2 * sgen-marksweep.c: The Mark & Sweep major collector.
5 * Mark Probst <mark.probst@gmail.com>
7 * Copyright 2009-2010 Novell, Inc.
8 * Copyright (C) 2012 Xamarin Inc
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Library General Public
12 * License 2.0 as published by the Free Software Foundation;
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Library General Public License for more details.
19 * You should have received a copy of the GNU Library General Public
20 * License 2.0 along with this library; if not, write to the Free
21 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
33 #include "mono/sgen/sgen-gc.h"
34 #include "mono/sgen/sgen-protocol.h"
35 #include "mono/sgen/sgen-cardtable.h"
36 #include "mono/sgen/sgen-memory-governor.h"
37 #include "mono/sgen/sgen-layout-stats.h"
38 #include "mono/sgen/sgen-pointer-queue.h"
39 #include "mono/sgen/sgen-pinning.h"
40 #include "mono/sgen/sgen-workers.h"
41 #include "mono/sgen/sgen-thread-pool.h"
42 #include "mono/sgen/sgen-client.h"
43 #include "mono/utils/mono-membar.h"
45 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
46 #define MS_BLOCK_SIZE ARCH_MIN_MS_BLOCK_SIZE
47 #define MS_BLOCK_SIZE_SHIFT ARCH_MIN_MS_BLOCK_SIZE_SHIFT
49 #define MS_BLOCK_SIZE_SHIFT 14 /* INT FASTENABLE */
50 #define MS_BLOCK_SIZE (1 << MS_BLOCK_SIZE_SHIFT)
52 #define MAJOR_SECTION_SIZE MS_BLOCK_SIZE
53 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
56 * Don't allocate single blocks, but alloc a contingent of this many
57 * blocks in one swoop. This must be a power of two.
59 #define MS_BLOCK_ALLOC_NUM 32
62 * Number of bytes before the first object in a block. At the start
63 * of a block is the MSBlockHeader, then opional padding, then come
64 * the objects, so this must be >= sizeof (MSBlockHeader).
66 #define MS_BLOCK_SKIP ((sizeof (MSBlockHeader) + 15) & ~15)
68 #define MS_BLOCK_FREE (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
70 #define MS_NUM_MARK_WORDS ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
73 * Blocks progress from one state to the next:
75 * SWEPT The block is fully swept. It might or might not be in
78 * MARKING The block might or might not contain live objects. If
79 * we're in between an initial collection pause and the
80 * finishing pause, the block might or might not be in a
83 * CHECKING The sweep thread is investigating the block to determine
84 * whether or not it contains live objects. The block is
87 * NEED_SWEEPING The block contains live objects but has not yet been
88 * swept. It also contains free slots. It is in a block
91 * SWEEPING The block is being swept. It might be in a free list.
98 BLOCK_STATE_NEED_SWEEPING,
102 typedef struct _MSBlockInfo MSBlockInfo;
103 struct _MSBlockInfo {
106 * FIXME: Do we even need this? It's only used during sweep and might be worth
107 * recalculating to save the space.
109 guint16 obj_size_index;
110 /* FIXME: Reduce this - it only needs a byte. */
111 volatile gint32 state;
112 unsigned int pinned : 1;
113 unsigned int has_references : 1;
114 unsigned int has_pinned : 1; /* means cannot evacuate */
115 unsigned int is_to_space : 1;
116 void ** volatile free_list;
117 MSBlockInfo * volatile next_free;
118 guint8 * volatile cardtable_mod_union;
119 mword mark_words [MS_NUM_MARK_WORDS];
122 #define MS_BLOCK_FOR_BLOCK_INFO(b) ((char*)(b))
124 #define MS_BLOCK_OBJ(b,i) ((void*)(MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i)))
125 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size) (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
126 #define MS_BLOCK_DATA_FOR_OBJ(o) ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
132 #define MS_BLOCK_FOR_OBJ(o) (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
134 /* object index will always be small */
135 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
137 //casting to int is fine since blocks are 32k
138 #define MS_CALC_MARK_BIT(w,b,o) do { \
139 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
140 if (sizeof (mword) == 4) { \
149 #define MS_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] & (ONE_P << (b)))
150 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
152 #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))
154 #define MS_BLOCK_OBJ_SIZE_FACTOR (pow (2.0, 1.0 / 3))
157 * This way we can lookup block object size indexes for sizes up to
158 * 256 bytes with a single load.
160 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES 32
162 static int *block_obj_sizes;
163 static int num_block_obj_sizes;
164 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
166 #define MS_BLOCK_FLAG_PINNED 1
167 #define MS_BLOCK_FLAG_REFS 2
169 #define MS_BLOCK_TYPE_MAX 4
171 static gboolean *evacuate_block_obj_sizes;
172 static float evacuation_threshold = 0.666f;
173 static float concurrent_evacuation_threshold = 0.666f;
174 static gboolean want_evacuation = FALSE;
176 static gboolean lazy_sweep = FALSE;
180 SWEEP_STATE_NEED_SWEEPING,
181 SWEEP_STATE_SWEEPING,
182 SWEEP_STATE_SWEEPING_AND_ITERATING,
183 SWEEP_STATE_COMPACTING
186 static volatile int sweep_state = SWEEP_STATE_SWEPT;
188 static gboolean concurrent_mark;
189 static gboolean concurrent_sweep = TRUE;
191 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl) SGEN_POINTER_IS_TAGGED_1 ((bl))
192 #define BLOCK_TAG_HAS_REFERENCES(bl) SGEN_POINTER_TAG_1 ((bl))
194 #define BLOCK_IS_TAGGED_CHECKING(bl) SGEN_POINTER_IS_TAGGED_2 ((bl))
195 #define BLOCK_TAG_CHECKING(bl) SGEN_POINTER_TAG_2 ((bl))
197 #define BLOCK_UNTAG(bl) SGEN_POINTER_UNTAG_12 ((bl))
199 #define BLOCK_TAG(bl) ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
201 /* all allocated blocks in the system */
202 static SgenPointerQueue allocated_blocks;
204 /* non-allocated block free-list */
205 static void *empty_blocks = NULL;
206 static size_t num_empty_blocks = 0;
208 #define FOREACH_BLOCK_NO_LOCK_CONDITION(cond,bl) { \
210 SGEN_ASSERT (0, (cond) && !sweep_in_progress (), "Can't iterate blocks while the world is running or sweep is in progress."); \
211 for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { \
212 (bl) = BLOCK_UNTAG (allocated_blocks.data [__index]);
213 #define FOREACH_BLOCK_NO_LOCK(bl) \
214 FOREACH_BLOCK_NO_LOCK_CONDITION(sgen_is_world_stopped (), bl)
215 #define FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK(bl,hr) { \
217 SGEN_ASSERT (0, sgen_is_world_stopped () && !sweep_in_progress (), "Can't iterate blocks while the world is running or sweep is in progress."); \
218 for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { \
219 (bl) = allocated_blocks.data [__index]; \
220 (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl)); \
221 (bl) = BLOCK_UNTAG ((bl));
222 #define END_FOREACH_BLOCK_NO_LOCK } }
224 static volatile size_t num_major_sections = 0;
226 * One free block list for each block object size. We add and remove blocks from these
227 * lists lock-free via CAS.
229 * Blocks accessed/removed from `free_block_lists`:
230 * from the mutator (with GC lock held)
231 * in nursery collections
232 * in non-concurrent major collections
233 * in the finishing pause of concurrent major collections (whole list is cleared)
235 * Blocks added to `free_block_lists`:
236 * in the sweeping thread
237 * during nursery collections
238 * from domain clearing (with the world stopped and no sweeping happening)
240 * The only item of those that doesn't require the GC lock is the sweep thread. The sweep
241 * thread only ever adds blocks to the free list, so the ABA problem can't occur.
243 static MSBlockInfo * volatile *free_block_lists [MS_BLOCK_TYPE_MAX];
245 static guint64 stat_major_blocks_alloced = 0;
246 static guint64 stat_major_blocks_freed = 0;
247 static guint64 stat_major_blocks_lazy_swept = 0;
248 static guint64 stat_major_objects_evacuated = 0;
250 #if SIZEOF_VOID_P != 8
251 static guint64 stat_major_blocks_freed_ideal = 0;
252 static guint64 stat_major_blocks_freed_less_ideal = 0;
253 static guint64 stat_major_blocks_freed_individual = 0;
254 static guint64 stat_major_blocks_alloced_less_ideal = 0;
257 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
258 static guint64 num_major_objects_marked = 0;
259 #define INC_NUM_MAJOR_OBJECTS_MARKED() (++num_major_objects_marked)
261 #define INC_NUM_MAJOR_OBJECTS_MARKED()
264 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
265 static mono_mutex_t scanned_objects_list_lock;
266 static SgenPointerQueue scanned_objects_list;
269 add_scanned_object (void *ptr)
271 if (!binary_protocol_is_enabled ())
274 mono_mutex_lock (&scanned_objects_list_lock);
275 sgen_pointer_queue_add (&scanned_objects_list, ptr);
276 mono_mutex_unlock (&scanned_objects_list_lock);
280 static gboolean sweep_block (MSBlockInfo *block);
283 ms_find_block_obj_size_index (size_t size)
286 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);
287 for (i = 0; i < num_block_obj_sizes; ++i)
288 if (block_obj_sizes [i] >= size)
290 g_error ("no object of size %zd\n", size);
294 #define FREE_BLOCKS_FROM(lists,p,r) (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
295 #define FREE_BLOCKS(p,r) (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
297 #define MS_BLOCK_OBJ_SIZE_INDEX(s) \
298 (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ? \
299 fast_block_obj_size_indexes [((s)+7)>>3] : \
300 ms_find_block_obj_size_index ((s)))
303 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
307 start = sgen_alloc_os_memory_aligned (nursery_size, nursery_align, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
309 start = sgen_alloc_os_memory (nursery_size, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
315 update_heap_boundaries_for_block (MSBlockInfo *block)
317 sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
324 ms_get_empty_block (void)
328 void *block, *empty, *next;
333 * We try allocating MS_BLOCK_ALLOC_NUM blocks first. If that's
334 * unsuccessful, we halve the number of blocks and try again, until we're at
335 * 1. If that doesn't work, either, we assert.
337 int alloc_num = MS_BLOCK_ALLOC_NUM;
339 p = sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE,
340 alloc_num == 1 ? "major heap section" : NULL);
346 for (i = 0; i < alloc_num; ++i) {
349 * We do the free list update one after the
350 * other so that other threads can use the new
351 * blocks as quickly as possible.
354 empty = empty_blocks;
355 *(void**)block = empty;
356 } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
360 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
362 stat_major_blocks_alloced += alloc_num;
363 #if SIZEOF_VOID_P != 8
364 if (alloc_num != MS_BLOCK_ALLOC_NUM)
365 stat_major_blocks_alloced_less_ideal += alloc_num;
370 empty = empty_blocks;
374 next = *(void**)block;
375 } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
377 SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
379 *(void**)block = NULL;
381 g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
387 * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
388 * list, where it will either be freed later on, or reused in nursery collections.
391 ms_free_block (void *block)
395 sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
396 memset (block, 0, MS_BLOCK_SIZE);
399 empty = empty_blocks;
400 *(void**)block = empty;
401 } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
403 SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
405 binary_protocol_block_free (block, MS_BLOCK_SIZE);
409 sweep_in_progress (void)
411 int state = sweep_state;
412 return state == SWEEP_STATE_SWEEPING ||
413 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
414 state == SWEEP_STATE_COMPACTING;
417 static inline gboolean
418 block_is_swept_or_marking (MSBlockInfo *block)
420 gint32 state = block->state;
421 return state == BLOCK_STATE_SWEPT || state == BLOCK_STATE_MARKING;
424 //#define MARKSWEEP_CONSISTENCY_CHECK
426 #ifdef MARKSWEEP_CONSISTENCY_CHECK
428 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
430 SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
431 for (; block; block = block->next_free) {
432 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
433 g_assert (block->obj_size == size);
434 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
436 /* blocks in the free lists must have at least
438 g_assert (block->free_list);
440 /* the block must be in the allocated_blocks array */
441 g_assert (sgen_pointer_queue_find (&allocated_blocks, BLOCK_TAG (block)) != (size_t)-1);
446 check_empty_blocks (void)
450 for (p = empty_blocks; p; p = *(void**)p)
452 g_assert (i == num_empty_blocks);
456 consistency_check (void)
461 /* check all blocks */
462 FOREACH_BLOCK_NO_LOCK (block) {
463 int count = MS_BLOCK_FREE / block->obj_size;
467 /* count number of free slots */
468 for (i = 0; i < count; ++i) {
469 void **obj = (void**) MS_BLOCK_OBJ (block, i);
470 if (!MS_OBJ_ALLOCED (obj, block))
474 /* check free list */
475 for (free = block->free_list; free; free = (void**)*free) {
476 g_assert (MS_BLOCK_FOR_OBJ (free) == block);
479 g_assert (num_free == 0);
481 /* check all mark words are zero */
482 if (!sgen_concurrent_collection_in_progress () && block_is_swept_or_marking (block)) {
483 for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
484 g_assert (block->mark_words [i] == 0);
486 } END_FOREACH_BLOCK_NO_LOCK;
488 /* check free blocks */
489 for (i = 0; i < num_block_obj_sizes; ++i) {
491 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
492 check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
495 check_empty_blocks ();
500 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
504 block->next_free = old = free_blocks [size_index];
505 } while (SGEN_CAS_PTR ((gpointer)&free_blocks [size_index], block, old) != old);
508 static void major_finish_sweep_checking (void);
511 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
513 int size = block_obj_sizes [size_index];
514 int count = MS_BLOCK_FREE / size;
516 MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
520 if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
523 info = (MSBlockInfo*)ms_get_empty_block ();
525 SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
527 info->obj_size = size;
528 info->obj_size_index = size_index;
529 info->pinned = pinned;
530 info->has_references = has_references;
531 info->has_pinned = pinned;
533 * Blocks that are to-space are not evacuated from. During an major collection
534 * blocks are allocated for two reasons: evacuating objects from the nursery and
535 * evacuating them from major blocks marked for evacuation. In both cases we don't
536 * want further evacuation.
538 info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD);
539 info->state = (info->is_to_space || sgen_concurrent_collection_in_progress ()) ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
540 SGEN_ASSERT (6, !sweep_in_progress () || info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
541 info->cardtable_mod_union = NULL;
543 update_heap_boundaries_for_block (info);
545 binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
547 /* build free list */
548 obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
549 info->free_list = (void**)obj_start;
550 /* we're skipping the last one - it must be nulled */
551 for (i = 0; i < count - 1; ++i) {
552 char *next_obj_start = obj_start + size;
553 *(void**)obj_start = next_obj_start;
554 obj_start = next_obj_start;
557 *(void**)obj_start = NULL;
559 add_free_block (free_blocks, size_index, info);
562 * This is the only place where the `allocated_blocks` array can potentially grow.
563 * We need to make sure concurrent sweep isn't running when that happens, so in that
564 * specific case we just wait for sweep to finish.
566 if (sgen_pointer_queue_will_grow (&allocated_blocks))
567 major_finish_sweep_checking ();
569 sgen_pointer_queue_add (&allocated_blocks, BLOCK_TAG (info));
571 SGEN_ATOMIC_ADD_P (num_major_sections, 1);
576 ptr_is_from_pinned_alloc (char *ptr)
580 FOREACH_BLOCK_NO_LOCK (block) {
581 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE)
582 return block->pinned;
583 } END_FOREACH_BLOCK_NO_LOCK;
588 ensure_can_access_block_free_list (MSBlockInfo *block)
592 switch (block->state) {
593 case BLOCK_STATE_SWEPT:
594 case BLOCK_STATE_MARKING:
596 case BLOCK_STATE_CHECKING:
597 SGEN_ASSERT (0, FALSE, "How did we get a block that's being checked from a free list?");
599 case BLOCK_STATE_NEED_SWEEPING:
600 if (sweep_block (block))
601 ++stat_major_blocks_lazy_swept;
603 case BLOCK_STATE_SWEEPING:
604 /* FIXME: do this more elegantly */
608 SGEN_ASSERT (0, FALSE, "Illegal block state");
615 unlink_slot_from_free_list_uncontested (MSBlockInfo * volatile *free_blocks, int size_index)
617 MSBlockInfo *block, *next_free_block;
618 void *obj, *next_free_slot;
621 block = free_blocks [size_index];
622 SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
624 ensure_can_access_block_free_list (block);
626 obj = block->free_list;
627 SGEN_ASSERT (6, obj, "block %p in free list had no available object to alloc from", block);
629 next_free_slot = *(void**)obj;
630 if (next_free_slot) {
631 block->free_list = next_free_slot;
635 next_free_block = block->next_free;
636 if (SGEN_CAS_PTR ((gpointer)&free_blocks [size_index], next_free_block, block) != block)
639 block->free_list = NULL;
640 block->next_free = NULL;
646 alloc_obj (GCVTable vtable, size_t size, gboolean pinned, gboolean has_references)
648 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
649 MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
652 if (!free_blocks [size_index]) {
653 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
657 obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
659 /* FIXME: assumes object layout */
660 *(GCVTable*)obj = vtable;
666 major_alloc_object (GCVTable vtable, size_t size, gboolean has_references)
668 return alloc_obj (vtable, size, FALSE, has_references);
672 * We're not freeing the block if it's empty. We leave that work for
673 * the next major collection.
675 * This is just called from the domain clearing code, which runs in a
676 * single thread and has the GC lock, so we don't need an extra lock.
679 free_object (GCObject *obj, size_t size, gboolean pinned)
681 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
683 gboolean in_free_list;
685 SGEN_ASSERT (9, sweep_state == SWEEP_STATE_SWEPT, "Should have waited for sweep to free objects.");
687 ensure_can_access_block_free_list (block);
688 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);
689 SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
690 MS_CALC_MARK_BIT (word, bit, obj);
691 SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set", obj);
693 memset (obj, 0, size);
695 in_free_list = !!block->free_list;
696 *(void**)obj = block->free_list;
697 block->free_list = (void**)obj;
700 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (pinned, block->has_references);
701 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
702 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);
703 add_free_block (free_blocks, size_index, block);
708 major_free_non_pinned_object (GCObject *obj, size_t size)
710 free_object (obj, size, FALSE);
713 /* size is a multiple of SGEN_ALLOC_ALIGN */
715 major_alloc_small_pinned_obj (GCVTable vtable, size_t size, gboolean has_references)
719 res = alloc_obj (vtable, size, TRUE, has_references);
720 /*If we failed to alloc memory, we better try releasing memory
721 *as pinned alloc is requested by the runtime.
724 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE);
725 res = alloc_obj (vtable, size, TRUE, has_references);
731 free_pinned_object (GCObject *obj, size_t size)
733 free_object (obj, size, TRUE);
737 * size is already rounded up and we hold the GC lock.
740 major_alloc_degraded (GCVTable vtable, size_t size)
742 GCObject *obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
743 if (G_LIKELY (obj)) {
744 HEAVY_STAT (++stat_objects_alloced_degraded);
745 HEAVY_STAT (stat_bytes_alloced_degraded += size);
751 * obj is some object. If it's not in the major heap (i.e. if it's in
752 * the nursery or LOS), return FALSE. Otherwise return whether it's
753 * been marked or copied.
756 major_is_object_live (GCObject *obj)
762 if (sgen_ptr_in_nursery (obj))
765 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
768 if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
771 /* now we know it's in a major block */
772 block = MS_BLOCK_FOR_OBJ (obj);
773 SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?", block);
774 MS_CALC_MARK_BIT (word, bit, obj);
775 return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
779 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
783 FOREACH_BLOCK_NO_LOCK (block) {
784 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
785 int count = MS_BLOCK_FREE / block->obj_size;
789 for (i = 0; i <= count; ++i) {
790 if (ptr >= (char*)MS_BLOCK_OBJ (block, i) && ptr < (char*)MS_BLOCK_OBJ (block, i + 1)) {
791 *start = MS_BLOCK_OBJ (block, i);
795 return !block->pinned;
797 } END_FOREACH_BLOCK_NO_LOCK;
802 try_set_sweep_state (int new, int expected)
804 int old = SGEN_CAS (&sweep_state, new, expected);
805 return old == expected;
809 set_sweep_state (int new, int expected)
811 gboolean success = try_set_sweep_state (new, expected);
812 SGEN_ASSERT (0, success, "Could not set sweep state.");
815 static gboolean ensure_block_is_checked_for_sweeping (int block_index, gboolean wait, gboolean *have_checked);
817 static SgenThreadPoolJob * volatile sweep_job;
820 major_finish_sweep_checking (void)
823 SgenThreadPoolJob *job;
826 switch (sweep_state) {
827 case SWEEP_STATE_SWEPT:
828 case SWEEP_STATE_NEED_SWEEPING:
830 case SWEEP_STATE_SWEEPING:
831 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
834 case SWEEP_STATE_SWEEPING_AND_ITERATING:
835 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
837 case SWEEP_STATE_COMPACTING:
840 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
845 * We're running with the world stopped and the only other thread doing work is the
846 * sweep thread, which doesn't add blocks to the array, so we can safely access
849 for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index)
850 ensure_block_is_checked_for_sweeping (block_index, FALSE, NULL);
852 set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
857 sgen_thread_pool_job_wait (job);
858 SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
859 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
863 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
865 gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
866 gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
867 gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
870 major_finish_sweep_checking ();
871 FOREACH_BLOCK_NO_LOCK (block) {
872 int count = MS_BLOCK_FREE / block->obj_size;
875 if (block->pinned && !pinned)
877 if (!block->pinned && !non_pinned)
879 if (sweep && lazy_sweep) {
881 SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
884 for (i = 0; i < count; ++i) {
885 void **obj = (void**) MS_BLOCK_OBJ (block, i);
887 * We've finished sweep checking, but if we're sweeping lazily and
888 * the flags don't require us to sweep, the block might still need
889 * sweeping. In that case, we need to consult the mark bits to tell
890 * us whether an object slot is live.
892 if (!block_is_swept_or_marking (block)) {
894 SGEN_ASSERT (6, !sweep && block->state == BLOCK_STATE_NEED_SWEEPING, "Has sweeping not finished?");
895 MS_CALC_MARK_BIT (word, bit, obj);
896 if (!MS_MARK_BIT (block, word, bit))
899 if (MS_OBJ_ALLOCED (obj, block))
900 callback ((GCObject*)obj, block->obj_size, data);
902 } END_FOREACH_BLOCK_NO_LOCK;
906 major_is_valid_object (char *object)
910 FOREACH_BLOCK_NO_LOCK (block) {
914 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
917 idx = MS_BLOCK_OBJ_INDEX (object, block);
918 obj = (char*)MS_BLOCK_OBJ (block, idx);
921 return MS_OBJ_ALLOCED (obj, block);
922 } END_FOREACH_BLOCK_NO_LOCK;
929 major_describe_pointer (char *ptr)
933 FOREACH_BLOCK_NO_LOCK (block) {
941 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
944 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
945 MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
947 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
948 obj = (char*)MS_BLOCK_OBJ (block, idx);
949 live = MS_OBJ_ALLOCED (obj, block);
950 vtable = live ? SGEN_LOAD_VTABLE ((GCObject*)obj) : NULL;
952 MS_CALC_MARK_BIT (w, b, obj);
953 marked = MS_MARK_BIT (block, w, b);
958 SGEN_LOG (0, "object");
960 SGEN_LOG (0, "dead-object");
963 SGEN_LOG (0, "interior-ptr offset %zd", ptr - obj);
965 SGEN_LOG (0, "dead-interior-ptr offset %zd", ptr - obj);
968 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
971 } END_FOREACH_BLOCK_NO_LOCK;
977 major_check_scan_starts (void)
982 major_dump_heap (FILE *heap_dump_file)
985 int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
986 int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
989 for (i = 0; i < num_block_obj_sizes; ++i)
990 slots_available [i] = slots_used [i] = 0;
992 FOREACH_BLOCK_NO_LOCK (block) {
993 int index = ms_find_block_obj_size_index (block->obj_size);
994 int count = MS_BLOCK_FREE / block->obj_size;
996 slots_available [index] += count;
997 for (i = 0; i < count; ++i) {
998 if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
999 ++slots_used [index];
1001 } END_FOREACH_BLOCK_NO_LOCK;
1003 fprintf (heap_dump_file, "<occupancies>\n");
1004 for (i = 0; i < num_block_obj_sizes; ++i) {
1005 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
1006 block_obj_sizes [i], slots_available [i], slots_used [i]);
1008 fprintf (heap_dump_file, "</occupancies>\n");
1010 FOREACH_BLOCK_NO_LOCK (block) {
1011 int count = MS_BLOCK_FREE / block->obj_size;
1015 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
1017 for (i = 0; i <= count; ++i) {
1018 if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
1023 sgen_dump_occupied (MS_BLOCK_OBJ (block, start), MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
1029 fprintf (heap_dump_file, "</section>\n");
1030 } END_FOREACH_BLOCK_NO_LOCK;
1034 get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate)
1036 guint8 *mod_union = block->cardtable_mod_union;
1042 mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1043 other = SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
1045 SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
1048 sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1052 static inline guint8*
1053 major_get_cardtable_mod_union_for_reference (char *ptr)
1055 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (ptr);
1056 size_t offset = sgen_card_table_get_card_offset (ptr, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1057 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
1058 SGEN_ASSERT (0, mod_union, "FIXME: optionally allocate the mod union if it's not here and CAS it in.");
1059 return &mod_union [offset];
1063 * Mark the mod-union card for `ptr`, which must be a reference within the object `obj`.
1066 mark_mod_union_card (GCObject *obj, void **ptr)
1068 int type = sgen_obj_get_descriptor (obj) & DESC_TYPE_MASK;
1069 if (sgen_safe_object_is_small (obj, type)) {
1070 guint8 *card_byte = major_get_cardtable_mod_union_for_reference ((char*)ptr);
1071 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?");
1074 sgen_los_mark_mod_union_card (obj, ptr);
1078 #define LOAD_VTABLE SGEN_LOAD_VTABLE
1080 #define MS_MARK_OBJECT_AND_ENQUEUE_CHECKED(obj,desc,block,queue) do { \
1081 int __word, __bit; \
1082 MS_CALC_MARK_BIT (__word, __bit, (obj)); \
1083 if (!MS_MARK_BIT ((block), __word, __bit) && MS_OBJ_ALLOCED ((obj), (block))) { \
1084 MS_SET_MARK_BIT ((block), __word, __bit); \
1085 if (sgen_gc_descr_has_references (desc)) \
1086 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1087 binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1088 INC_NUM_MAJOR_OBJECTS_MARKED (); \
1091 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do { \
1092 int __word, __bit; \
1093 MS_CALC_MARK_BIT (__word, __bit, (obj)); \
1094 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
1095 if (!MS_MARK_BIT ((block), __word, __bit)) { \
1096 MS_SET_MARK_BIT ((block), __word, __bit); \
1097 if (sgen_gc_descr_has_references (desc)) \
1098 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1099 binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1100 INC_NUM_MAJOR_OBJECTS_MARKED (); \
1105 pin_major_object (GCObject *obj, SgenGrayQueue *queue)
1109 if (concurrent_mark)
1110 g_assert_not_reached ();
1112 block = MS_BLOCK_FOR_OBJ (obj);
1113 block->has_pinned = TRUE;
1114 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1117 #include "sgen-major-copy-object.h"
1120 major_copy_or_mark_object_concurrent (GCObject **ptr, GCObject *obj, SgenGrayQueue *queue)
1122 SGEN_ASSERT (9, sgen_concurrent_collection_in_progress (), "Why are we scanning concurrently when there's no concurrent collection on?");
1123 SGEN_ASSERT (9, !sgen_workers_are_working () || sgen_thread_pool_is_thread_pool_thread (mono_native_thread_id_get ()), "We must not scan from two threads at the same time!");
1125 g_assert (!SGEN_OBJECT_IS_FORWARDED (obj));
1127 if (!sgen_ptr_in_nursery (obj)) {
1130 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
1132 if (objsize <= SGEN_MAX_SMALL_OBJ_SIZE) {
1133 MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
1134 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1136 if (sgen_los_object_is_pinned (obj))
1139 binary_protocol_mark (obj, SGEN_LOAD_VTABLE (obj), sgen_safe_object_get_size (obj));
1141 sgen_los_pin_object (obj);
1142 if (SGEN_OBJECT_HAS_REFERENCES (obj))
1143 GRAY_OBJECT_ENQUEUE (queue, obj, sgen_obj_get_descriptor (obj));
1144 INC_NUM_MAJOR_OBJECTS_MARKED ();
1150 major_get_and_reset_num_major_objects_marked (void)
1152 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1153 long long num = num_major_objects_marked;
1154 num_major_objects_marked = 0;
1161 #define PREFETCH_CARDS 1 /* BOOL FASTENABLE */
1163 #undef PREFETCH_CARDS
1166 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1167 #if defined(PLATFORM_MACOSX)
1168 #define GCC_VERSION (__GNUC__ * 10000 \
1169 + __GNUC_MINOR__ * 100 \
1170 + __GNUC_PATCHLEVEL__)
1171 #if GCC_VERSION <= 40300
1172 #undef PREFETCH_CARDS
1176 #ifdef HEAVY_STATISTICS
1177 static guint64 stat_optimized_copy;
1178 static guint64 stat_optimized_copy_nursery;
1179 static guint64 stat_optimized_copy_nursery_forwarded;
1180 static guint64 stat_optimized_copy_nursery_pinned;
1181 static guint64 stat_optimized_copy_major;
1182 static guint64 stat_optimized_copy_major_small_fast;
1183 static guint64 stat_optimized_copy_major_small_slow;
1184 static guint64 stat_optimized_copy_major_large;
1185 static guint64 stat_optimized_copy_major_forwarded;
1186 static guint64 stat_optimized_copy_major_small_evacuate;
1187 static guint64 stat_optimized_major_scan;
1188 static guint64 stat_optimized_major_scan_no_refs;
1190 static guint64 stat_drain_prefetch_fills;
1191 static guint64 stat_drain_prefetch_fill_failures;
1192 static guint64 stat_drain_loops;
1195 static void major_scan_object_with_evacuation (GCObject *start, mword desc, SgenGrayQueue *queue);
1197 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_no_evacuation
1198 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_no_evacuation
1199 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_no_evacuation
1200 #include "sgen-marksweep-drain-gray-stack.h"
1202 #define COPY_OR_MARK_WITH_EVACUATION
1203 #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_with_evacuation
1204 #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_with_evacuation
1205 #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_with_evacuation
1206 #include "sgen-marksweep-drain-gray-stack.h"
1209 drain_gray_stack (ScanCopyContext ctx)
1211 gboolean evacuation = FALSE;
1213 for (i = 0; i < num_block_obj_sizes; ++i) {
1214 if (evacuate_block_obj_sizes [i]) {
1221 return drain_gray_stack_with_evacuation (ctx);
1223 return drain_gray_stack_no_evacuation (ctx);
1226 #include "sgen-marksweep-scan-object-concurrent.h"
1229 major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue)
1231 major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1235 major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *queue)
1237 major_copy_or_mark_object_concurrent (ptr, *ptr, queue);
1241 major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue)
1243 major_copy_or_mark_object_no_evacuation (ptr, *ptr, queue);
1247 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1249 void **entry, **end;
1250 int last_index = -1;
1252 if (first_entry == last_entry)
1255 block->has_pinned = TRUE;
1257 entry = sgen_pinning_get_entry (first_entry);
1258 end = sgen_pinning_get_entry (last_entry);
1260 for (; entry < end; ++entry) {
1261 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1263 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));
1264 if (index == last_index)
1266 obj = MS_BLOCK_OBJ (block, index);
1267 MS_MARK_OBJECT_AND_ENQUEUE_CHECKED (obj, sgen_obj_get_descriptor (obj), block, queue);
1273 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1277 for (obj_index = 0; obj_index < count; ++obj_index) {
1279 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1281 MS_CALC_MARK_BIT (word, bit, obj);
1282 if (MS_MARK_BIT (block, word, bit)) {
1283 SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1285 /* an unmarked object */
1286 if (MS_OBJ_ALLOCED (obj, block)) {
1288 * FIXME: Merge consecutive
1289 * slots for lower reporting
1290 * overhead. Maybe memset
1291 * will also benefit?
1293 binary_protocol_empty (obj, obj_size);
1294 memset (obj, 0, obj_size);
1296 *(void**)obj = block->free_list;
1297 block->free_list = obj;
1302 static inline gboolean
1303 try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1305 gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
1306 gboolean success = old_state == expected_state;
1308 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
1313 set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1315 SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set");
1316 block->state = new_state;
1320 * If `block` needs sweeping, sweep it and return TRUE. Otherwise return FALSE.
1322 * Sweeping means iterating through the block's slots and building the free-list from the
1323 * unmarked ones. They will also be zeroed. The mark bits will be reset.
1326 sweep_block (MSBlockInfo *block)
1329 void *reversed = NULL;
1332 switch (block->state) {
1333 case BLOCK_STATE_SWEPT:
1335 case BLOCK_STATE_MARKING:
1336 case BLOCK_STATE_CHECKING:
1337 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1339 case BLOCK_STATE_SWEEPING:
1340 /* FIXME: Do this more elegantly */
1343 case BLOCK_STATE_NEED_SWEEPING:
1344 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1348 SGEN_ASSERT (0, FALSE, "Illegal block state");
1351 SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1353 count = MS_BLOCK_FREE / block->obj_size;
1355 block->free_list = NULL;
1357 /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1358 // FIXME: Add more sizes
1359 switch (block->obj_size) {
1361 sweep_block_for_size (block, count, 16);
1364 sweep_block_for_size (block, count, block->obj_size);
1368 /* reset mark bits */
1369 memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1371 /* Reverse free list so that it's in address order */
1373 while (block->free_list) {
1374 void *next = *(void**)block->free_list;
1375 *(void**)block->free_list = reversed;
1376 reversed = block->free_list;
1377 block->free_list = next;
1379 block->free_list = reversed;
1381 mono_memory_write_barrier ();
1383 set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1394 if (sizeof (mword) == sizeof (unsigned long))
1395 count += __builtin_popcountl (d);
1397 count += __builtin_popcount (d);
1407 /* statistics for evacuation */
1408 static size_t *sweep_slots_available;
1409 static size_t *sweep_slots_used;
1410 static size_t *sweep_num_blocks;
1412 static volatile size_t num_major_sections_before_sweep;
1413 static volatile size_t num_major_sections_freed_in_sweep;
1420 for (i = 0; i < num_block_obj_sizes; ++i)
1421 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1423 /* clear all the free lists */
1424 for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1425 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1427 for (j = 0; j < num_block_obj_sizes; ++j)
1428 free_blocks [j] = NULL;
1432 static void sweep_finish (void);
1435 * If `wait` is TRUE and the block is currently being checked, this function will wait until
1436 * the checking has finished.
1438 * Returns whether the block is still there. If `wait` is FALSE, the return value will not
1439 * be correct, i.e. must not be used.
1442 ensure_block_is_checked_for_sweeping (int block_index, gboolean wait, gboolean *have_checked)
1445 gboolean have_live = FALSE;
1446 gboolean have_free = FALSE;
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 *)&allocated_blocks.data [block_index];
1463 if (BLOCK_IS_TAGGED_CHECKING (tagged_block)) {
1466 /* FIXME: do this more elegantly */
1471 if (SGEN_CAS_PTR (&allocated_blocks.data [block_index], 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 sgen_card_table_free_mod_union (block->cardtable_mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1512 block->cardtable_mod_union = NULL;
1515 /* Count marked objects in the block */
1516 for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1517 nused += bitcount (block->mark_words [i]);
1525 int obj_size_index = block->obj_size_index;
1526 gboolean has_pinned = block->has_pinned;
1528 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1531 * FIXME: Go straight to SWEPT if there are no free slots. We need
1532 * to set the free slot list to NULL, though, and maybe update some
1536 sweep_block (block);
1539 ++sweep_num_blocks [obj_size_index];
1540 sweep_slots_used [obj_size_index] += nused;
1541 sweep_slots_available [obj_size_index] += count;
1545 * If there are free slots in the block, add
1546 * the block to the corresponding free list.
1549 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1552 SGEN_ASSERT (6, block->free_list, "How do we not have a free list when there are free slots?");
1554 add_free_block (free_blocks, obj_size_index, block);
1557 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1558 update_heap_boundaries_for_block (block);
1561 * Blocks without live objects are removed from the
1562 * block list and freed.
1564 SGEN_ASSERT (6, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1565 SGEN_ASSERT (6, allocated_blocks.data [block_index] == BLOCK_TAG_CHECKING (tagged_block), "How did the block move?");
1567 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1568 ms_free_block (block);
1570 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1572 tagged_block = NULL;
1576 allocated_blocks.data [block_index] = tagged_block;
1577 return !!tagged_block;
1581 sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1584 int num_blocks = num_major_sections_before_sweep;
1586 SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1587 SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1590 * We traverse the block array from high to low. Nursery collections will have to
1591 * cooperate with the sweep thread to finish sweeping, and they will traverse from
1592 * low to high, to avoid constantly colliding on the same blocks.
1594 for (block_index = num_blocks - 1; block_index >= 0; --block_index) {
1595 gboolean have_checked;
1598 * The block might have been freed by another thread doing some checking
1601 if (!ensure_block_is_checked_for_sweeping (block_index, TRUE, &have_checked))
1602 ++num_major_sections_freed_in_sweep;
1605 while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) {
1607 * The main GC thread is currently iterating over the block array to help us
1608 * finish the sweep. We have already finished, but we don't want to mess up
1609 * that iteration, so we just wait for it.
1614 if (SGEN_MAX_ASSERT_LEVEL >= 6) {
1615 for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1616 MSBlockInfo *block = BLOCK_UNTAG (allocated_blocks.data [block_index]);
1617 SGEN_ASSERT (6, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1621 sgen_pointer_queue_remove_nulls (&allocated_blocks);
1631 mword total_evacuate_heap = 0;
1632 mword total_evacuate_saved = 0;
1635 for (i = 0; i < num_block_obj_sizes; ++i) {
1636 float usage = (float)sweep_slots_used [i] / (float)sweep_slots_available [i];
1637 if (sweep_num_blocks [i] > 5 && usage < evacuation_threshold) {
1638 evacuate_block_obj_sizes [i] = TRUE;
1640 g_print ("slot size %d - %d of %d used\n",
1641 block_obj_sizes [i], slots_used [i], slots_available [i]);
1644 evacuate_block_obj_sizes [i] = FALSE;
1647 mword total_bytes = block_obj_sizes [i] * sweep_slots_available [i];
1648 total_evacuate_heap += total_bytes;
1649 if (evacuate_block_obj_sizes [i])
1650 total_evacuate_saved += total_bytes - block_obj_sizes [i] * sweep_slots_used [i];
1654 want_evacuation = (float)total_evacuate_saved / (float)total_evacuate_heap > (1 - concurrent_evacuation_threshold);
1656 set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1662 set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1666 SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?");
1668 num_major_sections_before_sweep = num_major_sections;
1669 num_major_sections_freed_in_sweep = 0;
1671 SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
1672 if (concurrent_sweep) {
1673 sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
1674 sgen_thread_pool_job_enqueue (sweep_job);
1676 sweep_job_func (NULL, NULL);
1681 major_have_swept (void)
1683 return sweep_state == SWEEP_STATE_SWEPT;
1686 static int count_pinned_ref;
1687 static int count_pinned_nonref;
1688 static int count_nonpinned_ref;
1689 static int count_nonpinned_nonref;
1692 count_nonpinned_callback (GCObject *obj, size_t size, void *data)
1694 GCVTable vtable = LOAD_VTABLE (obj);
1696 if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1697 ++count_nonpinned_ref;
1699 ++count_nonpinned_nonref;
1703 count_pinned_callback (GCObject *obj, size_t size, void *data)
1705 GCVTable vtable = LOAD_VTABLE (obj);
1707 if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1710 ++count_pinned_nonref;
1713 static G_GNUC_UNUSED void
1714 count_ref_nonref_objs (void)
1718 count_pinned_ref = 0;
1719 count_pinned_nonref = 0;
1720 count_nonpinned_ref = 0;
1721 count_nonpinned_nonref = 0;
1723 major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1724 major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1726 total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1728 g_print ("ref: %d pinned %d non-pinned non-ref: %d pinned %d non-pinned -- %.1f\n",
1729 count_pinned_ref, count_nonpinned_ref,
1730 count_pinned_nonref, count_nonpinned_nonref,
1731 (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1735 ms_calculate_block_obj_sizes (double factor, int *arr)
1742 * Have every possible slot size starting with the minimal
1743 * object size up to and including four times that size. Then
1744 * proceed by increasing geometrically with the given factor.
1747 for (int size = SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size <= 4 * SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size += SGEN_ALLOC_ALIGN) {
1749 arr [num_sizes] = size;
1753 target_size = (double)last_size;
1756 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1757 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1759 if (size != last_size) {
1761 arr [num_sizes] = size;
1766 target_size *= factor;
1767 } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1772 /* only valid during minor collections */
1773 static mword old_num_major_sections;
1776 major_start_nursery_collection (void)
1778 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1779 consistency_check ();
1782 old_num_major_sections = num_major_sections;
1786 major_finish_nursery_collection (void)
1788 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1789 consistency_check ();
1794 major_start_major_collection (void)
1799 major_finish_sweep_checking ();
1802 * Clear the free lists for block sizes where we do evacuation. For those block
1803 * sizes we will have to allocate new blocks.
1805 for (i = 0; i < num_block_obj_sizes; ++i) {
1806 if (!evacuate_block_obj_sizes [i])
1809 free_block_lists [0][i] = NULL;
1810 free_block_lists [MS_BLOCK_FLAG_REFS][i] = NULL;
1814 binary_protocol_sweep_begin (GENERATION_OLD, TRUE);
1816 /* Sweep all unswept blocks and set them to MARKING */
1817 FOREACH_BLOCK_NO_LOCK (block) {
1819 sweep_block (block);
1820 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1821 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1822 } END_FOREACH_BLOCK_NO_LOCK;
1825 binary_protocol_sweep_end (GENERATION_OLD, TRUE);
1827 set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1831 major_finish_major_collection (ScannedObjectCounts *counts)
1833 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1834 if (binary_protocol_is_enabled ()) {
1835 counts->num_scanned_objects = scanned_objects_list.next_slot;
1837 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1838 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1840 sgen_pointer_queue_clear (&scanned_objects_list);
1845 #if SIZEOF_VOID_P != 8
1847 compare_pointers (const void *va, const void *vb) {
1848 char *a = *(char**)va, *b = *(char**)vb;
1858 * This is called with sweep completed and the world stopped.
1861 major_free_swept_blocks (size_t allowance)
1863 /* FIXME: This is probably too much. It's assuming all objects are small. */
1864 size_t section_reserve = allowance / MS_BLOCK_SIZE;
1866 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1868 #if SIZEOF_VOID_P != 8
1870 int i, num_empty_blocks_orig, num_blocks, arr_length;
1872 void **empty_block_arr;
1873 void **rebuild_next;
1877 * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1878 * a VirtualAlloc ()-ed block.
1883 if (num_empty_blocks <= section_reserve)
1885 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1887 num_empty_blocks_orig = num_empty_blocks;
1888 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1889 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1890 if (!empty_block_arr)
1894 for (block = empty_blocks; block; block = *(void**)block)
1895 empty_block_arr [i++] = block;
1896 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1898 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1901 * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1902 * contiguous ones. If we do, we free them. If that's not enough to get to
1903 * section_reserve, we halve the number of contiguous blocks we're looking
1904 * for and have another go, until we're done with looking for pairs of
1905 * blocks, at which point we give up and go to the fallback.
1907 arr_length = num_empty_blocks_orig;
1908 num_blocks = MS_BLOCK_ALLOC_NUM;
1909 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1914 for (i = 0; i < arr_length; ++i) {
1916 void *block = empty_block_arr [i];
1917 SGEN_ASSERT (6, block, "we're not shifting correctly");
1919 empty_block_arr [dest] = block;
1921 * This is not strictly necessary, but we're
1924 empty_block_arr [i] = NULL;
1933 SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
1935 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
1940 if (d + 1 - first == num_blocks) {
1942 * We found num_blocks contiguous blocks. Free them
1943 * and null their array entries. As an optimization
1944 * we could, instead of nulling the entries, shift
1945 * the following entries over to the left, while
1949 sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
1950 for (j = first; j <= d; ++j)
1951 empty_block_arr [j] = NULL;
1955 num_empty_blocks -= num_blocks;
1957 stat_major_blocks_freed += num_blocks;
1958 if (num_blocks == MS_BLOCK_ALLOC_NUM)
1959 stat_major_blocks_freed_ideal += num_blocks;
1961 stat_major_blocks_freed_less_ideal += num_blocks;
1966 SGEN_ASSERT (6, dest <= i && dest <= arr_length, "array length is off");
1968 SGEN_ASSERT (6, arr_length == num_empty_blocks, "array length is off");
1973 /* rebuild empty_blocks free list */
1974 rebuild_next = (void**)&empty_blocks;
1975 for (i = 0; i < arr_length; ++i) {
1976 void *block = empty_block_arr [i];
1977 SGEN_ASSERT (6, block, "we're missing blocks");
1978 *rebuild_next = block;
1979 rebuild_next = (void**)block;
1981 *rebuild_next = NULL;
1984 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
1987 SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
1991 * This is our threshold. If there's not more empty than used blocks, we won't
1992 * release uncontiguous blocks, in fear of fragmenting the address space.
1994 if (num_empty_blocks <= num_major_sections)
1998 while (num_empty_blocks > section_reserve) {
1999 void *next = *(void**)empty_blocks;
2000 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
2001 empty_blocks = next;
2003 * Needs not be atomic because this is running
2008 ++stat_major_blocks_freed;
2009 #if SIZEOF_VOID_P != 8
2010 ++stat_major_blocks_freed_individual;
2016 major_pin_objects (SgenGrayQueue *queue)
2020 FOREACH_BLOCK_NO_LOCK (block) {
2021 size_t first_entry, last_entry;
2022 SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
2023 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
2024 &first_entry, &last_entry);
2025 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
2026 } END_FOREACH_BLOCK_NO_LOCK;
2030 major_init_to_space (void)
2035 major_report_pinned_memory_usage (void)
2037 g_assert_not_reached ();
2041 major_get_used_size (void)
2047 * We're holding the GC lock, but the sweep thread might be running. Make sure it's
2048 * finished, then we can iterate over the block array.
2050 major_finish_sweep_checking ();
2052 FOREACH_BLOCK_NO_LOCK_CONDITION (TRUE, block) {
2053 int count = MS_BLOCK_FREE / block->obj_size;
2055 size += count * block->obj_size;
2056 for (iter = block->free_list; iter; iter = (void**)*iter)
2057 size -= block->obj_size;
2058 } END_FOREACH_BLOCK_NO_LOCK;
2063 /* FIXME: return number of bytes, not of sections */
2065 get_num_major_sections (void)
2067 return num_major_sections;
2071 * Returns the number of bytes in blocks that were present when the last sweep was
2072 * initiated, and were not freed during the sweep. They are the basis for calculating the
2076 get_bytes_survived_last_sweep (void)
2078 SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2079 return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
2083 major_handle_gc_param (const char *opt)
2085 if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2086 const char *arg = strchr (opt, '=') + 1;
2087 int percentage = atoi (arg);
2088 if (percentage < 0 || percentage > 100) {
2089 fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2092 evacuation_threshold = (float)percentage / 100.0f;
2094 } else if (!strcmp (opt, "lazy-sweep")) {
2097 } else if (!strcmp (opt, "no-lazy-sweep")) {
2100 } else if (!strcmp (opt, "concurrent-sweep")) {
2101 concurrent_sweep = TRUE;
2103 } else if (!strcmp (opt, "no-concurrent-sweep")) {
2104 concurrent_sweep = FALSE;
2112 major_print_gc_param_usage (void)
2116 " evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2117 " (no-)lazy-sweep\n"
2118 " (no-)concurrent-sweep\n"
2123 * This callback is used to clear cards, move cards to the shadow table and do counting.
2126 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
2129 gboolean has_references;
2131 major_finish_sweep_checking ();
2132 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2134 callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2135 } END_FOREACH_BLOCK_NO_LOCK;
2138 #ifdef HEAVY_STATISTICS
2139 extern guint64 marked_cards;
2140 extern guint64 scanned_cards;
2141 extern guint64 scanned_objects;
2142 extern guint64 remarked_cards;
2145 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2147 * MS blocks are 16K aligned.
2148 * Cardtables are 4K aligned, at least.
2149 * This means that the cardtable of a given block is 32 bytes aligned.
2152 initial_skip_card (guint8 *card_data)
2154 mword *cards = (mword*)card_data;
2157 for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2163 if (i == CARD_WORDS_PER_BLOCK)
2164 return card_data + CARDS_PER_BLOCK;
2166 #if defined(__i386__) && defined(__GNUC__)
2167 return card_data + i * 4 + (__builtin_ffs (card) - 1) / 8;
2168 #elif defined(__x86_64__) && defined(__GNUC__)
2169 return card_data + i * 8 + (__builtin_ffsll (card) - 1) / 8;
2170 #elif defined(__s390x__) && defined(__GNUC__)
2171 return card_data + i * 8 + (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2173 for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2175 return &card_data [i];
2181 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2182 #define MS_BLOCK_OBJ_FAST(b,os,i) ((b) + MS_BLOCK_SKIP + (os) * (i))
2183 #define MS_OBJ_ALLOCED_FAST(o,b) (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2186 scan_card_table_for_block (MSBlockInfo *block, gboolean mod_union, ScanCopyContext ctx)
2188 SgenGrayQueue *queue = ctx.queue;
2189 ScanObjectFunc scan_func = ctx.ops->scan_object;
2190 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2191 guint8 cards_copy [CARDS_PER_BLOCK];
2193 gboolean small_objects;
2196 guint8 *card_data, *card_base;
2197 guint8 *card_data_end;
2198 char *scan_front = NULL;
2200 block_obj_size = block->obj_size;
2201 small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2203 block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2206 * This is safe in face of card aliasing for the following reason:
2208 * Major blocks are 16k aligned, or 32 cards aligned.
2209 * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2210 * sizes, they won't overflow the cardtable overlap modulus.
2213 card_data = card_base = block->cardtable_mod_union;
2215 * This happens when the nursery collection that precedes finishing
2216 * the concurrent collection allocates new major blocks.
2221 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2222 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2224 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2226 card_data = card_base = cards_copy;
2229 card_data_end = card_data + CARDS_PER_BLOCK;
2231 card_data += MS_BLOCK_SKIP >> CARD_BITS;
2233 card_data = initial_skip_card (card_data);
2234 while (card_data < card_data_end) {
2235 size_t card_index, first_object_index;
2238 char *first_obj, *obj;
2240 HEAVY_STAT (++scanned_cards);
2247 card_index = card_data - card_base;
2248 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2249 end = start + CARD_SIZE_IN_BYTES;
2251 if (!block_is_swept_or_marking (block))
2252 sweep_block (block);
2254 HEAVY_STAT (++marked_cards);
2257 sgen_card_table_prepare_card_for_scanning (card_data);
2260 * If the card we're looking at starts at or in the block header, we
2261 * must start at the first object in the block, without calculating
2262 * the index of the object we're hypothetically starting at, because
2263 * it would be negative.
2265 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2266 first_object_index = 0;
2268 first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2270 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2272 binary_protocol_card_scan (first_obj, end - first_obj);
2275 if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2279 /* FIXME: do this more efficiently */
2281 MS_CALC_MARK_BIT (w, b, obj);
2282 if (!MS_MARK_BIT (block, w, b))
2286 GCObject *object = (GCObject*)obj;
2288 if (small_objects) {
2289 HEAVY_STAT (++scanned_objects);
2290 scan_func (object, sgen_obj_get_descriptor (object), queue);
2292 size_t offset = sgen_card_table_get_card_offset (obj, block_start);
2293 sgen_cardtable_scan_object (object, block_obj_size, card_base + offset, mod_union, ctx);
2296 obj += block_obj_size;
2297 g_assert (scan_front <= obj);
2301 HEAVY_STAT (if (*card_data) ++remarked_cards);
2306 card_data = card_base + sgen_card_table_get_card_offset (obj, block_start);
2311 major_scan_card_table (gboolean mod_union, ScanCopyContext ctx)
2314 gboolean has_references;
2316 if (!concurrent_mark)
2317 g_assert (!mod_union);
2319 major_finish_sweep_checking ();
2320 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2321 #ifdef PREFETCH_CARDS
2322 int prefetch_index = __index + 6;
2323 if (prefetch_index < allocated_blocks.next_slot) {
2324 MSBlockInfo *prefetch_block = BLOCK_UNTAG (allocated_blocks.data [prefetch_index]);
2325 guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2326 PREFETCH_READ (prefetch_block);
2327 PREFETCH_WRITE (prefetch_cards);
2328 PREFETCH_WRITE (prefetch_cards + 32);
2332 if (!has_references)
2335 scan_card_table_for_block (block, mod_union, ctx);
2336 } END_FOREACH_BLOCK_NO_LOCK;
2340 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2343 gboolean has_references;
2344 long long total_cards = 0;
2345 long long marked_cards = 0;
2347 if (sweep_in_progress ()) {
2348 *num_total_cards = -1;
2349 *num_marked_cards = -1;
2353 FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2354 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2357 if (!has_references)
2360 total_cards += CARDS_PER_BLOCK;
2361 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2365 } END_FOREACH_BLOCK_NO_LOCK;
2367 *num_total_cards = total_cards;
2368 *num_marked_cards = marked_cards;
2372 update_cardtable_mod_union (void)
2376 FOREACH_BLOCK_NO_LOCK (block) {
2378 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
2379 sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2380 SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2381 } END_FOREACH_BLOCK_NO_LOCK;
2384 #undef pthread_create
2387 post_param_init (SgenMajorCollector *collector)
2389 collector->sweeps_lazily = lazy_sweep;
2390 collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
2394 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
2398 sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2400 num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2401 block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2402 ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2404 evacuate_block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2405 for (i = 0; i < num_block_obj_sizes; ++i)
2406 evacuate_block_obj_sizes [i] = FALSE;
2408 sweep_slots_available = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2409 sweep_slots_used = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2410 sweep_num_blocks = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2415 g_print ("block object sizes:\n");
2416 for (i = 0; i < num_block_obj_sizes; ++i)
2417 g_print ("%d\n", block_obj_sizes [i]);
2421 for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2422 free_block_lists [i] = sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2424 for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2425 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2426 for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2427 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2429 mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2430 mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2431 mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2432 mono_counters_register ("# major objects evacuated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_objects_evacuated);
2433 #if SIZEOF_VOID_P != 8
2434 mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2435 mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2436 mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2437 mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2440 collector->section_size = MAJOR_SECTION_SIZE;
2442 concurrent_mark = is_concurrent;
2443 collector->is_concurrent = is_concurrent;
2444 collector->needs_thread_pool = is_concurrent || concurrent_sweep;
2446 collector->want_synchronous_collection = &want_evacuation;
2448 collector->want_synchronous_collection = NULL;
2449 collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2450 collector->supports_cardtable = TRUE;
2452 collector->alloc_heap = major_alloc_heap;
2453 collector->is_object_live = major_is_object_live;
2454 collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2455 collector->alloc_degraded = major_alloc_degraded;
2457 collector->alloc_object = major_alloc_object;
2458 collector->free_pinned_object = free_pinned_object;
2459 collector->iterate_objects = major_iterate_objects;
2460 collector->free_non_pinned_object = major_free_non_pinned_object;
2461 collector->pin_objects = major_pin_objects;
2462 collector->pin_major_object = pin_major_object;
2463 collector->scan_card_table = major_scan_card_table;
2464 collector->iterate_live_block_ranges = (void*)(void*) major_iterate_live_block_ranges;
2465 if (is_concurrent) {
2466 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2467 collector->get_cardtable_mod_union_for_reference = major_get_cardtable_mod_union_for_reference;
2469 collector->init_to_space = major_init_to_space;
2470 collector->sweep = major_sweep;
2471 collector->have_swept = major_have_swept;
2472 collector->finish_sweeping = major_finish_sweep_checking;
2473 collector->free_swept_blocks = major_free_swept_blocks;
2474 collector->check_scan_starts = major_check_scan_starts;
2475 collector->dump_heap = major_dump_heap;
2476 collector->get_used_size = major_get_used_size;
2477 collector->start_nursery_collection = major_start_nursery_collection;
2478 collector->finish_nursery_collection = major_finish_nursery_collection;
2479 collector->start_major_collection = major_start_major_collection;
2480 collector->finish_major_collection = major_finish_major_collection;
2481 collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2482 collector->ptr_is_from_pinned_alloc = ptr_is_from_pinned_alloc;
2483 collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2484 collector->get_num_major_sections = get_num_major_sections;
2485 collector->get_bytes_survived_last_sweep = get_bytes_survived_last_sweep;
2486 collector->handle_gc_param = major_handle_gc_param;
2487 collector->print_gc_param_usage = major_print_gc_param_usage;
2488 collector->post_param_init = post_param_init;
2489 collector->is_valid_object = major_is_valid_object;
2490 collector->describe_pointer = major_describe_pointer;
2491 collector->count_cards = major_count_cards;
2493 collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
2494 collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
2495 if (is_concurrent) {
2496 collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2497 collector->major_ops_concurrent_start.scan_object = major_scan_object_no_mark_concurrent_start;
2499 collector->major_ops_concurrent.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2500 collector->major_ops_concurrent.scan_object = major_scan_object_no_mark_concurrent;
2502 collector->major_ops_concurrent_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_finish_canonical;
2503 collector->major_ops_concurrent_finish.scan_object = major_scan_object_no_evacuation;
2504 collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_concurrent_finish;
2507 #if !defined (FIXED_HEAP) && !defined (SGEN_PARALLEL_MARK)
2509 collector->drain_gray_stack = drain_gray_stack;
2511 #ifdef HEAVY_STATISTICS
2512 mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2513 mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2514 mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2515 mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2516 mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2517 mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2518 mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2519 mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2520 mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2521 mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2523 mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2524 mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2525 mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2529 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2530 mono_mutex_init (&scanned_objects_list_lock);
2533 SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2535 /*cardtable requires major pages to be 8 cards aligned*/
2536 g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2540 sgen_marksweep_init (SgenMajorCollector *collector)
2542 sgen_marksweep_init_internal (collector, FALSE);
2546 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2548 sgen_marksweep_init_internal (collector, TRUE);