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 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
  48 #define MS_BLOCK_SIZE   ARCH_MIN_MS_BLOCK_SIZE
  49 #define MS_BLOCK_SIZE_SHIFT     ARCH_MIN_MS_BLOCK_SIZE_SHIFT
  50 #else
  51 #define MS_BLOCK_SIZE_SHIFT     14      /* INT FASTENABLE */
  52 #define MS_BLOCK_SIZE           (1 << MS_BLOCK_SIZE_SHIFT)
  53 #endif
  54 #define MAJOR_SECTION_SIZE      MS_BLOCK_SIZE
  55 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
  56
  57 /*
  58  * Don't allocate single blocks, but alloc a contingent of this many
  59  * blocks in one swoop.  This must be a power of two.
  60  */
  61 #define MS_BLOCK_ALLOC_NUM      32
  62
  63 /*
  64  * Number of bytes before the first object in a block.  At the start
  65  * of a block is the MSBlockHeader, then opional padding, then come
  66  * the objects, so this must be >= sizeof (MSBlockHeader).
  67  */
  68 #define MS_BLOCK_SKIP   ((sizeof (MSBlockHeader) + 15) & ~15)
  69
  70 #define MS_BLOCK_FREE   (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
  71
  72 #define MS_NUM_MARK_WORDS       ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
  73
  74 typedef struct _MSBlockInfo MSBlockInfo;
  75 struct _MSBlockInfo {
  76         int obj_size;
  77         int obj_size_index;
  78         unsigned int pinned : 1;
  79         unsigned int has_references : 1;
  80         unsigned int has_pinned : 1;    /* means cannot evacuate */
  81         unsigned int is_to_space : 1;
  82         unsigned int swept : 1;
  83         void **free_list;
  84         MSBlockInfo *next_free;
  85         size_t pin_queue_first_entry;
  86         size_t pin_queue_last_entry;
  87         guint8 *cardtable_mod_union;
  88         mword mark_words [MS_NUM_MARK_WORDS];
  89 };
  90
  91 #define MS_BLOCK_FOR_BLOCK_INFO(b)      ((char*)(b))
  92
  93 #define MS_BLOCK_OBJ(b,i)               (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i))
  94 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size)             (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
  95 #define MS_BLOCK_DATA_FOR_OBJ(o)        ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
  96
  97 typedef struct {
  98         MSBlockInfo info;
  99 } MSBlockHeader;
 100
 101 #define MS_BLOCK_FOR_OBJ(o)             (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
 102
 103 /* object index will always be small */
 104 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
 105
 106 //casting to int is fine since blocks are 32k
 107 #define MS_CALC_MARK_BIT(w,b,o)         do {                            \
 108                 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
 109                 if (sizeof (mword) == 4) {                              \
 110                         (w) = i >> 5;                                   \
 111                         (b) = i & 31;                                   \
 112                 } else {                                                \
 113                         (w) = i >> 6;                                   \
 114                         (b) = i & 63;                                   \
 115                 }                                                       \
 116         } while (0)
 117
 118 #define MS_MARK_BIT(bl,w,b)     ((bl)->mark_words [(w)] & (ONE_P << (b)))
 119 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
 120
 121 #define MS_OBJ_ALLOCED(o,b)     (*(void**)(o) && (*(char**)(o) < MS_BLOCK_FOR_BLOCK_INFO (b) || *(char**)(o) >= MS_BLOCK_FOR_BLOCK_INFO (b) + MS_BLOCK_SIZE))
 122
 123 #define MS_BLOCK_OBJ_SIZE_FACTOR        (pow (2.0, 1.0 / 3))
 124
 125 /*
 126  * This way we can lookup block object size indexes for sizes up to
 127  * 256 bytes with a single load.
 128  */
 129 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES      32
 130
 131 static int *block_obj_sizes;
 132 static int num_block_obj_sizes;
 133 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
 134
 135 #define MS_BLOCK_FLAG_PINNED    1
 136 #define MS_BLOCK_FLAG_REFS      2
 137
 138 #define MS_BLOCK_TYPE_MAX       4
 139
 140 static gboolean *evacuate_block_obj_sizes;
 141 static float evacuation_threshold = 0.666f;
 142 static float concurrent_evacuation_threshold = 0.666f;
 143 static gboolean want_evacuation = FALSE;
 144
 145 static gboolean lazy_sweep = TRUE;
 146 static gboolean have_swept;
 147
 148 static gboolean concurrent_mark;
 149
 150 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl)      SGEN_POINTER_IS_TAGGED_1 ((bl))
 151 #define BLOCK_TAG_HAS_REFERENCES(bl)            SGEN_POINTER_TAG_1 ((bl))
 152 #define BLOCK_UNTAG_HAS_REFERENCES(bl)          SGEN_POINTER_UNTAG_1 ((bl))
 153
 154 #define BLOCK_TAG(bl)   ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
 155
 156 /* all allocated blocks in the system */
 157 static SgenPointerQueue allocated_blocks;
 158
 159 /* non-allocated block free-list */
 160 static void *empty_blocks = NULL;
 161 static size_t num_empty_blocks = 0;
 162
 163 #define FOREACH_BLOCK(bl)       { size_t __index; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [__index]);
 164 #define FOREACH_BLOCK_HAS_REFERENCES(bl,hr)     { size_t __index; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = allocated_blocks.data [__index]; (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl)); (bl) = BLOCK_UNTAG_HAS_REFERENCES ((bl));
 165 #define END_FOREACH_BLOCK       } }
 166 #define DELETE_BLOCK_IN_FOREACH()       (allocated_blocks.data [__index] = NULL)
 167
 168 static size_t num_major_sections = 0;
 169 /* one free block list for each block object size */
 170 static MSBlockInfo **free_block_lists [MS_BLOCK_TYPE_MAX];
 171
 172 static guint64 stat_major_blocks_alloced = 0;
 173 static guint64 stat_major_blocks_freed = 0;
 174 static guint64 stat_major_blocks_lazy_swept = 0;
 175 static guint64 stat_major_objects_evacuated = 0;
 176
 177 #if SIZEOF_VOID_P != 8
 178 static guint64 stat_major_blocks_freed_ideal = 0;
 179 static guint64 stat_major_blocks_freed_less_ideal = 0;
 180 static guint64 stat_major_blocks_freed_individual = 0;
 181 static guint64 stat_major_blocks_alloced_less_ideal = 0;
 182 #endif
 183
 184 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
 185 static guint64 num_major_objects_marked = 0;
 186 #define INC_NUM_MAJOR_OBJECTS_MARKED()  (++num_major_objects_marked)
 187 #else
 188 #define INC_NUM_MAJOR_OBJECTS_MARKED()
 189 #endif
 190
 191 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
 192 static mono_mutex_t scanned_objects_list_lock;
 193 static SgenPointerQueue scanned_objects_list;
 194
 195 static void
 196 add_scanned_object (void *ptr)
 197 {
 198         if (!binary_protocol_is_enabled ())
 199                 return;
 200
 201         mono_mutex_lock (&scanned_objects_list_lock);
 202         sgen_pointer_queue_add (&scanned_objects_list, ptr);
 203         mono_mutex_unlock (&scanned_objects_list_lock);
 204 }
 205 #endif
 206
 207 static void
 208 sweep_block (MSBlockInfo *block, gboolean during_major_collection);
 209
 210 static int
 211 ms_find_block_obj_size_index (size_t size)
 212 {
 213         int i;
 214         SGEN_ASSERT (9, size <= SGEN_MAX_SMALL_OBJ_SIZE, "size %d is bigger than max small object size %d", size, SGEN_MAX_SMALL_OBJ_SIZE);
 215         for (i = 0; i < num_block_obj_sizes; ++i)
 216                 if (block_obj_sizes [i] >= size)
 217                         return i;
 218         g_error ("no object of size %d\n", size);
 219 }
 220
 221 #define FREE_BLOCKS_FROM(lists,p,r)     (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
 222 #define FREE_BLOCKS(p,r)                (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
 223
 224 #define MS_BLOCK_OBJ_SIZE_INDEX(s)                              \
 225         (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ?      \
 226          fast_block_obj_size_indexes [((s)+7)>>3] :             \
 227          ms_find_block_obj_size_index ((s)))
 228
 229 static void*
 230 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
 231 {
 232         char *start;
 233         if (nursery_align)
 234                 start = sgen_alloc_os_memory_aligned (nursery_size, nursery_align, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
 235         else
 236                 start = sgen_alloc_os_memory (nursery_size, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
 237
 238         return start;
 239 }
 240
 241 static void
 242 update_heap_boundaries_for_block (MSBlockInfo *block)
 243 {
 244         sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
 245 }
 246
 247 static void*
 248 ms_get_empty_block (void)
 249 {
 250         char *p;
 251         int i;
 252         void *block, *empty, *next;
 253
 254  retry:
 255         if (!empty_blocks) {
 256                 /*
 257                  * We try allocating MS_BLOCK_ALLOC_NUM blocks first.  If that's
 258                  * unsuccessful, we halve the number of blocks and try again, until we're at
 259                  * 1.  If that doesn't work, either, we assert.
 260                  */
 261                 int alloc_num = MS_BLOCK_ALLOC_NUM;
 262                 for (;;) {
 263                         p = sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE,
 264                                         alloc_num == 1 ? "major heap section" : NULL);
 265                         if (p)
 266                                 break;
 267                         alloc_num >>= 1;
 268                 }
 269
 270                 for (i = 0; i < alloc_num; ++i) {
 271                         block = p;
 272                         /*
 273                          * We do the free list update one after the
 274                          * other so that other threads can use the new
 275                          * blocks as quickly as possible.
 276                          */
 277                         do {
 278                                 empty = empty_blocks;
 279                                 *(void**)block = empty;
 280                         } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
 281                         p += MS_BLOCK_SIZE;
 282                 }
 283
 284                 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
 285
 286                 stat_major_blocks_alloced += alloc_num;
 287 #if SIZEOF_VOID_P != 8
 288                 if (alloc_num != MS_BLOCK_ALLOC_NUM)
 289                         stat_major_blocks_alloced_less_ideal += alloc_num;
 290 #endif
 291         }
 292
 293         do {
 294                 empty = empty_blocks;
 295                 if (!empty)
 296                         goto retry;
 297                 block = empty;
 298                 next = *(void**)block;
 299         } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
 300
 301         SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
 302
 303         *(void**)block = NULL;
 304
 305         g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
 306
 307         return block;
 308 }
 309
 310 static void
 311 ms_free_block (void *block)
 312 {
 313         void *empty;
 314
 315         sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
 316         memset (block, 0, MS_BLOCK_SIZE);
 317
 318         do {
 319                 empty = empty_blocks;
 320                 *(void**)block = empty;
 321         } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
 322
 323         SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
 324 }
 325
 326 //#define MARKSWEEP_CONSISTENCY_CHECK
 327
 328 #ifdef MARKSWEEP_CONSISTENCY_CHECK
 329 static void
 330 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
 331 {
 332         MSBlockInfo *b;
 333
 334         for (; block; block = block->next_free) {
 335                 g_assert (block->obj_size == size);
 336                 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
 337
 338                 /* blocks in the free lists must have at least
 339                    one free slot */
 340                 if (block->swept)
 341                         g_assert (block->free_list);
 342
 343                 /* the block must be in the allocated_blocks array */
 344                 g_assert (sgen_pointer_queue_find (&allocated_blocks, BLOCK_TAG (block)) != (size_t)-1);
 345         }
 346 }
 347
 348 static void
 349 check_empty_blocks (void)
 350 {
 351         void *p;
 352         size_t i = 0;
 353         for (p = empty_blocks; p; p = *(void**)p)
 354                 ++i;
 355         g_assert (i == num_empty_blocks);
 356 }
 357
 358 static void
 359 consistency_check (void)
 360 {
 361         MSBlockInfo *block;
 362         int i;
 363
 364         /* check all blocks */
 365         FOREACH_BLOCK (block) {
 366                 int count = MS_BLOCK_FREE / block->obj_size;
 367                 int num_free = 0;
 368                 void **free;
 369
 370                 /* check block header */
 371                 g_assert (((MSBlockHeader*)block->block)->info == block);
 372
 373                 /* count number of free slots */
 374                 for (i = 0; i < count; ++i) {
 375                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 376                         if (!MS_OBJ_ALLOCED (obj, block))
 377                                 ++num_free;
 378                 }
 379
 380                 /* check free list */
 381                 for (free = block->free_list; free; free = (void**)*free) {
 382                         g_assert (MS_BLOCK_FOR_OBJ (free) == block);
 383                         --num_free;
 384                 }
 385                 g_assert (num_free == 0);
 386
 387                 /* check all mark words are zero */
 388                 if (block->swept) {
 389                         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
 390                                 g_assert (block->mark_words [i] == 0);
 391                 }
 392         } END_FOREACH_BLOCK;
 393
 394         /* check free blocks */
 395         for (i = 0; i < num_block_obj_sizes; ++i) {
 396                 int j;
 397                 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
 398                         check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
 399         }
 400
 401         check_empty_blocks ();
 402 }
 403 #endif
 404
 405 static gboolean
 406 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
 407 {
 408         int size = block_obj_sizes [size_index];
 409         int count = MS_BLOCK_FREE / size;
 410         MSBlockInfo *info;
 411         MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, has_references);
 412         char *obj_start;
 413         int i;
 414
 415         if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
 416                 return FALSE;
 417
 418         info = (MSBlockInfo*)ms_get_empty_block ();
 419
 420         SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
 421
 422         info->obj_size = size;
 423         info->obj_size_index = size_index;
 424         info->pinned = pinned;
 425         info->has_references = has_references;
 426         info->has_pinned = pinned;
 427         /*
 428          * Blocks that are to-space are not evacuated from.  During an major collection
 429          * blocks are allocated for two reasons: evacuating objects from the nursery and
 430          * evacuating them from major blocks marked for evacuation.  In both cases we don't
 431          * want further evacuation.
 432          */
 433         info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD);
 434         info->swept = 1;
 435         info->cardtable_mod_union = NULL;
 436
 437         update_heap_boundaries_for_block (info);
 438
 439         /* build free list */
 440         obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
 441         info->free_list = (void**)obj_start;
 442         /* we're skipping the last one - it must be nulled */
 443         for (i = 0; i < count - 1; ++i) {
 444                 char *next_obj_start = obj_start + size;
 445                 *(void**)obj_start = next_obj_start;
 446                 obj_start = next_obj_start;
 447         }
 448         /* the last one */
 449         *(void**)obj_start = NULL;
 450
 451         info->next_free = free_blocks [size_index];
 452         free_blocks [size_index] = info;
 453
 454         sgen_pointer_queue_add (&allocated_blocks, BLOCK_TAG (info));
 455
 456         ++num_major_sections;
 457         return TRUE;
 458 }
 459
 460 static gboolean
 461 obj_is_from_pinned_alloc (char *ptr)
 462 {
 463         MSBlockInfo *block;
 464
 465         FOREACH_BLOCK (block) {
 466                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE)
 467                         return block->pinned;
 468         } END_FOREACH_BLOCK;
 469         return FALSE;
 470 }
 471
 472 static void*
 473 unlink_slot_from_free_list_uncontested (MSBlockInfo **free_blocks, int size_index)
 474 {
 475         MSBlockInfo *block;
 476         void *obj;
 477
 478         block = free_blocks [size_index];
 479         SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
 480
 481         if (G_UNLIKELY (!block->swept)) {
 482                 stat_major_blocks_lazy_swept ++;
 483                 sweep_block (block, FALSE);
 484         }
 485
 486         obj = block->free_list;
 487         SGEN_ASSERT (9, obj, "block %p in free list had no available object to alloc from", block);
 488
 489         block->free_list = *(void**)obj;
 490         if (!block->free_list) {
 491                 free_blocks [size_index] = block->next_free;
 492                 block->next_free = NULL;
 493         }
 494
 495         return obj;
 496 }
 497
 498 static void*
 499 alloc_obj (MonoVTable *vtable, size_t size, gboolean pinned, gboolean has_references)
 500 {
 501         int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 502         MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, has_references);
 503         void *obj;
 504
 505         if (!free_blocks [size_index]) {
 506                 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
 507                         return NULL;
 508         }
 509
 510         obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
 511
 512         *(MonoVTable**)obj = vtable;
 513
 514         return obj;
 515 }
 516
 517 static void*
 518 major_alloc_object (MonoVTable *vtable, size_t size, gboolean has_references)
 519 {
 520         return alloc_obj (vtable, size, FALSE, has_references);
 521 }
 522
 523 /*
 524  * We're not freeing the block if it's empty.  We leave that work for
 525  * the next major collection.
 526  *
 527  * This is just called from the domain clearing code, which runs in a
 528  * single thread and has the GC lock, so we don't need an extra lock.
 529  */
 530 static void
 531 free_object (char *obj, size_t size, gboolean pinned)
 532 {
 533         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 534         int word, bit;
 535
 536         if (!block->swept)
 537                 sweep_block (block, FALSE);
 538         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);
 539         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
 540         MS_CALC_MARK_BIT (word, bit, obj);
 541         SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set");
 542         if (!block->free_list) {
 543                 MSBlockInfo **free_blocks = FREE_BLOCKS (pinned, block->has_references);
 544                 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 545                 SGEN_ASSERT (9, !block->next_free, "block %p doesn't have a free-list of object but belongs to a free-list of blocks");
 546                 block->next_free = free_blocks [size_index];
 547                 free_blocks [size_index] = block;
 548         }
 549         memset (obj, 0, size);
 550         *(void**)obj = block->free_list;
 551         block->free_list = (void**)obj;
 552 }
 553
 554 static void
 555 major_free_non_pinned_object (char *obj, size_t size)
 556 {
 557         free_object (obj, size, FALSE);
 558 }
 559
 560 /* size is a multiple of SGEN_ALLOC_ALIGN */
 561 static void*
 562 major_alloc_small_pinned_obj (MonoVTable *vtable, size_t size, gboolean has_references)
 563 {
 564         void *res;
 565
 566         res = alloc_obj (vtable, size, TRUE, has_references);
 567          /*If we failed to alloc memory, we better try releasing memory
 568           *as pinned alloc is requested by the runtime.
 569           */
 570          if (!res) {
 571                 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE);
 572                 res = alloc_obj (vtable, size, TRUE, has_references);
 573          }
 574          return res;
 575 }
 576
 577 static void
 578 free_pinned_object (char *obj, size_t size)
 579 {
 580         free_object (obj, size, TRUE);
 581 }
 582
 583 /*
 584  * size is already rounded up and we hold the GC lock.
 585  */
 586 static void*
 587 major_alloc_degraded (MonoVTable *vtable, size_t size)
 588 {
 589         void *obj;
 590         size_t old_num_sections;
 591
 592         old_num_sections = num_major_sections;
 593
 594         obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
 595         if (G_LIKELY (obj)) {
 596                 HEAVY_STAT (++stat_objects_alloced_degraded);
 597                 HEAVY_STAT (stat_bytes_alloced_degraded += size);
 598                 g_assert (num_major_sections >= old_num_sections);
 599                 sgen_register_major_sections_alloced (num_major_sections - old_num_sections);
 600         }
 601         return obj;
 602 }
 603
 604 /*
 605  * obj is some object.  If it's not in the major heap (i.e. if it's in
 606  * the nursery or LOS), return FALSE.  Otherwise return whether it's
 607  * been marked or copied.
 608  */
 609 static gboolean
 610 major_is_object_live (char *obj)
 611 {
 612         MSBlockInfo *block;
 613         int word, bit;
 614         mword objsize;
 615
 616         if (sgen_ptr_in_nursery (obj))
 617                 return FALSE;
 618
 619         objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
 620
 621         /* LOS */
 622         if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
 623                 return FALSE;
 624
 625         /* now we know it's in a major block */
 626         block = MS_BLOCK_FOR_OBJ (obj);
 627         SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?");
 628         MS_CALC_MARK_BIT (word, bit, obj);
 629         return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
 630 }
 631
 632 static gboolean
 633 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
 634 {
 635         MSBlockInfo *block;
 636
 637         FOREACH_BLOCK (block) {
 638                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
 639                         int count = MS_BLOCK_FREE / block->obj_size;
 640                         int i;
 641
 642                         *start = NULL;
 643                         for (i = 0; i <= count; ++i) {
 644                                 if (ptr >= MS_BLOCK_OBJ (block, i) && ptr < MS_BLOCK_OBJ (block, i + 1)) {
 645                                         *start = MS_BLOCK_OBJ (block, i);
 646                                         break;
 647                                 }
 648                         }
 649                         return !block->pinned;
 650                 }
 651         } END_FOREACH_BLOCK;
 652         return FALSE;
 653 }
 654
 655 static void
 656 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
 657 {
 658         gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
 659         gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
 660         gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
 661         MSBlockInfo *block;
 662
 663         FOREACH_BLOCK (block) {
 664                 int count = MS_BLOCK_FREE / block->obj_size;
 665                 int i;
 666
 667                 if (block->pinned && !pinned)
 668                         continue;
 669                 if (!block->pinned && !non_pinned)
 670                         continue;
 671                 if (sweep && lazy_sweep) {
 672                         sweep_block (block, FALSE);
 673                         SGEN_ASSERT (0, block->swept, "Block must be swept after sweeping");
 674                 }
 675
 676                 for (i = 0; i < count; ++i) {
 677                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 678                         if (!block->swept) {
 679                                 int word, bit;
 680                                 MS_CALC_MARK_BIT (word, bit, obj);
 681                                 if (!MS_MARK_BIT (block, word, bit))
 682                                         continue;
 683                         }
 684                         if (MS_OBJ_ALLOCED (obj, block))
 685                                 callback ((char*)obj, block->obj_size, data);
 686                 }
 687         } END_FOREACH_BLOCK;
 688 }
 689
 690 static gboolean
 691 major_is_valid_object (char *object)
 692 {
 693         MSBlockInfo *block;
 694
 695         FOREACH_BLOCK (block) {
 696                 int idx;
 697                 char *obj;
 698
 699                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
 700                         continue;
 701
 702                 idx = MS_BLOCK_OBJ_INDEX (object, block);
 703                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 704                 if (obj != object)
 705                         return FALSE;
 706                 return MS_OBJ_ALLOCED (obj, block);
 707         } END_FOREACH_BLOCK;
 708
 709         return FALSE;
 710 }
 711
 712
 713 static MonoVTable*
 714 major_describe_pointer (char *ptr)
 715 {
 716         MSBlockInfo *block;
 717
 718         FOREACH_BLOCK (block) {
 719                 int idx;
 720                 char *obj;
 721                 gboolean live;
 722                 MonoVTable *vtable;
 723                 int w, b;
 724                 gboolean marked;
 725
 726                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
 727                         continue;
 728
 729                 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
 730                         MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
 731
 732                 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
 733                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 734                 live = MS_OBJ_ALLOCED (obj, block);
 735                 vtable = live ? (MonoVTable*)SGEN_LOAD_VTABLE (obj) : NULL;
 736
 737                 MS_CALC_MARK_BIT (w, b, obj);
 738                 marked = MS_MARK_BIT (block, w, b);
 739
 740                 if (obj == ptr) {
 741                         SGEN_LOG (0, "\t(");
 742                         if (live)
 743                                 SGEN_LOG (0, "object");
 744                         else
 745                                 SGEN_LOG (0, "dead-object");
 746                 } else {
 747                         if (live)
 748                                 SGEN_LOG (0, "interior-ptr offset %td", ptr - obj);
 749                         else
 750                                 SGEN_LOG (0, "dead-interior-ptr offset %td", ptr - obj);
 751                 }
 752
 753                 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
 754
 755                 return vtable;
 756         } END_FOREACH_BLOCK;
 757
 758         return NULL;
 759 }
 760
 761 static void
 762 major_check_scan_starts (void)
 763 {
 764 }
 765
 766 static void
 767 major_dump_heap (FILE *heap_dump_file)
 768 {
 769         MSBlockInfo *block;
 770         int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
 771         int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
 772         int i;
 773
 774         for (i = 0; i < num_block_obj_sizes; ++i)
 775                 slots_available [i] = slots_used [i] = 0;
 776
 777         FOREACH_BLOCK (block) {
 778                 int index = ms_find_block_obj_size_index (block->obj_size);
 779                 int count = MS_BLOCK_FREE / block->obj_size;
 780
 781                 slots_available [index] += count;
 782                 for (i = 0; i < count; ++i) {
 783                         if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
 784                                 ++slots_used [index];
 785                 }
 786         } END_FOREACH_BLOCK;
 787
 788         fprintf (heap_dump_file, "<occupancies>\n");
 789         for (i = 0; i < num_block_obj_sizes; ++i) {
 790                 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
 791                                 block_obj_sizes [i], slots_available [i], slots_used [i]);
 792         }
 793         fprintf (heap_dump_file, "</occupancies>\n");
 794
 795         FOREACH_BLOCK (block) {
 796                 int count = MS_BLOCK_FREE / block->obj_size;
 797                 int i;
 798                 int start = -1;
 799
 800                 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
 801
 802                 for (i = 0; i <= count; ++i) {
 803                         if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
 804                                 if (start < 0)
 805                                         start = i;
 806                         } else {
 807                                 if (start >= 0) {
 808                                         sgen_dump_occupied (MS_BLOCK_OBJ (block, start), MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
 809                                         start = -1;
 810                                 }
 811                         }
 812                 }
 813
 814                 fprintf (heap_dump_file, "</section>\n");
 815         } END_FOREACH_BLOCK;
 816 }
 817
 818 #define LOAD_VTABLE     SGEN_LOAD_VTABLE
 819
 820 #define MS_MARK_OBJECT_AND_ENQUEUE_CHECKED(obj,desc,block,queue) do {   \
 821                 int __word, __bit;                                      \
 822                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
 823                 if (!MS_MARK_BIT ((block), __word, __bit) && MS_OBJ_ALLOCED ((obj), (block))) { \
 824                         MS_SET_MARK_BIT ((block), __word, __bit);       \
 825                         if (sgen_gc_descr_has_references (desc))                        \
 826                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
 827                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((MonoObject*)(obj))); \
 828                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
 829                 }                                                       \
 830         } while (0)
 831 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do {           \
 832                 int __word, __bit;                                      \
 833                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
 834                 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
 835                 if (!MS_MARK_BIT ((block), __word, __bit)) {            \
 836                         MS_SET_MARK_BIT ((block), __word, __bit);       \
 837                         if (sgen_gc_descr_has_references (desc))                        \
 838                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
 839                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((MonoObject*)(obj))); \
 840                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
 841                 }                                                       \
 842         } while (0)
 843
 844 static void
 845 pin_major_object (char *obj, SgenGrayQueue *queue)
 846 {
 847         MSBlockInfo *block;
 848
 849         if (concurrent_mark)
 850                 g_assert_not_reached ();
 851
 852         block = MS_BLOCK_FOR_OBJ (obj);
 853         block->has_pinned = TRUE;
 854         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
 855 }
 856
 857 #include "sgen-major-copy-object.h"
 858
 859 static void
 860 major_copy_or_mark_object_with_evacuation_concurrent (void **ptr, void *obj, SgenGrayQueue *queue)
 861 {
 862         SGEN_ASSERT (9, sgen_concurrent_collection_in_progress (), "Why are we scanning concurrently when there's no concurrent collection on?");
 863         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!");
 864
 865         g_assert (!SGEN_OBJECT_IS_FORWARDED (obj));
 866
 867         if (!sgen_ptr_in_nursery (obj)) {
 868                 mword objsize;
 869
 870                 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
 871
 872                 if (objsize <= SGEN_MAX_SMALL_OBJ_SIZE) {
 873                         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 874                         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
 875                 } else {
 876                         if (sgen_los_object_is_pinned (obj))
 877                                 return;
 878
 879 #ifdef ENABLE_DTRACE
 880                         if (G_UNLIKELY (MONO_GC_OBJ_PINNED_ENABLED ())) {
 881                                 MonoVTable *vt = (MonoVTable*)SGEN_LOAD_VTABLE (obj);
 882                                 MONO_GC_OBJ_PINNED ((mword)obj, sgen_safe_object_get_size (obj), vt->klass->name_space, vt->klass->name, GENERATION_OLD);
 883                         }
 884 #endif
 885
 886                         sgen_los_pin_object (obj);
 887                         if (SGEN_OBJECT_HAS_REFERENCES (obj))
 888                                 GRAY_OBJECT_ENQUEUE (queue, obj, sgen_obj_get_descriptor (obj));
 889                         INC_NUM_MAJOR_OBJECTS_MARKED ();
 890                 }
 891         }
 892 }
 893
 894 static long long
 895 major_get_and_reset_num_major_objects_marked (void)
 896 {
 897 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
 898         long long num = num_major_objects_marked;
 899         num_major_objects_marked = 0;
 900         return num;
 901 #else
 902         return 0;
 903 #endif
 904 }
 905
 906 #define PREFETCH_CARDS          1       /* BOOL FASTENABLE */
 907 #if !PREFETCH_CARDS
 908 #undef PREFETCH_CARDS
 909 #endif
 910
 911 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
 912 #if defined(PLATFORM_MACOSX)
 913 #define GCC_VERSION (__GNUC__ * 10000 \
 914                                + __GNUC_MINOR__ * 100 \
 915                                + __GNUC_PATCHLEVEL__)
 916 #if GCC_VERSION <= 40300
 917 #undef PREFETCH_CARDS
 918 #endif
 919 #endif
 920
 921 #ifdef HEAVY_STATISTICS
 922 static guint64 stat_optimized_copy;
 923 static guint64 stat_optimized_copy_nursery;
 924 static guint64 stat_optimized_copy_nursery_forwarded;
 925 static guint64 stat_optimized_copy_nursery_pinned;
 926 static guint64 stat_optimized_copy_major;
 927 static guint64 stat_optimized_copy_major_small_fast;
 928 static guint64 stat_optimized_copy_major_small_slow;
 929 static guint64 stat_optimized_copy_major_large;
 930 static guint64 stat_optimized_copy_major_forwarded;
 931 static guint64 stat_optimized_copy_major_small_evacuate;
 932 static guint64 stat_optimized_major_scan;
 933 static guint64 stat_optimized_major_scan_no_refs;
 934
 935 static guint64 stat_drain_prefetch_fills;
 936 static guint64 stat_drain_prefetch_fill_failures;
 937 static guint64 stat_drain_loops;
 938 #endif
 939
 940 static void major_scan_object_with_evacuation (char *start, mword desc, SgenGrayQueue *queue);
 941
 942 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_no_evacuation
 943 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_no_evacuation
 944 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_no_evacuation
 945 #include "sgen-marksweep-drain-gray-stack.h"
 946
 947 #define COPY_OR_MARK_WITH_EVACUATION
 948 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_with_evacuation
 949 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_with_evacuation
 950 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_with_evacuation
 951 #include "sgen-marksweep-drain-gray-stack.h"
 952
 953 static gboolean
 954 drain_gray_stack (ScanCopyContext ctx)
 955 {
 956         gboolean evacuation = FALSE;
 957         int i;
 958         for (i = 0; i < num_block_obj_sizes; ++i) {
 959                 if (evacuate_block_obj_sizes [i]) {
 960                         evacuation = TRUE;
 961                         break;
 962                 }
 963         }
 964
 965         if (evacuation)
 966                 return drain_gray_stack_with_evacuation (ctx);
 967         else
 968                 return drain_gray_stack_no_evacuation (ctx);
 969 }
 970
 971 #include "sgen-marksweep-scan-object-concurrent.h"
 972
 973 static void
 974 major_copy_or_mark_object_canonical (void **ptr, SgenGrayQueue *queue)
 975 {
 976         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
 977 }
 978
 979 static void
 980 major_copy_or_mark_object_concurrent_canonical (void **ptr, SgenGrayQueue *queue)
 981 {
 982         major_copy_or_mark_object_with_evacuation_concurrent (ptr, *ptr, queue);
 983 }
 984
 985 static void
 986 mark_pinned_objects_in_block (MSBlockInfo *block, SgenGrayQueue *queue)
 987 {
 988         void **entry, **end;
 989         int last_index = -1;
 990
 991         if (block->pin_queue_first_entry == block->pin_queue_last_entry)
 992                 return;
 993
 994         block->has_pinned = TRUE;
 995
 996         entry = sgen_pinning_get_entry (block->pin_queue_first_entry);
 997         end = sgen_pinning_get_entry (block->pin_queue_last_entry);
 998
 999         for (; entry < end; ++entry) {
1000                 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1001                 char *obj;
1002                 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);
1003                 if (index == last_index)
1004                         continue;
1005                 obj = MS_BLOCK_OBJ (block, index);
1006                 MS_MARK_OBJECT_AND_ENQUEUE_CHECKED (obj, sgen_obj_get_descriptor (obj), block, queue);
1007                 last_index = index;
1008         }
1009 }
1010
1011 static inline void
1012 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1013 {
1014         int obj_index;
1015
1016         for (obj_index = 0; obj_index < count; ++obj_index) {
1017                 int word, bit;
1018                 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1019
1020                 MS_CALC_MARK_BIT (word, bit, obj);
1021                 if (MS_MARK_BIT (block, word, bit)) {
1022                         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1023                 } else {
1024                         /* an unmarked object */
1025                         if (MS_OBJ_ALLOCED (obj, block)) {
1026                                 /*
1027                                  * FIXME: Merge consecutive
1028                                  * slots for lower reporting
1029                                  * overhead.  Maybe memset
1030                                  * will also benefit?
1031                                  */
1032                                 binary_protocol_empty (obj, obj_size);
1033                                 MONO_GC_MAJOR_SWEPT ((mword)obj, obj_size);
1034                                 memset (obj, 0, obj_size);
1035                         }
1036                         *(void**)obj = block->free_list;
1037                         block->free_list = obj;
1038                 }
1039         }
1040 }
1041
1042 /*
1043  * sweep_block:
1044  *
1045  *   Traverse BLOCK, freeing and zeroing unused objects.
1046  */
1047 static void
1048 sweep_block (MSBlockInfo *block, gboolean during_major_collection)
1049 {
1050         int count;
1051         void *reversed = NULL;
1052
1053         if (!during_major_collection)
1054                 g_assert (!sgen_concurrent_collection_in_progress ());
1055
1056         if (block->swept)
1057                 return;
1058
1059         count = MS_BLOCK_FREE / block->obj_size;
1060
1061         block->free_list = NULL;
1062
1063         /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1064         // FIXME: Add more sizes
1065         switch (block->obj_size) {
1066         case 16:
1067                 sweep_block_for_size (block, count, 16);
1068                 break;
1069         default:
1070                 sweep_block_for_size (block, count, block->obj_size);
1071                 break;
1072         }
1073
1074         /* reset mark bits */
1075         memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1076
1077         /* Reverse free list so that it's in address order */
1078         reversed = NULL;
1079         while (block->free_list) {
1080                 void *next = *(void**)block->free_list;
1081                 *(void**)block->free_list = reversed;
1082                 reversed = block->free_list;
1083                 block->free_list = next;
1084         }
1085         block->free_list = reversed;
1086
1087         block->swept = 1;
1088 }
1089
1090 static inline int
1091 bitcount (mword d)
1092 {
1093         int count = 0;
1094
1095 #ifdef __GNUC__
1096         if (sizeof (mword) == sizeof (unsigned long))
1097                 count += __builtin_popcountl (d);
1098         else
1099                 count += __builtin_popcount (d);
1100 #else
1101         while (d) {
1102                 count ++;
1103                 d &= (d - 1);
1104         }
1105 #endif
1106         return count;
1107 }
1108
1109 static void
1110 ms_sweep (void)
1111 {
1112         int i;
1113         MSBlockInfo *block;
1114
1115         /* statistics for evacuation */
1116         int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
1117         int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
1118         int *num_blocks = alloca (sizeof (int) * num_block_obj_sizes);
1119
1120         mword total_evacuate_heap = 0;
1121         mword total_evacuate_saved = 0;
1122
1123         for (i = 0; i < num_block_obj_sizes; ++i)
1124                 slots_available [i] = slots_used [i] = num_blocks [i] = 0;
1125
1126         /* clear all the free lists */
1127         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1128                 MSBlockInfo **free_blocks = free_block_lists [i];
1129                 int j;
1130                 for (j = 0; j < num_block_obj_sizes; ++j)
1131                         free_blocks [j] = NULL;
1132         }
1133
1134         /* traverse all blocks, free and zero unmarked objects */
1135         FOREACH_BLOCK (block) {
1136                 int count;
1137                 gboolean have_live = FALSE;
1138                 gboolean has_pinned;
1139                 gboolean have_free = FALSE;
1140                 int obj_size_index;
1141                 int nused = 0;
1142
1143                 obj_size_index = block->obj_size_index;
1144
1145                 has_pinned = block->has_pinned;
1146                 block->has_pinned = block->pinned;
1147
1148                 block->is_to_space = FALSE;
1149                 block->swept = 0;
1150
1151                 count = MS_BLOCK_FREE / block->obj_size;
1152
1153                 if (block->cardtable_mod_union) {
1154                         sgen_free_internal_dynamic (block->cardtable_mod_union, CARDS_PER_BLOCK, INTERNAL_MEM_CARDTABLE_MOD_UNION);
1155                         block->cardtable_mod_union = NULL;
1156                 }
1157
1158                 /* Count marked objects in the block */
1159                 for (i = 0; i < MS_NUM_MARK_WORDS; ++i) {
1160                         nused += bitcount (block->mark_words [i]);
1161                 }
1162                 if (nused) {
1163                         have_live = TRUE;
1164                 }
1165                 if (nused < count)
1166                         have_free = TRUE;
1167
1168                 if (!lazy_sweep)
1169                         sweep_block (block, TRUE);
1170
1171                 if (have_live) {
1172                         if (!has_pinned) {
1173                                 ++num_blocks [obj_size_index];
1174                                 slots_used [obj_size_index] += nused;
1175                                 slots_available [obj_size_index] += count;
1176                         }
1177
1178                         /*
1179                          * If there are free slots in the block, add
1180                          * the block to the corresponding free list.
1181                          */
1182                         if (have_free) {
1183                                 MSBlockInfo **free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1184                                 int index = MS_BLOCK_OBJ_SIZE_INDEX (block->obj_size);
1185                                 block->next_free = free_blocks [index];
1186                                 free_blocks [index] = block;
1187                         }
1188
1189                         update_heap_boundaries_for_block (block);
1190                 } else {
1191                         /*
1192                          * Blocks without live objects are removed from the
1193                          * block list and freed.
1194                          */
1195                         DELETE_BLOCK_IN_FOREACH ();
1196
1197                         binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1198                         ms_free_block (block);
1199
1200                         --num_major_sections;
1201                 }
1202         } END_FOREACH_BLOCK;
1203         sgen_pointer_queue_remove_nulls (&allocated_blocks);
1204
1205         for (i = 0; i < num_block_obj_sizes; ++i) {
1206                 float usage = (float)slots_used [i] / (float)slots_available [i];
1207                 if (num_blocks [i] > 5 && usage < evacuation_threshold) {
1208                         evacuate_block_obj_sizes [i] = TRUE;
1209                         /*
1210                         g_print ("slot size %d - %d of %d used\n",
1211                                         block_obj_sizes [i], slots_used [i], slots_available [i]);
1212                         */
1213                 } else {
1214                         evacuate_block_obj_sizes [i] = FALSE;
1215                 }
1216                 {
1217                         mword total_bytes = block_obj_sizes [i] * slots_available [i];
1218                         total_evacuate_heap += total_bytes;
1219                         if (evacuate_block_obj_sizes [i])
1220                                 total_evacuate_saved += total_bytes - block_obj_sizes [i] * slots_used [i];
1221                 }
1222         }
1223
1224         want_evacuation = (float)total_evacuate_saved / (float)total_evacuate_heap > (1 - concurrent_evacuation_threshold);
1225
1226         have_swept = TRUE;
1227 }
1228
1229 static void
1230 major_sweep (void)
1231 {
1232         ms_sweep ();
1233 }
1234
1235 static int count_pinned_ref;
1236 static int count_pinned_nonref;
1237 static int count_nonpinned_ref;
1238 static int count_nonpinned_nonref;
1239
1240 static void
1241 count_nonpinned_callback (char *obj, size_t size, void *data)
1242 {
1243         MonoVTable *vtable = (MonoVTable*)LOAD_VTABLE (obj);
1244
1245         if (vtable->klass->has_references)
1246                 ++count_nonpinned_ref;
1247         else
1248                 ++count_nonpinned_nonref;
1249 }
1250
1251 static void
1252 count_pinned_callback (char *obj, size_t size, void *data)
1253 {
1254         MonoVTable *vtable = (MonoVTable*)LOAD_VTABLE (obj);
1255
1256         if (vtable->klass->has_references)
1257                 ++count_pinned_ref;
1258         else
1259                 ++count_pinned_nonref;
1260 }
1261
1262 static G_GNUC_UNUSED void
1263 count_ref_nonref_objs (void)
1264 {
1265         int total;
1266
1267         count_pinned_ref = 0;
1268         count_pinned_nonref = 0;
1269         count_nonpinned_ref = 0;
1270         count_nonpinned_nonref = 0;
1271
1272         major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1273         major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1274
1275         total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1276
1277         g_print ("ref: %d pinned %d non-pinned   non-ref: %d pinned %d non-pinned  --  %.1f\n",
1278                         count_pinned_ref, count_nonpinned_ref,
1279                         count_pinned_nonref, count_nonpinned_nonref,
1280                         (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1281 }
1282
1283 static int
1284 ms_calculate_block_obj_sizes (double factor, int *arr)
1285 {
1286         double target_size = sizeof (MonoObject);
1287         int num_sizes = 0;
1288         int last_size = 0;
1289
1290         do {
1291                 int target_count = (int)ceil (MS_BLOCK_FREE / target_size);
1292                 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1293
1294                 if (size != last_size) {
1295                         if (arr)
1296                                 arr [num_sizes] = size;
1297                         ++num_sizes;
1298                         last_size = size;
1299                 }
1300
1301                 target_size *= factor;
1302         } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1303
1304         return num_sizes;
1305 }
1306
1307 /* only valid during minor collections */
1308 static mword old_num_major_sections;
1309
1310 static void
1311 major_start_nursery_collection (void)
1312 {
1313 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1314         consistency_check ();
1315 #endif
1316
1317         old_num_major_sections = num_major_sections;
1318 }
1319
1320 static void
1321 major_finish_nursery_collection (void)
1322 {
1323 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1324         consistency_check ();
1325 #endif
1326         sgen_register_major_sections_alloced (num_major_sections - old_num_major_sections);
1327 }
1328
1329 static void
1330 major_start_major_collection (void)
1331 {
1332         int i;
1333
1334         /* clear the free lists */
1335         for (i = 0; i < num_block_obj_sizes; ++i) {
1336                 if (!evacuate_block_obj_sizes [i])
1337                         continue;
1338
1339                 free_block_lists [0][i] = NULL;
1340                 free_block_lists [MS_BLOCK_FLAG_REFS][i] = NULL;
1341         }
1342
1343         // Sweep all unswept blocks
1344         if (lazy_sweep) {
1345                 MSBlockInfo *block;
1346
1347                 MONO_GC_SWEEP_BEGIN (GENERATION_OLD, TRUE);
1348
1349                 FOREACH_BLOCK (block) {
1350                         sweep_block (block, TRUE);
1351                 } END_FOREACH_BLOCK;
1352
1353                 MONO_GC_SWEEP_END (GENERATION_OLD, TRUE);
1354         }
1355 }
1356
1357 static void
1358 major_finish_major_collection (ScannedObjectCounts *counts)
1359 {
1360 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1361         if (binary_protocol_is_enabled ()) {
1362                 counts->num_scanned_objects = scanned_objects_list.next_slot;
1363
1364                 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1365                 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1366
1367                 sgen_pointer_queue_clear (&scanned_objects_list);
1368         }
1369 #endif
1370 }
1371
1372 #if SIZEOF_VOID_P != 8
1373 static int
1374 compare_pointers (const void *va, const void *vb) {
1375         char *a = *(char**)va, *b = *(char**)vb;
1376         if (a < b)
1377                 return -1;
1378         if (a > b)
1379                 return 1;
1380         return 0;
1381 }
1382 #endif
1383
1384 static void
1385 major_have_computer_minor_collection_allowance (void)
1386 {
1387         size_t section_reserve = sgen_get_minor_collection_allowance () / MS_BLOCK_SIZE;
1388
1389         g_assert (have_swept);
1390
1391 #if SIZEOF_VOID_P != 8
1392         {
1393                 int i, num_empty_blocks_orig, num_blocks, arr_length;
1394                 void *block;
1395                 void **empty_block_arr;
1396                 void **rebuild_next;
1397
1398 #ifdef TARGET_WIN32
1399                 /*
1400                  * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1401                  * a VirtualAlloc ()-ed block.
1402                  */
1403                 return;
1404 #endif
1405
1406                 if (num_empty_blocks <= section_reserve)
1407                         return;
1408                 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1409
1410                 num_empty_blocks_orig = num_empty_blocks;
1411                 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1412                                 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1413                 if (!empty_block_arr)
1414                         goto fallback;
1415
1416                 i = 0;
1417                 for (block = empty_blocks; block; block = *(void**)block)
1418                         empty_block_arr [i++] = block;
1419                 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1420
1421                 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1422
1423                 /*
1424                  * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1425                  * contiguous ones.  If we do, we free them.  If that's not enough to get to
1426                  * section_reserve, we halve the number of contiguous blocks we're looking
1427                  * for and have another go, until we're done with looking for pairs of
1428                  * blocks, at which point we give up and go to the fallback.
1429                  */
1430                 arr_length = num_empty_blocks_orig;
1431                 num_blocks = MS_BLOCK_ALLOC_NUM;
1432                 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1433                         int first = -1;
1434                         int dest = 0;
1435
1436                         dest = 0;
1437                         for (i = 0; i < arr_length; ++i) {
1438                                 int d = dest;
1439                                 void *block = empty_block_arr [i];
1440                                 SGEN_ASSERT (0, block, "we're not shifting correctly");
1441                                 if (i != dest) {
1442                                         empty_block_arr [dest] = block;
1443                                         /*
1444                                          * This is not strictly necessary, but we're
1445                                          * cautious.
1446                                          */
1447                                         empty_block_arr [i] = NULL;
1448                                 }
1449                                 ++dest;
1450
1451                                 if (first < 0) {
1452                                         first = d;
1453                                         continue;
1454                                 }
1455
1456                                 SGEN_ASSERT (0, first >= 0 && d > first, "algorithm is wrong");
1457
1458                                 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
1459                                         first = d;
1460                                         continue;
1461                                 }
1462
1463                                 if (d + 1 - first == num_blocks) {
1464                                         /*
1465                                          * We found num_blocks contiguous blocks.  Free them
1466                                          * and null their array entries.  As an optimization
1467                                          * we could, instead of nulling the entries, shift
1468                                          * the following entries over to the left, while
1469                                          * we're iterating.
1470                                          */
1471                                         int j;
1472                                         sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
1473                                         for (j = first; j <= d; ++j)
1474                                                 empty_block_arr [j] = NULL;
1475                                         dest = first;
1476                                         first = -1;
1477
1478                                         num_empty_blocks -= num_blocks;
1479
1480                                         stat_major_blocks_freed += num_blocks;
1481                                         if (num_blocks == MS_BLOCK_ALLOC_NUM)
1482                                                 stat_major_blocks_freed_ideal += num_blocks;
1483                                         else
1484                                                 stat_major_blocks_freed_less_ideal += num_blocks;
1485
1486                                 }
1487                         }
1488
1489                         SGEN_ASSERT (0, dest <= i && dest <= arr_length, "array length is off");
1490                         arr_length = dest;
1491                         SGEN_ASSERT (0, arr_length == num_empty_blocks, "array length is off");
1492
1493                         num_blocks >>= 1;
1494                 }
1495
1496                 /* rebuild empty_blocks free list */
1497                 rebuild_next = (void**)&empty_blocks;
1498                 for (i = 0; i < arr_length; ++i) {
1499                         void *block = empty_block_arr [i];
1500                         SGEN_ASSERT (0, block, "we're missing blocks");
1501                         *rebuild_next = block;
1502                         rebuild_next = (void**)block;
1503                 }
1504                 *rebuild_next = NULL;
1505
1506                 /* free array */
1507                 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
1508         }
1509
1510         SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
1511
1512  fallback:
1513         /*
1514          * This is our threshold.  If there's not more empty than used blocks, we won't
1515          * release uncontiguous blocks, in fear of fragmenting the address space.
1516          */
1517         if (num_empty_blocks <= num_major_sections)
1518                 return;
1519 #endif
1520
1521         while (num_empty_blocks > section_reserve) {
1522                 void *next = *(void**)empty_blocks;
1523                 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
1524                 empty_blocks = next;
1525                 /*
1526                  * Needs not be atomic because this is running
1527                  * single-threaded.
1528                  */
1529                 --num_empty_blocks;
1530
1531                 ++stat_major_blocks_freed;
1532 #if SIZEOF_VOID_P != 8
1533                 ++stat_major_blocks_freed_individual;
1534 #endif
1535         }
1536 }
1537
1538 static void
1539 major_find_pin_queue_start_ends (SgenGrayQueue *queue)
1540 {
1541         MSBlockInfo *block;
1542
1543         FOREACH_BLOCK (block) {
1544                 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
1545                                 &block->pin_queue_first_entry, &block->pin_queue_last_entry);
1546         } END_FOREACH_BLOCK;
1547 }
1548
1549 static void
1550 major_pin_objects (SgenGrayQueue *queue)
1551 {
1552         MSBlockInfo *block;
1553
1554         FOREACH_BLOCK (block) {
1555                 mark_pinned_objects_in_block (block, queue);
1556         } END_FOREACH_BLOCK;
1557 }
1558
1559 static void
1560 major_init_to_space (void)
1561 {
1562 }
1563
1564 static void
1565 major_report_pinned_memory_usage (void)
1566 {
1567         g_assert_not_reached ();
1568 }
1569
1570 static gint64
1571 major_get_used_size (void)
1572 {
1573         gint64 size = 0;
1574         MSBlockInfo *block;
1575
1576         FOREACH_BLOCK (block) {
1577                 int count = MS_BLOCK_FREE / block->obj_size;
1578                 void **iter;
1579                 size += count * block->obj_size;
1580                 for (iter = block->free_list; iter; iter = (void**)*iter)
1581                         size -= block->obj_size;
1582         } END_FOREACH_BLOCK;
1583
1584         return size;
1585 }
1586
1587 static size_t
1588 get_num_major_sections (void)
1589 {
1590         return num_major_sections;
1591 }
1592
1593 static gboolean
1594 major_handle_gc_param (const char *opt)
1595 {
1596         if (g_str_has_prefix (opt, "evacuation-threshold=")) {
1597                 const char *arg = strchr (opt, '=') + 1;
1598                 int percentage = atoi (arg);
1599                 if (percentage < 0 || percentage > 100) {
1600                         fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
1601                         exit (1);
1602                 }
1603                 evacuation_threshold = (float)percentage / 100.0f;
1604                 return TRUE;
1605         } else if (!strcmp (opt, "lazy-sweep")) {
1606                 lazy_sweep = TRUE;
1607                 return TRUE;
1608         } else if (!strcmp (opt, "no-lazy-sweep")) {
1609                 lazy_sweep = FALSE;
1610                 return TRUE;
1611         }
1612
1613         return FALSE;
1614 }
1615
1616 static void
1617 major_print_gc_param_usage (void)
1618 {
1619         fprintf (stderr,
1620                         ""
1621                         "  evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
1622                         "  (no-)lazy-sweep\n"
1623                         );
1624 }
1625
1626 static void
1627 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
1628 {
1629         MSBlockInfo *block;
1630         gboolean has_references;
1631
1632         FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1633                 if (has_references)
1634                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1635         } END_FOREACH_BLOCK;
1636 }
1637
1638 #ifdef HEAVY_STATISTICS
1639 extern guint64 marked_cards;
1640 extern guint64 scanned_cards;
1641 extern guint64 scanned_objects;
1642 extern guint64 remarked_cards;
1643 #endif
1644
1645 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
1646 /*
1647  * MS blocks are 16K aligned.
1648  * Cardtables are 4K aligned, at least.
1649  * This means that the cardtable of a given block is 32 bytes aligned.
1650  */
1651 static guint8*
1652 initial_skip_card (guint8 *card_data)
1653 {
1654         mword *cards = (mword*)card_data;
1655         mword card;
1656         int i;
1657         for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
1658                 card = cards [i];
1659                 if (card)
1660                         break;
1661         }
1662
1663         if (i == CARD_WORDS_PER_BLOCK)
1664                 return card_data + CARDS_PER_BLOCK;
1665
1666 #if defined(__i386__) && defined(__GNUC__)
1667         return card_data + i * 4 +  (__builtin_ffs (card) - 1) / 8;
1668 #elif defined(__x86_64__) && defined(__GNUC__)
1669         return card_data + i * 8 +  (__builtin_ffsll (card) - 1) / 8;
1670 #elif defined(__s390x__) && defined(__GNUC__)
1671         return card_data + i * 8 +  (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
1672 #else
1673         for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
1674                 if (card_data [i])
1675                         return &card_data [i];
1676         }
1677         return card_data;
1678 #endif
1679 }
1680
1681 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
1682 #define MS_BLOCK_OBJ_FAST(b,os,i)                       ((b) + MS_BLOCK_SKIP + (os) * (i))
1683 #define MS_OBJ_ALLOCED_FAST(o,b)                (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
1684
1685 static size_t
1686 card_offset (char *obj, char *base)
1687 {
1688         return (obj - base) >> CARD_BITS;
1689 }
1690
1691 static void
1692 major_scan_card_table (gboolean mod_union, SgenGrayQueue *queue)
1693 {
1694         MSBlockInfo *block;
1695         gboolean has_references;
1696         ScanObjectFunc scan_func = sgen_get_current_object_ops ()->scan_object;
1697
1698         if (!concurrent_mark)
1699                 g_assert (!mod_union);
1700
1701         FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1702 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
1703                 guint8 cards_copy [CARDS_PER_BLOCK];
1704 #endif
1705                 gboolean small_objects;
1706                 int block_obj_size;
1707                 char *block_start;
1708                 guint8 *card_data, *card_base;
1709                 guint8 *card_data_end;
1710                 char *scan_front = NULL;
1711
1712 #ifdef PREFETCH_CARDS
1713                 int prefetch_index = __index + 6;
1714                 if (prefetch_index < allocated_blocks.next_slot) {
1715                         MSBlockInfo *prefetch_block = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [prefetch_index]);
1716                         guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
1717                         PREFETCH_READ (prefetch_block);
1718                         PREFETCH_WRITE (prefetch_cards);
1719                         PREFETCH_WRITE (prefetch_cards + 32);
1720                 }
1721 #endif
1722
1723                 if (!has_references)
1724                         continue;
1725
1726                 block_obj_size = block->obj_size;
1727                 small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
1728
1729                 block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
1730
1731                 /*
1732                  * This is safe in face of card aliasing for the following reason:
1733                  *
1734                  * Major blocks are 16k aligned, or 32 cards aligned.
1735                  * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
1736                  * sizes, they won't overflow the cardtable overlap modulus.
1737                  */
1738                 if (mod_union) {
1739                         card_data = card_base = block->cardtable_mod_union;
1740                         /*
1741                          * This happens when the nursery collection that precedes finishing
1742                          * the concurrent collection allocates new major blocks.
1743                          */
1744                         if (!card_data)
1745                                 continue;
1746                 } else {
1747 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
1748                         card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
1749 #else
1750                         if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
1751                                 continue;
1752                         card_data = card_base = cards_copy;
1753 #endif
1754                 }
1755                 card_data_end = card_data + CARDS_PER_BLOCK;
1756
1757                 card_data += MS_BLOCK_SKIP >> CARD_BITS;
1758
1759                 card_data = initial_skip_card (card_data);
1760                 while (card_data < card_data_end) {
1761                         size_t card_index, first_object_index;
1762                         char *start;
1763                         char *end;
1764                         char *first_obj, *obj;
1765
1766                         HEAVY_STAT (++scanned_cards);
1767
1768                         if (!*card_data) {
1769                                 ++card_data;
1770                                 continue;
1771                         }
1772
1773                         card_index = card_data - card_base;
1774                         start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
1775                         end = start + CARD_SIZE_IN_BYTES;
1776
1777                         if (!block->swept)
1778                                 sweep_block (block, FALSE);
1779
1780                         HEAVY_STAT (++marked_cards);
1781
1782                         if (small_objects)
1783                                 sgen_card_table_prepare_card_for_scanning (card_data);
1784
1785                         /*
1786                          * If the card we're looking at starts at or in the block header, we
1787                          * must start at the first object in the block, without calculating
1788                          * the index of the object we're hypothetically starting at, because
1789                          * it would be negative.
1790                          */
1791                         if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
1792                                 first_object_index = 0;
1793                         else
1794                                 first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
1795
1796                         obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
1797
1798                         while (obj < end) {
1799                                 if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
1800                                         goto next_object;
1801
1802                                 if (mod_union) {
1803                                         /* FIXME: do this more efficiently */
1804                                         int w, b;
1805                                         MS_CALC_MARK_BIT (w, b, obj);
1806                                         if (!MS_MARK_BIT (block, w, b))
1807                                                 goto next_object;
1808                                 }
1809
1810                                 if (small_objects) {
1811                                         HEAVY_STAT (++scanned_objects);
1812                                         scan_func (obj, sgen_obj_get_descriptor (obj), queue);
1813                                 } else {
1814                                         size_t offset = card_offset (obj, block_start);
1815                                         sgen_cardtable_scan_object (obj, block_obj_size, card_base + offset, mod_union, queue);
1816                                 }
1817                         next_object:
1818                                 obj += block_obj_size;
1819                                 g_assert (scan_front <= obj);
1820                                 scan_front = obj;
1821                         }
1822
1823                         HEAVY_STAT (if (*card_data) ++remarked_cards);
1824                         binary_protocol_card_scan (first_obj, obj - first_obj);
1825
1826                         if (small_objects)
1827                                 ++card_data;
1828                         else
1829                                 card_data = card_base + card_offset (obj, block_start);
1830                 }
1831         } END_FOREACH_BLOCK;
1832 }
1833
1834 static void
1835 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
1836 {
1837         MSBlockInfo *block;
1838         gboolean has_references;
1839         long long total_cards = 0;
1840         long long marked_cards = 0;
1841
1842         FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
1843                 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
1844                 int i;
1845
1846                 if (!has_references)
1847                         continue;
1848
1849                 total_cards += CARDS_PER_BLOCK;
1850                 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
1851                         if (cards [i])
1852                                 ++marked_cards;
1853                 }
1854         } END_FOREACH_BLOCK;
1855
1856         *num_total_cards = total_cards;
1857         *num_marked_cards = marked_cards;
1858 }
1859
1860 static void
1861 update_cardtable_mod_union (void)
1862 {
1863         MSBlockInfo *block;
1864
1865         FOREACH_BLOCK (block) {
1866                 size_t num_cards;
1867
1868                 block->cardtable_mod_union = sgen_card_table_update_mod_union (block->cardtable_mod_union,
1869                                 MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
1870
1871                 SGEN_ASSERT (0, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
1872         } END_FOREACH_BLOCK;
1873 }
1874
1875 static guint8*
1876 major_get_cardtable_mod_union_for_object (char *obj)
1877 {
1878         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
1879         size_t offset = card_offset (obj, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1880         return &block->cardtable_mod_union [offset];
1881 }
1882
1883 static void
1884 alloc_free_block_lists (MSBlockInfo ***lists)
1885 {
1886         int i;
1887         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
1888                 lists [i] = sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
1889 }
1890
1891 #undef pthread_create
1892
1893 static void
1894 post_param_init (SgenMajorCollector *collector)
1895 {
1896         collector->sweeps_lazily = lazy_sweep;
1897 }
1898
1899 static void
1900 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
1901 {
1902         int i;
1903
1904         sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
1905
1906         num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
1907         block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
1908         ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
1909
1910         evacuate_block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
1911         for (i = 0; i < num_block_obj_sizes; ++i)
1912                 evacuate_block_obj_sizes [i] = FALSE;
1913
1914         /*
1915         {
1916                 int i;
1917                 g_print ("block object sizes:\n");
1918                 for (i = 0; i < num_block_obj_sizes; ++i)
1919                         g_print ("%d\n", block_obj_sizes [i]);
1920         }
1921         */
1922
1923         alloc_free_block_lists (free_block_lists);
1924
1925         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
1926                 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
1927         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
1928                 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
1929
1930         mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
1931         mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
1932         mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
1933         mono_counters_register ("# major objects evacuated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_objects_evacuated);
1934 #if SIZEOF_VOID_P != 8
1935         mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
1936         mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
1937         mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
1938         mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
1939 #endif
1940
1941         collector->section_size = MAJOR_SECTION_SIZE;
1942
1943         concurrent_mark = is_concurrent;
1944         if (is_concurrent) {
1945                 collector->is_concurrent = TRUE;
1946                 collector->want_synchronous_collection = &want_evacuation;
1947         } else {
1948                 collector->is_concurrent = FALSE;
1949                 collector->want_synchronous_collection = NULL;
1950         }
1951         collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
1952         collector->supports_cardtable = TRUE;
1953
1954         collector->have_swept = &have_swept;
1955
1956         collector->alloc_heap = major_alloc_heap;
1957         collector->is_object_live = major_is_object_live;
1958         collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
1959         collector->alloc_degraded = major_alloc_degraded;
1960
1961         collector->alloc_object = major_alloc_object;
1962         collector->free_pinned_object = free_pinned_object;
1963         collector->iterate_objects = major_iterate_objects;
1964         collector->free_non_pinned_object = major_free_non_pinned_object;
1965         collector->find_pin_queue_start_ends = major_find_pin_queue_start_ends;
1966         collector->pin_objects = major_pin_objects;
1967         collector->pin_major_object = pin_major_object;
1968         collector->scan_card_table = major_scan_card_table;
1969         collector->iterate_live_block_ranges = (void*)(void*) major_iterate_live_block_ranges;
1970         if (is_concurrent) {
1971                 collector->update_cardtable_mod_union = update_cardtable_mod_union;
1972                 collector->get_cardtable_mod_union_for_object = major_get_cardtable_mod_union_for_object;
1973         }
1974         collector->init_to_space = major_init_to_space;
1975         collector->sweep = major_sweep;
1976         collector->check_scan_starts = major_check_scan_starts;
1977         collector->dump_heap = major_dump_heap;
1978         collector->get_used_size = major_get_used_size;
1979         collector->start_nursery_collection = major_start_nursery_collection;
1980         collector->finish_nursery_collection = major_finish_nursery_collection;
1981         collector->start_major_collection = major_start_major_collection;
1982         collector->finish_major_collection = major_finish_major_collection;
1983         collector->have_computed_minor_collection_allowance = major_have_computer_minor_collection_allowance;
1984         collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
1985         collector->obj_is_from_pinned_alloc = obj_is_from_pinned_alloc;
1986         collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
1987         collector->get_num_major_sections = get_num_major_sections;
1988         collector->handle_gc_param = major_handle_gc_param;
1989         collector->print_gc_param_usage = major_print_gc_param_usage;
1990         collector->post_param_init = post_param_init;
1991         collector->is_valid_object = major_is_valid_object;
1992         collector->describe_pointer = major_describe_pointer;
1993         collector->count_cards = major_count_cards;
1994
1995         collector->major_ops.copy_or_mark_object = major_copy_or_mark_object_canonical;
1996         collector->major_ops.scan_object = major_scan_object_with_evacuation;
1997         if (is_concurrent) {
1998                 collector->major_concurrent_ops.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
1999                 collector->major_concurrent_ops.scan_object = major_scan_object_no_mark_concurrent;
2000                 collector->major_concurrent_ops.scan_vtype = major_scan_vtype_concurrent;
2001         }
2002
2003 #if !defined (FIXED_HEAP) && !defined (SGEN_PARALLEL_MARK)
2004         /* FIXME: this will not work with evacuation or the split nursery. */
2005         if (!is_concurrent)
2006                 collector->drain_gray_stack = drain_gray_stack;
2007
2008 #ifdef HEAVY_STATISTICS
2009         mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2010         mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2011         mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2012         mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2013         mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2014         mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2015         mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2016         mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2017         mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2018         mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2019
2020         mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2021         mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2022         mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2023 #endif
2024 #endif
2025
2026 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2027         mono_mutex_init (&scanned_objects_list_lock);
2028 #endif
2029
2030         SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2031
2032         /*cardtable requires major pages to be 8 cards aligned*/
2033         g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2034 }
2035
2036 void
2037 sgen_marksweep_init (SgenMajorCollector *collector)
2038 {
2039         sgen_marksweep_init_internal (collector, FALSE);
2040 }
2041
2042 void
2043 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2044 {
2045         sgen_marksweep_init_internal (collector, TRUE);
2046 }
2047
2048 #endif