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-array-list.h"
  40 #include "mono/sgen/sgen-pinning.h"
  41 #include "mono/sgen/sgen-workers.h"
  42 #include "mono/sgen/sgen-thread-pool.h"
  43 #include "mono/sgen/sgen-client.h"
  44 #include "mono/utils/mono-membar.h"
  45
  46 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
  47 #define MS_BLOCK_SIZE   ARCH_MIN_MS_BLOCK_SIZE
  48 #define MS_BLOCK_SIZE_SHIFT     ARCH_MIN_MS_BLOCK_SIZE_SHIFT
  49 #else
  50 #define MS_BLOCK_SIZE_SHIFT     14      /* INT FASTENABLE */
  51 #define MS_BLOCK_SIZE           (1 << MS_BLOCK_SIZE_SHIFT)
  52 #endif
  53 #define MAJOR_SECTION_SIZE      MS_BLOCK_SIZE
  54 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
  55
  56 /*
  57  * Don't allocate single blocks, but alloc a contingent of this many
  58  * blocks in one swoop.  This must be a power of two.
  59  */
  60 #define MS_BLOCK_ALLOC_NUM      32
  61
  62 /*
  63  * Number of bytes before the first object in a block.  At the start
  64  * of a block is the MSBlockHeader, then opional padding, then come
  65  * the objects, so this must be >= sizeof (MSBlockHeader).
  66  */
  67 #define MS_BLOCK_SKIP   ((sizeof (MSBlockHeader) + 15) & ~15)
  68
  69 #define MS_BLOCK_FREE   (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
  70
  71 #define MS_NUM_MARK_WORDS       ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
  72
  73 /*
  74  * Blocks progress from one state to the next:
  75  *
  76  * SWEPT           The block is fully swept.  It might or might not be in
  77  *                 a free list.
  78  *
  79  * MARKING         The block might or might not contain live objects.  If
  80  *                 we're in between an initial collection pause and the
  81  *                 finishing pause, the block might or might not be in a
  82  *                 free list.
  83  *
  84  * CHECKING        The sweep thread is investigating the block to determine
  85  *                 whether or not it contains live objects.  The block is
  86  *                 not in a free list.
  87  *
  88  * NEED_SWEEPING   The block contains live objects but has not yet been
  89  *                 swept.  It also contains free slots.  It is in a block
  90  *                 free list.
  91  *
  92  * SWEEPING        The block is being swept.  It might be in a free list.
  93  */
  94
  95 enum {
  96         BLOCK_STATE_SWEPT,
  97         BLOCK_STATE_MARKING,
  98         BLOCK_STATE_CHECKING,
  99         BLOCK_STATE_NEED_SWEEPING,
 100         BLOCK_STATE_SWEEPING
 101 };
 102
 103 typedef struct _MSBlockInfo MSBlockInfo;
 104 struct _MSBlockInfo {
 105         guint16 obj_size;
 106         /*
 107          * FIXME: Do we even need this? It's only used during sweep and might be worth
 108          * recalculating to save the space.
 109          */
 110         guint16 obj_size_index;
 111         /* FIXME: Reduce this - it only needs a byte. */
 112         volatile gint32 state;
 113         unsigned int pinned : 1;
 114         unsigned int has_references : 1;
 115         unsigned int has_pinned : 1;    /* means cannot evacuate */
 116         unsigned int is_to_space : 1;
 117         void ** volatile free_list;
 118         MSBlockInfo * volatile next_free;
 119         guint8 * volatile cardtable_mod_union;
 120         mword mark_words [MS_NUM_MARK_WORDS];
 121 };
 122
 123 #define MS_BLOCK_FOR_BLOCK_INFO(b)      ((char*)(b))
 124
 125 #define MS_BLOCK_OBJ(b,i)               ((GCObject *)(MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i)))
 126 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size)             (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
 127 #define MS_BLOCK_DATA_FOR_OBJ(o)        ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
 128
 129 typedef struct {
 130         MSBlockInfo info;
 131 } MSBlockHeader;
 132
 133 #define MS_BLOCK_FOR_OBJ(o)             (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
 134
 135 /* object index will always be small */
 136 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
 137
 138 //casting to int is fine since blocks are 32k
 139 #define MS_CALC_MARK_BIT(w,b,o)         do {                            \
 140                 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
 141                 if (sizeof (mword) == 4) {                              \
 142                         (w) = i >> 5;                                   \
 143                         (b) = i & 31;                                   \
 144                 } else {                                                \
 145                         (w) = i >> 6;                                   \
 146                         (b) = i & 63;                                   \
 147                 }                                                       \
 148         } while (0)
 149
 150 #define MS_MARK_BIT(bl,w,b)     ((bl)->mark_words [(w)] & (ONE_P << (b)))
 151 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
 152
 153 #define MS_OBJ_ALLOCED(o,b)     (*(void**)(o) && (*(char**)(o) < MS_BLOCK_FOR_BLOCK_INFO (b) || *(char**)(o) >= MS_BLOCK_FOR_BLOCK_INFO (b) + MS_BLOCK_SIZE))
 154
 155 #define MS_BLOCK_OBJ_SIZE_FACTOR        (pow (2.0, 1.0 / 3))
 156
 157 /*
 158  * This way we can lookup block object size indexes for sizes up to
 159  * 256 bytes with a single load.
 160  */
 161 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES      32
 162
 163 static int *block_obj_sizes;
 164 static int num_block_obj_sizes;
 165 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
 166
 167 #define MS_BLOCK_FLAG_PINNED    1
 168 #define MS_BLOCK_FLAG_REFS      2
 169
 170 #define MS_BLOCK_TYPE_MAX       4
 171
 172 static gboolean *evacuate_block_obj_sizes;
 173 static float evacuation_threshold = 0.666f;
 174
 175 static gboolean lazy_sweep = FALSE;
 176
 177 enum {
 178         SWEEP_STATE_SWEPT,
 179         SWEEP_STATE_NEED_SWEEPING,
 180         SWEEP_STATE_SWEEPING,
 181         SWEEP_STATE_SWEEPING_AND_ITERATING,
 182         SWEEP_STATE_COMPACTING
 183 };
 184
 185 static volatile int sweep_state = SWEEP_STATE_SWEPT;
 186
 187 static gboolean concurrent_mark;
 188 static gboolean concurrent_sweep = TRUE;
 189
 190 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl)      SGEN_POINTER_IS_TAGGED_1 ((bl))
 191 #define BLOCK_TAG_HAS_REFERENCES(bl)            SGEN_POINTER_TAG_1 ((bl))
 192
 193 #define BLOCK_IS_TAGGED_CHECKING(bl)            SGEN_POINTER_IS_TAGGED_2 ((bl))
 194 #define BLOCK_TAG_CHECKING(bl)                  SGEN_POINTER_TAG_2 ((bl))
 195
 196 #define BLOCK_UNTAG(bl)                         ((MSBlockInfo *)SGEN_POINTER_UNTAG_12 ((bl)))
 197
 198 #define BLOCK_TAG(bl)                           ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
 199
 200 /* all allocated blocks in the system */
 201 static SgenArrayList allocated_blocks = SGEN_ARRAY_LIST_INIT (NULL, NULL, NULL, INTERNAL_MEM_PIN_QUEUE);
 202
 203 /* non-allocated block free-list */
 204 static void *empty_blocks = NULL;
 205 static size_t num_empty_blocks = 0;
 206
 207 #define FOREACH_BLOCK_NO_LOCK(bl) {                                     \
 208         volatile gpointer *slot;                                                \
 209         SGEN_ASSERT (0, !sweep_in_progress (), "Can't iterate blocks while sweep is in progress."); \
 210         SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) {        \
 211                 (bl) = BLOCK_UNTAG (*slot);
 212 #define FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK(bl,hr) {                   \
 213         volatile gpointer *slot;                                                \
 214         SGEN_ASSERT (0, !sweep_in_progress (), "Can't iterate blocks while sweep is in progress."); \
 215         SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) {        \
 216                 (bl) = (MSBlockInfo *) (*slot);                 \
 217                 (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl));           \
 218                 (bl) = BLOCK_UNTAG ((bl));
 219 #define END_FOREACH_BLOCK_NO_LOCK       } SGEN_ARRAY_LIST_END_FOREACH_SLOT; }
 220
 221 static volatile size_t num_major_sections = 0;
 222 /*
 223  * One free block list for each block object size.  We add and remove blocks from these
 224  * lists lock-free via CAS.
 225  *
 226  * Blocks accessed/removed from `free_block_lists`:
 227  *   from the mutator (with GC lock held)
 228  *   in nursery collections
 229  *   in non-concurrent major collections
 230  *   in the finishing pause of concurrent major collections (whole list is cleared)
 231  *
 232  * Blocks added to `free_block_lists`:
 233  *   in the sweeping thread
 234  *   during nursery collections
 235  *   from domain clearing (with the world stopped and no sweeping happening)
 236  *
 237  * The only item of those that doesn't require the GC lock is the sweep thread.  The sweep
 238  * thread only ever adds blocks to the free list, so the ABA problem can't occur.
 239  */
 240 static MSBlockInfo * volatile *free_block_lists [MS_BLOCK_TYPE_MAX];
 241
 242 static guint64 stat_major_blocks_alloced = 0;
 243 static guint64 stat_major_blocks_freed = 0;
 244 static guint64 stat_major_blocks_lazy_swept = 0;
 245
 246 #if SIZEOF_VOID_P != 8
 247 static guint64 stat_major_blocks_freed_ideal = 0;
 248 static guint64 stat_major_blocks_freed_less_ideal = 0;
 249 static guint64 stat_major_blocks_freed_individual = 0;
 250 static guint64 stat_major_blocks_alloced_less_ideal = 0;
 251 #endif
 252
 253 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
 254 static guint64 num_major_objects_marked = 0;
 255 #define INC_NUM_MAJOR_OBJECTS_MARKED()  (++num_major_objects_marked)
 256 #else
 257 #define INC_NUM_MAJOR_OBJECTS_MARKED()
 258 #endif
 259
 260 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
 261 static mono_mutex_t scanned_objects_list_lock;
 262 static SgenPointerQueue scanned_objects_list;
 263
 264 static void
 265 add_scanned_object (void *ptr)
 266 {
 267         if (!binary_protocol_is_enabled ())
 268                 return;
 269
 270         mono_os_mutex_lock (&scanned_objects_list_lock);
 271         sgen_pointer_queue_add (&scanned_objects_list, ptr);
 272         mono_os_mutex_unlock (&scanned_objects_list_lock);
 273 }
 274 #endif
 275
 276 static gboolean sweep_block (MSBlockInfo *block);
 277
 278 static int
 279 ms_find_block_obj_size_index (size_t size)
 280 {
 281         int i;
 282         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);
 283         for (i = 0; i < num_block_obj_sizes; ++i)
 284                 if (block_obj_sizes [i] >= size)
 285                         return i;
 286         g_error ("no object of size %zd\n", size);
 287         return -1;
 288 }
 289
 290 #define FREE_BLOCKS_FROM(lists,p,r)     (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
 291 #define FREE_BLOCKS(p,r)                (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
 292
 293 #define MS_BLOCK_OBJ_SIZE_INDEX(s)                              \
 294         (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ?      \
 295          fast_block_obj_size_indexes [((s)+7)>>3] :             \
 296          ms_find_block_obj_size_index ((s)))
 297
 298 static void*
 299 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
 300 {
 301         char *start;
 302         if (nursery_align)
 303                 start = (char *)sgen_alloc_os_memory_aligned (nursery_size, nursery_align, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery");
 304         else
 305                 start = (char *)sgen_alloc_os_memory (nursery_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery");
 306
 307         return start;
 308 }
 309
 310 static void
 311 update_heap_boundaries_for_block (MSBlockInfo *block)
 312 {
 313         sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
 314 }
 315
 316 /*
 317  * Thread safe
 318  */
 319 static void*
 320 ms_get_empty_block (void)
 321 {
 322         char *p;
 323         int i;
 324         void *block, *empty, *next;
 325
 326  retry:
 327         if (!empty_blocks) {
 328                 /*
 329                  * We try allocating MS_BLOCK_ALLOC_NUM blocks first.  If that's
 330                  * unsuccessful, we halve the number of blocks and try again, until we're at
 331                  * 1.  If that doesn't work, either, we assert.
 332                  */
 333                 int alloc_num = MS_BLOCK_ALLOC_NUM;
 334                 for (;;) {
 335                         p = (char *)sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE,
 336                                 (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE),
 337                                 alloc_num == 1 ? "major heap section" : NULL);
 338                         if (p)
 339                                 break;
 340                         alloc_num >>= 1;
 341                 }
 342
 343                 for (i = 0; i < alloc_num; ++i) {
 344                         block = p;
 345                         /*
 346                          * We do the free list update one after the
 347                          * other so that other threads can use the new
 348                          * blocks as quickly as possible.
 349                          */
 350                         do {
 351                                 empty = empty_blocks;
 352                                 *(void**)block = empty;
 353                         } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
 354                         p += MS_BLOCK_SIZE;
 355                 }
 356
 357                 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
 358
 359                 stat_major_blocks_alloced += alloc_num;
 360 #if SIZEOF_VOID_P != 8
 361                 if (alloc_num != MS_BLOCK_ALLOC_NUM)
 362                         stat_major_blocks_alloced_less_ideal += alloc_num;
 363 #endif
 364         }
 365
 366         do {
 367                 empty = empty_blocks;
 368                 if (!empty)
 369                         goto retry;
 370                 block = empty;
 371                 next = *(void**)block;
 372         } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
 373
 374         SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
 375
 376         *(void**)block = NULL;
 377
 378         g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
 379
 380         return block;
 381 }
 382
 383 /*
 384  * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
 385  * list, where it will either be freed later on, or reused in nursery collections.
 386  */
 387 static void
 388 ms_free_block (void *block)
 389 {
 390         void *empty;
 391
 392         sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
 393         memset (block, 0, MS_BLOCK_SIZE);
 394
 395         do {
 396                 empty = empty_blocks;
 397                 *(void**)block = empty;
 398         } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
 399
 400         SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
 401
 402         binary_protocol_block_free (block, MS_BLOCK_SIZE);
 403 }
 404
 405 static gboolean
 406 sweep_in_progress (void)
 407 {
 408         int state = sweep_state;
 409         return state == SWEEP_STATE_SWEEPING ||
 410                 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
 411                 state == SWEEP_STATE_COMPACTING;
 412 }
 413
 414 static inline gboolean
 415 block_is_swept_or_marking (MSBlockInfo *block)
 416 {
 417         gint32 state = block->state;
 418         return state == BLOCK_STATE_SWEPT || state == BLOCK_STATE_MARKING;
 419 }
 420
 421 //#define MARKSWEEP_CONSISTENCY_CHECK
 422
 423 #ifdef MARKSWEEP_CONSISTENCY_CHECK
 424 static void
 425 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
 426 {
 427         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 428         for (; block; block = block->next_free) {
 429                 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
 430                 g_assert (block->obj_size == size);
 431                 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
 432
 433                 /* blocks in the free lists must have at least
 434                    one free slot */
 435                 g_assert (block->free_list);
 436
 437                 /* the block must be in the allocated_blocks array */
 438                 g_assert (sgen_array_list_find (&allocated_blocks, BLOCK_TAG (block)) != (guint32)-1);
 439         }
 440 }
 441
 442 static void
 443 check_empty_blocks (void)
 444 {
 445         void *p;
 446         size_t i = 0;
 447         for (p = empty_blocks; p; p = *(void**)p)
 448                 ++i;
 449         g_assert (i == num_empty_blocks);
 450 }
 451
 452 static void
 453 consistency_check (void)
 454 {
 455         MSBlockInfo *block;
 456         int i;
 457
 458         /* check all blocks */
 459         FOREACH_BLOCK_NO_LOCK (block) {
 460                 int count = MS_BLOCK_FREE / block->obj_size;
 461                 int num_free = 0;
 462                 void **free;
 463
 464                 /* count number of free slots */
 465                 for (i = 0; i < count; ++i) {
 466                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 467                         if (!MS_OBJ_ALLOCED (obj, block))
 468                                 ++num_free;
 469                 }
 470
 471                 /* check free list */
 472                 for (free = block->free_list; free; free = (void**)*free) {
 473                         g_assert (MS_BLOCK_FOR_OBJ (free) == block);
 474                         --num_free;
 475                 }
 476                 g_assert (num_free == 0);
 477
 478                 /* check all mark words are zero */
 479                 if (!sgen_concurrent_collection_in_progress () && block_is_swept_or_marking (block)) {
 480                         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
 481                                 g_assert (block->mark_words [i] == 0);
 482                 }
 483         } END_FOREACH_BLOCK_NO_LOCK;
 484
 485         /* check free blocks */
 486         for (i = 0; i < num_block_obj_sizes; ++i) {
 487                 int j;
 488                 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
 489                         check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
 490         }
 491
 492         check_empty_blocks ();
 493 }
 494 #endif
 495
 496 static void
 497 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
 498 {
 499         MSBlockInfo *old;
 500         do {
 501                 block->next_free = old = free_blocks [size_index];
 502         } while (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], block, old) != old);
 503 }
 504
 505 static void major_finish_sweep_checking (void);
 506
 507 static gboolean
 508 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
 509 {
 510         int size = block_obj_sizes [size_index];
 511         int count = MS_BLOCK_FREE / size;
 512         MSBlockInfo *info;
 513         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 514         char *obj_start;
 515         int i;
 516
 517         if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
 518                 return FALSE;
 519
 520         info = (MSBlockInfo*)ms_get_empty_block ();
 521
 522         SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
 523
 524         info->obj_size = size;
 525         info->obj_size_index = size_index;
 526         info->pinned = pinned;
 527         info->has_references = has_references;
 528         info->has_pinned = pinned;
 529         /*
 530          * Blocks that are to-space are not evacuated from.  During an major collection
 531          * blocks are allocated for two reasons: evacuating objects from the nursery and
 532          * evacuating them from major blocks marked for evacuation.  In both cases we don't
 533          * want further evacuation. We also don't want to evacuate objects allocated during
 534          * the concurrent mark since it would add pointless stress on the finishing pause.
 535          */
 536         info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD) || sgen_concurrent_collection_in_progress ();
 537         info->state = info->is_to_space ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
 538         SGEN_ASSERT (6, !sweep_in_progress () || info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
 539         info->cardtable_mod_union = NULL;
 540
 541         update_heap_boundaries_for_block (info);
 542
 543         binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
 544
 545         /* build free list */
 546         obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
 547         info->free_list = (void**)obj_start;
 548         /* we're skipping the last one - it must be nulled */
 549         for (i = 0; i < count - 1; ++i) {
 550                 char *next_obj_start = obj_start + size;
 551                 *(void**)obj_start = next_obj_start;
 552                 obj_start = next_obj_start;
 553         }
 554         /* the last one */
 555         *(void**)obj_start = NULL;
 556
 557         add_free_block (free_blocks, size_index, info);
 558
 559         /*
 560          * Adding to the allocated_blocks array is racy with the removal of nulls when
 561          * sweeping. We wait for sweep to finish to avoid that.
 562          *
 563          * The memory barrier here and in `sweep_job_func()` are required because we need
 564          * `allocated_blocks` synchronized between this and the sweep thread.
 565          */
 566         major_finish_sweep_checking ();
 567         mono_memory_barrier ();
 568
 569         sgen_array_list_add (&allocated_blocks, BLOCK_TAG (info), 0, FALSE);
 570
 571         SGEN_ATOMIC_ADD_P (num_major_sections, 1);
 572         return TRUE;
 573 }
 574
 575 static gboolean
 576 ptr_is_from_pinned_alloc (char *ptr)
 577 {
 578         MSBlockInfo *block;
 579
 580         FOREACH_BLOCK_NO_LOCK (block) {
 581                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE)
 582                         return block->pinned;
 583         } END_FOREACH_BLOCK_NO_LOCK;
 584         return FALSE;
 585 }
 586
 587 static void
 588 ensure_can_access_block_free_list (MSBlockInfo *block)
 589 {
 590  retry:
 591         for (;;) {
 592                 switch (block->state) {
 593                 case BLOCK_STATE_SWEPT:
 594                 case BLOCK_STATE_MARKING:
 595                         return;
 596                 case BLOCK_STATE_CHECKING:
 597                         SGEN_ASSERT (0, FALSE, "How did we get a block that's being checked from a free list?");
 598                         break;
 599                 case BLOCK_STATE_NEED_SWEEPING:
 600                         if (sweep_block (block))
 601                                 ++stat_major_blocks_lazy_swept;
 602                         break;
 603                 case BLOCK_STATE_SWEEPING:
 604                         /* FIXME: do this more elegantly */
 605                         g_usleep (100);
 606                         goto retry;
 607                 default:
 608                         SGEN_ASSERT (0, FALSE, "Illegal block state");
 609                         break;
 610                 }
 611         }
 612 }
 613
 614 static void*
 615 unlink_slot_from_free_list_uncontested (MSBlockInfo * volatile *free_blocks, int size_index)
 616 {
 617         MSBlockInfo *block, *next_free_block;
 618         void *obj, *next_free_slot;
 619
 620  retry:
 621         block = free_blocks [size_index];
 622         SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
 623
 624         ensure_can_access_block_free_list (block);
 625
 626         obj = block->free_list;
 627         SGEN_ASSERT (6, obj, "block %p in free list had no available object to alloc from", block);
 628
 629         next_free_slot = *(void**)obj;
 630         if (next_free_slot) {
 631                 block->free_list = (gpointer *)next_free_slot;
 632                 return obj;
 633         }
 634
 635         next_free_block = block->next_free;
 636         if (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], next_free_block, block) != block)
 637                 goto retry;
 638
 639         block->free_list = NULL;
 640         block->next_free = NULL;
 641
 642         return obj;
 643 }
 644
 645 static GCObject*
 646 alloc_obj (GCVTable vtable, size_t size, gboolean pinned, gboolean has_references)
 647 {
 648         int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 649         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 650         void *obj;
 651
 652         if (!free_blocks [size_index]) {
 653                 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
 654                         return NULL;
 655         }
 656
 657         obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
 658
 659         /* FIXME: assumes object layout */
 660         *(GCVTable*)obj = vtable;
 661
 662         return (GCObject *)obj;
 663 }
 664
 665 static GCObject*
 666 major_alloc_object (GCVTable vtable, size_t size, gboolean has_references)
 667 {
 668         return alloc_obj (vtable, size, FALSE, has_references);
 669 }
 670
 671 /*
 672  * We're not freeing the block if it's empty.  We leave that work for
 673  * the next major collection.
 674  *
 675  * This is just called from the domain clearing code, which runs in a
 676  * single thread and has the GC lock, so we don't need an extra lock.
 677  */
 678 static void
 679 free_object (GCObject *obj, size_t size, gboolean pinned)
 680 {
 681         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 682         int word, bit;
 683         gboolean in_free_list;
 684
 685         SGEN_ASSERT (9, sweep_state == SWEEP_STATE_SWEPT, "Should have waited for sweep to free objects.");
 686
 687         ensure_can_access_block_free_list (block);
 688         SGEN_ASSERT (9, (pinned && block->pinned) || (!pinned && !block->pinned), "free-object pinning mixup object %p pinned %d block %p pinned %d", obj, pinned, block, block->pinned);
 689         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
 690         MS_CALC_MARK_BIT (word, bit, obj);
 691         SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set", obj);
 692
 693         memset (obj, 0, size);
 694
 695         in_free_list = !!block->free_list;
 696         *(void**)obj = block->free_list;
 697         block->free_list = (void**)obj;
 698
 699         if (!in_free_list) {
 700                 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (pinned, block->has_references);
 701                 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 702                 SGEN_ASSERT (9, !block->next_free, "block %p doesn't have a free-list of object but belongs to a free-list of blocks", block);
 703                 add_free_block (free_blocks, size_index, block);
 704         }
 705 }
 706
 707 static void
 708 major_free_non_pinned_object (GCObject *obj, size_t size)
 709 {
 710         free_object (obj, size, FALSE);
 711 }
 712
 713 /* size is a multiple of SGEN_ALLOC_ALIGN */
 714 static GCObject*
 715 major_alloc_small_pinned_obj (GCVTable vtable, size_t size, gboolean has_references)
 716 {
 717         void *res;
 718
 719         res = alloc_obj (vtable, size, TRUE, has_references);
 720          /*If we failed to alloc memory, we better try releasing memory
 721           *as pinned alloc is requested by the runtime.
 722           */
 723          if (!res) {
 724                 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE);
 725                 res = alloc_obj (vtable, size, TRUE, has_references);
 726          }
 727          return (GCObject *)res;
 728 }
 729
 730 static void
 731 free_pinned_object (GCObject *obj, size_t size)
 732 {
 733         free_object (obj, size, TRUE);
 734 }
 735
 736 /*
 737  * size is already rounded up and we hold the GC lock.
 738  */
 739 static GCObject*
 740 major_alloc_degraded (GCVTable vtable, size_t size)
 741 {
 742         GCObject *obj;
 743
 744         obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
 745         if (G_LIKELY (obj)) {
 746                 HEAVY_STAT (++stat_objects_alloced_degraded);
 747                 HEAVY_STAT (stat_bytes_alloced_degraded += size);
 748         }
 749         return obj;
 750 }
 751
 752 /*
 753  * obj is some object.  If it's not in the major heap (i.e. if it's in
 754  * the nursery or LOS), return FALSE.  Otherwise return whether it's
 755  * been marked or copied.
 756  */
 757 static gboolean
 758 major_is_object_live (GCObject *obj)
 759 {
 760         MSBlockInfo *block;
 761         int word, bit;
 762         mword objsize;
 763
 764         if (sgen_ptr_in_nursery (obj))
 765                 return FALSE;
 766
 767         objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
 768
 769         /* LOS */
 770         if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
 771                 return FALSE;
 772
 773         /* now we know it's in a major block */
 774         block = MS_BLOCK_FOR_OBJ (obj);
 775         SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?", block);
 776         MS_CALC_MARK_BIT (word, bit, obj);
 777         return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
 778 }
 779
 780 static gboolean
 781 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
 782 {
 783         MSBlockInfo *block;
 784
 785         FOREACH_BLOCK_NO_LOCK (block) {
 786                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
 787                         int count = MS_BLOCK_FREE / block->obj_size;
 788                         int i;
 789
 790                         *start = NULL;
 791                         for (i = 0; i <= count; ++i) {
 792                                 if (ptr >= (char*)MS_BLOCK_OBJ (block, i) && ptr < (char*)MS_BLOCK_OBJ (block, i + 1)) {
 793                                         *start = (char *)MS_BLOCK_OBJ (block, i);
 794                                         break;
 795                                 }
 796                         }
 797                         return !block->pinned;
 798                 }
 799         } END_FOREACH_BLOCK_NO_LOCK;
 800         return FALSE;
 801 }
 802
 803 static gboolean
 804 try_set_sweep_state (int new_, int expected)
 805 {
 806         int old = SGEN_CAS (&sweep_state, new_, expected);
 807         return old == expected;
 808 }
 809
 810 static void
 811 set_sweep_state (int new_, int expected)
 812 {
 813         gboolean success = try_set_sweep_state (new_, expected);
 814         SGEN_ASSERT (0, success, "Could not set sweep state.");
 815 }
 816
 817 static gboolean ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked);
 818
 819 static SgenThreadPoolJob * volatile sweep_job;
 820
 821 static void
 822 major_finish_sweep_checking (void)
 823 {
 824         guint32 block_index;
 825         SgenThreadPoolJob *job;
 826
 827  retry:
 828         switch (sweep_state) {
 829         case SWEEP_STATE_SWEPT:
 830         case SWEEP_STATE_NEED_SWEEPING:
 831                 return;
 832         case SWEEP_STATE_SWEEPING:
 833                 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
 834                         break;
 835                 goto retry;
 836         case SWEEP_STATE_SWEEPING_AND_ITERATING:
 837                 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
 838                 goto retry;
 839         case SWEEP_STATE_COMPACTING:
 840                 goto wait;
 841         default:
 842                 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
 843                 break;
 844         }
 845
 846         /*
 847          * We're running with the world stopped and the only other thread doing work is the
 848          * sweep thread, which doesn't add blocks to the array, so we can safely access
 849          * `next_slot`.
 850          */
 851         for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index)
 852                 ensure_block_is_checked_for_sweeping (block_index, FALSE, NULL);
 853
 854         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
 855
 856  wait:
 857         job = sweep_job;
 858         if (job)
 859                 sgen_thread_pool_job_wait (job);
 860         SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
 861         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
 862 }
 863
 864 static void
 865 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
 866 {
 867         gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
 868         gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
 869         gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
 870         MSBlockInfo *block;
 871
 872         major_finish_sweep_checking ();
 873         FOREACH_BLOCK_NO_LOCK (block) {
 874                 int count = MS_BLOCK_FREE / block->obj_size;
 875                 int i;
 876
 877                 if (block->pinned && !pinned)
 878                         continue;
 879                 if (!block->pinned && !non_pinned)
 880                         continue;
 881                 if (sweep && lazy_sweep) {
 882                         sweep_block (block);
 883                         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
 884                 }
 885
 886                 for (i = 0; i < count; ++i) {
 887                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 888                         /*
 889                          * We've finished sweep checking, but if we're sweeping lazily and
 890                          * the flags don't require us to sweep, the block might still need
 891                          * sweeping.  In that case, we need to consult the mark bits to tell
 892                          * us whether an object slot is live.
 893                          */
 894                         if (!block_is_swept_or_marking (block)) {
 895                                 int word, bit;
 896                                 SGEN_ASSERT (6, !sweep && block->state == BLOCK_STATE_NEED_SWEEPING, "Has sweeping not finished?");
 897                                 MS_CALC_MARK_BIT (word, bit, obj);
 898                                 if (!MS_MARK_BIT (block, word, bit))
 899                                         continue;
 900                         }
 901                         if (MS_OBJ_ALLOCED (obj, block))
 902                                 callback ((GCObject*)obj, block->obj_size, data);
 903                 }
 904         } END_FOREACH_BLOCK_NO_LOCK;
 905 }
 906
 907 static gboolean
 908 major_is_valid_object (char *object)
 909 {
 910         MSBlockInfo *block;
 911
 912         FOREACH_BLOCK_NO_LOCK (block) {
 913                 int idx;
 914                 char *obj;
 915
 916                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
 917                         continue;
 918
 919                 idx = MS_BLOCK_OBJ_INDEX (object, block);
 920                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 921                 if (obj != object)
 922                         return FALSE;
 923                 return MS_OBJ_ALLOCED (obj, block);
 924         } END_FOREACH_BLOCK_NO_LOCK;
 925
 926         return FALSE;
 927 }
 928
 929
 930 static GCVTable
 931 major_describe_pointer (char *ptr)
 932 {
 933         MSBlockInfo *block;
 934
 935         FOREACH_BLOCK_NO_LOCK (block) {
 936                 int idx;
 937                 char *obj;
 938                 gboolean live;
 939                 GCVTable vtable;
 940                 int w, b;
 941                 gboolean marked;
 942
 943                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
 944                         continue;
 945
 946                 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
 947                         MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
 948
 949                 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
 950                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 951                 live = MS_OBJ_ALLOCED (obj, block);
 952                 vtable = live ? SGEN_LOAD_VTABLE ((GCObject*)obj) : NULL;
 953
 954                 MS_CALC_MARK_BIT (w, b, obj);
 955                 marked = MS_MARK_BIT (block, w, b);
 956
 957                 if (obj == ptr) {
 958                         SGEN_LOG (0, "\t(");
 959                         if (live)
 960                                 SGEN_LOG (0, "object");
 961                         else
 962                                 SGEN_LOG (0, "dead-object");
 963                 } else {
 964                         if (live)
 965                                 SGEN_LOG (0, "interior-ptr offset %zd", ptr - obj);
 966                         else
 967                                 SGEN_LOG (0, "dead-interior-ptr offset %zd", ptr - obj);
 968                 }
 969
 970                 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
 971
 972                 return vtable;
 973         } END_FOREACH_BLOCK_NO_LOCK;
 974
 975         return NULL;
 976 }
 977
 978 static void
 979 major_check_scan_starts (void)
 980 {
 981 }
 982
 983 static void
 984 major_dump_heap (FILE *heap_dump_file)
 985 {
 986         MSBlockInfo *block;
 987         int *slots_available = (int *)alloca (sizeof (int) * num_block_obj_sizes);
 988         int *slots_used = (int *)alloca (sizeof (int) * num_block_obj_sizes);
 989         int i;
 990
 991         for (i = 0; i < num_block_obj_sizes; ++i)
 992                 slots_available [i] = slots_used [i] = 0;
 993
 994         FOREACH_BLOCK_NO_LOCK (block) {
 995                 int index = ms_find_block_obj_size_index (block->obj_size);
 996                 int count = MS_BLOCK_FREE / block->obj_size;
 997
 998                 slots_available [index] += count;
 999                 for (i = 0; i < count; ++i) {
1000                         if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
1001                                 ++slots_used [index];
1002                 }
1003         } END_FOREACH_BLOCK_NO_LOCK;
1004
1005         fprintf (heap_dump_file, "<occupancies>\n");
1006         for (i = 0; i < num_block_obj_sizes; ++i) {
1007                 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
1008                                 block_obj_sizes [i], slots_available [i], slots_used [i]);
1009         }
1010         fprintf (heap_dump_file, "</occupancies>\n");
1011
1012         FOREACH_BLOCK_NO_LOCK (block) {
1013                 int count = MS_BLOCK_FREE / block->obj_size;
1014                 int i;
1015                 int start = -1;
1016
1017                 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
1018
1019                 for (i = 0; i <= count; ++i) {
1020                         if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
1021                                 if (start < 0)
1022                                         start = i;
1023                         } else {
1024                                 if (start >= 0) {
1025                                         sgen_dump_occupied ((char *)MS_BLOCK_OBJ (block, start), (char *)MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
1026                                         start = -1;
1027                                 }
1028                         }
1029                 }
1030
1031                 fprintf (heap_dump_file, "</section>\n");
1032         } END_FOREACH_BLOCK_NO_LOCK;
1033 }
1034
1035 static guint8*
1036 get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate)
1037 {
1038         guint8 *mod_union = block->cardtable_mod_union;
1039         guint8 *other;
1040         if (mod_union)
1041                 return mod_union;
1042         else if (!allocate)
1043                 return NULL;
1044         mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1045         other = (guint8 *)SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
1046         if (!other) {
1047                 SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
1048                 return mod_union;
1049         }
1050         sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1051         return other;
1052 }
1053
1054 static inline guint8*
1055 major_get_cardtable_mod_union_for_reference (char *ptr)
1056 {
1057         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (ptr);
1058         size_t offset = sgen_card_table_get_card_offset (ptr, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1059         guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
1060         SGEN_ASSERT (0, mod_union, "FIXME: optionally allocate the mod union if it's not here and CAS it in.");
1061         return &mod_union [offset];
1062 }
1063
1064 /*
1065  * Mark the mod-union card for `ptr`, which must be a reference within the object `obj`.
1066  */
1067 static void
1068 mark_mod_union_card (GCObject *obj, void **ptr, GCObject *value_obj)
1069 {
1070         int type = sgen_obj_get_descriptor (obj) & DESC_TYPE_MASK;
1071         if (sgen_safe_object_is_small (obj, type)) {
1072                 guint8 *card_byte = major_get_cardtable_mod_union_for_reference ((char*)ptr);
1073                 SGEN_ASSERT (0, MS_BLOCK_FOR_OBJ (obj) == MS_BLOCK_FOR_OBJ (ptr), "How can an object and a reference inside it not be in the same block?");
1074                 *card_byte = 1;
1075         } else {
1076                 sgen_los_mark_mod_union_card (obj, ptr);
1077         }
1078
1079         binary_protocol_mod_union_remset (obj, ptr, value_obj, SGEN_LOAD_VTABLE (value_obj));
1080 }
1081
1082 static inline gboolean
1083 major_block_is_evacuating (MSBlockInfo *block)
1084 {
1085         if (evacuate_block_obj_sizes [block->obj_size_index] &&
1086                         !block->has_pinned &&
1087                         !block->is_to_space)
1088                 return TRUE;
1089         return FALSE;
1090 }
1091
1092 #define LOAD_VTABLE     SGEN_LOAD_VTABLE
1093
1094 #define MS_MARK_OBJECT_AND_ENQUEUE_CHECKED(obj,desc,block,queue) do {   \
1095                 int __word, __bit;                                      \
1096                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
1097                 if (!MS_MARK_BIT ((block), __word, __bit) && MS_OBJ_ALLOCED ((obj), (block))) { \
1098                         MS_SET_MARK_BIT ((block), __word, __bit);       \
1099                         if (sgen_gc_descr_has_references (desc))                        \
1100                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1101                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1102                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
1103                 }                                                       \
1104         } while (0)
1105 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do {           \
1106                 int __word, __bit;                                      \
1107                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
1108                 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
1109                 if (!MS_MARK_BIT ((block), __word, __bit)) {            \
1110                         MS_SET_MARK_BIT ((block), __word, __bit);       \
1111                         if (sgen_gc_descr_has_references (desc))                        \
1112                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1113                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1114                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
1115                 }                                                       \
1116         } while (0)
1117
1118 static void
1119 pin_major_object (GCObject *obj, SgenGrayQueue *queue)
1120 {
1121         MSBlockInfo *block;
1122
1123         if (concurrent_mark)
1124                 g_assert_not_reached ();
1125
1126         block = MS_BLOCK_FOR_OBJ (obj);
1127         block->has_pinned = TRUE;
1128         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1129 }
1130
1131 #include "sgen-major-copy-object.h"
1132
1133 static long long
1134 major_get_and_reset_num_major_objects_marked (void)
1135 {
1136 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1137         long long num = num_major_objects_marked;
1138         num_major_objects_marked = 0;
1139         return num;
1140 #else
1141         return 0;
1142 #endif
1143 }
1144
1145 #define PREFETCH_CARDS          1       /* BOOL FASTENABLE */
1146 #if !PREFETCH_CARDS
1147 #undef PREFETCH_CARDS
1148 #endif
1149
1150 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1151 #if defined(PLATFORM_MACOSX)
1152 #define GCC_VERSION (__GNUC__ * 10000 \
1153                                + __GNUC_MINOR__ * 100 \
1154                                + __GNUC_PATCHLEVEL__)
1155 #if GCC_VERSION <= 40300
1156 #undef PREFETCH_CARDS
1157 #endif
1158 #endif
1159
1160 #ifdef HEAVY_STATISTICS
1161 static guint64 stat_optimized_copy;
1162 static guint64 stat_optimized_copy_nursery;
1163 static guint64 stat_optimized_copy_nursery_forwarded;
1164 static guint64 stat_optimized_copy_nursery_pinned;
1165 static guint64 stat_optimized_copy_major;
1166 static guint64 stat_optimized_copy_major_small_fast;
1167 static guint64 stat_optimized_copy_major_small_slow;
1168 static guint64 stat_optimized_copy_major_large;
1169 static guint64 stat_optimized_copy_major_forwarded;
1170 static guint64 stat_optimized_copy_major_small_evacuate;
1171 static guint64 stat_optimized_major_scan;
1172 static guint64 stat_optimized_major_scan_no_refs;
1173
1174 static guint64 stat_drain_prefetch_fills;
1175 static guint64 stat_drain_prefetch_fill_failures;
1176 static guint64 stat_drain_loops;
1177 #endif
1178
1179 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_no_evacuation
1180 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_no_evacuation
1181 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_no_evacuation
1182 #include "sgen-marksweep-drain-gray-stack.h"
1183
1184 #define COPY_OR_MARK_WITH_EVACUATION
1185 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_with_evacuation
1186 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_with_evacuation
1187 #define SCAN_VTYPE_FUNCTION_NAME        major_scan_vtype_with_evacuation
1188 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_with_evacuation
1189 #define SCAN_PTR_FIELD_FUNCTION_NAME    major_scan_ptr_field_with_evacuation
1190 #include "sgen-marksweep-drain-gray-stack.h"
1191
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 #define COPY_OR_MARK_CONCURRENT_WITH_EVACUATION
1199 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_concurrent_with_evacuation
1200 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_concurrent_with_evacuation
1201 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_concurrent_with_evacuation
1202 #include "sgen-marksweep-drain-gray-stack.h"
1203
1204 static inline gboolean
1205 major_is_evacuating (void)
1206 {
1207         int i;
1208         for (i = 0; i < num_block_obj_sizes; ++i) {
1209                 if (evacuate_block_obj_sizes [i]) {
1210                         return TRUE;
1211                 }
1212         }
1213
1214         return FALSE;
1215 }
1216
1217 static gboolean
1218 drain_gray_stack (SgenGrayQueue *queue)
1219 {
1220         if (major_is_evacuating ())
1221                 return drain_gray_stack_with_evacuation (queue);
1222         else
1223                 return drain_gray_stack_no_evacuation (queue);
1224 }
1225
1226 static gboolean
1227 drain_gray_stack_concurrent (SgenGrayQueue *queue)
1228 {
1229         if (major_is_evacuating ())
1230                 return drain_gray_stack_concurrent_with_evacuation (queue);
1231         else
1232                 return drain_gray_stack_concurrent_no_evacuation (queue);
1233 }
1234
1235 static void
1236 major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue)
1237 {
1238         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1239 }
1240
1241 static void
1242 major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *queue)
1243 {
1244         major_copy_or_mark_object_concurrent_with_evacuation (ptr, *ptr, queue);
1245 }
1246
1247 static void
1248 major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue)
1249 {
1250         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1251 }
1252
1253 static void
1254 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1255 {
1256         void **entry, **end;
1257         int last_index = -1;
1258
1259         if (first_entry == last_entry)
1260                 return;
1261
1262         block->has_pinned = TRUE;
1263
1264         entry = sgen_pinning_get_entry (first_entry);
1265         end = sgen_pinning_get_entry (last_entry);
1266
1267         for (; entry < end; ++entry) {
1268                 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1269                 GCObject *obj;
1270                 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));
1271                 if (index == last_index)
1272                         continue;
1273                 obj = MS_BLOCK_OBJ (block, index);
1274                 MS_MARK_OBJECT_AND_ENQUEUE_CHECKED (obj, sgen_obj_get_descriptor (obj), block, queue);
1275                 last_index = index;
1276         }
1277 }
1278
1279 static inline void
1280 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1281 {
1282         int obj_index;
1283
1284         for (obj_index = 0; obj_index < count; ++obj_index) {
1285                 int word, bit;
1286                 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1287
1288                 MS_CALC_MARK_BIT (word, bit, obj);
1289                 if (MS_MARK_BIT (block, word, bit)) {
1290                         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1291                 } else {
1292                         /* an unmarked object */
1293                         if (MS_OBJ_ALLOCED (obj, block)) {
1294                                 /*
1295                                  * FIXME: Merge consecutive
1296                                  * slots for lower reporting
1297                                  * overhead.  Maybe memset
1298                                  * will also benefit?
1299                                  */
1300                                 binary_protocol_empty (obj, obj_size);
1301                                 memset (obj, 0, obj_size);
1302                         }
1303                         *(void**)obj = block->free_list;
1304                         block->free_list = (void **)obj;
1305                 }
1306         }
1307 }
1308
1309 static inline gboolean
1310 try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1311 {
1312         gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
1313         gboolean success = old_state == expected_state;
1314         if (success)
1315                 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
1316         return success;
1317 }
1318
1319 static inline void
1320 set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1321 {
1322         SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set");
1323         block->state = new_state;
1324 }
1325
1326 /*
1327  * If `block` needs sweeping, sweep it and return TRUE.  Otherwise return FALSE.
1328  *
1329  * Sweeping means iterating through the block's slots and building the free-list from the
1330  * unmarked ones.  They will also be zeroed.  The mark bits will be reset.
1331  */
1332 static gboolean
1333 sweep_block (MSBlockInfo *block)
1334 {
1335         int count;
1336         void *reversed = NULL;
1337
1338  retry:
1339         switch (block->state) {
1340         case BLOCK_STATE_SWEPT:
1341                 return FALSE;
1342         case BLOCK_STATE_MARKING:
1343         case BLOCK_STATE_CHECKING:
1344                 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1345                 goto retry;
1346         case BLOCK_STATE_SWEEPING:
1347                 /* FIXME: Do this more elegantly */
1348                 g_usleep (100);
1349                 goto retry;
1350         case BLOCK_STATE_NEED_SWEEPING:
1351                 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1352                         goto retry;
1353                 break;
1354         default:
1355                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1356         }
1357
1358         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1359
1360         count = MS_BLOCK_FREE / block->obj_size;
1361
1362         block->free_list = NULL;
1363
1364         /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1365         // FIXME: Add more sizes
1366         switch (block->obj_size) {
1367         case 16:
1368                 sweep_block_for_size (block, count, 16);
1369                 break;
1370         default:
1371                 sweep_block_for_size (block, count, block->obj_size);
1372                 break;
1373         }
1374
1375         /* reset mark bits */
1376         memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1377
1378         /* Reverse free list so that it's in address order */
1379         reversed = NULL;
1380         while (block->free_list) {
1381                 void *next = *(void**)block->free_list;
1382                 *(void**)block->free_list = reversed;
1383                 reversed = block->free_list;
1384                 block->free_list = (void **)next;
1385         }
1386         block->free_list = (void **)reversed;
1387
1388         mono_memory_write_barrier ();
1389
1390         set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1391
1392         return TRUE;
1393 }
1394
1395 static inline int
1396 bitcount (mword d)
1397 {
1398         int count = 0;
1399
1400 #ifdef __GNUC__
1401         if (sizeof (mword) == 8)
1402                 count += __builtin_popcountll (d);
1403         else
1404                 count += __builtin_popcount (d);
1405 #else
1406         while (d) {
1407                 count ++;
1408                 d &= (d - 1);
1409         }
1410 #endif
1411         return count;
1412 }
1413
1414 /* statistics for evacuation */
1415 static size_t *sweep_slots_available;
1416 static size_t *sweep_slots_used;
1417 static size_t *sweep_num_blocks;
1418
1419 static volatile size_t num_major_sections_before_sweep;
1420 static volatile size_t num_major_sections_freed_in_sweep;
1421
1422 static void
1423 sweep_start (void)
1424 {
1425         int i;
1426
1427         for (i = 0; i < num_block_obj_sizes; ++i)
1428                 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1429
1430         /* clear all the free lists */
1431         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1432                 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1433                 int j;
1434                 for (j = 0; j < num_block_obj_sizes; ++j)
1435                         free_blocks [j] = NULL;
1436         }
1437 }
1438
1439 static void sweep_finish (void);
1440
1441 /*
1442  * If `wait` is TRUE and the block is currently being checked, this function will wait until
1443  * the checking has finished.
1444  *
1445  * Returns whether the block is still there.  If `wait` is FALSE, the return value will not
1446  * be correct, i.e. must not be used.
1447  */
1448 static gboolean
1449 ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked)
1450 {
1451         int count;
1452         gboolean have_live = FALSE;
1453         gboolean have_free = FALSE;
1454         int nused = 0;
1455         int block_state;
1456         int i;
1457         void *tagged_block;
1458         MSBlockInfo *block;
1459         volatile gpointer *block_slot = sgen_array_list_get_slot (&allocated_blocks, block_index);
1460
1461         SGEN_ASSERT (6, sweep_in_progress (), "Why do we call this function if there's no sweep in progress?");
1462
1463         if (have_checked)
1464                 *have_checked = FALSE;
1465
1466  retry:
1467         tagged_block = *(void * volatile *)block_slot;
1468         if (!tagged_block)
1469                 return FALSE;
1470
1471         if (BLOCK_IS_TAGGED_CHECKING (tagged_block)) {
1472                 if (!wait)
1473                         return FALSE;
1474                 /* FIXME: do this more elegantly */
1475                 g_usleep (100);
1476                 goto retry;
1477         }
1478
1479         if (SGEN_CAS_PTR (block_slot, BLOCK_TAG_CHECKING (tagged_block), tagged_block) != tagged_block)
1480                 goto retry;
1481
1482         block = BLOCK_UNTAG (tagged_block);
1483         block_state = block->state;
1484
1485         if (!sweep_in_progress ()) {
1486                 SGEN_ASSERT (6, block_state != BLOCK_STATE_SWEEPING && block_state != BLOCK_STATE_CHECKING, "Invalid block state.");
1487                 if (!lazy_sweep)
1488                         SGEN_ASSERT (6, block_state != BLOCK_STATE_NEED_SWEEPING, "Invalid block state.");
1489         }
1490
1491         switch (block_state) {
1492         case BLOCK_STATE_SWEPT:
1493         case BLOCK_STATE_NEED_SWEEPING:
1494         case BLOCK_STATE_SWEEPING:
1495                 goto done;
1496         case BLOCK_STATE_MARKING:
1497                 break;
1498         case BLOCK_STATE_CHECKING:
1499                 SGEN_ASSERT (0, FALSE, "We set the CHECKING bit - how can the stage be CHECKING?");
1500                 goto done;
1501         default:
1502                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1503                 break;
1504         }
1505
1506         SGEN_ASSERT (6, block->state == BLOCK_STATE_MARKING, "When we sweep all blocks must start out marking.");
1507         set_block_state (block, BLOCK_STATE_CHECKING, BLOCK_STATE_MARKING);
1508
1509         if (have_checked)
1510                 *have_checked = TRUE;
1511
1512         block->has_pinned = block->pinned;
1513
1514         block->is_to_space = FALSE;
1515
1516         count = MS_BLOCK_FREE / block->obj_size;
1517
1518         if (block->cardtable_mod_union) {
1519                 sgen_card_table_free_mod_union (block->cardtable_mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1520                 block->cardtable_mod_union = NULL;
1521         }
1522
1523         /* Count marked objects in the block */
1524         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1525                 nused += bitcount (block->mark_words [i]);
1526
1527         if (nused)
1528                 have_live = TRUE;
1529         if (nused < count)
1530                 have_free = TRUE;
1531
1532         if (have_live) {
1533                 int obj_size_index = block->obj_size_index;
1534                 gboolean has_pinned = block->has_pinned;
1535
1536                 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1537
1538                 /*
1539                  * FIXME: Go straight to SWEPT if there are no free slots.  We need
1540                  * to set the free slot list to NULL, though, and maybe update some
1541                  * statistics.
1542                  */
1543                 if (!lazy_sweep)
1544                         sweep_block (block);
1545
1546                 if (!has_pinned) {
1547                         ++sweep_num_blocks [obj_size_index];
1548                         sweep_slots_used [obj_size_index] += nused;
1549                         sweep_slots_available [obj_size_index] += count;
1550                 }
1551
1552                 /*
1553                  * If there are free slots in the block, add
1554                  * the block to the corresponding free list.
1555                  */
1556                 if (have_free) {
1557                         MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1558
1559                         if (!lazy_sweep)
1560                                 SGEN_ASSERT (6, block->free_list, "How do we not have a free list when there are free slots?");
1561
1562                         add_free_block (free_blocks, obj_size_index, block);
1563                 }
1564
1565                 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1566                 update_heap_boundaries_for_block (block);
1567         } else {
1568                 /*
1569                  * Blocks without live objects are removed from the
1570                  * block list and freed.
1571                  */
1572                 SGEN_ASSERT (6, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1573                 SGEN_ASSERT (6, *block_slot == BLOCK_TAG_CHECKING (tagged_block), "How did the block move?");
1574
1575                 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1576                 ms_free_block (block);
1577
1578                 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1579
1580                 tagged_block = NULL;
1581         }
1582
1583  done:
1584         *block_slot = tagged_block;
1585         return !!tagged_block;
1586 }
1587
1588 static void
1589 sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1590 {
1591         guint32 block_index;
1592         guint32 num_blocks = num_major_sections_before_sweep;
1593
1594         SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1595         SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1596
1597         /*
1598          * We traverse the block array from high to low.  Nursery collections will have to
1599          * cooperate with the sweep thread to finish sweeping, and they will traverse from
1600          * low to high, to avoid constantly colliding on the same blocks.
1601          */
1602         for (block_index = num_blocks; block_index-- > 0;) {
1603                 /*
1604                  * The block might have been freed by another thread doing some checking
1605                  * work.
1606                  */
1607                 if (!ensure_block_is_checked_for_sweeping (block_index, TRUE, NULL))
1608                         ++num_major_sections_freed_in_sweep;
1609         }
1610
1611         while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) {
1612                 /*
1613                  * The main GC thread is currently iterating over the block array to help us
1614                  * finish the sweep.  We have already finished, but we don't want to mess up
1615                  * that iteration, so we just wait for it.
1616                  */
1617                 g_usleep (100);
1618         }
1619
1620         if (SGEN_MAX_ASSERT_LEVEL >= 6) {
1621                 for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1622                         MSBlockInfo *block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, block_index));
1623                         SGEN_ASSERT (6, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1624                 }
1625         }
1626
1627         sgen_array_list_remove_nulls (&allocated_blocks);
1628
1629         sweep_finish ();
1630
1631         sweep_job = NULL;
1632 }
1633
1634 static void
1635 sweep_finish (void)
1636 {
1637         int i;
1638
1639         for (i = 0; i < num_block_obj_sizes; ++i) {
1640                 float usage = (float)sweep_slots_used [i] / (float)sweep_slots_available [i];
1641                 if (sweep_num_blocks [i] > 5 && usage < evacuation_threshold) {
1642                         evacuate_block_obj_sizes [i] = TRUE;
1643                         /*
1644                         g_print ("slot size %d - %d of %d used\n",
1645                                         block_obj_sizes [i], slots_used [i], slots_available [i]);
1646                         */
1647                 } else {
1648                         evacuate_block_obj_sizes [i] = FALSE;
1649                 }
1650         }
1651
1652         set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1653 }
1654
1655 static void
1656 major_sweep (void)
1657 {
1658         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1659
1660         sweep_start ();
1661
1662         SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?");
1663
1664         num_major_sections_before_sweep = num_major_sections;
1665         num_major_sections_freed_in_sweep = 0;
1666
1667         SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
1668         if (concurrent_sweep) {
1669                 sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
1670                 sgen_thread_pool_job_enqueue (sweep_job);
1671         } else {
1672                 sweep_job_func (NULL, NULL);
1673         }
1674 }
1675
1676 static gboolean
1677 major_have_swept (void)
1678 {
1679         return sweep_state == SWEEP_STATE_SWEPT;
1680 }
1681
1682 static int count_pinned_ref;
1683 static int count_pinned_nonref;
1684 static int count_nonpinned_ref;
1685 static int count_nonpinned_nonref;
1686
1687 static void
1688 count_nonpinned_callback (GCObject *obj, size_t size, void *data)
1689 {
1690         GCVTable vtable = LOAD_VTABLE (obj);
1691
1692         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1693                 ++count_nonpinned_ref;
1694         else
1695                 ++count_nonpinned_nonref;
1696 }
1697
1698 static void
1699 count_pinned_callback (GCObject *obj, size_t size, void *data)
1700 {
1701         GCVTable vtable = LOAD_VTABLE (obj);
1702
1703         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1704                 ++count_pinned_ref;
1705         else
1706                 ++count_pinned_nonref;
1707 }
1708
1709 static G_GNUC_UNUSED void
1710 count_ref_nonref_objs (void)
1711 {
1712         int total;
1713
1714         count_pinned_ref = 0;
1715         count_pinned_nonref = 0;
1716         count_nonpinned_ref = 0;
1717         count_nonpinned_nonref = 0;
1718
1719         major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1720         major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1721
1722         total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1723
1724         g_print ("ref: %d pinned %d non-pinned   non-ref: %d pinned %d non-pinned  --  %.1f\n",
1725                         count_pinned_ref, count_nonpinned_ref,
1726                         count_pinned_nonref, count_nonpinned_nonref,
1727                         (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1728 }
1729
1730 static int
1731 ms_calculate_block_obj_sizes (double factor, int *arr)
1732 {
1733         double target_size;
1734         int num_sizes = 0;
1735         int last_size = 0;
1736
1737         /*
1738          * Have every possible slot size starting with the minimal
1739          * object size up to and including four times that size.  Then
1740          * proceed by increasing geometrically with the given factor.
1741          */
1742
1743         for (int size = SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size <= 4 * SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size += SGEN_ALLOC_ALIGN) {
1744                 if (arr)
1745                         arr [num_sizes] = size;
1746                 ++num_sizes;
1747                 last_size = size;
1748         }
1749         target_size = (double)last_size;
1750
1751         do {
1752                 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1753                 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1754
1755                 if (size != last_size) {
1756                         if (arr)
1757                                 arr [num_sizes] = size;
1758                         ++num_sizes;
1759                         last_size = size;
1760                 }
1761
1762                 target_size *= factor;
1763         } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1764
1765         return num_sizes;
1766 }
1767
1768 /* only valid during minor collections */
1769 static mword old_num_major_sections;
1770
1771 static void
1772 major_start_nursery_collection (void)
1773 {
1774 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1775         consistency_check ();
1776 #endif
1777
1778         old_num_major_sections = num_major_sections;
1779 }
1780
1781 static void
1782 major_finish_nursery_collection (void)
1783 {
1784 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1785         consistency_check ();
1786 #endif
1787 }
1788
1789 static void
1790 major_start_major_collection (void)
1791 {
1792         MSBlockInfo *block;
1793         int i;
1794
1795         major_finish_sweep_checking ();
1796
1797         /*
1798          * Clear the free lists for block sizes where we do evacuation.  For those block
1799          * sizes we will have to allocate new blocks.
1800          */
1801         for (i = 0; i < num_block_obj_sizes; ++i) {
1802                 if (!evacuate_block_obj_sizes [i])
1803                         continue;
1804
1805                 binary_protocol_evacuating_blocks (block_obj_sizes [i]);
1806
1807                 free_block_lists [0][i] = NULL;
1808                 free_block_lists [MS_BLOCK_FLAG_REFS][i] = NULL;
1809         }
1810
1811         if (lazy_sweep)
1812                 binary_protocol_sweep_begin (GENERATION_OLD, TRUE);
1813
1814         /* Sweep all unswept blocks and set them to MARKING */
1815         FOREACH_BLOCK_NO_LOCK (block) {
1816                 if (lazy_sweep)
1817                         sweep_block (block);
1818                 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1819                 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1820         } END_FOREACH_BLOCK_NO_LOCK;
1821
1822         if (lazy_sweep)
1823                 binary_protocol_sweep_end (GENERATION_OLD, TRUE);
1824
1825         set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1826 }
1827
1828 static void
1829 major_finish_major_collection (ScannedObjectCounts *counts)
1830 {
1831 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1832         if (binary_protocol_is_enabled ()) {
1833                 counts->num_scanned_objects = scanned_objects_list.next_slot;
1834
1835                 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1836                 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1837
1838                 sgen_pointer_queue_clear (&scanned_objects_list);
1839         }
1840 #endif
1841 }
1842
1843 #if SIZEOF_VOID_P != 8
1844 static int
1845 compare_pointers (const void *va, const void *vb) {
1846         char *a = *(char**)va, *b = *(char**)vb;
1847         if (a < b)
1848                 return -1;
1849         if (a > b)
1850                 return 1;
1851         return 0;
1852 }
1853 #endif
1854
1855 /*
1856  * This is called with sweep completed and the world stopped.
1857  */
1858 static void
1859 major_free_swept_blocks (size_t allowance)
1860 {
1861         /* FIXME: This is probably too much.  It's assuming all objects are small. */
1862         size_t section_reserve = allowance / MS_BLOCK_SIZE;
1863
1864         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1865
1866 #ifdef TARGET_WIN32
1867                 /*
1868                  * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1869                  * a VirtualAlloc ()-ed block.
1870                  */
1871                 return;
1872 #endif
1873
1874 #if SIZEOF_VOID_P != 8
1875         {
1876                 int i, num_empty_blocks_orig, num_blocks, arr_length;
1877                 void *block;
1878                 void **empty_block_arr;
1879                 void **rebuild_next;
1880
1881                 if (num_empty_blocks <= section_reserve)
1882                         return;
1883                 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1884
1885                 num_empty_blocks_orig = num_empty_blocks;
1886                 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1887                                 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1888                 if (!empty_block_arr)
1889                         goto fallback;
1890
1891                 i = 0;
1892                 for (block = empty_blocks; block; block = *(void**)block)
1893                         empty_block_arr [i++] = block;
1894                 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1895
1896                 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1897
1898                 /*
1899                  * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1900                  * contiguous ones.  If we do, we free them.  If that's not enough to get to
1901                  * section_reserve, we halve the number of contiguous blocks we're looking
1902                  * for and have another go, until we're done with looking for pairs of
1903                  * blocks, at which point we give up and go to the fallback.
1904                  */
1905                 arr_length = num_empty_blocks_orig;
1906                 num_blocks = MS_BLOCK_ALLOC_NUM;
1907                 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1908                         int first = -1;
1909                         int dest = 0;
1910
1911                         dest = 0;
1912                         for (i = 0; i < arr_length; ++i) {
1913                                 int d = dest;
1914                                 void *block = empty_block_arr [i];
1915                                 SGEN_ASSERT (6, block, "we're not shifting correctly");
1916                                 if (i != dest) {
1917                                         empty_block_arr [dest] = block;
1918                                         /*
1919                                          * This is not strictly necessary, but we're
1920                                          * cautious.
1921                                          */
1922                                         empty_block_arr [i] = NULL;
1923                                 }
1924                                 ++dest;
1925
1926                                 if (first < 0) {
1927                                         first = d;
1928                                         continue;
1929                                 }
1930
1931                                 SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
1932
1933                                 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
1934                                         first = d;
1935                                         continue;
1936                                 }
1937
1938                                 if (d + 1 - first == num_blocks) {
1939                                         /*
1940                                          * We found num_blocks contiguous blocks.  Free them
1941                                          * and null their array entries.  As an optimization
1942                                          * we could, instead of nulling the entries, shift
1943                                          * the following entries over to the left, while
1944                                          * we're iterating.
1945                                          */
1946                                         int j;
1947                                         sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
1948                                         for (j = first; j <= d; ++j)
1949                                                 empty_block_arr [j] = NULL;
1950                                         dest = first;
1951                                         first = -1;
1952
1953                                         num_empty_blocks -= num_blocks;
1954
1955                                         stat_major_blocks_freed += num_blocks;
1956                                         if (num_blocks == MS_BLOCK_ALLOC_NUM)
1957                                                 stat_major_blocks_freed_ideal += num_blocks;
1958                                         else
1959                                                 stat_major_blocks_freed_less_ideal += num_blocks;
1960
1961                                 }
1962                         }
1963
1964                         SGEN_ASSERT (6, dest <= i && dest <= arr_length, "array length is off");
1965                         arr_length = dest;
1966                         SGEN_ASSERT (6, arr_length == num_empty_blocks, "array length is off");
1967
1968                         num_blocks >>= 1;
1969                 }
1970
1971                 /* rebuild empty_blocks free list */
1972                 rebuild_next = (void**)&empty_blocks;
1973                 for (i = 0; i < arr_length; ++i) {
1974                         void *block = empty_block_arr [i];
1975                         SGEN_ASSERT (6, block, "we're missing blocks");
1976                         *rebuild_next = block;
1977                         rebuild_next = (void**)block;
1978                 }
1979                 *rebuild_next = NULL;
1980
1981                 /* free array */
1982                 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
1983         }
1984
1985         SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
1986
1987  fallback:
1988         /*
1989          * This is our threshold.  If there's not more empty than used blocks, we won't
1990          * release uncontiguous blocks, in fear of fragmenting the address space.
1991          */
1992         if (num_empty_blocks <= num_major_sections)
1993                 return;
1994 #endif
1995
1996         while (num_empty_blocks > section_reserve) {
1997                 void *next = *(void**)empty_blocks;
1998                 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
1999                 empty_blocks = next;
2000                 /*
2001                  * Needs not be atomic because this is running
2002                  * single-threaded.
2003                  */
2004                 --num_empty_blocks;
2005
2006                 ++stat_major_blocks_freed;
2007 #if SIZEOF_VOID_P != 8
2008                 ++stat_major_blocks_freed_individual;
2009 #endif
2010         }
2011 }
2012
2013 static void
2014 major_pin_objects (SgenGrayQueue *queue)
2015 {
2016         MSBlockInfo *block;
2017
2018         FOREACH_BLOCK_NO_LOCK (block) {
2019                 size_t first_entry, last_entry;
2020                 SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
2021                 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
2022                                 &first_entry, &last_entry);
2023                 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
2024         } END_FOREACH_BLOCK_NO_LOCK;
2025 }
2026
2027 static void
2028 major_init_to_space (void)
2029 {
2030 }
2031
2032 static void
2033 major_report_pinned_memory_usage (void)
2034 {
2035         g_assert_not_reached ();
2036 }
2037
2038 static gint64
2039 major_get_used_size (void)
2040 {
2041         gint64 size = 0;
2042         MSBlockInfo *block;
2043
2044         /*
2045          * We're holding the GC lock, but the sweep thread might be running.  Make sure it's
2046          * finished, then we can iterate over the block array.
2047          */
2048         major_finish_sweep_checking ();
2049
2050         FOREACH_BLOCK_NO_LOCK (block) {
2051                 int count = MS_BLOCK_FREE / block->obj_size;
2052                 void **iter;
2053                 size += count * block->obj_size;
2054                 for (iter = block->free_list; iter; iter = (void**)*iter)
2055                         size -= block->obj_size;
2056         } END_FOREACH_BLOCK_NO_LOCK;
2057
2058         return size;
2059 }
2060
2061 /* FIXME: return number of bytes, not of sections */
2062 static size_t
2063 get_num_major_sections (void)
2064 {
2065         return num_major_sections;
2066 }
2067
2068 /*
2069  * Returns the number of bytes in blocks that were present when the last sweep was
2070  * initiated, and were not freed during the sweep.  They are the basis for calculating the
2071  * allowance.
2072  */
2073 static size_t
2074 get_bytes_survived_last_sweep (void)
2075 {
2076         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2077         return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
2078 }
2079
2080 static gboolean
2081 major_handle_gc_param (const char *opt)
2082 {
2083         if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2084                 const char *arg = strchr (opt, '=') + 1;
2085                 int percentage = atoi (arg);
2086                 if (percentage < 0 || percentage > 100) {
2087                         fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2088                         exit (1);
2089                 }
2090                 evacuation_threshold = (float)percentage / 100.0f;
2091                 return TRUE;
2092         } else if (!strcmp (opt, "lazy-sweep")) {
2093                 lazy_sweep = TRUE;
2094                 return TRUE;
2095         } else if (!strcmp (opt, "no-lazy-sweep")) {
2096                 lazy_sweep = FALSE;
2097                 return TRUE;
2098         } else if (!strcmp (opt, "concurrent-sweep")) {
2099                 concurrent_sweep = TRUE;
2100                 return TRUE;
2101         } else if (!strcmp (opt, "no-concurrent-sweep")) {
2102                 concurrent_sweep = FALSE;
2103                 return TRUE;
2104         }
2105
2106         return FALSE;
2107 }
2108
2109 static void
2110 major_print_gc_param_usage (void)
2111 {
2112         fprintf (stderr,
2113                         ""
2114                         "  evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2115                         "  (no-)lazy-sweep\n"
2116                         "  (no-)concurrent-sweep\n"
2117                         );
2118 }
2119
2120 /*
2121  * This callback is used to clear cards, move cards to the shadow table and do counting.
2122  */
2123 static void
2124 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
2125 {
2126         MSBlockInfo *block;
2127         gboolean has_references;
2128
2129         major_finish_sweep_checking ();
2130         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2131                 if (has_references)
2132                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2133         } END_FOREACH_BLOCK_NO_LOCK;
2134 }
2135
2136 #ifdef HEAVY_STATISTICS
2137 extern guint64 marked_cards;
2138 extern guint64 scanned_cards;
2139 extern guint64 scanned_objects;
2140 extern guint64 remarked_cards;
2141 #endif
2142
2143 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2144 /*
2145  * MS blocks are 16K aligned.
2146  * Cardtables are 4K aligned, at least.
2147  * This means that the cardtable of a given block is 32 bytes aligned.
2148  */
2149 static guint8*
2150 initial_skip_card (guint8 *card_data)
2151 {
2152         mword *cards = (mword*)card_data;
2153         mword card;
2154         int i;
2155         for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2156                 card = cards [i];
2157                 if (card)
2158                         break;
2159         }
2160
2161         if (i == CARD_WORDS_PER_BLOCK)
2162                 return card_data + CARDS_PER_BLOCK;
2163
2164 #if defined(__i386__) && defined(__GNUC__)
2165         return card_data + i * 4 +  (__builtin_ffs (card) - 1) / 8;
2166 #elif defined(__x86_64__) && defined(__GNUC__)
2167         return card_data + i * 8 +  (__builtin_ffsll (card) - 1) / 8;
2168 #elif defined(__s390x__) && defined(__GNUC__)
2169         return card_data + i * 8 +  (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2170 #else
2171         for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2172                 if (card_data [i])
2173                         return &card_data [i];
2174         }
2175         return card_data;
2176 #endif
2177 }
2178
2179 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2180 #define MS_BLOCK_OBJ_FAST(b,os,i)                       ((b) + MS_BLOCK_SKIP + (os) * (i))
2181 #define MS_OBJ_ALLOCED_FAST(o,b)                (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2182
2183 static void
2184 scan_card_table_for_block (MSBlockInfo *block, gboolean mod_union, ScanCopyContext ctx)
2185 {
2186         SgenGrayQueue *queue = ctx.queue;
2187         ScanObjectFunc scan_func = ctx.ops->scan_object;
2188 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2189         guint8 cards_copy [CARDS_PER_BLOCK];
2190 #endif
2191         gboolean small_objects;
2192         int block_obj_size;
2193         char *block_start;
2194         guint8 *card_data, *card_base;
2195         guint8 *card_data_end;
2196         char *scan_front = NULL;
2197
2198         block_obj_size = block->obj_size;
2199         small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2200
2201         block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2202
2203         /*
2204          * This is safe in face of card aliasing for the following reason:
2205          *
2206          * Major blocks are 16k aligned, or 32 cards aligned.
2207          * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2208          * sizes, they won't overflow the cardtable overlap modulus.
2209          */
2210         if (mod_union) {
2211                 card_data = card_base = block->cardtable_mod_union;
2212                 /*
2213                  * This happens when the nursery collection that precedes finishing
2214                  * the concurrent collection allocates new major blocks.
2215                  */
2216                 if (!card_data)
2217                         return;
2218         } else {
2219 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2220                 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2221 #else
2222                 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2223                         return;
2224                 card_data = card_base = cards_copy;
2225 #endif
2226         }
2227         card_data_end = card_data + CARDS_PER_BLOCK;
2228
2229         card_data += MS_BLOCK_SKIP >> CARD_BITS;
2230
2231         card_data = initial_skip_card (card_data);
2232         while (card_data < card_data_end) {
2233                 size_t card_index, first_object_index;
2234                 char *start;
2235                 char *end;
2236                 char *first_obj, *obj;
2237
2238                 HEAVY_STAT (++scanned_cards);
2239
2240                 if (!*card_data) {
2241                         ++card_data;
2242                         continue;
2243                 }
2244
2245                 card_index = card_data - card_base;
2246                 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2247                 end = start + CARD_SIZE_IN_BYTES;
2248
2249                 if (!block_is_swept_or_marking (block))
2250                         sweep_block (block);
2251
2252                 HEAVY_STAT (++marked_cards);
2253
2254                 if (small_objects)
2255                         sgen_card_table_prepare_card_for_scanning (card_data);
2256
2257                 /*
2258                  * If the card we're looking at starts at or in the block header, we
2259                  * must start at the first object in the block, without calculating
2260                  * the index of the object we're hypothetically starting at, because
2261                  * it would be negative.
2262                  */
2263                 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2264                         first_object_index = 0;
2265                 else
2266                         first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2267
2268                 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2269
2270                 binary_protocol_card_scan (first_obj, end - first_obj);
2271
2272                 while (obj < end) {
2273                         if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2274                                 goto next_object;
2275
2276                         if (mod_union) {
2277                                 /* FIXME: do this more efficiently */
2278                                 int w, b;
2279                                 MS_CALC_MARK_BIT (w, b, obj);
2280                                 if (!MS_MARK_BIT (block, w, b))
2281                                         goto next_object;
2282                         }
2283
2284                         GCObject *object = (GCObject*)obj;
2285
2286                         if (small_objects) {
2287                                 HEAVY_STAT (++scanned_objects);
2288                                 scan_func (object, sgen_obj_get_descriptor (object), queue);
2289                         } else {
2290                                 size_t offset = sgen_card_table_get_card_offset (obj, block_start);
2291                                 sgen_cardtable_scan_object (object, block_obj_size, card_base + offset, mod_union, ctx);
2292                         }
2293                 next_object:
2294                         obj += block_obj_size;
2295                         g_assert (scan_front <= obj);
2296                         scan_front = obj;
2297                 }
2298
2299                 HEAVY_STAT (if (*card_data) ++remarked_cards);
2300
2301                 if (small_objects)
2302                         ++card_data;
2303                 else
2304                         card_data = card_base + sgen_card_table_get_card_offset (obj, block_start);
2305         }
2306 }
2307
2308 static void
2309 major_scan_card_table (gboolean mod_union, ScanCopyContext ctx)
2310 {
2311         MSBlockInfo *block;
2312         gboolean has_references;
2313
2314         if (!concurrent_mark)
2315                 g_assert (!mod_union);
2316
2317         major_finish_sweep_checking ();
2318         binary_protocol_major_card_table_scan_start (sgen_timestamp (), mod_union);
2319         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2320 #ifdef PREFETCH_CARDS
2321                 int prefetch_index = __index + 6;
2322                 if (prefetch_index < allocated_blocks.next_slot) {
2323                         MSBlockInfo *prefetch_block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, prefetch_index));
2324                         PREFETCH_READ (prefetch_block);
2325                         if (!mod_union) {
2326                                 guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2327                                 PREFETCH_WRITE (prefetch_cards);
2328                                 PREFETCH_WRITE (prefetch_cards + 32);
2329                         }
2330                 }
2331 #endif
2332
2333                 if (!has_references)
2334                         continue;
2335
2336                 scan_card_table_for_block (block, mod_union, ctx);
2337         } END_FOREACH_BLOCK_NO_LOCK;
2338         binary_protocol_major_card_table_scan_end (sgen_timestamp (), mod_union);
2339 }
2340
2341 static void
2342 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2343 {
2344         MSBlockInfo *block;
2345         gboolean has_references;
2346         long long total_cards = 0;
2347         long long marked_cards = 0;
2348
2349         if (sweep_in_progress ()) {
2350                 *num_total_cards = -1;
2351                 *num_marked_cards = -1;
2352                 return;
2353         }
2354
2355         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2356                 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2357                 int i;
2358
2359                 if (!has_references)
2360                         continue;
2361
2362                 total_cards += CARDS_PER_BLOCK;
2363                 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2364                         if (cards [i])
2365                                 ++marked_cards;
2366                 }
2367         } END_FOREACH_BLOCK_NO_LOCK;
2368
2369         *num_total_cards = total_cards;
2370         *num_marked_cards = marked_cards;
2371 }
2372
2373 static void
2374 update_cardtable_mod_union (void)
2375 {
2376         MSBlockInfo *block;
2377
2378         FOREACH_BLOCK_NO_LOCK (block) {
2379                 size_t num_cards;
2380                 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
2381                 sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2382                 SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2383         } END_FOREACH_BLOCK_NO_LOCK;
2384 }
2385
2386 #undef pthread_create
2387
2388 static void
2389 post_param_init (SgenMajorCollector *collector)
2390 {
2391         collector->sweeps_lazily = lazy_sweep;
2392         collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
2393 }
2394
2395 static void
2396 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
2397 {
2398         int i;
2399
2400         sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2401
2402         num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2403         block_obj_sizes = (int *)sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2404         ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2405
2406         evacuate_block_obj_sizes = (gboolean *)sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2407         for (i = 0; i < num_block_obj_sizes; ++i)
2408                 evacuate_block_obj_sizes [i] = FALSE;
2409
2410         sweep_slots_available = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2411         sweep_slots_used = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2412         sweep_num_blocks = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2413
2414         /*
2415         {
2416                 int i;
2417                 g_print ("block object sizes:\n");
2418                 for (i = 0; i < num_block_obj_sizes; ++i)
2419                         g_print ("%d\n", block_obj_sizes [i]);
2420         }
2421         */
2422
2423         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2424                 free_block_lists [i] = (MSBlockInfo *volatile *)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2425
2426         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2427                 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2428         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2429                 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2430
2431         mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2432         mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2433         mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2434 #if SIZEOF_VOID_P != 8
2435         mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2436         mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2437         mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2438         mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2439 #endif
2440
2441         collector->section_size = MAJOR_SECTION_SIZE;
2442
2443         concurrent_mark = is_concurrent;
2444         collector->is_concurrent = is_concurrent;
2445         collector->needs_thread_pool = is_concurrent || concurrent_sweep;
2446         collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2447         collector->supports_cardtable = TRUE;
2448
2449         collector->alloc_heap = major_alloc_heap;
2450         collector->is_object_live = major_is_object_live;
2451         collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2452         collector->alloc_degraded = major_alloc_degraded;
2453
2454         collector->alloc_object = major_alloc_object;
2455         collector->free_pinned_object = free_pinned_object;
2456         collector->iterate_objects = major_iterate_objects;
2457         collector->free_non_pinned_object = major_free_non_pinned_object;
2458         collector->pin_objects = major_pin_objects;
2459         collector->pin_major_object = pin_major_object;
2460         collector->scan_card_table = major_scan_card_table;
2461         collector->iterate_live_block_ranges = major_iterate_live_block_ranges;
2462         if (is_concurrent) {
2463                 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2464                 collector->get_cardtable_mod_union_for_reference = major_get_cardtable_mod_union_for_reference;
2465         }
2466         collector->init_to_space = major_init_to_space;
2467         collector->sweep = major_sweep;
2468         collector->have_swept = major_have_swept;
2469         collector->finish_sweeping = major_finish_sweep_checking;
2470         collector->free_swept_blocks = major_free_swept_blocks;
2471         collector->check_scan_starts = major_check_scan_starts;
2472         collector->dump_heap = major_dump_heap;
2473         collector->get_used_size = major_get_used_size;
2474         collector->start_nursery_collection = major_start_nursery_collection;
2475         collector->finish_nursery_collection = major_finish_nursery_collection;
2476         collector->start_major_collection = major_start_major_collection;
2477         collector->finish_major_collection = major_finish_major_collection;
2478         collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2479         collector->ptr_is_from_pinned_alloc = ptr_is_from_pinned_alloc;
2480         collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2481         collector->get_num_major_sections = get_num_major_sections;
2482         collector->get_bytes_survived_last_sweep = get_bytes_survived_last_sweep;
2483         collector->handle_gc_param = major_handle_gc_param;
2484         collector->print_gc_param_usage = major_print_gc_param_usage;
2485         collector->post_param_init = post_param_init;
2486         collector->is_valid_object = major_is_valid_object;
2487         collector->describe_pointer = major_describe_pointer;
2488         collector->count_cards = major_count_cards;
2489
2490         collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
2491         collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
2492         collector->major_ops_serial.drain_gray_stack = drain_gray_stack;
2493         if (is_concurrent) {
2494                 collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2495                 collector->major_ops_concurrent_start.scan_object = major_scan_object_concurrent_with_evacuation;
2496                 collector->major_ops_concurrent_start.drain_gray_stack = drain_gray_stack_concurrent;
2497
2498                 collector->major_ops_concurrent_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_finish_canonical;
2499                 collector->major_ops_concurrent_finish.scan_object = major_scan_object_with_evacuation;
2500                 collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_with_evacuation;
2501                 collector->major_ops_concurrent_finish.scan_ptr_field = major_scan_ptr_field_with_evacuation;
2502                 collector->major_ops_concurrent_finish.drain_gray_stack = drain_gray_stack;
2503         }
2504
2505 #ifdef HEAVY_STATISTICS
2506         mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2507         mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2508         mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2509         mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2510         mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2511         mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2512         mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2513         mono_counters_register ("Optimized copy major small evacuate", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_evacuate);
2514         mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2515         mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2516         mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2517
2518         mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2519         mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2520         mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2521 #endif
2522
2523 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2524         mono_os_mutex_init (&scanned_objects_list_lock);
2525 #endif
2526
2527         SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2528
2529         /*cardtable requires major pages to be 8 cards aligned*/
2530         g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2531 }
2532
2533 void
2534 sgen_marksweep_init (SgenMajorCollector *collector)
2535 {
2536         sgen_marksweep_init_internal (collector, FALSE);
2537 }
2538
2539 void
2540 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2541 {
2542         sgen_marksweep_init_internal (collector, TRUE);
2543 }
2544
2545 #endif