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 #if SIZEOF_VOID_P != 8
 244 static guint64 stat_major_blocks_freed_ideal = 0;
 245 static guint64 stat_major_blocks_freed_less_ideal = 0;
 246 static guint64 stat_major_blocks_freed_individual = 0;
 247 static guint64 stat_major_blocks_alloced_less_ideal = 0;
 248 #endif
 249
 250 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
 251 static guint64 num_major_objects_marked = 0;
 252 #define INC_NUM_MAJOR_OBJECTS_MARKED()  (++num_major_objects_marked)
 253 #else
 254 #define INC_NUM_MAJOR_OBJECTS_MARKED()
 255 #endif
 256
 257 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
 258 static mono_mutex_t scanned_objects_list_lock;
 259 static SgenPointerQueue scanned_objects_list;
 260
 261 static void
 262 add_scanned_object (void *ptr)
 263 {
 264         if (!binary_protocol_is_enabled ())
 265                 return;
 266
 267         mono_os_mutex_lock (&scanned_objects_list_lock);
 268         sgen_pointer_queue_add (&scanned_objects_list, ptr);
 269         mono_os_mutex_unlock (&scanned_objects_list_lock);
 270 }
 271 #endif
 272
 273 static gboolean sweep_block (MSBlockInfo *block);
 274
 275 static int
 276 ms_find_block_obj_size_index (size_t size)
 277 {
 278         int i;
 279         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);
 280         for (i = 0; i < num_block_obj_sizes; ++i)
 281                 if (block_obj_sizes [i] >= size)
 282                         return i;
 283         g_error ("no object of size %zd\n", size);
 284         return -1;
 285 }
 286
 287 #define FREE_BLOCKS_FROM(lists,p,r)     (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)])
 288 #define FREE_BLOCKS(p,r)                (FREE_BLOCKS_FROM (free_block_lists, (p), (r)))
 289
 290 #define MS_BLOCK_OBJ_SIZE_INDEX(s)                              \
 291         (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ?      \
 292          fast_block_obj_size_indexes [((s)+7)>>3] :             \
 293          ms_find_block_obj_size_index ((s)))
 294
 295 static void*
 296 major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits)
 297 {
 298         char *start;
 299         if (nursery_align)
 300                 start = (char *)sgen_alloc_os_memory_aligned (nursery_size, nursery_align, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery");
 301         else
 302                 start = (char *)sgen_alloc_os_memory (nursery_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery");
 303
 304         return start;
 305 }
 306
 307 static void
 308 update_heap_boundaries_for_block (MSBlockInfo *block)
 309 {
 310         sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE);
 311 }
 312
 313 /*
 314  * Thread safe
 315  */
 316 static void*
 317 ms_get_empty_block (void)
 318 {
 319         char *p;
 320         int i;
 321         void *block, *empty, *next;
 322
 323  retry:
 324         if (!empty_blocks) {
 325                 /*
 326                  * We try allocating MS_BLOCK_ALLOC_NUM blocks first.  If that's
 327                  * unsuccessful, we halve the number of blocks and try again, until we're at
 328                  * 1.  If that doesn't work, either, we assert.
 329                  */
 330                 int alloc_num = MS_BLOCK_ALLOC_NUM;
 331                 for (;;) {
 332                         p = (char *)sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE,
 333                                 (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE),
 334                                 alloc_num == 1 ? "major heap section" : NULL);
 335                         if (p)
 336                                 break;
 337                         alloc_num >>= 1;
 338                 }
 339
 340                 for (i = 0; i < alloc_num; ++i) {
 341                         block = p;
 342                         /*
 343                          * We do the free list update one after the
 344                          * other so that other threads can use the new
 345                          * blocks as quickly as possible.
 346                          */
 347                         do {
 348                                 empty = empty_blocks;
 349                                 *(void**)block = empty;
 350                         } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
 351                         p += MS_BLOCK_SIZE;
 352                 }
 353
 354                 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
 355
 356                 stat_major_blocks_alloced += alloc_num;
 357 #if SIZEOF_VOID_P != 8
 358                 if (alloc_num != MS_BLOCK_ALLOC_NUM)
 359                         stat_major_blocks_alloced_less_ideal += alloc_num;
 360 #endif
 361         }
 362
 363         do {
 364                 empty = empty_blocks;
 365                 if (!empty)
 366                         goto retry;
 367                 block = empty;
 368                 next = *(void**)block;
 369         } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
 370
 371         SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
 372
 373         *(void**)block = NULL;
 374
 375         g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
 376
 377         return block;
 378 }
 379
 380 /*
 381  * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
 382  * list, where it will either be freed later on, or reused in nursery collections.
 383  */
 384 static void
 385 ms_free_block (MSBlockInfo *info)
 386 {
 387         void *empty;
 388         char *block = MS_BLOCK_FOR_BLOCK_INFO (info);
 389
 390         sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
 391         if (info->cardtable_mod_union)
 392                 sgen_card_table_free_mod_union (info->cardtable_mod_union, block, MS_BLOCK_SIZE);
 393         memset (block, 0, MS_BLOCK_SIZE);
 394
 395         do {
 396                 empty = empty_blocks;
 397                 *(void**)block = empty;
 398         } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
 399
 400         SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
 401
 402         binary_protocol_block_free (block, MS_BLOCK_SIZE);
 403 }
 404
 405 static gboolean
 406 sweep_in_progress (void)
 407 {
 408         int state = sweep_state;
 409         return state == SWEEP_STATE_SWEEPING ||
 410                 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
 411                 state == SWEEP_STATE_COMPACTING;
 412 }
 413
 414 static inline gboolean
 415 block_is_swept_or_marking (MSBlockInfo *block)
 416 {
 417         gint32 state = block->state;
 418         return state == BLOCK_STATE_SWEPT || state == BLOCK_STATE_MARKING;
 419 }
 420
 421 //#define MARKSWEEP_CONSISTENCY_CHECK
 422
 423 #ifdef MARKSWEEP_CONSISTENCY_CHECK
 424 static void
 425 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
 426 {
 427         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 428         for (; block; block = block->next_free) {
 429                 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
 430                 g_assert (block->obj_size == size);
 431                 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
 432
 433                 /* blocks in the free lists must have at least
 434                    one free slot */
 435                 g_assert (block->free_list);
 436
 437                 /* the block must be in the allocated_blocks array */
 438                 g_assert (sgen_array_list_find (&allocated_blocks, BLOCK_TAG (block)) != (guint32)-1);
 439         }
 440 }
 441
 442 static void
 443 check_empty_blocks (void)
 444 {
 445         void *p;
 446         size_t i = 0;
 447         for (p = empty_blocks; p; p = *(void**)p)
 448                 ++i;
 449         g_assert (i == num_empty_blocks);
 450 }
 451
 452 static void
 453 consistency_check (void)
 454 {
 455         MSBlockInfo *block;
 456         int i;
 457
 458         /* check all blocks */
 459         FOREACH_BLOCK_NO_LOCK (block) {
 460                 int count = MS_BLOCK_FREE / block->obj_size;
 461                 int num_free = 0;
 462                 void **free;
 463
 464                 /* count number of free slots */
 465                 for (i = 0; i < count; ++i) {
 466                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 467                         if (!MS_OBJ_ALLOCED (obj, block))
 468                                 ++num_free;
 469                 }
 470
 471                 /* check free list */
 472                 for (free = block->free_list; free; free = (void**)*free) {
 473                         g_assert (MS_BLOCK_FOR_OBJ (free) == block);
 474                         --num_free;
 475                 }
 476                 g_assert (num_free == 0);
 477
 478                 /* check all mark words are zero */
 479                 if (!sgen_concurrent_collection_in_progress () && block_is_swept_or_marking (block)) {
 480                         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
 481                                 g_assert (block->mark_words [i] == 0);
 482                 }
 483         } END_FOREACH_BLOCK_NO_LOCK;
 484
 485         /* check free blocks */
 486         for (i = 0; i < num_block_obj_sizes; ++i) {
 487                 int j;
 488                 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
 489                         check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
 490         }
 491
 492         check_empty_blocks ();
 493 }
 494 #endif
 495
 496 static void
 497 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
 498 {
 499         MSBlockInfo *old;
 500         do {
 501                 block->next_free = old = free_blocks [size_index];
 502         } while (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], block, old) != old);
 503 }
 504
 505 static void major_finish_sweep_checking (void);
 506
 507 static gboolean
 508 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
 509 {
 510         int size = block_obj_sizes [size_index];
 511         int count = MS_BLOCK_FREE / size;
 512         MSBlockInfo *info;
 513         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 514         char *obj_start;
 515         int i;
 516
 517         if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
 518                 return FALSE;
 519
 520         info = (MSBlockInfo*)ms_get_empty_block ();
 521
 522         SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
 523
 524         info->obj_size = size;
 525         info->obj_size_index = size_index;
 526         info->pinned = pinned;
 527         info->has_references = has_references;
 528         info->has_pinned = pinned;
 529         /*
 530          * Blocks that are to-space are not evacuated from.  During an major collection
 531          * blocks are allocated for two reasons: evacuating objects from the nursery and
 532          * evacuating them from major blocks marked for evacuation.  In both cases we don't
 533          * want further evacuation. We also don't want to evacuate objects allocated during
 534          * the concurrent mark since it would add pointless stress on the finishing pause.
 535          */
 536         info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD) || sgen_concurrent_collection_in_progress ();
 537         info->state = info->is_to_space ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
 538         SGEN_ASSERT (6, !sweep_in_progress () || info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
 539         info->cardtable_mod_union = NULL;
 540
 541         update_heap_boundaries_for_block (info);
 542
 543         binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
 544
 545         /* build free list */
 546         obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
 547         info->free_list = (void**)obj_start;
 548         /* we're skipping the last one - it must be nulled */
 549         for (i = 0; i < count - 1; ++i) {
 550                 char *next_obj_start = obj_start + size;
 551                 *(void**)obj_start = next_obj_start;
 552                 obj_start = next_obj_start;
 553         }
 554         /* the last one */
 555         *(void**)obj_start = NULL;
 556
 557         add_free_block (free_blocks, size_index, info);
 558
 559         sgen_array_list_add (&allocated_blocks, BLOCK_TAG (info), 0, FALSE);
 560
 561         SGEN_ATOMIC_ADD_P (num_major_sections, 1);
 562         return TRUE;
 563 }
 564
 565 static gboolean
 566 ptr_is_in_major_block (char *ptr, char **start, gboolean *pinned)
 567 {
 568         MSBlockInfo *block;
 569
 570         FOREACH_BLOCK_NO_LOCK (block) {
 571                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
 572                         int count = MS_BLOCK_FREE / block->obj_size;
 573                         int i;
 574
 575                         if (start)
 576                                 *start = NULL;
 577                         for (i = 0; i <= count; ++i) {
 578                                 if (ptr >= (char*)MS_BLOCK_OBJ (block, i) && ptr < (char*)MS_BLOCK_OBJ (block, i + 1)) {
 579                                         if (start)
 580                                                 *start = (char *)MS_BLOCK_OBJ (block, i);
 581                                         break;
 582                                 }
 583                         }
 584                         if (pinned)
 585                                 *pinned = block->pinned;
 586                         return TRUE;
 587                 }
 588         } END_FOREACH_BLOCK_NO_LOCK;
 589         return FALSE;
 590 }
 591
 592 static gboolean
 593 ptr_is_from_pinned_alloc (char *ptr)
 594 {
 595         gboolean pinned;
 596         if (ptr_is_in_major_block (ptr, NULL, &pinned))
 597                 return pinned;
 598         return FALSE;
 599 }
 600
 601 static void
 602 ensure_can_access_block_free_list (MSBlockInfo *block)
 603 {
 604  retry:
 605         for (;;) {
 606                 switch (block->state) {
 607                 case BLOCK_STATE_SWEPT:
 608                 case BLOCK_STATE_MARKING:
 609                         return;
 610                 case BLOCK_STATE_CHECKING:
 611                         SGEN_ASSERT (0, FALSE, "How did we get a block that's being checked from a free list?");
 612                         break;
 613                 case BLOCK_STATE_NEED_SWEEPING:
 614                         if (sweep_block (block))
 615                                 ++stat_major_blocks_lazy_swept;
 616                         break;
 617                 case BLOCK_STATE_SWEEPING:
 618                         /* FIXME: do this more elegantly */
 619                         g_usleep (100);
 620                         goto retry;
 621                 default:
 622                         SGEN_ASSERT (0, FALSE, "Illegal block state");
 623                         break;
 624                 }
 625         }
 626 }
 627
 628 static void*
 629 unlink_slot_from_free_list_uncontested (MSBlockInfo * volatile *free_blocks, int size_index)
 630 {
 631         MSBlockInfo *block, *next_free_block;
 632         void *obj, *next_free_slot;
 633
 634  retry:
 635         block = free_blocks [size_index];
 636         SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
 637
 638         ensure_can_access_block_free_list (block);
 639
 640         obj = block->free_list;
 641         SGEN_ASSERT (6, obj, "block %p in free list had no available object to alloc from", block);
 642
 643         next_free_slot = *(void**)obj;
 644         if (next_free_slot) {
 645                 block->free_list = (gpointer *)next_free_slot;
 646                 return obj;
 647         }
 648
 649         next_free_block = block->next_free;
 650         if (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], next_free_block, block) != block)
 651                 goto retry;
 652
 653         block->free_list = NULL;
 654         block->next_free = NULL;
 655
 656         return obj;
 657 }
 658
 659 static GCObject*
 660 alloc_obj (GCVTable vtable, size_t size, gboolean pinned, gboolean has_references)
 661 {
 662         int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 663         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 664         void *obj;
 665
 666         if (!free_blocks [size_index]) {
 667                 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
 668                         return NULL;
 669         }
 670
 671         obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
 672
 673         /* FIXME: assumes object layout */
 674         *(GCVTable*)obj = vtable;
 675
 676         total_allocated_major += block_obj_sizes [size_index];
 677
 678         return (GCObject *)obj;
 679 }
 680
 681 static GCObject*
 682 major_alloc_object (GCVTable vtable, size_t size, gboolean has_references)
 683 {
 684         return alloc_obj (vtable, size, FALSE, has_references);
 685 }
 686
 687 /*
 688  * We're not freeing the block if it's empty.  We leave that work for
 689  * the next major collection.
 690  *
 691  * This is just called from the domain clearing code, which runs in a
 692  * single thread and has the GC lock, so we don't need an extra lock.
 693  */
 694 static void
 695 free_object (GCObject *obj, size_t size, gboolean pinned)
 696 {
 697         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 698         int word, bit;
 699         gboolean in_free_list;
 700
 701         SGEN_ASSERT (9, sweep_state == SWEEP_STATE_SWEPT, "Should have waited for sweep to free objects.");
 702
 703         ensure_can_access_block_free_list (block);
 704         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);
 705         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
 706         MS_CALC_MARK_BIT (word, bit, obj);
 707         SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set", obj);
 708
 709         memset (obj, 0, size);
 710
 711         in_free_list = !!block->free_list;
 712         *(void**)obj = block->free_list;
 713         block->free_list = (void**)obj;
 714
 715         if (!in_free_list) {
 716                 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (pinned, block->has_references);
 717                 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 718                 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);
 719                 add_free_block (free_blocks, size_index, block);
 720         }
 721 }
 722
 723 static void
 724 major_free_non_pinned_object (GCObject *obj, size_t size)
 725 {
 726         free_object (obj, size, FALSE);
 727 }
 728
 729 /* size is a multiple of SGEN_ALLOC_ALIGN */
 730 static GCObject*
 731 major_alloc_small_pinned_obj (GCVTable vtable, size_t size, gboolean has_references)
 732 {
 733         void *res;
 734
 735         res = alloc_obj (vtable, size, TRUE, has_references);
 736          /*If we failed to alloc memory, we better try releasing memory
 737           *as pinned alloc is requested by the runtime.
 738           */
 739          if (!res) {
 740                 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE, TRUE);
 741                 res = alloc_obj (vtable, size, TRUE, has_references);
 742          }
 743          return (GCObject *)res;
 744 }
 745
 746 static void
 747 free_pinned_object (GCObject *obj, size_t size)
 748 {
 749         free_object (obj, size, TRUE);
 750 }
 751
 752 /*
 753  * size is already rounded up and we hold the GC lock.
 754  */
 755 static GCObject*
 756 major_alloc_degraded (GCVTable vtable, size_t size)
 757 {
 758         GCObject *obj;
 759
 760         obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
 761         if (G_LIKELY (obj)) {
 762                 HEAVY_STAT (++stat_objects_alloced_degraded);
 763                 HEAVY_STAT (stat_bytes_alloced_degraded += size);
 764         }
 765         return obj;
 766 }
 767
 768 /*
 769  * obj is some object.  If it's not in the major heap (i.e. if it's in
 770  * the nursery or LOS), return FALSE.  Otherwise return whether it's
 771  * been marked or copied.
 772  */
 773 static gboolean
 774 major_is_object_live (GCObject *obj)
 775 {
 776         MSBlockInfo *block;
 777         int word, bit;
 778         mword objsize;
 779
 780         if (sgen_ptr_in_nursery (obj))
 781                 return FALSE;
 782
 783         objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size (obj));
 784
 785         /* LOS */
 786         if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
 787                 return FALSE;
 788
 789         /* now we know it's in a major block */
 790         block = MS_BLOCK_FOR_OBJ (obj);
 791         SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?", block);
 792         MS_CALC_MARK_BIT (word, bit, obj);
 793         return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
 794 }
 795
 796 static gboolean
 797 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
 798 {
 799         gboolean pinned;
 800         if (ptr_is_in_major_block (ptr, start, &pinned))
 801                 return !pinned;
 802         return FALSE;
 803 }
 804
 805 static gboolean
 806 try_set_sweep_state (int new_, int expected)
 807 {
 808         int old = SGEN_CAS (&sweep_state, new_, expected);
 809         return old == expected;
 810 }
 811
 812 static void
 813 set_sweep_state (int new_, int expected)
 814 {
 815         gboolean success = try_set_sweep_state (new_, expected);
 816         SGEN_ASSERT (0, success, "Could not set sweep state.");
 817 }
 818
 819 static gboolean ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked);
 820
 821 static SgenThreadPoolJob * volatile sweep_job;
 822 static SgenThreadPoolJob * volatile sweep_blocks_job;
 823
 824 static void
 825 major_finish_sweep_checking (void)
 826 {
 827         guint32 block_index;
 828         SgenThreadPoolJob *job;
 829
 830  retry:
 831         switch (sweep_state) {
 832         case SWEEP_STATE_SWEPT:
 833         case SWEEP_STATE_NEED_SWEEPING:
 834                 return;
 835         case SWEEP_STATE_SWEEPING:
 836                 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
 837                         break;
 838                 goto retry;
 839         case SWEEP_STATE_SWEEPING_AND_ITERATING:
 840                 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
 841                 goto retry;
 842         case SWEEP_STATE_COMPACTING:
 843                 goto wait;
 844         default:
 845                 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
 846                 break;
 847         }
 848
 849         /*
 850          * We're running with the world stopped and the only other thread doing work is the
 851          * sweep thread, which doesn't add blocks to the array, so we can safely access
 852          * `next_slot`.
 853          */
 854         for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index)
 855                 ensure_block_is_checked_for_sweeping (block_index, FALSE, NULL);
 856
 857         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
 858
 859  wait:
 860         job = sweep_job;
 861         if (job)
 862                 sgen_thread_pool_job_wait (job);
 863         SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?");
 864         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?");
 865 }
 866
 867 static void
 868 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
 869 {
 870         gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
 871         gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
 872         gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
 873         MSBlockInfo *block;
 874
 875         major_finish_sweep_checking ();
 876         FOREACH_BLOCK_NO_LOCK (block) {
 877                 int count = MS_BLOCK_FREE / block->obj_size;
 878                 int i;
 879
 880                 if (block->pinned && !pinned)
 881                         continue;
 882                 if (!block->pinned && !non_pinned)
 883                         continue;
 884                 if (sweep && lazy_sweep) {
 885                         sweep_block (block);
 886                         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
 887                 }
 888
 889                 for (i = 0; i < count; ++i) {
 890                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 891                         /*
 892                          * We've finished sweep checking, but if we're sweeping lazily and
 893                          * the flags don't require us to sweep, the block might still need
 894                          * sweeping.  In that case, we need to consult the mark bits to tell
 895                          * us whether an object slot is live.
 896                          */
 897                         if (!block_is_swept_or_marking (block)) {
 898                                 int word, bit;
 899                                 SGEN_ASSERT (6, !sweep && block->state == BLOCK_STATE_NEED_SWEEPING, "Has sweeping not finished?");
 900                                 MS_CALC_MARK_BIT (word, bit, obj);
 901                                 if (!MS_MARK_BIT (block, word, bit))
 902                                         continue;
 903                         }
 904                         if (MS_OBJ_ALLOCED (obj, block))
 905                                 callback ((GCObject*)obj, block->obj_size, data);
 906                 }
 907         } END_FOREACH_BLOCK_NO_LOCK;
 908 }
 909
 910 static gboolean
 911 major_is_valid_object (char *object)
 912 {
 913         MSBlockInfo *block;
 914
 915         FOREACH_BLOCK_NO_LOCK (block) {
 916                 int idx;
 917                 char *obj;
 918
 919                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
 920                         continue;
 921
 922                 idx = MS_BLOCK_OBJ_INDEX (object, block);
 923                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 924                 if (obj != object)
 925                         return FALSE;
 926                 return MS_OBJ_ALLOCED (obj, block);
 927         } END_FOREACH_BLOCK_NO_LOCK;
 928
 929         return FALSE;
 930 }
 931
 932
 933 static GCVTable
 934 major_describe_pointer (char *ptr)
 935 {
 936         MSBlockInfo *block;
 937
 938         FOREACH_BLOCK_NO_LOCK (block) {
 939                 int idx;
 940                 char *obj;
 941                 gboolean live;
 942                 GCVTable vtable;
 943                 int w, b;
 944                 gboolean marked;
 945
 946                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
 947                         continue;
 948
 949                 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
 950                         MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
 951
 952                 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
 953                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 954                 live = MS_OBJ_ALLOCED (obj, block);
 955                 vtable = live ? SGEN_LOAD_VTABLE ((GCObject*)obj) : NULL;
 956
 957                 MS_CALC_MARK_BIT (w, b, obj);
 958                 marked = MS_MARK_BIT (block, w, b);
 959
 960                 if (obj == ptr) {
 961                         SGEN_LOG (0, "\t(");
 962                         if (live)
 963                                 SGEN_LOG (0, "object");
 964                         else
 965                                 SGEN_LOG (0, "dead-object");
 966                 } else {
 967                         if (live)
 968                                 SGEN_LOG (0, "interior-ptr offset %zd", ptr - obj);
 969                         else
 970                                 SGEN_LOG (0, "dead-interior-ptr offset %zd", ptr - obj);
 971                 }
 972
 973                 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
 974
 975                 return vtable;
 976         } END_FOREACH_BLOCK_NO_LOCK;
 977
 978         return NULL;
 979 }
 980
 981 static void
 982 major_check_scan_starts (void)
 983 {
 984 }
 985
 986 static void
 987 major_dump_heap (FILE *heap_dump_file)
 988 {
 989         MSBlockInfo *block;
 990         int *slots_available = (int *)alloca (sizeof (int) * num_block_obj_sizes);
 991         int *slots_used = (int *)alloca (sizeof (int) * num_block_obj_sizes);
 992         int i;
 993
 994         for (i = 0; i < num_block_obj_sizes; ++i)
 995                 slots_available [i] = slots_used [i] = 0;
 996
 997         FOREACH_BLOCK_NO_LOCK (block) {
 998                 int index = ms_find_block_obj_size_index (block->obj_size);
 999                 int count = MS_BLOCK_FREE / block->obj_size;
1000
1001                 slots_available [index] += count;
1002                 for (i = 0; i < count; ++i) {
1003                         if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
1004                                 ++slots_used [index];
1005                 }
1006         } END_FOREACH_BLOCK_NO_LOCK;
1007
1008         fprintf (heap_dump_file, "<occupancies>\n");
1009         for (i = 0; i < num_block_obj_sizes; ++i) {
1010                 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
1011                                 block_obj_sizes [i], slots_available [i], slots_used [i]);
1012         }
1013         fprintf (heap_dump_file, "</occupancies>\n");
1014
1015         FOREACH_BLOCK_NO_LOCK (block) {
1016                 int count = MS_BLOCK_FREE / block->obj_size;
1017                 int i;
1018                 int start = -1;
1019
1020                 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
1021
1022                 for (i = 0; i <= count; ++i) {
1023                         if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
1024                                 if (start < 0)
1025                                         start = i;
1026                         } else {
1027                                 if (start >= 0) {
1028                                         sgen_dump_occupied ((char *)MS_BLOCK_OBJ (block, start), (char *)MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
1029                                         start = -1;
1030                                 }
1031                         }
1032                 }
1033
1034                 fprintf (heap_dump_file, "</section>\n");
1035         } END_FOREACH_BLOCK_NO_LOCK;
1036 }
1037
1038 static guint8*
1039 get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate)
1040 {
1041         guint8 *mod_union = block->cardtable_mod_union;
1042         guint8 *other;
1043         if (mod_union)
1044                 return mod_union;
1045         else if (!allocate)
1046                 return NULL;
1047         mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1048         other = (guint8 *)SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL);
1049         if (!other) {
1050                 SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?");
1051                 return mod_union;
1052         }
1053         sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
1054         return other;
1055 }
1056
1057 static inline guint8*
1058 major_get_cardtable_mod_union_for_reference (char *ptr)
1059 {
1060         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (ptr);
1061         size_t offset = sgen_card_table_get_card_offset (ptr, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
1062         guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
1063         SGEN_ASSERT (0, mod_union, "FIXME: optionally allocate the mod union if it's not here and CAS it in.");
1064         return &mod_union [offset];
1065 }
1066
1067 /*
1068  * Mark the mod-union card for `ptr`, which must be a reference within the object `obj`.
1069  */
1070 static void
1071 mark_mod_union_card (GCObject *obj, void **ptr, GCObject *value_obj)
1072 {
1073         int type = sgen_obj_get_descriptor (obj) & DESC_TYPE_MASK;
1074         if (sgen_safe_object_is_small (obj, type)) {
1075                 guint8 *card_byte = major_get_cardtable_mod_union_for_reference ((char*)ptr);
1076                 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?");
1077                 *card_byte = 1;
1078         } else {
1079                 sgen_los_mark_mod_union_card (obj, ptr);
1080         }
1081         binary_protocol_mod_union_remset (obj, ptr, value_obj, SGEN_LOAD_VTABLE (value_obj));
1082 }
1083
1084 static inline gboolean
1085 major_block_is_evacuating (MSBlockInfo *block)
1086 {
1087         if (evacuate_block_obj_sizes [block->obj_size_index] &&
1088                         !block->has_pinned &&
1089                         !block->is_to_space)
1090                 return TRUE;
1091         return FALSE;
1092 }
1093
1094 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do {           \
1095                 int __word, __bit;                                      \
1096                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
1097                 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
1098                 if (!MS_MARK_BIT ((block), __word, __bit)) {            \
1099                         MS_SET_MARK_BIT ((block), __word, __bit);       \
1100                         if (sgen_gc_descr_has_references (desc))                        \
1101                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
1102                         binary_protocol_mark ((obj), (gpointer)SGEN_LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \
1103                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
1104                 }                                                       \
1105         } while (0)
1106
1107 static void
1108 pin_major_object (GCObject *obj, SgenGrayQueue *queue)
1109 {
1110         MSBlockInfo *block;
1111
1112         if (concurrent_mark)
1113                 g_assert_not_reached ();
1114
1115         block = MS_BLOCK_FOR_OBJ (obj);
1116         block->has_pinned = TRUE;
1117         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1118 }
1119
1120 #include "sgen-major-copy-object.h"
1121
1122 static long long
1123 major_get_and_reset_num_major_objects_marked (void)
1124 {
1125 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1126         long long num = num_major_objects_marked;
1127         num_major_objects_marked = 0;
1128         return num;
1129 #else
1130         return 0;
1131 #endif
1132 }
1133
1134 #define PREFETCH_CARDS          1       /* BOOL FASTENABLE */
1135 #if !PREFETCH_CARDS
1136 #undef PREFETCH_CARDS
1137 #endif
1138
1139 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1140 #if defined(PLATFORM_MACOSX)
1141 #if MONO_GNUC_VERSION <= 40300
1142 #undef PREFETCH_CARDS
1143 #endif
1144 #endif
1145
1146 #ifdef HEAVY_STATISTICS
1147 static guint64 stat_optimized_copy;
1148 static guint64 stat_optimized_copy_nursery;
1149 static guint64 stat_optimized_copy_nursery_forwarded;
1150 static guint64 stat_optimized_copy_nursery_pinned;
1151 static guint64 stat_optimized_copy_major;
1152 static guint64 stat_optimized_copy_major_small_fast;
1153 static guint64 stat_optimized_copy_major_small_slow;
1154 static guint64 stat_optimized_copy_major_large;
1155 static guint64 stat_optimized_copy_major_forwarded;
1156 static guint64 stat_optimized_copy_major_small_evacuate;
1157 static guint64 stat_optimized_major_scan;
1158 static guint64 stat_optimized_major_scan_no_refs;
1159
1160 static guint64 stat_drain_prefetch_fills;
1161 static guint64 stat_drain_prefetch_fill_failures;
1162 static guint64 stat_drain_loops;
1163 #endif
1164
1165 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_no_evacuation
1166 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_no_evacuation
1167 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_no_evacuation
1168 #include "sgen-marksweep-drain-gray-stack.h"
1169
1170 #define COPY_OR_MARK_WITH_EVACUATION
1171 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_with_evacuation
1172 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_with_evacuation
1173 #define SCAN_VTYPE_FUNCTION_NAME        major_scan_vtype_with_evacuation
1174 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_with_evacuation
1175 #define SCAN_PTR_FIELD_FUNCTION_NAME    major_scan_ptr_field_with_evacuation
1176 #include "sgen-marksweep-drain-gray-stack.h"
1177
1178 #define COPY_OR_MARK_CONCURRENT
1179 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_concurrent_no_evacuation
1180 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_concurrent_no_evacuation
1181 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_concurrent_no_evacuation
1182 #include "sgen-marksweep-drain-gray-stack.h"
1183
1184 #define COPY_OR_MARK_CONCURRENT_WITH_EVACUATION
1185 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_concurrent_with_evacuation
1186 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_concurrent_with_evacuation
1187 #define SCAN_VTYPE_FUNCTION_NAME        major_scan_vtype_concurrent_with_evacuation
1188 #define SCAN_PTR_FIELD_FUNCTION_NAME    major_scan_ptr_field_concurrent_with_evacuation
1189 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_concurrent_with_evacuation
1190 #include "sgen-marksweep-drain-gray-stack.h"
1191
1192 static inline gboolean
1193 major_is_evacuating (void)
1194 {
1195         int i;
1196         for (i = 0; i < num_block_obj_sizes; ++i) {
1197                 if (evacuate_block_obj_sizes [i]) {
1198                         return TRUE;
1199                 }
1200         }
1201
1202         return FALSE;
1203 }
1204
1205 static gboolean
1206 drain_gray_stack (SgenGrayQueue *queue)
1207 {
1208         if (major_is_evacuating ())
1209                 return drain_gray_stack_with_evacuation (queue);
1210         else
1211                 return drain_gray_stack_no_evacuation (queue);
1212 }
1213
1214 static gboolean
1215 drain_gray_stack_concurrent (SgenGrayQueue *queue)
1216 {
1217         if (major_is_evacuating ())
1218                 return drain_gray_stack_concurrent_with_evacuation (queue);
1219         else
1220                 return drain_gray_stack_concurrent_no_evacuation (queue);
1221 }
1222
1223 static void
1224 major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue)
1225 {
1226         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1227 }
1228
1229 static void
1230 major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *queue)
1231 {
1232         major_copy_or_mark_object_concurrent_with_evacuation (ptr, *ptr, queue);
1233 }
1234
1235 static void
1236 major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue)
1237 {
1238         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1239 }
1240
1241 static void
1242 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1243 {
1244         void **entry, **end;
1245         int last_index = -1;
1246
1247         if (first_entry == last_entry)
1248                 return;
1249
1250         entry = sgen_pinning_get_entry (first_entry);
1251         end = sgen_pinning_get_entry (last_entry);
1252
1253         for (; entry < end; ++entry) {
1254                 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1255                 GCObject *obj;
1256                 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));
1257                 if (index == last_index)
1258                         continue;
1259                 obj = MS_BLOCK_OBJ (block, index);
1260                 if (!MS_OBJ_ALLOCED (obj, block))
1261                         continue;
1262                 MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1263                 sgen_pin_stats_register_object (obj, GENERATION_OLD);
1264                 last_index = index;
1265         }
1266
1267         /*
1268          * There might have been potential pinning "pointers" into this block, but none of
1269          * them pointed to occupied slots, in which case we don't have to pin the block.
1270          */
1271         if (last_index >= 0)
1272                 block->has_pinned = TRUE;
1273 }
1274
1275 static inline void
1276 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1277 {
1278         int obj_index;
1279
1280         for (obj_index = 0; obj_index < count; ++obj_index) {
1281                 int word, bit;
1282                 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1283
1284                 MS_CALC_MARK_BIT (word, bit, obj);
1285                 if (MS_MARK_BIT (block, word, bit)) {
1286                         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1287                 } else {
1288                         /* an unmarked object */
1289                         if (MS_OBJ_ALLOCED (obj, block)) {
1290                                 /*
1291                                  * FIXME: Merge consecutive
1292                                  * slots for lower reporting
1293                                  * overhead.  Maybe memset
1294                                  * will also benefit?
1295                                  */
1296                                 binary_protocol_empty (obj, obj_size);
1297                                 memset (obj, 0, obj_size);
1298                         }
1299                         *(void**)obj = block->free_list;
1300                         block->free_list = (void **)obj;
1301                 }
1302         }
1303 }
1304
1305 static inline gboolean
1306 try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1307 {
1308         gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
1309         gboolean success = old_state == expected_state;
1310         if (success)
1311                 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
1312         return success;
1313 }
1314
1315 static inline void
1316 set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1317 {
1318         SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set");
1319         block->state = new_state;
1320         binary_protocol_block_set_state (block, MS_BLOCK_SIZE, expected_state, new_state);
1321 }
1322
1323 /*
1324  * If `block` needs sweeping, sweep it and return TRUE.  Otherwise return FALSE.
1325  *
1326  * Sweeping means iterating through the block's slots and building the free-list from the
1327  * unmarked ones.  They will also be zeroed.  The mark bits will be reset.
1328  */
1329 static gboolean
1330 sweep_block (MSBlockInfo *block)
1331 {
1332         int count;
1333         void *reversed = NULL;
1334
1335  retry:
1336         switch (block->state) {
1337         case BLOCK_STATE_SWEPT:
1338                 return FALSE;
1339         case BLOCK_STATE_MARKING:
1340         case BLOCK_STATE_CHECKING:
1341                 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1342                 goto retry;
1343         case BLOCK_STATE_SWEEPING:
1344                 /* FIXME: Do this more elegantly */
1345                 g_usleep (100);
1346                 goto retry;
1347         case BLOCK_STATE_NEED_SWEEPING:
1348                 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1349                         goto retry;
1350                 break;
1351         default:
1352                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1353         }
1354
1355         SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1356
1357         count = MS_BLOCK_FREE / block->obj_size;
1358
1359         block->free_list = NULL;
1360
1361         /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1362         // FIXME: Add more sizes
1363         switch (block->obj_size) {
1364         case 16:
1365                 sweep_block_for_size (block, count, 16);
1366                 break;
1367         default:
1368                 sweep_block_for_size (block, count, block->obj_size);
1369                 break;
1370         }
1371
1372         /* reset mark bits */
1373         memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1374
1375         /* Reverse free list so that it's in address order */
1376         reversed = NULL;
1377         while (block->free_list) {
1378                 void *next = *(void**)block->free_list;
1379                 *(void**)block->free_list = reversed;
1380                 reversed = block->free_list;
1381                 block->free_list = (void **)next;
1382         }
1383         block->free_list = (void **)reversed;
1384
1385         mono_memory_write_barrier ();
1386
1387         set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1388
1389         return TRUE;
1390 }
1391
1392 static inline int
1393 bitcount (mword d)
1394 {
1395         int count = 0;
1396
1397 #ifdef __GNUC__
1398         if (sizeof (mword) == 8)
1399                 count += __builtin_popcountll (d);
1400         else
1401                 count += __builtin_popcount (d);
1402 #else
1403         while (d) {
1404                 count ++;
1405                 d &= (d - 1);
1406         }
1407 #endif
1408         return count;
1409 }
1410
1411 /* statistics for evacuation */
1412 static size_t *sweep_slots_available;
1413 static size_t *sweep_slots_used;
1414 static size_t *sweep_num_blocks;
1415
1416 static volatile size_t num_major_sections_before_sweep;
1417 static volatile size_t num_major_sections_freed_in_sweep;
1418
1419 static void
1420 sweep_start (void)
1421 {
1422         int i;
1423
1424         for (i = 0; i < num_block_obj_sizes; ++i)
1425                 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1426
1427         /* clear all the free lists */
1428         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1429                 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1430                 int j;
1431                 for (j = 0; j < num_block_obj_sizes; ++j)
1432                         free_blocks [j] = NULL;
1433         }
1434
1435         sgen_array_list_remove_nulls (&allocated_blocks);
1436 }
1437
1438 static void sweep_finish (void);
1439
1440 /*
1441  * If `wait` is TRUE and the block is currently being checked, this function will wait until
1442  * the checking has finished.
1443  *
1444  * Returns whether the block is still there.  If `wait` is FALSE, the return value will not
1445  * be correct, i.e. must not be used.
1446  */
1447 static gboolean
1448 ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboolean *have_checked)
1449 {
1450         int count;
1451         gboolean have_live = FALSE;
1452         gboolean have_free = FALSE;
1453         int nused = 0;
1454         int block_state;
1455         int i;
1456         void *tagged_block;
1457         MSBlockInfo *block;
1458         volatile gpointer *block_slot = sgen_array_list_get_slot (&allocated_blocks, block_index);
1459
1460         SGEN_ASSERT (6, sweep_in_progress (), "Why do we call this function if there's no sweep in progress?");
1461
1462         if (have_checked)
1463                 *have_checked = FALSE;
1464
1465  retry:
1466         tagged_block = *(void * volatile *)block_slot;
1467         if (!tagged_block)
1468                 return FALSE;
1469
1470         if (BLOCK_IS_TAGGED_CHECKING (tagged_block)) {
1471                 if (!wait)
1472                         return FALSE;
1473                 /* FIXME: do this more elegantly */
1474                 g_usleep (100);
1475                 goto retry;
1476         }
1477
1478         if (SGEN_CAS_PTR (block_slot, BLOCK_TAG_CHECKING (tagged_block), tagged_block) != tagged_block)
1479                 goto retry;
1480
1481         block = BLOCK_UNTAG (tagged_block);
1482         block_state = block->state;
1483
1484         if (!sweep_in_progress ()) {
1485                 SGEN_ASSERT (6, block_state != BLOCK_STATE_SWEEPING && block_state != BLOCK_STATE_CHECKING, "Invalid block state.");
1486                 if (!lazy_sweep)
1487                         SGEN_ASSERT (6, block_state != BLOCK_STATE_NEED_SWEEPING, "Invalid block state.");
1488         }
1489
1490         switch (block_state) {
1491         case BLOCK_STATE_SWEPT:
1492         case BLOCK_STATE_NEED_SWEEPING:
1493         case BLOCK_STATE_SWEEPING:
1494                 goto done;
1495         case BLOCK_STATE_MARKING:
1496                 break;
1497         case BLOCK_STATE_CHECKING:
1498                 SGEN_ASSERT (0, FALSE, "We set the CHECKING bit - how can the stage be CHECKING?");
1499                 goto done;
1500         default:
1501                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1502                 break;
1503         }
1504
1505         SGEN_ASSERT (6, block->state == BLOCK_STATE_MARKING, "When we sweep all blocks must start out marking.");
1506         set_block_state (block, BLOCK_STATE_CHECKING, BLOCK_STATE_MARKING);
1507
1508         if (have_checked)
1509                 *have_checked = TRUE;
1510
1511         block->has_pinned = block->pinned;
1512
1513         block->is_to_space = FALSE;
1514
1515         count = MS_BLOCK_FREE / block->obj_size;
1516
1517         if (block->cardtable_mod_union)
1518                 memset (block->cardtable_mod_union, 0, CARDS_PER_BLOCK);
1519
1520         /* Count marked objects in the block */
1521         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1522                 nused += bitcount (block->mark_words [i]);
1523
1524         block->nused = nused;
1525         if (nused)
1526                 have_live = TRUE;
1527         if (nused < count)
1528                 have_free = TRUE;
1529
1530         if (have_live) {
1531                 int obj_size_index = block->obj_size_index;
1532                 gboolean has_pinned = block->has_pinned;
1533
1534                 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1535
1536                 /*
1537                  * FIXME: Go straight to SWEPT if there are no free slots.  We need
1538                  * to set the free slot list to NULL, though, and maybe update some
1539                  * statistics.
1540                  */
1541                 if (!lazy_sweep)
1542                         sweep_block (block);
1543
1544                 if (!has_pinned) {
1545                         ++sweep_num_blocks [obj_size_index];
1546                         sweep_slots_used [obj_size_index] += nused;
1547                         sweep_slots_available [obj_size_index] += count;
1548                 }
1549
1550                 /*
1551                  * If there are free slots in the block, add
1552                  * the block to the corresponding free list.
1553                  */
1554                 if (have_free) {
1555                         MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1556
1557                         if (!lazy_sweep)
1558                                 SGEN_ASSERT (6, block->free_list, "How do we not have a free list when there are free slots?");
1559
1560                         add_free_block (free_blocks, obj_size_index, block);
1561                 }
1562
1563                 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1564                 update_heap_boundaries_for_block (block);
1565         } else {
1566                 /*
1567                  * Blocks without live objects are removed from the
1568                  * block list and freed.
1569                  */
1570                 SGEN_ASSERT (6, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1571                 SGEN_ASSERT (6, *block_slot == BLOCK_TAG_CHECKING (tagged_block), "How did the block move?");
1572
1573                 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1574                 ms_free_block (block);
1575
1576                 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1577
1578                 tagged_block = NULL;
1579         }
1580
1581  done:
1582         /*
1583          * Once the block is written back without the checking bit other threads are
1584          * free to access it. Make sure the block state is visible before we write it
1585          * back.
1586          */
1587         mono_memory_write_barrier ();
1588         *block_slot = tagged_block;
1589         return !!tagged_block;
1590 }
1591
1592 static void
1593 sweep_blocks_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1594 {
1595         volatile gpointer *slot;
1596         MSBlockInfo *bl;
1597
1598         SGEN_ARRAY_LIST_FOREACH_SLOT (&allocated_blocks, slot) {
1599                 bl = BLOCK_UNTAG (*slot);
1600                 if (bl)
1601                         sweep_block (bl);
1602         } SGEN_ARRAY_LIST_END_FOREACH_SLOT;
1603
1604         mono_memory_write_barrier ();
1605
1606         sweep_blocks_job = NULL;
1607 }
1608
1609 static void
1610 sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job)
1611 {
1612         guint32 block_index;
1613         guint32 num_blocks = num_major_sections_before_sweep;
1614
1615         SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1616         SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1617
1618         /*
1619          * We traverse the block array from high to low.  Nursery collections will have to
1620          * cooperate with the sweep thread to finish sweeping, and they will traverse from
1621          * low to high, to avoid constantly colliding on the same blocks.
1622          */
1623         for (block_index = num_blocks; block_index-- > 0;) {
1624                 /*
1625                  * The block might have been freed by another thread doing some checking
1626                  * work.
1627                  */
1628                 if (!ensure_block_is_checked_for_sweeping (block_index, TRUE, NULL))
1629                         ++num_major_sections_freed_in_sweep;
1630         }
1631
1632         while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) {
1633                 /*
1634                  * The main GC thread is currently iterating over the block array to help us
1635                  * finish the sweep.  We have already finished, but we don't want to mess up
1636                  * that iteration, so we just wait for it.
1637                  */
1638                 g_usleep (100);
1639         }
1640
1641         if (SGEN_MAX_ASSERT_LEVEL >= 6) {
1642                 for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1643                         MSBlockInfo *block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, block_index));
1644                         SGEN_ASSERT (6, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1645                 }
1646         }
1647
1648         /*
1649          * Concurrently sweep all the blocks to reduce workload during minor
1650          * pauses where we need certain blocks to be swept. At the start of
1651          * the next major we need all blocks to be swept anyway.
1652          */
1653         if (concurrent_sweep && lazy_sweep) {
1654                 sweep_blocks_job = sgen_thread_pool_job_alloc ("sweep_blocks", sweep_blocks_job_func, sizeof (SgenThreadPoolJob));
1655                 sgen_thread_pool_job_enqueue (sweep_blocks_job);
1656         }
1657
1658         sweep_finish ();
1659
1660         sweep_job = NULL;
1661 }
1662
1663 static void
1664 sweep_finish (void)
1665 {
1666         mword used_slots_size = 0;
1667         int i;
1668
1669         for (i = 0; i < num_block_obj_sizes; ++i) {
1670                 float usage = (float)sweep_slots_used [i] / (float)sweep_slots_available [i];
1671                 if (sweep_num_blocks [i] > 5 && usage < evacuation_threshold) {
1672                         evacuate_block_obj_sizes [i] = TRUE;
1673                         /*
1674                         g_print ("slot size %d - %d of %d used\n",
1675                                         block_obj_sizes [i], slots_used [i], slots_available [i]);
1676                         */
1677                 } else {
1678                         evacuate_block_obj_sizes [i] = FALSE;
1679                 }
1680
1681                 used_slots_size += sweep_slots_used [i] * block_obj_sizes [i];
1682         }
1683
1684         sgen_memgov_major_post_sweep (used_slots_size);
1685
1686         set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1687         if (concurrent_sweep)
1688                 binary_protocol_concurrent_sweep_end (sgen_timestamp ());
1689 }
1690
1691 static void
1692 major_sweep (void)
1693 {
1694         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1695
1696         sweep_start ();
1697
1698         SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?");
1699
1700         num_major_sections_before_sweep = num_major_sections;
1701         num_major_sections_freed_in_sweep = 0;
1702
1703         SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?");
1704         if (concurrent_sweep) {
1705                 sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob));
1706                 sgen_thread_pool_job_enqueue (sweep_job);
1707         } else {
1708                 sweep_job_func (NULL, NULL);
1709         }
1710 }
1711
1712 static gboolean
1713 major_have_swept (void)
1714 {
1715         return sweep_state == SWEEP_STATE_SWEPT;
1716 }
1717
1718 static int count_pinned_ref;
1719 static int count_pinned_nonref;
1720 static int count_nonpinned_ref;
1721 static int count_nonpinned_nonref;
1722
1723 static void
1724 count_nonpinned_callback (GCObject *obj, size_t size, void *data)
1725 {
1726         GCVTable vtable = SGEN_LOAD_VTABLE (obj);
1727
1728         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1729                 ++count_nonpinned_ref;
1730         else
1731                 ++count_nonpinned_nonref;
1732 }
1733
1734 static void
1735 count_pinned_callback (GCObject *obj, size_t size, void *data)
1736 {
1737         GCVTable vtable = SGEN_LOAD_VTABLE (obj);
1738
1739         if (SGEN_VTABLE_HAS_REFERENCES (vtable))
1740                 ++count_pinned_ref;
1741         else
1742                 ++count_pinned_nonref;
1743 }
1744
1745 static G_GNUC_UNUSED void
1746 count_ref_nonref_objs (void)
1747 {
1748         int total;
1749
1750         count_pinned_ref = 0;
1751         count_pinned_nonref = 0;
1752         count_nonpinned_ref = 0;
1753         count_nonpinned_nonref = 0;
1754
1755         major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1756         major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1757
1758         total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1759
1760         g_print ("ref: %d pinned %d non-pinned   non-ref: %d pinned %d non-pinned  --  %.1f\n",
1761                         count_pinned_ref, count_nonpinned_ref,
1762                         count_pinned_nonref, count_nonpinned_nonref,
1763                         (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1764 }
1765
1766 static int
1767 ms_calculate_block_obj_sizes (double factor, int *arr)
1768 {
1769         double target_size;
1770         int num_sizes = 0;
1771         int last_size = 0;
1772
1773         /*
1774          * Have every possible slot size starting with the minimal
1775          * object size up to and including four times that size.  Then
1776          * proceed by increasing geometrically with the given factor.
1777          */
1778
1779         for (int size = SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size <= 4 * SGEN_CLIENT_MINIMUM_OBJECT_SIZE; size += SGEN_ALLOC_ALIGN) {
1780                 if (arr)
1781                         arr [num_sizes] = size;
1782                 ++num_sizes;
1783                 last_size = size;
1784         }
1785         target_size = (double)last_size;
1786
1787         do {
1788                 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1789                 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1790
1791                 if (size != last_size) {
1792                         if (arr)
1793                                 arr [num_sizes] = size;
1794                         ++num_sizes;
1795                         last_size = size;
1796                 }
1797
1798                 target_size *= factor;
1799         } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1800
1801         return num_sizes;
1802 }
1803
1804 /* only valid during minor collections */
1805 static mword old_num_major_sections;
1806
1807 static void
1808 major_start_nursery_collection (void)
1809 {
1810 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1811         consistency_check ();
1812 #endif
1813
1814         old_num_major_sections = num_major_sections;
1815 }
1816
1817 static void
1818 major_finish_nursery_collection (void)
1819 {
1820 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1821         consistency_check ();
1822 #endif
1823 }
1824
1825 static int
1826 block_usage_comparer (const void *bl1, const void *bl2)
1827 {
1828         const gint16 nused1 = (*(MSBlockInfo**)bl1)->nused;
1829         const gint16 nused2 = (*(MSBlockInfo**)bl2)->nused;
1830
1831         return nused2 - nused1;
1832 }
1833
1834 static void
1835 sgen_evacuation_freelist_blocks (MSBlockInfo * volatile *block_list, int size_index)
1836 {
1837         MSBlockInfo **evacuated_blocks;
1838         size_t index = 0, count, num_blocks = 0, num_used = 0;
1839         MSBlockInfo *info;
1840         MSBlockInfo * volatile *prev;
1841
1842         for (info = *block_list; info != NULL; info = info->next_free) {
1843                 num_blocks++;
1844                 num_used += info->nused;
1845         }
1846
1847         /*
1848          * We have a set of blocks in the freelist which will be evacuated. Instead
1849          * of evacuating all of the blocks into new ones, we traverse the freelist
1850          * sorting it by the number of occupied slots, evacuating the objects from
1851          * blocks with fewer used slots into fuller blocks.
1852          *
1853          * The number of used slots is set at the end of the previous sweep. Since
1854          * we sequentially unlink slots from blocks, except for the head of the
1855          * freelist, for blocks on the freelist, the number of used slots is the same
1856          * as at the end of the previous sweep.
1857          */
1858         evacuated_blocks = (MSBlockInfo**)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_blocks, INTERNAL_MEM_TEMPORARY, TRUE);
1859
1860         for (info = *block_list; info != NULL; info = info->next_free) {
1861                 evacuated_blocks [index++] = info;
1862         }
1863
1864         SGEN_ASSERT (0, num_blocks == index, "Why did the freelist change ?");
1865
1866         sgen_qsort (evacuated_blocks, num_blocks, sizeof (gpointer), block_usage_comparer);
1867
1868         /*
1869          * Form a new freelist with the fullest blocks. These blocks will also be
1870          * marked as to_space so we don't evacuate from them.
1871          */
1872         count = MS_BLOCK_FREE / block_obj_sizes [size_index];
1873         prev = block_list;
1874         for (index = 0; index < (num_used + count - 1) / count; index++) {
1875                 SGEN_ASSERT (0, index < num_blocks, "Why do we need more blocks for compaction than we already had ?");
1876                 info = evacuated_blocks [index];
1877                 info->is_to_space = TRUE;
1878                 *prev = info;
1879                 prev = &info->next_free;
1880         }
1881         *prev = NULL;
1882
1883         sgen_free_internal_dynamic (evacuated_blocks, sizeof (MSBlockInfo*) * num_blocks, INTERNAL_MEM_TEMPORARY);
1884 }
1885
1886 static void
1887 major_start_major_collection (void)
1888 {
1889         MSBlockInfo *block;
1890         int i;
1891
1892         major_finish_sweep_checking ();
1893
1894         /*
1895          * Clear the free lists for block sizes where we do evacuation.  For those block
1896          * sizes we will have to allocate new blocks.
1897          */
1898         for (i = 0; i < num_block_obj_sizes; ++i) {
1899                 if (!evacuate_block_obj_sizes [i])
1900                         continue;
1901
1902                 binary_protocol_evacuating_blocks (block_obj_sizes [i]);
1903
1904                 sgen_evacuation_freelist_blocks (&free_block_lists [0][i], i);
1905                 sgen_evacuation_freelist_blocks (&free_block_lists [MS_BLOCK_FLAG_REFS][i], i);
1906         }
1907
1908         if (lazy_sweep && concurrent_sweep) {
1909                 /*
1910                  * sweep_blocks_job is created before sweep_finish, which we wait for above
1911                  * (major_finish_sweep_checking). After the end of sweep, if we don't have
1912                  * sweep_blocks_job set, it means that it has already been run.
1913                  */
1914                 SgenThreadPoolJob *job = sweep_blocks_job;
1915                 if (job)
1916                         sgen_thread_pool_job_wait (job);
1917         }
1918
1919         if (lazy_sweep && !concurrent_sweep)
1920                 binary_protocol_sweep_begin (GENERATION_OLD, TRUE);
1921         /* Sweep all unswept blocks and set them to MARKING */
1922         FOREACH_BLOCK_NO_LOCK (block) {
1923                 if (lazy_sweep && !concurrent_sweep)
1924                         sweep_block (block);
1925                 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1926                 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1927                 /*
1928                  * Swept blocks that have a null free_list are full. Evacuation is not
1929                  * effective on these blocks since we expect them to have high usage anyway,
1930                  * given that the survival rate for majors is relatively high.
1931                  */
1932                 if (evacuate_block_obj_sizes [block->obj_size_index] && !block->free_list)
1933                         block->is_to_space = TRUE;
1934         } END_FOREACH_BLOCK_NO_LOCK;
1935         if (lazy_sweep && !concurrent_sweep)
1936                 binary_protocol_sweep_end (GENERATION_OLD, TRUE);
1937
1938         set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1939 }
1940
1941 static void
1942 major_finish_major_collection (ScannedObjectCounts *counts)
1943 {
1944 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1945         if (binary_protocol_is_enabled ()) {
1946                 counts->num_scanned_objects = scanned_objects_list.next_slot;
1947
1948                 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1949                 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1950
1951                 sgen_pointer_queue_clear (&scanned_objects_list);
1952         }
1953 #endif
1954 }
1955
1956 #if SIZEOF_VOID_P != 8
1957 static int
1958 compare_pointers (const void *va, const void *vb) {
1959         char *a = *(char**)va, *b = *(char**)vb;
1960         if (a < b)
1961                 return -1;
1962         if (a > b)
1963                 return 1;
1964         return 0;
1965 }
1966 #endif
1967
1968 /*
1969  * This is called with sweep completed and the world stopped.
1970  */
1971 static void
1972 major_free_swept_blocks (size_t allowance)
1973 {
1974         /* FIXME: This is probably too much.  It's assuming all objects are small. */
1975         size_t section_reserve = allowance / MS_BLOCK_SIZE;
1976
1977         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1978
1979 #ifdef TARGET_WIN32
1980                 /*
1981                  * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1982                  * a VirtualAlloc ()-ed block.
1983                  */
1984                 return;
1985 #endif
1986
1987 #if SIZEOF_VOID_P != 8
1988         {
1989                 int i, num_empty_blocks_orig, num_blocks, arr_length;
1990                 void *block;
1991                 void **empty_block_arr;
1992                 void **rebuild_next;
1993
1994                 if (num_empty_blocks <= section_reserve)
1995                         return;
1996                 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1997
1998                 num_empty_blocks_orig = num_empty_blocks;
1999                 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
2000                                 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
2001                 if (!empty_block_arr)
2002                         goto fallback;
2003
2004                 i = 0;
2005                 for (block = empty_blocks; block; block = *(void**)block)
2006                         empty_block_arr [i++] = block;
2007                 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
2008
2009                 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
2010
2011                 /*
2012                  * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
2013                  * contiguous ones.  If we do, we free them.  If that's not enough to get to
2014                  * section_reserve, we halve the number of contiguous blocks we're looking
2015                  * for and have another go, until we're done with looking for pairs of
2016                  * blocks, at which point we give up and go to the fallback.
2017                  */
2018                 arr_length = num_empty_blocks_orig;
2019                 num_blocks = MS_BLOCK_ALLOC_NUM;
2020                 while (num_empty_blocks > section_reserve && num_blocks > 1) {
2021                         int first = -1;
2022                         int dest = 0;
2023
2024                         dest = 0;
2025                         for (i = 0; i < arr_length; ++i) {
2026                                 int d = dest;
2027                                 void *block = empty_block_arr [i];
2028                                 SGEN_ASSERT (6, block, "we're not shifting correctly");
2029                                 if (i != dest) {
2030                                         empty_block_arr [dest] = block;
2031                                         /*
2032                                          * This is not strictly necessary, but we're
2033                                          * cautious.
2034                                          */
2035                                         empty_block_arr [i] = NULL;
2036                                 }
2037                                 ++dest;
2038
2039                                 if (first < 0) {
2040                                         first = d;
2041                                         continue;
2042                                 }
2043
2044                                 SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong");
2045
2046                                 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
2047                                         first = d;
2048                                         continue;
2049                                 }
2050
2051                                 if (d + 1 - first == num_blocks) {
2052                                         /*
2053                                          * We found num_blocks contiguous blocks.  Free them
2054                                          * and null their array entries.  As an optimization
2055                                          * we could, instead of nulling the entries, shift
2056                                          * the following entries over to the left, while
2057                                          * we're iterating.
2058                                          */
2059                                         int j;
2060                                         sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
2061                                         for (j = first; j <= d; ++j)
2062                                                 empty_block_arr [j] = NULL;
2063                                         dest = first;
2064                                         first = -1;
2065
2066                                         num_empty_blocks -= num_blocks;
2067
2068                                         stat_major_blocks_freed += num_blocks;
2069                                         if (num_blocks == MS_BLOCK_ALLOC_NUM)
2070                                                 stat_major_blocks_freed_ideal += num_blocks;
2071                                         else
2072                                                 stat_major_blocks_freed_less_ideal += num_blocks;
2073
2074                                 }
2075                         }
2076
2077                         SGEN_ASSERT (6, dest <= i && dest <= arr_length, "array length is off");
2078                         arr_length = dest;
2079                         SGEN_ASSERT (6, arr_length == num_empty_blocks, "array length is off");
2080
2081                         num_blocks >>= 1;
2082                 }
2083
2084                 /* rebuild empty_blocks free list */
2085                 rebuild_next = (void**)&empty_blocks;
2086                 for (i = 0; i < arr_length; ++i) {
2087                         void *block = empty_block_arr [i];
2088                         SGEN_ASSERT (6, block, "we're missing blocks");
2089                         *rebuild_next = block;
2090                         rebuild_next = (void**)block;
2091                 }
2092                 *rebuild_next = NULL;
2093
2094                 /* free array */
2095                 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
2096         }
2097
2098         SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
2099
2100  fallback:
2101         /*
2102          * This is our threshold.  If there's not more empty than used blocks, we won't
2103          * release uncontiguous blocks, in fear of fragmenting the address space.
2104          */
2105         if (num_empty_blocks <= num_major_sections)
2106                 return;
2107 #endif
2108
2109         while (num_empty_blocks > section_reserve) {
2110                 void *next = *(void**)empty_blocks;
2111                 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
2112                 empty_blocks = next;
2113                 /*
2114                  * Needs not be atomic because this is running
2115                  * single-threaded.
2116                  */
2117                 --num_empty_blocks;
2118
2119                 ++stat_major_blocks_freed;
2120 #if SIZEOF_VOID_P != 8
2121                 ++stat_major_blocks_freed_individual;
2122 #endif
2123         }
2124 }
2125
2126 static void
2127 major_pin_objects (SgenGrayQueue *queue)
2128 {
2129         MSBlockInfo *block;
2130
2131         FOREACH_BLOCK_NO_LOCK (block) {
2132                 size_t first_entry, last_entry;
2133                 SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning.");
2134                 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
2135                                 &first_entry, &last_entry);
2136                 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
2137         } END_FOREACH_BLOCK_NO_LOCK;
2138 }
2139
2140 static void
2141 major_init_to_space (void)
2142 {
2143 }
2144
2145 static void
2146 major_report_pinned_memory_usage (void)
2147 {
2148         g_assert_not_reached ();
2149 }
2150
2151 static gint64
2152 major_get_used_size (void)
2153 {
2154         gint64 size = 0;
2155         MSBlockInfo *block;
2156
2157         /*
2158          * We're holding the GC lock, but the sweep thread might be running.  Make sure it's
2159          * finished, then we can iterate over the block array.
2160          */
2161         major_finish_sweep_checking ();
2162
2163         FOREACH_BLOCK_NO_LOCK (block) {
2164                 int count = MS_BLOCK_FREE / block->obj_size;
2165                 void **iter;
2166                 size += count * block->obj_size;
2167                 for (iter = block->free_list; iter; iter = (void**)*iter)
2168                         size -= block->obj_size;
2169         } END_FOREACH_BLOCK_NO_LOCK;
2170
2171         return size;
2172 }
2173
2174 /* FIXME: return number of bytes, not of sections */
2175 static size_t
2176 get_num_major_sections (void)
2177 {
2178         return num_major_sections;
2179 }
2180
2181 /*
2182  * Returns the number of bytes in blocks that were present when the last sweep was
2183  * initiated, and were not freed during the sweep.  They are the basis for calculating the
2184  * allowance.
2185  */
2186 static size_t
2187 get_bytes_survived_last_sweep (void)
2188 {
2189         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2190         return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE;
2191 }
2192
2193 static gboolean
2194 major_handle_gc_param (const char *opt)
2195 {
2196         if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2197                 const char *arg = strchr (opt, '=') + 1;
2198                 int percentage = atoi (arg);
2199                 if (percentage < 0 || percentage > 100) {
2200                         fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2201                         exit (1);
2202                 }
2203                 evacuation_threshold = (float)percentage / 100.0f;
2204                 return TRUE;
2205         } else if (!strcmp (opt, "lazy-sweep")) {
2206                 lazy_sweep = TRUE;
2207                 return TRUE;
2208         } else if (!strcmp (opt, "no-lazy-sweep")) {
2209                 lazy_sweep = FALSE;
2210                 return TRUE;
2211         } else if (!strcmp (opt, "concurrent-sweep")) {
2212                 concurrent_sweep = TRUE;
2213                 return TRUE;
2214         } else if (!strcmp (opt, "no-concurrent-sweep")) {
2215                 concurrent_sweep = FALSE;
2216                 return TRUE;
2217         }
2218
2219         return FALSE;
2220 }
2221
2222 static void
2223 major_print_gc_param_usage (void)
2224 {
2225         fprintf (stderr,
2226                         ""
2227                         "  evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2228                         "  (no-)lazy-sweep\n"
2229                         "  (no-)concurrent-sweep\n"
2230                         );
2231 }
2232
2233 /*
2234  * This callback is used to clear cards, move cards to the shadow table and do counting.
2235  */
2236 static void
2237 major_iterate_block_ranges (sgen_cardtable_block_callback callback)
2238 {
2239         MSBlockInfo *block;
2240         gboolean has_references;
2241
2242         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2243                 if (has_references)
2244                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2245         } END_FOREACH_BLOCK_NO_LOCK;
2246 }
2247
2248 static void
2249 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback)
2250 {
2251         MSBlockInfo *block;
2252         gboolean has_references;
2253
2254         major_finish_sweep_checking ();
2255         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2256                 if (has_references)
2257                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2258         } END_FOREACH_BLOCK_NO_LOCK;
2259 }
2260
2261 #ifdef HEAVY_STATISTICS
2262 extern guint64 marked_cards;
2263 extern guint64 scanned_cards;
2264 extern guint64 scanned_objects;
2265 extern guint64 remarked_cards;
2266 #endif
2267
2268 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2269 /*
2270  * MS blocks are 16K aligned.
2271  * Cardtables are 4K aligned, at least.
2272  * This means that the cardtable of a given block is 32 bytes aligned.
2273  */
2274 static guint8*
2275 initial_skip_card (guint8 *card_data)
2276 {
2277         mword *cards = (mword*)card_data;
2278         mword card;
2279         int i;
2280         for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2281                 card = cards [i];
2282                 if (card)
2283                         break;
2284         }
2285
2286         if (i == CARD_WORDS_PER_BLOCK)
2287                 return card_data + CARDS_PER_BLOCK;
2288
2289 #if defined(__i386__) && defined(__GNUC__)
2290         return card_data + i * 4 +  (__builtin_ffs (card) - 1) / 8;
2291 #elif defined(__x86_64__) && defined(__GNUC__)
2292         return card_data + i * 8 +  (__builtin_ffsll (card) - 1) / 8;
2293 #elif defined(__s390x__) && defined(__GNUC__)
2294         return card_data + i * 8 +  (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2295 #else
2296         for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2297                 if (card_data [i])
2298                         return &card_data [i];
2299         }
2300         return card_data;
2301 #endif
2302 }
2303
2304 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2305 #define MS_BLOCK_OBJ_FAST(b,os,i)                       ((b) + MS_BLOCK_SKIP + (os) * (i))
2306 #define MS_OBJ_ALLOCED_FAST(o,b)                (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2307
2308 static void
2309 scan_card_table_for_block (MSBlockInfo *block, CardTableScanType scan_type, ScanCopyContext ctx)
2310 {
2311         SgenGrayQueue *queue = ctx.queue;
2312         ScanObjectFunc scan_func = ctx.ops->scan_object;
2313 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2314         guint8 cards_copy [CARDS_PER_BLOCK];
2315 #endif
2316         guint8 cards_preclean [CARDS_PER_BLOCK];
2317         gboolean small_objects;
2318         int block_obj_size;
2319         char *block_start;
2320         guint8 *card_data, *card_base;
2321         guint8 *card_data_end;
2322         char *scan_front = NULL;
2323
2324         /* The concurrent mark doesn't enter evacuating blocks */
2325         if (scan_type == CARDTABLE_SCAN_MOD_UNION_PRECLEAN && major_block_is_evacuating (block))
2326                 return;
2327
2328         block_obj_size = block->obj_size;
2329         small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2330
2331         block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2332
2333         /*
2334          * This is safe in face of card aliasing for the following reason:
2335          *
2336          * Major blocks are 16k aligned, or 32 cards aligned.
2337          * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2338          * sizes, they won't overflow the cardtable overlap modulus.
2339          */
2340         if (scan_type & CARDTABLE_SCAN_MOD_UNION) {
2341                 card_data = card_base = block->cardtable_mod_union;
2342                 /*
2343                  * This happens when the nursery collection that precedes finishing
2344                  * the concurrent collection allocates new major blocks.
2345                  */
2346                 if (!card_data)
2347                         return;
2348
2349                 if (scan_type == CARDTABLE_SCAN_MOD_UNION_PRECLEAN) {
2350                         sgen_card_table_preclean_mod_union (card_data, cards_preclean, CARDS_PER_BLOCK);
2351                         card_data = card_base = cards_preclean;
2352                 }
2353         } else {
2354 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2355                 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2356 #else
2357                 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2358                         return;
2359                 card_data = card_base = cards_copy;
2360 #endif
2361         }
2362         card_data_end = card_data + CARDS_PER_BLOCK;
2363
2364         card_data += MS_BLOCK_SKIP >> CARD_BITS;
2365
2366         card_data = initial_skip_card (card_data);
2367         while (card_data < card_data_end) {
2368                 size_t card_index, first_object_index;
2369                 char *start;
2370                 char *end;
2371                 char *first_obj, *obj;
2372
2373                 HEAVY_STAT (++scanned_cards);
2374
2375                 if (!*card_data) {
2376                         ++card_data;
2377                         continue;
2378                 }
2379
2380                 card_index = card_data - card_base;
2381                 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2382                 end = start + CARD_SIZE_IN_BYTES;
2383
2384                 if (!block_is_swept_or_marking (block))
2385                         sweep_block (block);
2386
2387                 HEAVY_STAT (++marked_cards);
2388
2389                 if (small_objects)
2390                         sgen_card_table_prepare_card_for_scanning (card_data);
2391
2392                 /*
2393                  * If the card we're looking at starts at or in the block header, we
2394                  * must start at the first object in the block, without calculating
2395                  * the index of the object we're hypothetically starting at, because
2396                  * it would be negative.
2397                  */
2398                 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2399                         first_object_index = 0;
2400                 else
2401                         first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2402
2403                 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2404
2405                 binary_protocol_card_scan (first_obj, end - first_obj);
2406
2407                 while (obj < end) {
2408                         if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2409                                 goto next_object;
2410
2411                         if (scan_type & CARDTABLE_SCAN_MOD_UNION) {
2412                                 /* FIXME: do this more efficiently */
2413                                 int w, b;
2414                                 MS_CALC_MARK_BIT (w, b, obj);
2415                                 if (!MS_MARK_BIT (block, w, b))
2416                                         goto next_object;
2417                         }
2418
2419                         GCObject *object = (GCObject*)obj;
2420
2421                         if (small_objects) {
2422                                 HEAVY_STAT (++scanned_objects);
2423                                 scan_func (object, sgen_obj_get_descriptor (object), queue);
2424                         } else {
2425                                 size_t offset = sgen_card_table_get_card_offset (obj, block_start);
2426                                 sgen_cardtable_scan_object (object, block_obj_size, card_base + offset, ctx);
2427                         }
2428                 next_object:
2429                         obj += block_obj_size;
2430                         g_assert (scan_front <= obj);
2431                         scan_front = obj;
2432                 }
2433
2434                 HEAVY_STAT (if (*card_data) ++remarked_cards);
2435
2436                 if (small_objects)
2437                         ++card_data;
2438                 else
2439                         card_data = card_base + sgen_card_table_get_card_offset (obj, block_start);
2440         }
2441 }
2442
2443 static void
2444 major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx)
2445 {
2446         MSBlockInfo *block;
2447         gboolean has_references, was_sweeping, skip_scan;
2448
2449         if (!concurrent_mark)
2450                 g_assert (scan_type == CARDTABLE_SCAN_GLOBAL);
2451
2452         if (scan_type != CARDTABLE_SCAN_GLOBAL)
2453                 SGEN_ASSERT (0, !sweep_in_progress (), "Sweep should be finished when we scan mod union card table");
2454         was_sweeping = sweep_in_progress ();
2455
2456         binary_protocol_major_card_table_scan_start (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
2457         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2458 #ifdef PREFETCH_CARDS
2459                 int prefetch_index = __index + 6;
2460                 if (prefetch_index < allocated_blocks.next_slot) {
2461                         MSBlockInfo *prefetch_block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, prefetch_index));
2462                         PREFETCH_READ (prefetch_block);
2463                         if (scan_type == CARDTABLE_SCAN_GLOBAL) {
2464                                 guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2465                                 PREFETCH_WRITE (prefetch_cards);
2466                                 PREFETCH_WRITE (prefetch_cards + 32);
2467                         }
2468                 }
2469 #endif
2470
2471                 if (!has_references)
2472                         continue;
2473                 skip_scan = FALSE;
2474
2475                 if (scan_type == CARDTABLE_SCAN_GLOBAL) {
2476                         gpointer *card_start = (gpointer*) sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (block));
2477                         gboolean has_dirty_cards = FALSE;
2478                         int i;
2479                         for (i = 0; i < CARDS_PER_BLOCK / sizeof(gpointer); i++) {
2480                                 if (card_start [i]) {
2481                                         has_dirty_cards = TRUE;
2482                                         break;
2483                                 }
2484                         }
2485                         if (!has_dirty_cards) {
2486                                 skip_scan = TRUE;
2487                         } else {
2488                                 /*
2489                                  * After the start of the concurrent collections, blocks change state
2490                                  * to marking. We should not sweep it in that case. We can't race with
2491                                  * sweep start since we are in a nursery collection. Also avoid CAS-ing
2492                                  */
2493                                 if (sweep_in_progress ()) {
2494                                         skip_scan = !ensure_block_is_checked_for_sweeping (__index, TRUE, NULL);
2495                                 } else if (was_sweeping) {
2496                                         /* Recheck in case sweep finished after dereferencing the slot */
2497                                         skip_scan = *sgen_array_list_get_slot (&allocated_blocks, __index) == 0;
2498                                 }
2499                         }
2500                 }
2501                 if (!skip_scan)
2502                         scan_card_table_for_block (block, scan_type, ctx);
2503         } END_FOREACH_BLOCK_NO_LOCK;
2504         binary_protocol_major_card_table_scan_end (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
2505 }
2506
2507 static void
2508 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2509 {
2510         MSBlockInfo *block;
2511         gboolean has_references;
2512         long long total_cards = 0;
2513         long long marked_cards = 0;
2514
2515         if (sweep_in_progress ()) {
2516                 *num_total_cards = -1;
2517                 *num_marked_cards = -1;
2518                 return;
2519         }
2520
2521         FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
2522                 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2523                 int i;
2524
2525                 if (!has_references)
2526                         continue;
2527
2528                 total_cards += CARDS_PER_BLOCK;
2529                 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2530                         if (cards [i])
2531                                 ++marked_cards;
2532                 }
2533         } END_FOREACH_BLOCK_NO_LOCK;
2534
2535         *num_total_cards = total_cards;
2536         *num_marked_cards = marked_cards;
2537 }
2538
2539 static void
2540 update_cardtable_mod_union (void)
2541 {
2542         MSBlockInfo *block;
2543
2544         FOREACH_BLOCK_NO_LOCK (block) {
2545                 gpointer *card_start = (gpointer*) sgen_card_table_get_card_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (block));
2546                 gboolean has_dirty_cards = FALSE;
2547                 int i;
2548                 for (i = 0; i < CARDS_PER_BLOCK / sizeof(gpointer); i++) {
2549                         if (card_start [i]) {
2550                                 has_dirty_cards = TRUE;
2551                                 break;
2552                         }
2553                 }
2554                 if (has_dirty_cards) {
2555                         size_t num_cards;
2556                         guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE);
2557                         sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2558                         SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2559                 }
2560         } END_FOREACH_BLOCK_NO_LOCK;
2561 }
2562
2563 #undef pthread_create
2564
2565 static void
2566 post_param_init (SgenMajorCollector *collector)
2567 {
2568         collector->sweeps_lazily = lazy_sweep;
2569         collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
2570 }
2571
2572 static void
2573 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
2574 {
2575         int i;
2576
2577         sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2578
2579         num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2580         block_obj_sizes = (int *)sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2581         ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2582
2583         evacuate_block_obj_sizes = (gboolean *)sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2584         for (i = 0; i < num_block_obj_sizes; ++i)
2585                 evacuate_block_obj_sizes [i] = FALSE;
2586
2587         sweep_slots_available = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2588         sweep_slots_used = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2589         sweep_num_blocks = (size_t *)sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2590
2591         /*
2592         {
2593                 int i;
2594                 g_print ("block object sizes:\n");
2595                 for (i = 0; i < num_block_obj_sizes; ++i)
2596                         g_print ("%d\n", block_obj_sizes [i]);
2597         }
2598         */
2599
2600         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2601                 free_block_lists [i] = (MSBlockInfo *volatile *)sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2602
2603         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2604                 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2605         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2606                 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2607
2608         mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2609         mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2610         mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2611 #if SIZEOF_VOID_P != 8
2612         mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2613         mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2614         mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2615         mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2616 #endif
2617
2618         collector->section_size = MAJOR_SECTION_SIZE;
2619
2620         concurrent_mark = is_concurrent;
2621         collector->is_concurrent = is_concurrent;
2622         collector->needs_thread_pool = is_concurrent || concurrent_sweep;
2623         collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2624         collector->supports_cardtable = TRUE;
2625
2626         collector->alloc_heap = major_alloc_heap;
2627         collector->is_object_live = major_is_object_live;
2628         collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2629         collector->alloc_degraded = major_alloc_degraded;
2630
2631         collector->alloc_object = major_alloc_object;
2632         collector->free_pinned_object = free_pinned_object;
2633         collector->iterate_objects = major_iterate_objects;
2634         collector->free_non_pinned_object = major_free_non_pinned_object;
2635         collector->pin_objects = major_pin_objects;
2636         collector->pin_major_object = pin_major_object;
2637         collector->scan_card_table = major_scan_card_table;
2638         collector->iterate_live_block_ranges = major_iterate_live_block_ranges;
2639         collector->iterate_block_ranges = major_iterate_block_ranges;
2640         if (is_concurrent) {
2641                 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2642                 collector->get_cardtable_mod_union_for_reference = major_get_cardtable_mod_union_for_reference;
2643         }
2644         collector->init_to_space = major_init_to_space;
2645         collector->sweep = major_sweep;
2646         collector->have_swept = major_have_swept;
2647         collector->finish_sweeping = major_finish_sweep_checking;
2648         collector->free_swept_blocks = major_free_swept_blocks;
2649         collector->check_scan_starts = major_check_scan_starts;
2650         collector->dump_heap = major_dump_heap;
2651         collector->get_used_size = major_get_used_size;
2652         collector->start_nursery_collection = major_start_nursery_collection;
2653         collector->finish_nursery_collection = major_finish_nursery_collection;
2654         collector->start_major_collection = major_start_major_collection;
2655         collector->finish_major_collection = major_finish_major_collection;
2656         collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2657         collector->ptr_is_from_pinned_alloc = ptr_is_from_pinned_alloc;
2658         collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2659         collector->get_num_major_sections = get_num_major_sections;
2660         collector->get_bytes_survived_last_sweep = get_bytes_survived_last_sweep;
2661         collector->handle_gc_param = major_handle_gc_param;
2662         collector->print_gc_param_usage = major_print_gc_param_usage;
2663         collector->post_param_init = post_param_init;
2664         collector->is_valid_object = major_is_valid_object;
2665         collector->describe_pointer = major_describe_pointer;
2666         collector->count_cards = major_count_cards;
2667
2668         collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical;
2669         collector->major_ops_serial.scan_object = major_scan_object_with_evacuation;
2670         collector->major_ops_serial.drain_gray_stack = drain_gray_stack;
2671         if (is_concurrent) {
2672                 collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2673                 collector->major_ops_concurrent_start.scan_object = major_scan_object_concurrent_with_evacuation;
2674                 collector->major_ops_concurrent_start.scan_vtype = major_scan_vtype_concurrent_with_evacuation;
2675                 collector->major_ops_concurrent_start.scan_ptr_field = major_scan_ptr_field_concurrent_with_evacuation;
2676                 collector->major_ops_concurrent_start.drain_gray_stack = drain_gray_stack_concurrent;
2677
2678                 collector->major_ops_concurrent_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_finish_canonical;
2679                 collector->major_ops_concurrent_finish.scan_object = major_scan_object_with_evacuation;
2680                 collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_with_evacuation;
2681                 collector->major_ops_concurrent_finish.scan_ptr_field = major_scan_ptr_field_with_evacuation;
2682                 collector->major_ops_concurrent_finish.drain_gray_stack = drain_gray_stack;
2683         }
2684
2685 #ifdef HEAVY_STATISTICS
2686         mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2687         mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2688         mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2689         mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2690         mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2691         mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2692         mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2693         mono_counters_register ("Optimized copy major small evacuate", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_evacuate);
2694         mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2695         mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2696         mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2697
2698         mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2699         mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2700         mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2701 #endif
2702
2703 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2704         mono_os_mutex_init (&scanned_objects_list_lock);
2705 #endif
2706
2707         SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2708
2709         /*cardtable requires major pages to be 8 cards aligned*/
2710         g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2711 }
2712
2713 void
2714 sgen_marksweep_init (SgenMajorCollector *collector)
2715 {
2716         sgen_marksweep_init_internal (collector, FALSE);
2717 }
2718
2719 void
2720 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2721 {
2722         sgen_marksweep_init_internal (collector, TRUE);
2723 }
2724
2725 #endif