freed_in_sweep;
1410 for (i = 0; i < num_block_obj_sizes; ++i)
1411 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1413 /* clear all the free lists */
1414 for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1415 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1417 for (j = 0; j < num_block_obj_sizes; ++j)
1418 free_blocks [j] = NULL;
1422 static void sweep_finish (void);
1425 * If `wait` is TRUE and the block is currently being checked, this function will wait until
1426 * the checking has finished.
1428 * Returns whether the block is still there. If `wait` is FALSE, the return value will not
1429 * be correct, i.e. must not be used.
1432 ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked)
1435 gboolean have_live = FALSE;
1436 gboolean have_free = FALSE;
1442 volatile gpointer *block_slot = sgen_array_list_get_slot (&allocated_blocks, block_index);
1444 SGEN_ASSERT (6, sweep_in_progress (), "Why do we call this function if there's no sweep in progress?");
1447 *have_checked = FALSE;
1450 tagged_block = *(void * volatile *)block_slot;
1454 if (BLOCK_IS_TAGGED_CHECKING (tagged_block)) {
1457 /* FIXME: do this more elegantly */
1462 if (SGEN_CAS_PTR (block_slot, BLOCK_TAG_CHECKING (tagged_block), tagged_block) != tagged_block)
1465 block = BLOCK_UNTAG (tagged_block);
1466 block_state = block->state;
1468 if (!sweep_in_progress ()) {
1469 SGEN_ASSERT (6, block_state != BLOCK_STATE_SWEEPING && block_state != BLOCK_STATE_CHECKING, "Invalid block state.");
1471 SGEN_ASSERT (6, block_state != BLOCK_STATE_NEED_SWEEPING, "Invalid block state.");
1474 switch (block_state) {
1475 case BLOCK_STATE_SWEPT:
1476 case BLOCK_STATE_NEED_SWEEPING:
1477 case BLOCK_STATE_SWEEPING:
1479 case BLOCK_STATE_MARKING:
1481 case BLOCK_STATE_CHECKING:
1482 SGEN_ASSERT (0, FALSE, "We set the CHECKING bit - how can the stage be CHECKING?");
1485 SGEN_ASSERT (0, FALSE, "Illegal block state");
1489 SGEN_ASSERT (6, block->state == BLOCK_STATE_MARKING, "When we sweep all blocks must start out marking.");
1490 set_block_state (block, BLOCK_STATE_CHECKING, BLOCK_STATE_MARKING);
1493 *have_checked = TRUE;
1495 block->has_pinned = block->pinned;
1497 block->is_to_space = FALSE;
1499 count = MS_BLOCK_FREE / block->obj_size;
1501 if (block->cardtable_mod_union)
1502 memset (block->cardtable_mod_union, 0, CARDS_PER_BLOCK);
1504 /* Count marked objects in the block */
1505 for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1506 nused += bitcount (block->mark_words [i]);
1508 block->nused = nused;
1515 int obj_size_index = block->obj_size_index;
1516 gboolean has_pinned = block->has_pinned;
1518 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1521 * FIXME: Go straight to SWEPT if there are no free slots. We need
1522 * to set the free slot list to NULL, though, and maybe update some
1526 sweep_block (block);
1529 ++sweep_num_blocks [obj_size_index];
1530 sweep_slots_used [obj_size_index] += nused;
1531 sweep_slots_available [obj_size_index] += count;
1535 * If there are free slots in the block, add
1536 * the block to the corresponding free list.
1539 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1542 SGEN_ASSERT (6, block->free_list, "How do we not have a free list when there are free slots?");
1544 add_free_block (free_blocks, obj_size_index, block);
1547 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1548 update_heap_boundaries_for_block (block);
1551 * Blocks without live objects are removed from the
1552 * block list and freed.
1554 SGEN_ASSERT (6, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1555 SGEN_ASSERT (6, *block_slot == BLOCK_TAG_CHECKING (tagged_block), "How did the block move?");
1557 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1558 ms_free_block (block);
1560 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1561 SGEN_ATOMIC_ADD_P (num_major_sections_freed_in_sweep, 1);
1563 tagged_block = NULL;
1568 * Once the block is written back without the checking bit other threads are
1569 * free to access it. Make sure the block state is visible before we write it
1572 mono_memory_write_barrier ();
1573 *block_slot = tagged_block;
1574 return !!tagged_block;
1578 sweep_blocks_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1580 volatile gpointer *slot;
1583 SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) {
1584 bl = BLOCK_UNTAG (*slot);
1587 } SGEN_ARRAY_LIST_END_FOREACH_SLOT;
1589 mono_memory_write_barrier ();
1591 sweep_blocks_job = NULL;
1595 sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1597 guint32 block_index;
1598 guint32 num_blocks = num_major_sections_before_sweep;
1600 SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1601 SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1604 * We traverse the block array from high to low. Nursery collections will have to
1605 * cooperate with the sweep thread to finish sweeping, and they will traverse from
1606 * low to high, to avoid constantly colliding on the same blocks.
1608 for (block_index = allocated_blocks.next_slot; block_index-- > 0;) {
1609 ensure_block_is_checked_for_sweeping (block_index, TRUE, NULL);
1612 while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) {
1614 * The main GC thread is currently iterating over the block array to help us
1615 * finish the sweep. We have already finished, but we don't want to mess up
1616 * that iteration, so we just wait for it.
1621 if (SGEN_MAX_ASSERT_LEVEL >= 6) {
1622 for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1623 MSBlockInfo *block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, block_index));
1624 SGEN_ASSERT (6, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1629 * Concurrently sweep all the blocks to reduce workload during minor
1630 * pauses where we need certain blocks to be swept. At the start of
1631 * the next major we need all blocks to be swept anyway.
1633 if (concurrent_sweep && lazy_sweep) {
1634 sweep_blocks_job = sgen_thread_pool_job_alloc ("sweep_blocks", sweep_blocks_job_func, sizeof (SgenThreadPoolJob));
1635 sgen_thread_pool_job_enqueue (sweep_blocks_job);
1646 mword used_slots_size = 0;
1649 for (i = 0; i < num_block_obj_sizes; ++i) {
1650 float usage = (float)sweep_slots_used [i] / (float)sweep_slots_available [i];
1651 if (sweep_num_blocks [i] > 5 && usage < evacuation_threshold) {
1652 evacuate_block_obj_sizes [i] = TRUE;
1654 g_print ("slot size %d - %d of %d used\n",
1655 block_obj_sizes [i], slots_used [i], slots_available [i]);
1658 evacuate_block_obj_sizes [i] = FALSE;
1661 used_slots_size += sweep_slots_used [i] * block_obj_sizes [i];
1664 sgen_memgov_major_post_sweep (used_slots_size);
1666 set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1667 if (concurrent_sweep)
1668 binary_protocol_concurrent_sweep_end (sgen_timestamp ());
1674 set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1678 num_major_sections_before_sweep = num_major_sections;
1679 num_major_sections_freed_in_sweep = 0;
1681 SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
1682 if (concurrent_sweep) {
1683 sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
1684 sgen_thread_pool_job_enqueue (sweep_job);
1686 sweep_job_func (NULL, NULL);
1691 major_have_swept (void)
1693 return sweep_state == SWEEP_STATE_SWEPT;
1696 static int count_pinned_ref;
1697 static int count_pinned_nonref;
1698 static int count_nonpinned_ref;
1699 static int count_nonpinned_nonref;
1702 count_nonpinned_callback (GCObject *obj, size_t size, void *data)
1704 GCVTable vtable = SGEN_LOAD_VTABLE (obj);
1706 if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1707 ++count_nonpinned_ref;
1709 ++count_nonpinned_nonref;
1713 count_pinned_callback (GCObject *obj, size_t size, void *data)
1715 GCVTable vtable = SGEN_LOAD_VTABLE (obj);
1717 if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1720 ++count_pinned_nonref;
1723 static G_GNUC_UNUSED void
1724 count_ref_nonref_objs (void)
1728 count_pinned_ref = 0;
1729 count_pinned_nonref = 0;
1730 count_nonpinned_ref = 0;
1731 count_nonpinned_nonref = 0;
1733 major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1734 major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1736 total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1738 g_print ("ref: %d pinned %d non-pinned non-ref: %d pinned %d non-pinned -- %.1f\n",
1739 count_pinned_ref, count_nonpinned_ref,
1740 count_pinned_nonref, count_nonpinned_nonref,
1741 (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1745 ms_calculate_block_obj_sizes (double factor, int *arr)
1752 * Have every possible slot size starting with the minimal
1753 * object size up to and including four times that size. Then
1754 * proceed by increasing geometrically with the given factor.
1757 for (int size = SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size <= 4 * SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size += SGEN_ALLOC_ALIGN) {
1759 arr [num_sizes] = size;
1763 target_size = (double)last_size;
1766 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1767 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1769 if (size != last_size) {
1771 arr [num_sizes] = size;
1776 target_size *= factor;
1777 } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1782 /* only valid during minor collections */
1783 static mword old_num_major_sections;
1786 major_start_nursery_collection (void)
1788 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1789 consistency_check ();
1792 old_num_major_sections = num_major_sections;
1796 major_finish_nursery_collection (void)
1798 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1799 consistency_check ();
1804 block_usage_comparer (const void *bl1, const void *bl2)
1806 const gint16 nused1 = (*(MSBlockInfo**)bl1)->nused;
1807 const gint16 nused2 = (*(MSBlockInfo**)bl2)->nused;
1809 return nused2 - nused1;
1813 sgen_evacuation_freelist_blocks (MSBlockInfo * volatile *block_list, int size_index)
1815 MSBlockInfo **evacuated_blocks;
1816 size_t index = 0, count, num_blocks = 0, num_used = 0;
1818 MSBlockInfo * volatile *prev;
1820 for (info = *block_list; info != NULL; info = info->next_free) {
1822 num_used += info->nused;
1826 * We have a set of blocks in the freelist which will be evacuated. Instead
1827 * of evacuating all of the blocks into new ones, we traverse the freelist
1828 * sorting it by the number of occupied slots, evacuating the objects from
1829 * blocks with fewer used slots into fuller blocks.
1831 * The number of used slots is set at the end of the previous sweep. Since
1832 * we sequentially unlink slots from blocks, except for the head of the
1833 * freelist, for blocks on the freelist, the number of used slots is the same
1834 * as at the end of the previous sweep.
1836 evacuated_blocks = (MSBlockInfo**)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_blocks, INTERNAL_MEM_TEMPORARY, TRUE);
1838 for (info = *block_list; info != NULL; info = info->next_free) {
1839 evacuated_blocks [index++] = info;
1842 SGEN_ASSERT (0, num_blocks == index, "Why did the freelist change ?");
1844 sgen_qsort (evacuated_blocks, num_blocks, sizeof (gpointer), block_usage_comparer);
1847 * Form a new freelist with the fullest blocks. These blocks will also be
1848 * marked as to_space so we don't evacuate from them.
1850 count = MS_BLOCK_FREE / block_obj_sizes [size_index];
1852 for (index = 0; index < (num_used + count - 1) / count; index++) {
1853 SGEN_ASSERT (0, index < num_blocks, "Why do we need more blocks for compaction than we already had ?");
1854 info = evacuated_blocks [index];
1855 info->is_to_space = TRUE;
1857 prev = &info->next_free;
1861 sgen_free_internal_dynamic (evacuated_blocks, sizeof (MSBlockInfo*) * num_blocks, INTERNAL_MEM_TEMPORARY);
1865 major_start_major_collection (void)
1870 major_finish_sweep_checking ();
1873 * Clear the free lists for block sizes where we do evacuation. For those block
1874 * sizes we will have to allocate new blocks.
1876 for (i = 0; i < num_block_obj_sizes; ++i) {
1877 if (!evacuate_block_obj_sizes [i])
1880 binary_protocol_evacuating_blocks (block_obj_sizes [i]);
1882 sgen_evacuation_freelist_blocks (&free_block_lists [0][i], i);
1883 sgen_evacuation_freelist_blocks (&free_block_lists [MS_BLOCK_FLAG_REFS][i], i);
1886 if (lazy_sweep && concurrent_sweep) {
1888 * sweep_blocks_job is created before sweep_finish, which we wait for above
1889 * (major_finish_sweep_checking). After the end of sweep, if we don't have
1890 * sweep_blocks_job set, it means that it has already been run.
1892 SgenThreadPoolJob *job = sweep_blocks_job;
1894 sgen_thread_pool_job_wait (job);
1897 if (lazy_sweep && !concurrent_sweep)
1898 binary_protocol_sweep_begin (GENERATION_OLD, TRUE);
1899 /* Sweep all unswept blocks and set them to MARKING */
1900 FOREACH_BLOCK_NO_LOCK (block) {
1901 if (lazy_sweep && !concurrent_sweep)
1902 sweep_block (block);
1903 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1904 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1906 * Swept blocks that have a null free_list are full. Evacuation is not
1907 * effective on these blocks since we expect them to have high usage anyway,
1908 * given that the survival rate for majors is relatively high.
1910 if (evacuate_block_obj_sizes [block->obj_size_index] && !block->free_list)
1911 block->is_to_space = TRUE;
1912 } END_FOREACH_BLOCK_NO_LOCK;
1913 if (lazy_sweep && !concurrent_sweep)
1914 binary_protocol_sweep_end (GENERATION_OLD, TRUE);
1916 set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1920 major_finish_major_collection (ScannedObjectCounts *counts)
1922 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1923 if (binary_protocol_is_enabled ()) {
1924 counts->num_scanned_objects = scanned_objects_list.next_slot;
1926 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1927 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1929 sgen_pointer_queue_clear (&scanned_objects_list);
1935 compare_pointers (const void *va, const void *vb) {
1936 char *a = *(char**)va, *b = *(char**)vb;
1945 * This is called with sweep completed and the world stopped.
1948 major_free_swept_blocks (size_t section_reserve)
1950 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1954 * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1955 * a VirtualAlloc ()-ed block.
1961 int i, num_empty_blocks_orig, num_blocks, arr_length;
1963 void **empty_block_arr;
1964 void **rebuild_next;
1966 if (num_empty_blocks <= section_reserve)
1968 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1970 num_empty_blocks_orig = num_empty_blocks;
1971 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1972 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1973 if (!empty_block_arr)
1977 for (block = empty_blocks; block; block = *(void**)block)
1978 empty_block_arr [i++] = block;
1979 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1981 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1984 * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1985 * contiguous ones. If we do, we free them. If that's not enough to get to
1986 * section_reserve, we halve the number of contiguous blocks we're looking
1987 * for and have another go, until we're done with looking for pairs of
1988 * blocks, at which point we give up and go to the fallback.
1990 arr_length = num_empty_blocks_orig;
1991 num_blocks = MS_BLOCK_ALLOC_NUM;
1992 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1997 for (i = 0; i < arr_length; ++i) {
1999 void *block = empty_block_arr [i];
2000 SGEN_ASSERT (6, block, "we're not shifting correctly");
2002 empty_block_arr [dest] = block;
2004 * This is not strictly necessary, but we're
2007 empty_block_arr [i] = NULL;
2016 SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
2018 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
2023 if (d + 1 - first == num_blocks) {
2025 * We found num_blocks contiguous blocks. Free them
2026 * and null their array entries. As an optimization
2027 * we could, instead of nulling the entries, shift
2028 * the following entries over to the left, while
2032 sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
2033 for (j = first; j <= d; ++j)
2034 empty_block_arr [j] = NULL;
2038 num_empty_blocks -= num_blocks;
2040 stat_major_blocks_freed += num_blocks;
2041 if (num_blocks == MS_BLOCK_ALLOC_NUM)
2042 stat_major_blocks_freed_ideal += num_blocks;
2044 stat_major_blocks_freed_less_ideal += num_blocks;
2049 SGEN_ASSERT (6, dest <= i && dest <= arr_length, "array length is off");
2051 SGEN_ASSERT (6, arr_length == num_empty_blocks, "array length is off");
2056 /* rebuild empty_blocks free list */
2057 rebuild_next = (void**)&empty_blocks;
2058 for (i = 0; i < arr_length; ++i) {
2059 void *block = empty_block_arr [i];
2060 SGEN_ASSERT (6, block, "we're missing blocks");
2061 *rebuild_next = block;
2062 rebuild_next = (void**)block;
2064 *rebuild_next = NULL;
2067 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
2070 SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
2074 * This is our threshold. If there's not more empty than used blocks, we won't
2075 * release uncontiguous blocks, in fear of fragmenting the address space.
2077 if (num_empty_blocks <= num_major_sections)
2080 while (num_empty_blocks > section_reserve) {
2081 void *next = *(void**)empty_blocks;
2082 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
2083 empty_blocks = next;
2085 * Needs not be atomic because this is running
2090 ++stat_major_blocks_freed;
2091 ++stat_major_blocks_freed_individual;
2096 major_pin_objects (SgenGrayQueue *queue)
2100 FOREACH_BLOCK_NO_LOCK (block) {
2101 size_t first_entry, last_entry;
2102 SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
2103 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
2104 &first_entry, &last_entry);
2105 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
2106 } END_FOREACH_BLOCK_NO_LOCK;
2110 major_init_to_space (void)
2115 major_report_pinned_memory_usage (void)
2117 g_assert_not_reached ();
2121 major_get_used_size (void)
2127 * We're holding the GC lock, but the sweep thread might be running. Make sure it's
2128 * finished, then we can iterate over the block array.
2130 major_finish_sweep_checking ();
2132 FOREACH_BLOCK_NO_LOCK (block) {
2133 int count = MS_BLOCK_FREE / block->obj_size;
2135 size += count * block->obj_size;
2136 for (iter = block->free_list; iter; iter = (void**)*iter)
2137 size -= block->obj_size;
2138 } END_FOREACH_BLOCK_NO_LOCK;
2143 /* FIXME: return number of bytes, not of sections */
2145 get_num_major_sections (void)
2147 return num_major_sections;
2151 * Returns the number of bytes in blocks that were present when the last sweep was
2152 * initiated, and were not freed during the sweep. They are the basis for calculating the
2156 get_bytes_survived_last_sweep (void)
2158 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2159 return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
2163 major_handle_gc_param (const char *opt)
2165 if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2166 const char *arg = strchr (opt, '=') + 1;
2167 int percentage = atoi (arg);
2168 if (percentage < 0 || percentage > 100) {
2169 fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2172 evacuation_threshold = (float)percentage / 100.0f;
2174 } else if (!strcmp (opt, "lazy-sweep")) {
2177 } else if (!strcmp (opt, "no-lazy-sweep")) {
2180 } else if (!strcmp (opt, "concurrent-sweep")) {
2181 concurrent_sweep = TRUE;
2183 } else if (!strcmp (opt, "no-concurrent-sweep")) {
2184 concurrent_sweep = FALSE;
2192 major_print_gc_param_usage (void)
2196 " evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2197 " (no-)lazy-sweep\n"
2198 " (no-)concurrent-sweep\n"
2203 * This callback is used to clear cards, move cards to the shadow table and do counting.
2206 major_iterate_block_ranges (sgen_cardtable_block_callback callback)
2209 gboolean has_references;
2211 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2213 callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2214 } END_FOREACH_BLOCK_NO_LOCK;
2218 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
2221 gboolean has_references;
2223 major_finish_sweep_checking ();
2224 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2226 callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2227 } END_FOREACH_BLOCK_NO_LOCK;
2230 #ifdef HEAVY_STATISTICS
2231 extern guint64 marked_cards;
2232 extern guint64 scanned_cards;
2233 extern guint64 scanned_objects;
2234 extern guint64 remarked_cards;
2237 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2239 * MS blocks are 16K aligned.
2240 * Cardtables are 4K aligned, at least.
2241 * This means that the cardtable of a given block is 32 bytes aligned.
2244 initial_skip_card (guint8 *card_data)
2246 mword *cards = (mword*)card_data;
2249 for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2255 if (i == CARD_WORDS_PER_BLOCK)
2256 return card_data + CARDS_PER_BLOCK;
2258 #if defined(__i386__) && defined(__GNUC__)
2259 return card_data + i * 4 + (__builtin_ffs (card) - 1) / 8;
2260 #elif defined(__x86_64__) && defined(__GNUC__)
2261 return card_data + i * 8 + (__builtin_ffsll (card) - 1) / 8;
2262 #elif defined(__s390x__) && defined(__GNUC__)
2263 return card_data + i * 8 + (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2265 for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2267 return &card_data [i];
2273 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2274 #define MS_BLOCK_OBJ_FAST(b,os,i) ((b) + MS_BLOCK_SKIP + (os) * (i))
2275 #define MS_OBJ_ALLOCED_FAST(o,b) (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2278 scan_card_table_for_block (MSBlockInfo *block, CardTableScanType scan_type, ScanCopyContext ctx)
2280 SgenGrayQueue *queue = ctx.queue;
2281 ScanObjectFunc scan_func = ctx.ops->scan_object;
2282 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2283 guint8 cards_copy [CARDS_PER_BLOCK];
2285 guint8 cards_preclean [CARDS_PER_BLOCK];
2286 gboolean small_objects;
2289 guint8 *card_data, *card_base;
2290 guint8 *card_data_end;
2291 char *scan_front = NULL;
2293 /* The concurrent mark doesn't enter evacuating blocks */
2294 if (scan_type == CARDTABLE_SCAN_MOD_UNION_PRECLEAN && major_block_is_evacuating (block))
2297 block_obj_size = block->obj_size;
2298 small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2300 block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2303 * This is safe in face of card aliasing for the following reason:
2305 * Major blocks are 16k aligned, or 32 cards aligned.
2306 * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2307 * sizes, they won't overflow the cardtable overlap modulus.
2309 if (scan_type & CARDTABLE_SCAN_MOD_UNION) {
2310 card_data = card_base = block->cardtable_mod_union;
2312 * This happens when the nursery collection that precedes finishing
2313 * the concurrent collection allocates new major blocks.
2318 if (scan_type == CARDTABLE_SCAN_MOD_UNION_PRECLEAN) {
2319 sgen_card_table_preclean_mod_union (card_data, cards_preclean, CARDS_PER_BLOCK);
2320 card_data = card_base = cards_preclean;
2323 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2324 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2326 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2328 card_data = card_base = cards_copy;
2331 card_data_end = card_data + CARDS_PER_BLOCK;
2333 card_data += MS_BLOCK_SKIP >> CARD_BITS;
2335 card_data = initial_skip_card (card_data);
2336 while (card_data < card_data_end) {
2337 size_t card_index, first_object_index;
2340 char *first_obj, *obj;
2342 HEAVY_STAT (++scanned_cards);
2349 card_index = card_data - card_base;
2350 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2351 end = start + CARD_SIZE_IN_BYTES;
2353 if (!block_is_swept_or_marking (block))
2354 sweep_block (block);
2356 HEAVY_STAT (++marked_cards);
2359 sgen_card_table_prepare_card_for_scanning (card_data);
2362 * If the card we're looking at starts at or in the block header, we
2363 * must start at the first object in the block, without calculating
2364 * the index of the object we're hypothetically starting at, because
2365 * it would be negative.
2367 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2368 first_object_index = 0;
2370 first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2372 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2374 binary_protocol_card_scan (first_obj, end - first_obj);
2377 if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2380 if (scan_type & CARDTABLE_SCAN_MOD_UNION) {
2381 /* FIXME: do this more efficiently */
2383 MS_CALC_MARK_BIT (w, b, obj);
2384 if (!MS_MARK_BIT (block, w, b))
2388 GCObject *object = (GCObject*)obj;
2390 if (small_objects) {
2391 HEAVY_STAT (++scanned_objects);
2392 scan_func (object, sgen_obj_get_descriptor (object), queue);
2394 size_t offset = sgen_card_table_get_card_offset (obj, block_start);
2395 sgen_cardtable_scan_object (object, block_obj_size, card_base + offset, ctx);
2398 obj += block_obj_size;
2399 g_assert (scan_front <= obj);
2403 HEAVY_STAT (if (*card_data) ++remarked_cards);
2408 card_data = card_base + sgen_card_table_get_card_offset (obj, block_start);
2413 major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx)
2416 gboolean has_references, was_sweeping, skip_scan;
2418 if (!concurrent_mark)
2419 g_assert (scan_type == CARDTABLE_SCAN_GLOBAL);
2421 if (scan_type != CARDTABLE_SCAN_GLOBAL)
2422 SGEN_ASSERT (0, !sweep_in_progress (), "Sweep should be finished when we scan mod union card table");
2423 was_sweeping = sweep_in_progress ();
2425 binary_protocol_major_card_table_scan_start (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
2426 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2427 #ifdef PREFETCH_CARDS
2428 int prefetch_index = __index + 6;
2429 if (prefetch_index < allocated_blocks.next_slot) {
2430 MSBlockInfo *prefetch_block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, prefetch_index));
2431 PREFETCH_READ (prefetch_block);
2432 if (scan_type == CARDTABLE_SCAN_GLOBAL) {
2433 guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2434 PREFETCH_WRITE (prefetch_cards);
2435 PREFETCH_WRITE (prefetch_cards + 32);
2440 if (!has_references)
2444 if (scan_type == CARDTABLE_SCAN_GLOBAL) {
2445 gpointer *card_start = (gpointer*) sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (block));
2446 gboolean has_dirty_cards = FALSE;
2448 for (i = 0; i < CARDS_PER_BLOCK / sizeof(gpointer); i++) {
2449 if (card_start [i]) {
2450 has_dirty_cards = TRUE;
2454 if (!has_dirty_cards) {
2458 * After the start of the concurrent collections, blocks change state
2459 * to marking. We should not sweep it in that case. We can't race with
2460 * sweep start since we are in a nursery collection. Also avoid CAS-ing
2462 if (sweep_in_progress ()) {
2463 skip_scan = !ensure_block_is_checked_for_sweeping (__index, TRUE, NULL);
2464 } else if (was_sweeping) {
2465 /* Recheck in case sweep finished after dereferencing the slot */
2466 skip_scan = *sgen_array_list_get_slot (&allocated_blocks, __index) == 0;
2471 scan_card_table_for_block (block, scan_type, ctx);
2472 } END_FOREACH_BLOCK_NO_LOCK;
2473 binary_protocol_major_card_table_scan_end (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
2477 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2480 gboolean has_references;
2481 long long total_cards = 0;
2482 long long marked_cards = 0;
2484 if (sweep_in_progress ()) {
2485 *num_total_cards = -1;
2486 *num_marked_cards = -1;
2490 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2491 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2494 if (!has_references)
2497 total_cards += CARDS_PER_BLOCK;
2498 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2502 } END_FOREACH_BLOCK_NO_LOCK;
2504 *num_total_cards = total_cards;
2505 *num_marked_cards = marked_cards;
2509 update_cardtable_mod_union (void)
2513 FOREACH_BLOCK_NO_LOCK (block) {
2514 gpointer *card_start = (gpointer*) sgen_card_table_get_card_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (block));
2515 gboolean has_dirty_cards = FALSE;
2517 for (i = 0; i < CARDS_PER_BLOCK / sizeof(gpointer); i++) {
2518 if (card_start [i]) {
2519 has_dirty_cards = TRUE;
2523 if (has_dirty_cards) {
2525 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
2526 sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2527 SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2529 } END_FOREACH_BLOCK_NO_LOCK;
2532 #undef pthread_create
2535 post_param_init (SgenMajorCollector *collector)
2537 collector->sweeps_lazily = lazy_sweep;
2538 collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
2542 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent, gboolean is_parallel)
2546 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2548 num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2549 block_obj_sizes = (int *)sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2550 ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2552 evacuate_block_obj_sizes = (gboolean *)sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2553 for (i = 0; i < num_block_obj_sizes; ++i)
2554 evacuate_block_obj_sizes [i] = FALSE;
2556 sweep_slots_available = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2557 sweep_slots_used = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2558 sweep_num_blocks = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2563 g_print ("block object sizes:\n");
2564 for (i = 0; i < num_block_obj_sizes; ++i)
2565 g_print ("%d\n", block_obj_sizes [i]);
2569 for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2570 free_block_lists [i] = (MSBlockInfo *volatile *)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2572 for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2573 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2574 for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2575 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2577 mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2578 mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2579 mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2580 mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2581 mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2582 mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2583 mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2585 collector->section_size = MAJOR_SECTION_SIZE;
2587 concurrent_mark = is_concurrent;
2588 collector->is_concurrent = is_concurrent;
2589 collector->is_parallel = is_parallel;
2590 collector->needs_thread_pool = is_concurrent || concurrent_sweep;
2591 collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2592 collector->supports_cardtable = TRUE;
2594 collector->alloc_heap = major_alloc_heap;
2595 collector->is_object_live = major_is_object_live;
2596 collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2597 collector->alloc_degraded = major_alloc_degraded;
2599 collector->alloc_object = major_alloc_object;
2600 collector->free_pinned_object = free_pinned_object;
2601 collector->iterate_objects = major_iterate_objects;
2602 collector->free_non_pinned_object = major_free_non_pinned_object;
2603 collector->pin_objects = major_pin_objects;
2604 collector->pin_major_object = pin_major_object;
2605 collector->scan_card_table = major_scan_card_table;
2606 collector->iterate_live_block_ranges = major_iterate_live_block_ranges;
2607 collector->iterate_block_ranges = major_iterate_block_ranges;
2608 if (is_concurrent) {
2609 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2610 collector->get_cardtable_mod_union_for_reference = major_get_cardtable_mod_union_for_reference;
2612 collector->init_to_space = major_init_to_space;
2613 collector->sweep = major_sweep;
2614 collector->have_swept = major_have_swept;
2615 collector->finish_sweeping = major_finish_sweep_checking;
2616 collector->free_swept_blocks = major_free_swept_blocks;
2617 collector->check_scan_starts = major_check_scan_starts;
2618 collector->dump_heap = major_dump_heap;
2619 collector->get_used_size = major_get_used_size;
2620 collector->start_nursery_collection = major_start_nursery_collection;
2621 collector->finish_nursery_collection = major_finish_nursery_collection;
2622 collector->start_major_collection = major_start_major_collection;
2623 collector->finish_major_collection = major_finish_major_collection;
2624 collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2625 collector->ptr_is_from_pinned_alloc = ptr_is_from_pinned_alloc;
2626 collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2627 collector->get_num_major_sections = get_num_major_sections;
2628 collector->get_bytes_survived_last_sweep = get_bytes_survived_last_sweep;
2629 collector->handle_gc_param = major_handle_gc_param;
2630 collector->print_gc_param_usage = major_print_gc_param_usage;
2631 collector->post_param_init = post_param_init;
2632 collector->is_valid_object = major_is_valid_object;
2633 collector->describe_pointer = major_describe_pointer;
2634 collector->count_cards = major_count_cards;
2636 collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
2637 collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
2638 collector->major_ops_serial.drain_gray_stack = drain_gray_stack;
2639 if (is_concurrent) {
2640 collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2641 collector->major_ops_concurrent_start.scan_object = major_scan_object_concurrent_with_evacuation;
2642 collector->major_ops_concurrent_start.scan_vtype = major_scan_vtype_concurrent_with_evacuation;
2643 collector->major_ops_concurrent_start.scan_ptr_field = major_scan_ptr_field_concurrent_with_evacuation;
2644 collector->major_ops_concurrent_start.drain_gray_stack = drain_gray_stack_concurrent;
2646 collector->major_ops_concurrent_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_finish_canonical;
2647 collector->major_ops_concurrent_finish.scan_object = major_scan_object_with_evacuation;
2648 collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_with_evacuation;
2649 collector->major_ops_concurrent_finish.scan_ptr_field = major_scan_ptr_field_with_evacuation;
2650 collector->major_ops_concurrent_finish.drain_gray_stack = drain_gray_stack;
2653 /* FIXME use parallel obj ops */
2654 collector->major_ops_conc_par_start = collector->major_ops_concurrent_start;
2655 collector->major_ops_conc_par_finish = collector->major_ops_concurrent_finish;
2659 #ifdef HEAVY_STATISTICS
2660 mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2661 mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2662 mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2663 mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2664 mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2665 mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2666 mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2667 mono_counters_register ("Optimized copy major small evacuate", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_evacuate);
2668 mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2669 mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2670 mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2672 mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2673 mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2674 mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2677 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2678 mono_os_mutex_init (&scanned_objects_list_lock);
2681 SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2683 /*cardtable requires major pages to be 8 cards aligned*/
2684 g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2688 sgen_marksweep_init (SgenMajorCollector *collector)
2690 sgen_marksweep_init_internal (collector, FALSE, FALSE);
2694 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2696 sgen_marksweep_init_internal (collector, TRUE, FALSE);
2700 sgen_marksweep_conc_par_init (SgenMajorCollector *collector)
2702 sgen_marksweep_init_internal (collector, TRUE, TRUE);