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  * Licensed under the MIT license. See LICENSE file in the project root for full license information.
  11  */
  12
  13 #include "config.h"
  14
  15 #ifdef HAVE_SGEN_GC
  16
  17 #include <math.h>
  18 #include <errno.h>
  19 #include <string.h>
  20 #include <stdlib.h>
  21
  22 #include "mono/sgen/sgen-gc.h"
  23 #include "mono/sgen/sgen-protocol.h"
  24 #include "mono/sgen/sgen-cardtable.h"
  25 #include "mono/sgen/sgen-memory-governor.h"
  26 #include "mono/sgen/sgen-layout-stats.h"
  27 #include "mono/sgen/sgen-pointer-queue.h"
  28 #include "mono/sgen/sgen-array-list.h"
  29 #include "mono/sgen/sgen-pinning.h"
  30 #include "mono/sgen/sgen-workers.h"
  31 #include "mono/sgen/sgen-thread-pool.h"
  32 #include "mono/sgen/sgen-client.h"
  33 #include "mono/utils/mono-memory-model.h"
  34
  35 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
  36 #define MS_BLOCK_SIZE   ARCH_MIN_MS_BLOCK_SIZE
  37 #define MS_BLOCK_SIZE_SHIFT     ARCH_MIN_MS_BLOCK_SIZE_SHIFT
  38 #else
  39 #define MS_BLOCK_SIZE_SHIFT     14      /* INT FASTENABLE */
  40 #define MS_BLOCK_SIZE           (1 << MS_BLOCK_SIZE_SHIFT)
  41 #endif
  42 #define MAJOR_SECTION_SIZE      MS_BLOCK_SIZE
  43 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
  44
  45 /*
  46  * Don't allocate single blocks, but alloc a contingent of this many
  47  * blocks in one swoop.  This must be a power of two.
  48  */
  49 #define MS_BLOCK_ALLOC_NUM      32
  50
  51 /*
  52  * Number of bytes before the first object in a block.  At the start
  53  * of a block is the MSBlockHeader, then opional padding, then come
  54  * the objects, so this must be >= sizeof (MSBlockHeader).
  55  */
  56 #define MS_BLOCK_SKIP   ((sizeof (MSBlockHeader) + 15) & ~15)
  57
  58 #define MS_BLOCK_FREE   (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
  59
  60 #define MS_NUM_MARK_WORDS       ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
  61
  62 /*
  63  * Blocks progress from one state to the next:
  64  *
  65  * SWEPT           The block is fully swept.  It might or might not be in
  66  *                 a free list.
  67  *
  68  * MARKING         The block might or might not contain live objects.  If
  69  *                 we're in between an initial collection pause and the
  70  *                 finishing pause, the block might or might not be in a
  71  *                 free list.
  72  *
  73  * CHECKING        The sweep thread is investigating the block to determine
  74  *                 whether or not it contains live objects.  The block is
  75  *                 not in a free list.
  76  *
  77  * NEED_SWEEPING   The block contains live objects but has not yet been
  78  *                 swept.  It also contains free slots.  It is in a block
  79  *                 free list.
  80  *
  81  * SWEEPING        The block is being swept.  It might be in a free list.
  82  */
  83
  84 enum {
  85         BLOCK_STATE_SWEPT,
  86         BLOCK_STATE_MARKING,
  87         BLOCK_STATE_CHECKING,
  88         BLOCK_STATE_NEED_SWEEPING,
  89         BLOCK_STATE_SWEEPING
  90 };
  91
  92 typedef struct _MSBlockInfo MSBlockInfo;
  93 struct _MSBlockInfo {
  94         guint16 obj_size;
  95         /*
  96          * FIXME: Do we even need this? It's only used during sweep and might be worth
  97          * recalculating to save the space.
  98          */
  99         guint16 obj_size_index;
 100         /* FIXME: Reduce this - it only needs a byte. */
 101         volatile gint32 state;
 102         gint16 nused;
 103         unsigned int pinned : 1;
 104         unsigned int has_references : 1;
 105         unsigned int has_pinned : 1;    /* means cannot evacuate */
 106         unsigned int is_to_space : 1;
 107         void ** volatile free_list;
 108         MSBlockInfo * volatile next_free;
 109         guint8 * volatile cardtable_mod_union;
 110         mword mark_words [MS_NUM_MARK_WORDS];
 111 };
 112
 113 #define MS_BLOCK_FOR_BLOCK_INFO(b)      ((char*)(b))
 114
 115 #define MS_BLOCK_OBJ(b,i)               ((GCObject *)(MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i)))
 116 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size)             (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
 117 #define MS_BLOCK_DATA_FOR_OBJ(o)        ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
 118
 119 typedef struct {
 120         MSBlockInfo info;
 121 } MSBlockHeader;
 122
 123 #define MS_BLOCK_FOR_OBJ(o)             (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
 124
 125 /* object index will always be small */
 126 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
 127
 128 //casting to int is fine since blocks are 32k
 129 #define MS_CALC_MARK_BIT(w,b,o)         do {                            \
 130                 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
 131                 if (sizeof (mword) == 4) {                              \
 132                         (w) = i >> 5;                                   \
 133                         (b) = i & 31;                                   \
 134                 } else {                                                \
 135                         (w) = i >> 6;                                   \
 136                         (b) = i & 63;                                   \
 137                 }                                                       \
 138         } while (0)
 139
 140 #define MS_MARK_BIT(bl,w,b)     ((bl)->mark_words [(w)] & (ONE_P << (b)))
 141 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
 142
 143 #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))
 144
 145 #define MS_BLOCK_OBJ_SIZE_FACTOR        (pow (2.0, 1.0 / 3))
 146
 147 /*
 148  * This way we can lookup block object size indexes for sizes up to
 149  * 256 bytes with a single load.
 150  */
 151 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES      32
 152
 153 static int *block_obj_sizes;
 154 static int num_block_obj_sizes;
 155 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
 156
 157 #define MS_BLOCK_FLAG_PINNED    1
 158 #define MS_BLOCK_FLAG_REFS      2
 159
 160 #define MS_BLOCK_TYPE_MAX       4
 161
 162 static gboolean *evacuate_block_obj_sizes;
 163 static float evacuation_threshold = 0.666f;
 164
 165 static gboolean lazy_sweep = TRUE;
 166
 167 enum {
 168         SWEEP_STATE_SWEPT,
 169         SWEEP_STATE_NEED_SWEEPING,
 170         SWEEP_STATE_SWEEPING,
 171         SWEEP_STATE_SWEEPING_AND_ITERATING,
 172         SWEEP_STATE_COMPACTING
 173 };
 174
 175 static volatile int sweep_state = SWEEP_STATE_SWEPT;
 176
 177 static gboolean concurrent_mark;
 178 static gboolean concurrent_sweep = TRUE;
 179
 180 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl)      SGEN_POINTER_IS_TAGGED_1 ((bl))
 181 #define BLOCK_TAG_HAS_REFERENCES(bl)            SGEN_POINTER_TAG_1 ((bl))
 182
 183 #define BLOCK_IS_TAGGED_CHECKING(bl)            SGEN_POINTER_IS_TAGGED_2 ((bl))
 184 #define BLOCK_TAG_CHECKING(bl)                  SGEN_POINTER_TAG_2 ((bl))
 185
 186 #define BLOCK_UNTAG(bl)                         ((MSBlockInfo *)SGEN_POINTER_UNTAG_12 ((bl)))
 187
 188 #define BLOCK_TAG(bl)                           ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
 189
 190 /* all allocated blocks in the system */
 191 static SgenArrayList allocated_blocks = SGEN_ARRAY_LIST_INIT (NULL, NULL, NULL, INTERNAL_MEM_PIN_QUEUE);
 192
 193 /* non-allocated block free-list */
 194 static void *empty_blocks = NULL;
 195 static size_t num_empty_blocks = 0;
 196
 197 /*
 198  * We can iterate the block list also while sweep is in progress but we
 199  * need to account for blocks that will be checked for sweeping and even
 200  * freed in the process.
 201  */
 202 #define FOREACH_BLOCK_NO_LOCK(bl) {                                     \
 203         volatile gpointer *slot;                                                \
 204         SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) {        \
 205                 (bl) = BLOCK_UNTAG (*slot);                             \
 206                 if (!(bl))                                              \
 207                         continue;
 208 #define FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK(bl,hr) {                   \
 209         volatile gpointer *slot;                                                \
 210         SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) {        \
 211                 (bl) = (MSBlockInfo *) (*slot);                 \
 212                 if (!(bl))                                              \
 213                         continue;                                       \
 214                 (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl));           \
 215                 (bl) = BLOCK_UNTAG ((bl));
 216 #define END_FOREACH_BLOCK_NO_LOCK       } SGEN_ARRAY_LIST_END_FOREACH_SLOT; }
 217
 218 static volatile size_t num_major_sections = 0;
 219 /*
 220  * One free block list for each block object size.  We add and remove blocks from these
 221  * lists lock-free via CAS.
 222  *
 223  * Blocks accessed/removed from `free_block_lists`:
 224  *   from the mutator (with GC lock held)
 225  *   in nursery collections
 226  *   in non-concurrent major collections
 227  *   in the finishing pause of concurrent major collections (whole list is cleared)
 228  *
 229  * Blocks added to `free_block_lists`:
 230  *   in the sweeping thread
 231  *   during nursery collections
 232  *   from domain clearing (with the world stopped and no sweeping happening)
 233  *
 234  * The only item of those that doesn't require the GC lock is the sweep thread.  The sweep
 235  * thread only ever adds blocks to the free list, so the ABA problem can't occur.
 236  */
 237 static MSBlockInfo * volatile *free_block_lists [MS_BLOCK_TYPE_MAX];
 238
 239 static guint64 stat_major_blocks_alloced = 0;
 240 static guint64 stat_major_blocks_freed = 0;
 241 static guint64 stat_major_blocks_lazy_swept = 0;
 242
 243 static guint64 stat_major_blocks_freed_ideal = 0;
 244 static guint64 stat_major_blocks_freed_less_ideal = 0;
 245 static guint64 stat_major_blocks_freed_individual = 0;
 246 static guint64 stat_major_blocks_alloced_less_ideal = 0;
 247
 248 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
 249 static guint64 num_major_objects_marked = 0;
 250 #define INC_NUM_MAJOR_OBJECTS_MARKED()  (++num_major_objects_marked)
 251 #else
 252 #define INC_NUM_MAJOR_OBJECTS_MARKED()
 253 #endif
 254
 255 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
 256 static mono_mutex_t scanned_objects_list_lock;
 257 static SgenPointerQueue scanned_objects_list;
 258
 259 static void
 260 add_scanned_object (void *ptr)
 261 {
 262         if (!binary_protocol_is_enabled ())
 263                 return;
 264
 265         mono_os_mutex_lock (&scanned_objects_list_lock);
 266         sgen_pointer_queue_add (&scanned_objects_list, ptr);
 267         mono_os_mutex_unlock (&scanned_objects_list_lock);
 268 }
 269 #endif
 270
 271 static gboolean sweep_block (MSBlockInfo *block);
 272
 273 static int
 274 ms_find_block_obj_size_index (size_t size)
 275 {
 276         int i;
 277         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);
 278         for (i = 0; i < num_block_obj_sizes; ++i)
 279                 if (block_obj_sizes [i] >= size)
 280                         return i;
 281         g_error ("no object of size %zd\n", size);
 282         return -1;
 283 }
 284
 285 #define FREE_BLOCKS_FROM(lists,p,r)     (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
 286 #define FREE_BLOCKS(p,r)                (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
 287
 288 #define MS_BLOCK_OBJ_SIZE_INDEX(s)                              \
 289         (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ?      \
 290          fast_block_obj_size_indexes [((s)+7)>>3] :             \
 291          ms_find_block_obj_size_index ((s)))
 292
 293 static void*
 294 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
 295 {
 296         char *start;
 297         if (nursery_align)
 298                 start = (char *)sgen_alloc_os_memory_aligned (nursery_size, nursery_align, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery", MONO_MEM_ACCOUNT_SGEN_NURSERY);
 299         else
 300                 start = (char *)sgen_alloc_os_memory (nursery_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery", MONO_MEM_ACCOUNT_SGEN_NURSERY);
 301
 302         return start;
 303 }
 304
 305 static void
 306 update_heap_boundaries_for_block (MSBlockInfo *block)
 307 {
 308         sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
 309 }
 310
 311 /*
 312  * Thread safe
 313  */
 314 static void*
 315 ms_get_empty_block (void)
 316 {
 317         char *p;
 318         int i;
 319         void *block, *empty, *next;
 320
 321  retry:
 322         if (!empty_blocks) {
 323                 /*
 324                  * We try allocating MS_BLOCK_ALLOC_NUM blocks first.  If that's
 325                  * unsuccessful, we halve the number of blocks and try again, until we're at
 326                  * 1.  If that doesn't work, either, we assert.
 327                  */
 328                 int alloc_num = MS_BLOCK_ALLOC_NUM;
 329                 for (;;) {
 330                         p = (char *)sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE,
 331                                 (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE),
 332                                 alloc_num == 1 ? "major heap section" : NULL, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
 333                         if (p)
 334                                 break;
 335                         alloc_num >>= 1;
 336                 }
 337
 338                 for (i = 0; i < alloc_num; ++i) {
 339                         block = p;
 340                         /*
 341                          * We do the free list update one after the
 342                          * other so that other threads can use the new
 343                          * blocks as quickly as possible.
 344                          */
 345                         do {
 346                                 empty = empty_blocks;
 347                                 *(void**)block = empty;
 348                         } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
 349                         p += MS_BLOCK_SIZE;
 350                 }
 351
 352                 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
 353
 354                 stat_major_blocks_alloced += alloc_num;
 355 #if SIZEOF_VOID_P != 8
 356                 if (alloc_num != MS_BLOCK_ALLOC_NUM)
 357                         stat_major_blocks_alloced_less_ideal += alloc_num;
 358 #endif
 359         }
 360
 361         do {
 362                 empty = empty_blocks;
 363                 if (!empty)
 364                         goto retry;
 365                 block = empty;
 366                 next = *(void**)block;
 367         } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
 368
 369         SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
 370
 371         *(void**)block = NULL;
 372
 373         g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
 374
 375         return block;
 376 }
 377
 378 /*
 379  * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
 380  * list, where it will either be freed later on, or reused in nursery collections.
 381  */
 382 static void
 383 ms_free_block (MSBlockInfo *info)
 384 {
 385         void *empty;
 386         char *block = MS_BLOCK_FOR_BLOCK_INFO (info);
 387
 388         sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
 389         if (info->cardtable_mod_union)
 390                 sgen_card_table_free_mod_union (info->cardtable_mod_union, block, MS_BLOCK_SIZE);
 391         memset (block, 0, MS_BLOCK_SIZE);
 392
 393         do {
 394                 empty = empty_blocks;
 395                 *(void**)block = empty;
 396         } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
 397
 398         SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
 399
 400         binary_protocol_block_free (block, MS_BLOCK_SIZE);
 401 }
 402
 403 static gboolean
 404 sweep_in_progress (void)
 405 {
 406         int state = sweep_state;
 407         return state == SWEEP_STATE_SWEEPING ||
 408                 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
 409                 state == SWEEP_STATE_COMPACTING;
 410 }
 411
 412 static inline gboolean
 413 block_is_swept_or_marking (MSBlockInfo *block)
 414 {
 415         gint32 state = block->state;
 416         return state == BLOCK_STATE_SWEPT || state == BLOCK_STATE_MARKING;
 417 }
 418
 419 //#define MARKSWEEP_CONSISTENCY_CHECK
 420
 421 #ifdef MARKSWEEP_CONSISTENCY_CHECK
 422 static void
 423 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
 424 {
 425         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 426         for (; block; block = block->next_free) {
 427                 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
 428                 g_assert (block->obj_size == size);
 429                 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
 430
 431                 /* blocks in the free lists must have at least
 432                    one free slot */
 433                 g_assert (block->free_list);
 434
 435                 /* the block must be in the allocated_blocks array */
 436                 g_assert (sgen_array_list_find (&allocated_blocks, BLOCK_TAG (block)) != (guint32)-1);
 437         }
 438 }
 439
 440 static void
 441 check_empty_blocks (void)
 442 {
 443         void *p;
 444         size_t i = 0;
 445         for (p = empty_blocks; p; p = *(void**)p)
 446                 ++i;
 447         g_assert (i == num_empty_blocks);
 448 }
 449
 450 static void
 451 consistency_check (void)
 452 {
 453         MSBlockInfo *block;
 454         int i;
 455
 456         /* check all blocks */
 457         FOREACH_BLOCK_NO_LOCK (block) {
 458                 int count = MS_BLOCK_FREE / block->obj_size;
 459                 int num_free = 0;
 460                 void **free;
 461
 462                 /* count number of free slots */
 463                 for (i = 0; i < count; ++i) {
 464                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 465                         if (!MS_OBJ_ALLOCED (obj, block))
 466                                 ++num_free;
 467                 }
 468
 469                 /* check free list */
 470                 for (free = block->free_list; free; free = (void**)*free) {
 471                         g_assert (MS_BLOCK_FOR_OBJ (free) == block);
 472                         --num_free;
 473                 }
 474                 g_assert (num_free == 0);
 475
 476                 /* check all mark words are zero */
 477                 if (!sgen_concurrent_collection_in_progress () && block_is_swept_or_marking (block)) {
 478                         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
 479                                 g_assert (block->mark_words [i] == 0);
 480                 }
 481         } END_FOREACH_BLOCK_NO_LOCK;
 482
 483         /* check free blocks */
 484         for (i = 0; i < num_block_obj_sizes; ++i) {
 485                 int j;
 486                 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
 487                         check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
 488         }
 489
 490         check_empty_blocks ();
 491 }
 492 #endif
 493
 494 static void
 495 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
 496 {
 497         MSBlockInfo *old;
 498         do {
 499                 block->next_free = old = free_blocks [size_index];
 500         } while (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], block, old) != old);
 501 }
 502
 503 static void major_finish_sweep_checking (void);
 504
 505 static gboolean
 506 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
 507 {
 508         int size = block_obj_sizes [size_index];
 509         int count = MS_BLOCK_FREE / size;
 510         MSBlockInfo *info;
 511         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 512         char *obj_start;
 513         int i;
 514
 515         if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
 516                 return FALSE;
 517
 518         info = (MSBlockInfo*)ms_get_empty_block ();
 519
 520         SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
 521
 522         info->obj_size = size;
 523         info->obj_size_index = size_index;
 524         info->pinned = pinned;
 525         info->has_references = has_references;
 526         info->has_pinned = pinned;
 527         /*
 528          * Blocks that are to-space are not evacuated from.  During an major collection
 529          * blocks are allocated for two reasons: evacuating objects from the nursery and
 530          * evacuating them from major blocks marked for evacuation.  In both cases we don't
 531          * want further evacuation. We also don't want to evacuate objects allocated during
 532          * the concurrent mark since it would add pointless stress on the finishing pause.
 533          */
 534         info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD) || sgen_concurrent_collection_in_progress ();
 535         info->state = info->is_to_space ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
 536         SGEN_ASSERT (6, !sweep_in_progress () || info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
 537         info->cardtable_mod_union = NULL;
 538
 539         update_heap_boundaries_for_block (info);
 540
 541         binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
 542
 543         /* build free list */
 544         obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
 545         info->free_list = (void**)obj_start;
 546         /* we're skipping the last one - it must be nulled */
 547         for (i = 0; i < count - 1; ++i) {
 548                 char *next_obj_start = obj_start + size;
 549                 *(void**)obj_start = next_obj_start;
 550                 obj_start = next_obj_start;
 551         }
 552         /* the last one */
 553         *(void**)obj_start = NULL;
 554
 555         add_free_block (free_blocks, size_index, info);
 556
 557         sgen_array_list_add (&allocated_blocks, BLOCK_TAG (info), 0, FALSE);
 558
 559         SGEN_ATOMIC_ADD_P (num_major_sections, 1);
 560         return TRUE;
 561 }
 562
 563 static gboolean
 564 ptr_is_in_major_block (char *ptr, char **start, gboolean *pinned)
 565 {
 566         MSBlockInfo *block;
 567
 568         FOREACH_BLOCK_NO_LOCK (block) {
 569                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
 570                         int count = MS_BLOCK_FREE / block->obj_size;
 571                         int i;
 572
 573                         if (start)
 574                                 *start = NULL;
 575                         for (i = 0; i <= count; ++i) {
 576                                 if (ptr >= (char*)MS_BLOCK_OBJ (block, i) && ptr < (char*)MS_BLOCK_OBJ (block, i + 1)) {
 577                                         if (start)
 578                                                 *start = (char *)MS_BLOCK_OBJ (block, i);
 579                                         break;
 580                                 }
 581                         }
 582                         if (pinned)
 583                                 *pinned = block->pinned;
 584                         return TRUE;
 585                 }
 586         } END_FOREACH_BLOCK_NO_LOCK;
 587         return FALSE;
 588 }
 589
 590 static gboolean
 591 ptr_is_from_pinned_alloc (char *ptr)
 592 {
 593         gboolean pinned;
 594         if (ptr_is_in_major_block (ptr, NULL, &pinned))
 595                 return pinned;
 596         return FALSE;
 597 }
 598
 599 static void
 600 ensure_can_access_block_free_list (MSBlockInfo *block)
 601 {
 602  retry:
 603         for (;;) {
 604                 switch (block->state) {
 605                 case BLOCK_STATE_SWEPT:
 606                 case BLOCK_STATE_MARKING:
 607                         return;
 608                 case BLOCK_STATE_CHECKING:
 609                         SGEN_ASSERT (0, FALSE, "How did we get a block that's being checked from a free list?");
 610                         break;
 611                 case BLOCK_STATE_NEED_SWEEPING:
 612                         if (sweep_block (block))
 613                                 ++stat_major_blocks_lazy_swept;
 614                         break;
 615                 case BLOCK_STATE_SWEEPING:
 616                         /* FIXME: do this more elegantly */
 617                         g_usleep (100);
 618                         goto retry;
 619                 default:
 620                         SGEN_ASSERT (0, FALSE, "Illegal block state");
 621                         break;
 622                 }
 623         }
 624 }
 625
 626 static void*
 627 unlink_slot_from_free_list_uncontested (MSBlockInfo * volatile *free_blocks, int size_index)
 628 {
 629         MSBlockInfo *block, *next_free_block;
 630         void *obj, *next_free_slot;
 631
 632  retry:
 633         block = free_blocks [size_index];
 634         SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
 635
 636         ensure_can_access_block_free_list (block);
 637
 638         obj = block->free_list;
 639         SGEN_ASSERT (6, obj, "block %p in free list had no available object to alloc from", block);
 640
 641         next_free_slot = *(void**)obj;
 642         if (next_free_slot) {
 643                 block->free_list = (gpointer *)next_free_slot;
 644                 return obj;
 645         }
 646
 647         next_free_block = block->next_free;
 648         if (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], next_free_block, block) != block)
 649                 goto retry;
 650
 651         block->free_list = NULL;
 652         block->next_free = NULL;
 653
 654         return obj;
 655 }
 656
 657 static GCObject*
 658 alloc_obj (GCVTable vtable, size_t size, gboolean pinned, gboolean has_references)
 659 {
 660         int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 661         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 662         void *obj;
 663
 664         if (!free_blocks [size_index]) {
 665                 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
 666                         return NULL;
 667         }
 668
 669         obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
 670
 671         /* FIXME: assumes object layout */
 672         *(GCVTable*)obj = vtable;
 673
 674         total_allocated_major += block_obj_sizes [size_index];
 675
 676         return (GCObject *)obj;
 677 }
 678
 679 static GCObject*
 680 major_alloc_object (GCVTable vtable, size_t size, gboolean has_references)
 681 {
 682         return alloc_obj (vtable, size, FALSE, has_references);
 683 }
 684
 685 /*
 686  * We're not freeing the block if it's empty.  We leave that work for
 687  * the next major collection.
 688  *
 689  * This is just called from the domain clearing code, which runs in a
 690  * single thread and has the GC lock, so we don't need an extra lock.
 691  */
 692 static void
 693 free_object (GCObject *obj, size_t size, gboolean pinned)
 694 {
 695         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 696         int word, bit;
 697         gboolean in_free_list;
 698
 699         SGEN_ASSERT (9, sweep_state == SWEEP_STATE_SWEPT, "Should have waited for sweep to free objects.");
 700
 701         ensure_can_access_block_free_list (block);
 702         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);
 703         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
 704         MS_CALC_MARK_BIT (word, bit, obj);
 705         SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set", obj);
 706
 707         memset (obj, 0, size);
 708
 709         in_free_list = !!block->free_list;
 710         *(void**)obj = block->free_list;
 711         block->free_list = (void**)obj;
 712
 713         if (!in_free_list) {
 714                 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (pinned, block->has_references);
 715                 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 716                 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);
 717                 add_free_block (free_blocks, size_index, block);
 718         }
 719 }
 720
 721 static void
 722 major_free_non_pinned_object (GCObject *obj, size_t size)
 723 {
 724         free_object (obj, size, FALSE);
 725 }
 726
 727 /* size is a multiple of SGEN_ALLOC_ALIGN */
 728 static GCObject*
 729 major_alloc_small_pinned_obj (GCVTable vtable, size_t size, gboolean has_references)
 730 {
 731         void *res;
 732
 733         res = alloc_obj (vtable, size, TRUE, has_references);
 734          /*If we failed to alloc memory, we better try releasing memory
 735           *as pinned alloc is requested by the runtime.
 736           */
 737          if (!res) {
 738                 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE, TRUE);
 739                 res = alloc_obj (vtable, size, TRUE, has_references);
 740          }
 741          return (GCObject *)res;
 742 }
 743
 744 static void
 745 free_pinned_object (GCObject *obj, size_t size)
 746 {
 747         free_object (obj, size, TRUE);
 748 }
 749
 750 /*
 751  * size is already rounded up and we hold the GC lock.
 752  */
 753 static GCObject*
 754 major_alloc_degraded (GCVTable vtable, size_t size)
 755 {
 756         GCObject *obj;
 757
 758         obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
 759         if (G_LIKELY (obj)) {
 760                 HEAVY_STAT (++stat_objects_alloced_degraded);
 761                 HEAVY_STAT (stat_bytes_alloced_degraded += size);
 762         }
 763         return obj;
 764 }
 765
 766 /*
 767  * obj is some object.  If it's not in the major heap (i.e. if it's in
 768  * the nursery or LOS), return FALSE.  Otherwise return whether it's
 769  * been marked or copied.
 770  */
 771 static gboolean
 772 major_is_object_live (GCObject *obj)
 773 {
 774         MSBlockInfo *block;
 775         int word, bit;
 776         mword objsize;
 777
 778         if (sgen_ptr_in_nursery (obj))
 779                 return FALSE;
 780
 781         objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
 782
 783         /* LOS */
 784         if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
 785                 return FALSE;
 786
 787         /* now we know it's in a major block */
 788         block = MS_BLOCK_FOR_OBJ (obj);
 789         SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?", block);
 790         MS_CALC_MARK_BIT (word, bit, obj);
 791         return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
 792 }
 793
 794 static gboolean
 795 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
 796 {
 797         gboolean pinned;
 798         if (ptr_is_in_major_block (ptr, start, &pinned))
 799                 return !pinned;
 800         return FALSE;
 801 }
 802
 803 static gboolean
 804 try_set_sweep_state (int new_, int expected)
 805 {
 806         int old = SGEN_CAS (&sweep_state, new_, expected);
 807         return old == expected;
 808 }
 809
 810 static void
 811 set_sweep_state (int new_, int expected)
 812 {
 813         gboolean success = try_set_sweep_state (new_, expected);
 814         SGEN_ASSERT (0, success, "Could not set sweep state.");
 815 }
 816
 817 static gboolean ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked);
 818
 819 static SgenThreadPoolJob * volatile sweep_job;
 820 static SgenThreadPoolJob * volatile sweep_blocks_job;
 821
 822 static void
 823 major_finish_sweep_checking (void)
 824 {
 825         guint32 block_index;
 826         SgenThreadPoolJob *job;
 827
 828  retry:
 829         switch (sweep_state) {
 830         case SWEEP_STATE_SWEPT:
 831         case SWEEP_STATE_NEED_SWEEPING:
 832                 return;
 833         case SWEEP_STATE_SWEEPING:
 834                 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
 835                         break;
 836                 goto retry;
 837         case SWEEP_STATE_SWEEPING_AND_ITERATING:
 838                 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
 839                 goto retry;
 840         case SWEEP_STATE_COMPACTING:
 841                 goto wait;
 842         default:
 843                 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
 844                 break;
 845         }
 846
 847         /*
 848          * We're running with the world stopped and the only other thread doing work is the
 849          * sweep thread, which doesn't add blocks to the array, so we can safely access
 850          * `next_slot`.
 851          */
 852         for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index)
 853                 ensure_block_is_checked_for_sweeping (block_index, FALSE, NULL);
 854
 855         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
 856
 857  wait:
 858         job = sweep_job;
 859         if (job)
 860                 sgen_thread_pool_job_wait (job);
 861         SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
 862         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
 863 }
 864
 865 static void
 866 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
 867 {
 868         gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
 869         gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
 870         gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
 871         MSBlockInfo *block;
 872
 873         /* No actual sweeping will take place if we are in the middle of a major collection. */
 874         major_finish_sweep_checking ();
 875         FOREACH_BLOCK_NO_LOCK (block) {
 876                 int count = MS_BLOCK_FREE / block->obj_size;
 877                 int i;
 878
 879                 if (block->pinned && !pinned)
 880                         continue;
 881                 if (!block->pinned && !non_pinned)
 882                         continue;
 883                 if (sweep && lazy_sweep && !block_is_swept_or_marking (block)) {
 884                         sweep_block (block);
 885                         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
 886                 }
 887
 888                 for (i = 0; i < count; ++i) {
 889                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 890                         if (MS_OBJ_ALLOCED (obj, block))
 891                                 callback ((GCObject*)obj, block->obj_size, data);
 892                 }
 893         } END_FOREACH_BLOCK_NO_LOCK;
 894 }
 895
 896 static gboolean
 897 major_is_valid_object (char *object)
 898 {
 899         MSBlockInfo *block;
 900
 901         FOREACH_BLOCK_NO_LOCK (block) {
 902                 int idx;
 903                 char *obj;
 904
 905                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
 906                         continue;
 907
 908                 idx = MS_BLOCK_OBJ_INDEX (object, block);
 909                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 910                 if (obj != object)
 911                         return FALSE;
 912                 return MS_OBJ_ALLOCED (obj, block);
 913         } END_FOREACH_BLOCK_NO_LOCK;
 914
 915         return FALSE;
 916 }
 917
 918
 919 static GCVTable
 920 major_describe_pointer (char *ptr)
 921 {
 922         MSBlockInfo *block;
 923
 924         FOREACH_BLOCK_NO_LOCK (block) {
 925                 int idx;
 926                 char *obj;
 927                 gboolean live;
 928                 GCVTable vtable;
 929                 int w, b;
 930                 gboolean marked;
 931
 932                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
 933                         continue;
 934
 935                 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
 936                         MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
 937
 938                 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
 939                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 940                 live = MS_OBJ_ALLOCED (obj, block);
 941                 vtable = live ? SGEN_LOAD_VTABLE ((GCObject*)obj) : NULL;
 942
 943                 MS_CALC_MARK_BIT (w, b, obj);
 944                 marked = MS_MARK_BIT (block, w, b);
 945
 946                 if (obj == ptr) {
 947                         SGEN_LOG (0, "\t(");
 948                         if (live)
 949                                 SGEN_LOG (0, "object");
 950                         else
 951                                 SGEN_LOG (0, "dead-object");
 952                 } else {
 953                         if (live)
 954                                 SGEN_LOG (0, "interior-ptr offset %zd", ptr - obj);
 955                         else
 956                                 SGEN_LOG (0, "dead-interior-ptr offset %zd", ptr - obj);
 957                 }
 958
 959                 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
 960
 961                 return vtable;
 962         } END_FOREACH_BLOCK_NO_LOCK;
 963
 964         return NULL;
 965 }
 966
 967 static void
 968 major_check_scan_starts (void)
 969 {
 970 }
 971
 972 static void
 973 major_dump_heap (FILE *heap_dump_file)
 974 {
 975         MSBlockInfo *block;
 976         int *slots_available = (int *)alloca (sizeof (int) * num_block_obj_sizes);
 977         int *slots_used = (int *)alloca (sizeof (int) * num_block_obj_sizes);
 978         int i;
 979
 980         for (i = 0; i < num_block_obj_sizes; ++i)
 981                 slots_available [i] = slots_used [i] = 0;
 982
 983         FOREACH_BLOCK_NO_LOCK (block) {
 984                 int index = ms_find_block_obj_size_index (block->obj_size);
 985                 int count = MS_BLOCK_FREE / block->obj_size;
 986
 987                 slots_available [index] += count;
 988                 for (i = 0; i < count; ++i) {
 989                         if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
 990                                 ++slots_used [index];
 991                 }
 992         } END_FOREACH_BLOCK_NO_LOCK;
 993
 994         fprintf (heap_dump_file, "<occupancies>\n");
 995         for (i = 0; i < num_block_obj_sizes; ++i) {
 996                 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
 997                                 block_obj_sizes [i], slots_available [i], slots_used [i]);
 998         }
 999         fprintf (heap_dump_file, "</occupancies>\n");
1000
1001         FOREACH_BLOCK_NO_LOCK (block) {
1002                 int count = MS_BLOCK_FREE / block->obj_size;
1003                 int i;
1004                 int start = -1;
1005
1006                 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
1007
1008                 for (i = 0; i <= count; ++i) {
1009                         if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
1010                                 if (start < 0)
1011                                         start = i;
1012                         } else {
1013                                 if (start >= 0) {
1014                                         sgen_dump_occupied ((char *)MS_BLOCK_OBJ (block, start), (char *)MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
1015                                         start = -1;
1016                                 }
1017                         }
1018                 }
1019
1020                 fprintf (heap_dump_file, "</section>\n");
1021         } END_FOREACH_BLOCK_NO_LOCK;
1022 }
1023
1024 static guint8*
1025 get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate)
1026 {
1027         guint8 *mod_union = block->cardtable_mod_union;
1028         guint8 *other;
1029         if (mod_union)
1030                 return mod_union;
1031         else if (!allocate)
1032                 return NULL;
1033         mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1034         other = (guint8 *)SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
1035         if (!other) {
1036                 SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
1037                 return mod_union;
1038         }
1039         sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1040         return other;
1041 }
1042
1043 static inline guint8*
1044 major_get_cardtable_mod_union_for_reference (char *ptr)
1045 {
1046         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (ptr);
1047         size_t offset = sgen_card_table_get_card_offset (ptr, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1048         guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
1049         SGEN_ASSERT (0, mod_union, "FIXME: optionally allocate the mod union if it's not here and CAS it in.");
1050         return &mod_union [offset];
1051 }
1052
1053 /*
1054  * Mark the mod-union card for `ptr`, which must be a reference within the object `obj`.
1055  */
1056 static void
1057 mark_mod_union_card (GCObject *obj, void **ptr, GCObject *value_obj)
1058 {
1059         int type = sgen_obj_get_descriptor (obj) & DESC_TYPE_MASK;
1060         if (sgen_safe_object_is_small (obj, type)) {
1061                 guint8 *card_byte = major_get_cardtable_mod_union_for_reference ((char*)ptr);
1062                 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?");
1063                 *card_byte = 1;
1064         } else {
1065                 sgen_los_mark_mod_union_card (obj, ptr);
1066         }
1067         binary_protocol_mod_union_remset (obj, ptr, value_obj, SGEN_LOAD_VTABLE (value_obj));
1068 }
1069
1070 static inline gboolean
1071 major_block_is_evacuating (MSBlockInfo *block)
1072 {
1073         if (evacuate_block_obj_sizes [block->obj_size_index] &&
1074                         !block->has_pinned &&
1075                         !block->is_to_space)
1076                 return TRUE;
1077         return FALSE;
1078 }
1079
1080 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do {           \
1081                 int __word, __bit;                                      \
1082                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
1083                 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
1084                 if (!MS_MARK_BIT ((block), __word, __bit)) {            \
1085                         MS_SET_MARK_BIT ((block), __word, __bit);       \
1086                         if (sgen_gc_descr_has_references (desc))                        \
1087                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1088                         binary_protocol_mark ((obj), (gpointer)SGEN_LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1089                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
1090                 }                                                       \
1091         } while (0)
1092
1093 static void
1094 pin_major_object (GCObject *obj, SgenGrayQueue *queue)
1095 {
1096         MSBlockInfo *block;
1097
1098         if (concurrent_mark)
1099                 g_assert_not_reached ();
1100
1101         block = MS_BLOCK_FOR_OBJ (obj);
1102         block->has_pinned = TRUE;
1103         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1104 }
1105
1106 #include "sgen-major-copy-object.h"
1107
1108 static long long
1109 major_get_and_reset_num_major_objects_marked (void)
1110 {
1111 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1112         long long num = num_major_objects_marked;
1113         num_major_objects_marked = 0;
1114         return num;
1115 #else
1116         return 0;
1117 #endif
1118 }
1119
1120 #define PREFETCH_CARDS          1       /* BOOL FASTENABLE */
1121 #if !PREFETCH_CARDS
1122 #undef PREFETCH_CARDS
1123 #endif
1124
1125 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1126 #if defined(PLATFORM_MACOSX)
1127 #if MONO_GNUC_VERSION <= 40300
1128 #undef PREFETCH_CARDS
1129 #endif
1130 #endif
1131
1132 #ifdef HEAVY_STATISTICS
1133 static guint64 stat_optimized_copy;
1134 static guint64 stat_optimized_copy_nursery;
1135 static guint64 stat_optimized_copy_nursery_forwarded;
1136 static guint64 stat_optimized_copy_nursery_pinned;
1137 static guint64 stat_optimized_copy_major;
1138 static guint64 stat_optimized_copy_major_small_fast;
1139 static guint64 stat_optimized_copy_major_small_slow;
1140 static guint64 stat_optimized_copy_major_large;
1141 static guint64 stat_optimized_copy_major_forwarded;
1142 static guint64 stat_optimized_copy_major_small_evacuate;
1143 static guint64 stat_optimized_major_scan;
1144 static guint64 stat_optimized_major_scan_no_refs;
1145
1146 static guint64 stat_drain_prefetch_fills;
1147 static guint64 stat_drain_prefetch_fill_failures;
1148 static guint64 stat_drain_loops;
1149 #endif
1150
1151 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_no_evacuation
1152 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_no_evacuation
1153 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_no_evacuation
1154 #include "sgen-marksweep-drain-gray-stack.h"
1155
1156 #define COPY_OR_MARK_WITH_EVACUATION
1157 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_with_evacuation
1158 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_with_evacuation
1159 #define SCAN_VTYPE_FUNCTION_NAME        major_scan_vtype_with_evacuation
1160 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_with_evacuation
1161 #define SCAN_PTR_FIELD_FUNCTION_NAME    major_scan_ptr_field_with_evacuation
1162 #include "sgen-marksweep-drain-gray-stack.h"
1163
1164 #define COPY_OR_MARK_CONCURRENT
1165 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_concurrent_no_evacuation
1166 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_concurrent_no_evacuation
1167 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_concurrent_no_evacuation
1168 #include "sgen-marksweep-drain-gray-stack.h"
1169
1170 #define COPY_OR_MARK_CONCURRENT_WITH_EVACUATION
1171 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_concurrent_with_evacuation
1172 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_concurrent_with_evacuation
1173 #define SCAN_VTYPE_FUNCTION_NAME        major_scan_vtype_concurrent_with_evacuation
1174 #define SCAN_PTR_FIELD_FUNCTION_NAME    major_scan_ptr_field_concurrent_with_evacuation
1175 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_concurrent_with_evacuation
1176 #include "sgen-marksweep-drain-gray-stack.h"
1177
1178 static inline gboolean
1179 major_is_evacuating (void)
1180 {
1181         int i;
1182         for (i = 0; i < num_block_obj_sizes; ++i) {
1183                 if (evacuate_block_obj_sizes [i]) {
1184                         return TRUE;
1185                 }
1186         }
1187
1188         return FALSE;
1189 }
1190
1191 static gboolean
1192 drain_gray_stack (SgenGrayQueue *queue)
1193 {
1194         if (major_is_evacuating ())
1195                 return drain_gray_stack_with_evacuation (queue);
1196         else
1197                 return drain_gray_stack_no_evacuation (queue);
1198 }
1199
1200 static gboolean
1201 drain_gray_stack_concurrent (SgenGrayQueue *queue)
1202 {
1203         if (major_is_evacuating ())
1204                 return drain_gray_stack_concurrent_with_evacuation (queue);
1205         else
1206                 return drain_gray_stack_concurrent_no_evacuation (queue);
1207 }
1208
1209 static void
1210 major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue)
1211 {
1212         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1213 }
1214
1215 static void
1216 major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *queue)
1217 {
1218         major_copy_or_mark_object_concurrent_with_evacuation (ptr, *ptr, queue);
1219 }
1220
1221 static void
1222 major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue)
1223 {
1224         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1225 }
1226
1227 static void
1228 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1229 {
1230         void **entry, **end;
1231         int last_index = -1;
1232
1233         if (first_entry == last_entry)
1234                 return;
1235
1236         entry = sgen_pinning_get_entry (first_entry);
1237         end = sgen_pinning_get_entry (last_entry);
1238
1239         for (; entry < end; ++entry) {
1240                 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1241                 GCObject *obj;
1242                 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));
1243                 if (index == last_index)
1244                         continue;
1245                 obj = MS_BLOCK_OBJ (block, index);
1246                 if (!MS_OBJ_ALLOCED (obj, block))
1247                         continue;
1248                 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1249                 sgen_pin_stats_register_object (obj, GENERATION_OLD);
1250                 last_index = index;
1251         }
1252
1253         /*
1254          * There might have been potential pinning "pointers" into this block, but none of
1255          * them pointed to occupied slots, in which case we don't have to pin the block.
1256          */
1257         if (last_index >= 0)
1258                 block->has_pinned = TRUE;
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 = (void **)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         binary_protocol_block_set_state (block, MS_BLOCK_SIZE, expected_state, new_state);
1307 }
1308
1309 /*
1310  * If `block` needs sweeping, sweep it and return TRUE.  Otherwise return FALSE.
1311  *
1312  * Sweeping means iterating through the block's slots and building the free-list from the
1313  * unmarked ones.  They will also be zeroed.  The mark bits will be reset.
1314  */
1315 static gboolean
1316 sweep_block (MSBlockInfo *block)
1317 {
1318         int count;
1319         void *reversed = NULL;
1320
1321  retry:
1322         switch (block->state) {
1323         case BLOCK_STATE_SWEPT:
1324                 return FALSE;
1325         case BLOCK_STATE_MARKING:
1326         case BLOCK_STATE_CHECKING:
1327                 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1328                 goto retry;
1329         case BLOCK_STATE_SWEEPING:
1330                 /* FIXME: Do this more elegantly */
1331                 g_usleep (100);
1332                 goto retry;
1333         case BLOCK_STATE_NEED_SWEEPING:
1334                 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1335                         goto retry;
1336                 break;
1337         default:
1338                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1339         }
1340
1341         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1342
1343         count = MS_BLOCK_FREE / block->obj_size;
1344
1345         block->free_list = NULL;
1346
1347         /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1348         // FIXME: Add more sizes
1349         switch (block->obj_size) {
1350         case 16:
1351                 sweep_block_for_size (block, count, 16);
1352                 break;
1353         default:
1354                 sweep_block_for_size (block, count, block->obj_size);
1355                 break;
1356         }
1357
1358         /* reset mark bits */
1359         memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1360
1361         /* Reverse free list so that it's in address order */
1362         reversed = NULL;
1363         while (block->free_list) {
1364                 void *next = *(void**)block->free_list;
1365                 *(void**)block->free_list = reversed;
1366                 reversed = block->free_list;
1367                 block->free_list = (void **)next;
1368         }
1369         block->free_list = (void **)reversed;
1370
1371         mono_memory_write_barrier ();
1372
1373         set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1374
1375         return TRUE;
1376 }
1377
1378 static inline int
1379 bitcount (mword d)
1380 {
1381         int count = 0;
1382
1383 #ifdef __GNUC__
1384         if (sizeof (mword) == 8)
1385                 count += __builtin_popcountll (d);
1386         else
1387                 count += __builtin_popcount (d);
1388 #else
1389         while (d) {
1390                 count ++;
1391                 d &= (d - 1);
1392         }
1393 #endif
1394         return count;
1395 }
1396
1397 /* statistics for evacuation */
1398 static size_t *sweep_slots_available;
1399 static size_t *sweep_slots_used;
1400 static size_t *sweep_num_blocks;
1401
1402 static volatile size_t num_major_sections_before_sweep;
1403 static volatile size_t num_major_sections_freed_in_sweep;
1404
1405 static void
1406 sweep_start (void)
1407 {
1408         int i;
1409
1410         for (i = 0; i < num_block_obj_sizes; ++i)
1411                 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1412
1413         /* clear all the free lists */
1414         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1415                 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1416                 int j;
1417                 for (j = 0; j < num_block_obj_sizes; ++j)
1418                         free_blocks [j] = NULL;
1419         }
1420
1421         sgen_array_list_remove_nulls (&allocated_blocks);
1422 }
1423
1424 static void sweep_finish (void);
1425
1426 /*
1427  * If `wait` is TRUE and the block is currently being checked, this function will wait until
1428  * the checking has finished.
1429  *
1430  * Returns whether the block is still there.  If `wait` is FALSE, the return value will not
1431  * be correct, i.e. must not be used.
1432  */
1433 static gboolean
1434 ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked)
1435 {
1436         int count;
1437         gboolean have_live = FALSE;
1438         gboolean have_free = FALSE;
1439         int nused = 0;
1440         int block_state;
1441         int i;
1442         void *tagged_block;
1443         MSBlockInfo *block;
1444         volatile gpointer *block_slot = sgen_array_list_get_slot (&allocated_blocks, block_index);
1445
1446         SGEN_ASSERT (6, sweep_in_progress (), "Why do we call this function if there's no sweep in progress?");
1447
1448         if (have_checked)
1449                 *have_checked = FALSE;
1450
1451  retry:
1452         tagged_block = *(void * volatile *)block_slot;
1453         if (!tagged_block)
1454                 return FALSE;
1455
1456         if (BLOCK_IS_TAGGED_CHECKING (tagged_block)) {
1457                 if (!wait)
1458                         return FALSE;
1459                 /* FIXME: do this more elegantly */
1460                 g_usleep (100);
1461                 goto retry;
1462         }
1463
1464         if (SGEN_CAS_PTR (block_slot, BLOCK_TAG_CHECKING (tagged_block), tagged_block) != tagged_block)
1465                 goto retry;
1466
1467         block = BLOCK_UNTAG (tagged_block);
1468         block_state = block->state;
1469
1470         if (!sweep_in_progress ()) {
1471                 SGEN_ASSERT (6, block_state != BLOCK_STATE_SWEEPING && block_state != BLOCK_STATE_CHECKING, "Invalid block state.");
1472                 if (!lazy_sweep)
1473                         SGEN_ASSERT (6, block_state != BLOCK_STATE_NEED_SWEEPING, "Invalid block state.");
1474         }
1475
1476         switch (block_state) {
1477         case BLOCK_STATE_SWEPT:
1478         case BLOCK_STATE_NEED_SWEEPING:
1479         case BLOCK_STATE_SWEEPING:
1480                 goto done;
1481         case BLOCK_STATE_MARKING:
1482                 break;
1483         case BLOCK_STATE_CHECKING:
1484                 SGEN_ASSERT (0, FALSE, "We set the CHECKING bit - how can the stage be CHECKING?");
1485                 goto done;
1486         default:
1487                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1488                 break;
1489         }
1490
1491         SGEN_ASSERT (6, block->state == BLOCK_STATE_MARKING, "When we sweep all blocks must start out marking.");
1492         set_block_state (block, BLOCK_STATE_CHECKING, BLOCK_STATE_MARKING);
1493
1494         if (have_checked)
1495                 *have_checked = TRUE;
1496
1497         block->has_pinned = block->pinned;
1498
1499         block->is_to_space = FALSE;
1500
1501         count = MS_BLOCK_FREE / block->obj_size;
1502
1503         if (block->cardtable_mod_union)
1504                 memset (block->cardtable_mod_union, 0, CARDS_PER_BLOCK);
1505
1506         /* Count marked objects in the block */
1507         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1508                 nused += bitcount (block->mark_words [i]);
1509
1510         block->nused = nused;
1511         if (nused)
1512                 have_live = TRUE;
1513         if (nused < count)
1514                 have_free = TRUE;
1515
1516         if (have_live) {
1517                 int obj_size_index = block->obj_size_index;
1518                 gboolean has_pinned = block->has_pinned;
1519
1520                 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1521
1522                 /*
1523                  * FIXME: Go straight to SWEPT if there are no free slots.  We need
1524                  * to set the free slot list to NULL, though, and maybe update some
1525                  * statistics.
1526                  */
1527                 if (!lazy_sweep)
1528                         sweep_block (block);
1529
1530                 if (!has_pinned) {
1531                         ++sweep_num_blocks [obj_size_index];
1532                         sweep_slots_used [obj_size_index] += nused;
1533                         sweep_slots_available [obj_size_index] += count;
1534                 }
1535
1536                 /*
1537                  * If there are free slots in the block, add
1538                  * the block to the corresponding free list.
1539                  */
1540                 if (have_free) {
1541                         MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1542
1543                         if (!lazy_sweep)
1544                                 SGEN_ASSERT (6, block->free_list, "How do we not have a free list when there are free slots?");
1545
1546                         add_free_block (free_blocks, obj_size_index, block);
1547                 }
1548
1549                 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1550                 update_heap_boundaries_for_block (block);
1551         } else {
1552                 /*
1553                  * Blocks without live objects are removed from the
1554                  * block list and freed.
1555                  */
1556                 SGEN_ASSERT (6, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1557                 SGEN_ASSERT (6, *block_slot == BLOCK_TAG_CHECKING (tagged_block), "How did the block move?");
1558
1559                 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1560                 ms_free_block (block);
1561
1562                 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1563
1564                 tagged_block = NULL;
1565         }
1566
1567  done:
1568         /*
1569          * Once the block is written back without the checking bit other threads are
1570          * free to access it. Make sure the block state is visible before we write it
1571          * back.
1572          */
1573         mono_memory_write_barrier ();
1574         *block_slot = tagged_block;
1575         return !!tagged_block;
1576 }
1577
1578 static void
1579 sweep_blocks_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1580 {
1581         volatile gpointer *slot;
1582         MSBlockInfo *bl;
1583
1584         SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) {
1585                 bl = BLOCK_UNTAG (*slot);
1586                 if (bl)
1587                         sweep_block (bl);
1588         } SGEN_ARRAY_LIST_END_FOREACH_SLOT;
1589
1590         mono_memory_write_barrier ();
1591
1592         sweep_blocks_job = NULL;
1593 }
1594
1595 static void
1596 sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1597 {
1598         guint32 block_index;
1599         guint32 num_blocks = num_major_sections_before_sweep;
1600
1601         SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1602         SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1603
1604         /*
1605          * We traverse the block array from high to low.  Nursery collections will have to
1606          * cooperate with the sweep thread to finish sweeping, and they will traverse from
1607          * low to high, to avoid constantly colliding on the same blocks.
1608          */
1609         for (block_index = num_blocks; block_index-- > 0;) {
1610                 /*
1611                  * The block might have been freed by another thread doing some checking
1612                  * work.
1613                  */
1614                 if (!ensure_block_is_checked_for_sweeping (block_index, TRUE, NULL))
1615                         ++num_major_sections_freed_in_sweep;
1616         }
1617
1618         while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) {
1619                 /*
1620                  * The main GC thread is currently iterating over the block array to help us
1621                  * finish the sweep.  We have already finished, but we don't want to mess up
1622                  * that iteration, so we just wait for it.
1623                  */
1624                 g_usleep (100);
1625         }
1626
1627         if (SGEN_MAX_ASSERT_LEVEL >= 6) {
1628                 for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1629                         MSBlockInfo *block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, block_index));
1630                         SGEN_ASSERT (6, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1631                 }
1632         }
1633
1634         /*
1635          * Concurrently sweep all the blocks to reduce workload during minor
1636          * pauses where we need certain blocks to be swept. At the start of
1637          * the next major we need all blocks to be swept anyway.
1638          */
1639         if (concurrent_sweep && lazy_sweep) {
1640                 sweep_blocks_job = sgen_thread_pool_job_alloc ("sweep_blocks", sweep_blocks_job_func, sizeof (SgenThreadPoolJob));
1641                 sgen_thread_pool_job_enqueue (sweep_blocks_job);
1642         }
1643
1644         sweep_finish ();
1645
1646         sweep_job = NULL;
1647 }
1648
1649 static void
1650 sweep_finish (void)
1651 {
1652         mword used_slots_size = 0;
1653         int i;
1654
1655         for (i = 0; i < num_block_obj_sizes; ++i) {
1656                 float usage = (float)sweep_slots_used [i] / (float)sweep_slots_available [i];
1657                 if (sweep_num_blocks [i] > 5 && usage < evacuation_threshold) {
1658                         evacuate_block_obj_sizes [i] = TRUE;
1659                         /*
1660                         g_print ("slot size %d - %d of %d used\n",
1661                                         block_obj_sizes [i], slots_used [i], slots_available [i]);
1662                         */
1663                 } else {
1664                         evacuate_block_obj_sizes [i] = FALSE;
1665                 }
1666
1667                 used_slots_size += sweep_slots_used [i] * block_obj_sizes [i];
1668         }
1669
1670         sgen_memgov_major_post_sweep (used_slots_size);
1671
1672         set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1673         if (concurrent_sweep)
1674                 binary_protocol_concurrent_sweep_end (sgen_timestamp ());
1675 }
1676
1677 static void
1678 major_sweep (void)
1679 {
1680         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1681
1682         sweep_start ();
1683
1684         SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?");
1685
1686         num_major_sections_before_sweep = num_major_sections;
1687         num_major_sections_freed_in_sweep = 0;
1688
1689         SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
1690         if (concurrent_sweep) {
1691                 sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
1692                 sgen_thread_pool_job_enqueue (sweep_job);
1693         } else {
1694                 sweep_job_func (NULL, NULL);
1695         }
1696 }
1697
1698 static gboolean
1699 major_have_swept (void)
1700 {
1701         return sweep_state == SWEEP_STATE_SWEPT;
1702 }
1703
1704 static int count_pinned_ref;
1705 static int count_pinned_nonref;
1706 static int count_nonpinned_ref;
1707 static int count_nonpinned_nonref;
1708
1709 static void
1710 count_nonpinned_callback (GCObject *obj, size_t size, void *data)
1711 {
1712         GCVTable vtable = SGEN_LOAD_VTABLE (obj);
1713
1714         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1715                 ++count_nonpinned_ref;
1716         else
1717                 ++count_nonpinned_nonref;
1718 }
1719
1720 static void
1721 count_pinned_callback (GCObject *obj, size_t size, void *data)
1722 {
1723         GCVTable vtable = SGEN_LOAD_VTABLE (obj);
1724
1725         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1726                 ++count_pinned_ref;
1727         else
1728                 ++count_pinned_nonref;
1729 }
1730
1731 static G_GNUC_UNUSED void
1732 count_ref_nonref_objs (void)
1733 {
1734         int total;
1735
1736         count_pinned_ref = 0;
1737         count_pinned_nonref = 0;
1738         count_nonpinned_ref = 0;
1739         count_nonpinned_nonref = 0;
1740
1741         major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1742         major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1743
1744         total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1745
1746         g_print ("ref: %d pinned %d non-pinned   non-ref: %d pinned %d non-pinned  --  %.1f\n",
1747                         count_pinned_ref, count_nonpinned_ref,
1748                         count_pinned_nonref, count_nonpinned_nonref,
1749                         (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1750 }
1751
1752 static int
1753 ms_calculate_block_obj_sizes (double factor, int *arr)
1754 {
1755         double target_size;
1756         int num_sizes = 0;
1757         int last_size = 0;
1758
1759         /*
1760          * Have every possible slot size starting with the minimal
1761          * object size up to and including four times that size.  Then
1762          * proceed by increasing geometrically with the given factor.
1763          */
1764
1765         for (int size = SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size <= 4 * SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size += SGEN_ALLOC_ALIGN) {
1766                 if (arr)
1767                         arr [num_sizes] = size;
1768                 ++num_sizes;
1769                 last_size = size;
1770         }
1771         target_size = (double)last_size;
1772
1773         do {
1774                 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1775                 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1776
1777                 if (size != last_size) {
1778                         if (arr)
1779                                 arr [num_sizes] = size;
1780                         ++num_sizes;
1781                         last_size = size;
1782                 }
1783
1784                 target_size *= factor;
1785         } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1786
1787         return num_sizes;
1788 }
1789
1790 /* only valid during minor collections */
1791 static mword old_num_major_sections;
1792
1793 static void
1794 major_start_nursery_collection (void)
1795 {
1796 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1797         consistency_check ();
1798 #endif
1799
1800         old_num_major_sections = num_major_sections;
1801 }
1802
1803 static void
1804 major_finish_nursery_collection (void)
1805 {
1806 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1807         consistency_check ();
1808 #endif
1809 }
1810
1811 static int
1812 block_usage_comparer (const void *bl1, const void *bl2)
1813 {
1814         const gint16 nused1 = (*(MSBlockInfo**)bl1)->nused;
1815         const gint16 nused2 = (*(MSBlockInfo**)bl2)->nused;
1816
1817         return nused2 - nused1;
1818 }
1819
1820 static void
1821 sgen_evacuation_freelist_blocks (MSBlockInfo * volatile *block_list, int size_index)
1822 {
1823         MSBlockInfo **evacuated_blocks;
1824         size_t index = 0, count, num_blocks = 0, num_used = 0;
1825         MSBlockInfo *info;
1826         MSBlockInfo * volatile *prev;
1827
1828         for (info = *block_list; info != NULL; info = info->next_free) {
1829                 num_blocks++;
1830                 num_used += info->nused;
1831         }
1832
1833         /*
1834          * We have a set of blocks in the freelist which will be evacuated. Instead
1835          * of evacuating all of the blocks into new ones, we traverse the freelist
1836          * sorting it by the number of occupied slots, evacuating the objects from
1837          * blocks with fewer used slots into fuller blocks.
1838          *
1839          * The number of used slots is set at the end of the previous sweep. Since
1840          * we sequentially unlink slots from blocks, except for the head of the
1841          * freelist, for blocks on the freelist, the number of used slots is the same
1842          * as at the end of the previous sweep.
1843          */
1844         evacuated_blocks = (MSBlockInfo**)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_blocks, INTERNAL_MEM_TEMPORARY, TRUE);
1845
1846         for (info = *block_list; info != NULL; info = info->next_free) {
1847                 evacuated_blocks [index++] = info;
1848         }
1849
1850         SGEN_ASSERT (0, num_blocks == index, "Why did the freelist change ?");
1851
1852         sgen_qsort (evacuated_blocks, num_blocks, sizeof (gpointer), block_usage_comparer);
1853
1854         /*
1855          * Form a new freelist with the fullest blocks. These blocks will also be
1856          * marked as to_space so we don't evacuate from them.
1857          */
1858         count = MS_BLOCK_FREE / block_obj_sizes [size_index];
1859         prev = block_list;
1860         for (index = 0; index < (num_used + count - 1) / count; index++) {
1861                 SGEN_ASSERT (0, index < num_blocks, "Why do we need more blocks for compaction than we already had ?");
1862                 info = evacuated_blocks [index];
1863                 info->is_to_space = TRUE;
1864                 *prev = info;
1865                 prev = &info->next_free;
1866         }
1867         *prev = NULL;
1868
1869         sgen_free_internal_dynamic (evacuated_blocks, sizeof (MSBlockInfo*) * num_blocks, INTERNAL_MEM_TEMPORARY);
1870 }
1871
1872 static void
1873 major_start_major_collection (void)
1874 {
1875         MSBlockInfo *block;
1876         int i;
1877
1878         major_finish_sweep_checking ();
1879
1880         /*
1881          * Clear the free lists for block sizes where we do evacuation.  For those block
1882          * sizes we will have to allocate new blocks.
1883          */
1884         for (i = 0; i < num_block_obj_sizes; ++i) {
1885                 if (!evacuate_block_obj_sizes [i])
1886                         continue;
1887
1888                 binary_protocol_evacuating_blocks (block_obj_sizes [i]);
1889
1890                 sgen_evacuation_freelist_blocks (&free_block_lists [0][i], i);
1891                 sgen_evacuation_freelist_blocks (&free_block_lists [MS_BLOCK_FLAG_REFS][i], i);
1892         }
1893
1894         if (lazy_sweep && concurrent_sweep) {
1895                 /*
1896                  * sweep_blocks_job is created before sweep_finish, which we wait for above
1897                  * (major_finish_sweep_checking). After the end of sweep, if we don't have
1898                  * sweep_blocks_job set, it means that it has already been run.
1899                  */
1900                 SgenThreadPoolJob *job = sweep_blocks_job;
1901                 if (job)
1902                         sgen_thread_pool_job_wait (job);
1903         }
1904
1905         if (lazy_sweep && !concurrent_sweep)
1906                 binary_protocol_sweep_begin (GENERATION_OLD, TRUE);
1907         /* Sweep all unswept blocks and set them to MARKING */
1908         FOREACH_BLOCK_NO_LOCK (block) {
1909                 if (lazy_sweep && !concurrent_sweep)
1910                         sweep_block (block);
1911                 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1912                 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1913                 /*
1914                  * Swept blocks that have a null free_list are full. Evacuation is not
1915                  * effective on these blocks since we expect them to have high usage anyway,
1916                  * given that the survival rate for majors is relatively high.
1917                  */
1918                 if (evacuate_block_obj_sizes [block->obj_size_index] && !block->free_list)
1919                         block->is_to_space = TRUE;
1920         } END_FOREACH_BLOCK_NO_LOCK;
1921         if (lazy_sweep && !concurrent_sweep)
1922                 binary_protocol_sweep_end (GENERATION_OLD, TRUE);
1923
1924         set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1925 }
1926
1927 static void
1928 major_finish_major_collection (ScannedObjectCounts *counts)
1929 {
1930 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1931         if (binary_protocol_is_enabled ()) {
1932                 counts->num_scanned_objects = scanned_objects_list.next_slot;
1933
1934                 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1935                 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1936
1937                 sgen_pointer_queue_clear (&scanned_objects_list);
1938         }
1939 #endif
1940 }
1941
1942 static int
1943 compare_pointers (const void *va, const void *vb) {
1944         char *a = *(char**)va, *b = *(char**)vb;
1945         if (a < b)
1946                 return -1;
1947         if (a > b)
1948                 return 1;
1949         return 0;
1950 }
1951
1952 /*
1953  * This is called with sweep completed and the world stopped.
1954  */
1955 static void
1956 major_free_swept_blocks (size_t section_reserve)
1957 {
1958         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1959
1960 #ifdef TARGET_WIN32
1961                 /*
1962                  * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1963                  * a VirtualAlloc ()-ed block.
1964                  */
1965                 return;
1966 #endif
1967
1968         {
1969                 int i, num_empty_blocks_orig, num_blocks, arr_length;
1970                 void *block;
1971                 void **empty_block_arr;
1972                 void **rebuild_next;
1973
1974                 if (num_empty_blocks <= section_reserve)
1975                         return;
1976                 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1977
1978                 num_empty_blocks_orig = num_empty_blocks;
1979                 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1980                                 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1981                 if (!empty_block_arr)
1982                         goto fallback;
1983
1984                 i = 0;
1985                 for (block = empty_blocks; block; block = *(void**)block)
1986                         empty_block_arr [i++] = block;
1987                 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1988
1989                 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1990
1991                 /*
1992                  * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1993                  * contiguous ones.  If we do, we free them.  If that's not enough to get to
1994                  * section_reserve, we halve the number of contiguous blocks we're looking
1995                  * for and have another go, until we're done with looking for pairs of
1996                  * blocks, at which point we give up and go to the fallback.
1997                  */
1998                 arr_length = num_empty_blocks_orig;
1999                 num_blocks = MS_BLOCK_ALLOC_NUM;
2000                 while (num_empty_blocks > section_reserve && num_blocks > 1) {
2001                         int first = -1;
2002                         int dest = 0;
2003
2004                         dest = 0;
2005                         for (i = 0; i < arr_length; ++i) {
2006                                 int d = dest;
2007                                 void *block = empty_block_arr [i];
2008                                 SGEN_ASSERT (6, block, "we're not shifting correctly");
2009                                 if (i != dest) {
2010                                         empty_block_arr [dest] = block;
2011                                         /*
2012                                          * This is not strictly necessary, but we're
2013                                          * cautious.
2014                                          */
2015                                         empty_block_arr [i] = NULL;
2016                                 }
2017                                 ++dest;
2018
2019                                 if (first < 0) {
2020                                         first = d;
2021                                         continue;
2022                                 }
2023
2024                                 SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
2025
2026                                 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
2027                                         first = d;
2028                                         continue;
2029                                 }
2030
2031                                 if (d + 1 - first == num_blocks) {
2032                                         /*
2033                                          * We found num_blocks contiguous blocks.  Free them
2034                                          * and null their array entries.  As an optimization
2035                                          * we could, instead of nulling the entries, shift
2036                                          * the following entries over to the left, while
2037                                          * we're iterating.
2038                                          */
2039                                         int j;
2040                                         sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
2041                                         for (j = first; j <= d; ++j)
2042                                                 empty_block_arr [j] = NULL;
2043                                         dest = first;
2044                                         first = -1;
2045
2046                                         num_empty_blocks -= num_blocks;
2047
2048                                         stat_major_blocks_freed += num_blocks;
2049                                         if (num_blocks == MS_BLOCK_ALLOC_NUM)
2050                                                 stat_major_blocks_freed_ideal += num_blocks;
2051                                         else
2052                                                 stat_major_blocks_freed_less_ideal += num_blocks;
2053
2054                                 }
2055                         }
2056
2057                         SGEN_ASSERT (6, dest <= i && dest <= arr_length, "array length is off");
2058                         arr_length = dest;
2059                         SGEN_ASSERT (6, arr_length == num_empty_blocks, "array length is off");
2060
2061                         num_blocks >>= 1;
2062                 }
2063
2064                 /* rebuild empty_blocks free list */
2065                 rebuild_next = (void**)&empty_blocks;
2066                 for (i = 0; i < arr_length; ++i) {
2067                         void *block = empty_block_arr [i];
2068                         SGEN_ASSERT (6, block, "we're missing blocks");
2069                         *rebuild_next = block;
2070                         rebuild_next = (void**)block;
2071                 }
2072                 *rebuild_next = NULL;
2073
2074                 /* free array */
2075                 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
2076         }
2077
2078         SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
2079
2080  fallback:
2081         /*
2082          * This is our threshold.  If there's not more empty than used blocks, we won't
2083          * release uncontiguous blocks, in fear of fragmenting the address space.
2084          */
2085         if (num_empty_blocks <= num_major_sections)
2086                 return;
2087
2088         while (num_empty_blocks > section_reserve) {
2089                 void *next = *(void**)empty_blocks;
2090                 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP);
2091                 empty_blocks = next;
2092                 /*
2093                  * Needs not be atomic because this is running
2094                  * single-threaded.
2095                  */
2096                 --num_empty_blocks;
2097
2098                 ++stat_major_blocks_freed;
2099                 ++stat_major_blocks_freed_individual;
2100         }
2101 }
2102
2103 static void
2104 major_pin_objects (SgenGrayQueue *queue)
2105 {
2106         MSBlockInfo *block;
2107
2108         FOREACH_BLOCK_NO_LOCK (block) {
2109                 size_t first_entry, last_entry;
2110                 SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
2111                 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
2112                                 &first_entry, &last_entry);
2113                 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
2114         } END_FOREACH_BLOCK_NO_LOCK;
2115 }
2116
2117 static void
2118 major_init_to_space (void)
2119 {
2120 }
2121
2122 static void
2123 major_report_pinned_memory_usage (void)
2124 {
2125         g_assert_not_reached ();
2126 }
2127
2128 static gint64
2129 major_get_used_size (void)
2130 {
2131         gint64 size = 0;
2132         MSBlockInfo *block;
2133
2134         /*
2135          * We're holding the GC lock, but the sweep thread might be running.  Make sure it's
2136          * finished, then we can iterate over the block array.
2137          */
2138         major_finish_sweep_checking ();
2139
2140         FOREACH_BLOCK_NO_LOCK (block) {
2141                 int count = MS_BLOCK_FREE / block->obj_size;
2142                 void **iter;
2143                 size += count * block->obj_size;
2144                 for (iter = block->free_list; iter; iter = (void**)*iter)
2145                         size -= block->obj_size;
2146         } END_FOREACH_BLOCK_NO_LOCK;
2147
2148         return size;
2149 }
2150
2151 /* FIXME: return number of bytes, not of sections */
2152 static size_t
2153 get_num_major_sections (void)
2154 {
2155         return num_major_sections;
2156 }
2157
2158 /*
2159  * Returns the number of bytes in blocks that were present when the last sweep was
2160  * initiated, and were not freed during the sweep.  They are the basis for calculating the
2161  * allowance.
2162  */
2163 static size_t
2164 get_bytes_survived_last_sweep (void)
2165 {
2166         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2167         return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
2168 }
2169
2170 static gboolean
2171 major_handle_gc_param (const char *opt)
2172 {
2173         if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2174                 const char *arg = strchr (opt, '=') + 1;
2175                 int percentage = atoi (arg);
2176                 if (percentage < 0 || percentage > 100) {
2177                         fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2178                         exit (1);
2179                 }
2180                 evacuation_threshold = (float)percentage / 100.0f;
2181                 return TRUE;
2182         } else if (!strcmp (opt, "lazy-sweep")) {
2183                 lazy_sweep = TRUE;
2184                 return TRUE;
2185         } else if (!strcmp (opt, "no-lazy-sweep")) {
2186                 lazy_sweep = FALSE;
2187                 return TRUE;
2188         } else if (!strcmp (opt, "concurrent-sweep")) {
2189                 concurrent_sweep = TRUE;
2190                 return TRUE;
2191         } else if (!strcmp (opt, "no-concurrent-sweep")) {
2192                 concurrent_sweep = FALSE;
2193                 return TRUE;
2194         }
2195
2196         return FALSE;
2197 }
2198
2199 static void
2200 major_print_gc_param_usage (void)
2201 {
2202         fprintf (stderr,
2203                         ""
2204                         "  evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2205                         "  (no-)lazy-sweep\n"
2206                         "  (no-)concurrent-sweep\n"
2207                         );
2208 }
2209
2210 /*
2211  * This callback is used to clear cards, move cards to the shadow table and do counting.
2212  */
2213 static void
2214 major_iterate_block_ranges (sgen_cardtable_block_callback callback)
2215 {
2216         MSBlockInfo *block;
2217         gboolean has_references;
2218
2219         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2220                 if (has_references)
2221                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2222         } END_FOREACH_BLOCK_NO_LOCK;
2223 }
2224
2225 static void
2226 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
2227 {
2228         MSBlockInfo *block;
2229         gboolean has_references;
2230
2231         major_finish_sweep_checking ();
2232         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2233                 if (has_references)
2234                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2235         } END_FOREACH_BLOCK_NO_LOCK;
2236 }
2237
2238 #ifdef HEAVY_STATISTICS
2239 extern guint64 marked_cards;
2240 extern guint64 scanned_cards;
2241 extern guint64 scanned_objects;
2242 extern guint64 remarked_cards;
2243 #endif
2244
2245 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2246 /*
2247  * MS blocks are 16K aligned.
2248  * Cardtables are 4K aligned, at least.
2249  * This means that the cardtable of a given block is 32 bytes aligned.
2250  */
2251 static guint8*
2252 initial_skip_card (guint8 *card_data)
2253 {
2254         mword *cards = (mword*)card_data;
2255         mword card;
2256         int i;
2257         for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2258                 card = cards [i];
2259                 if (card)
2260                         break;
2261         }
2262
2263         if (i == CARD_WORDS_PER_BLOCK)
2264                 return card_data + CARDS_PER_BLOCK;
2265
2266 #if defined(__i386__) && defined(__GNUC__)
2267         return card_data + i * 4 +  (__builtin_ffs (card) - 1) / 8;
2268 #elif defined(__x86_64__) && defined(__GNUC__)
2269         return card_data + i * 8 +  (__builtin_ffsll (card) - 1) / 8;
2270 #elif defined(__s390x__) && defined(__GNUC__)
2271         return card_data + i * 8 +  (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2272 #else
2273         for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2274                 if (card_data [i])
2275                         return &card_data [i];
2276         }
2277         return card_data;
2278 #endif
2279 }
2280
2281 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2282 #define MS_BLOCK_OBJ_FAST(b,os,i)                       ((b) + MS_BLOCK_SKIP + (os) * (i))
2283 #define MS_OBJ_ALLOCED_FAST(o,b)                (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2284
2285 static void
2286 scan_card_table_for_block (MSBlockInfo *block, CardTableScanType scan_type, ScanCopyContext ctx)
2287 {
2288         SgenGrayQueue *queue = ctx.queue;
2289         ScanObjectFunc scan_func = ctx.ops->scan_object;
2290 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2291         guint8 cards_copy [CARDS_PER_BLOCK];
2292 #endif
2293         guint8 cards_preclean [CARDS_PER_BLOCK];
2294         gboolean small_objects;
2295         int block_obj_size;
2296         char *block_start;
2297         guint8 *card_data, *card_base;
2298         guint8 *card_data_end;
2299         char *scan_front = NULL;
2300
2301         /* The concurrent mark doesn't enter evacuating blocks */
2302         if (scan_type == CARDTABLE_SCAN_MOD_UNION_PRECLEAN && major_block_is_evacuating (block))
2303                 return;
2304
2305         block_obj_size = block->obj_size;
2306         small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2307
2308         block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2309
2310         /*
2311          * This is safe in face of card aliasing for the following reason:
2312          *
2313          * Major blocks are 16k aligned, or 32 cards aligned.
2314          * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2315          * sizes, they won't overflow the cardtable overlap modulus.
2316          */
2317         if (scan_type & CARDTABLE_SCAN_MOD_UNION) {
2318                 card_data = card_base = block->cardtable_mod_union;
2319                 /*
2320                  * This happens when the nursery collection that precedes finishing
2321                  * the concurrent collection allocates new major blocks.
2322                  */
2323                 if (!card_data)
2324                         return;
2325
2326                 if (scan_type == CARDTABLE_SCAN_MOD_UNION_PRECLEAN) {
2327                         sgen_card_table_preclean_mod_union (card_data, cards_preclean, CARDS_PER_BLOCK);
2328                         card_data = card_base = cards_preclean;
2329                 }
2330         } else {
2331 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2332                 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2333 #else
2334                 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2335                         return;
2336                 card_data = card_base = cards_copy;
2337 #endif
2338         }
2339         card_data_end = card_data + CARDS_PER_BLOCK;
2340
2341         card_data += MS_BLOCK_SKIP >> CARD_BITS;
2342
2343         card_data = initial_skip_card (card_data);
2344         while (card_data < card_data_end) {
2345                 size_t card_index, first_object_index;
2346                 char *start;
2347                 char *end;
2348                 char *first_obj, *obj;
2349
2350                 HEAVY_STAT (++scanned_cards);
2351
2352                 if (!*card_data) {
2353                         ++card_data;
2354                         continue;
2355                 }
2356
2357                 card_index = card_data - card_base;
2358                 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2359                 end = start + CARD_SIZE_IN_BYTES;
2360
2361                 if (!block_is_swept_or_marking (block))
2362                         sweep_block (block);
2363
2364                 HEAVY_STAT (++marked_cards);
2365
2366                 if (small_objects)
2367                         sgen_card_table_prepare_card_for_scanning (card_data);
2368
2369                 /*
2370                  * If the card we're looking at starts at or in the block header, we
2371                  * must start at the first object in the block, without calculating
2372                  * the index of the object we're hypothetically starting at, because
2373                  * it would be negative.
2374                  */
2375                 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2376                         first_object_index = 0;
2377                 else
2378                         first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2379
2380                 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2381
2382                 binary_protocol_card_scan (first_obj, end - first_obj);
2383
2384                 while (obj < end) {
2385                         if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2386                                 goto next_object;
2387
2388                         if (scan_type & CARDTABLE_SCAN_MOD_UNION) {
2389                                 /* FIXME: do this more efficiently */
2390                                 int w, b;
2391                                 MS_CALC_MARK_BIT (w, b, obj);
2392                                 if (!MS_MARK_BIT (block, w, b))
2393                                         goto next_object;
2394                         }
2395
2396                         GCObject *object = (GCObject*)obj;
2397
2398                         if (small_objects) {
2399                                 HEAVY_STAT (++scanned_objects);
2400                                 scan_func (object, sgen_obj_get_descriptor (object), queue);
2401                         } else {
2402                                 size_t offset = sgen_card_table_get_card_offset (obj, block_start);
2403                                 sgen_cardtable_scan_object (object, block_obj_size, card_base + offset, ctx);
2404                         }
2405                 next_object:
2406                         obj += block_obj_size;
2407                         g_assert (scan_front <= obj);
2408                         scan_front = obj;
2409                 }
2410
2411                 HEAVY_STAT (if (*card_data) ++remarked_cards);
2412
2413                 if (small_objects)
2414                         ++card_data;
2415                 else
2416                         card_data = card_base + sgen_card_table_get_card_offset (obj, block_start);
2417         }
2418 }
2419
2420 static void
2421 major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx)
2422 {
2423         MSBlockInfo *block;
2424         gboolean has_references, was_sweeping, skip_scan;
2425
2426         if (!concurrent_mark)
2427                 g_assert (scan_type == CARDTABLE_SCAN_GLOBAL);
2428
2429         if (scan_type != CARDTABLE_SCAN_GLOBAL)
2430                 SGEN_ASSERT (0, !sweep_in_progress (), "Sweep should be finished when we scan mod union card table");
2431         was_sweeping = sweep_in_progress ();
2432
2433         binary_protocol_major_card_table_scan_start (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
2434         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2435 #ifdef PREFETCH_CARDS
2436                 int prefetch_index = __index + 6;
2437                 if (prefetch_index < allocated_blocks.next_slot) {
2438                         MSBlockInfo *prefetch_block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, prefetch_index));
2439                         PREFETCH_READ (prefetch_block);
2440                         if (scan_type == CARDTABLE_SCAN_GLOBAL) {
2441                                 guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2442                                 PREFETCH_WRITE (prefetch_cards);
2443                                 PREFETCH_WRITE (prefetch_cards + 32);
2444                         }
2445                 }
2446 #endif
2447
2448                 if (!has_references)
2449                         continue;
2450                 skip_scan = FALSE;
2451
2452                 if (scan_type == CARDTABLE_SCAN_GLOBAL) {
2453                         gpointer *card_start = (gpointer*) sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (block));
2454                         gboolean has_dirty_cards = FALSE;
2455                         int i;
2456                         for (i = 0; i < CARDS_PER_BLOCK / sizeof(gpointer); i++) {
2457                                 if (card_start [i]) {
2458                                         has_dirty_cards = TRUE;
2459                                         break;
2460                                 }
2461                         }
2462                         if (!has_dirty_cards) {
2463                                 skip_scan = TRUE;
2464                         } else {
2465                                 /*
2466                                  * After the start of the concurrent collections, blocks change state
2467                                  * to marking. We should not sweep it in that case. We can't race with
2468                                  * sweep start since we are in a nursery collection. Also avoid CAS-ing
2469                                  */
2470                                 if (sweep_in_progress ()) {
2471                                         skip_scan = !ensure_block_is_checked_for_sweeping (__index, TRUE, NULL);
2472                                 } else if (was_sweeping) {
2473                                         /* Recheck in case sweep finished after dereferencing the slot */
2474                                         skip_scan = *sgen_array_list_get_slot (&allocated_blocks, __index) == 0;
2475                                 }
2476                         }
2477                 }
2478                 if (!skip_scan)
2479                         scan_card_table_for_block (block, scan_type, ctx);
2480         } END_FOREACH_BLOCK_NO_LOCK;
2481         binary_protocol_major_card_table_scan_end (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
2482 }
2483
2484 static void
2485 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2486 {
2487         MSBlockInfo *block;
2488         gboolean has_references;
2489         long long total_cards = 0;
2490         long long marked_cards = 0;
2491
2492         if (sweep_in_progress ()) {
2493                 *num_total_cards = -1;
2494                 *num_marked_cards = -1;
2495                 return;
2496         }
2497
2498         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2499                 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2500                 int i;
2501
2502                 if (!has_references)
2503                         continue;
2504
2505                 total_cards += CARDS_PER_BLOCK;
2506                 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2507                         if (cards [i])
2508                                 ++marked_cards;
2509                 }
2510         } END_FOREACH_BLOCK_NO_LOCK;
2511
2512         *num_total_cards = total_cards;
2513         *num_marked_cards = marked_cards;
2514 }
2515
2516 static void
2517 update_cardtable_mod_union (void)
2518 {
2519         MSBlockInfo *block;
2520
2521         FOREACH_BLOCK_NO_LOCK (block) {
2522                 gpointer *card_start = (gpointer*) sgen_card_table_get_card_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (block));
2523                 gboolean has_dirty_cards = FALSE;
2524                 int i;
2525                 for (i = 0; i < CARDS_PER_BLOCK / sizeof(gpointer); i++) {
2526                         if (card_start [i]) {
2527                                 has_dirty_cards = TRUE;
2528                                 break;
2529                         }
2530                 }
2531                 if (has_dirty_cards) {
2532                         size_t num_cards;
2533                         guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
2534                         sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2535                         SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2536                 }
2537         } END_FOREACH_BLOCK_NO_LOCK;
2538 }
2539
2540 #undef pthread_create
2541
2542 static void
2543 post_param_init (SgenMajorCollector *collector)
2544 {
2545         collector->sweeps_lazily = lazy_sweep;
2546         collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
2547 }
2548
2549 static void
2550 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
2551 {
2552         int i;
2553
2554         sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2555
2556         num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2557         block_obj_sizes = (int *)sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2558         ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2559
2560         evacuate_block_obj_sizes = (gboolean *)sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2561         for (i = 0; i < num_block_obj_sizes; ++i)
2562                 evacuate_block_obj_sizes [i] = FALSE;
2563
2564         sweep_slots_available = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2565         sweep_slots_used = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2566         sweep_num_blocks = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2567
2568         /*
2569         {
2570                 int i;
2571                 g_print ("block object sizes:\n");
2572                 for (i = 0; i < num_block_obj_sizes; ++i)
2573                         g_print ("%d\n", block_obj_sizes [i]);
2574         }
2575         */
2576
2577         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2578                 free_block_lists [i] = (MSBlockInfo *volatile *)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2579
2580         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2581                 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2582         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2583                 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2584
2585         mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2586         mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2587         mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2588         mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2589         mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2590         mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2591         mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2592
2593         collector->section_size = MAJOR_SECTION_SIZE;
2594
2595         concurrent_mark = is_concurrent;
2596         collector->is_concurrent = is_concurrent;
2597         collector->needs_thread_pool = is_concurrent || concurrent_sweep;
2598         collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2599         collector->supports_cardtable = TRUE;
2600
2601         collector->alloc_heap = major_alloc_heap;
2602         collector->is_object_live = major_is_object_live;
2603         collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2604         collector->alloc_degraded = major_alloc_degraded;
2605
2606         collector->alloc_object = major_alloc_object;
2607         collector->free_pinned_object = free_pinned_object;
2608         collector->iterate_objects = major_iterate_objects;
2609         collector->free_non_pinned_object = major_free_non_pinned_object;
2610         collector->pin_objects = major_pin_objects;
2611         collector->pin_major_object = pin_major_object;
2612         collector->scan_card_table = major_scan_card_table;
2613         collector->iterate_live_block_ranges = major_iterate_live_block_ranges;
2614         collector->iterate_block_ranges = major_iterate_block_ranges;
2615         if (is_concurrent) {
2616                 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2617                 collector->get_cardtable_mod_union_for_reference = major_get_cardtable_mod_union_for_reference;
2618         }
2619         collector->init_to_space = major_init_to_space;
2620         collector->sweep = major_sweep;
2621         collector->have_swept = major_have_swept;
2622         collector->finish_sweeping = major_finish_sweep_checking;
2623         collector->free_swept_blocks = major_free_swept_blocks;
2624         collector->check_scan_starts = major_check_scan_starts;
2625         collector->dump_heap = major_dump_heap;
2626         collector->get_used_size = major_get_used_size;
2627         collector->start_nursery_collection = major_start_nursery_collection;
2628         collector->finish_nursery_collection = major_finish_nursery_collection;
2629         collector->start_major_collection = major_start_major_collection;
2630         collector->finish_major_collection = major_finish_major_collection;
2631         collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2632         collector->ptr_is_from_pinned_alloc = ptr_is_from_pinned_alloc;
2633         collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2634         collector->get_num_major_sections = get_num_major_sections;
2635         collector->get_bytes_survived_last_sweep = get_bytes_survived_last_sweep;
2636         collector->handle_gc_param = major_handle_gc_param;
2637         collector->print_gc_param_usage = major_print_gc_param_usage;
2638         collector->post_param_init = post_param_init;
2639         collector->is_valid_object = major_is_valid_object;
2640         collector->describe_pointer = major_describe_pointer;
2641         collector->count_cards = major_count_cards;
2642
2643         collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
2644         collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
2645         collector->major_ops_serial.drain_gray_stack = drain_gray_stack;
2646         if (is_concurrent) {
2647                 collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2648                 collector->major_ops_concurrent_start.scan_object = major_scan_object_concurrent_with_evacuation;
2649                 collector->major_ops_concurrent_start.scan_vtype = major_scan_vtype_concurrent_with_evacuation;
2650                 collector->major_ops_concurrent_start.scan_ptr_field = major_scan_ptr_field_concurrent_with_evacuation;
2651                 collector->major_ops_concurrent_start.drain_gray_stack = drain_gray_stack_concurrent;
2652
2653                 collector->major_ops_concurrent_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_finish_canonical;
2654                 collector->major_ops_concurrent_finish.scan_object = major_scan_object_with_evacuation;
2655                 collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_with_evacuation;
2656                 collector->major_ops_concurrent_finish.scan_ptr_field = major_scan_ptr_field_with_evacuation;
2657                 collector->major_ops_concurrent_finish.drain_gray_stack = drain_gray_stack;
2658         }
2659
2660 #ifdef HEAVY_STATISTICS
2661         mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2662         mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2663         mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2664         mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2665         mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2666         mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2667         mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2668         mono_counters_register ("Optimized copy major small evacuate", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_evacuate);
2669         mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2670         mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2671         mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2672
2673         mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2674         mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2675         mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2676 #endif
2677
2678 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2679         mono_os_mutex_init (&scanned_objects_list_lock);
2680 #endif
2681
2682         SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2683
2684         /*cardtable requires major pages to be 8 cards aligned*/
2685         g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2686 }
2687
2688 void
2689 sgen_marksweep_init (SgenMajorCollector *collector)
2690 {
2691         sgen_marksweep_init_internal (collector, FALSE);
2692 }
2693
2694 void
2695 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2696 {
2697         sgen_marksweep_init_internal (collector, TRUE);
2698 }
2699
2700 #endif