mono/sgen/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 #include <string.h>
  31 #include <stdlib.h>
  32
  33 #include "mono/sgen/sgen-gc.h"
  34 #include "mono/sgen/sgen-protocol.h"
  35 #include "mono/sgen/sgen-cardtable.h"
  36 #include "mono/sgen/sgen-memory-governor.h"
  37 #include "mono/sgen/sgen-layout-stats.h"
  38 #include "mono/sgen/sgen-pointer-queue.h"
  39 #include "mono/sgen/sgen-pinning.h"
  40 #include "mono/sgen/sgen-workers.h"
  41 #include "mono/sgen/sgen-thread-pool.h"
  42 #include "mono/sgen/sgen-client.h"
  43 #include "mono/utils/mono-membar.h"
  44
  45 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
  46 #define MS_BLOCK_SIZE   ARCH_MIN_MS_BLOCK_SIZE
  47 #define MS_BLOCK_SIZE_SHIFT     ARCH_MIN_MS_BLOCK_SIZE_SHIFT
  48 #else
  49 #define MS_BLOCK_SIZE_SHIFT     14      /* INT FASTENABLE */
  50 #define MS_BLOCK_SIZE           (1 << MS_BLOCK_SIZE_SHIFT)
  51 #endif
  52 #define MAJOR_SECTION_SIZE      MS_BLOCK_SIZE
  53 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
  54
  55 /*
  56  * Don't allocate single blocks, but alloc a contingent of this many
  57  * blocks in one swoop.  This must be a power of two.
  58  */
  59 #define MS_BLOCK_ALLOC_NUM      32
  60
  61 /*
  62  * Number of bytes before the first object in a block.  At the start
  63  * of a block is the MSBlockHeader, then opional padding, then come
  64  * the objects, so this must be >= sizeof (MSBlockHeader).
  65  */
  66 #define MS_BLOCK_SKIP   ((sizeof (MSBlockHeader) + 15) & ~15)
  67
  68 #define MS_BLOCK_FREE   (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
  69
  70 #define MS_NUM_MARK_WORDS       ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
  71
  72 /*
  73  * Blocks progress from one state to the next:
  74  *
  75  * SWEPT           The block is fully swept.  It might or might not be in
  76  *                 a free list.
  77  *
  78  * MARKING         The block might or might not contain live objects.  If
  79  *                 we're in between an initial collection pause and the
  80  *                 finishing pause, the block might or might not be in a
  81  *                 free list.
  82  *
  83  * CHECKING        The sweep thread is investigating the block to determine
  84  *                 whether or not it contains live objects.  The block is
  85  *                 not in a free list.
  86  *
  87  * NEED_SWEEPING   The block contains live objects but has not yet been
  88  *                 swept.  It also contains free slots.  It is in a block
  89  *                 free list.
  90  *
  91  * SWEEPING        The block is being swept.  It might be in a free list.
  92  */
  93
  94 enum {
  95         BLOCK_STATE_SWEPT,
  96         BLOCK_STATE_MARKING,
  97         BLOCK_STATE_CHECKING,
  98         BLOCK_STATE_NEED_SWEEPING,
  99         BLOCK_STATE_SWEEPING
 100 };
 101
 102 typedef struct _MSBlockInfo MSBlockInfo;
 103 struct _MSBlockInfo {
 104         guint16 obj_size;
 105         /*
 106          * FIXME: Do we even need this? It's only used during sweep and might be worth
 107          * recalculating to save the space.
 108          */
 109         guint16 obj_size_index;
 110         /* FIXME: Reduce this - it only needs a byte. */
 111         volatile gint32 state;
 112         unsigned int pinned : 1;
 113         unsigned int has_references : 1;
 114         unsigned int has_pinned : 1;    /* means cannot evacuate */
 115         unsigned int is_to_space : 1;
 116         void ** volatile free_list;
 117         MSBlockInfo * volatile next_free;
 118         guint8 * volatile cardtable_mod_union;
 119         mword mark_words [MS_NUM_MARK_WORDS];
 120 };
 121
 122 #define MS_BLOCK_FOR_BLOCK_INFO(b)      ((char*)(b))
 123
 124 #define MS_BLOCK_OBJ(b,i)               ((GCObject *)(MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i)))
 125 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size)             (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
 126 #define MS_BLOCK_DATA_FOR_OBJ(o)        ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
 127
 128 typedef struct {
 129         MSBlockInfo info;
 130 } MSBlockHeader;
 131
 132 #define MS_BLOCK_FOR_OBJ(o)             (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
 133
 134 /* object index will always be small */
 135 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
 136
 137 //casting to int is fine since blocks are 32k
 138 #define MS_CALC_MARK_BIT(w,b,o)         do {                            \
 139                 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
 140                 if (sizeof (mword) == 4) {                              \
 141                         (w) = i >> 5;                                   \
 142                         (b) = i & 31;                                   \
 143                 } else {                                                \
 144                         (w) = i >> 6;                                   \
 145                         (b) = i & 63;                                   \
 146                 }                                                       \
 147         } while (0)
 148
 149 #define MS_MARK_BIT(bl,w,b)     ((bl)->mark_words [(w)] & (ONE_P << (b)))
 150 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
 151
 152 #define MS_OBJ_ALLOCED(o,b)     (*(void**)(o) && (*(char**)(o) < MS_BLOCK_FOR_BLOCK_INFO (b) || *(char**)(o) >= MS_BLOCK_FOR_BLOCK_INFO (b) + MS_BLOCK_SIZE))
 153
 154 #define MS_BLOCK_OBJ_SIZE_FACTOR        (pow (2.0, 1.0 / 3))
 155
 156 /*
 157  * This way we can lookup block object size indexes for sizes up to
 158  * 256 bytes with a single load.
 159  */
 160 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES      32
 161
 162 static int *block_obj_sizes;
 163 static int num_block_obj_sizes;
 164 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
 165
 166 #define MS_BLOCK_FLAG_PINNED    1
 167 #define MS_BLOCK_FLAG_REFS      2
 168
 169 #define MS_BLOCK_TYPE_MAX       4
 170
 171 static gboolean *evacuate_block_obj_sizes;
 172 static float evacuation_threshold = 0.666f;
 173
 174 static gboolean lazy_sweep = FALSE;
 175
 176 enum {
 177         SWEEP_STATE_SWEPT,
 178         SWEEP_STATE_NEED_SWEEPING,
 179         SWEEP_STATE_SWEEPING,
 180         SWEEP_STATE_SWEEPING_AND_ITERATING,
 181         SWEEP_STATE_COMPACTING
 182 };
 183
 184 static volatile int sweep_state = SWEEP_STATE_SWEPT;
 185
 186 static gboolean concurrent_mark;
 187 static gboolean concurrent_sweep = TRUE;
 188
 189 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl)      SGEN_POINTER_IS_TAGGED_1 ((bl))
 190 #define BLOCK_TAG_HAS_REFERENCES(bl)            SGEN_POINTER_TAG_1 ((bl))
 191
 192 #define BLOCK_IS_TAGGED_CHECKING(bl)            SGEN_POINTER_IS_TAGGED_2 ((bl))
 193 #define BLOCK_TAG_CHECKING(bl)                  SGEN_POINTER_TAG_2 ((bl))
 194
 195 #define BLOCK_UNTAG(bl)                         ((MSBlockInfo *)SGEN_POINTER_UNTAG_12 ((bl)))
 196
 197 #define BLOCK_TAG(bl)                           ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
 198
 199 /* all allocated blocks in the system */
 200 static SgenPointerQueue allocated_blocks;
 201
 202 /* non-allocated block free-list */
 203 static void *empty_blocks = NULL;
 204 static size_t num_empty_blocks = 0;
 205
 206 #define FOREACH_BLOCK_NO_LOCK_CONDITION(cond,bl) {                      \
 207         size_t __index;                                                 \
 208         SGEN_ASSERT (0, (cond) && !sweep_in_progress (), "Can't iterate blocks while the world is running or sweep is in progress."); \
 209         for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { \
 210                 (bl) = BLOCK_UNTAG (allocated_blocks.data [__index]);
 211 #define FOREACH_BLOCK_NO_LOCK(bl)                                       \
 212         FOREACH_BLOCK_NO_LOCK_CONDITION(sgen_is_world_stopped (), bl)
 213 #define FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK(bl,hr) {                   \
 214         size_t __index;                                                 \
 215         SGEN_ASSERT (0, sgen_is_world_stopped () && !sweep_in_progress (), "Can't iterate blocks while the world is running or sweep is in progress."); \
 216         for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { \
 217                 (bl) = (MSBlockInfo *)allocated_blocks.data [__index];                  \
 218                 (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl));           \
 219                 (bl) = BLOCK_UNTAG ((bl));
 220 #define END_FOREACH_BLOCK_NO_LOCK       } }
 221
 222 static volatile size_t num_major_sections = 0;
 223 /*
 224  * One free block list for each block object size.  We add and remove blocks from these
 225  * lists lock-free via CAS.
 226  *
 227  * Blocks accessed/removed from `free_block_lists`:
 228  *   from the mutator (with GC lock held)
 229  *   in nursery collections
 230  *   in non-concurrent major collections
 231  *   in the finishing pause of concurrent major collections (whole list is cleared)
 232  *
 233  * Blocks added to `free_block_lists`:
 234  *   in the sweeping thread
 235  *   during nursery collections
 236  *   from domain clearing (with the world stopped and no sweeping happening)
 237  *
 238  * The only item of those that doesn't require the GC lock is the sweep thread.  The sweep
 239  * thread only ever adds blocks to the free list, so the ABA problem can't occur.
 240  */
 241 static MSBlockInfo * volatile *free_block_lists [MS_BLOCK_TYPE_MAX];
 242
 243 static guint64 stat_major_blocks_alloced = 0;
 244 static guint64 stat_major_blocks_freed = 0;
 245 static guint64 stat_major_blocks_lazy_swept = 0;
 246
 247 #if SIZEOF_VOID_P != 8
 248 static guint64 stat_major_blocks_freed_ideal = 0;
 249 static guint64 stat_major_blocks_freed_less_ideal = 0;
 250 static guint64 stat_major_blocks_freed_individual = 0;
 251 static guint64 stat_major_blocks_alloced_less_ideal = 0;
 252 #endif
 253
 254 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
 255 static guint64 num_major_objects_marked = 0;
 256 #define INC_NUM_MAJOR_OBJECTS_MARKED()  (++num_major_objects_marked)
 257 #else
 258 #define INC_NUM_MAJOR_OBJECTS_MARKED()
 259 #endif
 260
 261 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
 262 static mono_mutex_t scanned_objects_list_lock;
 263 static SgenPointerQueue scanned_objects_list;
 264
 265 static void
 266 add_scanned_object (void *ptr)
 267 {
 268         if (!binary_protocol_is_enabled ())
 269                 return;
 270
 271         mono_os_mutex_lock (&scanned_objects_list_lock);
 272         sgen_pointer_queue_add (&scanned_objects_list, ptr);
 273         mono_os_mutex_unlock (&scanned_objects_list_lock);
 274 }
 275 #endif
 276
 277 static gboolean sweep_block (MSBlockInfo *block);
 278
 279 static int
 280 ms_find_block_obj_size_index (size_t size)
 281 {
 282         int i;
 283         SGEN_ASSERT (9, size <= SGEN_MAX_SMALL_OBJ_SIZE, "size %zd is bigger than max small object size %d", size, SGEN_MAX_SMALL_OBJ_SIZE);
 284         for (i = 0; i < num_block_obj_sizes; ++i)
 285                 if (block_obj_sizes [i] >= size)
 286                         return i;
 287         g_error ("no object of size %zd\n", size);
 288         return -1;
 289 }
 290
 291 #define FREE_BLOCKS_FROM(lists,p,r)     (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
 292 #define FREE_BLOCKS(p,r)                (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
 293
 294 #define MS_BLOCK_OBJ_SIZE_INDEX(s)                              \
 295         (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ?      \
 296          fast_block_obj_size_indexes [((s)+7)>>3] :             \
 297          ms_find_block_obj_size_index ((s)))
 298
 299 static void*
 300 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
 301 {
 302         char *start;
 303         if (nursery_align)
 304                 start = (char *)sgen_alloc_os_memory_aligned (nursery_size, nursery_align, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery");
 305         else
 306                 start = (char *)sgen_alloc_os_memory (nursery_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery");
 307
 308         return start;
 309 }
 310
 311 static void
 312 update_heap_boundaries_for_block (MSBlockInfo *block)
 313 {
 314         sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
 315 }
 316
 317 /*
 318  * Thread safe
 319  */
 320 static void*
 321 ms_get_empty_block (void)
 322 {
 323         char *p;
 324         int i;
 325         void *block, *empty, *next;
 326
 327  retry:
 328         if (!empty_blocks) {
 329                 /*
 330                  * We try allocating MS_BLOCK_ALLOC_NUM blocks first.  If that's
 331                  * unsuccessful, we halve the number of blocks and try again, until we're at
 332                  * 1.  If that doesn't work, either, we assert.
 333                  */
 334                 int alloc_num = MS_BLOCK_ALLOC_NUM;
 335                 for (;;) {
 336                         p = (char *)sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE,
 337                                 (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE),
 338                                 alloc_num == 1 ? "major heap section" : NULL);
 339                         if (p)
 340                                 break;
 341                         alloc_num >>= 1;
 342                 }
 343
 344                 for (i = 0; i < alloc_num; ++i) {
 345                         block = p;
 346                         /*
 347                          * We do the free list update one after the
 348                          * other so that other threads can use the new
 349                          * blocks as quickly as possible.
 350                          */
 351                         do {
 352                                 empty = empty_blocks;
 353                                 *(void**)block = empty;
 354                         } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
 355                         p += MS_BLOCK_SIZE;
 356                 }
 357
 358                 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
 359
 360                 stat_major_blocks_alloced += alloc_num;
 361 #if SIZEOF_VOID_P != 8
 362                 if (alloc_num != MS_BLOCK_ALLOC_NUM)
 363                         stat_major_blocks_alloced_less_ideal += alloc_num;
 364 #endif
 365         }
 366
 367         do {
 368                 empty = empty_blocks;
 369                 if (!empty)
 370                         goto retry;
 371                 block = empty;
 372                 next = *(void**)block;
 373         } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
 374
 375         SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
 376
 377         *(void**)block = NULL;
 378
 379         g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
 380
 381         return block;
 382 }
 383
 384 /*
 385  * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
 386  * list, where it will either be freed later on, or reused in nursery collections.
 387  */
 388 static void
 389 ms_free_block (void *block)
 390 {
 391         void *empty;
 392
 393         sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
 394         memset (block, 0, MS_BLOCK_SIZE);
 395
 396         do {
 397                 empty = empty_blocks;
 398                 *(void**)block = empty;
 399         } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
 400
 401         SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
 402
 403         binary_protocol_block_free (block, MS_BLOCK_SIZE);
 404 }
 405
 406 static gboolean
 407 sweep_in_progress (void)
 408 {
 409         int state = sweep_state;
 410         return state == SWEEP_STATE_SWEEPING ||
 411                 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
 412                 state == SWEEP_STATE_COMPACTING;
 413 }
 414
 415 static inline gboolean
 416 block_is_swept_or_marking (MSBlockInfo *block)
 417 {
 418         gint32 state = block->state;
 419         return state == BLOCK_STATE_SWEPT || state == BLOCK_STATE_MARKING;
 420 }
 421
 422 //#define MARKSWEEP_CONSISTENCY_CHECK
 423
 424 #ifdef MARKSWEEP_CONSISTENCY_CHECK
 425 static void
 426 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
 427 {
 428         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 429         for (; block; block = block->next_free) {
 430                 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
 431                 g_assert (block->obj_size == size);
 432                 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
 433
 434                 /* blocks in the free lists must have at least
 435                    one free slot */
 436                 g_assert (block->free_list);
 437
 438                 /* the block must be in the allocated_blocks array */
 439                 g_assert (sgen_pointer_queue_find (&allocated_blocks, BLOCK_TAG (block)) != (size_t)-1);
 440         }
 441 }
 442
 443 static void
 444 check_empty_blocks (void)
 445 {
 446         void *p;
 447         size_t i = 0;
 448         for (p = empty_blocks; p; p = *(void**)p)
 449                 ++i;
 450         g_assert (i == num_empty_blocks);
 451 }
 452
 453 static void
 454 consistency_check (void)
 455 {
 456         MSBlockInfo *block;
 457         int i;
 458
 459         /* check all blocks */
 460         FOREACH_BLOCK_NO_LOCK (block) {
 461                 int count = MS_BLOCK_FREE / block->obj_size;
 462                 int num_free = 0;
 463                 void **free;
 464
 465                 /* count number of free slots */
 466                 for (i = 0; i < count; ++i) {
 467                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 468                         if (!MS_OBJ_ALLOCED (obj, block))
 469                                 ++num_free;
 470                 }
 471
 472                 /* check free list */
 473                 for (free = block->free_list; free; free = (void**)*free) {
 474                         g_assert (MS_BLOCK_FOR_OBJ (free) == block);
 475                         --num_free;
 476                 }
 477                 g_assert (num_free == 0);
 478
 479                 /* check all mark words are zero */
 480                 if (!sgen_concurrent_collection_in_progress () && block_is_swept_or_marking (block)) {
 481                         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
 482                                 g_assert (block->mark_words [i] == 0);
 483                 }
 484         } END_FOREACH_BLOCK_NO_LOCK;
 485
 486         /* check free blocks */
 487         for (i = 0; i < num_block_obj_sizes; ++i) {
 488                 int j;
 489                 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
 490                         check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
 491         }
 492
 493         check_empty_blocks ();
 494 }
 495 #endif
 496
 497 static void
 498 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
 499 {
 500         MSBlockInfo *old;
 501         do {
 502                 block->next_free = old = free_blocks [size_index];
 503         } while (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], block, old) != old);
 504 }
 505
 506 static void major_finish_sweep_checking (void);
 507
 508 static gboolean
 509 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
 510 {
 511         int size = block_obj_sizes [size_index];
 512         int count = MS_BLOCK_FREE / size;
 513         MSBlockInfo *info;
 514         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 515         char *obj_start;
 516         int i;
 517
 518         if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
 519                 return FALSE;
 520
 521         info = (MSBlockInfo*)ms_get_empty_block ();
 522
 523         SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
 524
 525         info->obj_size = size;
 526         info->obj_size_index = size_index;
 527         info->pinned = pinned;
 528         info->has_references = has_references;
 529         info->has_pinned = pinned;
 530         /*
 531          * Blocks that are to-space are not evacuated from.  During an major collection
 532          * blocks are allocated for two reasons: evacuating objects from the nursery and
 533          * evacuating them from major blocks marked for evacuation.  In both cases we don't
 534          * want further evacuation. We also don't want to evacuate objects allocated during
 535          * the concurrent mark since it would add pointless stress on the finishing pause.
 536          */
 537         info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD) || sgen_concurrent_collection_in_progress ();
 538         info->state = info->is_to_space ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
 539         SGEN_ASSERT (6, !sweep_in_progress () || info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
 540         info->cardtable_mod_union = NULL;
 541
 542         update_heap_boundaries_for_block (info);
 543
 544         binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
 545
 546         /* build free list */
 547         obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
 548         info->free_list = (void**)obj_start;
 549         /* we're skipping the last one - it must be nulled */
 550         for (i = 0; i < count - 1; ++i) {
 551                 char *next_obj_start = obj_start + size;
 552                 *(void**)obj_start = next_obj_start;
 553                 obj_start = next_obj_start;
 554         }
 555         /* the last one */
 556         *(void**)obj_start = NULL;
 557
 558         add_free_block (free_blocks, size_index, info);
 559
 560         /*
 561          * Adding to the allocated_blocks array is racy with the removal of nulls when
 562          * sweeping. We wait for sweep to finish to avoid that.
 563          *
 564          * The memory barrier here and in `sweep_job_func()` are required because we need
 565          * `allocated_blocks` synchronized between this and the sweep thread.
 566          */
 567         major_finish_sweep_checking ();
 568         mono_memory_barrier ();
 569
 570         sgen_pointer_queue_add (&allocated_blocks, BLOCK_TAG (info));
 571
 572         SGEN_ATOMIC_ADD_P (num_major_sections, 1);
 573         return TRUE;
 574 }
 575
 576 static gboolean
 577 ptr_is_from_pinned_alloc (char *ptr)
 578 {
 579         MSBlockInfo *block;
 580
 581         FOREACH_BLOCK_NO_LOCK (block) {
 582                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE)
 583                         return block->pinned;
 584         } END_FOREACH_BLOCK_NO_LOCK;
 585         return FALSE;
 586 }
 587
 588 static void
 589 ensure_can_access_block_free_list (MSBlockInfo *block)
 590 {
 591  retry:
 592         for (;;) {
 593                 switch (block->state) {
 594                 case BLOCK_STATE_SWEPT:
 595                 case BLOCK_STATE_MARKING:
 596                         return;
 597                 case BLOCK_STATE_CHECKING:
 598                         SGEN_ASSERT (0, FALSE, "How did we get a block that's being checked from a free list?");
 599                         break;
 600                 case BLOCK_STATE_NEED_SWEEPING:
 601                         if (sweep_block (block))
 602                                 ++stat_major_blocks_lazy_swept;
 603                         break;
 604                 case BLOCK_STATE_SWEEPING:
 605                         /* FIXME: do this more elegantly */
 606                         g_usleep (100);
 607                         goto retry;
 608                 default:
 609                         SGEN_ASSERT (0, FALSE, "Illegal block state");
 610                         break;
 611                 }
 612         }
 613 }
 614
 615 static void*
 616 unlink_slot_from_free_list_uncontested (MSBlockInfo * volatile *free_blocks, int size_index)
 617 {
 618         MSBlockInfo *block, *next_free_block;
 619         void *obj, *next_free_slot;
 620
 621  retry:
 622         block = free_blocks [size_index];
 623         SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
 624
 625         ensure_can_access_block_free_list (block);
 626
 627         obj = block->free_list;
 628         SGEN_ASSERT (6, obj, "block %p in free list had no available object to alloc from", block);
 629
 630         next_free_slot = *(void**)obj;
 631         if (next_free_slot) {
 632                 block->free_list = (gpointer *)next_free_slot;
 633                 return obj;
 634         }
 635
 636         next_free_block = block->next_free;
 637         if (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], next_free_block, block) != block)
 638                 goto retry;
 639
 640         block->free_list = NULL;
 641         block->next_free = NULL;
 642
 643         return obj;
 644 }
 645
 646 static GCObject*
 647 alloc_obj (GCVTable vtable, size_t size, gboolean pinned, gboolean has_references)
 648 {
 649         int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 650         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 651         void *obj;
 652
 653         if (!free_blocks [size_index]) {
 654                 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
 655                         return NULL;
 656         }
 657
 658         obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
 659
 660         /* FIXME: assumes object layout */
 661         *(GCVTable*)obj = vtable;
 662
 663         return (GCObject *)obj;
 664 }
 665
 666 static GCObject*
 667 major_alloc_object (GCVTable vtable, size_t size, gboolean has_references)
 668 {
 669         return alloc_obj (vtable, size, FALSE, has_references);
 670 }
 671
 672 /*
 673  * We're not freeing the block if it's empty.  We leave that work for
 674  * the next major collection.
 675  *
 676  * This is just called from the domain clearing code, which runs in a
 677  * single thread and has the GC lock, so we don't need an extra lock.
 678  */
 679 static void
 680 free_object (GCObject *obj, size_t size, gboolean pinned)
 681 {
 682         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 683         int word, bit;
 684         gboolean in_free_list;
 685
 686         SGEN_ASSERT (9, sweep_state == SWEEP_STATE_SWEPT, "Should have waited for sweep to free objects.");
 687
 688         ensure_can_access_block_free_list (block);
 689         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);
 690         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
 691         MS_CALC_MARK_BIT (word, bit, obj);
 692         SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set", obj);
 693
 694         memset (obj, 0, size);
 695
 696         in_free_list = !!block->free_list;
 697         *(void**)obj = block->free_list;
 698         block->free_list = (void**)obj;
 699
 700         if (!in_free_list) {
 701                 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (pinned, block->has_references);
 702                 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 703                 SGEN_ASSERT (9, !block->next_free, "block %p doesn't have a free-list of object but belongs to a free-list of blocks", block);
 704                 add_free_block (free_blocks, size_index, block);
 705         }
 706 }
 707
 708 static void
 709 major_free_non_pinned_object (GCObject *obj, size_t size)
 710 {
 711         free_object (obj, size, FALSE);
 712 }
 713
 714 /* size is a multiple of SGEN_ALLOC_ALIGN */
 715 static GCObject*
 716 major_alloc_small_pinned_obj (GCVTable vtable, size_t size, gboolean has_references)
 717 {
 718         void *res;
 719
 720         res = alloc_obj (vtable, size, TRUE, has_references);
 721          /*If we failed to alloc memory, we better try releasing memory
 722           *as pinned alloc is requested by the runtime.
 723           */
 724          if (!res) {
 725                 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE);
 726                 res = alloc_obj (vtable, size, TRUE, has_references);
 727          }
 728          return (GCObject *)res;
 729 }
 730
 731 static void
 732 free_pinned_object (GCObject *obj, size_t size)
 733 {
 734         free_object (obj, size, TRUE);
 735 }
 736
 737 /*
 738  * size is already rounded up and we hold the GC lock.
 739  */
 740 static GCObject*
 741 major_alloc_degraded (GCVTable vtable, size_t size)
 742 {
 743         GCObject *obj;
 744
 745         obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
 746         if (G_LIKELY (obj)) {
 747                 HEAVY_STAT (++stat_objects_alloced_degraded);
 748                 HEAVY_STAT (stat_bytes_alloced_degraded += size);
 749         }
 750         return obj;
 751 }
 752
 753 /*
 754  * obj is some object.  If it's not in the major heap (i.e. if it's in
 755  * the nursery or LOS), return FALSE.  Otherwise return whether it's
 756  * been marked or copied.
 757  */
 758 static gboolean
 759 major_is_object_live (GCObject *obj)
 760 {
 761         MSBlockInfo *block;
 762         int word, bit;
 763         mword objsize;
 764
 765         if (sgen_ptr_in_nursery (obj))
 766                 return FALSE;
 767
 768         objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
 769
 770         /* LOS */
 771         if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
 772                 return FALSE;
 773
 774         /* now we know it's in a major block */
 775         block = MS_BLOCK_FOR_OBJ (obj);
 776         SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?", block);
 777         MS_CALC_MARK_BIT (word, bit, obj);
 778         return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
 779 }
 780
 781 static gboolean
 782 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
 783 {
 784         MSBlockInfo *block;
 785
 786         FOREACH_BLOCK_NO_LOCK (block) {
 787                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
 788                         int count = MS_BLOCK_FREE / block->obj_size;
 789                         int i;
 790
 791                         *start = NULL;
 792                         for (i = 0; i <= count; ++i) {
 793                                 if (ptr >= (char*)MS_BLOCK_OBJ (block, i) && ptr < (char*)MS_BLOCK_OBJ (block, i + 1)) {
 794                                         *start = (char *)MS_BLOCK_OBJ (block, i);
 795                                         break;
 796                                 }
 797                         }
 798                         return !block->pinned;
 799                 }
 800         } END_FOREACH_BLOCK_NO_LOCK;
 801         return FALSE;
 802 }
 803
 804 static gboolean
 805 try_set_sweep_state (int new_, int expected)
 806 {
 807         int old = SGEN_CAS (&sweep_state, new_, expected);
 808         return old == expected;
 809 }
 810
 811 static void
 812 set_sweep_state (int new_, int expected)
 813 {
 814         gboolean success = try_set_sweep_state (new_, expected);
 815         SGEN_ASSERT (0, success, "Could not set sweep state.");
 816 }
 817
 818 static gboolean ensure_block_is_checked_for_sweeping (int block_index, gboolean wait, gboolean *have_checked);
 819
 820 static SgenThreadPoolJob * volatile sweep_job;
 821
 822 static void
 823 major_finish_sweep_checking (void)
 824 {
 825         int block_index;
 826         SgenThreadPoolJob *job;
 827
 828  retry:
 829         switch (sweep_state) {
 830         case SWEEP_STATE_SWEPT:
 831         case SWEEP_STATE_NEED_SWEEPING:
 832                 return;
 833         case SWEEP_STATE_SWEEPING:
 834                 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
 835                         break;
 836                 goto retry;
 837         case SWEEP_STATE_SWEEPING_AND_ITERATING:
 838                 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
 839                 goto retry;
 840         case SWEEP_STATE_COMPACTING:
 841                 goto wait;
 842         default:
 843                 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
 844                 break;
 845         }
 846
 847         /*
 848          * We're running with the world stopped and the only other thread doing work is the
 849          * sweep thread, which doesn't add blocks to the array, so we can safely access
 850          * `next_slot`.
 851          */
 852         for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index)
 853                 ensure_block_is_checked_for_sweeping (block_index, FALSE, NULL);
 854
 855         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
 856
 857  wait:
 858         job = sweep_job;
 859         if (job)
 860                 sgen_thread_pool_job_wait (job);
 861         SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
 862         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
 863 }
 864
 865 static void
 866 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
 867 {
 868         gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
 869         gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
 870         gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
 871         MSBlockInfo *block;
 872
 873         major_finish_sweep_checking ();
 874         FOREACH_BLOCK_NO_LOCK (block) {
 875                 int count = MS_BLOCK_FREE / block->obj_size;
 876                 int i;
 877
 878                 if (block->pinned && !pinned)
 879                         continue;
 880                 if (!block->pinned && !non_pinned)
 881                         continue;
 882                 if (sweep && lazy_sweep) {
 883                         sweep_block (block);
 884                         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
 885                 }
 886
 887                 for (i = 0; i < count; ++i) {
 888                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 889                         /*
 890                          * We've finished sweep checking, but if we're sweeping lazily and
 891                          * the flags don't require us to sweep, the block might still need
 892                          * sweeping.  In that case, we need to consult the mark bits to tell
 893                          * us whether an object slot is live.
 894                          */
 895                         if (!block_is_swept_or_marking (block)) {
 896                                 int word, bit;
 897                                 SGEN_ASSERT (6, !sweep && block->state == BLOCK_STATE_NEED_SWEEPING, "Has sweeping not finished?");
 898                                 MS_CALC_MARK_BIT (word, bit, obj);
 899                                 if (!MS_MARK_BIT (block, word, bit))
 900                                         continue;
 901                         }
 902                         if (MS_OBJ_ALLOCED (obj, block))
 903                                 callback ((GCObject*)obj, block->obj_size, data);
 904                 }
 905         } END_FOREACH_BLOCK_NO_LOCK;
 906 }
 907
 908 static gboolean
 909 major_is_valid_object (char *object)
 910 {
 911         MSBlockInfo *block;
 912
 913         FOREACH_BLOCK_NO_LOCK (block) {
 914                 int idx;
 915                 char *obj;
 916
 917                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
 918                         continue;
 919
 920                 idx = MS_BLOCK_OBJ_INDEX (object, block);
 921                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 922                 if (obj != object)
 923                         return FALSE;
 924                 return MS_OBJ_ALLOCED (obj, block);
 925         } END_FOREACH_BLOCK_NO_LOCK;
 926
 927         return FALSE;
 928 }
 929
 930
 931 static GCVTable
 932 major_describe_pointer (char *ptr)
 933 {
 934         MSBlockInfo *block;
 935
 936         FOREACH_BLOCK_NO_LOCK (block) {
 937                 int idx;
 938                 char *obj;
 939                 gboolean live;
 940                 GCVTable vtable;
 941                 int w, b;
 942                 gboolean marked;
 943
 944                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
 945                         continue;
 946
 947                 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
 948                         MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
 949
 950                 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
 951                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 952                 live = MS_OBJ_ALLOCED (obj, block);
 953                 vtable = live ? SGEN_LOAD_VTABLE ((GCObject*)obj) : NULL;
 954
 955                 MS_CALC_MARK_BIT (w, b, obj);
 956                 marked = MS_MARK_BIT (block, w, b);
 957
 958                 if (obj == ptr) {
 959                         SGEN_LOG (0, "\t(");
 960                         if (live)
 961                                 SGEN_LOG (0, "object");
 962                         else
 963                                 SGEN_LOG (0, "dead-object");
 964                 } else {
 965                         if (live)
 966                                 SGEN_LOG (0, "interior-ptr offset %zd", ptr - obj);
 967                         else
 968                                 SGEN_LOG (0, "dead-interior-ptr offset %zd", ptr - obj);
 969                 }
 970
 971                 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
 972
 973                 return vtable;
 974         } END_FOREACH_BLOCK_NO_LOCK;
 975
 976         return NULL;
 977 }
 978
 979 static void
 980 major_check_scan_starts (void)
 981 {
 982 }
 983
 984 static void
 985 major_dump_heap (FILE *heap_dump_file)
 986 {
 987         MSBlockInfo *block;
 988         int *slots_available = (int *)alloca (sizeof (int) * num_block_obj_sizes);
 989         int *slots_used = (int *)alloca (sizeof (int) * num_block_obj_sizes);
 990         int i;
 991
 992         for (i = 0; i < num_block_obj_sizes; ++i)
 993                 slots_available [i] = slots_used [i] = 0;
 994
 995         FOREACH_BLOCK_NO_LOCK (block) {
 996                 int index = ms_find_block_obj_size_index (block->obj_size);
 997                 int count = MS_BLOCK_FREE / block->obj_size;
 998
 999                 slots_available [index] += count;
1000                 for (i = 0; i < count; ++i) {
1001                         if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
1002                                 ++slots_used [index];
1003                 }
1004         } END_FOREACH_BLOCK_NO_LOCK;
1005
1006         fprintf (heap_dump_file, "<occupancies>\n");
1007         for (i = 0; i < num_block_obj_sizes; ++i) {
1008                 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
1009                                 block_obj_sizes [i], slots_available [i], slots_used [i]);
1010         }
1011         fprintf (heap_dump_file, "</occupancies>\n");
1012
1013         FOREACH_BLOCK_NO_LOCK (block) {
1014                 int count = MS_BLOCK_FREE / block->obj_size;
1015                 int i;
1016                 int start = -1;
1017
1018                 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
1019
1020                 for (i = 0; i <= count; ++i) {
1021                         if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
1022                                 if (start < 0)
1023                                         start = i;
1024                         } else {
1025                                 if (start >= 0) {
1026                                         sgen_dump_occupied ((char *)MS_BLOCK_OBJ (block, start), (char *)MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
1027                                         start = -1;
1028                                 }
1029                         }
1030                 }
1031
1032                 fprintf (heap_dump_file, "</section>\n");
1033         } END_FOREACH_BLOCK_NO_LOCK;
1034 }
1035
1036 static guint8*
1037 get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate)
1038 {
1039         guint8 *mod_union = block->cardtable_mod_union;
1040         guint8 *other;
1041         if (mod_union)
1042                 return mod_union;
1043         else if (!allocate)
1044                 return NULL;
1045         mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1046         other = (guint8 *)SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
1047         if (!other) {
1048                 SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
1049                 return mod_union;
1050         }
1051         sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1052         return other;
1053 }
1054
1055 static inline guint8*
1056 major_get_cardtable_mod_union_for_reference (char *ptr)
1057 {
1058         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (ptr);
1059         size_t offset = sgen_card_table_get_card_offset (ptr, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1060         guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
1061         SGEN_ASSERT (0, mod_union, "FIXME: optionally allocate the mod union if it's not here and CAS it in.");
1062         return &mod_union [offset];
1063 }
1064
1065 /*
1066  * Mark the mod-union card for `ptr`, which must be a reference within the object `obj`.
1067  */
1068 static void
1069 mark_mod_union_card (GCObject *obj, void **ptr, GCObject *value_obj)
1070 {
1071         int type = sgen_obj_get_descriptor (obj) & DESC_TYPE_MASK;
1072         if (sgen_safe_object_is_small (obj, type)) {
1073                 guint8 *card_byte = major_get_cardtable_mod_union_for_reference ((char*)ptr);
1074                 SGEN_ASSERT (0, MS_BLOCK_FOR_OBJ (obj) == MS_BLOCK_FOR_OBJ (ptr), "How can an object and a reference inside it not be in the same block?");
1075                 *card_byte = 1;
1076         } else {
1077                 sgen_los_mark_mod_union_card (obj, ptr);
1078         }
1079
1080         binary_protocol_mod_union_remset (obj, ptr, value_obj, SGEN_LOAD_VTABLE (value_obj));
1081 }
1082
1083 static inline gboolean
1084 major_block_is_evacuating (MSBlockInfo *block)
1085 {
1086         if (evacuate_block_obj_sizes [block->obj_size_index] &&
1087                         !block->has_pinned &&
1088                         !block->is_to_space)
1089                 return TRUE;
1090         return FALSE;
1091 }
1092
1093 #define LOAD_VTABLE     SGEN_LOAD_VTABLE
1094
1095 #define MS_MARK_OBJECT_AND_ENQUEUE_CHECKED(obj,desc,block,queue) do {   \
1096                 int __word, __bit;                                      \
1097                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
1098                 if (!MS_MARK_BIT ((block), __word, __bit) && MS_OBJ_ALLOCED ((obj), (block))) { \
1099                         MS_SET_MARK_BIT ((block), __word, __bit);       \
1100                         if (sgen_gc_descr_has_references (desc))                        \
1101                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1102                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1103                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
1104                 }                                                       \
1105         } while (0)
1106 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do {           \
1107                 int __word, __bit;                                      \
1108                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
1109                 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
1110                 if (!MS_MARK_BIT ((block), __word, __bit)) {            \
1111                         MS_SET_MARK_BIT ((block), __word, __bit);       \
1112                         if (sgen_gc_descr_has_references (desc))                        \
1113                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1114                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1115                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
1116                 }                                                       \
1117         } while (0)
1118
1119 static void
1120 pin_major_object (GCObject *obj, SgenGrayQueue *queue)
1121 {
1122         MSBlockInfo *block;
1123
1124         if (concurrent_mark)
1125                 g_assert_not_reached ();
1126
1127         block = MS_BLOCK_FOR_OBJ (obj);
1128         block->has_pinned = TRUE;
1129         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1130 }
1131
1132 #include "sgen-major-copy-object.h"
1133
1134 static long long
1135 major_get_and_reset_num_major_objects_marked (void)
1136 {
1137 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1138         long long num = num_major_objects_marked;
1139         num_major_objects_marked = 0;
1140         return num;
1141 #else
1142         return 0;
1143 #endif
1144 }
1145
1146 #define PREFETCH_CARDS          1       /* BOOL FASTENABLE */
1147 #if !PREFETCH_CARDS
1148 #undef PREFETCH_CARDS
1149 #endif
1150
1151 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1152 #if defined(PLATFORM_MACOSX)
1153 #define GCC_VERSION (__GNUC__ * 10000 \
1154                                + __GNUC_MINOR__ * 100 \
1155                                + __GNUC_PATCHLEVEL__)
1156 #if GCC_VERSION <= 40300
1157 #undef PREFETCH_CARDS
1158 #endif
1159 #endif
1160
1161 #ifdef HEAVY_STATISTICS
1162 static guint64 stat_optimized_copy;
1163 static guint64 stat_optimized_copy_nursery;
1164 static guint64 stat_optimized_copy_nursery_forwarded;
1165 static guint64 stat_optimized_copy_nursery_pinned;
1166 static guint64 stat_optimized_copy_major;
1167 static guint64 stat_optimized_copy_major_small_fast;
1168 static guint64 stat_optimized_copy_major_small_slow;
1169 static guint64 stat_optimized_copy_major_large;
1170 static guint64 stat_optimized_copy_major_forwarded;
1171 static guint64 stat_optimized_copy_major_small_evacuate;
1172 static guint64 stat_optimized_major_scan;
1173 static guint64 stat_optimized_major_scan_no_refs;
1174
1175 static guint64 stat_drain_prefetch_fills;
1176 static guint64 stat_drain_prefetch_fill_failures;
1177 static guint64 stat_drain_loops;
1178 #endif
1179
1180 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_no_evacuation
1181 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_no_evacuation
1182 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_no_evacuation
1183 #include "sgen-marksweep-drain-gray-stack.h"
1184
1185 #define COPY_OR_MARK_WITH_EVACUATION
1186 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_with_evacuation
1187 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_with_evacuation
1188 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_with_evacuation
1189 #include "sgen-marksweep-drain-gray-stack.h"
1190
1191 #undef COPY_OR_MARK_WITH_EVACUATION
1192 #define COPY_OR_MARK_CONCURRENT
1193 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_concurrent_no_evacuation
1194 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_concurrent_no_evacuation
1195 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_concurrent_no_evacuation
1196 #include "sgen-marksweep-drain-gray-stack.h"
1197
1198 #undef COPY_OR_MARK_CONCURRENT
1199 #define COPY_OR_MARK_CONCURRENT_WITH_EVACUATION
1200 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_concurrent_with_evacuation
1201 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_concurrent_with_evacuation
1202 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_concurrent_with_evacuation
1203 #include "sgen-marksweep-drain-gray-stack.h"
1204
1205 static inline gboolean
1206 major_is_evacuating (void)
1207 {
1208         int i;
1209         for (i = 0; i < num_block_obj_sizes; ++i) {
1210                 if (evacuate_block_obj_sizes [i]) {
1211                         return TRUE;
1212                 }
1213         }
1214
1215         return FALSE;
1216 }
1217
1218 static gboolean
1219 drain_gray_stack (SgenGrayQueue *queue)
1220 {
1221         if (major_is_evacuating ())
1222                 return drain_gray_stack_with_evacuation (queue);
1223         else
1224                 return drain_gray_stack_no_evacuation (queue);
1225 }
1226
1227 static gboolean
1228 drain_gray_stack_concurrent (SgenGrayQueue *queue)
1229 {
1230         if (major_is_evacuating ())
1231                 return drain_gray_stack_concurrent_with_evacuation (queue);
1232         else
1233                 return drain_gray_stack_concurrent_no_evacuation (queue);
1234 }
1235
1236 #include "sgen-marksweep-scan-object-concurrent.h"
1237
1238 static void
1239 major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue)
1240 {
1241         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1242 }
1243
1244 static void
1245 major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *queue)
1246 {
1247         major_copy_or_mark_object_concurrent_with_evacuation (ptr, *ptr, queue);
1248 }
1249
1250 static void
1251 major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue)
1252 {
1253         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1254 }
1255
1256 static void
1257 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1258 {
1259         void **entry, **end;
1260         int last_index = -1;
1261
1262         if (first_entry == last_entry)
1263                 return;
1264
1265         block->has_pinned = TRUE;
1266
1267         entry = sgen_pinning_get_entry (first_entry);
1268         end = sgen_pinning_get_entry (last_entry);
1269
1270         for (; entry < end; ++entry) {
1271                 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1272                 GCObject *obj;
1273                 SGEN_ASSERT (9, index >= 0 && index < MS_BLOCK_FREE / block->obj_size, "invalid object %p index %d max-index %d", *entry, index, (int)(MS_BLOCK_FREE / block->obj_size));
1274                 if (index == last_index)
1275                         continue;
1276                 obj = MS_BLOCK_OBJ (block, index);
1277                 MS_MARK_OBJECT_AND_ENQUEUE_CHECKED (obj, sgen_obj_get_descriptor (obj), block, queue);
1278                 last_index = index;
1279         }
1280 }
1281
1282 static inline void
1283 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1284 {
1285         int obj_index;
1286
1287         for (obj_index = 0; obj_index < count; ++obj_index) {
1288                 int word, bit;
1289                 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1290
1291                 MS_CALC_MARK_BIT (word, bit, obj);
1292                 if (MS_MARK_BIT (block, word, bit)) {
1293                         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1294                 } else {
1295                         /* an unmarked object */
1296                         if (MS_OBJ_ALLOCED (obj, block)) {
1297                                 /*
1298                                  * FIXME: Merge consecutive
1299                                  * slots for lower reporting
1300                                  * overhead.  Maybe memset
1301                                  * will also benefit?
1302                                  */
1303                                 binary_protocol_empty (obj, obj_size);
1304                                 memset (obj, 0, obj_size);
1305                         }
1306                         *(void**)obj = block->free_list;
1307                         block->free_list = (void **)obj;
1308                 }
1309         }
1310 }
1311
1312 static inline gboolean
1313 try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1314 {
1315         gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
1316         gboolean success = old_state == expected_state;
1317         if (success)
1318                 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
1319         return success;
1320 }
1321
1322 static inline void
1323 set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1324 {
1325         SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set");
1326         block->state = new_state;
1327 }
1328
1329 /*
1330  * If `block` needs sweeping, sweep it and return TRUE.  Otherwise return FALSE.
1331  *
1332  * Sweeping means iterating through the block's slots and building the free-list from the
1333  * unmarked ones.  They will also be zeroed.  The mark bits will be reset.
1334  */
1335 static gboolean
1336 sweep_block (MSBlockInfo *block)
1337 {
1338         int count;
1339         void *reversed = NULL;
1340
1341  retry:
1342         switch (block->state) {
1343         case BLOCK_STATE_SWEPT:
1344                 return FALSE;
1345         case BLOCK_STATE_MARKING:
1346         case BLOCK_STATE_CHECKING:
1347                 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1348                 goto retry;
1349         case BLOCK_STATE_SWEEPING:
1350                 /* FIXME: Do this more elegantly */
1351                 g_usleep (100);
1352                 goto retry;
1353         case BLOCK_STATE_NEED_SWEEPING:
1354                 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1355                         goto retry;
1356                 break;
1357         default:
1358                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1359         }
1360
1361         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1362
1363         count = MS_BLOCK_FREE / block->obj_size;
1364
1365         block->free_list = NULL;
1366
1367         /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1368         // FIXME: Add more sizes
1369         switch (block->obj_size) {
1370         case 16:
1371                 sweep_block_for_size (block, count, 16);
1372                 break;
1373         default:
1374                 sweep_block_for_size (block, count, block->obj_size);
1375                 break;
1376         }
1377
1378         /* reset mark bits */
1379         memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1380
1381         /* Reverse free list so that it's in address order */
1382         reversed = NULL;
1383         while (block->free_list) {
1384                 void *next = *(void**)block->free_list;
1385                 *(void**)block->free_list = reversed;
1386                 reversed = block->free_list;
1387                 block->free_list = (void **)next;
1388         }
1389         block->free_list = (void **)reversed;
1390
1391         mono_memory_write_barrier ();
1392
1393         set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1394
1395         return TRUE;
1396 }
1397
1398 static inline int
1399 bitcount (mword d)
1400 {
1401         int count = 0;
1402
1403 #ifdef __GNUC__
1404         if (sizeof (mword) == 8)
1405                 count += __builtin_popcountll (d);
1406         else
1407                 count += __builtin_popcount (d);
1408 #else
1409         while (d) {
1410                 count ++;
1411                 d &= (d - 1);
1412         }
1413 #endif
1414         return count;
1415 }
1416
1417 /* statistics for evacuation */
1418 static size_t *sweep_slots_available;
1419 static size_t *sweep_slots_used;
1420 static size_t *sweep_num_blocks;
1421
1422 static volatile size_t num_major_sections_before_sweep;
1423 static volatile size_t num_major_sections_freed_in_sweep;
1424
1425 static void
1426 sweep_start (void)
1427 {
1428         int i;
1429
1430         for (i = 0; i < num_block_obj_sizes; ++i)
1431                 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1432
1433         /* clear all the free lists */
1434         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1435                 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1436                 int j;
1437                 for (j = 0; j < num_block_obj_sizes; ++j)
1438                         free_blocks [j] = NULL;
1439         }
1440 }
1441
1442 static void sweep_finish (void);
1443
1444 /*
1445  * If `wait` is TRUE and the block is currently being checked, this function will wait until
1446  * the checking has finished.
1447  *
1448  * Returns whether the block is still there.  If `wait` is FALSE, the return value will not
1449  * be correct, i.e. must not be used.
1450  */
1451 static gboolean
1452 ensure_block_is_checked_for_sweeping (int block_index, gboolean wait, gboolean *have_checked)
1453 {
1454         int count;
1455         gboolean have_live = FALSE;
1456         gboolean have_free = FALSE;
1457         int nused = 0;
1458         int block_state;
1459         int i;
1460         void *tagged_block;
1461         MSBlockInfo *block;
1462
1463         SGEN_ASSERT (6, sweep_in_progress (), "Why do we call this function if there's no sweep in progress?");
1464
1465         if (have_checked)
1466                 *have_checked = FALSE;
1467
1468  retry:
1469         tagged_block = *(void * volatile *)&allocated_blocks.data [block_index];
1470         if (!tagged_block)
1471                 return FALSE;
1472
1473         if (BLOCK_IS_TAGGED_CHECKING (tagged_block)) {
1474                 if (!wait)
1475                         return FALSE;
1476                 /* FIXME: do this more elegantly */
1477                 g_usleep (100);
1478                 goto retry;
1479         }
1480
1481         if (SGEN_CAS_PTR (&allocated_blocks.data [block_index], BLOCK_TAG_CHECKING (tagged_block), tagged_block) != tagged_block)
1482                 goto retry;
1483
1484         block = BLOCK_UNTAG (tagged_block);
1485         block_state = block->state;
1486
1487         if (!sweep_in_progress ()) {
1488                 SGEN_ASSERT (6, block_state != BLOCK_STATE_SWEEPING && block_state != BLOCK_STATE_CHECKING, "Invalid block state.");
1489                 if (!lazy_sweep)
1490                         SGEN_ASSERT (6, block_state != BLOCK_STATE_NEED_SWEEPING, "Invalid block state.");
1491         }
1492
1493         switch (block_state) {
1494         case BLOCK_STATE_SWEPT:
1495         case BLOCK_STATE_NEED_SWEEPING:
1496         case BLOCK_STATE_SWEEPING:
1497                 goto done;
1498         case BLOCK_STATE_MARKING:
1499                 break;
1500         case BLOCK_STATE_CHECKING:
1501                 SGEN_ASSERT (0, FALSE, "We set the CHECKING bit - how can the stage be CHECKING?");
1502                 goto done;
1503         default:
1504                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1505                 break;
1506         }
1507
1508         SGEN_ASSERT (6, block->state == BLOCK_STATE_MARKING, "When we sweep all blocks must start out marking.");
1509         set_block_state (block, BLOCK_STATE_CHECKING, BLOCK_STATE_MARKING);
1510
1511         if (have_checked)
1512                 *have_checked = TRUE;
1513
1514         block->has_pinned = block->pinned;
1515
1516         block->is_to_space = FALSE;
1517
1518         count = MS_BLOCK_FREE / block->obj_size;
1519
1520         if (block->cardtable_mod_union) {
1521                 sgen_card_table_free_mod_union (block->cardtable_mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1522                 block->cardtable_mod_union = NULL;
1523         }
1524
1525         /* Count marked objects in the block */
1526         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1527                 nused += bitcount (block->mark_words [i]);
1528
1529         if (nused)
1530                 have_live = TRUE;
1531         if (nused < count)
1532                 have_free = TRUE;
1533
1534         if (have_live) {
1535                 int obj_size_index = block->obj_size_index;
1536                 gboolean has_pinned = block->has_pinned;
1537
1538                 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1539
1540                 /*
1541                  * FIXME: Go straight to SWEPT if there are no free slots.  We need
1542                  * to set the free slot list to NULL, though, and maybe update some
1543                  * statistics.
1544                  */
1545                 if (!lazy_sweep)
1546                         sweep_block (block);
1547
1548                 if (!has_pinned) {
1549                         ++sweep_num_blocks [obj_size_index];
1550                         sweep_slots_used [obj_size_index] += nused;
1551                         sweep_slots_available [obj_size_index] += count;
1552                 }
1553
1554                 /*
1555                  * If there are free slots in the block, add
1556                  * the block to the corresponding free list.
1557                  */
1558                 if (have_free) {
1559                         MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1560
1561                         if (!lazy_sweep)
1562                                 SGEN_ASSERT (6, block->free_list, "How do we not have a free list when there are free slots?");
1563
1564                         add_free_block (free_blocks, obj_size_index, block);
1565                 }
1566
1567                 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1568                 update_heap_boundaries_for_block (block);
1569         } else {
1570                 /*
1571                  * Blocks without live objects are removed from the
1572                  * block list and freed.
1573                  */
1574                 SGEN_ASSERT (6, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1575                 SGEN_ASSERT (6, allocated_blocks.data [block_index] == BLOCK_TAG_CHECKING (tagged_block), "How did the block move?");
1576
1577                 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1578                 ms_free_block (block);
1579
1580                 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1581
1582                 tagged_block = NULL;
1583         }
1584
1585  done:
1586         allocated_blocks.data [block_index] = tagged_block;
1587         return !!tagged_block;
1588 }
1589
1590 static void
1591 sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1592 {
1593         int block_index;
1594         int num_blocks = num_major_sections_before_sweep;
1595
1596         SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1597         SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1598
1599         /*
1600          * We traverse the block array from high to low.  Nursery collections will have to
1601          * cooperate with the sweep thread to finish sweeping, and they will traverse from
1602          * low to high, to avoid constantly colliding on the same blocks.
1603          */
1604         for (block_index = num_blocks - 1; block_index >= 0; --block_index) {
1605                 gboolean have_checked;
1606
1607                 /*
1608                  * The block might have been freed by another thread doing some checking
1609                  * work.
1610                  */
1611                 if (!ensure_block_is_checked_for_sweeping (block_index, TRUE, &have_checked))
1612                         ++num_major_sections_freed_in_sweep;
1613         }
1614
1615         while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) {
1616                 /*
1617                  * The main GC thread is currently iterating over the block array to help us
1618                  * finish the sweep.  We have already finished, but we don't want to mess up
1619                  * that iteration, so we just wait for it.
1620                  */
1621                 g_usleep (100);
1622         }
1623
1624         if (SGEN_MAX_ASSERT_LEVEL >= 6) {
1625                 for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1626                         MSBlockInfo *block = BLOCK_UNTAG (allocated_blocks.data [block_index]);
1627                         SGEN_ASSERT (6, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1628                 }
1629         }
1630
1631         sgen_pointer_queue_remove_nulls (&allocated_blocks);
1632         mono_memory_barrier ();
1633
1634         sweep_finish ();
1635
1636         sweep_job = NULL;
1637 }
1638
1639 static void
1640 sweep_finish (void)
1641 {
1642         int i;
1643
1644         for (i = 0; i < num_block_obj_sizes; ++i) {
1645                 float usage = (float)sweep_slots_used [i] / (float)sweep_slots_available [i];
1646                 if (sweep_num_blocks [i] > 5 && usage < evacuation_threshold) {
1647                         evacuate_block_obj_sizes [i] = TRUE;
1648                         /*
1649                         g_print ("slot size %d - %d of %d used\n",
1650                                         block_obj_sizes [i], slots_used [i], slots_available [i]);
1651                         */
1652                 } else {
1653                         evacuate_block_obj_sizes [i] = FALSE;
1654                 }
1655         }
1656
1657         set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1658 }
1659
1660 static void
1661 major_sweep (void)
1662 {
1663         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1664
1665         sweep_start ();
1666
1667         SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?");
1668
1669         num_major_sections_before_sweep = num_major_sections;
1670         num_major_sections_freed_in_sweep = 0;
1671
1672         SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
1673         if (concurrent_sweep) {
1674                 sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
1675                 sgen_thread_pool_job_enqueue (sweep_job);
1676         } else {
1677                 sweep_job_func (NULL, NULL);
1678         }
1679 }
1680
1681 static gboolean
1682 major_have_swept (void)
1683 {
1684         return sweep_state == SWEEP_STATE_SWEPT;
1685 }
1686
1687 static int count_pinned_ref;
1688 static int count_pinned_nonref;
1689 static int count_nonpinned_ref;
1690 static int count_nonpinned_nonref;
1691
1692 static void
1693 count_nonpinned_callback (GCObject *obj, size_t size, void *data)
1694 {
1695         GCVTable vtable = LOAD_VTABLE (obj);
1696
1697         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1698                 ++count_nonpinned_ref;
1699         else
1700                 ++count_nonpinned_nonref;
1701 }
1702
1703 static void
1704 count_pinned_callback (GCObject *obj, size_t size, void *data)
1705 {
1706         GCVTable vtable = LOAD_VTABLE (obj);
1707
1708         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1709                 ++count_pinned_ref;
1710         else
1711                 ++count_pinned_nonref;
1712 }
1713
1714 static G_GNUC_UNUSED void
1715 count_ref_nonref_objs (void)
1716 {
1717         int total;
1718
1719         count_pinned_ref = 0;
1720         count_pinned_nonref = 0;
1721         count_nonpinned_ref = 0;
1722         count_nonpinned_nonref = 0;
1723
1724         major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1725         major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1726
1727         total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1728
1729         g_print ("ref: %d pinned %d non-pinned   non-ref: %d pinned %d non-pinned  --  %.1f\n",
1730                         count_pinned_ref, count_nonpinned_ref,
1731                         count_pinned_nonref, count_nonpinned_nonref,
1732                         (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1733 }
1734
1735 static int
1736 ms_calculate_block_obj_sizes (double factor, int *arr)
1737 {
1738         double target_size;
1739         int num_sizes = 0;
1740         int last_size = 0;
1741
1742         /*
1743          * Have every possible slot size starting with the minimal
1744          * object size up to and including four times that size.  Then
1745          * proceed by increasing geometrically with the given factor.
1746          */
1747
1748         for (int size = SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size <= 4 * SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size += SGEN_ALLOC_ALIGN) {
1749                 if (arr)
1750                         arr [num_sizes] = size;
1751                 ++num_sizes;
1752                 last_size = size;
1753         }
1754         target_size = (double)last_size;
1755
1756         do {
1757                 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1758                 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1759
1760                 if (size != last_size) {
1761                         if (arr)
1762                                 arr [num_sizes] = size;
1763                         ++num_sizes;
1764                         last_size = size;
1765                 }
1766
1767                 target_size *= factor;
1768         } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1769
1770         return num_sizes;
1771 }
1772
1773 /* only valid during minor collections */
1774 static mword old_num_major_sections;
1775
1776 static void
1777 major_start_nursery_collection (void)
1778 {
1779 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1780         consistency_check ();
1781 #endif
1782
1783         old_num_major_sections = num_major_sections;
1784 }
1785
1786 static void
1787 major_finish_nursery_collection (void)
1788 {
1789 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1790         consistency_check ();
1791 #endif
1792 }
1793
1794 static void
1795 major_start_major_collection (void)
1796 {
1797         MSBlockInfo *block;
1798         int i;
1799
1800         major_finish_sweep_checking ();
1801
1802         /*
1803          * Clear the free lists for block sizes where we do evacuation.  For those block
1804          * sizes we will have to allocate new blocks.
1805          */
1806         for (i = 0; i < num_block_obj_sizes; ++i) {
1807                 if (!evacuate_block_obj_sizes [i])
1808                         continue;
1809
1810                 binary_protocol_evacuating_blocks (block_obj_sizes [i]);
1811
1812                 free_block_lists [0][i] = NULL;
1813                 free_block_lists [MS_BLOCK_FLAG_REFS][i] = NULL;
1814         }
1815
1816         if (lazy_sweep)
1817                 binary_protocol_sweep_begin (GENERATION_OLD, TRUE);
1818
1819         /* Sweep all unswept blocks and set them to MARKING */
1820         FOREACH_BLOCK_NO_LOCK (block) {
1821                 if (lazy_sweep)
1822                         sweep_block (block);
1823                 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1824                 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1825         } END_FOREACH_BLOCK_NO_LOCK;
1826
1827         if (lazy_sweep)
1828                 binary_protocol_sweep_end (GENERATION_OLD, TRUE);
1829
1830         set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1831 }
1832
1833 static void
1834 major_finish_major_collection (ScannedObjectCounts *counts)
1835 {
1836 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1837         if (binary_protocol_is_enabled ()) {
1838                 counts->num_scanned_objects = scanned_objects_list.next_slot;
1839
1840                 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1841                 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1842
1843                 sgen_pointer_queue_clear (&scanned_objects_list);
1844         }
1845 #endif
1846 }
1847
1848 #if SIZEOF_VOID_P != 8
1849 static int
1850 compare_pointers (const void *va, const void *vb) {
1851         char *a = *(char**)va, *b = *(char**)vb;
1852         if (a < b)
1853                 return -1;
1854         if (a > b)
1855                 return 1;
1856         return 0;
1857 }
1858 #endif
1859
1860 /*
1861  * This is called with sweep completed and the world stopped.
1862  */
1863 static void
1864 major_free_swept_blocks (size_t allowance)
1865 {
1866         /* FIXME: This is probably too much.  It's assuming all objects are small. */
1867         size_t section_reserve = allowance / MS_BLOCK_SIZE;
1868
1869         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1870
1871 #ifdef TARGET_WIN32
1872                 /*
1873                  * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1874                  * a VirtualAlloc ()-ed block.
1875                  */
1876                 return;
1877 #endif
1878
1879 #if SIZEOF_VOID_P != 8
1880         {
1881                 int i, num_empty_blocks_orig, num_blocks, arr_length;
1882                 void *block;
1883                 void **empty_block_arr;
1884                 void **rebuild_next;
1885
1886                 if (num_empty_blocks <= section_reserve)
1887                         return;
1888                 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1889
1890                 num_empty_blocks_orig = num_empty_blocks;
1891                 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1892                                 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1893                 if (!empty_block_arr)
1894                         goto fallback;
1895
1896                 i = 0;
1897                 for (block = empty_blocks; block; block = *(void**)block)
1898                         empty_block_arr [i++] = block;
1899                 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1900
1901                 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1902
1903                 /*
1904                  * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1905                  * contiguous ones.  If we do, we free them.  If that's not enough to get to
1906                  * section_reserve, we halve the number of contiguous blocks we're looking
1907                  * for and have another go, until we're done with looking for pairs of
1908                  * blocks, at which point we give up and go to the fallback.
1909                  */
1910                 arr_length = num_empty_blocks_orig;
1911                 num_blocks = MS_BLOCK_ALLOC_NUM;
1912                 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1913                         int first = -1;
1914                         int dest = 0;
1915
1916                         dest = 0;
1917                         for (i = 0; i < arr_length; ++i) {
1918                                 int d = dest;
1919                                 void *block = empty_block_arr [i];
1920                                 SGEN_ASSERT (6, block, "we're not shifting correctly");
1921                                 if (i != dest) {
1922                                         empty_block_arr [dest] = block;
1923                                         /*
1924                                          * This is not strictly necessary, but we're
1925                                          * cautious.
1926                                          */
1927                                         empty_block_arr [i] = NULL;
1928                                 }
1929                                 ++dest;
1930
1931                                 if (first < 0) {
1932                                         first = d;
1933                                         continue;
1934                                 }
1935
1936                                 SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
1937
1938                                 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
1939                                         first = d;
1940                                         continue;
1941                                 }
1942
1943                                 if (d + 1 - first == num_blocks) {
1944                                         /*
1945                                          * We found num_blocks contiguous blocks.  Free them
1946                                          * and null their array entries.  As an optimization
1947                                          * we could, instead of nulling the entries, shift
1948                                          * the following entries over to the left, while
1949                                          * we're iterating.
1950                                          */
1951                                         int j;
1952                                         sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
1953                                         for (j = first; j <= d; ++j)
1954                                                 empty_block_arr [j] = NULL;
1955                                         dest = first;
1956                                         first = -1;
1957
1958                                         num_empty_blocks -= num_blocks;
1959
1960                                         stat_major_blocks_freed += num_blocks;
1961                                         if (num_blocks == MS_BLOCK_ALLOC_NUM)
1962                                                 stat_major_blocks_freed_ideal += num_blocks;
1963                                         else
1964                                                 stat_major_blocks_freed_less_ideal += num_blocks;
1965
1966                                 }
1967                         }
1968
1969                         SGEN_ASSERT (6, dest <= i && dest <= arr_length, "array length is off");
1970                         arr_length = dest;
1971                         SGEN_ASSERT (6, arr_length == num_empty_blocks, "array length is off");
1972
1973                         num_blocks >>= 1;
1974                 }
1975
1976                 /* rebuild empty_blocks free list */
1977                 rebuild_next = (void**)&empty_blocks;
1978                 for (i = 0; i < arr_length; ++i) {
1979                         void *block = empty_block_arr [i];
1980                         SGEN_ASSERT (6, block, "we're missing blocks");
1981                         *rebuild_next = block;
1982                         rebuild_next = (void**)block;
1983                 }
1984                 *rebuild_next = NULL;
1985
1986                 /* free array */
1987                 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
1988         }
1989
1990         SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
1991
1992  fallback:
1993         /*
1994          * This is our threshold.  If there's not more empty than used blocks, we won't
1995          * release uncontiguous blocks, in fear of fragmenting the address space.
1996          */
1997         if (num_empty_blocks <= num_major_sections)
1998                 return;
1999 #endif
2000
2001         while (num_empty_blocks > section_reserve) {
2002                 void *next = *(void**)empty_blocks;
2003                 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
2004                 empty_blocks = next;
2005                 /*
2006                  * Needs not be atomic because this is running
2007                  * single-threaded.
2008                  */
2009                 --num_empty_blocks;
2010
2011                 ++stat_major_blocks_freed;
2012 #if SIZEOF_VOID_P != 8
2013                 ++stat_major_blocks_freed_individual;
2014 #endif
2015         }
2016 }
2017
2018 static void
2019 major_pin_objects (SgenGrayQueue *queue)
2020 {
2021         MSBlockInfo *block;
2022
2023         FOREACH_BLOCK_NO_LOCK (block) {
2024                 size_t first_entry, last_entry;
2025                 SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
2026                 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
2027                                 &first_entry, &last_entry);
2028                 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
2029         } END_FOREACH_BLOCK_NO_LOCK;
2030 }
2031
2032 static void
2033 major_init_to_space (void)
2034 {
2035 }
2036
2037 static void
2038 major_report_pinned_memory_usage (void)
2039 {
2040         g_assert_not_reached ();
2041 }
2042
2043 static gint64
2044 major_get_used_size (void)
2045 {
2046         gint64 size = 0;
2047         MSBlockInfo *block;
2048
2049         /*
2050          * We're holding the GC lock, but the sweep thread might be running.  Make sure it's
2051          * finished, then we can iterate over the block array.
2052          */
2053         major_finish_sweep_checking ();
2054
2055         FOREACH_BLOCK_NO_LOCK_CONDITION (TRUE, block) {
2056                 int count = MS_BLOCK_FREE / block->obj_size;
2057                 void **iter;
2058                 size += count * block->obj_size;
2059                 for (iter = block->free_list; iter; iter = (void**)*iter)
2060                         size -= block->obj_size;
2061         } END_FOREACH_BLOCK_NO_LOCK;
2062
2063         return size;
2064 }
2065
2066 /* FIXME: return number of bytes, not of sections */
2067 static size_t
2068 get_num_major_sections (void)
2069 {
2070         return num_major_sections;
2071 }
2072
2073 /*
2074  * Returns the number of bytes in blocks that were present when the last sweep was
2075  * initiated, and were not freed during the sweep.  They are the basis for calculating the
2076  * allowance.
2077  */
2078 static size_t
2079 get_bytes_survived_last_sweep (void)
2080 {
2081         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2082         return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
2083 }
2084
2085 static gboolean
2086 major_handle_gc_param (const char *opt)
2087 {
2088         if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2089                 const char *arg = strchr (opt, '=') + 1;
2090                 int percentage = atoi (arg);
2091                 if (percentage < 0 || percentage > 100) {
2092                         fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2093                         exit (1);
2094                 }
2095                 evacuation_threshold = (float)percentage / 100.0f;
2096                 return TRUE;
2097         } else if (!strcmp (opt, "lazy-sweep")) {
2098                 lazy_sweep = TRUE;
2099                 return TRUE;
2100         } else if (!strcmp (opt, "no-lazy-sweep")) {
2101                 lazy_sweep = FALSE;
2102                 return TRUE;
2103         } else if (!strcmp (opt, "concurrent-sweep")) {
2104                 concurrent_sweep = TRUE;
2105                 return TRUE;
2106         } else if (!strcmp (opt, "no-concurrent-sweep")) {
2107                 concurrent_sweep = FALSE;
2108                 return TRUE;
2109         }
2110
2111         return FALSE;
2112 }
2113
2114 static void
2115 major_print_gc_param_usage (void)
2116 {
2117         fprintf (stderr,
2118                         ""
2119                         "  evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2120                         "  (no-)lazy-sweep\n"
2121                         "  (no-)concurrent-sweep\n"
2122                         );
2123 }
2124
2125 /*
2126  * This callback is used to clear cards, move cards to the shadow table and do counting.
2127  */
2128 static void
2129 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
2130 {
2131         MSBlockInfo *block;
2132         gboolean has_references;
2133
2134         major_finish_sweep_checking ();
2135         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2136                 if (has_references)
2137                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2138         } END_FOREACH_BLOCK_NO_LOCK;
2139 }
2140
2141 #ifdef HEAVY_STATISTICS
2142 extern guint64 marked_cards;
2143 extern guint64 scanned_cards;
2144 extern guint64 scanned_objects;
2145 extern guint64 remarked_cards;
2146 #endif
2147
2148 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2149 /*
2150  * MS blocks are 16K aligned.
2151  * Cardtables are 4K aligned, at least.
2152  * This means that the cardtable of a given block is 32 bytes aligned.
2153  */
2154 static guint8*
2155 initial_skip_card (guint8 *card_data)
2156 {
2157         mword *cards = (mword*)card_data;
2158         mword card;
2159         int i;
2160         for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2161                 card = cards [i];
2162                 if (card)
2163                         break;
2164         }
2165
2166         if (i == CARD_WORDS_PER_BLOCK)
2167                 return card_data + CARDS_PER_BLOCK;
2168
2169 #if defined(__i386__) && defined(__GNUC__)
2170         return card_data + i * 4 +  (__builtin_ffs (card) - 1) / 8;
2171 #elif defined(__x86_64__) && defined(__GNUC__)
2172         return card_data + i * 8 +  (__builtin_ffsll (card) - 1) / 8;
2173 #elif defined(__s390x__) && defined(__GNUC__)
2174         return card_data + i * 8 +  (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2175 #else
2176         for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2177                 if (card_data [i])
2178                         return &card_data [i];
2179         }
2180         return card_data;
2181 #endif
2182 }
2183
2184 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2185 #define MS_BLOCK_OBJ_FAST(b,os,i)                       ((b) + MS_BLOCK_SKIP + (os) * (i))
2186 #define MS_OBJ_ALLOCED_FAST(o,b)                (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2187
2188 static void
2189 scan_card_table_for_block (MSBlockInfo *block, gboolean mod_union, ScanCopyContext ctx)
2190 {
2191         SgenGrayQueue *queue = ctx.queue;
2192         ScanObjectFunc scan_func = ctx.ops->scan_object;
2193 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2194         guint8 cards_copy [CARDS_PER_BLOCK];
2195 #endif
2196         gboolean small_objects;
2197         int block_obj_size;
2198         char *block_start;
2199         guint8 *card_data, *card_base;
2200         guint8 *card_data_end;
2201         char *scan_front = NULL;
2202
2203         block_obj_size = block->obj_size;
2204         small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2205
2206         block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2207
2208         /*
2209          * This is safe in face of card aliasing for the following reason:
2210          *
2211          * Major blocks are 16k aligned, or 32 cards aligned.
2212          * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2213          * sizes, they won't overflow the cardtable overlap modulus.
2214          */
2215         if (mod_union) {
2216                 card_data = card_base = block->cardtable_mod_union;
2217                 /*
2218                  * This happens when the nursery collection that precedes finishing
2219                  * the concurrent collection allocates new major blocks.
2220                  */
2221                 if (!card_data)
2222                         return;
2223         } else {
2224 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2225                 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2226 #else
2227                 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2228                         return;
2229                 card_data = card_base = cards_copy;
2230 #endif
2231         }
2232         card_data_end = card_data + CARDS_PER_BLOCK;
2233
2234         card_data += MS_BLOCK_SKIP >> CARD_BITS;
2235
2236         card_data = initial_skip_card (card_data);
2237         while (card_data < card_data_end) {
2238                 size_t card_index, first_object_index;
2239                 char *start;
2240                 char *end;
2241                 char *first_obj, *obj;
2242
2243                 HEAVY_STAT (++scanned_cards);
2244
2245                 if (!*card_data) {
2246                         ++card_data;
2247                         continue;
2248                 }
2249
2250                 card_index = card_data - card_base;
2251                 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2252                 end = start + CARD_SIZE_IN_BYTES;
2253
2254                 if (!block_is_swept_or_marking (block))
2255                         sweep_block (block);
2256
2257                 HEAVY_STAT (++marked_cards);
2258
2259                 if (small_objects)
2260                         sgen_card_table_prepare_card_for_scanning (card_data);
2261
2262                 /*
2263                  * If the card we're looking at starts at or in the block header, we
2264                  * must start at the first object in the block, without calculating
2265                  * the index of the object we're hypothetically starting at, because
2266                  * it would be negative.
2267                  */
2268                 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2269                         first_object_index = 0;
2270                 else
2271                         first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2272
2273                 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2274
2275                 binary_protocol_card_scan (first_obj, end - first_obj);
2276
2277                 while (obj < end) {
2278                         if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2279                                 goto next_object;
2280
2281                         if (mod_union) {
2282                                 /* FIXME: do this more efficiently */
2283                                 int w, b;
2284                                 MS_CALC_MARK_BIT (w, b, obj);
2285                                 if (!MS_MARK_BIT (block, w, b))
2286                                         goto next_object;
2287                         }
2288
2289                         GCObject *object = (GCObject*)obj;
2290
2291                         if (small_objects) {
2292                                 HEAVY_STAT (++scanned_objects);
2293                                 scan_func (object, sgen_obj_get_descriptor (object), queue);
2294                         } else {
2295                                 size_t offset = sgen_card_table_get_card_offset (obj, block_start);
2296                                 sgen_cardtable_scan_object (object, block_obj_size, card_base + offset, mod_union, ctx);
2297                         }
2298                 next_object:
2299                         obj += block_obj_size;
2300                         g_assert (scan_front <= obj);
2301                         scan_front = obj;
2302                 }
2303
2304                 HEAVY_STAT (if (*card_data) ++remarked_cards);
2305
2306                 if (small_objects)
2307                         ++card_data;
2308                 else
2309                         card_data = card_base + sgen_card_table_get_card_offset (obj, block_start);
2310         }
2311 }
2312
2313 static void
2314 major_scan_card_table (gboolean mod_union, ScanCopyContext ctx)
2315 {
2316         MSBlockInfo *block;
2317         gboolean has_references;
2318
2319         if (!concurrent_mark)
2320                 g_assert (!mod_union);
2321
2322         major_finish_sweep_checking ();
2323         binary_protocol_major_card_table_scan_start (sgen_timestamp (), mod_union);
2324         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2325 #ifdef PREFETCH_CARDS
2326                 int prefetch_index = __index + 6;
2327                 if (prefetch_index < allocated_blocks.next_slot) {
2328                         MSBlockInfo *prefetch_block = BLOCK_UNTAG (allocated_blocks.data [prefetch_index]);
2329                         guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2330                         PREFETCH_READ (prefetch_block);
2331                         PREFETCH_WRITE (prefetch_cards);
2332                         PREFETCH_WRITE (prefetch_cards + 32);
2333                 }
2334 #endif
2335
2336                 if (!has_references)
2337                         continue;
2338
2339                 scan_card_table_for_block (block, mod_union, ctx);
2340         } END_FOREACH_BLOCK_NO_LOCK;
2341         binary_protocol_major_card_table_scan_end (sgen_timestamp (), mod_union);
2342 }
2343
2344 static void
2345 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2346 {
2347         MSBlockInfo *block;
2348         gboolean has_references;
2349         long long total_cards = 0;
2350         long long marked_cards = 0;
2351
2352         if (sweep_in_progress ()) {
2353                 *num_total_cards = -1;
2354                 *num_marked_cards = -1;
2355                 return;
2356         }
2357
2358         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2359                 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2360                 int i;
2361
2362                 if (!has_references)
2363                         continue;
2364
2365                 total_cards += CARDS_PER_BLOCK;
2366                 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2367                         if (cards [i])
2368                                 ++marked_cards;
2369                 }
2370         } END_FOREACH_BLOCK_NO_LOCK;
2371
2372         *num_total_cards = total_cards;
2373         *num_marked_cards = marked_cards;
2374 }
2375
2376 static void
2377 update_cardtable_mod_union (void)
2378 {
2379         MSBlockInfo *block;
2380
2381         FOREACH_BLOCK_NO_LOCK (block) {
2382                 size_t num_cards;
2383                 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
2384                 sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2385                 SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2386         } END_FOREACH_BLOCK_NO_LOCK;
2387 }
2388
2389 #undef pthread_create
2390
2391 static void
2392 post_param_init (SgenMajorCollector *collector)
2393 {
2394         collector->sweeps_lazily = lazy_sweep;
2395         collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
2396 }
2397
2398 static void
2399 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
2400 {
2401         int i;
2402
2403         sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2404
2405         num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2406         block_obj_sizes = (int *)sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2407         ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2408
2409         evacuate_block_obj_sizes = (gboolean *)sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2410         for (i = 0; i < num_block_obj_sizes; ++i)
2411                 evacuate_block_obj_sizes [i] = FALSE;
2412
2413         sweep_slots_available = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2414         sweep_slots_used = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2415         sweep_num_blocks = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2416
2417         /*
2418         {
2419                 int i;
2420                 g_print ("block object sizes:\n");
2421                 for (i = 0; i < num_block_obj_sizes; ++i)
2422                         g_print ("%d\n", block_obj_sizes [i]);
2423         }
2424         */
2425
2426         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2427                 free_block_lists [i] = (MSBlockInfo *volatile *)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2428
2429         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2430                 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2431         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2432                 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2433
2434         mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2435         mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2436         mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2437 #if SIZEOF_VOID_P != 8
2438         mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2439         mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2440         mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2441         mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2442 #endif
2443
2444         collector->section_size = MAJOR_SECTION_SIZE;
2445
2446         concurrent_mark = is_concurrent;
2447         collector->is_concurrent = is_concurrent;
2448         collector->needs_thread_pool = is_concurrent || concurrent_sweep;
2449         collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2450         collector->supports_cardtable = TRUE;
2451
2452         collector->alloc_heap = major_alloc_heap;
2453         collector->is_object_live = major_is_object_live;
2454         collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2455         collector->alloc_degraded = major_alloc_degraded;
2456
2457         collector->alloc_object = major_alloc_object;
2458         collector->free_pinned_object = free_pinned_object;
2459         collector->iterate_objects = major_iterate_objects;
2460         collector->free_non_pinned_object = major_free_non_pinned_object;
2461         collector->pin_objects = major_pin_objects;
2462         collector->pin_major_object = pin_major_object;
2463         collector->scan_card_table = major_scan_card_table;
2464         collector->iterate_live_block_ranges = major_iterate_live_block_ranges;
2465         if (is_concurrent) {
2466                 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2467                 collector->get_cardtable_mod_union_for_reference = major_get_cardtable_mod_union_for_reference;
2468         }
2469         collector->init_to_space = major_init_to_space;
2470         collector->sweep = major_sweep;
2471         collector->have_swept = major_have_swept;
2472         collector->finish_sweeping = major_finish_sweep_checking;
2473         collector->free_swept_blocks = major_free_swept_blocks;
2474         collector->check_scan_starts = major_check_scan_starts;
2475         collector->dump_heap = major_dump_heap;
2476         collector->get_used_size = major_get_used_size;
2477         collector->start_nursery_collection = major_start_nursery_collection;
2478         collector->finish_nursery_collection = major_finish_nursery_collection;
2479         collector->start_major_collection = major_start_major_collection;
2480         collector->finish_major_collection = major_finish_major_collection;
2481         collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2482         collector->ptr_is_from_pinned_alloc = ptr_is_from_pinned_alloc;
2483         collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2484         collector->get_num_major_sections = get_num_major_sections;
2485         collector->get_bytes_survived_last_sweep = get_bytes_survived_last_sweep;
2486         collector->handle_gc_param = major_handle_gc_param;
2487         collector->print_gc_param_usage = major_print_gc_param_usage;
2488         collector->post_param_init = post_param_init;
2489         collector->is_valid_object = major_is_valid_object;
2490         collector->describe_pointer = major_describe_pointer;
2491         collector->count_cards = major_count_cards;
2492
2493         collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
2494         collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
2495         collector->major_ops_serial.drain_gray_stack = drain_gray_stack;
2496         if (is_concurrent) {
2497                 collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2498                 collector->major_ops_concurrent_start.scan_object = major_scan_object_concurrent_with_evacuation;
2499                 collector->major_ops_concurrent_start.drain_gray_stack = drain_gray_stack_concurrent;
2500
2501                 collector->major_ops_concurrent_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_finish_canonical;
2502                 collector->major_ops_concurrent_finish.scan_object = major_scan_object_with_evacuation;
2503                 collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_concurrent_finish;
2504                 collector->major_ops_concurrent_finish.drain_gray_stack = drain_gray_stack;
2505         }
2506
2507 #ifdef HEAVY_STATISTICS
2508         mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2509         mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2510         mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2511         mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2512         mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2513         mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2514         mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2515         mono_counters_register ("Optimized copy major small evacuate", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_evacuate);
2516         mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2517         mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2518         mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2519
2520         mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2521         mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2522         mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2523 #endif
2524
2525 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2526         mono_os_mutex_init (&scanned_objects_list_lock);
2527 #endif
2528
2529         SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2530
2531         /*cardtable requires major pages to be 8 cards aligned*/
2532         g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2533 }
2534
2535 void
2536 sgen_marksweep_init (SgenMajorCollector *collector)
2537 {
2538         sgen_marksweep_init_internal (collector, FALSE);
2539 }
2540
2541 void
2542 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2543 {
2544         sgen_marksweep_init_internal (collector, TRUE);
2545 }
2546
2547 #endif