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