mono/metadata/sgen-marksweep.c

   1 /*
   2  * sgen-marksweep.c: The Mark & Sweep major collector.
   3  *
   4  * Author:
   5  *      Mark Probst <mark.probst@gmail.com>
   6  *
   7  * Copyright 2009-2010 Novell, Inc.
   8  * Copyright (C) 2012 Xamarin Inc
   9  *
  10  * This library is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Library General Public
  12  * License 2.0 as published by the Free Software Foundation;
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Library General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Library General Public
  20  * License 2.0 along with this library; if not, write to the Free
  21  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #include "config.h"
  25
  26 #ifdef HAVE_SGEN_GC
  27
  28 #include <math.h>
  29 #include <errno.h>
  30
  31 #include "utils/mono-counters.h"
  32 #include "utils/mono-semaphore.h"
  33 #include "utils/mono-time.h"
  34 #include "metadata/object-internals.h"
  35 #include "metadata/profiler-private.h"
  36
  37 #include "metadata/sgen-gc.h"
  38 #include "metadata/sgen-protocol.h"
  39 #include "metadata/sgen-cardtable.h"
  40 #include "metadata/sgen-memory-governor.h"
  41 #include "metadata/sgen-layout-stats.h"
  42 #include "metadata/gc-internal.h"
  43 #include "metadata/sgen-pointer-queue.h"
  44 #include "metadata/sgen-pinning.h"
  45 #include "metadata/sgen-workers.h"
  46
  47 #if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT)
  48 #define MS_BLOCK_SIZE   ARCH_MIN_MS_BLOCK_SIZE
  49 #define MS_BLOCK_SIZE_SHIFT     ARCH_MIN_MS_BLOCK_SIZE_SHIFT
  50 #else
  51 #define MS_BLOCK_SIZE_SHIFT     14      /* INT FASTENABLE */
  52 #define MS_BLOCK_SIZE           (1 << MS_BLOCK_SIZE_SHIFT)
  53 #endif
  54 #define MAJOR_SECTION_SIZE      MS_BLOCK_SIZE
  55 #define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES)
  56
  57 /*
  58  * Don't allocate single blocks, but alloc a contingent of this many
  59  * blocks in one swoop.  This must be a power of two.
  60  */
  61 #define MS_BLOCK_ALLOC_NUM      32
  62
  63 /*
  64  * Number of bytes before the first object in a block.  At the start
  65  * of a block is the MSBlockHeader, then opional padding, then come
  66  * the objects, so this must be >= sizeof (MSBlockHeader).
  67  */
  68 #define MS_BLOCK_SKIP   ((sizeof (MSBlockHeader) + 15) & ~15)
  69
  70 #define MS_BLOCK_FREE   (MS_BLOCK_SIZE - MS_BLOCK_SKIP)
  71
  72 #define MS_NUM_MARK_WORDS       ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8))
  73
  74 /*
  75  * Blocks progress from one state to the next:
  76  *
  77  * SWEPT           The block is fully swept.  It might or might not be in
  78  *                 a free list.
  79  *
  80  * NOT_SWEPT       The block might or might not contain live objects.  If
  81  *                 we're in between an initial collection pause and the
  82  *                 finishing pause, the block might or might not be in a
  83  *                 free list.
  84  *
  85  * CHECKING        The sweep thread is investigating the block to determine
  86  *                 whether or not it contains live objects.  The block is
  87  *                 not in a free list.
  88  *
  89  * NEED_SWEEPING   The block contains live objects but has not yet been
  90  *                 swept.  It also contains free slots.  It is in a block
  91  *                 free list.
  92  *
  93  * SWEEPING        The block is being swept.  It might be in a free list.
  94  */
  95
  96 enum {
  97         BLOCK_STATE_SWEPT,
  98         BLOCK_STATE_MARKING,
  99         BLOCK_STATE_CHECKING,
 100         BLOCK_STATE_NEED_SWEEPING,
 101         BLOCK_STATE_SWEEPING
 102 };
 103
 104 typedef struct _MSBlockInfo MSBlockInfo;
 105 struct _MSBlockInfo {
 106         guint16 obj_size;
 107         /*
 108          * FIXME: Do we even need this? It's only used during sweep and might be worth
 109          * recalculating to save the space.
 110          */
 111         guint16 obj_size_index;
 112         /* FIXME: reduce this */
 113         volatile gint32 state;
 114         unsigned int pinned : 1;
 115         unsigned int has_references : 1;
 116         unsigned int has_pinned : 1;    /* means cannot evacuate */
 117         unsigned int is_to_space : 1;
 118         void ** volatile free_list;
 119         MSBlockInfo * volatile next_free;
 120         guint8 *cardtable_mod_union;
 121         mword mark_words [MS_NUM_MARK_WORDS];
 122 };
 123
 124 #define MS_BLOCK_FOR_BLOCK_INFO(b)      ((char*)(b))
 125
 126 #define MS_BLOCK_OBJ(b,i)               (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i))
 127 #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size)             (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i))
 128 #define MS_BLOCK_DATA_FOR_OBJ(o)        ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1)))
 129
 130 typedef struct {
 131         MSBlockInfo info;
 132 } MSBlockHeader;
 133
 134 #define MS_BLOCK_FOR_OBJ(o)             (&((MSBlockHeader*)MS_BLOCK_DATA_FOR_OBJ ((o)))->info)
 135
 136 /* object index will always be small */
 137 #define MS_BLOCK_OBJ_INDEX(o,b) ((int)(((char*)(o) - (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP)) / (b)->obj_size))
 138
 139 //casting to int is fine since blocks are 32k
 140 #define MS_CALC_MARK_BIT(w,b,o)         do {                            \
 141                 int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \
 142                 if (sizeof (mword) == 4) {                              \
 143                         (w) = i >> 5;                                   \
 144                         (b) = i & 31;                                   \
 145                 } else {                                                \
 146                         (w) = i >> 6;                                   \
 147                         (b) = i & 63;                                   \
 148                 }                                                       \
 149         } while (0)
 150
 151 #define MS_MARK_BIT(bl,w,b)     ((bl)->mark_words [(w)] & (ONE_P << (b)))
 152 #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b)))
 153
 154 #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))
 155
 156 #define MS_BLOCK_OBJ_SIZE_FACTOR        (pow (2.0, 1.0 / 3))
 157
 158 /*
 159  * This way we can lookup block object size indexes for sizes up to
 160  * 256 bytes with a single load.
 161  */
 162 #define MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES      32
 163
 164 static int *block_obj_sizes;
 165 static int num_block_obj_sizes;
 166 static int fast_block_obj_size_indexes [MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES];
 167
 168 #define MS_BLOCK_FLAG_PINNED    1
 169 #define MS_BLOCK_FLAG_REFS      2
 170
 171 #define MS_BLOCK_TYPE_MAX       4
 172
 173 static gboolean *evacuate_block_obj_sizes;
 174 static float evacuation_threshold = 0.666f;
 175 static float concurrent_evacuation_threshold = 0.666f;
 176 static gboolean want_evacuation = FALSE;
 177
 178 static gboolean lazy_sweep = FALSE;
 179
 180 enum {
 181         SWEEP_STATE_SWEPT,
 182         SWEEP_STATE_NEED_SWEEPING,
 183         SWEEP_STATE_SWEEPING,
 184         SWEEP_STATE_SWEEPING_AND_ITERATING,
 185         SWEEP_STATE_COMPACTING
 186 };
 187
 188 static volatile int sweep_state = SWEEP_STATE_SWEPT;
 189
 190 static gboolean concurrent_mark;
 191
 192 #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl)      SGEN_POINTER_IS_TAGGED_1 ((bl))
 193 #define BLOCK_TAG_HAS_REFERENCES(bl)            SGEN_POINTER_TAG_1 ((bl))
 194 #define BLOCK_UNTAG_HAS_REFERENCES(bl)          SGEN_POINTER_UNTAG_1 ((bl))
 195
 196 #define BLOCK_TAG(bl)   ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl))
 197
 198 /* all allocated blocks in the system */
 199 static SgenPointerQueue allocated_blocks;
 200 static mono_mutex_t allocated_blocks_lock;
 201
 202 #define LOCK_ALLOCATED_BLOCKS   mono_mutex_lock (&allocated_blocks_lock)
 203 #define UNLOCK_ALLOCATED_BLOCKS mono_mutex_unlock (&allocated_blocks_lock)
 204
 205 /* non-allocated block free-list */
 206 static void *empty_blocks = NULL;
 207 static size_t num_empty_blocks = 0;
 208
 209 #define FOREACH_BLOCK(bl)       { size_t __index; LOCK_ALLOCATED_BLOCKS; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [__index]);
 210 #define FOREACH_BLOCK_HAS_REFERENCES(bl,hr)     { size_t __index; LOCK_ALLOCATED_BLOCKS; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = allocated_blocks.data [__index]; (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl)); (bl) = BLOCK_UNTAG_HAS_REFERENCES ((bl));
 211 #define END_FOREACH_BLOCK       } UNLOCK_ALLOCATED_BLOCKS; }
 212
 213 #define FOREACH_BLOCK_NO_LOCK(bl)       { size_t __index; for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [__index]);
 214 #define FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK(bl,hr)     { size_t __index; SGEN_ASSERT (0, sgen_is_world_stopped (), "Can't iterate blocks without lock when world is running."); for (__index = 0; __index < allocated_blocks.next_slot; ++__index) { (bl) = allocated_blocks.data [__index]; (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl)); (bl) = BLOCK_UNTAG_HAS_REFERENCES ((bl));
 215 #define END_FOREACH_BLOCK_NO_LOCK       } }
 216
 217 static volatile size_t num_major_sections = 0;
 218 /*
 219  * One free block list for each block object size.  We add and remove blocks from these
 220  * lists lock-free via CAS.
 221  *
 222  * Blocks accessed/removed from `free_block_lists`:
 223  *   from the mutator (with GC lock held)
 224  *   in nursery collections
 225  *   in non-concurrent major collections
 226  *   in the finishing pause of concurrent major collections (whole list is cleared)
 227  *
 228  * Blocks added to `free_block_lists`:
 229  *   in the sweeping thread
 230  *   during nursery collections
 231  *   from domain clearing (with the world stopped and no sweeping happening)
 232  *
 233  * The only item of those that doesn't require the GC lock is the sweep thread.  The sweep
 234  * thread only ever adds blocks to the free list, so the ABA problem can't occur.
 235  */
 236 static MSBlockInfo * volatile *free_block_lists [MS_BLOCK_TYPE_MAX];
 237
 238 static guint64 stat_major_blocks_alloced = 0;
 239 static guint64 stat_major_blocks_freed = 0;
 240 static guint64 stat_major_blocks_lazy_swept = 0;
 241 static guint64 stat_major_objects_evacuated = 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_mutex_lock (&scanned_objects_list_lock);
 268         sgen_pointer_queue_add (&scanned_objects_list, ptr);
 269         mono_mutex_unlock (&scanned_objects_list_lock);
 270 }
 271 #endif
 272
 273 static void
 274 sweep_block (MSBlockInfo *block);
 275
 276 static int
 277 ms_find_block_obj_size_index (size_t size)
 278 {
 279         int i;
 280         SGEN_ASSERT (9, size <= SGEN_MAX_SMALL_OBJ_SIZE, "size %d is bigger than max small object size %d", size, SGEN_MAX_SMALL_OBJ_SIZE);
 281         for (i = 0; i < num_block_obj_sizes; ++i)
 282                 if (block_obj_sizes [i] >= size)
 283                         return i;
 284         g_error ("no object of size %d\n", size);
 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 = sgen_alloc_os_memory_aligned (nursery_size, nursery_align, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE, "nursery");
 301         else
 302                 start = sgen_alloc_os_memory (nursery_size, 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 = sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE,
 333                                         alloc_num == 1 ? "major heap section" : NULL);
 334                         if (p)
 335                                 break;
 336                         alloc_num >>= 1;
 337                 }
 338
 339                 for (i = 0; i < alloc_num; ++i) {
 340                         block = p;
 341                         /*
 342                          * We do the free list update one after the
 343                          * other so that other threads can use the new
 344                          * blocks as quickly as possible.
 345                          */
 346                         do {
 347                                 empty = empty_blocks;
 348                                 *(void**)block = empty;
 349                         } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty);
 350                         p += MS_BLOCK_SIZE;
 351                 }
 352
 353                 SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num);
 354
 355                 stat_major_blocks_alloced += alloc_num;
 356 #if SIZEOF_VOID_P != 8
 357                 if (alloc_num != MS_BLOCK_ALLOC_NUM)
 358                         stat_major_blocks_alloced_less_ideal += alloc_num;
 359 #endif
 360         }
 361
 362         do {
 363                 empty = empty_blocks;
 364                 if (!empty)
 365                         goto retry;
 366                 block = empty;
 367                 next = *(void**)block;
 368         } while (SGEN_CAS_PTR (&empty_blocks, next, empty) != empty);
 369
 370         SGEN_ATOMIC_ADD_P (num_empty_blocks, -1);
 371
 372         *(void**)block = NULL;
 373
 374         g_assert (!((mword)block & (MS_BLOCK_SIZE - 1)));
 375
 376         return block;
 377 }
 378
 379 /*
 380  * This doesn't actually free a block immediately, but enqueues it into the `empty_blocks`
 381  * list, where it will either be freed later on, or reused in nursery collections.
 382  */
 383 static void
 384 ms_free_block (void *block)
 385 {
 386         void *empty;
 387
 388         sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR);
 389         memset (block, 0, MS_BLOCK_SIZE);
 390
 391         do {
 392                 empty = empty_blocks;
 393                 *(void**)block = empty;
 394         } while (SGEN_CAS_PTR (&empty_blocks, block, empty) != empty);
 395
 396         SGEN_ATOMIC_ADD_P (num_empty_blocks, 1);
 397
 398         binary_protocol_block_free (block, MS_BLOCK_SIZE);
 399 }
 400
 401 static gboolean
 402 sweep_in_progress (void)
 403 {
 404         int state = sweep_state;
 405         return state == SWEEP_STATE_SWEEPING ||
 406                 state == SWEEP_STATE_SWEEPING_AND_ITERATING ||
 407                 state == SWEEP_STATE_COMPACTING;
 408 }
 409
 410 //#define MARKSWEEP_CONSISTENCY_CHECK
 411
 412 #ifdef MARKSWEEP_CONSISTENCY_CHECK
 413 static void
 414 check_block_free_list (MSBlockInfo *block, int size, gboolean pinned)
 415 {
 416         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 417         for (; block; block = block->next_free) {
 418                 SGEN_ASSERT (0, block->state != BLOCK_STATE_CHECKING, "Can't have a block we're checking in a free list.");
 419                 g_assert (block->obj_size == size);
 420                 g_assert ((pinned && block->pinned) || (!pinned && !block->pinned));
 421
 422                 /* blocks in the free lists must have at least
 423                    one free slot */
 424                 g_assert (block->free_list);
 425
 426                 /* the block must be in the allocated_blocks array */
 427                 g_assert (sgen_pointer_queue_find (&allocated_blocks, BLOCK_TAG (block)) != (size_t)-1);
 428         }
 429 }
 430
 431 static void
 432 check_empty_blocks (void)
 433 {
 434         void *p;
 435         size_t i = 0;
 436         for (p = empty_blocks; p; p = *(void**)p)
 437                 ++i;
 438         g_assert (i == num_empty_blocks);
 439 }
 440
 441 static void
 442 consistency_check (void)
 443 {
 444         MSBlockInfo *block;
 445         int i;
 446
 447         /* check all blocks */
 448         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 449         FOREACH_BLOCK_NO_LOCK (block) {
 450                 int count = MS_BLOCK_FREE / block->obj_size;
 451                 int num_free = 0;
 452                 void **free;
 453
 454                 /* count number of free slots */
 455                 for (i = 0; i < count; ++i) {
 456                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 457                         if (!MS_OBJ_ALLOCED (obj, block))
 458                                 ++num_free;
 459                 }
 460
 461                 /* check free list */
 462                 for (free = block->free_list; free; free = (void**)*free) {
 463                         g_assert (MS_BLOCK_FOR_OBJ (free) == block);
 464                         --num_free;
 465                 }
 466                 g_assert (num_free == 0);
 467
 468                 /* check all mark words are zero */
 469                 if (!sgen_concurrent_collection_in_progress () && (block->state == BLOCK_STATE_SWEPT || block->state == BLOCK_STATE_MARKING)) {
 470                         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
 471                                 g_assert (block->mark_words [i] == 0);
 472                 }
 473         } END_FOREACH_BLOCK_NO_LOCK;
 474
 475         /* check free blocks */
 476         for (i = 0; i < num_block_obj_sizes; ++i) {
 477                 int j;
 478                 for (j = 0; j < MS_BLOCK_TYPE_MAX; ++j)
 479                         check_block_free_list (free_block_lists [j][i], block_obj_sizes [i], j & MS_BLOCK_FLAG_PINNED);
 480         }
 481
 482         check_empty_blocks ();
 483 }
 484 #endif
 485
 486 static void
 487 add_free_block (MSBlockInfo * volatile *free_blocks, int size_index, MSBlockInfo *block)
 488 {
 489         MSBlockInfo *old;
 490         do {
 491                 block->next_free = old = free_blocks [size_index];
 492         } while (SGEN_CAS_PTR ((gpointer)&free_blocks [size_index], block, old) != old);
 493 }
 494
 495 static gboolean
 496 ms_alloc_block (int size_index, gboolean pinned, gboolean has_references)
 497 {
 498         int size = block_obj_sizes [size_index];
 499         int count = MS_BLOCK_FREE / size;
 500         MSBlockInfo *info;
 501         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 502         char *obj_start;
 503         int i;
 504
 505         if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR))
 506                 return FALSE;
 507
 508         info = (MSBlockInfo*)ms_get_empty_block ();
 509
 510         SGEN_ASSERT (9, count >= 2, "block with %d objects, it must hold at least 2", count);
 511
 512         info->obj_size = size;
 513         info->obj_size_index = size_index;
 514         info->pinned = pinned;
 515         info->has_references = has_references;
 516         info->has_pinned = pinned;
 517         /*
 518          * Blocks that are to-space are not evacuated from.  During an major collection
 519          * blocks are allocated for two reasons: evacuating objects from the nursery and
 520          * evacuating them from major blocks marked for evacuation.  In both cases we don't
 521          * want further evacuation.
 522          */
 523         info->is_to_space = (sgen_get_current_collection_generation () == GENERATION_OLD);
 524         info->state = (info->is_to_space || sgen_concurrent_collection_in_progress ()) ? BLOCK_STATE_MARKING : BLOCK_STATE_SWEPT;
 525         if (sweep_in_progress ())
 526                 SGEN_ASSERT (0, info->state == BLOCK_STATE_SWEPT, "How do we add a new block to be swept while sweeping?");
 527         info->cardtable_mod_union = NULL;
 528
 529         update_heap_boundaries_for_block (info);
 530
 531         binary_protocol_block_alloc (info, MS_BLOCK_SIZE);
 532
 533         /* build free list */
 534         obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP;
 535         info->free_list = (void**)obj_start;
 536         /* we're skipping the last one - it must be nulled */
 537         for (i = 0; i < count - 1; ++i) {
 538                 char *next_obj_start = obj_start + size;
 539                 *(void**)obj_start = next_obj_start;
 540                 obj_start = next_obj_start;
 541         }
 542         /* the last one */
 543         *(void**)obj_start = NULL;
 544
 545         add_free_block (free_blocks, size_index, info);
 546
 547         LOCK_ALLOCATED_BLOCKS;
 548         sgen_pointer_queue_add (&allocated_blocks, BLOCK_TAG (info));
 549         UNLOCK_ALLOCATED_BLOCKS;
 550
 551         SGEN_ATOMIC_ADD_P (num_major_sections, 1);
 552         return TRUE;
 553 }
 554
 555 static gboolean
 556 obj_is_from_pinned_alloc (char *ptr)
 557 {
 558         MSBlockInfo *block;
 559
 560         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 561         FOREACH_BLOCK_NO_LOCK (block) {
 562                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE)
 563                         return block->pinned;
 564         } END_FOREACH_BLOCK_NO_LOCK;
 565         return FALSE;
 566 }
 567
 568 static void
 569 ensure_can_access_block_free_list (MSBlockInfo *block)
 570 {
 571  retry:
 572         for (;;) {
 573                 switch (block->state) {
 574                 case BLOCK_STATE_SWEPT:
 575                 case BLOCK_STATE_MARKING:
 576                         return;
 577                 case BLOCK_STATE_CHECKING:
 578                         SGEN_ASSERT (0, FALSE, "How did we get a block that's being checked from a free list?");
 579                         break;
 580                 case BLOCK_STATE_NEED_SWEEPING:
 581                         stat_major_blocks_lazy_swept ++;
 582                         sweep_block (block);
 583                         break;
 584                 case BLOCK_STATE_SWEEPING:
 585                         /* FIXME: do this more elegantly */
 586                         g_usleep (100);
 587                         goto retry;
 588                 default:
 589                         SGEN_ASSERT (0, FALSE, "Illegal block state");
 590                         break;
 591                 }
 592         }
 593 }
 594
 595 static void*
 596 unlink_slot_from_free_list_uncontested (MSBlockInfo * volatile *free_blocks, int size_index)
 597 {
 598         MSBlockInfo *block, *next_free_block;
 599         void *obj, *next_free_slot;
 600
 601  retry:
 602         block = free_blocks [size_index];
 603         SGEN_ASSERT (9, block, "no free block to unlink from free_blocks %p size_index %d", free_blocks, size_index);
 604
 605         ensure_can_access_block_free_list (block);
 606
 607         obj = block->free_list;
 608         SGEN_ASSERT (0, obj, "block %p in free list had no available object to alloc from", block);
 609
 610         next_free_slot = *(void**)obj;
 611         if (next_free_slot) {
 612                 block->free_list = next_free_slot;
 613                 return obj;
 614         }
 615
 616         next_free_block = block->next_free;
 617         if (SGEN_CAS_PTR ((gpointer)&free_blocks [size_index], next_free_block, block) != block)
 618                 goto retry;
 619
 620         block->free_list = NULL;
 621         block->next_free = NULL;
 622
 623         return obj;
 624 }
 625
 626 static void*
 627 alloc_obj (MonoVTable *vtable, size_t size, gboolean pinned, gboolean has_references)
 628 {
 629         int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 630         MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (pinned, has_references);
 631         void *obj;
 632
 633         if (!free_blocks [size_index]) {
 634                 if (G_UNLIKELY (!ms_alloc_block (size_index, pinned, has_references)))
 635                         return NULL;
 636         }
 637
 638         obj = unlink_slot_from_free_list_uncontested (free_blocks, size_index);
 639
 640         *(MonoVTable**)obj = vtable;
 641
 642         return obj;
 643 }
 644
 645 static void*
 646 major_alloc_object (MonoVTable *vtable, size_t size, gboolean has_references)
 647 {
 648         return alloc_obj (vtable, size, FALSE, has_references);
 649 }
 650
 651 /*
 652  * We're not freeing the block if it's empty.  We leave that work for
 653  * the next major collection.
 654  *
 655  * This is just called from the domain clearing code, which runs in a
 656  * single thread and has the GC lock, so we don't need an extra lock.
 657  */
 658 static void
 659 free_object (char *obj, size_t size, gboolean pinned)
 660 {
 661         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
 662         int word, bit;
 663         gboolean in_free_list;
 664
 665         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Should have waited for sweep to free objects.");
 666
 667         ensure_can_access_block_free_list (block);
 668         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);
 669         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p is already free", obj);
 670         MS_CALC_MARK_BIT (word, bit, obj);
 671         SGEN_ASSERT (9, !MS_MARK_BIT (block, word, bit), "object %p has mark bit set");
 672
 673         memset (obj, 0, size);
 674
 675         in_free_list = !!block->free_list;
 676         *(void**)obj = block->free_list;
 677         block->free_list = (void**)obj;
 678
 679         if (!in_free_list) {
 680                 MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (pinned, block->has_references);
 681                 int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size);
 682                 SGEN_ASSERT (9, !block->next_free, "block %p doesn't have a free-list of object but belongs to a free-list of blocks");
 683                 add_free_block (free_blocks, size_index, block);
 684         }
 685 }
 686
 687 static void
 688 major_free_non_pinned_object (char *obj, size_t size)
 689 {
 690         free_object (obj, size, FALSE);
 691 }
 692
 693 /* size is a multiple of SGEN_ALLOC_ALIGN */
 694 static void*
 695 major_alloc_small_pinned_obj (MonoVTable *vtable, size_t size, gboolean has_references)
 696 {
 697         void *res;
 698
 699         res = alloc_obj (vtable, size, TRUE, has_references);
 700          /*If we failed to alloc memory, we better try releasing memory
 701           *as pinned alloc is requested by the runtime.
 702           */
 703          if (!res) {
 704                 sgen_perform_collection (0, GENERATION_OLD, "pinned alloc failure", TRUE);
 705                 res = alloc_obj (vtable, size, TRUE, has_references);
 706          }
 707          return res;
 708 }
 709
 710 static void
 711 free_pinned_object (char *obj, size_t size)
 712 {
 713         free_object (obj, size, TRUE);
 714 }
 715
 716 /*
 717  * size is already rounded up and we hold the GC lock.
 718  */
 719 static void*
 720 major_alloc_degraded (MonoVTable *vtable, size_t size)
 721 {
 722         void *obj = alloc_obj (vtable, size, FALSE, SGEN_VTABLE_HAS_REFERENCES (vtable));
 723         if (G_LIKELY (obj)) {
 724                 HEAVY_STAT (++stat_objects_alloced_degraded);
 725                 HEAVY_STAT (stat_bytes_alloced_degraded += size);
 726         }
 727         return obj;
 728 }
 729
 730 /*
 731  * obj is some object.  If it's not in the major heap (i.e. if it's in
 732  * the nursery or LOS), return FALSE.  Otherwise return whether it's
 733  * been marked or copied.
 734  */
 735 static gboolean
 736 major_is_object_live (char *obj)
 737 {
 738         MSBlockInfo *block;
 739         int word, bit;
 740         mword objsize;
 741
 742         if (sgen_ptr_in_nursery (obj))
 743                 return FALSE;
 744
 745         objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
 746
 747         /* LOS */
 748         if (objsize > SGEN_MAX_SMALL_OBJ_SIZE)
 749                 return FALSE;
 750
 751         /* now we know it's in a major block */
 752         block = MS_BLOCK_FOR_OBJ (obj);
 753         SGEN_ASSERT (9, !block->pinned, "block %p is pinned, BTW why is this bad?");
 754         MS_CALC_MARK_BIT (word, bit, obj);
 755         return MS_MARK_BIT (block, word, bit) ? TRUE : FALSE;
 756 }
 757
 758 static gboolean
 759 major_ptr_is_in_non_pinned_space (char *ptr, char **start)
 760 {
 761         MSBlockInfo *block;
 762
 763         SGEN_ASSERT (0, !sweep_in_progress (), "Can't examine allocated blocks during sweep");
 764         FOREACH_BLOCK_NO_LOCK (block) {
 765                 if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) {
 766                         int count = MS_BLOCK_FREE / block->obj_size;
 767                         int i;
 768
 769                         *start = NULL;
 770                         for (i = 0; i <= count; ++i) {
 771                                 if (ptr >= MS_BLOCK_OBJ (block, i) && ptr < MS_BLOCK_OBJ (block, i + 1)) {
 772                                         *start = MS_BLOCK_OBJ (block, i);
 773                                         break;
 774                                 }
 775                         }
 776                         return !block->pinned;
 777                 }
 778         } END_FOREACH_BLOCK_NO_LOCK;
 779         return FALSE;
 780 }
 781
 782 static void
 783 major_finish_sweeping (void)
 784 {
 785         if (!sweep_in_progress ())
 786                 return;
 787
 788         while (sweep_in_progress ())
 789                 g_usleep (100);
 790 }
 791
 792 static void
 793 major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc callback, void *data)
 794 {
 795         gboolean sweep = flags & ITERATE_OBJECTS_SWEEP;
 796         gboolean non_pinned = flags & ITERATE_OBJECTS_NON_PINNED;
 797         gboolean pinned = flags & ITERATE_OBJECTS_PINNED;
 798         MSBlockInfo *block;
 799
 800         major_finish_sweeping ();
 801         FOREACH_BLOCK (block) {
 802                 int count = MS_BLOCK_FREE / block->obj_size;
 803                 int i;
 804
 805                 if (block->pinned && !pinned)
 806                         continue;
 807                 if (!block->pinned && !non_pinned)
 808                         continue;
 809                 if (sweep && lazy_sweep) {
 810                         /* FIXME: We can't just call `sweep_block` willy-nilly. */
 811                         sweep_block (block);
 812                         SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping");
 813                 }
 814
 815                 for (i = 0; i < count; ++i) {
 816                         void **obj = (void**) MS_BLOCK_OBJ (block, i);
 817                         /* FIXME: This condition is probably incorrect. */
 818                         if (block->state != BLOCK_STATE_SWEPT && block->state != BLOCK_STATE_MARKING) {
 819                                 int word, bit;
 820                                 MS_CALC_MARK_BIT (word, bit, obj);
 821                                 if (!MS_MARK_BIT (block, word, bit))
 822                                         continue;
 823                         }
 824                         if (MS_OBJ_ALLOCED (obj, block))
 825                                 callback ((char*)obj, block->obj_size, data);
 826                 }
 827         } END_FOREACH_BLOCK;
 828 }
 829
 830 static gboolean
 831 major_is_valid_object (char *object)
 832 {
 833         MSBlockInfo *block;
 834
 835         SGEN_ASSERT (0, !sweep_in_progress (), "Can't iterate blocks during sweep");
 836         FOREACH_BLOCK_NO_LOCK (block) {
 837                 int idx;
 838                 char *obj;
 839
 840                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object))
 841                         continue;
 842
 843                 idx = MS_BLOCK_OBJ_INDEX (object, block);
 844                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 845                 if (obj != object)
 846                         return FALSE;
 847                 return MS_OBJ_ALLOCED (obj, block);
 848         } END_FOREACH_BLOCK_NO_LOCK;
 849
 850         return FALSE;
 851 }
 852
 853
 854 static MonoVTable*
 855 major_describe_pointer (char *ptr)
 856 {
 857         MSBlockInfo *block;
 858
 859         FOREACH_BLOCK_NO_LOCK (block) {
 860                 int idx;
 861                 char *obj;
 862                 gboolean live;
 863                 MonoVTable *vtable;
 864                 int w, b;
 865                 gboolean marked;
 866
 867                 if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr))
 868                         continue;
 869
 870                 SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n",
 871                         MS_BLOCK_FOR_BLOCK_INFO (block), block->obj_size, block->pinned, block->has_references);
 872
 873                 idx = MS_BLOCK_OBJ_INDEX (ptr, block);
 874                 obj = (char*)MS_BLOCK_OBJ (block, idx);
 875                 live = MS_OBJ_ALLOCED (obj, block);
 876                 vtable = live ? (MonoVTable*)SGEN_LOAD_VTABLE (obj) : NULL;
 877
 878                 MS_CALC_MARK_BIT (w, b, obj);
 879                 marked = MS_MARK_BIT (block, w, b);
 880
 881                 if (obj == ptr) {
 882                         SGEN_LOG (0, "\t(");
 883                         if (live)
 884                                 SGEN_LOG (0, "object");
 885                         else
 886                                 SGEN_LOG (0, "dead-object");
 887                 } else {
 888                         if (live)
 889                                 SGEN_LOG (0, "interior-ptr offset %td", ptr - obj);
 890                         else
 891                                 SGEN_LOG (0, "dead-interior-ptr offset %td", ptr - obj);
 892                 }
 893
 894                 SGEN_LOG (0, " marked %d)\n", marked ? 1 : 0);
 895
 896                 return vtable;
 897         } END_FOREACH_BLOCK_NO_LOCK;
 898
 899         return NULL;
 900 }
 901
 902 static void
 903 major_check_scan_starts (void)
 904 {
 905 }
 906
 907 static void
 908 major_dump_heap (FILE *heap_dump_file)
 909 {
 910         MSBlockInfo *block;
 911         int *slots_available = alloca (sizeof (int) * num_block_obj_sizes);
 912         int *slots_used = alloca (sizeof (int) * num_block_obj_sizes);
 913         int i;
 914
 915         for (i = 0; i < num_block_obj_sizes; ++i)
 916                 slots_available [i] = slots_used [i] = 0;
 917
 918         SGEN_ASSERT (0, !sweep_in_progress (), "Can't iterate blocks during sweep");
 919         FOREACH_BLOCK (block) {
 920                 int index = ms_find_block_obj_size_index (block->obj_size);
 921                 int count = MS_BLOCK_FREE / block->obj_size;
 922
 923                 slots_available [index] += count;
 924                 for (i = 0; i < count; ++i) {
 925                         if (MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block))
 926                                 ++slots_used [index];
 927                 }
 928         } END_FOREACH_BLOCK;
 929
 930         fprintf (heap_dump_file, "<occupancies>\n");
 931         for (i = 0; i < num_block_obj_sizes; ++i) {
 932                 fprintf (heap_dump_file, "<occupancy size=\"%d\" available=\"%d\" used=\"%d\" />\n",
 933                                 block_obj_sizes [i], slots_available [i], slots_used [i]);
 934         }
 935         fprintf (heap_dump_file, "</occupancies>\n");
 936
 937         FOREACH_BLOCK (block) {
 938                 int count = MS_BLOCK_FREE / block->obj_size;
 939                 int i;
 940                 int start = -1;
 941
 942                 fprintf (heap_dump_file, "<section type=\"%s\" size=\"%zu\">\n", "old", (size_t)MS_BLOCK_FREE);
 943
 944                 for (i = 0; i <= count; ++i) {
 945                         if ((i < count) && MS_OBJ_ALLOCED (MS_BLOCK_OBJ (block, i), block)) {
 946                                 if (start < 0)
 947                                         start = i;
 948                         } else {
 949                                 if (start >= 0) {
 950                                         sgen_dump_occupied (MS_BLOCK_OBJ (block, start), MS_BLOCK_OBJ (block, i), MS_BLOCK_FOR_BLOCK_INFO (block));
 951                                         start = -1;
 952                                 }
 953                         }
 954                 }
 955
 956                 fprintf (heap_dump_file, "</section>\n");
 957         } END_FOREACH_BLOCK;
 958 }
 959
 960 #define LOAD_VTABLE     SGEN_LOAD_VTABLE
 961
 962 #define MS_MARK_OBJECT_AND_ENQUEUE_CHECKED(obj,desc,block,queue) do {   \
 963                 int __word, __bit;                                      \
 964                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
 965                 if (!MS_MARK_BIT ((block), __word, __bit) && MS_OBJ_ALLOCED ((obj), (block))) { \
 966                         MS_SET_MARK_BIT ((block), __word, __bit);       \
 967                         if (sgen_gc_descr_has_references (desc))                        \
 968                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
 969                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((MonoObject*)(obj))); \
 970                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
 971                 }                                                       \
 972         } while (0)
 973 #define MS_MARK_OBJECT_AND_ENQUEUE(obj,desc,block,queue) do {           \
 974                 int __word, __bit;                                      \
 975                 MS_CALC_MARK_BIT (__word, __bit, (obj));                \
 976                 SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \
 977                 if (!MS_MARK_BIT ((block), __word, __bit)) {            \
 978                         MS_SET_MARK_BIT ((block), __word, __bit);       \
 979                         if (sgen_gc_descr_has_references (desc))                        \
 980                                 GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \
 981                         binary_protocol_mark ((obj), (gpointer)LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((MonoObject*)(obj))); \
 982                         INC_NUM_MAJOR_OBJECTS_MARKED ();                \
 983                 }                                                       \
 984         } while (0)
 985
 986 static void
 987 pin_major_object (char *obj, SgenGrayQueue *queue)
 988 {
 989         MSBlockInfo *block;
 990
 991         if (concurrent_mark)
 992                 g_assert_not_reached ();
 993
 994         block = MS_BLOCK_FOR_OBJ (obj);
 995         block->has_pinned = TRUE;
 996         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
 997 }
 998
 999 #include "sgen-major-copy-object.h"
1000
1001 static void
1002 major_copy_or_mark_object_with_evacuation_concurrent (void **ptr, void *obj, SgenGrayQueue *queue)
1003 {
1004         SGEN_ASSERT (9, sgen_concurrent_collection_in_progress (), "Why are we scanning concurrently when there's no concurrent collection on?");
1005         SGEN_ASSERT (9, !sgen_workers_are_working () || sgen_is_worker_thread (mono_native_thread_id_get ()), "We must not scan from two threads at the same time!");
1006
1007         g_assert (!SGEN_OBJECT_IS_FORWARDED (obj));
1008
1009         if (!sgen_ptr_in_nursery (obj)) {
1010                 mword objsize;
1011
1012                 objsize = SGEN_ALIGN_UP (sgen_safe_object_get_size ((MonoObject*)obj));
1013
1014                 if (objsize <= SGEN_MAX_SMALL_OBJ_SIZE) {
1015                         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
1016                         MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue);
1017                 } else {
1018                         if (sgen_los_object_is_pinned (obj))
1019                                 return;
1020
1021 #ifdef ENABLE_DTRACE
1022                         if (G_UNLIKELY (MONO_GC_OBJ_PINNED_ENABLED ())) {
1023                                 MonoVTable *vt = (MonoVTable*)SGEN_LOAD_VTABLE (obj);
1024                                 MONO_GC_OBJ_PINNED ((mword)obj, sgen_safe_object_get_size (obj), vt->klass->name_space, vt->klass->name, GENERATION_OLD);
1025                         }
1026 #endif
1027
1028                         sgen_los_pin_object (obj);
1029                         if (SGEN_OBJECT_HAS_REFERENCES (obj))
1030                                 GRAY_OBJECT_ENQUEUE (queue, obj, sgen_obj_get_descriptor (obj));
1031                         INC_NUM_MAJOR_OBJECTS_MARKED ();
1032                 }
1033         }
1034 }
1035
1036 static long long
1037 major_get_and_reset_num_major_objects_marked (void)
1038 {
1039 #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED
1040         long long num = num_major_objects_marked;
1041         num_major_objects_marked = 0;
1042         return num;
1043 #else
1044         return 0;
1045 #endif
1046 }
1047
1048 #define PREFETCH_CARDS          1       /* BOOL FASTENABLE */
1049 #if !PREFETCH_CARDS
1050 #undef PREFETCH_CARDS
1051 #endif
1052
1053 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
1054 #if defined(PLATFORM_MACOSX)
1055 #define GCC_VERSION (__GNUC__ * 10000 \
1056                                + __GNUC_MINOR__ * 100 \
1057                                + __GNUC_PATCHLEVEL__)
1058 #if GCC_VERSION <= 40300
1059 #undef PREFETCH_CARDS
1060 #endif
1061 #endif
1062
1063 #ifdef HEAVY_STATISTICS
1064 static guint64 stat_optimized_copy;
1065 static guint64 stat_optimized_copy_nursery;
1066 static guint64 stat_optimized_copy_nursery_forwarded;
1067 static guint64 stat_optimized_copy_nursery_pinned;
1068 static guint64 stat_optimized_copy_major;
1069 static guint64 stat_optimized_copy_major_small_fast;
1070 static guint64 stat_optimized_copy_major_small_slow;
1071 static guint64 stat_optimized_copy_major_large;
1072 static guint64 stat_optimized_copy_major_forwarded;
1073 static guint64 stat_optimized_copy_major_small_evacuate;
1074 static guint64 stat_optimized_major_scan;
1075 static guint64 stat_optimized_major_scan_no_refs;
1076
1077 static guint64 stat_drain_prefetch_fills;
1078 static guint64 stat_drain_prefetch_fill_failures;
1079 static guint64 stat_drain_loops;
1080 #endif
1081
1082 static void major_scan_object_with_evacuation (char *start, mword desc, SgenGrayQueue *queue);
1083
1084 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_no_evacuation
1085 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_no_evacuation
1086 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_no_evacuation
1087 #include "sgen-marksweep-drain-gray-stack.h"
1088
1089 #define COPY_OR_MARK_WITH_EVACUATION
1090 #define COPY_OR_MARK_FUNCTION_NAME      major_copy_or_mark_object_with_evacuation
1091 #define SCAN_OBJECT_FUNCTION_NAME       major_scan_object_with_evacuation
1092 #define DRAIN_GRAY_STACK_FUNCTION_NAME  drain_gray_stack_with_evacuation
1093 #include "sgen-marksweep-drain-gray-stack.h"
1094
1095 static gboolean
1096 drain_gray_stack (ScanCopyContext ctx)
1097 {
1098         gboolean evacuation = FALSE;
1099         int i;
1100         for (i = 0; i < num_block_obj_sizes; ++i) {
1101                 if (evacuate_block_obj_sizes [i]) {
1102                         evacuation = TRUE;
1103                         break;
1104                 }
1105         }
1106
1107         if (evacuation)
1108                 return drain_gray_stack_with_evacuation (ctx);
1109         else
1110                 return drain_gray_stack_no_evacuation (ctx);
1111 }
1112
1113 #include "sgen-marksweep-scan-object-concurrent.h"
1114
1115 static void
1116 major_copy_or_mark_object_canonical (void **ptr, SgenGrayQueue *queue)
1117 {
1118         major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue);
1119 }
1120
1121 static void
1122 major_copy_or_mark_object_concurrent_canonical (void **ptr, SgenGrayQueue *queue)
1123 {
1124         major_copy_or_mark_object_with_evacuation_concurrent (ptr, *ptr, queue);
1125 }
1126
1127 static void
1128 mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue)
1129 {
1130         void **entry, **end;
1131         int last_index = -1;
1132
1133         if (first_entry == last_entry)
1134                 return;
1135
1136         block->has_pinned = TRUE;
1137
1138         entry = sgen_pinning_get_entry (first_entry);
1139         end = sgen_pinning_get_entry (last_entry);
1140
1141         for (; entry < end; ++entry) {
1142                 int index = MS_BLOCK_OBJ_INDEX (*entry, block);
1143                 char *obj;
1144                 SGEN_ASSERT (9, index >= 0 && index < MS_BLOCK_FREE / block->obj_size, "invalid object %p index %d max-index %d", *entry, index, MS_BLOCK_FREE / block->obj_size);
1145                 if (index == last_index)
1146                         continue;
1147                 obj = MS_BLOCK_OBJ (block, index);
1148                 MS_MARK_OBJECT_AND_ENQUEUE_CHECKED (obj, sgen_obj_get_descriptor (obj), block, queue);
1149                 last_index = index;
1150         }
1151 }
1152
1153 static inline void
1154 sweep_block_for_size (MSBlockInfo *block, int count, int obj_size)
1155 {
1156         int obj_index;
1157
1158         for (obj_index = 0; obj_index < count; ++obj_index) {
1159                 int word, bit;
1160                 void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size);
1161
1162                 MS_CALC_MARK_BIT (word, bit, obj);
1163                 if (MS_MARK_BIT (block, word, bit)) {
1164                         SGEN_ASSERT (9, MS_OBJ_ALLOCED (obj, block), "object %p not allocated", obj);
1165                 } else {
1166                         /* an unmarked object */
1167                         if (MS_OBJ_ALLOCED (obj, block)) {
1168                                 /*
1169                                  * FIXME: Merge consecutive
1170                                  * slots for lower reporting
1171                                  * overhead.  Maybe memset
1172                                  * will also benefit?
1173                                  */
1174                                 binary_protocol_empty (obj, obj_size);
1175                                 MONO_GC_MAJOR_SWEPT ((mword)obj, obj_size);
1176                                 memset (obj, 0, obj_size);
1177                         }
1178                         *(void**)obj = block->free_list;
1179                         block->free_list = obj;
1180                 }
1181         }
1182 }
1183
1184 static gboolean
1185 try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1186 {
1187         gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state);
1188         gboolean success = old_state == expected_state;
1189         if (success)
1190                 binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state);
1191         return success;
1192 }
1193
1194 /*
1195  * FIXME: This only CASes to catch errors.  It's not needed for correctness.
1196  */
1197 static void
1198 set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state)
1199 {
1200         gboolean success = try_set_block_state (block, new_state, expected_state);
1201         SGEN_ASSERT (0, success, "Couldn't set block state");
1202         SGEN_ASSERT (0, block->state == new_state, "Block state incorrect after set");
1203 }
1204
1205 /*
1206  * sweep_block:
1207  *
1208  *   Traverse BLOCK, freeing and zeroing unused objects.
1209  */
1210 static void
1211 sweep_block (MSBlockInfo *block)
1212 {
1213         int count;
1214         void *reversed = NULL;
1215
1216  retry:
1217         switch (block->state) {
1218         case BLOCK_STATE_SWEPT:
1219                 return;
1220         case BLOCK_STATE_MARKING:
1221         case BLOCK_STATE_CHECKING:
1222                 SGEN_ASSERT (0, FALSE, "How did we get to sweep a block that's being marked or being checked?");
1223                 goto retry;
1224         case BLOCK_STATE_SWEEPING:
1225                 /* FIXME: Pick another block or whatever */
1226                 g_usleep (100);
1227                 goto retry;
1228         case BLOCK_STATE_NEED_SWEEPING:
1229                 if (!try_set_block_state (block, BLOCK_STATE_SWEEPING, BLOCK_STATE_NEED_SWEEPING))
1230                         goto retry;
1231                 break;
1232         default:
1233                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1234         }
1235
1236         SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEEPING, "How did we get here without setting state to sweeping?");
1237
1238         count = MS_BLOCK_FREE / block->obj_size;
1239
1240         block->free_list = NULL;
1241
1242         /* Use inline instances specialized to constant sizes, this allows the compiler to replace the memset calls with inline code */
1243         // FIXME: Add more sizes
1244         switch (block->obj_size) {
1245         case 16:
1246                 sweep_block_for_size (block, count, 16);
1247                 break;
1248         default:
1249                 sweep_block_for_size (block, count, block->obj_size);
1250                 break;
1251         }
1252
1253         /* reset mark bits */
1254         memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS);
1255
1256         /* Reverse free list so that it's in address order */
1257         reversed = NULL;
1258         while (block->free_list) {
1259                 void *next = *(void**)block->free_list;
1260                 *(void**)block->free_list = reversed;
1261                 reversed = block->free_list;
1262                 block->free_list = next;
1263         }
1264         block->free_list = reversed;
1265
1266         mono_memory_write_barrier ();
1267
1268         set_block_state (block, BLOCK_STATE_SWEPT, BLOCK_STATE_SWEEPING);
1269 }
1270
1271 static inline int
1272 bitcount (mword d)
1273 {
1274         int count = 0;
1275
1276 #ifdef __GNUC__
1277         if (sizeof (mword) == sizeof (unsigned long))
1278                 count += __builtin_popcountl (d);
1279         else
1280                 count += __builtin_popcount (d);
1281 #else
1282         while (d) {
1283                 count ++;
1284                 d &= (d - 1);
1285         }
1286 #endif
1287         return count;
1288 }
1289
1290 /* statistics for evacuation */
1291 static size_t *sweep_slots_available;
1292 static size_t *sweep_slots_used;
1293 static size_t *sweep_num_blocks;
1294
1295 static volatile size_t num_major_sections_before_sweep;
1296 static volatile size_t num_major_sections_freed_in_sweep;
1297
1298 static void
1299 sweep_start (void)
1300 {
1301         int i;
1302
1303         for (i = 0; i < num_block_obj_sizes; ++i)
1304                 sweep_slots_available [i] = sweep_slots_used [i] = sweep_num_blocks [i] = 0;
1305
1306         /* clear all the free lists */
1307         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i) {
1308                 MSBlockInfo * volatile *free_blocks = free_block_lists [i];
1309                 int j;
1310                 for (j = 0; j < num_block_obj_sizes; ++j)
1311                         free_blocks [j] = NULL;
1312         }
1313 }
1314
1315 static void sweep_finish (void);
1316
1317 /*
1318  * LOCKING: The allocated blocks lock must be held when entering this function.  `block`
1319  * must have been loaded from the array with the lock held.  This function will unlock the
1320  * lock.
1321  *
1322  * Returns whether the block is still there.
1323  */
1324 static gboolean
1325 ensure_block_is_checked_for_sweeping (MSBlockInfo *block, int block_index, gboolean *have_checked)
1326 {
1327         int count;
1328         gboolean have_live = FALSE;
1329         gboolean have_free = FALSE;
1330         int nused = 0;
1331         int block_state = block->state;
1332         int i;
1333
1334         if (have_checked)
1335                 *have_checked = FALSE;
1336
1337         if (!sweep_in_progress ()) {
1338                 SGEN_ASSERT (0, block_state != BLOCK_STATE_SWEEPING && block_state != BLOCK_STATE_CHECKING, "Invalid block state.");
1339                 if (!lazy_sweep)
1340                         SGEN_ASSERT (0, block_state != BLOCK_STATE_NEED_SWEEPING, "Invalid block state.");
1341         }
1342
1343  retry:
1344         switch (block_state) {
1345         case BLOCK_STATE_SWEPT:
1346         case BLOCK_STATE_NEED_SWEEPING:
1347         case BLOCK_STATE_SWEEPING:
1348                 UNLOCK_ALLOCATED_BLOCKS;
1349                 return TRUE;
1350         case BLOCK_STATE_MARKING:
1351                 if (sweep_in_progress ())
1352                         break;
1353                 UNLOCK_ALLOCATED_BLOCKS;
1354                 return TRUE;
1355         case BLOCK_STATE_CHECKING: {
1356                 MSBlockInfo *block_before = block;
1357                 /*
1358                  * FIXME: do this more elegantly.
1359                  *
1360                  * Also, when we're called from the sweep thread, we don't actually have to
1361                  * wait for it to finish, because the sweep thread doesn't use the block.
1362                  * However, the sweep thread needs to know when all the blocks have been
1363                  * checked (so it can set the global sweep state to SWEPT), so we'd have to
1364                  * do some kind of accounting if we don't wait.
1365                  */
1366                 UNLOCK_ALLOCATED_BLOCKS;
1367                 g_usleep (100);
1368                 LOCK_ALLOCATED_BLOCKS;
1369                 block = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [block_index]);
1370                 if (!block) {
1371                         UNLOCK_ALLOCATED_BLOCKS;
1372                         return FALSE;
1373                 }
1374                 SGEN_ASSERT (0, block == block_before, "How did the block get exchanged for a different one?");
1375                 block_state = block->state;
1376                 goto retry;
1377         }
1378         default:
1379                 SGEN_ASSERT (0, FALSE, "Illegal block state");
1380                 break;
1381         }
1382
1383         SGEN_ASSERT (0, block->state == BLOCK_STATE_MARKING, "When we sweep all blocks must start out marking.");
1384         set_block_state (block, BLOCK_STATE_CHECKING, BLOCK_STATE_MARKING);
1385         UNLOCK_ALLOCATED_BLOCKS;
1386
1387         if (have_checked)
1388                 *have_checked = TRUE;
1389
1390         block->has_pinned = block->pinned;
1391
1392         block->is_to_space = FALSE;
1393
1394         count = MS_BLOCK_FREE / block->obj_size;
1395
1396         if (block->cardtable_mod_union) {
1397                 sgen_free_internal_dynamic (block->cardtable_mod_union, CARDS_PER_BLOCK, INTERNAL_MEM_CARDTABLE_MOD_UNION);
1398                 block->cardtable_mod_union = NULL;
1399         }
1400
1401         /* Count marked objects in the block */
1402         for (i = 0; i < MS_NUM_MARK_WORDS; ++i)
1403                 nused += bitcount (block->mark_words [i]);
1404
1405         if (nused)
1406                 have_live = TRUE;
1407         if (nused < count)
1408                 have_free = TRUE;
1409
1410         if (have_live) {
1411                 int obj_size_index = block->obj_size_index;
1412                 gboolean has_pinned = block->has_pinned;
1413
1414                 set_block_state (block, BLOCK_STATE_NEED_SWEEPING, BLOCK_STATE_CHECKING);
1415
1416                 /*
1417                  * FIXME: Go straight to SWEPT if there are no free slots.  We need
1418                  * to set the free slot list to NULL, though, and maybe update some
1419                  * statistics.
1420                  */
1421                 if (!lazy_sweep)
1422                         sweep_block (block);
1423
1424                 if (!has_pinned) {
1425                         ++sweep_num_blocks [obj_size_index];
1426                         sweep_slots_used [obj_size_index] += nused;
1427                         sweep_slots_available [obj_size_index] += count;
1428                 }
1429
1430                 /*
1431                  * If there are free slots in the block, add
1432                  * the block to the corresponding free list.
1433                  */
1434                 if (have_free) {
1435                         MSBlockInfo * volatile *free_blocks = FREE_BLOCKS (block->pinned, block->has_references);
1436                         int index = MS_BLOCK_OBJ_SIZE_INDEX (block->obj_size);
1437
1438                         if (!lazy_sweep)
1439                                 SGEN_ASSERT (0, block->free_list, "How do we not have a free list when there are free slots?");
1440
1441                         add_free_block (free_blocks, index, block);
1442                 }
1443
1444                 /* FIXME: Do we need the heap boundaries while we do nursery collections? */
1445                 update_heap_boundaries_for_block (block);
1446
1447                 return TRUE;
1448         } else {
1449                 /*
1450                  * Blocks without live objects are removed from the
1451                  * block list and freed.
1452                  */
1453                 LOCK_ALLOCATED_BLOCKS;
1454                 SGEN_ASSERT (0, block_index < allocated_blocks.next_slot, "How did the number of blocks shrink?");
1455                 SGEN_ASSERT (0, BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [block_index]) == block, "How did the block move?");
1456                 allocated_blocks.data [block_index] = NULL;
1457                 UNLOCK_ALLOCATED_BLOCKS;
1458
1459                 binary_protocol_empty (MS_BLOCK_OBJ (block, 0), (char*)MS_BLOCK_OBJ (block, count) - (char*)MS_BLOCK_OBJ (block, 0));
1460                 ms_free_block (block);
1461
1462                 SGEN_ATOMIC_ADD_P (num_major_sections, -1);
1463
1464                 return FALSE;
1465         }
1466 }
1467
1468 static gboolean
1469 try_set_sweep_state (int new, int expected)
1470 {
1471         int old = SGEN_CAS (&sweep_state, new, expected);
1472         return old == expected;
1473 }
1474
1475 static void
1476 set_sweep_state (int new, int expected)
1477 {
1478         gboolean success = try_set_sweep_state (new, expected);
1479         SGEN_ASSERT (0, success, "Could not set sweep state.");
1480 }
1481
1482 static mono_native_thread_return_t
1483 sweep_loop_thread_func (void *dummy)
1484 {
1485         int block_index;
1486         int num_blocks = num_major_sections_before_sweep;
1487         int small_id = mono_thread_info_register_small_id ();
1488
1489         SGEN_ASSERT (0, sweep_in_progress (), "Sweep thread called with wrong state");
1490         SGEN_ASSERT (0, num_blocks <= allocated_blocks.next_slot, "How did we lose blocks?");
1491
1492         /*
1493          * We traverse the block array from high to low.  Nursery collections will have to
1494          * cooperate with the sweep thread to finish sweeping, and they will traverse from
1495          * low to high, to avoid constantly colliding on the same blocks.
1496          */
1497         for (block_index = num_blocks - 1; block_index >= 0; --block_index) {
1498                 MSBlockInfo *block;
1499                 gboolean have_checked;
1500
1501                 LOCK_ALLOCATED_BLOCKS;
1502                 block = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [block_index]);
1503
1504                 /*
1505                  * The block might have been freed by another thread doing some checking
1506                  * work.
1507                  */
1508                 if (!block) {
1509                         UNLOCK_ALLOCATED_BLOCKS;
1510                         ++num_major_sections_freed_in_sweep;
1511                         continue;
1512                 }
1513
1514                 if (block->state == BLOCK_STATE_SWEPT) {
1515                         UNLOCK_ALLOCATED_BLOCKS;
1516                         continue;
1517                 }
1518
1519                 ensure_block_is_checked_for_sweeping (block, block_index, &have_checked);
1520         }
1521
1522         while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING))
1523                 g_usleep (100);
1524
1525         LOCK_ALLOCATED_BLOCKS;
1526         for (block_index = num_blocks; block_index < allocated_blocks.next_slot; ++block_index) {
1527                 MSBlockInfo *block = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [block_index]);
1528                 SGEN_ASSERT (0, block && block->state == BLOCK_STATE_SWEPT, "How did a new block to be swept get added while swept?");
1529         }
1530
1531         sgen_pointer_queue_remove_nulls (&allocated_blocks);
1532         UNLOCK_ALLOCATED_BLOCKS;
1533
1534         sweep_finish ();
1535
1536         mono_thread_small_id_free (small_id);
1537
1538         return NULL;
1539 }
1540
1541 static void
1542 sweep_finish (void)
1543 {
1544         mword total_evacuate_heap = 0;
1545         mword total_evacuate_saved = 0;
1546         int i;
1547
1548         for (i = 0; i < num_block_obj_sizes; ++i) {
1549                 float usage = (float)sweep_slots_used [i] / (float)sweep_slots_available [i];
1550                 if (sweep_num_blocks [i] > 5 && usage < evacuation_threshold) {
1551                         evacuate_block_obj_sizes [i] = TRUE;
1552                         /*
1553                         g_print ("slot size %d - %d of %d used\n",
1554                                         block_obj_sizes [i], slots_used [i], slots_available [i]);
1555                         */
1556                 } else {
1557                         evacuate_block_obj_sizes [i] = FALSE;
1558                 }
1559                 {
1560                         mword total_bytes = block_obj_sizes [i] * sweep_slots_available [i];
1561                         total_evacuate_heap += total_bytes;
1562                         if (evacuate_block_obj_sizes [i])
1563                                 total_evacuate_saved += total_bytes - block_obj_sizes [i] * sweep_slots_used [i];
1564                 }
1565         }
1566
1567         want_evacuation = (float)total_evacuate_saved / (float)total_evacuate_heap > (1 - concurrent_evacuation_threshold);
1568
1569         set_sweep_state (SWEEP_STATE_SWEPT, SWEEP_STATE_COMPACTING);
1570 }
1571
1572 static MonoNativeThreadId sweep_loop_thread;
1573
1574 static void
1575 major_sweep (void)
1576 {
1577         set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_NEED_SWEEPING);
1578
1579         sweep_start ();
1580
1581         SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?");
1582
1583         num_major_sections_before_sweep = num_major_sections;
1584         num_major_sections_freed_in_sweep = 0;
1585
1586         if (TRUE /*concurrent_mark*/) {
1587                 /*
1588                  * FIXME: We can't create a thread while the world is stopped because it
1589                  * might deadlock.  `finalizer-wait.exe` exposes this.
1590                  */
1591                 mono_native_thread_create (&sweep_loop_thread, sweep_loop_thread_func, NULL);
1592         } else {
1593                 sweep_loop_thread_func (NULL);
1594         }
1595 }
1596
1597 static gboolean
1598 major_have_swept (void)
1599 {
1600         return sweep_state == SWEEP_STATE_SWEPT;
1601 }
1602
1603 static int count_pinned_ref;
1604 static int count_pinned_nonref;
1605 static int count_nonpinned_ref;
1606 static int count_nonpinned_nonref;
1607
1608 static void
1609 count_nonpinned_callback (char *obj, size_t size, void *data)
1610 {
1611         MonoVTable *vtable = (MonoVTable*)LOAD_VTABLE (obj);
1612
1613         if (vtable->klass->has_references)
1614                 ++count_nonpinned_ref;
1615         else
1616                 ++count_nonpinned_nonref;
1617 }
1618
1619 static void
1620 count_pinned_callback (char *obj, size_t size, void *data)
1621 {
1622         MonoVTable *vtable = (MonoVTable*)LOAD_VTABLE (obj);
1623
1624         if (vtable->klass->has_references)
1625                 ++count_pinned_ref;
1626         else
1627                 ++count_pinned_nonref;
1628 }
1629
1630 static G_GNUC_UNUSED void
1631 count_ref_nonref_objs (void)
1632 {
1633         int total;
1634
1635         count_pinned_ref = 0;
1636         count_pinned_nonref = 0;
1637         count_nonpinned_ref = 0;
1638         count_nonpinned_nonref = 0;
1639
1640         major_iterate_objects (ITERATE_OBJECTS_SWEEP_NON_PINNED, count_nonpinned_callback, NULL);
1641         major_iterate_objects (ITERATE_OBJECTS_SWEEP_PINNED, count_pinned_callback, NULL);
1642
1643         total = count_pinned_nonref + count_nonpinned_nonref + count_pinned_ref + count_nonpinned_ref;
1644
1645         g_print ("ref: %d pinned %d non-pinned   non-ref: %d pinned %d non-pinned  --  %.1f\n",
1646                         count_pinned_ref, count_nonpinned_ref,
1647                         count_pinned_nonref, count_nonpinned_nonref,
1648                         (count_pinned_nonref + count_nonpinned_nonref) * 100.0 / total);
1649 }
1650
1651 static int
1652 ms_calculate_block_obj_sizes (double factor, int *arr)
1653 {
1654         double target_size;
1655         int num_sizes = 0;
1656         int last_size = 0;
1657
1658         /*
1659          * Have every possible slot size starting with the minimal
1660          * object size up to and including four times that size.  Then
1661          * proceed by increasing geometrically with the given factor.
1662          */
1663
1664         for (int size = sizeof (MonoObject); size <= 4 * sizeof (MonoObject); size += SGEN_ALLOC_ALIGN) {
1665                 if (arr)
1666                         arr [num_sizes] = size;
1667                 ++num_sizes;
1668                 last_size = size;
1669         }
1670         target_size = (double)last_size;
1671
1672         do {
1673                 int target_count = (int)floor (MS_BLOCK_FREE / target_size);
1674                 int size = MIN ((MS_BLOCK_FREE / target_count) & ~(SGEN_ALLOC_ALIGN - 1), SGEN_MAX_SMALL_OBJ_SIZE);
1675
1676                 if (size != last_size) {
1677                         if (arr)
1678                                 arr [num_sizes] = size;
1679                         ++num_sizes;
1680                         last_size = size;
1681                 }
1682
1683                 target_size *= factor;
1684         } while (last_size < SGEN_MAX_SMALL_OBJ_SIZE);
1685
1686         return num_sizes;
1687 }
1688
1689 /* only valid during minor collections */
1690 static mword old_num_major_sections;
1691
1692 static void
1693 major_start_nursery_collection (void)
1694 {
1695 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1696         consistency_check ();
1697 #endif
1698
1699         old_num_major_sections = num_major_sections;
1700
1701         if (sweep_in_progress ())
1702                 g_print ("sweeping during nursery collection\n");
1703 }
1704
1705 static void
1706 major_finish_nursery_collection (void)
1707 {
1708 #ifdef MARKSWEEP_CONSISTENCY_CHECK
1709         consistency_check ();
1710 #endif
1711 }
1712
1713 static void
1714 major_start_major_collection (void)
1715 {
1716         MSBlockInfo *block;
1717         int i;
1718
1719         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Major collection on unswept heap");
1720
1721         /*
1722          * Clear the free lists for block sizes where we do evacuation.  For those block
1723          * sizes we will have to allocate new blocks.
1724          */
1725         for (i = 0; i < num_block_obj_sizes; ++i) {
1726                 if (!evacuate_block_obj_sizes [i])
1727                         continue;
1728
1729                 free_block_lists [0][i] = NULL;
1730                 free_block_lists [MS_BLOCK_FLAG_REFS][i] = NULL;
1731         }
1732
1733         // Sweep all unswept blocks
1734         if (lazy_sweep) {
1735                 MONO_GC_SWEEP_BEGIN (GENERATION_OLD, TRUE);
1736
1737                 FOREACH_BLOCK (block) {
1738                         sweep_block (block);
1739                 } END_FOREACH_BLOCK;
1740
1741                 MONO_GC_SWEEP_END (GENERATION_OLD, TRUE);
1742         }
1743
1744         /* FIXME: Just do one iteration over the blocks in this function. */
1745         FOREACH_BLOCK (block) {
1746                 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT, "All blocks must be swept when we're pinning.");
1747                 /*
1748                  * FIXME: We don't need CAS here because there's still only one thread doing
1749                  * stuff.
1750                  */
1751                 set_block_state (block, BLOCK_STATE_MARKING, BLOCK_STATE_SWEPT);
1752         } END_FOREACH_BLOCK;
1753
1754         set_sweep_state (SWEEP_STATE_NEED_SWEEPING, SWEEP_STATE_SWEPT);
1755 }
1756
1757 static void
1758 major_finish_major_collection (ScannedObjectCounts *counts)
1759 {
1760 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
1761         if (binary_protocol_is_enabled ()) {
1762                 counts->num_scanned_objects = scanned_objects_list.next_slot;
1763
1764                 sgen_pointer_queue_sort_uniq (&scanned_objects_list);
1765                 counts->num_unique_scanned_objects = scanned_objects_list.next_slot;
1766
1767                 sgen_pointer_queue_clear (&scanned_objects_list);
1768         }
1769 #endif
1770 }
1771
1772 #if SIZEOF_VOID_P != 8
1773 static int
1774 compare_pointers (const void *va, const void *vb) {
1775         char *a = *(char**)va, *b = *(char**)vb;
1776         if (a < b)
1777                 return -1;
1778         if (a > b)
1779                 return 1;
1780         return 0;
1781 }
1782 #endif
1783
1784 /*
1785  * This is called with sweep completed and the world stopped.
1786  */
1787 static void
1788 major_free_swept_blocks (void)
1789 {
1790         size_t section_reserve = sgen_get_minor_collection_allowance () / MS_BLOCK_SIZE;
1791
1792         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
1793
1794 #if SIZEOF_VOID_P != 8
1795         {
1796                 int i, num_empty_blocks_orig, num_blocks, arr_length;
1797                 void *block;
1798                 void **empty_block_arr;
1799                 void **rebuild_next;
1800
1801 #ifdef TARGET_WIN32
1802                 /*
1803                  * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
1804                  * a VirtualAlloc ()-ed block.
1805                  */
1806                 return;
1807 #endif
1808
1809                 if (num_empty_blocks <= section_reserve)
1810                         return;
1811                 SGEN_ASSERT (0, num_empty_blocks > 0, "section reserve can't be negative");
1812
1813                 num_empty_blocks_orig = num_empty_blocks;
1814                 empty_block_arr = (void**)sgen_alloc_internal_dynamic (sizeof (void*) * num_empty_blocks_orig,
1815                                 INTERNAL_MEM_MS_BLOCK_INFO_SORT, FALSE);
1816                 if (!empty_block_arr)
1817                         goto fallback;
1818
1819                 i = 0;
1820                 for (block = empty_blocks; block; block = *(void**)block)
1821                         empty_block_arr [i++] = block;
1822                 SGEN_ASSERT (0, i == num_empty_blocks, "empty block count wrong");
1823
1824                 sgen_qsort (empty_block_arr, num_empty_blocks, sizeof (void*), compare_pointers);
1825
1826                 /*
1827                  * We iterate over the free blocks, trying to find MS_BLOCK_ALLOC_NUM
1828                  * contiguous ones.  If we do, we free them.  If that's not enough to get to
1829                  * section_reserve, we halve the number of contiguous blocks we're looking
1830                  * for and have another go, until we're done with looking for pairs of
1831                  * blocks, at which point we give up and go to the fallback.
1832                  */
1833                 arr_length = num_empty_blocks_orig;
1834                 num_blocks = MS_BLOCK_ALLOC_NUM;
1835                 while (num_empty_blocks > section_reserve && num_blocks > 1) {
1836                         int first = -1;
1837                         int dest = 0;
1838
1839                         dest = 0;
1840                         for (i = 0; i < arr_length; ++i) {
1841                                 int d = dest;
1842                                 void *block = empty_block_arr [i];
1843                                 SGEN_ASSERT (0, block, "we're not shifting correctly");
1844                                 if (i != dest) {
1845                                         empty_block_arr [dest] = block;
1846                                         /*
1847                                          * This is not strictly necessary, but we're
1848                                          * cautious.
1849                                          */
1850                                         empty_block_arr [i] = NULL;
1851                                 }
1852                                 ++dest;
1853
1854                                 if (first < 0) {
1855                                         first = d;
1856                                         continue;
1857                                 }
1858
1859                                 SGEN_ASSERT (0, first >= 0 && d > first, "algorithm is wrong");
1860
1861                                 if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) {
1862                                         first = d;
1863                                         continue;
1864                                 }
1865
1866                                 if (d + 1 - first == num_blocks) {
1867                                         /*
1868                                          * We found num_blocks contiguous blocks.  Free them
1869                                          * and null their array entries.  As an optimization
1870                                          * we could, instead of nulling the entries, shift
1871                                          * the following entries over to the left, while
1872                                          * we're iterating.
1873                                          */
1874                                         int j;
1875                                         sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP);
1876                                         for (j = first; j <= d; ++j)
1877                                                 empty_block_arr [j] = NULL;
1878                                         dest = first;
1879                                         first = -1;
1880
1881                                         num_empty_blocks -= num_blocks;
1882
1883                                         stat_major_blocks_freed += num_blocks;
1884                                         if (num_blocks == MS_BLOCK_ALLOC_NUM)
1885                                                 stat_major_blocks_freed_ideal += num_blocks;
1886                                         else
1887                                                 stat_major_blocks_freed_less_ideal += num_blocks;
1888
1889                                 }
1890                         }
1891
1892                         SGEN_ASSERT (0, dest <= i && dest <= arr_length, "array length is off");
1893                         arr_length = dest;
1894                         SGEN_ASSERT (0, arr_length == num_empty_blocks, "array length is off");
1895
1896                         num_blocks >>= 1;
1897                 }
1898
1899                 /* rebuild empty_blocks free list */
1900                 rebuild_next = (void**)&empty_blocks;
1901                 for (i = 0; i < arr_length; ++i) {
1902                         void *block = empty_block_arr [i];
1903                         SGEN_ASSERT (0, block, "we're missing blocks");
1904                         *rebuild_next = block;
1905                         rebuild_next = (void**)block;
1906                 }
1907                 *rebuild_next = NULL;
1908
1909                 /* free array */
1910                 sgen_free_internal_dynamic (empty_block_arr, sizeof (void*) * num_empty_blocks_orig, INTERNAL_MEM_MS_BLOCK_INFO_SORT);
1911         }
1912
1913         SGEN_ASSERT (0, num_empty_blocks >= 0, "we freed more blocks than we had in the first place?");
1914
1915  fallback:
1916         /*
1917          * This is our threshold.  If there's not more empty than used blocks, we won't
1918          * release uncontiguous blocks, in fear of fragmenting the address space.
1919          */
1920         if (num_empty_blocks <= num_major_sections)
1921                 return;
1922 #endif
1923
1924         while (num_empty_blocks > section_reserve) {
1925                 void *next = *(void**)empty_blocks;
1926                 sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP);
1927                 empty_blocks = next;
1928                 /*
1929                  * Needs not be atomic because this is running
1930                  * single-threaded.
1931                  */
1932                 --num_empty_blocks;
1933
1934                 ++stat_major_blocks_freed;
1935 #if SIZEOF_VOID_P != 8
1936                 ++stat_major_blocks_freed_individual;
1937 #endif
1938         }
1939 }
1940
1941 /* FIXME: Unify `major_find_pin_queue_start_ends` and `major_pin_objects`. */
1942 static void
1943 major_pin_objects (SgenGrayQueue *queue)
1944 {
1945         MSBlockInfo *block;
1946
1947         SGEN_ASSERT (0, !sweep_in_progress (), "Cannot iterate blocks during sweep");
1948         FOREACH_BLOCK (block) {
1949                 size_t first_entry, last_entry;
1950                 SGEN_ASSERT (0, block->state == BLOCK_STATE_SWEPT || block->state == BLOCK_STATE_MARKING, "All blocks must be swept when we're pinning.");
1951                 sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE,
1952                                 &first_entry, &last_entry);
1953                 mark_pinned_objects_in_block (block, first_entry, last_entry, queue);
1954         } END_FOREACH_BLOCK;
1955 }
1956
1957 static void
1958 major_init_to_space (void)
1959 {
1960 }
1961
1962 static void
1963 major_report_pinned_memory_usage (void)
1964 {
1965         g_assert_not_reached ();
1966 }
1967
1968 static gint64
1969 major_get_used_size (void)
1970 {
1971         gint64 size = 0;
1972         MSBlockInfo *block;
1973
1974         SGEN_ASSERT (0, !sweep_in_progress (), "Cannot iterate blocks during sweep");
1975         FOREACH_BLOCK (block) {
1976                 int count = MS_BLOCK_FREE / block->obj_size;
1977                 void **iter;
1978                 size += count * block->obj_size;
1979                 for (iter = block->free_list; iter; iter = (void**)*iter)
1980                         size -= block->obj_size;
1981         } END_FOREACH_BLOCK;
1982
1983         return size;
1984 }
1985
1986 static size_t
1987 get_num_major_sections (void)
1988 {
1989         return num_major_sections;
1990 }
1991
1992 /*
1993  * Returns the number of major sections that were present when the last sweep was initiated,
1994  * and were not freed during the sweep.  They are the basis for calculating the allowance.
1995  */
1996 static size_t
1997 get_num_major_unswept_old_sections (void)
1998 {
1999         SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep");
2000         return num_major_sections_before_sweep - num_major_sections_freed_in_sweep;
2001 }
2002
2003 static gboolean
2004 major_handle_gc_param (const char *opt)
2005 {
2006         if (g_str_has_prefix (opt, "evacuation-threshold=")) {
2007                 const char *arg = strchr (opt, '=') + 1;
2008                 int percentage = atoi (arg);
2009                 if (percentage < 0 || percentage > 100) {
2010                         fprintf (stderr, "evacuation-threshold must be an integer in the range 0-100.\n");
2011                         exit (1);
2012                 }
2013                 evacuation_threshold = (float)percentage / 100.0f;
2014                 return TRUE;
2015         } else if (!strcmp (opt, "lazy-sweep")) {
2016                 lazy_sweep = TRUE;
2017                 return TRUE;
2018         } else if (!strcmp (opt, "no-lazy-sweep")) {
2019                 lazy_sweep = FALSE;
2020                 return TRUE;
2021         }
2022
2023         return FALSE;
2024 }
2025
2026 static void
2027 major_print_gc_param_usage (void)
2028 {
2029         fprintf (stderr,
2030                         ""
2031                         "  evacuation-threshold=P (where P is a percentage, an integer in 0-100)\n"
2032                         "  (no-)lazy-sweep\n"
2033                         );
2034 }
2035
2036 /*
2037  * This callback is used to clear cards, move cards to the shadow table and do counting.
2038  */
2039 static void
2040 major_iterate_live_block_ranges (sgen_cardtable_block_callback callback, gboolean requires_sweep)
2041 {
2042         MSBlockInfo *block;
2043         gboolean has_references;
2044
2045         if (requires_sweep)
2046                 major_finish_sweeping ();
2047
2048         /*
2049          * FIXME: Don't take the lock for the whole allocated blocks array because we're
2050          * stopping the sweep thread.
2051          */
2052         FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
2053                 if (has_references)
2054                         callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE);
2055         } END_FOREACH_BLOCK;
2056 }
2057
2058 #ifdef HEAVY_STATISTICS
2059 extern guint64 marked_cards;
2060 extern guint64 scanned_cards;
2061 extern guint64 scanned_objects;
2062 extern guint64 remarked_cards;
2063 #endif
2064
2065 #define CARD_WORDS_PER_BLOCK (CARDS_PER_BLOCK / SIZEOF_VOID_P)
2066 /*
2067  * MS blocks are 16K aligned.
2068  * Cardtables are 4K aligned, at least.
2069  * This means that the cardtable of a given block is 32 bytes aligned.
2070  */
2071 static guint8*
2072 initial_skip_card (guint8 *card_data)
2073 {
2074         mword *cards = (mword*)card_data;
2075         mword card;
2076         int i;
2077         for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) {
2078                 card = cards [i];
2079                 if (card)
2080                         break;
2081         }
2082
2083         if (i == CARD_WORDS_PER_BLOCK)
2084                 return card_data + CARDS_PER_BLOCK;
2085
2086 #if defined(__i386__) && defined(__GNUC__)
2087         return card_data + i * 4 +  (__builtin_ffs (card) - 1) / 8;
2088 #elif defined(__x86_64__) && defined(__GNUC__)
2089         return card_data + i * 8 +  (__builtin_ffsll (card) - 1) / 8;
2090 #elif defined(__s390x__) && defined(__GNUC__)
2091         return card_data + i * 8 +  (__builtin_ffsll (GUINT64_TO_LE(card)) - 1) / 8;
2092 #else
2093         for (i = i * SIZEOF_VOID_P; i < CARDS_PER_BLOCK; ++i) {
2094                 if (card_data [i])
2095                         return &card_data [i];
2096         }
2097         return card_data;
2098 #endif
2099 }
2100
2101 #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os))
2102 #define MS_BLOCK_OBJ_FAST(b,os,i)                       ((b) + MS_BLOCK_SKIP + (os) * (i))
2103 #define MS_OBJ_ALLOCED_FAST(o,b)                (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE))
2104
2105 static size_t
2106 card_offset (char *obj, char *base)
2107 {
2108         return (obj - base) >> CARD_BITS;
2109 }
2110
2111 static void
2112 scan_card_table_for_block (MSBlockInfo *block, gboolean mod_union, ScanObjectFunc scan_func, SgenGrayQueue *queue)
2113 {
2114 #ifndef SGEN_HAVE_OVERLAPPING_CARDS
2115         guint8 cards_copy [CARDS_PER_BLOCK];
2116 #endif
2117         gboolean small_objects;
2118         int block_obj_size;
2119         char *block_start;
2120         guint8 *card_data, *card_base;
2121         guint8 *card_data_end;
2122         char *scan_front = NULL;
2123
2124         block_obj_size = block->obj_size;
2125         small_objects = block_obj_size < CARD_SIZE_IN_BYTES;
2126
2127         block_start = MS_BLOCK_FOR_BLOCK_INFO (block);
2128
2129         /*
2130          * This is safe in face of card aliasing for the following reason:
2131          *
2132          * Major blocks are 16k aligned, or 32 cards aligned.
2133          * Cards aliasing happens in powers of two, so as long as major blocks are aligned to their
2134          * sizes, they won't overflow the cardtable overlap modulus.
2135          */
2136         if (mod_union) {
2137                 card_data = card_base = block->cardtable_mod_union;
2138                 /*
2139                  * This happens when the nursery collection that precedes finishing
2140                  * the concurrent collection allocates new major blocks.
2141                  */
2142                 if (!card_data)
2143                         return;
2144         } else {
2145 #ifdef SGEN_HAVE_OVERLAPPING_CARDS
2146                 card_data = card_base = sgen_card_table_get_card_scan_address ((mword)block_start);
2147 #else
2148                 if (!sgen_card_table_get_card_data (cards_copy, (mword)block_start, CARDS_PER_BLOCK))
2149                         return;
2150                 card_data = card_base = cards_copy;
2151 #endif
2152         }
2153         card_data_end = card_data + CARDS_PER_BLOCK;
2154
2155         card_data += MS_BLOCK_SKIP >> CARD_BITS;
2156
2157         card_data = initial_skip_card (card_data);
2158         while (card_data < card_data_end) {
2159                 size_t card_index, first_object_index;
2160                 char *start;
2161                 char *end;
2162                 char *first_obj, *obj;
2163
2164                 HEAVY_STAT (++scanned_cards);
2165
2166                 if (!*card_data) {
2167                         ++card_data;
2168                         continue;
2169                 }
2170
2171                 card_index = card_data - card_base;
2172                 start = (char*)(block_start + card_index * CARD_SIZE_IN_BYTES);
2173                 end = start + CARD_SIZE_IN_BYTES;
2174
2175                 if (block->state != BLOCK_STATE_SWEPT && block->state != BLOCK_STATE_MARKING)
2176                         sweep_block (block);
2177
2178                 HEAVY_STAT (++marked_cards);
2179
2180                 if (small_objects)
2181                         sgen_card_table_prepare_card_for_scanning (card_data);
2182
2183                 /*
2184                  * If the card we're looking at starts at or in the block header, we
2185                  * must start at the first object in the block, without calculating
2186                  * the index of the object we're hypothetically starting at, because
2187                  * it would be negative.
2188                  */
2189                 if (card_index <= (MS_BLOCK_SKIP >> CARD_BITS))
2190                         first_object_index = 0;
2191                 else
2192                         first_object_index = MS_BLOCK_OBJ_INDEX_FAST (start, block_start, block_obj_size);
2193
2194                 obj = first_obj = (char*)MS_BLOCK_OBJ_FAST (block_start, block_obj_size, first_object_index);
2195
2196                 binary_protocol_card_scan (first_obj, end - first_obj);
2197
2198                 while (obj < end) {
2199                         if (obj < scan_front || !MS_OBJ_ALLOCED_FAST (obj, block_start))
2200                                 goto next_object;
2201
2202                         if (mod_union) {
2203                                 /* FIXME: do this more efficiently */
2204                                 int w, b;
2205                                 MS_CALC_MARK_BIT (w, b, obj);
2206                                 if (!MS_MARK_BIT (block, w, b))
2207                                         goto next_object;
2208                         }
2209
2210                         if (small_objects) {
2211                                 HEAVY_STAT (++scanned_objects);
2212                                 scan_func (obj, sgen_obj_get_descriptor (obj), queue);
2213                         } else {
2214                                 size_t offset = card_offset (obj, block_start);
2215                                 sgen_cardtable_scan_object (obj, block_obj_size, card_base + offset, mod_union, queue);
2216                         }
2217                 next_object:
2218                         obj += block_obj_size;
2219                         g_assert (scan_front <= obj);
2220                         scan_front = obj;
2221                 }
2222
2223                 HEAVY_STAT (if (*card_data) ++remarked_cards);
2224
2225                 if (small_objects)
2226                         ++card_data;
2227                 else
2228                         card_data = card_base + card_offset (obj, block_start);
2229         }
2230 }
2231
2232 static void
2233 major_scan_card_table (gboolean mod_union, SgenGrayQueue *queue)
2234 {
2235         ScanObjectFunc scan_func = sgen_get_current_object_ops ()->scan_object;
2236         int block_index;
2237         gboolean do_sweep_checking = sweep_state == SWEEP_STATE_SWEEPING;
2238
2239         if (!concurrent_mark)
2240                 g_assert (!mod_union);
2241
2242  retry:
2243         switch (sweep_state) {
2244         case SWEEP_STATE_SWEPT:
2245         case SWEEP_STATE_NEED_SWEEPING:
2246                 break;
2247         case SWEEP_STATE_SWEEPING:
2248                 if (try_set_sweep_state (SWEEP_STATE_SWEEPING_AND_ITERATING, SWEEP_STATE_SWEEPING))
2249                         break;
2250                 goto retry;
2251         case SWEEP_STATE_SWEEPING_AND_ITERATING:
2252                 SGEN_ASSERT (0, FALSE, "Is there another minor collection running?");
2253                 goto retry;
2254         case SWEEP_STATE_COMPACTING:
2255                 g_usleep (100);
2256                 goto retry;
2257         default:
2258                 SGEN_ASSERT (0, FALSE, "Invalid sweep state.");
2259                 break;
2260         }
2261
2262         //major_finish_sweeping ();
2263
2264         /*
2265          * We're running with the world stopped and the only other thread doing work is the
2266          * sweep thread, which doesn't add blocks to the array, so we can safely access
2267          * `next_slot` without locking.
2268          */
2269         for (block_index = 0; block_index < allocated_blocks.next_slot; ++block_index) {
2270                 //gboolean has_references;
2271                 void *tagged_block;
2272                 MSBlockInfo *block;
2273
2274 #ifdef PREFETCH_CARDS
2275                 int prefetch_index = block_index + 6;
2276                 if (prefetch_index < allocated_blocks.next_slot) {
2277                         MSBlockInfo *prefetch_block = BLOCK_UNTAG_HAS_REFERENCES (allocated_blocks.data [prefetch_index]);
2278                         guint8 *prefetch_cards = sgen_card_table_get_card_scan_address ((mword)MS_BLOCK_FOR_BLOCK_INFO (prefetch_block));
2279                         PREFETCH_READ (prefetch_block);
2280                         PREFETCH_WRITE (prefetch_cards);
2281                         PREFETCH_WRITE (prefetch_cards + 32);
2282                 }
2283 #endif
2284
2285                 if (do_sweep_checking) {
2286                         LOCK_ALLOCATED_BLOCKS;
2287                         tagged_block = allocated_blocks.data [block_index];
2288                         block = BLOCK_UNTAG_HAS_REFERENCES (tagged_block);
2289
2290                         if (!block || !BLOCK_IS_TAGGED_HAS_REFERENCES (tagged_block)) {
2291                                 UNLOCK_ALLOCATED_BLOCKS;
2292                                 continue;
2293                         }
2294
2295                         if (!ensure_block_is_checked_for_sweeping (block, block_index, NULL))
2296                                 continue;
2297                 } else {
2298                         tagged_block = allocated_blocks.data [block_index];
2299                         block = BLOCK_UNTAG_HAS_REFERENCES (tagged_block);
2300                         SGEN_ASSERT (0, block, "Why are there holes in the block array when we're not sweeping?");
2301
2302                         if (!BLOCK_IS_TAGGED_HAS_REFERENCES (tagged_block))
2303                                 continue;
2304                 }
2305
2306                 scan_card_table_for_block (block, mod_union, scan_func, queue);
2307         }
2308
2309         if (sweep_state == SWEEP_STATE_SWEEPING_AND_ITERATING)
2310                 set_sweep_state (SWEEP_STATE_SWEEPING, SWEEP_STATE_SWEEPING_AND_ITERATING);
2311 }
2312
2313 static void
2314 major_count_cards (long long *num_total_cards, long long *num_marked_cards)
2315 {
2316         MSBlockInfo *block;
2317         gboolean has_references;
2318         long long total_cards = 0;
2319         long long marked_cards = 0;
2320
2321         if (sweep_in_progress ()) {
2322                 *num_total_cards = -1;
2323                 *num_marked_cards = -1;
2324         }
2325
2326         FOREACH_BLOCK_HAS_REFERENCES (block, has_references) {
2327                 guint8 *cards = sgen_card_table_get_card_scan_address ((mword) MS_BLOCK_FOR_BLOCK_INFO (block));
2328                 int i;
2329
2330                 if (!has_references)
2331                         continue;
2332
2333                 total_cards += CARDS_PER_BLOCK;
2334                 for (i = 0; i < CARDS_PER_BLOCK; ++i) {
2335                         if (cards [i])
2336                                 ++marked_cards;
2337                 }
2338         } END_FOREACH_BLOCK;
2339
2340         *num_total_cards = total_cards;
2341         *num_marked_cards = marked_cards;
2342 }
2343
2344 static void
2345 update_cardtable_mod_union (void)
2346 {
2347         MSBlockInfo *block;
2348
2349         SGEN_ASSERT (0, !sweep_in_progress (), "Cannot iterate blocks during sweep");
2350         FOREACH_BLOCK (block) {
2351                 size_t num_cards;
2352
2353                 block->cardtable_mod_union = sgen_card_table_update_mod_union (block->cardtable_mod_union,
2354                                 MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards);
2355
2356                 SGEN_ASSERT (0, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong");
2357         } END_FOREACH_BLOCK;
2358 }
2359
2360 static guint8*
2361 major_get_cardtable_mod_union_for_object (char *obj)
2362 {
2363         MSBlockInfo *block = MS_BLOCK_FOR_OBJ (obj);
2364         size_t offset = card_offset (obj, (char*)sgen_card_table_align_pointer (MS_BLOCK_FOR_BLOCK_INFO (block)));
2365         return &block->cardtable_mod_union [offset];
2366 }
2367
2368 #undef pthread_create
2369
2370 static void
2371 post_param_init (SgenMajorCollector *collector)
2372 {
2373         collector->sweeps_lazily = lazy_sweep;
2374 }
2375
2376 static void
2377 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent)
2378 {
2379         int i;
2380
2381         sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo));
2382
2383         num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL);
2384         block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2385         ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, block_obj_sizes);
2386
2387         evacuate_block_obj_sizes = sgen_alloc_internal_dynamic (sizeof (gboolean) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2388         for (i = 0; i < num_block_obj_sizes; ++i)
2389                 evacuate_block_obj_sizes [i] = FALSE;
2390
2391         sweep_slots_available = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2392         sweep_slots_used = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2393         sweep_num_blocks = sgen_alloc_internal_dynamic (sizeof (size_t) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2394
2395         /*
2396         {
2397                 int i;
2398                 g_print ("block object sizes:\n");
2399                 for (i = 0; i < num_block_obj_sizes; ++i)
2400                         g_print ("%d\n", block_obj_sizes [i]);
2401         }
2402         */
2403
2404         for (i = 0; i < MS_BLOCK_TYPE_MAX; ++i)
2405                 free_block_lists [i] = sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE);
2406
2407         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES; ++i)
2408                 fast_block_obj_size_indexes [i] = ms_find_block_obj_size_index (i * 8);
2409         for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i)
2410                 g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i));
2411
2412         mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced);
2413         mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed);
2414         mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept);
2415         mono_counters_register ("# major objects evacuated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_objects_evacuated);
2416 #if SIZEOF_VOID_P != 8
2417         mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal);
2418         mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal);
2419         mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual);
2420         mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal);
2421 #endif
2422
2423         collector->section_size = MAJOR_SECTION_SIZE;
2424
2425         concurrent_mark = is_concurrent;
2426         if (is_concurrent) {
2427                 collector->is_concurrent = TRUE;
2428                 collector->want_synchronous_collection = &want_evacuation;
2429         } else {
2430                 collector->is_concurrent = FALSE;
2431                 collector->want_synchronous_collection = NULL;
2432         }
2433         collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked;
2434         collector->supports_cardtable = TRUE;
2435
2436         collector->alloc_heap = major_alloc_heap;
2437         collector->is_object_live = major_is_object_live;
2438         collector->alloc_small_pinned_obj = major_alloc_small_pinned_obj;
2439         collector->alloc_degraded = major_alloc_degraded;
2440
2441         collector->alloc_object = major_alloc_object;
2442         collector->free_pinned_object = free_pinned_object;
2443         collector->iterate_objects = major_iterate_objects;
2444         collector->free_non_pinned_object = major_free_non_pinned_object;
2445         collector->pin_objects = major_pin_objects;
2446         collector->pin_major_object = pin_major_object;
2447         collector->scan_card_table = major_scan_card_table;
2448         collector->iterate_live_block_ranges = (void*)(void*) major_iterate_live_block_ranges;
2449         if (is_concurrent) {
2450                 collector->update_cardtable_mod_union = update_cardtable_mod_union;
2451                 collector->get_cardtable_mod_union_for_object = major_get_cardtable_mod_union_for_object;
2452         }
2453         collector->init_to_space = major_init_to_space;
2454         collector->sweep = major_sweep;
2455         collector->have_swept = major_have_swept;
2456         collector->finish_sweeping = major_finish_sweeping;
2457         collector->free_swept_blocks = major_free_swept_blocks;
2458         collector->check_scan_starts = major_check_scan_starts;
2459         collector->dump_heap = major_dump_heap;
2460         collector->get_used_size = major_get_used_size;
2461         collector->start_nursery_collection = major_start_nursery_collection;
2462         collector->finish_nursery_collection = major_finish_nursery_collection;
2463         collector->start_major_collection = major_start_major_collection;
2464         collector->finish_major_collection = major_finish_major_collection;
2465         collector->ptr_is_in_non_pinned_space = major_ptr_is_in_non_pinned_space;
2466         collector->obj_is_from_pinned_alloc = obj_is_from_pinned_alloc;
2467         collector->report_pinned_memory_usage = major_report_pinned_memory_usage;
2468         collector->get_num_major_sections = get_num_major_sections;
2469         collector->get_num_major_unswept_old_sections = get_num_major_unswept_old_sections;
2470         collector->handle_gc_param = major_handle_gc_param;
2471         collector->print_gc_param_usage = major_print_gc_param_usage;
2472         collector->post_param_init = post_param_init;
2473         collector->is_valid_object = major_is_valid_object;
2474         collector->describe_pointer = major_describe_pointer;
2475         collector->count_cards = major_count_cards;
2476
2477         collector->major_ops.copy_or_mark_object = major_copy_or_mark_object_canonical;
2478         collector->major_ops.scan_object = major_scan_object_with_evacuation;
2479         if (is_concurrent) {
2480                 collector->major_concurrent_ops.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical;
2481                 collector->major_concurrent_ops.scan_object = major_scan_object_no_mark_concurrent;
2482                 collector->major_concurrent_ops.scan_vtype = major_scan_vtype_concurrent;
2483         }
2484
2485 #if !defined (FIXED_HEAP) && !defined (SGEN_PARALLEL_MARK)
2486         /* FIXME: this will not work with evacuation or the split nursery. */
2487         if (!is_concurrent)
2488                 collector->drain_gray_stack = drain_gray_stack;
2489
2490 #ifdef HEAVY_STATISTICS
2491         mono_counters_register ("Optimized copy", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy);
2492         mono_counters_register ("Optimized copy nursery", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery);
2493         mono_counters_register ("Optimized copy nursery forwarded", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_forwarded);
2494         mono_counters_register ("Optimized copy nursery pinned", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_nursery_pinned);
2495         mono_counters_register ("Optimized copy major", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major);
2496         mono_counters_register ("Optimized copy major small fast", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_fast);
2497         mono_counters_register ("Optimized copy major small slow", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_small_slow);
2498         mono_counters_register ("Optimized copy major large", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_copy_major_large);
2499         mono_counters_register ("Optimized major scan", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan);
2500         mono_counters_register ("Optimized major scan no refs", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_optimized_major_scan_no_refs);
2501
2502         mono_counters_register ("Gray stack drain loops", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_loops);
2503         mono_counters_register ("Gray stack prefetch fills", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fills);
2504         mono_counters_register ("Gray stack prefetch failures", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_drain_prefetch_fill_failures);
2505 #endif
2506 #endif
2507
2508         mono_mutex_init (&allocated_blocks_lock);
2509
2510 #ifdef SGEN_HEAVY_BINARY_PROTOCOL
2511         mono_mutex_init (&scanned_objects_list_lock);
2512 #endif
2513
2514         SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2");
2515
2516         /*cardtable requires major pages to be 8 cards aligned*/
2517         g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0);
2518 }
2519
2520 void
2521 sgen_marksweep_init (SgenMajorCollector *collector)
2522 {
2523         sgen_marksweep_init_internal (collector, FALSE);
2524 }
2525
2526 void
2527 sgen_marksweep_conc_init (SgenMajorCollector *collector)
2528 {
2529         sgen_marksweep_init_internal (collector, TRUE);
2530 }
2531
2532 #endif