mono/sgen/sgen-marksweep.c

   1 /*
   2  * sgen-marksweep.c: The Mark & Sweep major collector.
   3  *
   4  * Author:
   5  *      Mark Probst <mark.probst@gmail.com>
   6  *
   7  * Copyright 2009-2010 Novell, Inc.
   8  * Copyright (C) 2012 Xamarin Inc
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Library General Public
  12  * License 2.0 as published by the Free Software Foundation;
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Library General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Library General Public
  20  * License 2.0 along with this library; if not, write to the Free
  21  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #include "config.h"
  25
  26 #ifdef HAVE_SGEN_GC
  27
  28 #include <math.h>
  29 #include <errno.h>
  30 #include <string.h>
  31 #include <stdlib.h>
  32
  33 #include "mono/sgen/sgen-gc.h"
  34 #include "mono/sgen/sgen-protocol.h"
  35 #include "mono/sgen/sgen-cardtable.h"
  36 #include "mono/sgen/sgen-memory-governor.h"
  37 #include "mono/sgen/sgen-layout-stats.h"
  38 #include "mono/sgen/sgen-pointer-queue.h"
  39 #include "mono/sgen/sgen-pinning.h"
  40 #include "mono/sgen/sgen-workers.h"
  41 #include "mono/sgen/sgen-thread-pool.h"
  42 #include "mono/sgen/sgen-client.h"
  43 #include "mono/utils/mono-membar.h"
  44
  45 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
  46 #define MS_BLOCK_SIZE   ARCH_MIN_MS_BLOCK_SIZE
  47 #define MS_BLOCK_SIZE_SHIFT     ARCH_MIN_MS_BLOCK_SIZE_SHIFT
  48 #else
  49 #define MS_BLOCK_SIZE_SHIFT     14      /* INT FASTENABLE */
  50 #define MS_BLOCK_SIZE           (1 << MS_BLOCK_SIZE_SHIFT)
  51 #endif
  52 #define MAJOR_SECTION_SIZE      MS_BLOCK_SIZE
  53 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
  54
  55 /*
  56  * Don't allocate single blocks, but alloc a contingent of this many
  57  * blocks in one swoop.  This must be a power of two.
  58  */
  59 #define MS_BLOCK_ALLOC_NUM      32
  60
  61 /*
  62  * Number of bytes before the first object in a block.  At the start
  63  * of a block is the MSBlockHeader, then opional padding, then come
  64  * the objects, so this must be >= sizeof (MSBlockHeader).
  65  */
  66 #define MS_BLOCK_SKIP   ((sizeof (MSBlockHeader) + 15) & ~15)
  67
  68 #define MS_BLOCK_FREE   (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
  69
  70 #define MS_NUM_MARK_WORDS       ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
  71
  72 /*
  73  * Blocks progress from one state to the next:
  74  *
  75  * SWEPT           The block is fully swept.  It might or might not be in
  76  *                 a free list.
  77  *
  78  * MARKING         The block might or might not contain live objects.  If
  79  *                 we're in between an initial collection pause and the
  80  *                 finishing pause, the block might or might not be in a
  81  *                 free list.
  82  *
  83  * CHECKING        The sweep thread is investigating the block to determine
  84  *                 whether or not it contains live objects.  The block is
  85  *                 not in a free list.
  86  *
  87  * NEED_SWEEPING   The block contains live objects but has not yet been
  88  *                 swept.  It also contains free slots.  It is in a block
  89  *                 free list.
  90  *
  91  * SWEEPING        The block is being swept.  It might be in a free list.
  92  */
  93
  94 enum {
  95         BLOCK_STATE_SWEPT,
  96         BLOCK_STATE_MARKING,
  97         BLOCK_STATE_CHECKING,
  98         BLOCK_STATE_NEED_SWEEPING,
  99         BLOCK_STATE_SWEEPING
 100 };
 101
 102 typedef struct _MSBlockInfo MSBlockInfo;
 103 struct _MSBlockInfo {
 104         guint16 obj_size;
 105         /*
 106          * FIXME: Do we even need this? It's only used during sweep and might be worth
 107          * recalculating to save the space.
 108          */
 109         guint16 obj_size_index;
 110         /* FIXME: Reduce this - it only needs a byte. */
 111         volatile gint32 state;
 112         unsigned int pinned : 1;
 113         unsigned int has_references : 1;
 114         unsigned int has_pinned : 1;    /* means cannot evacuate */
 115         unsigned int is_to_space : 1;
 116         void ** volatile free_list;
 117         MSBlockInfo * volatile next_free;
 118         guint8 * volatile cardtable_mod_union;
 119         mword mark_words [MS_NUM_MARK_WORDS];
 120 };
 121
 122 #define MS_BLOCK_FOR_BLOCK_INFO(b)      ((char*)(b))
 123
 124 #define MS_BLOCK_OBJ(b,i)               ((void*)(MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i)))
 125 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size)             (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
 126 #define MS_BLOCK_DATA_FOR_OBJ(o)        ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
 127
 128 typedef struct {
 129         MSBlockInfo info;
 130 } MSBlockHeader;
 131
 132 #define MS_BLOCK_FOR_OBJ(o)             (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
 133
 134 /* object index will always be small */
 135 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
 136
 137 //casting to int is fine since blocks are 32k
 138 #define MS_CALC_MARK_BIT(w,b,o)         do {                            \
 139                 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
 140                 if (sizeof (mword) == 4) {                              \
 141                         (w) = i >> 5;                                   \
 142                         (b) = i & 31;                                   \
 143                 } else {                                                \
 144                         (w) = i >> 6;                                   \
 145                         (b) = i & 63;                                   \
 146                 }                                                       \
 147         } while (0)
 148
 149 #define MS_MARK_BIT(bl,w,b)     ((bl)->mark_words [(w)] & (ONE_P << (b)))
 150 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
 151
 152 #define MS_OBJ_ALLOCED(o,b)     (*(void**)(o) && (*(char**)(o) < MS_BLOCK_FOR_BLOCK_INFO (b) || *(char**)(o) >= MS_BLOCK_FOR_BLOCK_INFO (b) + MS_BLOCK_SIZE))
 153
 154 #define MS_BLOCK_OBJ_SIZE_FACTOR        (pow (2.0, 1.0 / 3))
 155
 156 /*
 157  * This way we can lookup block object size indexes for sizes up to
 158  * 256 bytes with a single load.
 159  */
 160 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES      32
 161
 162 static int *block_obj_sizes;
 163 static int num_block_obj_sizes;
 164 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
 165
 166 #define MS_BLOCK_FLAG_PINNED    1
 167 #define MS_BLOCK_FLAG_REFS      2
 168
 169 #define MS_BLOCK_TYPE_MAX       4
 170
 171 static gboolean *evacuate_block_obj_sizes;
 172 static float evacuation_threshold = 0.666f;
 173 static float concurrent_evacuation_threshold = 0.666f;
 174 static gboolean want_evacuation = FALSE;
 175
 176 static gboolean lazy_sweep = FALSE;
 177
 178 enum {
 179         SWEEP_STATE_SWEPT,
 180         SWEEP_STATE_NEED_SWEEPING,
 181         SWEEP_STATE_SWEEPING,
 182         SWEEP_STATE_SWEEPING_AND_ITERATING,
 183         SWEEP_STATE_COMPACTING
 184 };
 185
 186 static volatile int sweep_state = SWEEP_STATE_SWEPT;
 187
 188 static gboolean concurrent_mark;
 189 static gboolean concurrent_sweep = TRUE;
 190
 191 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl)      SGEN_POINTER_IS_TAGGED_1 ((bl))
 192 #define BLOCK_TAG_HAS_REFERENCES(bl)            SGEN_POINTER_TAG_1 ((bl))
 193
 194 #define BLOCK_IS_TAGGED_CHECKING(bl)            SGEN_POINTER_IS_TAGGED_2 ((bl))
 195 #define BLOCK_TAG_CHECKING(bl)                  SGEN_POINTER_TAG_2 ((bl))
 196
 197 #define BLOCK_UNTAG(bl)                         SGEN_POINTER_UNTAG_12 ((bl))
 198
 199 #define BLOCK_TAG(bl)                           ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
 200
 201 /* all allocated blocks in the system */
 202 static SgenPointerQueue allocated_blocks;
 203
 204 /* non-allocated block free-list */
 205 static void *empty_blocks = NULL;
 206 static size_t num_empty_blocks = 0;
 207
 208 #define FOREACH_BLOCK_NO_LOCK_CONDITION(cond,bl) {                      \
 209         size_t __index;                                                 \
 210         SGEN_ASSERT (0, (cond) && !sweep_in_progress (), "Can't iterate blocks while the world is running or sweep is in progress."); \
 211         for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { \
 212                 (bl) = BLOCK_UNTAG (allocated_blocks.data [__index]);
 213 #define FOREACH_BLOCK_NO_LOCK(bl)                                       \
 214         FOREACH_BLOCK_NO_LOCK_CONDITION(sgen_is_world_stopped (), bl)
 215 #define FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK(bl,hr) {                   \
 216         size_t __index;                                                 \
 217         SGEN_ASSERT (0, sgen_is_world_stopped () && !sweep_in_progress (), "Can't iterate blocks while the world is running or sweep is in progress."); \
 218         for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { \
 219                 (bl) = allocated_blocks.data [__index];                 \
 220                 (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl));           \
 221                 (bl) = BLOCK_UNTAG ((bl));
 222 #define END_FOREACH_BLOCK_NO_LOCK       } }
 223
 224 static volatile size_t num_major_sections = 0;
 225 /*
 226  * One free block list for each block object size.  We add and remove blocks from these
 227  * lists lock-free via CAS.
 228  *
 229  * Blocks accessed/removed from `free_block_lists`:
 230  *   from the mutator (with GC lock held)
 231  *   in nursery collections
 232  *   in non-concurrent major collections
 233  *   in the finishing pause of concurrent major collections (whole list is cleared)
 234  *
 235  * Blocks added to `free_block_lists`:
 236  *   in the sweeping thread
 237  *   during nursery collections
 238  *   from domain clearing (with the world stopped and no sweeping happening)
 239  *
 240  * The only item of those that doesn't require the GC lock is the sweep thread.  The sweep
 241  * thread only ever adds blocks to the free list, so the ABA problem can't occur.
 242  */
 243 static MSBlockInfo * volatile *free_block_lists [MS_BLOCK_TYPE_MAX];
 244
 245 static guint64 stat_major_blocks_alloced = 0;
 246 static guint64 stat_major_blocks_freed = 0;
 247 static guint64 stat_major_blocks_lazy_swept = 0;
 248 static guint64 stat_major_objects_evacuated = 0;
 249
 250 #if SIZEOF_VOID_P != 8
 251 static guint64 stat_major_blocks_freed_ideal = 0;
 252 static guint64 stat_major_blocks_freed_less_ideal = 0;
 253 static guint64 stat_major_blocks_freed_individual = 0;
 254 static guint64 stat_major_blocks_alloced_less_ideal = 0;
 255 #endif
 256
 257 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
 258 static guint64 num_major_objects_marked = 0;
 259 #define INC_NUM_MAJOR_OBJECTS_MARKED()  (++num_major_objects_marked)
 260 #else
 261 #define INC_NUM_MAJOR_OBJECTS_MARKED()
 262 #endif
 263
 264 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
 265 static mono_mutex_t scanned_objects_list_lock;
 266 static SgenPointerQueue scanned_objects_list;
 267
 268 static void
 269 add_scanned_object (void *ptr)
 270 {
 271         if (!binary_protocol_is_enabled ())
 272                 return;
 273
 274         mono_mutex_lock (&scanned_objects_list_lock);
 275         sgen_pointer_queue_add (&scanned_objects_list, ptr);
 276         mono_mutex_unlock (&scanned_objects_list_lock);
 277 }
 278 #endif
 279
 280 static gboolean sweep_block (MSBlockInfo *block);
 281
 282 static int
 283 ms_find_block_obj_size_index (size_t size)
 284 {
 285         int i;
 286         SGEN_ASSERT (9, size <= SGEN_MAX_SMALL_OBJ_SIZE, "size %zd is bigger than max small object size %d", size, SGEN_MAX_SMALL_OBJ_SIZE);
 287         for (i = 0; i < num_block_obj_sizes; ++i)
 288                 if (block_obj_sizes [i] >= size)
 289                         return i;
 290         g_error ("no object of size %zd\n", size);
 291         return -1;
 292 }
 293
 294 #define FREE_BLOCKS_FROM(lists,p,r)     (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
 295 #define FREE_BLOCKS(p,r)                (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
 296
 297 #define MS_BLOCK_OBJ_SIZE_INDEX(s)                              \
 298         (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ?      \
 299          fast_block_obj_size_indexes [((s)+7)>>3] :             \
 300          ms_find_block_obj_size_index ((s)))
 301
 302 static void*
 303 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
 304 {
 305         char *start;
 306         if (nursery_align)
 307                 start = sgen_alloc_os_memory_aligned (nursery_size, nursery_align, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
 308         else
 309                 start = sgen_alloc_os_memory (nursery_size, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
 310
 311         return start;
 312 }
 313
 314 static void
 315 update_heap_boundaries_for_block (MSBlockInfo *block)
 316 {
 317         sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
 318 }
 319
 320 /*
 321  * Thread safe
 322  */
 323 static void*
 324 ms_get_empty_block (void)
 325 {
 326         char *p;
 327         int i;
 328         void *block, *empty, *next;
 329
 330  retry:
 331         if (!empty_blocks) {
 332                 /*
 333                  * We try allocating MS_BLOCK_ALLOC_NUM blocks first.  If that's
 334                  * unsuccessful, we halve the number of blocks and try again, until we're at
 335                  * 1.  If that doesn't work, either, we assert.
 336                  */
 337                 int alloc_num = MS_BLOCK_ALLOC_NUM;
 338                 for (;;) {
 339                         p = sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE,
 340                                         alloc_num == 1 ? "major heap section" : NULL);
 341                         if (p)
 342                                 break;
 343                         alloc_num >>= 1;
 344                 }
 345
 346                 for (i = 0; i < alloc_num; ++i) {
 347                         block = p;
 348                         /*
 349                          * We do the free list update one after the
 350                          * other so that other threads can use the new
 351                          * blocks as quickly as possible.
 352                          */
 353                         do {
 354                                 empty = empty_blocks;
 355                                 *(void**)block = empty;
 356                         } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
 357                         p += MS_BLOCK_SIZE;
 358                 }
 359
 360                 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
 361
 362                 stat_major_blocks_alloced += alloc_num;
 363 #if SIZEOF_VOID_P != 8
 364                 if (alloc_num != MS_BLOCK_ALLOC_NUM)
 365                         stat_major_blocks_alloced_less_ideal += alloc_num;
 366 #endif
 367         }
 368
 369         do {
 370                 empty = empty_blocks;
 371                 if (!empty)
 372                         goto retry;
 373                 block = empty;
 374                 next = *(void**)block;
 375         } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
 376
 377         SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
 378
 379         *(void**)block = NULL;
 380
 381         g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
 382
 383         return block;
 384 }
 385
 386 /*
 387  * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
 388  * list, where it will either be freed later on, or reused in nursery collections.
 389  */
 390 static void
 391 ms_free_block (void *block)
 392 {
 393         void *empty;
 394
 395         sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
 396         memset (block, 0, MS_BLOCK_SIZE);
 397
 398         do {
 399                 empty = empty_blocks;
 400                 *(void**)block = empty;
 401         } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
 402
 403         SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
 404
 405         binary_protocol_block_free (block, MS_BLOCK_SIZE);
 406 }
 407
 408 static gboolean
 409 sweep_in_progress (void)
 410 {
 411         int state = sweep_state;
 412         return state == SWEEP_STATE_SWEEPING ||
 413                 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
 414                 state == SWEEP_STATE_COMPACTING;
 415 }
 416
 417 static inline gboolean
 418 block_is_swept_or_marking (MSBlockInfo *block)
 419 {
 420         gint32 state = block->state;
 421         return state == BLOCK_STATE_SWEPT || state == BLOCK_STATE_MARKING;
 422 }
 423
 424 //#define MARKSWEEP_CONSISTENCY_CHECK
 425
 426 #ifdef MARKSWEEP_CONSISTENCY_CHECK
 427 static void
 428 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
 429 {
 430         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 431         for (; block; block = block->next_free) {
 432                 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
 433                 g_assert (block->obj_size == size);
 434                 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
 435
 436                 /* blocks in the free lists must have at least
 437                    one free slot */
 438                 g_assert (block->free_list);
 439
 440                 /* the block must be in the allocated_blocks array */
 441                 g_assert (sgen_pointer_queue_find (&allocated_blocks, BLOCK_TAG (block)) != (size_t)-1);
 442         }
 443 }
 444
 445 static void
 446 check_empty_blocks (void)
 447 {
 448         void *p;
 449         size_t i = 0;
 450         for (p = empty_blocks; p; p = *(void**)p)
 451                 ++i;
 452         g_assert (i == num_empty_blocks);
 453 }
 454
 455 static void
 456 consistency_check (void)
 457 {
 458         MSBlockInfo *block;
 459         int i;
 460
 461         /* check all blocks */
 462         FOREACH_BLOCK_NO_LOCK (block) {
 463                 int count = MS_BLOCK_FREE / block->obj_size;
 464                 int num_free = 0;
 465                 void **free;
 466
 467                 /* count number of free slots */
 468                 for (i = 0; i < count; ++i) {
 469                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 470                         if (!MS_OBJ_ALLOCED (obj, block))
 471                                 ++num_free;
 472                 }
 473
 474                 /* check free list */
 475                 for (free = block->free_list; free; free = (void**)*free) {
 476                         g_assert (MS_BLOCK_FOR_OBJ (free) == block);
 477                         --num_free;
 478                 }
 479                 g_assert (num_free == 0);
 480
 481                 /* check all mark words are zero */
 482                 if (!sgen_concurrent_collection_in_progress () && block_is_swept_or_marking (block)) {
 483                         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
 484                                 g_assert (block->mark_words [i] == 0);
 485                 }
 486         } END_FOREACH_BLOCK_NO_LOCK;
 487
 488         /* check free blocks */
 489         for (i = 0; i < num_block_obj_sizes; ++i) {
 490                 int j;
 491                 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
 492                         check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
 493         }
 494
 495         check_empty_blocks ();
 496 }
 497 #endif
 498
 499 static void
 500 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
 501 {
 502         MSBlockInfo *old;
 503         do {
 504                 block->next_free = old = free_blocks [size_index];
 505         } while (SGEN_CAS_PTR ((gpointer)&free_blocks [size_index], block, old) != old);
 506 }
 507
 508 static void major_finish_sweep_checking (void);
 509
 510 static gboolean
 511 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
 512 {
 513         int size = block_obj_sizes [size_index];
 514         int count = MS_BLOCK_FREE / size;
 515         MSBlockInfo *info;
 516         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 517         char *obj_start;
 518         int i;
 519
 520         if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
 521                 return FALSE;
 522
 523         info = (MSBlockInfo*)ms_get_empty_block ();
 524
 525         SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
 526
 527         info->obj_size = size;
 528         info->obj_size_index = size_index;
 529         info->pinned = pinned;
 530         info->has_references = has_references;
 531         info->has_pinned = pinned;
 532         /*
 533          * Blocks that are to-space are not evacuated from.  During an major collection
 534          * blocks are allocated for two reasons: evacuating objects from the nursery and
 535          * evacuating them from major blocks marked for evacuation.  In both cases we don't
 536          * want further evacuation.
 537          */
 538         info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD);
 539         info->state = (info->is_to_space || sgen_concurrent_collection_in_progress ()) ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
 540         SGEN_ASSERT (6, !sweep_in_progress () || info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
 541         info->cardtable_mod_union = NULL;
 542
 543         update_heap_boundaries_for_block (info);
 544
 545         binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
 546
 547         /* build free list */
 548         obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
 549         info->free_list = (void**)obj_start;
 550         /* we're skipping the last one - it must be nulled */
 551         for (i = 0; i < count - 1; ++i) {
 552                 char *next_obj_start = obj_start + size;
 553                 *(void**)obj_start = next_obj_start;
 554                 obj_start = next_obj_start;
 555         }
 556         /* the last one */
 557         *(void**)obj_start = NULL;
 558
 559         add_free_block (free_blocks, size_index, info);
 560
 561         /*
 562          * This is the only place where the `allocated_blocks` array can potentially grow.
 563          * We need to make sure concurrent sweep isn't running when that happens, so in that
 564          * specific case we just wait for sweep to finish.
 565          */
 566         if (sgen_pointer_queue_will_grow (&allocated_blocks))
 567                 major_finish_sweep_checking ();
 568
 569         sgen_pointer_queue_add (&allocated_blocks, BLOCK_TAG (info));
 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 = next_free_slot;
 632                 return obj;
 633         }
 634
 635         next_free_block = block->next_free;
 636         if (SGEN_CAS_PTR ((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 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 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         major_finish_sweep_checking ();
 745
 746         obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
 747         if (G_LIKELY (obj)) {
 748                 HEAVY_STAT (++stat_objects_alloced_degraded);
 749                 HEAVY_STAT (stat_bytes_alloced_degraded += size);
 750         }
 751         return obj;
 752 }
 753
 754 /*
 755  * obj is some object.  If it's not in the major heap (i.e. if it's in
 756  * the nursery or LOS), return FALSE.  Otherwise return whether it's
 757  * been marked or copied.
 758  */
 759 static gboolean
 760 major_is_object_live (GCObject *obj)
 761 {
 762         MSBlockInfo *block;
 763         int word, bit;
 764         mword objsize;
 765
 766         if (sgen_ptr_in_nursery (obj))
 767                 return FALSE;
 768
 769         objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
 770
 771         /* LOS */
 772         if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
 773                 return FALSE;
 774
 775         /* now we know it's in a major block */
 776         block = MS_BLOCK_FOR_OBJ (obj);
 777         SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?", block);
 778         MS_CALC_MARK_BIT (word, bit, obj);
 779         return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
 780 }
 781
 782 static gboolean
 783 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
 784 {
 785         MSBlockInfo *block;
 786
 787         FOREACH_BLOCK_NO_LOCK (block) {
 788                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
 789                         int count = MS_BLOCK_FREE / block->obj_size;
 790                         int i;
 791
 792                         *start = NULL;
 793                         for (i = 0; i <= count; ++i) {
 794                                 if (ptr >= (char*)MS_BLOCK_OBJ (block, i) && ptr < (char*)MS_BLOCK_OBJ (block, i + 1)) {
 795                                         *start = MS_BLOCK_OBJ (block, i);
 796                                         break;
 797                                 }
 798                         }
 799                         return !block->pinned;
 800                 }
 801         } END_FOREACH_BLOCK_NO_LOCK;
 802         return FALSE;
 803 }
 804
 805 static gboolean
 806 try_set_sweep_state (int new, int expected)
 807 {
 808         int old = SGEN_CAS (&sweep_state, new, expected);
 809         return old == expected;
 810 }
 811
 812 static void
 813 set_sweep_state (int new, int expected)
 814 {
 815         gboolean success = try_set_sweep_state (new, expected);
 816         SGEN_ASSERT (0, success, "Could not set sweep state.");
 817 }
 818
 819 static gboolean ensure_block_is_checked_for_sweeping (int block_index, gboolean wait, gboolean *have_checked);
 820
 821 static SgenThreadPoolJob * volatile sweep_job;
 822
 823 static void
 824 major_finish_sweep_checking (void)
 825 {
 826         int block_index;
 827         SgenThreadPoolJob *job;
 828
 829  retry:
 830         switch (sweep_state) {
 831         case SWEEP_STATE_SWEPT:
 832         case SWEEP_STATE_NEED_SWEEPING:
 833                 return;
 834         case SWEEP_STATE_SWEEPING:
 835                 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
 836                         break;
 837                 goto retry;
 838         case SWEEP_STATE_SWEEPING_AND_ITERATING:
 839                 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
 840                 goto retry;
 841         case SWEEP_STATE_COMPACTING:
 842                 goto wait;
 843         default:
 844                 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
 845                 break;
 846         }
 847
 848         /*
 849          * We're running with the world stopped and the only other thread doing work is the
 850          * sweep thread, which doesn't add blocks to the array, so we can safely access
 851          * `next_slot`.
 852          */
 853         for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index)
 854                 ensure_block_is_checked_for_sweeping (block_index, FALSE, NULL);
 855
 856         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
 857
 858  wait:
 859         job = sweep_job;
 860         if (job)
 861                 sgen_thread_pool_job_wait (job);
 862         SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
 863         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
 864 }
 865
 866 static void
 867 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
 868 {
 869         gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
 870         gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
 871         gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
 872         MSBlockInfo *block;
 873
 874         major_finish_sweep_checking ();
 875         FOREACH_BLOCK_NO_LOCK (block) {
 876                 int count = MS_BLOCK_FREE / block->obj_size;
 877                 int i;
 878
 879                 if (block->pinned && !pinned)
 880                         continue;
 881                 if (!block->pinned && !non_pinned)
 882                         continue;
 883                 if (sweep && lazy_sweep) {
 884                         sweep_block (block);
 885                         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
 886                 }
 887
 888                 for (i = 0; i < count; ++i) {
 889                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 890                         /*
 891                          * We've finished sweep checking, but if we're sweeping lazily and
 892                          * the flags don't require us to sweep, the block might still need
 893                          * sweeping.  In that case, we need to consult the mark bits to tell
 894                          * us whether an object slot is live.
 895                          */
 896                         if (!block_is_swept_or_marking (block)) {
 897                                 int word, bit;
 898                                 SGEN_ASSERT (6, !sweep && block->state == BLOCK_STATE_NEED_SWEEPING, "Has sweeping not finished?");
 899                                 MS_CALC_MARK_BIT (word, bit, obj);
 900                                 if (!MS_MARK_BIT (block, word, bit))
 901                                         continue;
 902                         }
 903                         if (MS_OBJ_ALLOCED (obj, block))
 904                                 callback ((GCObject*)obj, block->obj_size, data);
 905                 }
 906         } END_FOREACH_BLOCK_NO_LOCK;
 907 }
 908
 909 static gboolean
 910 major_is_valid_object (char *object)
 911 {
 912         MSBlockInfo *block;
 913
 914         FOREACH_BLOCK_NO_LOCK (block) {
 915                 int idx;
 916                 char *obj;
 917
 918                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
 919                         continue;
 920
 921                 idx = MS_BLOCK_OBJ_INDEX (object, block);
 922                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 923                 if (obj != object)
 924                         return FALSE;
 925                 return MS_OBJ_ALLOCED (obj, block);
 926         } END_FOREACH_BLOCK_NO_LOCK;
 927
 928         return FALSE;
 929 }
 930
 931
 932 static GCVTable
 933 major_describe_pointer (char *ptr)
 934 {
 935         MSBlockInfo *block;
 936
 937         FOREACH_BLOCK_NO_LOCK (block) {
 938                 int idx;
 939                 char *obj;
 940                 gboolean live;
 941                 GCVTable vtable;
 942                 int w, b;
 943                 gboolean marked;
 944
 945                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
 946                         continue;
 947
 948                 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
 949                         MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
 950
 951                 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
 952                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 953                 live = MS_OBJ_ALLOCED (obj, block);
 954                 vtable = live ? SGEN_LOAD_VTABLE ((GCObject*)obj) : NULL;
 955
 956                 MS_CALC_MARK_BIT (w, b, obj);
 957                 marked = MS_MARK_BIT (block, w, b);
 958
 959                 if (obj == ptr) {
 960                         SGEN_LOG (0, "\t(");
 961                         if (live)
 962                                 SGEN_LOG (0, "object");
 963                         else
 964                                 SGEN_LOG (0, "dead-object");
 965                 } else {
 966                         if (live)
 967                                 SGEN_LOG (0, "interior-ptr offset %zd", ptr - obj);
 968                         else
 969                                 SGEN_LOG (0, "dead-interior-ptr offset %zd", ptr - obj);
 970                 }
 971
 972                 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
 973
 974                 return vtable;
 975         } END_FOREACH_BLOCK_NO_LOCK;
 976
 977         return NULL;
 978 }
 979
 980 static void
 981 major_check_scan_starts (void)
 982 {
 983 }
 984
 985 static void
 986 major_dump_heap (FILE *heap_dump_file)
 987 {
 988         MSBlockInfo *block;
 989         int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
 990         int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
 991         int i;
 992
 993         for (i = 0; i < num_block_obj_sizes; ++i)
 994                 slots_available [i] = slots_used [i] = 0;
 995
 996         FOREACH_BLOCK_NO_LOCK (block) {
 997                 int index = ms_find_block_obj_size_index (block->obj_size);
 998                 int count = MS_BLOCK_FREE / block->obj_size;
 999
1000                 slots_available [index] += count;
1001                 for (i = 0; i < count; ++i) {
1002                         if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
1003                                 ++slots_used [index];
1004                 }
1005         } END_FOREACH_BLOCK_NO_LOCK;
1006
1007         fprintf (heap_dump_file, "<occupancies>\n");
1008         for (i = 0; i < num_block_obj_sizes; ++i) {
1009                 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
1010                                 block_obj_sizes [i], slots_available [i], slots_used [i]);
1011         }
1012         fprintf (heap_dump_file, "</occupancies>\n");
1013
1014         FOREACH_BLOCK_NO_LOCK (block) {
1015                 int count = MS_BLOCK_FREE / block->obj_size;
1016                 int i;
1017                 int start = -1;
1018
1019                 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
1020
1021                 for (i = 0; i <= count; ++i) {
1022                         if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
1023                                 if (start < 0)
1024                                         start = i;
1025                         } else {
1026                                 if (start >= 0) {
1027                                         sgen_dump_occupied (MS_BLOCK_OBJ (block, start), MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
1028                                         start = -1;
1029                                 }
1030                         }
1031                 }
1032
1033                 fprintf (heap_dump_file, "</section>\n");
1034         } END_FOREACH_BLOCK_NO_LOCK;
1035 }
1036
1037 static guint8*
1038 get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate)
1039 {
1040         guint8 *mod_union = block->cardtable_mod_union;
1041         guint8 *other;
1042         if (mod_union)
1043                 return mod_union;
1044         else if (!allocate)
1045                 return NULL;
1046         mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1047         other = SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
1048         if (!other) {
1049                 SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
1050                 return mod_union;
1051         }
1052         sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1053         return other;
1054 }
1055
1056 static inline guint8*
1057 major_get_cardtable_mod_union_for_reference (char *ptr)
1058 {
1059         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (ptr);
1060         size_t offset = sgen_card_table_get_card_offset (ptr, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1061         guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
1062         SGEN_ASSERT (0, mod_union, "FIXME: optionally allocate the mod union if it's not here and CAS it in.");
1063         return &mod_union [offset];
1064 }
1065
1066 /*
1067  * Mark the mod-union card for `ptr`, which must be a reference within the object `obj`.
1068  */
1069 static void
1070 mark_mod_union_card (GCObject *obj, void **ptr)
1071 {
1072         int type = sgen_obj_get_descriptor (obj) & DESC_TYPE_MASK;
1073         if (sgen_safe_object_is_small (obj, type)) {
1074                 guint8 *card_byte = major_get_cardtable_mod_union_for_reference ((char*)ptr);
1075                 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?");
1076                 *card_byte = 1;
1077         } else {
1078                 sgen_los_mark_mod_union_card (obj, ptr);
1079         }
1080 }
1081
1082 #define LOAD_VTABLE     SGEN_LOAD_VTABLE
1083
1084 #define MS_MARK_OBJECT_AND_ENQUEUE_CHECKED(obj,desc,block,queue) do {   \
1085                 int __word, __bit;                                      \
1086                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
1087                 if (!MS_MARK_BIT ((block), __word, __bit) && MS_OBJ_ALLOCED ((obj), (block))) { \
1088                         MS_SET_MARK_BIT ((block), __word, __bit);       \
1089                         if (sgen_gc_descr_has_references (desc))                        \
1090                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1091                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1092                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
1093                 }                                                       \
1094         } while (0)
1095 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do {           \
1096                 int __word, __bit;                                      \
1097                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
1098                 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
1099                 if (!MS_MARK_BIT ((block), __word, __bit)) {            \
1100                         MS_SET_MARK_BIT ((block), __word, __bit);       \
1101                         if (sgen_gc_descr_has_references (desc))                        \
1102                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1103                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1104                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
1105                 }                                                       \
1106         } while (0)
1107
1108 static void
1109 pin_major_object (GCObject *obj, SgenGrayQueue *queue)
1110 {
1111         MSBlockInfo *block;
1112
1113         if (concurrent_mark)
1114                 g_assert_not_reached ();
1115
1116         block = MS_BLOCK_FOR_OBJ (obj);
1117         block->has_pinned = TRUE;
1118         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1119 }
1120
1121 #include "sgen-major-copy-object.h"
1122
1123 static void
1124 major_copy_or_mark_object_concurrent (GCObject **ptr, GCObject *obj, SgenGrayQueue *queue)
1125 {
1126         SGEN_ASSERT (9, sgen_concurrent_collection_in_progress (), "Why are we scanning concurrently when there's no concurrent collection on?");
1127         SGEN_ASSERT (9, !sgen_workers_are_working () || sgen_thread_pool_is_thread_pool_thread (mono_native_thread_id_get ()), "We must not scan from two threads at the same time!");
1128
1129         g_assert (!SGEN_OBJECT_IS_FORWARDED (obj));
1130
1131         if (!sgen_ptr_in_nursery (obj)) {
1132                 mword objsize;
1133
1134                 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
1135
1136                 if (objsize <= SGEN_MAX_SMALL_OBJ_SIZE) {
1137                         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
1138                         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1139                 } else {
1140                         if (sgen_los_object_is_pinned (obj))
1141                                 return;
1142
1143                         binary_protocol_mark (obj, SGEN_LOAD_VTABLE (obj), sgen_safe_object_get_size (obj));
1144
1145                         sgen_los_pin_object (obj);
1146                         if (SGEN_OBJECT_HAS_REFERENCES (obj))
1147                                 GRAY_OBJECT_ENQUEUE (queue, obj, sgen_obj_get_descriptor (obj));
1148                         INC_NUM_MAJOR_OBJECTS_MARKED ();
1149                 }
1150         }
1151 }
1152
1153 static long long
1154 major_get_and_reset_num_major_objects_marked (void)
1155 {
1156 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1157         long long num = num_major_objects_marked;
1158         num_major_objects_marked = 0;
1159         return num;
1160 #else
1161         return 0;
1162 #endif
1163 }
1164
1165 #define PREFETCH_CARDS          1       /* BOOL FASTENABLE */
1166 #if !PREFETCH_CARDS
1167 #undef PREFETCH_CARDS
1168 #endif
1169
1170 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1171 #if defined(PLATFORM_MACOSX)
1172 #define GCC_VERSION (__GNUC__ * 10000 \
1173                                + __GNUC_MINOR__ * 100 \
1174                                + __GNUC_PATCHLEVEL__)
1175 #if GCC_VERSION <= 40300
1176 #undef PREFETCH_CARDS
1177 #endif
1178 #endif
1179
1180 #ifdef HEAVY_STATISTICS
1181 static guint64 stat_optimized_copy;
1182 static guint64 stat_optimized_copy_nursery;
1183 static guint64 stat_optimized_copy_nursery_forwarded;
1184 static guint64 stat_optimized_copy_nursery_pinned;
1185 static guint64 stat_optimized_copy_major;
1186 static guint64 stat_optimized_copy_major_small_fast;
1187 static guint64 stat_optimized_copy_major_small_slow;
1188 static guint64 stat_optimized_copy_major_large;
1189 static guint64 stat_optimized_copy_major_forwarded;
1190 static guint64 stat_optimized_copy_major_small_evacuate;
1191 static guint64 stat_optimized_major_scan;
1192 static guint64 stat_optimized_major_scan_no_refs;
1193
1194 static guint64 stat_drain_prefetch_fills;
1195 static guint64 stat_drain_prefetch_fill_failures;
1196 static guint64 stat_drain_loops;
1197 #endif
1198
1199 static void major_scan_object_with_evacuation (GCObject *start, mword desc, SgenGrayQueue *queue);
1200
1201 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_no_evacuation
1202 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_no_evacuation
1203 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_no_evacuation
1204 #include "sgen-marksweep-drain-gray-stack.h"
1205
1206 #define COPY_OR_MARK_WITH_EVACUATION
1207 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_with_evacuation
1208 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_with_evacuation
1209 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_with_evacuation
1210 #include "sgen-marksweep-drain-gray-stack.h"
1211
1212 static gboolean
1213 drain_gray_stack (ScanCopyContext ctx)
1214 {
1215         gboolean evacuation = FALSE;
1216         int i;
1217         for (i = 0; i < num_block_obj_sizes; ++i) {
1218                 if (evacuate_block_obj_sizes [i]) {
1219                         evacuation = TRUE;
1220                         break;
1221                 }
1222         }
1223
1224         if (evacuation)
1225                 return drain_gray_stack_with_evacuation (ctx);
1226         else
1227                 return drain_gray_stack_no_evacuation (ctx);
1228 }
1229
1230 #include "sgen-marksweep-scan-object-concurrent.h"
1231
1232 static void
1233 major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue)
1234 {
1235         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1236 }
1237
1238 static void
1239 major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *queue)
1240 {
1241         major_copy_or_mark_object_concurrent (ptr, *ptr, queue);
1242 }
1243
1244 static void
1245 major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue)
1246 {
1247         major_copy_or_mark_object_no_evacuation (ptr, *ptr, queue);
1248 }
1249
1250 static void
1251 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1252 {
1253         void **entry, **end;
1254         int last_index = -1;
1255
1256         if (first_entry == last_entry)
1257                 return;
1258
1259         block->has_pinned = TRUE;
1260
1261         entry = sgen_pinning_get_entry (first_entry);
1262         end = sgen_pinning_get_entry (last_entry);
1263
1264         for (; entry < end; ++entry) {
1265                 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1266                 GCObject *obj;
1267                 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));
1268                 if (index == last_index)
1269                         continue;
1270                 obj = MS_BLOCK_OBJ (block, index);
1271                 MS_MARK_OBJECT_AND_ENQUEUE_CHECKED (obj, sgen_obj_get_descriptor (obj), block, queue);
1272                 last_index = index;
1273         }
1274 }
1275
1276 static inline void
1277 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1278 {
1279         int obj_index;
1280
1281         for (obj_index = 0; obj_index < count; ++obj_index) {
1282                 int word, bit;
1283                 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1284
1285                 MS_CALC_MARK_BIT (word, bit, obj);
1286                 if (MS_MARK_BIT (block, word, bit)) {
1287                         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1288                 } else {
1289                         /* an unmarked object */
1290                         if (MS_OBJ_ALLOCED (obj, block)) {
1291                                 /*
1292                                  * FIXME: Merge consecutive
1293                                  * slots for lower reporting
1294                                  * overhead.  Maybe memset
1295                                  * will also benefit?
1296                                  */
1297                                 binary_protocol_empty (obj, obj_size);
1298                                 memset (obj, 0, obj_size);
1299                         }
1300                         *(void**)obj = block->free_list;
1301                         block->free_list = obj;
1302                 }
1303         }
1304 }
1305
1306 static inline gboolean
1307 try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1308 {
1309         gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
1310         gboolean success = old_state == expected_state;
1311         if (success)
1312                 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
1313         return success;
1314 }
1315
1316 static inline void
1317 set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1318 {
1319         SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set");
1320         block->state = new_state;
1321 }
1322
1323 /*
1324  * If `block` needs sweeping, sweep it and return TRUE.  Otherwise return FALSE.
1325  *
1326  * Sweeping means iterating through the block's slots and building the free-list from the
1327  * unmarked ones.  They will also be zeroed.  The mark bits will be reset.
1328  */
1329 static gboolean
1330 sweep_block (MSBlockInfo *block)
1331 {
1332         int count;
1333         void *reversed = NULL;
1334
1335  retry:
1336         switch (block->state) {
1337         case BLOCK_STATE_SWEPT:
1338                 return FALSE;
1339         case BLOCK_STATE_MARKING:
1340         case BLOCK_STATE_CHECKING:
1341                 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1342                 goto retry;
1343         case BLOCK_STATE_SWEEPING:
1344                 /* FIXME: Do this more elegantly */
1345                 g_usleep (100);
1346                 goto retry;
1347         case BLOCK_STATE_NEED_SWEEPING:
1348                 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1349                         goto retry;
1350                 break;
1351         default:
1352                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1353         }
1354
1355         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1356
1357         count = MS_BLOCK_FREE / block->obj_size;
1358
1359         block->free_list = NULL;
1360
1361         /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1362         // FIXME: Add more sizes
1363         switch (block->obj_size) {
1364         case 16:
1365                 sweep_block_for_size (block, count, 16);
1366                 break;
1367         default:
1368                 sweep_block_for_size (block, count, block->obj_size);
1369                 break;
1370         }
1371
1372         /* reset mark bits */
1373         memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1374
1375         /* Reverse free list so that it's in address order */
1376         reversed = NULL;
1377         while (block->free_list) {
1378                 void *next = *(void**)block->free_list;
1379                 *(void**)block->free_list = reversed;
1380                 reversed = block->free_list;
1381                 block->free_list = next;
1382         }
1383         block->free_list = reversed;
1384
1385         mono_memory_write_barrier ();
1386
1387         set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1388
1389         return TRUE;
1390 }
1391
1392 static inline int
1393 bitcount (mword d)
1394 {
1395         int count = 0;
1396
1397 #ifdef __GNUC__
1398         if (sizeof (mword) == sizeof (unsigned long))
1399                 count += __builtin_popcountl (d);
1400         else
1401                 count += __builtin_popcount (d);
1402 #else
1403         while (d) {
1404                 count ++;
1405                 d &= (d - 1);
1406         }
1407 #endif
1408         return count;
1409 }
1410
1411 /* statistics for evacuation */
1412 static size_t *sweep_slots_available;
1413 static size_t *sweep_slots_used;
1414 static size_t *sweep_num_blocks;
1415
1416 static volatile size_t num_major_sections_before_sweep;
1417 static volatile size_t num_major_sections_freed_in_sweep;
1418
1419 static void
1420 sweep_start (void)
1421 {
1422         int i;
1423
1424         for (i = 0; i < num_block_obj_sizes; ++i)
1425                 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1426
1427         /* clear all the free lists */
1428         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1429                 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1430                 int j;
1431                 for (j = 0; j < num_block_obj_sizes; ++j)
1432                         free_blocks [j] = NULL;
1433         }
1434 }
1435
1436 static void sweep_finish (void);
1437
1438 /*
1439  * If `wait` is TRUE and the block is currently being checked, this function will wait until
1440  * the checking has finished.
1441  *
1442  * Returns whether the block is still there.  If `wait` is FALSE, the return value will not
1443  * be correct, i.e. must not be used.
1444  */
1445 static gboolean
1446 ensure_block_is_checked_for_sweeping (int block_index, gboolean wait, gboolean *have_checked)
1447 {
1448         int count;
1449         gboolean have_live = FALSE;
1450         gboolean have_free = FALSE;
1451         int nused = 0;
1452         int block_state;
1453         int i;
1454         void *tagged_block;
1455         MSBlockInfo *block;
1456
1457         SGEN_ASSERT (6, sweep_in_progress (), "Why do we call this function if there's no sweep in progress?");
1458
1459         if (have_checked)
1460                 *have_checked = FALSE;
1461
1462  retry:
1463         tagged_block = *(void * volatile *)&allocated_blocks.data [block_index];
1464         if (!tagged_block)
1465                 return FALSE;
1466
1467         if (BLOCK_IS_TAGGED_CHECKING (tagged_block)) {
1468                 if (!wait)
1469                         return FALSE;
1470                 /* FIXME: do this more elegantly */
1471                 g_usleep (100);
1472                 goto retry;
1473         }
1474
1475         if (SGEN_CAS_PTR (&allocated_blocks.data [block_index], BLOCK_TAG_CHECKING (tagged_block), tagged_block) != tagged_block)
1476                 goto retry;
1477
1478         block = BLOCK_UNTAG (tagged_block);
1479         block_state = block->state;
1480
1481         if (!sweep_in_progress ()) {
1482                 SGEN_ASSERT (6, block_state != BLOCK_STATE_SWEEPING && block_state != BLOCK_STATE_CHECKING, "Invalid block state.");
1483                 if (!lazy_sweep)
1484                         SGEN_ASSERT (6, block_state != BLOCK_STATE_NEED_SWEEPING, "Invalid block state.");
1485         }
1486
1487         switch (block_state) {
1488         case BLOCK_STATE_SWEPT:
1489         case BLOCK_STATE_NEED_SWEEPING:
1490         case BLOCK_STATE_SWEEPING:
1491                 goto done;
1492         case BLOCK_STATE_MARKING:
1493                 break;
1494         case BLOCK_STATE_CHECKING:
1495                 SGEN_ASSERT (0, FALSE, "We set the CHECKING bit - how can the stage be CHECKING?");
1496                 goto done;
1497         default:
1498                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1499                 break;
1500         }
1501
1502         SGEN_ASSERT (6, block->state == BLOCK_STATE_MARKING, "When we sweep all blocks must start out marking.");
1503         set_block_state (block, BLOCK_STATE_CHECKING, BLOCK_STATE_MARKING);
1504
1505         if (have_checked)
1506                 *have_checked = TRUE;
1507
1508         block->has_pinned = block->pinned;
1509
1510         block->is_to_space = FALSE;
1511
1512         count = MS_BLOCK_FREE / block->obj_size;
1513
1514         if (block->cardtable_mod_union) {
1515                 sgen_card_table_free_mod_union (block->cardtable_mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1516                 block->cardtable_mod_union = NULL;
1517         }
1518
1519         /* Count marked objects in the block */
1520         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1521                 nused += bitcount (block->mark_words [i]);
1522
1523         if (nused)
1524                 have_live = TRUE;
1525         if (nused < count)
1526                 have_free = TRUE;
1527
1528         if (have_live) {
1529                 int obj_size_index = block->obj_size_index;
1530                 gboolean has_pinned = block->has_pinned;
1531
1532                 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1533
1534                 /*
1535                  * FIXME: Go straight to SWEPT if there are no free slots.  We need
1536                  * to set the free slot list to NULL, though, and maybe update some
1537                  * statistics.
1538                  */
1539                 if (!lazy_sweep)
1540                         sweep_block (block);
1541
1542                 if (!has_pinned) {
1543                         ++sweep_num_blocks [obj_size_index];
1544                         sweep_slots_used [obj_size_index] += nused;
1545                         sweep_slots_available [obj_size_index] += count;
1546                 }
1547
1548                 /*
1549                  * If there are free slots in the block, add
1550                  * the block to the corresponding free list.
1551                  */
1552                 if (have_free) {
1553                         MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1554
1555                         if (!lazy_sweep)
1556                                 SGEN_ASSERT (6, block->free_list, "How do we not have a free list when there are free slots?");
1557
1558                         add_free_block (free_blocks, obj_size_index, block);
1559                 }
1560
1561                 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1562                 update_heap_boundaries_for_block (block);
1563         } else {
1564                 /*
1565                  * Blocks without live objects are removed from the
1566                  * block list and freed.
1567                  */
1568                 SGEN_ASSERT (6, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1569                 SGEN_ASSERT (6, allocated_blocks.data [block_index] == BLOCK_TAG_CHECKING (tagged_block), "How did the block move?");
1570
1571                 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1572                 ms_free_block (block);
1573
1574                 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1575
1576                 tagged_block = NULL;
1577         }
1578
1579  done:
1580         allocated_blocks.data [block_index] = tagged_block;
1581         return !!tagged_block;
1582 }
1583
1584 static void
1585 sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1586 {
1587         int block_index;
1588         int num_blocks = num_major_sections_before_sweep;
1589
1590         SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1591         SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1592
1593         /*
1594          * We traverse the block array from high to low.  Nursery collections will have to
1595          * cooperate with the sweep thread to finish sweeping, and they will traverse from
1596          * low to high, to avoid constantly colliding on the same blocks.
1597          */
1598         for (block_index = num_blocks - 1; block_index >= 0; --block_index) {
1599                 gboolean have_checked;
1600
1601                 /*
1602                  * The block might have been freed by another thread doing some checking
1603                  * work.
1604                  */
1605                 if (!ensure_block_is_checked_for_sweeping (block_index, TRUE, &have_checked))
1606                         ++num_major_sections_freed_in_sweep;
1607         }
1608
1609         while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) {
1610                 /*
1611                  * The main GC thread is currently iterating over the block array to help us
1612                  * finish the sweep.  We have already finished, but we don't want to mess up
1613                  * that iteration, so we just wait for it.
1614                  */
1615                 g_usleep (100);
1616         }
1617
1618         if (SGEN_MAX_ASSERT_LEVEL >= 6) {
1619                 for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1620                         MSBlockInfo *block = BLOCK_UNTAG (allocated_blocks.data [block_index]);
1621                         SGEN_ASSERT (6, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1622                 }
1623         }
1624
1625         sgen_pointer_queue_remove_nulls (&allocated_blocks);
1626
1627         sweep_finish ();
1628
1629         sweep_job = NULL;
1630 }
1631
1632 static void
1633 sweep_finish (void)
1634 {
1635         mword total_evacuate_heap = 0;
1636         mword total_evacuate_saved = 0;
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                         mword total_bytes = block_obj_sizes [i] * sweep_slots_available [i];
1652                         total_evacuate_heap += total_bytes;
1653                         if (evacuate_block_obj_sizes [i])
1654                                 total_evacuate_saved += total_bytes - block_obj_sizes [i] * sweep_slots_used [i];
1655                 }
1656         }
1657
1658         want_evacuation = (float)total_evacuate_saved / (float)total_evacuate_heap > (1 - concurrent_evacuation_threshold);
1659
1660         set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1661 }
1662
1663 static void
1664 major_sweep (void)
1665 {
1666         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1667
1668         sweep_start ();
1669
1670         SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?");
1671
1672         num_major_sections_before_sweep = num_major_sections;
1673         num_major_sections_freed_in_sweep = 0;
1674
1675         SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
1676         if (concurrent_sweep) {
1677                 sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
1678                 sgen_thread_pool_job_enqueue (sweep_job);
1679         } else {
1680                 sweep_job_func (NULL, NULL);
1681         }
1682 }
1683
1684 static gboolean
1685 major_have_swept (void)
1686 {
1687         return sweep_state == SWEEP_STATE_SWEPT;
1688 }
1689
1690 static int count_pinned_ref;
1691 static int count_pinned_nonref;
1692 static int count_nonpinned_ref;
1693 static int count_nonpinned_nonref;
1694
1695 static void
1696 count_nonpinned_callback (GCObject *obj, size_t size, void *data)
1697 {
1698         GCVTable vtable = LOAD_VTABLE (obj);
1699
1700         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1701                 ++count_nonpinned_ref;
1702         else
1703                 ++count_nonpinned_nonref;
1704 }
1705
1706 static void
1707 count_pinned_callback (GCObject *obj, size_t size, void *data)
1708 {
1709         GCVTable vtable = LOAD_VTABLE (obj);
1710
1711         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1712                 ++count_pinned_ref;
1713         else
1714                 ++count_pinned_nonref;
1715 }
1716
1717 static G_GNUC_UNUSED void
1718 count_ref_nonref_objs (void)
1719 {
1720         int total;
1721
1722         count_pinned_ref = 0;
1723         count_pinned_nonref = 0;
1724         count_nonpinned_ref = 0;
1725         count_nonpinned_nonref = 0;
1726
1727         major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1728         major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1729
1730         total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1731
1732         g_print ("ref: %d pinned %d non-pinned   non-ref: %d pinned %d non-pinned  --  %.1f\n",
1733                         count_pinned_ref, count_nonpinned_ref,
1734                         count_pinned_nonref, count_nonpinned_nonref,
1735                         (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1736 }
1737
1738 static int
1739 ms_calculate_block_obj_sizes (double factor, int *arr)
1740 {
1741         double target_size;
1742         int num_sizes = 0;
1743         int last_size = 0;
1744
1745         /*
1746          * Have every possible slot size starting with the minimal
1747          * object size up to and including four times that size.  Then
1748          * proceed by increasing geometrically with the given factor.
1749          */
1750
1751         for (int size = SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size <= 4 * SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size += SGEN_ALLOC_ALIGN) {
1752                 if (arr)
1753                         arr [num_sizes] = size;
1754                 ++num_sizes;
1755                 last_size = size;
1756         }
1757         target_size = (double)last_size;
1758
1759         do {
1760                 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1761                 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1762
1763                 if (size != last_size) {
1764                         if (arr)
1765                                 arr [num_sizes] = size;
1766                         ++num_sizes;
1767                         last_size = size;
1768                 }
1769
1770                 target_size *= factor;
1771         } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1772
1773         return num_sizes;
1774 }
1775
1776 /* only valid during minor collections */
1777 static mword old_num_major_sections;
1778
1779 static void
1780 major_start_nursery_collection (void)
1781 {
1782 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1783         consistency_check ();
1784 #endif
1785
1786         old_num_major_sections = num_major_sections;
1787 }
1788
1789 static void
1790 major_finish_nursery_collection (void)
1791 {
1792 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1793         consistency_check ();
1794 #endif
1795 }
1796
1797 static void
1798 major_start_major_collection (void)
1799 {
1800         MSBlockInfo *block;
1801         int i;
1802
1803         major_finish_sweep_checking ();
1804
1805         /*
1806          * Clear the free lists for block sizes where we do evacuation.  For those block
1807          * sizes we will have to allocate new blocks.
1808          */
1809         for (i = 0; i < num_block_obj_sizes; ++i) {
1810                 if (!evacuate_block_obj_sizes [i])
1811                         continue;
1812
1813                 free_block_lists [0][i] = NULL;
1814                 free_block_lists [MS_BLOCK_FLAG_REFS][i] = NULL;
1815         }
1816
1817         if (lazy_sweep)
1818                 binary_protocol_sweep_begin (GENERATION_OLD, TRUE);
1819
1820         /* Sweep all unswept blocks and set them to MARKING */
1821         FOREACH_BLOCK_NO_LOCK (block) {
1822                 if (lazy_sweep)
1823                         sweep_block (block);
1824                 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1825                 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1826         } END_FOREACH_BLOCK_NO_LOCK;
1827
1828         if (lazy_sweep)
1829                 binary_protocol_sweep_end (GENERATION_OLD, TRUE);
1830
1831         set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1832 }
1833
1834 static void
1835 major_finish_major_collection (ScannedObjectCounts *counts)
1836 {
1837 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1838         if (binary_protocol_is_enabled ()) {
1839                 counts->num_scanned_objects = scanned_objects_list.next_slot;
1840
1841                 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1842                 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1843
1844                 sgen_pointer_queue_clear (&scanned_objects_list);
1845         }
1846 #endif
1847 }
1848
1849 #if SIZEOF_VOID_P != 8
1850 static int
1851 compare_pointers (const void *va, const void *vb) {
1852         char *a = *(char**)va, *b = *(char**)vb;
1853         if (a < b)
1854                 return -1;
1855         if (a > b)
1856                 return 1;
1857         return 0;
1858 }
1859 #endif
1860
1861 /*
1862  * This is called with sweep completed and the world stopped.
1863  */
1864 static void
1865 major_free_swept_blocks (size_t allowance)
1866 {
1867         /* FIXME: This is probably too much.  It's assuming all objects are small. */
1868         size_t section_reserve = allowance / MS_BLOCK_SIZE;
1869
1870         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1871
1872 #if SIZEOF_VOID_P != 8
1873         {
1874                 int i, num_empty_blocks_orig, num_blocks, arr_length;
1875                 void *block;
1876                 void **empty_block_arr;
1877                 void **rebuild_next;
1878
1879 #ifdef TARGET_WIN32
1880                 /*
1881                  * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1882                  * a VirtualAlloc ()-ed block.
1883                  */
1884                 return;
1885 #endif
1886
1887                 if (num_empty_blocks <= section_reserve)
1888                         return;
1889                 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1890
1891                 num_empty_blocks_orig = num_empty_blocks;
1892                 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1893                                 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1894                 if (!empty_block_arr)
1895                         goto fallback;
1896
1897                 i = 0;
1898                 for (block = empty_blocks; block; block = *(void**)block)
1899                         empty_block_arr [i++] = block;
1900                 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1901
1902                 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1903
1904                 /*
1905                  * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1906                  * contiguous ones.  If we do, we free them.  If that's not enough to get to
1907                  * section_reserve, we halve the number of contiguous blocks we're looking
1908                  * for and have another go, until we're done with looking for pairs of
1909                  * blocks, at which point we give up and go to the fallback.
1910                  */
1911                 arr_length = num_empty_blocks_orig;
1912                 num_blocks = MS_BLOCK_ALLOC_NUM;
1913                 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1914                         int first = -1;
1915                         int dest = 0;
1916
1917                         dest = 0;
1918                         for (i = 0; i < arr_length; ++i) {
1919                                 int d = dest;
1920                                 void *block = empty_block_arr [i];
1921                                 SGEN_ASSERT (6, block, "we're not shifting correctly");
1922                                 if (i != dest) {
1923                                         empty_block_arr [dest] = block;
1924                                         /*
1925                                          * This is not strictly necessary, but we're
1926                                          * cautious.
1927                                          */
1928                                         empty_block_arr [i] = NULL;
1929                                 }
1930                                 ++dest;
1931
1932                                 if (first < 0) {
1933                                         first = d;
1934                                         continue;
1935                                 }
1936
1937                                 SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
1938
1939                                 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
1940                                         first = d;
1941                                         continue;
1942                                 }
1943
1944                                 if (d + 1 - first == num_blocks) {
1945                                         /*
1946                                          * We found num_blocks contiguous blocks.  Free them
1947                                          * and null their array entries.  As an optimization
1948                                          * we could, instead of nulling the entries, shift
1949                                          * the following entries over to the left, while
1950                                          * we're iterating.
1951                                          */
1952                                         int j;
1953                                         sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
1954                                         for (j = first; j <= d; ++j)
1955                                                 empty_block_arr [j] = NULL;
1956                                         dest = first;
1957                                         first = -1;
1958
1959                                         num_empty_blocks -= num_blocks;
1960
1961                                         stat_major_blocks_freed += num_blocks;
1962                                         if (num_blocks == MS_BLOCK_ALLOC_NUM)
1963                                                 stat_major_blocks_freed_ideal += num_blocks;
1964                                         else
1965                                                 stat_major_blocks_freed_less_ideal += num_blocks;
1966
1967                                 }
1968                         }
1969
1970                         SGEN_ASSERT (6, dest <= i && dest <= arr_length, "array length is off");
1971                         arr_length = dest;
1972                         SGEN_ASSERT (6, arr_length == num_empty_blocks, "array length is off");
1973
1974                         num_blocks >>= 1;
1975                 }
1976
1977                 /* rebuild empty_blocks free list */
1978                 rebuild_next = (void**)&empty_blocks;
1979                 for (i = 0; i < arr_length; ++i) {
1980                         void *block = empty_block_arr [i];
1981                         SGEN_ASSERT (6, block, "we're missing blocks");
1982                         *rebuild_next = block;
1983                         rebuild_next = (void**)block;
1984                 }
1985                 *rebuild_next = NULL;
1986
1987                 /* free array */
1988                 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
1989         }
1990
1991         SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
1992
1993  fallback:
1994         /*
1995          * This is our threshold.  If there's not more empty than used blocks, we won't
1996          * release uncontiguous blocks, in fear of fragmenting the address space.
1997          */
1998         if (num_empty_blocks <= num_major_sections)
1999                 return;
2000 #endif
2001
2002         while (num_empty_blocks > section_reserve) {
2003                 void *next = *(void**)empty_blocks;
2004                 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
2005                 empty_blocks = next;
2006                 /*
2007                  * Needs not be atomic because this is running
2008                  * single-threaded.
2009                  */
2010                 --num_empty_blocks;
2011
2012                 ++stat_major_blocks_freed;
2013 #if SIZEOF_VOID_P != 8
2014                 ++stat_major_blocks_freed_individual;
2015 #endif
2016         }
2017 }
2018
2019 static void
2020 major_pin_objects (SgenGrayQueue *queue)
2021 {
2022         MSBlockInfo *block;
2023
2024         FOREACH_BLOCK_NO_LOCK (block) {
2025                 size_t first_entry, last_entry;
2026                 SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
2027                 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
2028                                 &first_entry, &last_entry);
2029                 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
2030         } END_FOREACH_BLOCK_NO_LOCK;
2031 }
2032
2033 static void
2034 major_init_to_space (void)
2035 {
2036 }
2037
2038 static void
2039 major_report_pinned_memory_usage (void)
2040 {
2041         g_assert_not_reached ();
2042 }
2043
2044 static gint64
2045 major_get_used_size (void)
2046 {
2047         gint64 size = 0;
2048         MSBlockInfo *block;
2049
2050         /*
2051          * We're holding the GC lock, but the sweep thread might be running.  Make sure it's
2052          * finished, then we can iterate over the block array.
2053          */
2054         major_finish_sweep_checking ();
2055
2056         FOREACH_BLOCK_NO_LOCK_CONDITION (TRUE, block) {
2057                 int count = MS_BLOCK_FREE / block->obj_size;
2058                 void **iter;
2059                 size += count * block->obj_size;
2060                 for (iter = block->free_list; iter; iter = (void**)*iter)
2061                         size -= block->obj_size;
2062         } END_FOREACH_BLOCK_NO_LOCK;
2063
2064         return size;
2065 }
2066
2067 /* FIXME: return number of bytes, not of sections */
2068 static size_t
2069 get_num_major_sections (void)
2070 {
2071         return num_major_sections;
2072 }
2073
2074 /*
2075  * Returns the number of bytes in blocks that were present when the last sweep was
2076  * initiated, and were not freed during the sweep.  They are the basis for calculating the
2077  * allowance.
2078  */
2079 static size_t
2080 get_bytes_survived_last_sweep (void)
2081 {
2082         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2083         return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
2084 }
2085
2086 static gboolean
2087 major_handle_gc_param (const char *opt)
2088 {
2089         if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2090                 const char *arg = strchr (opt, '=') + 1;
2091                 int percentage = atoi (arg);
2092                 if (percentage < 0 || percentage > 100) {
2093                         fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2094                         exit (1);
2095                 }
2096                 evacuation_threshold = (float)percentage / 100.0f;
2097                 return TRUE;
2098         } else if (!strcmp (opt, "lazy-sweep")) {
2099                 lazy_sweep = TRUE;
2100                 return TRUE;
2101         } else if (!strcmp (opt, "no-lazy-sweep")) {
2102                 lazy_sweep = FALSE;
2103                 return TRUE;
2104         } else if (!strcmp (opt, "concurrent-sweep")) {
2105                 concurrent_sweep = TRUE;
2106                 return TRUE;
2107         } else if (!strcmp (opt, "no-concurrent-sweep")) {
2108                 concurrent_sweep = FALSE;
2109                 return TRUE;
2110         }
2111
2112         return FALSE;
2113 }
2114
2115 static void
2116 major_print_gc_param_usage (void)
2117 {
2118         fprintf (stderr,
2119                         ""
2120                         "  evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2121                         "  (no-)lazy-sweep\n"
2122                         "  (no-)concurrent-sweep\n"
2123                         );
2124 }
2125
2126 /*
2127  * This callback is used to clear cards, move cards to the shadow table and do counting.
2128  */
2129 static void
2130 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
2131 {
2132         MSBlockInfo *block;
2133         gboolean has_references;
2134
2135         major_finish_sweep_checking ();
2136         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2137                 if (has_references)
2138                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2139         } END_FOREACH_BLOCK_NO_LOCK;
2140 }
2141
2142 #ifdef HEAVY_STATISTICS
2143 extern guint64 marked_cards;
2144 extern guint64 scanned_cards;
2145 extern guint64 scanned_objects;
2146 extern guint64 remarked_cards;
2147 #endif
2148
2149 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2150 /*
2151  * MS blocks are 16K aligned.
2152  * Cardtables are 4K aligned, at least.
2153  * This means that the cardtable of a given block is 32 bytes aligned.
2154  */
2155 static guint8*
2156 initial_skip_card (guint8 *card_data)
2157 {
2158         mword *cards = (mword*)card_data;
2159         mword card;
2160         int i;
2161         for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2162                 card = cards [i];
2163                 if (card)
2164                         break;
2165         }
2166
2167         if (i == CARD_WORDS_PER_BLOCK)
2168                 return card_data + CARDS_PER_BLOCK;
2169
2170 #if defined(__i386__) && defined(__GNUC__)
2171         return card_data + i * 4 +  (__builtin_ffs (card) - 1) / 8;
2172 #elif defined(__x86_64__) && defined(__GNUC__)
2173         return card_data + i * 8 +  (__builtin_ffsll (card) - 1) / 8;
2174 #elif defined(__s390x__) && defined(__GNUC__)
2175         return card_data + i * 8 +  (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2176 #else
2177         for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2178                 if (card_data [i])
2179                         return &card_data [i];
2180         }
2181         return card_data;
2182 #endif
2183 }
2184
2185 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2186 #define MS_BLOCK_OBJ_FAST(b,os,i)                       ((b) + MS_BLOCK_SKIP + (os) * (i))
2187 #define MS_OBJ_ALLOCED_FAST(o,b)                (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2188
2189 static void
2190 scan_card_table_for_block (MSBlockInfo *block, gboolean mod_union, ScanCopyContext ctx)
2191 {
2192         SgenGrayQueue *queue = ctx.queue;
2193         ScanObjectFunc scan_func = ctx.ops->scan_object;
2194 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2195         guint8 cards_copy [CARDS_PER_BLOCK];
2196 #endif
2197         gboolean small_objects;
2198         int block_obj_size;
2199         char *block_start;
2200         guint8 *card_data, *card_base;
2201         guint8 *card_data_end;
2202         char *scan_front = NULL;
2203
2204         block_obj_size = block->obj_size;
2205         small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2206
2207         block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2208
2209         /*
2210          * This is safe in face of card aliasing for the following reason:
2211          *
2212          * Major blocks are 16k aligned, or 32 cards aligned.
2213          * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2214          * sizes, they won't overflow the cardtable overlap modulus.
2215          */
2216         if (mod_union) {
2217                 card_data = card_base = block->cardtable_mod_union;
2218                 /*
2219                  * This happens when the nursery collection that precedes finishing
2220                  * the concurrent collection allocates new major blocks.
2221                  */
2222                 if (!card_data)
2223                         return;
2224         } else {
2225 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2226                 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2227 #else
2228                 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2229                         return;
2230                 card_data = card_base = cards_copy;
2231 #endif
2232         }
2233         card_data_end = card_data + CARDS_PER_BLOCK;
2234
2235         card_data += MS_BLOCK_SKIP >> CARD_BITS;
2236
2237         card_data = initial_skip_card (card_data);
2238         while (card_data < card_data_end) {
2239                 size_t card_index, first_object_index;
2240                 char *start;
2241                 char *end;
2242                 char *first_obj, *obj;
2243
2244                 HEAVY_STAT (++scanned_cards);
2245
2246                 if (!*card_data) {
2247                         ++card_data;
2248                         continue;
2249                 }
2250
2251                 card_index = card_data - card_base;
2252                 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2253                 end = start + CARD_SIZE_IN_BYTES;
2254
2255                 if (!block_is_swept_or_marking (block))
2256                         sweep_block (block);
2257
2258                 HEAVY_STAT (++marked_cards);
2259
2260                 if (small_objects)
2261                         sgen_card_table_prepare_card_for_scanning (card_data);
2262
2263                 /*
2264                  * If the card we're looking at starts at or in the block header, we
2265                  * must start at the first object in the block, without calculating
2266                  * the index of the object we're hypothetically starting at, because
2267                  * it would be negative.
2268                  */
2269                 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2270                         first_object_index = 0;
2271                 else
2272                         first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2273
2274                 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2275
2276                 binary_protocol_card_scan (first_obj, end - first_obj);
2277
2278                 while (obj < end) {
2279                         if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2280                                 goto next_object;
2281
2282                         if (mod_union) {
2283                                 /* FIXME: do this more efficiently */
2284                                 int w, b;
2285                                 MS_CALC_MARK_BIT (w, b, obj);
2286                                 if (!MS_MARK_BIT (block, w, b))
2287                                         goto next_object;
2288                         }
2289
2290                         GCObject *object = (GCObject*)obj;
2291
2292                         if (small_objects) {
2293                                 HEAVY_STAT (++scanned_objects);
2294                                 scan_func (object, sgen_obj_get_descriptor (object), queue);
2295                         } else {
2296                                 size_t offset = sgen_card_table_get_card_offset (obj, block_start);
2297                                 sgen_cardtable_scan_object (object, block_obj_size, card_base + offset, mod_union, ctx);
2298                         }
2299                 next_object:
2300                         obj += block_obj_size;
2301                         g_assert (scan_front <= obj);
2302                         scan_front = obj;
2303                 }
2304
2305                 HEAVY_STAT (if (*card_data) ++remarked_cards);
2306
2307                 if (small_objects)
2308                         ++card_data;
2309                 else
2310                         card_data = card_base + sgen_card_table_get_card_offset (obj, block_start);
2311         }
2312 }
2313
2314 static void
2315 major_scan_card_table (gboolean mod_union, ScanCopyContext ctx)
2316 {
2317         MSBlockInfo *block;
2318         gboolean has_references;
2319
2320         if (!concurrent_mark)
2321                 g_assert (!mod_union);
2322
2323         major_finish_sweep_checking ();
2324         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2325 #ifdef PREFETCH_CARDS
2326                 int prefetch_index = __index + 6;
2327                 if (prefetch_index < allocated_blocks.next_slot) {
2328                         MSBlockInfo *prefetch_block = BLOCK_UNTAG (allocated_blocks.data [prefetch_index]);
2329                         guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2330                         PREFETCH_READ (prefetch_block);
2331                         PREFETCH_WRITE (prefetch_cards);
2332                         PREFETCH_WRITE (prefetch_cards + 32);
2333                 }
2334 #endif
2335
2336                 if (!has_references)
2337                         continue;
2338
2339                 scan_card_table_for_block (block, mod_union, ctx);
2340         } END_FOREACH_BLOCK_NO_LOCK;
2341 }
2342
2343 static void
2344 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2345 {
2346         MSBlockInfo *block;
2347         gboolean has_references;
2348         long long total_cards = 0;
2349         long long marked_cards = 0;
2350
2351         if (sweep_in_progress ()) {
2352                 *num_total_cards = -1;
2353                 *num_marked_cards = -1;
2354                 return;
2355         }
2356
2357         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2358                 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2359                 int i;
2360
2361                 if (!has_references)
2362                         continue;
2363
2364                 total_cards += CARDS_PER_BLOCK;
2365                 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2366                         if (cards [i])
2367                                 ++marked_cards;
2368                 }
2369         } END_FOREACH_BLOCK_NO_LOCK;
2370
2371         *num_total_cards = total_cards;
2372         *num_marked_cards = marked_cards;
2373 }
2374
2375 static void
2376 update_cardtable_mod_union (void)
2377 {
2378         MSBlockInfo *block;
2379
2380         FOREACH_BLOCK_NO_LOCK (block) {
2381                 size_t num_cards;
2382                 guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
2383                 sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2384                 SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2385         } END_FOREACH_BLOCK_NO_LOCK;
2386 }
2387
2388 #undef pthread_create
2389
2390 static void
2391 post_param_init (SgenMajorCollector *collector)
2392 {
2393         collector->sweeps_lazily = lazy_sweep;
2394         collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
2395 }
2396
2397 static void
2398 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
2399 {
2400         int i;
2401
2402         sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2403
2404         num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2405         block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2406         ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2407
2408         evacuate_block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2409         for (i = 0; i < num_block_obj_sizes; ++i)
2410                 evacuate_block_obj_sizes [i] = FALSE;
2411
2412         sweep_slots_available = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2413         sweep_slots_used = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2414         sweep_num_blocks = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2415
2416         /*
2417         {
2418                 int i;
2419                 g_print ("block object sizes:\n");
2420                 for (i = 0; i < num_block_obj_sizes; ++i)
2421                         g_print ("%d\n", block_obj_sizes [i]);
2422         }
2423         */
2424
2425         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2426                 free_block_lists [i] = sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2427
2428         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2429                 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2430         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2431                 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2432
2433         mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2434         mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2435         mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2436         mono_counters_register ("# major objects evacuated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_objects_evacuated);
2437 #if SIZEOF_VOID_P != 8
2438         mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2439         mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2440         mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2441         mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2442 #endif
2443
2444         collector->section_size = MAJOR_SECTION_SIZE;
2445
2446         concurrent_mark = is_concurrent;
2447         collector->is_concurrent = is_concurrent;
2448         collector->needs_thread_pool = is_concurrent || concurrent_sweep;
2449         if (is_concurrent)
2450                 collector->want_synchronous_collection = &want_evacuation;
2451         else
2452                 collector->want_synchronous_collection = NULL;
2453         collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2454         collector->supports_cardtable = TRUE;
2455
2456         collector->alloc_heap = major_alloc_heap;
2457         collector->is_object_live = major_is_object_live;
2458         collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2459         collector->alloc_degraded = major_alloc_degraded;
2460
2461         collector->alloc_object = major_alloc_object;
2462         collector->free_pinned_object = free_pinned_object;
2463         collector->iterate_objects = major_iterate_objects;
2464         collector->free_non_pinned_object = major_free_non_pinned_object;
2465         collector->pin_objects = major_pin_objects;
2466         collector->pin_major_object = pin_major_object;
2467         collector->scan_card_table = major_scan_card_table;
2468         collector->iterate_live_block_ranges = major_iterate_live_block_ranges;
2469         if (is_concurrent) {
2470                 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2471                 collector->get_cardtable_mod_union_for_reference = major_get_cardtable_mod_union_for_reference;
2472         }
2473         collector->init_to_space = major_init_to_space;
2474         collector->sweep = major_sweep;
2475         collector->have_swept = major_have_swept;
2476         collector->finish_sweeping = major_finish_sweep_checking;
2477         collector->free_swept_blocks = major_free_swept_blocks;
2478         collector->check_scan_starts = major_check_scan_starts;
2479         collector->dump_heap = major_dump_heap;
2480         collector->get_used_size = major_get_used_size;
2481         collector->start_nursery_collection = major_start_nursery_collection;
2482         collector->finish_nursery_collection = major_finish_nursery_collection;
2483         collector->start_major_collection = major_start_major_collection;
2484         collector->finish_major_collection = major_finish_major_collection;
2485         collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2486         collector->ptr_is_from_pinned_alloc = ptr_is_from_pinned_alloc;
2487         collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2488         collector->get_num_major_sections = get_num_major_sections;
2489         collector->get_bytes_survived_last_sweep = get_bytes_survived_last_sweep;
2490         collector->handle_gc_param = major_handle_gc_param;
2491         collector->print_gc_param_usage = major_print_gc_param_usage;
2492         collector->post_param_init = post_param_init;
2493         collector->is_valid_object = major_is_valid_object;
2494         collector->describe_pointer = major_describe_pointer;
2495         collector->count_cards = major_count_cards;
2496
2497         collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
2498         collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
2499         if (is_concurrent) {
2500                 collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2501                 collector->major_ops_concurrent_start.scan_object = major_scan_object_no_mark_concurrent_start;
2502
2503                 collector->major_ops_concurrent.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2504                 collector->major_ops_concurrent.scan_object = major_scan_object_no_mark_concurrent;
2505
2506                 collector->major_ops_concurrent_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_finish_canonical;
2507                 collector->major_ops_concurrent_finish.scan_object = major_scan_object_no_evacuation;
2508                 collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_concurrent_finish;
2509         }
2510
2511 #if !defined (FIXED_HEAP) && !defined (SGEN_PARALLEL_MARK)
2512         if (!is_concurrent)
2513                 collector->drain_gray_stack = drain_gray_stack;
2514
2515 #ifdef HEAVY_STATISTICS
2516         mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2517         mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2518         mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2519         mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2520         mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2521         mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2522         mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2523         mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2524         mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2525         mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2526
2527         mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2528         mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2529         mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2530 #endif
2531 #endif
2532
2533 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2534         mono_mutex_init (&scanned_objects_list_lock);
2535 #endif
2536
2537         SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2538
2539         /*cardtable requires major pages to be 8 cards aligned*/
2540         g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2541 }
2542
2543 void
2544 sgen_marksweep_init (SgenMajorCollector *collector)
2545 {
2546         sgen_marksweep_init_internal (collector, FALSE);
2547 }
2548
2549 void
2550 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2551 {
2552         sgen_marksweep_init_internal (collector, TRUE);
2553 }
2554
2555 #endif