mono/metadata/sgen-marksweep.c

   1 /*
   2  * sgen-marksweep.c: The Mark & Sweep major collector.
   3  *
   4  * Author:
   5  *      Mark Probst <mark.probst@gmail.com>
   6  *
   7  * Copyright 2009-2010 Novell, Inc.
   8  * Copyright (C) 2012 Xamarin Inc
   9  *
  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;
  13  *
  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.
  18  *
  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.
  22  */
  23
  24 #include "config.h"
  25
  26 #ifdef HAVE_SGEN_GC
  27
  28 #include <math.h>
  29 #include <errno.h>
  30
  31 #include "utils/mono-counters.h"
  32 #include "utils/mono-semaphore.h"
  33 #include "utils/mono-time.h"
  34 #include "metadata/object-internals.h"
  35 #include "metadata/profiler-private.h"
  36
  37 #include "metadata/sgen-gc.h"
  38 #include "metadata/sgen-protocol.h"
  39 #include "metadata/sgen-cardtable.h"
  40 #include "metadata/sgen-memory-governor.h"
  41 #include "metadata/sgen-layout-stats.h"
  42 #include "metadata/gc-internal.h"
  43 #include "metadata/sgen-pointer-queue.h"
  44 #include "metadata/sgen-pinning.h"
  45 #include "metadata/sgen-workers.h"
  46
  47 #define SGEN_HAVE_CONCURRENT_MARK
  48
  49 #define MS_BLOCK_SIZE   (16*1024)
  50 #define MS_BLOCK_SIZE_SHIFT     14
  51 #define MAJOR_SECTION_SIZE      MS_BLOCK_SIZE
  52 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
  53
  54 /*
  55  * Don't allocate single blocks, but alloc a contingent of this many
  56  * blocks in one swoop.  This must be a power of two.
  57  */
  58 #define MS_BLOCK_ALLOC_NUM      32
  59
  60 /*
  61  * Number of bytes before the first object in a block.  At the start
  62  * of a block is the MSBlockHeader, then opional padding, then come
  63  * the objects, so this must be >= sizeof (MSBlockHeader).
  64  */
  65 #define MS_BLOCK_SKIP   16
  66
  67 #define MS_BLOCK_FREE   (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
  68
  69 #define MS_NUM_MARK_WORDS       ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
  70
  71 #if SGEN_MAX_SMALL_OBJ_SIZE > MS_BLOCK_FREE / 2
  72 #error MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2
  73 #endif
  74
  75 typedef struct _MSBlockInfo MSBlockInfo;
  76 struct _MSBlockInfo {
  77         int obj_size;
  78         int obj_size_index;
  79         unsigned int pinned : 1;
  80         unsigned int has_references : 1;
  81         unsigned int has_pinned : 1;    /* means cannot evacuate */
  82         unsigned int is_to_space : 1;
  83         unsigned int swept : 1;
  84         char *block;
  85         void **free_list;
  86         MSBlockInfo *next_free;
  87         size_t pin_queue_first_entry;
  88         size_t pin_queue_last_entry;
  89 #ifdef SGEN_HAVE_CONCURRENT_MARK
  90         guint8 *cardtable_mod_union;
  91 #endif
  92         mword mark_words [MS_NUM_MARK_WORDS];
  93 };
  94
  95 #define MS_BLOCK_FOR_BLOCK_INFO(b)      ((b)->block)
  96
  97 #define MS_BLOCK_OBJ(b,i)               (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i))
  98 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size)             (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
  99 #define MS_BLOCK_DATA_FOR_OBJ(o)        ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
 100
 101 typedef struct {
 102         MSBlockInfo *info;
 103 } MSBlockHeader;
 104
 105 #define MS_BLOCK_FOR_OBJ(o)             (((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
 106
 107 /* object index will always be small */
 108 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
 109
 110 //casting to int is fine since blocks are 32k
 111 #define MS_CALC_MARK_BIT(w,b,o)         do {                            \
 112                 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
 113                 if (sizeof (mword) == 4) {                              \
 114                         (w) = i >> 5;                                   \
 115                         (b) = i & 31;                                   \
 116                 } else {                                                \
 117                         (w) = i >> 6;                                   \
 118                         (b) = i & 63;                                   \
 119                 }                                                       \
 120         } while (0)
 121
 122 #define MS_MARK_BIT(bl,w,b)     ((bl)->mark_words [(w)] & (ONE_P << (b)))
 123 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
 124
 125 #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))
 126
 127 #define MS_BLOCK_OBJ_SIZE_FACTOR        (sqrt (2.0))
 128
 129 /*
 130  * This way we can lookup block object size indexes for sizes up to
 131  * 256 bytes with a single load.
 132  */
 133 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES      32
 134
 135 static int *block_obj_sizes;
 136 static int num_block_obj_sizes;
 137 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
 138
 139 #define MS_BLOCK_FLAG_PINNED    1
 140 #define MS_BLOCK_FLAG_REFS      2
 141
 142 #define MS_BLOCK_TYPE_MAX       4
 143
 144 static gboolean *evacuate_block_obj_sizes;
 145 static float evacuation_threshold = 0.666f;
 146 #ifdef SGEN_HAVE_CONCURRENT_MARK
 147 static float concurrent_evacuation_threshold = 0.666f;
 148 static gboolean want_evacuation = FALSE;
 149 #endif
 150
 151 static gboolean lazy_sweep = TRUE;
 152 static gboolean have_swept;
 153
 154 #ifdef SGEN_HAVE_CONCURRENT_MARK
 155 static gboolean concurrent_mark;
 156 #endif
 157
 158 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl)      SGEN_POINTER_IS_TAGGED_1 ((bl))
 159 #define BLOCK_TAG_HAS_REFERENCES(bl)            SGEN_POINTER_TAG_1 ((bl))
 160 #define BLOCK_UNTAG_HAS_REFERENCES(bl)          SGEN_POINTER_UNTAG_1 ((bl))
 161
 162 #define BLOCK_TAG(bl)   ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
 163
 164 /* all allocated blocks in the system */
 165 static SgenPointerQueue allocated_blocks;
 166
 167 /* non-allocated block free-list */
 168 static void *empty_blocks = NULL;
 169 static size_t num_empty_blocks = 0;
 170
 171 #define FOREACH_BLOCK(bl)       { size_t __index; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [__index]);
 172 #define FOREACH_BLOCK_HAS_REFERENCES(bl,hr)     { size_t __index; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = allocated_blocks.data [__index]; (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl)); (bl) = BLOCK_UNTAG_HAS_REFERENCES ((bl));
 173 #define END_FOREACH_BLOCK       } }
 174 #define DELETE_BLOCK_IN_FOREACH()       (allocated_blocks.data [__index] = NULL)
 175
 176 static size_t num_major_sections = 0;
 177 /* one free block list for each block object size */
 178 static MSBlockInfo **free_block_lists [MS_BLOCK_TYPE_MAX];
 179
 180 static long long stat_major_blocks_alloced = 0;
 181 static long long stat_major_blocks_freed = 0;
 182 static long long stat_major_blocks_lazy_swept = 0;
 183 static long long stat_major_objects_evacuated = 0;
 184
 185 #if SIZEOF_VOID_P != 8
 186 static long long stat_major_blocks_freed_ideal = 0;
 187 static long long stat_major_blocks_freed_less_ideal = 0;
 188 static long long stat_major_blocks_freed_individual = 0;
 189 static long long stat_major_blocks_alloced_less_ideal = 0;
 190 #endif
 191
 192 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
 193 static long long num_major_objects_marked = 0;
 194 #define INC_NUM_MAJOR_OBJECTS_MARKED()  (++num_major_objects_marked)
 195 #else
 196 #define INC_NUM_MAJOR_OBJECTS_MARKED()
 197 #endif
 198
 199 static void
 200 sweep_block (MSBlockInfo *block, gboolean during_major_collection);
 201
 202 static int
 203 ms_find_block_obj_size_index (size_t size)
 204 {
 205         int i;
 206         SGEN_ASSERT (9, size <= SGEN_MAX_SMALL_OBJ_SIZE, "size %d is bigger than max small object size %d", size, SGEN_MAX_SMALL_OBJ_SIZE);
 207         for (i = 0; i < num_block_obj_sizes; ++i)
 208                 if (block_obj_sizes [i] >= size)
 209                         return i;
 210         g_error ("no object of size %d\n", size);
 211 }
 212
 213 #define FREE_BLOCKS_FROM(lists,p,r)     (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
 214 #define FREE_BLOCKS(p,r)                (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
 215
 216 #define MS_BLOCK_OBJ_SIZE_INDEX(s)                              \
 217         (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ?      \
 218          fast_block_obj_size_indexes [((s)+7)>>3] :             \
 219          ms_find_block_obj_size_index ((s)))
 220
 221 static void*
 222 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
 223 {
 224         char *start;
 225         if (nursery_align)
 226                 start = sgen_alloc_os_memory_aligned (nursery_size, nursery_align, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
 227         else
 228                 start = sgen_alloc_os_memory (nursery_size, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
 229
 230         return start;
 231 }
 232
 233 static void
 234 update_heap_boundaries_for_block (MSBlockInfo *block)
 235 {
 236         sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
 237 }
 238
 239 static void*
 240 ms_get_empty_block (void)
 241 {
 242         char *p;
 243         int i;
 244         void *block, *empty, *next;
 245
 246  retry:
 247         if (!empty_blocks) {
 248                 /*
 249                  * We try allocating MS_BLOCK_ALLOC_NUM blocks first.  If that's
 250                  * unsuccessful, we halve the number of blocks and try again, until we're at
 251                  * 1.  If that doesn't work, either, we assert.
 252                  */
 253                 int alloc_num = MS_BLOCK_ALLOC_NUM;
 254                 for (;;) {
 255                         p = sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE,
 256                                         alloc_num == 1 ? "major heap section" : NULL);
 257                         if (p)
 258                                 break;
 259                         alloc_num >>= 1;
 260                 }
 261
 262                 for (i = 0; i < alloc_num; ++i) {
 263                         block = p;
 264                         /*
 265                          * We do the free list update one after the
 266                          * other so that other threads can use the new
 267                          * blocks as quickly as possible.
 268                          */
 269                         do {
 270                                 empty = empty_blocks;
 271                                 *(void**)block = empty;
 272                         } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
 273                         p += MS_BLOCK_SIZE;
 274                 }
 275
 276                 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
 277
 278                 stat_major_blocks_alloced += alloc_num;
 279 #if SIZEOF_VOID_P != 8
 280                 if (alloc_num != MS_BLOCK_ALLOC_NUM)
 281                         stat_major_blocks_alloced_less_ideal += alloc_num;
 282 #endif
 283         }
 284
 285         do {
 286                 empty = empty_blocks;
 287                 if (!empty)
 288                         goto retry;
 289                 block = empty;
 290                 next = *(void**)block;
 291         } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
 292
 293         SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
 294
 295         *(void**)block = NULL;
 296
 297         g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
 298
 299         return block;
 300 }
 301
 302 static void
 303 ms_free_block (void *block)
 304 {
 305         void *empty;
 306
 307         sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
 308         memset (block, 0, MS_BLOCK_SIZE);
 309
 310         do {
 311                 empty = empty_blocks;
 312                 *(void**)block = empty;
 313         } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
 314
 315         SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
 316 }
 317
 318 //#define MARKSWEEP_CONSISTENCY_CHECK
 319
 320 #ifdef MARKSWEEP_CONSISTENCY_CHECK
 321 static void
 322 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
 323 {
 324         MSBlockInfo *b;
 325
 326         for (; block; block = block->next_free) {
 327                 g_assert (block->obj_size == size);
 328                 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
 329
 330                 /* blocks in the free lists must have at least
 331                    one free slot */
 332                 if (block->swept)
 333                         g_assert (block->free_list);
 334
 335                 /* the block must be in the allocated_blocks array */
 336                 g_assert (sgen_pointer_queue_find (&allocated_blocks, BLOCK_TAG (block)) != (size_t)-1);
 337         }
 338 }
 339
 340 static void
 341 check_empty_blocks (void)
 342 {
 343         void *p;
 344         size_t i = 0;
 345         for (p = empty_blocks; p; p = *(void**)p)
 346                 ++i;
 347         g_assert (i == num_empty_blocks);
 348 }
 349
 350 static void
 351 consistency_check (void)
 352 {
 353         MSBlockInfo *block;
 354         int i;
 355
 356         /* check all blocks */
 357         FOREACH_BLOCK (block) {
 358                 int count = MS_BLOCK_FREE / block->obj_size;
 359                 int num_free = 0;
 360                 void **free;
 361
 362                 /* check block header */
 363                 g_assert (((MSBlockHeader*)block->block)->info == block);
 364
 365                 /* count number of free slots */
 366                 for (i = 0; i < count; ++i) {
 367                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 368                         if (!MS_OBJ_ALLOCED (obj, block))
 369                                 ++num_free;
 370                 }
 371
 372                 /* check free list */
 373                 for (free = block->free_list; free; free = (void**)*free) {
 374                         g_assert (MS_BLOCK_FOR_OBJ (free) == block);
 375                         --num_free;
 376                 }
 377                 g_assert (num_free == 0);
 378
 379                 /* check all mark words are zero */
 380                 if (block->swept) {
 381                         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
 382                                 g_assert (block->mark_words [i] == 0);
 383                 }
 384         } END_FOREACH_BLOCK;
 385
 386         /* check free blocks */
 387         for (i = 0; i < num_block_obj_sizes; ++i) {
 388                 int j;
 389                 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
 390                         check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
 391         }
 392
 393         check_empty_blocks ();
 394 }
 395 #endif
 396
 397 static gboolean
 398 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
 399 {
 400         int size = block_obj_sizes [size_index];
 401         int count = MS_BLOCK_FREE / size;
 402         MSBlockInfo *info;
 403         MSBlockHeader *header;
 404         MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, has_references);
 405         char *obj_start;
 406         int i;
 407
 408         if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
 409                 return FALSE;
 410
 411         info = sgen_alloc_internal (INTERNAL_MEM_MS_BLOCK_INFO);
 412
 413         SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
 414
 415         info->obj_size = size;
 416         info->obj_size_index = size_index;
 417         info->pinned = pinned;
 418         info->has_references = has_references;
 419         info->has_pinned = pinned;
 420         /*
 421          * Blocks that are to-space are not evacuated from.  During an major collection
 422          * blocks are allocated for two reasons: evacuating objects from the nursery and
 423          * evacuating them from major blocks marked for evacuation.  In both cases we don't
 424          * want further evacuation.
 425          */
 426         info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD);
 427         info->swept = 1;
 428         info->block = ms_get_empty_block ();
 429
 430         header = (MSBlockHeader*) info->block;
 431         header->info = info;
 432 #ifdef SGEN_HAVE_CONCURRENT_MARK
 433         info->cardtable_mod_union = NULL;
 434 #endif
 435
 436         update_heap_boundaries_for_block (info);
 437
 438         /* build free list */
 439         obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
 440         info->free_list = (void**)obj_start;
 441         /* we're skipping the last one - it must be nulled */
 442         for (i = 0; i < count - 1; ++i) {
 443                 char *next_obj_start = obj_start + size;
 444                 *(void**)obj_start = next_obj_start;
 445                 obj_start = next_obj_start;
 446         }
 447         /* the last one */
 448         *(void**)obj_start = NULL;
 449
 450         info->next_free = free_blocks [size_index];
 451         free_blocks [size_index] = info;
 452
 453         sgen_pointer_queue_add (&allocated_blocks, BLOCK_TAG (info));
 454
 455         ++num_major_sections;
 456         return TRUE;
 457 }
 458
 459 static gboolean
 460 obj_is_from_pinned_alloc (char *ptr)
 461 {
 462         MSBlockInfo *block;
 463
 464         FOREACH_BLOCK (block) {
 465                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE)
 466                         return block->pinned;
 467         } END_FOREACH_BLOCK;
 468         return FALSE;
 469 }
 470
 471 static void*
 472 unlink_slot_from_free_list_uncontested (MSBlockInfo **free_blocks, int size_index)
 473 {
 474         MSBlockInfo *block;
 475         void *obj;
 476
 477         block = free_blocks [size_index];
 478         SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
 479
 480         if (G_UNLIKELY (!block->swept)) {
 481                 stat_major_blocks_lazy_swept ++;
 482                 sweep_block (block, FALSE);
 483         }
 484
 485         obj = block->free_list;
 486         SGEN_ASSERT (9, obj, "block %p in free list had no available object to alloc from", block);
 487
 488         block->free_list = *(void**)obj;
 489         if (!block->free_list) {
 490                 free_blocks [size_index] = block->next_free;
 491                 block->next_free = NULL;
 492         }
 493
 494         return obj;
 495 }
 496
 497 static void*
 498 alloc_obj (MonoVTable *vtable, size_t size, gboolean pinned, gboolean has_references)
 499 {
 500         int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 501         MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, has_references);
 502         void *obj;
 503
 504         if (!free_blocks [size_index]) {
 505                 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
 506                         return NULL;
 507         }
 508
 509         obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
 510
 511         *(MonoVTable**)obj = vtable;
 512
 513         return obj;
 514 }
 515
 516 static void*
 517 major_alloc_object (MonoVTable *vtable, size_t size, gboolean has_references)
 518 {
 519         return alloc_obj (vtable, size, FALSE, has_references);
 520 }
 521
 522 /*
 523  * We're not freeing the block if it's empty.  We leave that work for
 524  * the next major collection.
 525  *
 526  * This is just called from the domain clearing code, which runs in a
 527  * single thread and has the GC lock, so we don't need an extra lock.
 528  */
 529 static void
 530 free_object (char *obj, size_t size, gboolean pinned)
 531 {
 532         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 533         int word, bit;
 534
 535         if (!block->swept)
 536                 sweep_block (block, FALSE);
 537         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);
 538         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
 539         MS_CALC_MARK_BIT (word, bit, obj);
 540         SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set");
 541         if (!block->free_list) {
 542                 MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, block->has_references);
 543                 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 544                 SGEN_ASSERT (9, !block->next_free, "block %p doesn't have a free-list of object but belongs to a free-list of blocks");
 545                 block->next_free = free_blocks [size_index];
 546                 free_blocks [size_index] = block;
 547         }
 548         memset (obj, 0, size);
 549         *(void**)obj = block->free_list;
 550         block->free_list = (void**)obj;
 551 }
 552
 553 static void
 554 major_free_non_pinned_object (char *obj, size_t size)
 555 {
 556         free_object (obj, size, FALSE);
 557 }
 558
 559 /* size is a multiple of SGEN_ALLOC_ALIGN */
 560 static void*
 561 major_alloc_small_pinned_obj (MonoVTable *vtable, size_t size, gboolean has_references)
 562 {
 563         void *res;
 564
 565         res = alloc_obj (vtable, size, TRUE, has_references);
 566          /*If we failed to alloc memory, we better try releasing memory
 567           *as pinned alloc is requested by the runtime.
 568           */
 569          if (!res) {
 570                 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE);
 571                 res = alloc_obj (vtable, size, TRUE, has_references);
 572          }
 573          return res;
 574 }
 575
 576 static void
 577 free_pinned_object (char *obj, size_t size)
 578 {
 579         free_object (obj, size, TRUE);
 580 }
 581
 582 /*
 583  * size is already rounded up and we hold the GC lock.
 584  */
 585 static void*
 586 major_alloc_degraded (MonoVTable *vtable, size_t size)
 587 {
 588         void *obj;
 589         size_t old_num_sections;
 590
 591         old_num_sections = num_major_sections;
 592
 593         obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
 594         if (G_LIKELY (obj)) {
 595                 HEAVY_STAT (++stat_objects_alloced_degraded);
 596                 HEAVY_STAT (stat_bytes_alloced_degraded += size);
 597                 g_assert (num_major_sections >= old_num_sections);
 598                 sgen_register_major_sections_alloced (num_major_sections - old_num_sections);
 599         }
 600         return obj;
 601 }
 602
 603 /*
 604  * obj is some object.  If it's not in the major heap (i.e. if it's in
 605  * the nursery or LOS), return FALSE.  Otherwise return whether it's
 606  * been marked or copied.
 607  */
 608 static gboolean
 609 major_is_object_live (char *obj)
 610 {
 611         MSBlockInfo *block;
 612         int word, bit;
 613         mword objsize;
 614
 615         if (sgen_ptr_in_nursery (obj))
 616                 return FALSE;
 617
 618         objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
 619
 620         /* LOS */
 621         if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
 622                 return FALSE;
 623
 624         /* now we know it's in a major block */
 625         block = MS_BLOCK_FOR_OBJ (obj);
 626         SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?");
 627         MS_CALC_MARK_BIT (word, bit, obj);
 628         return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
 629 }
 630
 631 static gboolean
 632 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
 633 {
 634         MSBlockInfo *block;
 635
 636         FOREACH_BLOCK (block) {
 637                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
 638                         int count = MS_BLOCK_FREE / block->obj_size;
 639                         int i;
 640
 641                         *start = NULL;
 642                         for (i = 0; i <= count; ++i) {
 643                                 if (ptr >= MS_BLOCK_OBJ (block, i) && ptr < MS_BLOCK_OBJ (block, i + 1)) {
 644                                         *start = MS_BLOCK_OBJ (block, i);
 645                                         break;
 646                                 }
 647                         }
 648                         return !block->pinned;
 649                 }
 650         } END_FOREACH_BLOCK;
 651         return FALSE;
 652 }
 653
 654 static void
 655 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
 656 {
 657         gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
 658         gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
 659         gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
 660         MSBlockInfo *block;
 661
 662         FOREACH_BLOCK (block) {
 663                 int count = MS_BLOCK_FREE / block->obj_size;
 664                 int i;
 665
 666                 if (block->pinned && !pinned)
 667                         continue;
 668                 if (!block->pinned && !non_pinned)
 669                         continue;
 670                 if (sweep && lazy_sweep) {
 671                         sweep_block (block, FALSE);
 672                         SGEN_ASSERT (0, block->swept, "Block must be swept after sweeping");
 673                 }
 674
 675                 for (i = 0; i < count; ++i) {
 676                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 677                         if (!block->swept) {
 678                                 int word, bit;
 679                                 MS_CALC_MARK_BIT (word, bit, obj);
 680                                 if (!MS_MARK_BIT (block, word, bit))
 681                                         continue;
 682                         }
 683                         if (MS_OBJ_ALLOCED (obj, block))
 684                                 callback ((char*)obj, block->obj_size, data);
 685                 }
 686         } END_FOREACH_BLOCK;
 687 }
 688
 689 static gboolean
 690 major_is_valid_object (char *object)
 691 {
 692         MSBlockInfo *block;
 693
 694         FOREACH_BLOCK (block) {
 695                 int idx;
 696                 char *obj;
 697
 698                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
 699                         continue;
 700
 701                 idx = MS_BLOCK_OBJ_INDEX (object, block);
 702                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 703                 if (obj != object)
 704                         return FALSE;
 705                 return MS_OBJ_ALLOCED (obj, block);
 706         } END_FOREACH_BLOCK;
 707
 708         return FALSE;
 709 }
 710
 711
 712 static MonoVTable*
 713 major_describe_pointer (char *ptr)
 714 {
 715         MSBlockInfo *block;
 716
 717         FOREACH_BLOCK (block) {
 718                 int idx;
 719                 char *obj;
 720                 gboolean live;
 721                 MonoVTable *vtable;
 722                 int w, b;
 723                 gboolean marked;
 724
 725                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
 726                         continue;
 727
 728                 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
 729                         MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
 730
 731                 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
 732                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 733                 live = MS_OBJ_ALLOCED (obj, block);
 734                 vtable = live ? (MonoVTable*)SGEN_LOAD_VTABLE (obj) : NULL;
 735
 736                 MS_CALC_MARK_BIT (w, b, obj);
 737                 marked = MS_MARK_BIT (block, w, b);
 738
 739                 if (obj == ptr) {
 740                         SGEN_LOG (0, "\t(");
 741                         if (live)
 742                                 SGEN_LOG (0, "object");
 743                         else
 744                                 SGEN_LOG (0, "dead-object");
 745                 } else {
 746                         if (live)
 747                                 SGEN_LOG (0, "interior-ptr offset %td", ptr - obj);
 748                         else
 749                                 SGEN_LOG (0, "dead-interior-ptr offset %td", ptr - obj);
 750                 }
 751
 752                 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
 753
 754                 return vtable;
 755         } END_FOREACH_BLOCK;
 756
 757         return NULL;
 758 }
 759
 760 static void
 761 major_check_scan_starts (void)
 762 {
 763 }
 764
 765 static void
 766 major_dump_heap (FILE *heap_dump_file)
 767 {
 768         MSBlockInfo *block;
 769         int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
 770         int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
 771         int i;
 772
 773         for (i = 0; i < num_block_obj_sizes; ++i)
 774                 slots_available [i] = slots_used [i] = 0;
 775
 776         FOREACH_BLOCK (block) {
 777                 int index = ms_find_block_obj_size_index (block->obj_size);
 778                 int count = MS_BLOCK_FREE / block->obj_size;
 779
 780                 slots_available [index] += count;
 781                 for (i = 0; i < count; ++i) {
 782                         if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
 783                                 ++slots_used [index];
 784                 }
 785         } END_FOREACH_BLOCK;
 786
 787         fprintf (heap_dump_file, "<occupancies>\n");
 788         for (i = 0; i < num_block_obj_sizes; ++i) {
 789                 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
 790                                 block_obj_sizes [i], slots_available [i], slots_used [i]);
 791         }
 792         fprintf (heap_dump_file, "</occupancies>\n");
 793
 794         FOREACH_BLOCK (block) {
 795                 int count = MS_BLOCK_FREE / block->obj_size;
 796                 int i;
 797                 int start = -1;
 798
 799                 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
 800
 801                 for (i = 0; i <= count; ++i) {
 802                         if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
 803                                 if (start < 0)
 804                                         start = i;
 805                         } else {
 806                                 if (start >= 0) {
 807                                         sgen_dump_occupied (MS_BLOCK_OBJ (block, start), MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
 808                                         start = -1;
 809                                 }
 810                         }
 811                 }
 812
 813                 fprintf (heap_dump_file, "</section>\n");
 814         } END_FOREACH_BLOCK;
 815 }
 816
 817 #define LOAD_VTABLE     SGEN_LOAD_VTABLE
 818
 819 #define MS_MARK_OBJECT_AND_ENQUEUE_CHECKED(obj,desc,block,queue) do {   \
 820                 int __word, __bit;                                      \
 821                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
 822                 if (!MS_MARK_BIT ((block), __word, __bit) && MS_OBJ_ALLOCED ((obj), (block))) { \
 823                         MS_SET_MARK_BIT ((block), __word, __bit);       \
 824                         if ((block)->has_references)                    \
 825                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
 826                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((MonoObject*)(obj))); \
 827                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
 828                 }                                                       \
 829         } while (0)
 830 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do {           \
 831                 int __word, __bit;                                      \
 832                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
 833                 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj);       \
 834                 if (!MS_MARK_BIT ((block), __word, __bit)) {            \
 835                         MS_SET_MARK_BIT ((block), __word, __bit);       \
 836                         if ((block)->has_references)                    \
 837                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
 838                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((MonoObject*)(obj))); \
 839                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
 840                 }                                                       \
 841         } while (0)
 842
 843 static void
 844 pin_major_object (char *obj, SgenGrayQueue *queue)
 845 {
 846         MSBlockInfo *block;
 847
 848 #ifdef SGEN_HAVE_CONCURRENT_MARK
 849         if (concurrent_mark)
 850                 g_assert_not_reached ();
 851 #endif
 852
 853         block = MS_BLOCK_FOR_OBJ (obj);
 854         block->has_pinned = TRUE;
 855         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
 856 }
 857
 858 #include "sgen-major-copy-object.h"
 859
 860 #ifdef SGEN_HAVE_CONCURRENT_MARK
 861 static void
 862 major_copy_or_mark_object_concurrent (void **ptr, void *obj, SgenGrayQueue *queue)
 863 {
 864         SGEN_ASSERT (9, sgen_concurrent_collection_in_progress (), "Why are we scanning concurrently when there's no concurrent collection on?");
 865         SGEN_ASSERT (9, !sgen_workers_are_working () || sgen_is_worker_thread (mono_native_thread_id_get ()), "We must not scan from two threads at the same time!");
 866
 867         g_assert (!SGEN_OBJECT_IS_FORWARDED (obj));
 868
 869         if (!sgen_ptr_in_nursery (obj)) {
 870                 mword objsize;
 871
 872                 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
 873
 874                 if (objsize <= SGEN_MAX_SMALL_OBJ_SIZE) {
 875                         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 876                         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
 877                 } else {
 878                         if (sgen_los_object_is_pinned (obj))
 879                                 return;
 880
 881 #ifdef ENABLE_DTRACE
 882                         if (G_UNLIKELY (MONO_GC_OBJ_PINNED_ENABLED ())) {
 883                                 MonoVTable *vt = (MonoVTable*)SGEN_LOAD_VTABLE (obj);
 884                                 MONO_GC_OBJ_PINNED ((mword)obj, sgen_safe_object_get_size (obj), vt->klass->name_space, vt->klass->name, GENERATION_OLD);
 885                         }
 886 #endif
 887
 888                         sgen_los_pin_object (obj);
 889                         if (SGEN_OBJECT_HAS_REFERENCES (obj))
 890                                 GRAY_OBJECT_ENQUEUE (queue, obj, sgen_obj_get_descriptor (obj));
 891                         INC_NUM_MAJOR_OBJECTS_MARKED ();
 892                 }
 893         }
 894 }
 895 #endif
 896
 897 static void
 898 major_copy_or_mark_object (void **ptr, void *obj, SgenGrayQueue *queue)
 899 {
 900         MSBlockInfo *block;
 901
 902         HEAVY_STAT (++stat_copy_object_called_major);
 903
 904         SGEN_ASSERT (9, !sgen_concurrent_collection_in_progress (), "Why are we scanning non-concurrently when there's a concurrent collection on?");
 905
 906         SGEN_ASSERT (9, obj, "null object from pointer %p", ptr);
 907         SGEN_ASSERT (9, current_collection_generation == GENERATION_OLD, "old gen parallel allocator called from a %d collection", current_collection_generation);
 908
 909         if (sgen_ptr_in_nursery (obj)) {
 910                 int word, bit;
 911                 char *forwarded, *old_obj;
 912
 913                 if ((forwarded = SGEN_OBJECT_IS_FORWARDED (obj))) {
 914                         SGEN_UPDATE_REFERENCE (ptr, forwarded);
 915                         return;
 916                 }
 917                 if (SGEN_OBJECT_IS_PINNED (obj))
 918                         return;
 919
 920                 /* An object in the nursery To Space has already been copied and grayed. Nothing to do. */
 921                 if (sgen_nursery_is_to_space (obj))
 922                         return;
 923
 924                 HEAVY_STAT (++stat_objects_copied_major);
 925
 926         do_copy_object:
 927                 old_obj = obj;
 928                 obj = copy_object_no_checks (obj, queue);
 929                 if (G_UNLIKELY (old_obj == obj)) {
 930                         /*If we fail to evacuate an object we just stop doing it for a given block size as all other will surely fail too.*/
 931                         if (!sgen_ptr_in_nursery (obj)) {
 932                                 int size_index;
 933                                 block = MS_BLOCK_FOR_OBJ (obj);
 934                                 size_index = block->obj_size_index;
 935                                 evacuate_block_obj_sizes [size_index] = FALSE;
 936                                 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
 937                         }
 938                         return;
 939                 }
 940                 SGEN_UPDATE_REFERENCE (ptr, obj);
 941
 942                 /*
 943                  * FIXME: See comment for copy_object_no_checks().  If
 944                  * we have that, we can let the allocation function
 945                  * give us the block info, too, and we won't have to
 946                  * re-fetch it.
 947                  *
 948                  * FIXME (2): We should rework this to avoid all those nursery checks.
 949                  */
 950                 /*
 951                  * For the split nursery allocator the object might
 952                  * still be in the nursery despite having being
 953                  * promoted, in which case we can't mark it.
 954                  */
 955                 if (!sgen_ptr_in_nursery (obj)) {
 956                         block = MS_BLOCK_FOR_OBJ (obj);
 957                         MS_CALC_MARK_BIT (word, bit, obj);
 958                         SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p already marked", obj);
 959                         MS_SET_MARK_BIT (block, word, bit);
 960                         binary_protocol_mark (obj, (gpointer)LOAD_VTABLE (obj), sgen_safe_object_get_size ((MonoObject*)obj));
 961                 }
 962         } else {
 963                 char *forwarded;
 964                 mword objsize;
 965
 966                 /*
 967                  * If we have don't have a fixed heap we cannot know
 968                  * whether an object is in the LOS or in the small
 969                  * object major heap without checking its size.  To do
 970                  * that, however, we need to know that we actually
 971                  * have a valid object, not a forwarding pointer, so
 972                  * we have to do this check first.
 973                  */
 974                 if ((forwarded = SGEN_OBJECT_IS_FORWARDED (obj))) {
 975                         SGEN_UPDATE_REFERENCE (ptr, forwarded);
 976                         return;
 977                 }
 978
 979                 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
 980
 981                 if (objsize <= SGEN_MAX_SMALL_OBJ_SIZE) {
 982                         int size_index;
 983                         gboolean evacuate;
 984
 985                         block = MS_BLOCK_FOR_OBJ (obj);
 986                         size_index = block->obj_size_index;
 987                         evacuate = evacuate_block_obj_sizes [size_index];
 988
 989                         if (evacuate && !block->has_pinned) {
 990                                 g_assert (!SGEN_OBJECT_IS_PINNED (obj));
 991                                 if (block->is_to_space)
 992                                         return;
 993                                 HEAVY_STAT (++stat_major_objects_evacuated);
 994                                 goto do_copy_object;
 995                         } else {
 996                                 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
 997                         }
 998                 } else {
 999                         if (sgen_los_object_is_pinned (obj))
1000                                 return;
1001                         binary_protocol_pin (obj, (gpointer)SGEN_LOAD_VTABLE (obj), sgen_safe_object_get_size ((MonoObject*)obj));
1002
1003 #ifdef ENABLE_DTRACE
1004                         if (G_UNLIKELY (MONO_GC_OBJ_PINNED_ENABLED ())) {
1005                                 MonoVTable *vt = (MonoVTable*)SGEN_LOAD_VTABLE (obj);
1006                                 MONO_GC_OBJ_PINNED ((mword)obj, sgen_safe_object_get_size (obj), vt->klass->name_space, vt->klass->name, GENERATION_OLD);
1007                         }
1008 #endif
1009
1010                         sgen_los_pin_object (obj);
1011                         if (SGEN_OBJECT_HAS_REFERENCES (obj))
1012                                 GRAY_OBJECT_ENQUEUE (queue, obj, sgen_obj_get_descriptor (obj));
1013                 }
1014         }
1015 }
1016
1017 static void
1018 major_copy_or_mark_object_canonical (void **ptr, SgenGrayQueue *queue)
1019 {
1020         major_copy_or_mark_object (ptr, *ptr, queue);
1021 }
1022
1023 #ifdef SGEN_HAVE_CONCURRENT_MARK
1024 static void
1025 major_copy_or_mark_object_concurrent_canonical (void **ptr, SgenGrayQueue *queue)
1026 {
1027         major_copy_or_mark_object_concurrent (ptr, *ptr, queue);
1028 }
1029 #endif
1030
1031 static long long
1032 major_get_and_reset_num_major_objects_marked (void)
1033 {
1034 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1035         long long num = num_major_objects_marked;
1036         num_major_objects_marked = 0;
1037         return num;
1038 #else
1039         return 0;
1040 #endif
1041 }
1042
1043 #include "sgen-major-scan-object.h"
1044
1045 #ifdef SGEN_HAVE_CONCURRENT_MARK
1046 #define SCAN_FOR_CONCURRENT_MARK
1047 #include "sgen-major-scan-object.h"
1048 #undef SCAN_FOR_CONCURRENT_MARK
1049 #endif
1050
1051 static void
1052 mark_pinned_objects_in_block (MSBlockInfo *block, SgenGrayQueue *queue)
1053 {
1054         void **entry, **end;
1055         int last_index = -1;
1056
1057         if (block->pin_queue_first_entry == block->pin_queue_last_entry)
1058                 return;
1059
1060         block->has_pinned = TRUE;
1061
1062         entry = sgen_pinning_get_entry (block->pin_queue_first_entry);
1063         end = sgen_pinning_get_entry (block->pin_queue_last_entry);
1064
1065         for (; entry < end; ++entry) {
1066                 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1067                 char *obj;
1068                 SGEN_ASSERT (9, index >= 0 && index < MS_BLOCK_FREE / block->obj_size, "invalid object %p index %d max-index %d", *entry, index, MS_BLOCK_FREE / block->obj_size);
1069                 if (index == last_index)
1070                         continue;
1071                 obj = MS_BLOCK_OBJ (block, index);
1072                 MS_MARK_OBJECT_AND_ENQUEUE_CHECKED (obj, sgen_obj_get_descriptor (obj), block, queue);
1073                 last_index = index;
1074         }
1075 }
1076
1077 static inline void
1078 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1079 {
1080         int obj_index;
1081
1082         for (obj_index = 0; obj_index < count; ++obj_index) {
1083                 int word, bit;
1084                 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1085
1086                 MS_CALC_MARK_BIT (word, bit, obj);
1087                 if (MS_MARK_BIT (block, word, bit)) {
1088                         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1089                 } else {
1090                         /* an unmarked object */
1091                         if (MS_OBJ_ALLOCED (obj, block)) {
1092                                 /*
1093                                  * FIXME: Merge consecutive
1094                                  * slots for lower reporting
1095                                  * overhead.  Maybe memset
1096                                  * will also benefit?
1097                                  */
1098                                 binary_protocol_empty (obj, obj_size);
1099                                 MONO_GC_MAJOR_SWEPT ((mword)obj, obj_size);
1100                                 memset (obj, 0, obj_size);
1101                         }
1102                         *(void**)obj = block->free_list;
1103                         block->free_list = obj;
1104                 }
1105         }
1106 }
1107
1108 /*
1109  * sweep_block:
1110  *
1111  *   Traverse BLOCK, freeing and zeroing unused objects.
1112  */
1113 static void
1114 sweep_block (MSBlockInfo *block, gboolean during_major_collection)
1115 {
1116         int count;
1117         void *reversed = NULL;
1118
1119         if (!during_major_collection)
1120                 g_assert (!sgen_concurrent_collection_in_progress ());
1121
1122         if (block->swept)
1123                 return;
1124
1125         count = MS_BLOCK_FREE / block->obj_size;
1126
1127         block->free_list = NULL;
1128
1129         /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1130         // FIXME: Add more sizes
1131         switch (block->obj_size) {
1132         case 16:
1133                 sweep_block_for_size (block, count, 16);
1134                 break;
1135         default:
1136                 sweep_block_for_size (block, count, block->obj_size);
1137                 break;
1138         }
1139
1140         /* reset mark bits */
1141         memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1142
1143         /* Reverse free list so that it's in address order */
1144         reversed = NULL;
1145         while (block->free_list) {
1146                 void *next = *(void**)block->free_list;
1147                 *(void**)block->free_list = reversed;
1148                 reversed = block->free_list;
1149                 block->free_list = next;
1150         }
1151         block->free_list = reversed;
1152
1153         block->swept = 1;
1154 }
1155
1156 static inline int
1157 bitcount (mword d)
1158 {
1159         int count = 0;
1160
1161 #ifdef __GNUC__
1162         if (sizeof (mword) == sizeof (unsigned long))
1163                 count += __builtin_popcountl (d);
1164         else
1165                 count += __builtin_popcount (d);
1166 #else
1167         while (d) {
1168                 count ++;
1169                 d &= (d - 1);
1170         }
1171 #endif
1172         return count;
1173 }
1174
1175 static void
1176 ms_sweep (void)
1177 {
1178         int i;
1179         MSBlockInfo *block;
1180
1181         /* statistics for evacuation */
1182         int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
1183         int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
1184         int *num_blocks = alloca (sizeof (int) * num_block_obj_sizes);
1185
1186 #ifdef SGEN_HAVE_CONCURRENT_MARK
1187         mword total_evacuate_heap = 0;
1188         mword total_evacuate_saved = 0;
1189 #endif
1190
1191         for (i = 0; i < num_block_obj_sizes; ++i)
1192                 slots_available [i] = slots_used [i] = num_blocks [i] = 0;
1193
1194         /* clear all the free lists */
1195         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1196                 MSBlockInfo **free_blocks = free_block_lists [i];
1197                 int j;
1198                 for (j = 0; j < num_block_obj_sizes; ++j)
1199                         free_blocks [j] = NULL;
1200         }
1201
1202         /* traverse all blocks, free and zero unmarked objects */
1203         FOREACH_BLOCK (block) {
1204                 int count;
1205                 gboolean have_live = FALSE;
1206                 gboolean has_pinned;
1207                 gboolean have_free = FALSE;
1208                 int obj_size_index;
1209                 int nused = 0;
1210
1211                 obj_size_index = block->obj_size_index;
1212
1213                 has_pinned = block->has_pinned;
1214                 block->has_pinned = block->pinned;
1215
1216                 block->is_to_space = FALSE;
1217                 block->swept = 0;
1218
1219                 count = MS_BLOCK_FREE / block->obj_size;
1220
1221 #ifdef SGEN_HAVE_CONCURRENT_MARK
1222                 if (block->cardtable_mod_union) {
1223                         sgen_free_internal_dynamic (block->cardtable_mod_union, CARDS_PER_BLOCK, INTERNAL_MEM_CARDTABLE_MOD_UNION);
1224                         block->cardtable_mod_union = NULL;
1225                 }
1226 #endif
1227
1228                 /* Count marked objects in the block */
1229                 for (i = 0; i < MS_NUM_MARK_WORDS; ++i) {
1230                         nused += bitcount (block->mark_words [i]);
1231                 }
1232                 if (nused) {
1233                         have_live = TRUE;
1234                 }
1235                 if (nused < count)
1236                         have_free = TRUE;
1237
1238                 if (!lazy_sweep)
1239                         sweep_block (block, TRUE);
1240
1241                 if (have_live) {
1242                         if (!has_pinned) {
1243                                 ++num_blocks [obj_size_index];
1244                                 slots_used [obj_size_index] += nused;
1245                                 slots_available [obj_size_index] += count;
1246                         }
1247
1248                         /*
1249                          * If there are free slots in the block, add
1250                          * the block to the corresponding free list.
1251                          */
1252                         if (have_free) {
1253                                 MSBlockInfo **free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1254                                 int index = MS_BLOCK_OBJ_SIZE_INDEX (block->obj_size);
1255                                 block->next_free = free_blocks [index];
1256                                 free_blocks [index] = block;
1257                         }
1258
1259                         update_heap_boundaries_for_block (block);
1260                 } else {
1261                         /*
1262                          * Blocks without live objects are removed from the
1263                          * block list and freed.
1264                          */
1265                         DELETE_BLOCK_IN_FOREACH ();
1266
1267                         binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1268                         ms_free_block (block->block);
1269                         sgen_free_internal (block, INTERNAL_MEM_MS_BLOCK_INFO);
1270
1271                         --num_major_sections;
1272                 }
1273         } END_FOREACH_BLOCK;
1274         sgen_pointer_queue_remove_nulls (&allocated_blocks);
1275
1276         for (i = 0; i < num_block_obj_sizes; ++i) {
1277                 float usage = (float)slots_used [i] / (float)slots_available [i];
1278                 if (num_blocks [i] > 5 && usage < evacuation_threshold) {
1279                         evacuate_block_obj_sizes [i] = TRUE;
1280                         /*
1281                         g_print ("slot size %d - %d of %d used\n",
1282                                         block_obj_sizes [i], slots_used [i], slots_available [i]);
1283                         */
1284                 } else {
1285                         evacuate_block_obj_sizes [i] = FALSE;
1286                 }
1287 #ifdef SGEN_HAVE_CONCURRENT_MARK
1288                 {
1289                         mword total_bytes = block_obj_sizes [i] * slots_available [i];
1290                         total_evacuate_heap += total_bytes;
1291                         if (evacuate_block_obj_sizes [i])
1292                                 total_evacuate_saved += total_bytes - block_obj_sizes [i] * slots_used [i];
1293                 }
1294 #endif
1295         }
1296
1297 #ifdef SGEN_HAVE_CONCURRENT_MARK
1298         want_evacuation = (float)total_evacuate_saved / (float)total_evacuate_heap > (1 - concurrent_evacuation_threshold);
1299 #endif
1300
1301         have_swept = TRUE;
1302 }
1303
1304 static void
1305 major_sweep (void)
1306 {
1307         ms_sweep ();
1308 }
1309
1310 static int count_pinned_ref;
1311 static int count_pinned_nonref;
1312 static int count_nonpinned_ref;
1313 static int count_nonpinned_nonref;
1314
1315 static void
1316 count_nonpinned_callback (char *obj, size_t size, void *data)
1317 {
1318         MonoVTable *vtable = (MonoVTable*)LOAD_VTABLE (obj);
1319
1320         if (vtable->klass->has_references)
1321                 ++count_nonpinned_ref;
1322         else
1323                 ++count_nonpinned_nonref;
1324 }
1325
1326 static void
1327 count_pinned_callback (char *obj, size_t size, void *data)
1328 {
1329         MonoVTable *vtable = (MonoVTable*)LOAD_VTABLE (obj);
1330
1331         if (vtable->klass->has_references)
1332                 ++count_pinned_ref;
1333         else
1334                 ++count_pinned_nonref;
1335 }
1336
1337 static G_GNUC_UNUSED void
1338 count_ref_nonref_objs (void)
1339 {
1340         int total;
1341
1342         count_pinned_ref = 0;
1343         count_pinned_nonref = 0;
1344         count_nonpinned_ref = 0;
1345         count_nonpinned_nonref = 0;
1346
1347         major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1348         major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1349
1350         total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1351
1352         g_print ("ref: %d pinned %d non-pinned   non-ref: %d pinned %d non-pinned  --  %.1f\n",
1353                         count_pinned_ref, count_nonpinned_ref,
1354                         count_pinned_nonref, count_nonpinned_nonref,
1355                         (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1356 }
1357
1358 static int
1359 ms_calculate_block_obj_sizes (double factor, int *arr)
1360 {
1361         double target_size = sizeof (MonoObject);
1362         int num_sizes = 0;
1363         int last_size = 0;
1364
1365         do {
1366                 int target_count = (int)ceil (MS_BLOCK_FREE / target_size);
1367                 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1368
1369                 if (size != last_size) {
1370                         if (arr)
1371                                 arr [num_sizes] = size;
1372                         ++num_sizes;
1373                         last_size = size;
1374                 }
1375
1376                 target_size *= factor;
1377         } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1378
1379         return num_sizes;
1380 }
1381
1382 /* only valid during minor collections */
1383 static mword old_num_major_sections;
1384
1385 static void
1386 major_start_nursery_collection (void)
1387 {
1388 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1389         consistency_check ();
1390 #endif
1391
1392         old_num_major_sections = num_major_sections;
1393 }
1394
1395 static void
1396 major_finish_nursery_collection (void)
1397 {
1398 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1399         consistency_check ();
1400 #endif
1401         sgen_register_major_sections_alloced (num_major_sections - old_num_major_sections);
1402 }
1403
1404 static void
1405 major_start_major_collection (void)
1406 {
1407         int i;
1408
1409         /* clear the free lists */
1410         for (i = 0; i < num_block_obj_sizes; ++i) {
1411                 if (!evacuate_block_obj_sizes [i])
1412                         continue;
1413
1414                 free_block_lists [0][i] = NULL;
1415                 free_block_lists [MS_BLOCK_FLAG_REFS][i] = NULL;
1416         }
1417
1418         // Sweep all unswept blocks
1419         if (lazy_sweep) {
1420                 MSBlockInfo *block;
1421
1422                 MONO_GC_SWEEP_BEGIN (GENERATION_OLD, TRUE);
1423
1424                 FOREACH_BLOCK (block) {
1425                         sweep_block (block, TRUE);
1426                 } END_FOREACH_BLOCK;
1427
1428                 MONO_GC_SWEEP_END (GENERATION_OLD, TRUE);
1429         }
1430 }
1431
1432 static void
1433 major_finish_major_collection (void)
1434 {
1435 }
1436
1437 #if SIZEOF_VOID_P != 8
1438 static int
1439 compare_pointers (const void *va, const void *vb) {
1440         char *a = *(char**)va, *b = *(char**)vb;
1441         if (a < b)
1442                 return -1;
1443         if (a > b)
1444                 return 1;
1445         return 0;
1446 }
1447 #endif
1448
1449 static void
1450 major_have_computer_minor_collection_allowance (void)
1451 {
1452         size_t section_reserve = sgen_get_minor_collection_allowance () / MS_BLOCK_SIZE;
1453
1454         g_assert (have_swept);
1455
1456 #if SIZEOF_VOID_P != 8
1457         {
1458                 int i, num_empty_blocks_orig, num_blocks, arr_length;
1459                 void *block;
1460                 void **empty_block_arr;
1461                 void **rebuild_next;
1462
1463 #ifdef TARGET_WIN32
1464                 /*
1465                  * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1466                  * a VirtualAlloc ()-ed block.
1467                  */
1468                 return;
1469 #endif
1470
1471                 if (num_empty_blocks <= section_reserve)
1472                         return;
1473                 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1474
1475                 num_empty_blocks_orig = num_empty_blocks;
1476                 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1477                                 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1478                 if (!empty_block_arr)
1479                         goto fallback;
1480
1481                 i = 0;
1482                 for (block = empty_blocks; block; block = *(void**)block)
1483                         empty_block_arr [i++] = block;
1484                 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1485
1486                 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1487
1488                 /*
1489                  * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1490                  * contiguous ones.  If we do, we free them.  If that's not enough to get to
1491                  * section_reserve, we halve the number of contiguous blocks we're looking
1492                  * for and have another go, until we're done with looking for pairs of
1493                  * blocks, at which point we give up and go to the fallback.
1494                  */
1495                 arr_length = num_empty_blocks_orig;
1496                 num_blocks = MS_BLOCK_ALLOC_NUM;
1497                 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1498                         int first = -1;
1499                         int dest = 0;
1500
1501                         dest = 0;
1502                         for (i = 0; i < arr_length; ++i) {
1503                                 int d = dest;
1504                                 void *block = empty_block_arr [i];
1505                                 SGEN_ASSERT (0, block, "we're not shifting correctly");
1506                                 if (i != dest) {
1507                                         empty_block_arr [dest] = block;
1508                                         /*
1509                                          * This is not strictly necessary, but we're
1510                                          * cautious.
1511                                          */
1512                                         empty_block_arr [i] = NULL;
1513                                 }
1514                                 ++dest;
1515
1516                                 if (first < 0) {
1517                                         first = d;
1518                                         continue;
1519                                 }
1520
1521                                 SGEN_ASSERT (0, first >= 0 && d > first, "algorithm is wrong");
1522
1523                                 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
1524                                         first = d;
1525                                         continue;
1526                                 }
1527
1528                                 if (d + 1 - first == num_blocks) {
1529                                         /*
1530                                          * We found num_blocks contiguous blocks.  Free them
1531                                          * and null their array entries.  As an optimization
1532                                          * we could, instead of nulling the entries, shift
1533                                          * the following entries over to the left, while
1534                                          * we're iterating.
1535                                          */
1536                                         int j;
1537                                         sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
1538                                         for (j = first; j <= d; ++j)
1539                                                 empty_block_arr [j] = NULL;
1540                                         dest = first;
1541                                         first = -1;
1542
1543                                         num_empty_blocks -= num_blocks;
1544
1545                                         stat_major_blocks_freed += num_blocks;
1546                                         if (num_blocks == MS_BLOCK_ALLOC_NUM)
1547                                                 stat_major_blocks_freed_ideal += num_blocks;
1548                                         else
1549                                                 stat_major_blocks_freed_less_ideal += num_blocks;
1550
1551                                 }
1552                         }
1553
1554                         SGEN_ASSERT (0, dest <= i && dest <= arr_length, "array length is off");
1555                         arr_length = dest;
1556                         SGEN_ASSERT (0, arr_length == num_empty_blocks, "array length is off");
1557
1558                         num_blocks >>= 1;
1559                 }
1560
1561                 /* rebuild empty_blocks free list */
1562                 rebuild_next = (void**)&empty_blocks;
1563                 for (i = 0; i < arr_length; ++i) {
1564                         void *block = empty_block_arr [i];
1565                         SGEN_ASSERT (0, block, "we're missing blocks");
1566                         *rebuild_next = block;
1567                         rebuild_next = (void**)block;
1568                 }
1569                 *rebuild_next = NULL;
1570
1571                 /* free array */
1572                 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
1573         }
1574
1575         SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
1576
1577  fallback:
1578         /*
1579          * This is our threshold.  If there's not more empty than used blocks, we won't
1580          * release uncontiguous blocks, in fear of fragmenting the address space.
1581          */
1582         if (num_empty_blocks <= num_major_sections)
1583                 return;
1584 #endif
1585
1586         while (num_empty_blocks > section_reserve) {
1587                 void *next = *(void**)empty_blocks;
1588                 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
1589                 empty_blocks = next;
1590                 /*
1591                  * Needs not be atomic because this is running
1592                  * single-threaded.
1593                  */
1594                 --num_empty_blocks;
1595
1596                 ++stat_major_blocks_freed;
1597 #if SIZEOF_VOID_P != 8
1598                 ++stat_major_blocks_freed_individual;
1599 #endif
1600         }
1601 }
1602
1603 static void
1604 major_find_pin_queue_start_ends (SgenGrayQueue *queue)
1605 {
1606         MSBlockInfo *block;
1607
1608         FOREACH_BLOCK (block) {
1609                 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
1610                                 &block->pin_queue_first_entry, &block->pin_queue_last_entry);
1611         } END_FOREACH_BLOCK;
1612 }
1613
1614 static void
1615 major_pin_objects (SgenGrayQueue *queue)
1616 {
1617         MSBlockInfo *block;
1618
1619         FOREACH_BLOCK (block) {
1620                 mark_pinned_objects_in_block (block, queue);
1621         } END_FOREACH_BLOCK;
1622 }
1623
1624 static void
1625 major_init_to_space (void)
1626 {
1627 }
1628
1629 static void
1630 major_report_pinned_memory_usage (void)
1631 {
1632         g_assert_not_reached ();
1633 }
1634
1635 static gint64
1636 major_get_used_size (void)
1637 {
1638         gint64 size = 0;
1639         MSBlockInfo *block;
1640
1641         FOREACH_BLOCK (block) {
1642                 int count = MS_BLOCK_FREE / block->obj_size;
1643                 void **iter;
1644                 size += count * block->obj_size;
1645                 for (iter = block->free_list; iter; iter = (void**)*iter)
1646                         size -= block->obj_size;
1647         } END_FOREACH_BLOCK;
1648
1649         return size;
1650 }
1651
1652 static size_t
1653 get_num_major_sections (void)
1654 {
1655         return num_major_sections;
1656 }
1657
1658 static gboolean
1659 major_handle_gc_param (const char *opt)
1660 {
1661         if (g_str_has_prefix (opt, "evacuation-threshold=")) {
1662                 const char *arg = strchr (opt, '=') + 1;
1663                 int percentage = atoi (arg);
1664                 if (percentage < 0 || percentage > 100) {
1665                         fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
1666                         exit (1);
1667                 }
1668                 evacuation_threshold = (float)percentage / 100.0f;
1669                 return TRUE;
1670         } else if (!strcmp (opt, "lazy-sweep")) {
1671                 lazy_sweep = TRUE;
1672                 return TRUE;
1673         } else if (!strcmp (opt, "no-lazy-sweep")) {
1674                 lazy_sweep = FALSE;
1675                 return TRUE;
1676         }
1677
1678         return FALSE;
1679 }
1680
1681 static void
1682 major_print_gc_param_usage (void)
1683 {
1684         fprintf (stderr,
1685                         ""
1686                         "  evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
1687                         "  (no-)lazy-sweep\n"
1688                         );
1689 }
1690
1691 static void
1692 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
1693 {
1694         MSBlockInfo *block;
1695         gboolean has_references;
1696
1697         FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1698                 if (has_references)
1699                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1700         } END_FOREACH_BLOCK;
1701 }
1702
1703 #ifdef HEAVY_STATISTICS
1704 extern long long marked_cards;
1705 extern long long scanned_cards;
1706 extern long long scanned_objects;
1707 extern long long remarked_cards;
1708 #endif
1709
1710 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
1711 /*
1712  * MS blocks are 16K aligned.
1713  * Cardtables are 4K aligned, at least.
1714  * This means that the cardtable of a given block is 32 bytes aligned.
1715  */
1716 static guint8*
1717 initial_skip_card (guint8 *card_data)
1718 {
1719         mword *cards = (mword*)card_data;
1720         mword card;
1721         int i;
1722         for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
1723                 card = cards [i];
1724                 if (card)
1725                         break;
1726         }
1727
1728         if (i == CARD_WORDS_PER_BLOCK)
1729                 return card_data + CARDS_PER_BLOCK;
1730
1731 #if defined(__i386__) && defined(__GNUC__)
1732         return card_data + i * 4 +  (__builtin_ffs (card) - 1) / 8;
1733 #elif defined(__x86_64__) && defined(__GNUC__)
1734         return card_data + i * 8 +  (__builtin_ffsll (card) - 1) / 8;
1735 #elif defined(__s390x__) && defined(__GNUC__)
1736         return card_data + i * 8 +  (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
1737 #else
1738         for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
1739                 if (card_data [i])
1740                         return &card_data [i];
1741         }
1742         return card_data;
1743 #endif
1744 }
1745
1746
1747 static G_GNUC_UNUSED guint8*
1748 skip_card (guint8 *card_data, guint8 *card_data_end)
1749 {
1750         while (card_data < card_data_end && !*card_data)
1751                 ++card_data;
1752         return card_data;
1753 }
1754
1755 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
1756 #define MS_BLOCK_OBJ_FAST(b,os,i)                       ((b) + MS_BLOCK_SKIP + (os) * (i))
1757 #define MS_OBJ_ALLOCED_FAST(o,b)                (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
1758
1759 static void
1760 major_scan_card_table (gboolean mod_union, SgenGrayQueue *queue)
1761 {
1762         MSBlockInfo *block;
1763         gboolean has_references;
1764         ScanObjectFunc scan_func = sgen_get_current_object_ops ()->scan_object;
1765
1766 #ifdef SGEN_HAVE_CONCURRENT_MARK
1767         if (!concurrent_mark)
1768                 g_assert (!mod_union);
1769 #else
1770         g_assert (!mod_union);
1771 #endif
1772
1773         FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1774                 int block_obj_size;
1775                 char *block_start;
1776
1777                 if (!has_references)
1778                         continue;
1779
1780                 block_obj_size = block->obj_size;
1781                 block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
1782
1783                 if (block_obj_size >= CARD_SIZE_IN_BYTES) {
1784                         guint8 *cards;
1785 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
1786                         guint8 cards_data [CARDS_PER_BLOCK];
1787 #endif
1788                         char *obj, *end, *base;
1789
1790                         if (mod_union) {
1791 #ifdef SGEN_HAVE_CONCURRENT_MARK
1792                                 cards = block->cardtable_mod_union;
1793                                 /*
1794                                  * This happens when the nursery
1795                                  * collection that precedes finishing
1796                                  * the concurrent collection allocates
1797                                  * new major blocks.
1798                                  */
1799                                 if (!cards)
1800                                         continue;
1801 #endif
1802                         } else {
1803                         /*We can avoid the extra copy since the remark cardtable was cleaned before */
1804 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
1805                                 cards = sgen_card_table_get_card_scan_address ((mword)block_start);
1806 #else
1807                                 cards = cards_data;
1808                                 if (!sgen_card_table_get_card_data (cards_data, (mword)block_start, CARDS_PER_BLOCK))
1809                                         continue;
1810 #endif
1811                         }
1812
1813                         obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, 0);
1814                         end = block_start + MS_BLOCK_SIZE;
1815                         base = sgen_card_table_align_pointer (obj);
1816
1817                         while (obj < end) {
1818                                 size_t card_offset;
1819
1820                                 if (!block->swept)
1821                                         sweep_block (block, FALSE);
1822
1823                                 if (!MS_OBJ_ALLOCED_FAST (obj, block_start))
1824                                         goto next_large;
1825
1826                                 if (mod_union) {
1827                                         /* FIXME: do this more efficiently */
1828                                         int w, b;
1829                                         MS_CALC_MARK_BIT (w, b, obj);
1830                                         if (!MS_MARK_BIT (block, w, b))
1831                                                 goto next_large;
1832                                 }
1833
1834                                 card_offset = (obj - base) >> CARD_BITS;
1835                                 sgen_cardtable_scan_object (obj, block_obj_size, cards + card_offset, mod_union, queue);
1836
1837                         next_large:
1838                                 obj += block_obj_size;
1839                         }
1840                 } else {
1841                         guint8 *card_data, *card_base;
1842                         guint8 *card_data_end;
1843
1844                         /*
1845                          * This is safe in face of card aliasing for the following reason:
1846                          *
1847                          * Major blocks are 16k aligned, or 32 cards aligned.
1848                          * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
1849                          * sizes, they won't overflow the cardtable overlap modulus.
1850                          */
1851                         if (mod_union) {
1852 #ifdef SGEN_HAVE_CONCURRENT_MARK
1853                                 card_data = card_base = block->cardtable_mod_union;
1854                                 /*
1855                                  * This happens when the nursery
1856                                  * collection that precedes finishing
1857                                  * the concurrent collection allocates
1858                                  * new major blocks.
1859                                  */
1860                                 if (!card_data)
1861                                         continue;
1862 #else
1863                                 g_assert_not_reached ();
1864                                 card_data = NULL;
1865 #endif
1866                         } else {
1867                                 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
1868                         }
1869                         card_data_end = card_data + CARDS_PER_BLOCK;
1870
1871                         for (card_data = initial_skip_card (card_data); card_data < card_data_end; ++card_data) { //card_data = skip_card (card_data + 1, card_data_end)) {
1872                                 size_t index;
1873                                 size_t idx = card_data - card_base;
1874                                 char *start = (char*)(block_start + idx * CARD_SIZE_IN_BYTES);
1875                                 char *end = start + CARD_SIZE_IN_BYTES;
1876                                 char *first_obj, *obj;
1877
1878                                 HEAVY_STAT (++scanned_cards);
1879
1880                                 if (!*card_data)
1881                                         continue;
1882
1883                                 if (!block->swept)
1884                                         sweep_block (block, FALSE);
1885
1886                                 HEAVY_STAT (++marked_cards);
1887
1888                                 sgen_card_table_prepare_card_for_scanning (card_data);
1889
1890                                 if (idx == 0)
1891                                         index = 0;
1892                                 else
1893                                         index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
1894
1895                                 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, index);
1896                                 while (obj < end) {
1897                                         if (!MS_OBJ_ALLOCED_FAST (obj, block_start))
1898                                                 goto next_small;
1899
1900                                         if (mod_union) {
1901                                                 /* FIXME: do this more efficiently */
1902                                                 int w, b;
1903                                                 MS_CALC_MARK_BIT (w, b, obj);
1904                                                 if (!MS_MARK_BIT (block, w, b))
1905                                                         goto next_small;
1906                                         }
1907
1908                                         HEAVY_STAT (++scanned_objects);
1909                                         scan_func (obj, sgen_obj_get_descriptor (obj), queue);
1910                                 next_small:
1911                                         obj += block_obj_size;
1912                                 }
1913                                 HEAVY_STAT (if (*card_data) ++remarked_cards);
1914                                 binary_protocol_card_scan (first_obj, obj - first_obj);
1915                         }
1916                 }
1917         } END_FOREACH_BLOCK;
1918 }
1919
1920 static void
1921 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
1922 {
1923         MSBlockInfo *block;
1924         gboolean has_references;
1925         long long total_cards = 0;
1926         long long marked_cards = 0;
1927
1928         FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1929                 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
1930                 int i;
1931
1932                 if (!has_references)
1933                         continue;
1934
1935                 total_cards += CARDS_PER_BLOCK;
1936                 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
1937                         if (cards [i])
1938                                 ++marked_cards;
1939                 }
1940         } END_FOREACH_BLOCK;
1941
1942         *num_total_cards = total_cards;
1943         *num_marked_cards = marked_cards;
1944 }
1945
1946 #ifdef SGEN_HAVE_CONCURRENT_MARK
1947 static void
1948 update_cardtable_mod_union (void)
1949 {
1950         MSBlockInfo *block;
1951
1952         FOREACH_BLOCK (block) {
1953                 size_t num_cards;
1954
1955                 block->cardtable_mod_union = sgen_card_table_update_mod_union (block->cardtable_mod_union,
1956                                 MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
1957
1958                 SGEN_ASSERT (0, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
1959         } END_FOREACH_BLOCK;
1960 }
1961
1962 static guint8*
1963 major_get_cardtable_mod_union_for_object (char *obj)
1964 {
1965         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
1966         return &block->cardtable_mod_union [(obj - (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block))) >> CARD_BITS];
1967 }
1968 #endif
1969
1970 static void
1971 alloc_free_block_lists (MSBlockInfo ***lists)
1972 {
1973         int i;
1974         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
1975                 lists [i] = sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
1976 }
1977
1978 #undef pthread_create
1979
1980 static void
1981 post_param_init (SgenMajorCollector *collector)
1982 {
1983         collector->sweeps_lazily = lazy_sweep;
1984 }
1985
1986 #ifdef SGEN_HAVE_CONCURRENT_MARK
1987 static void
1988 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
1989 #else // SGEN_HAVE_CONCURRENT_MARK
1990 #error unknown configuration
1991 #endif // SGEN_HAVE_CONCURRENT_MARK
1992 {
1993         int i;
1994
1995         sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
1996
1997         num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
1998         block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
1999         ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2000
2001         evacuate_block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2002         for (i = 0; i < num_block_obj_sizes; ++i)
2003                 evacuate_block_obj_sizes [i] = FALSE;
2004
2005         /*
2006         {
2007                 int i;
2008                 g_print ("block object sizes:\n");
2009                 for (i = 0; i < num_block_obj_sizes; ++i)
2010                         g_print ("%d\n", block_obj_sizes [i]);
2011         }
2012         */
2013
2014         alloc_free_block_lists (free_block_lists);
2015
2016         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2017                 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2018         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2019                 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2020
2021         mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_major_blocks_alloced);
2022         mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_major_blocks_freed);
2023         mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_major_blocks_lazy_swept);
2024         mono_counters_register ("# major objects evacuated", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_major_objects_evacuated);
2025 #if SIZEOF_VOID_P != 8
2026         mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_major_blocks_freed_ideal);
2027         mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_major_blocks_freed_less_ideal);
2028         mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_major_blocks_freed_individual);
2029         mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_major_blocks_alloced_less_ideal);
2030 #endif
2031
2032         collector->section_size = MAJOR_SECTION_SIZE;
2033
2034 #ifdef SGEN_HAVE_CONCURRENT_MARK
2035         concurrent_mark = is_concurrent;
2036         if (is_concurrent) {
2037                 collector->is_concurrent = TRUE;
2038                 collector->want_synchronous_collection = &want_evacuation;
2039         } else
2040 #endif
2041         {
2042                 collector->is_concurrent = FALSE;
2043                 collector->want_synchronous_collection = NULL;
2044         }
2045         collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2046         collector->supports_cardtable = TRUE;
2047
2048         collector->have_swept = &have_swept;
2049
2050         collector->alloc_heap = major_alloc_heap;
2051         collector->is_object_live = major_is_object_live;
2052         collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2053         collector->alloc_degraded = major_alloc_degraded;
2054
2055         collector->alloc_object = major_alloc_object;
2056         collector->free_pinned_object = free_pinned_object;
2057         collector->iterate_objects = major_iterate_objects;
2058         collector->free_non_pinned_object = major_free_non_pinned_object;
2059         collector->find_pin_queue_start_ends = major_find_pin_queue_start_ends;
2060         collector->pin_objects = major_pin_objects;
2061         collector->pin_major_object = pin_major_object;
2062         collector->scan_card_table = major_scan_card_table;
2063         collector->iterate_live_block_ranges = (void*)(void*) major_iterate_live_block_ranges;
2064 #ifdef SGEN_HAVE_CONCURRENT_MARK
2065         if (is_concurrent) {
2066                 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2067                 collector->get_cardtable_mod_union_for_object = major_get_cardtable_mod_union_for_object;
2068         }
2069 #endif
2070         collector->init_to_space = major_init_to_space;
2071         collector->sweep = major_sweep;
2072         collector->check_scan_starts = major_check_scan_starts;
2073         collector->dump_heap = major_dump_heap;
2074         collector->get_used_size = major_get_used_size;
2075         collector->start_nursery_collection = major_start_nursery_collection;
2076         collector->finish_nursery_collection = major_finish_nursery_collection;
2077         collector->start_major_collection = major_start_major_collection;
2078         collector->finish_major_collection = major_finish_major_collection;
2079         collector->have_computed_minor_collection_allowance = major_have_computer_minor_collection_allowance;
2080         collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2081         collector->obj_is_from_pinned_alloc = obj_is_from_pinned_alloc;
2082         collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2083         collector->get_num_major_sections = get_num_major_sections;
2084         collector->handle_gc_param = major_handle_gc_param;
2085         collector->print_gc_param_usage = major_print_gc_param_usage;
2086         collector->post_param_init = post_param_init;
2087         collector->is_valid_object = major_is_valid_object;
2088         collector->describe_pointer = major_describe_pointer;
2089         collector->count_cards = major_count_cards;
2090
2091         collector->major_ops.copy_or_mark_object = major_copy_or_mark_object_canonical;
2092         collector->major_ops.scan_object = major_scan_object;
2093 #ifdef SGEN_HAVE_CONCURRENT_MARK
2094         if (is_concurrent) {
2095                 collector->major_concurrent_ops.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2096                 collector->major_concurrent_ops.scan_object = major_scan_object_concurrent;
2097                 collector->major_concurrent_ops.scan_vtype = major_scan_vtype_concurrent;
2098         }
2099 #endif
2100
2101         /*cardtable requires major pages to be 8 cards aligned*/
2102         g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2103 }
2104
2105 #ifdef SGEN_HAVE_CONCURRENT_MARK
2106 void
2107 sgen_marksweep_init (SgenMajorCollector *collector)
2108 {
2109         sgen_marksweep_init_internal (collector, FALSE);
2110 }
2111
2112 void
2113 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2114 {
2115         sgen_marksweep_init_internal (collector, TRUE);
2116 }
2117 #endif
2118
2119 #endif