X-Git-Url: http://wien.tomnetworks.com/gitweb/?a=blobdiff_plain;f=mono%2Fsgen%2Fsgen-marksweep.c;h=7711679e3cb8239c3b06fdccf20f5c57cb84a82b;hb=3ea09f9c8f10097e50b94702c1e4a06cb7517c4c;hp=a215c44502c12b226359ff45fddafe511f569374;hpb=89ce2b94f3f9cec00446168096076499c1c1bde4;p=mono.git diff --git a/mono/sgen/sgen-marksweep.c b/mono/sgen/sgen-marksweep.c index a215c44502c..7711679e3cb 100644 --- a/mono/sgen/sgen-marksweep.c +++ b/mono/sgen/sgen-marksweep.c @@ -1,5 +1,6 @@ -/* - * sgen-marksweep.c: The Mark & Sweep major collector. +/** + * \file + * The Mark & Sweep major collector. * * Author: * Mark Probst @@ -32,15 +33,15 @@ #include "mono/sgen/sgen-client.h" #include "mono/utils/mono-memory-model.h" -#if defined(ARCH_MIN_MS_BLOCK_SIZE) && defined(ARCH_MIN_MS_BLOCK_SIZE_SHIFT) -#define MS_BLOCK_SIZE ARCH_MIN_MS_BLOCK_SIZE -#define MS_BLOCK_SIZE_SHIFT ARCH_MIN_MS_BLOCK_SIZE_SHIFT -#else -#define MS_BLOCK_SIZE_SHIFT 14 /* INT FASTENABLE */ -#define MS_BLOCK_SIZE (1 << MS_BLOCK_SIZE_SHIFT) -#endif -#define MAJOR_SECTION_SIZE MS_BLOCK_SIZE -#define CARDS_PER_BLOCK (MS_BLOCK_SIZE / CARD_SIZE_IN_BYTES) +static int ms_block_size; + +/* + * Blocks must be at least this size, meaning that if we detect a + * page size lower than this, we'll use this instead. + */ +#define MS_BLOCK_SIZE_MIN (1024 * 16) + +#define CARDS_PER_BLOCK (ms_block_size / CARD_SIZE_IN_BYTES) /* * Don't allocate single blocks, but alloc a contingent of this many @@ -48,16 +49,22 @@ */ #define MS_BLOCK_ALLOC_NUM 32 +#define MS_NUM_MARK_WORDS ((ms_block_size / SGEN_ALLOC_ALIGN + sizeof (guint32) * 8 - 1) / (sizeof (guint32) * 8)) + +/* + * Use this instead of sizeof (MSBlockInfo) since the mark_words + * array size depends on page size at runtime. + */ +#define SIZEOF_MS_BLOCK_INFO (sizeof (MSBlockInfo) + sizeof (guint32) * (MS_NUM_MARK_WORDS - MONO_ZERO_LEN_ARRAY)) + /* * Number of bytes before the first object in a block. At the start * of a block is the MSBlockHeader, then opional padding, then come - * the objects, so this must be >= sizeof (MSBlockHeader). + * the objects, so this must be >= SIZEOF_MS_BLOCK_INFO. */ -#define MS_BLOCK_SKIP ((sizeof (MSBlockHeader) + 15) & ~15) - -#define MS_BLOCK_FREE (MS_BLOCK_SIZE - MS_BLOCK_SKIP) +#define MS_BLOCK_SKIP ((SIZEOF_MS_BLOCK_INFO + 15) & ~15) -#define MS_NUM_MARK_WORDS ((MS_BLOCK_SIZE / SGEN_ALLOC_ALIGN + sizeof (mword) * 8 - 1) / (sizeof (mword) * 8)) +#define MS_BLOCK_FREE (ms_block_size - MS_BLOCK_SKIP) /* * Blocks progress from one state to the next: @@ -107,14 +114,14 @@ struct _MSBlockInfo { void ** volatile free_list; MSBlockInfo * volatile next_free; guint8 * volatile cardtable_mod_union; - mword mark_words [MS_NUM_MARK_WORDS]; + guint32 mark_words [MONO_ZERO_LEN_ARRAY]; }; #define MS_BLOCK_FOR_BLOCK_INFO(b) ((char*)(b)) #define MS_BLOCK_OBJ(b,i) ((GCObject *)(MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (b)->obj_size * (i))) #define MS_BLOCK_OBJ_FOR_SIZE(b,i,obj_size) (MS_BLOCK_FOR_BLOCK_INFO(b) + MS_BLOCK_SKIP + (obj_size) * (i)) -#define MS_BLOCK_DATA_FOR_OBJ(o) ((char*)((mword)(o) & ~(mword)(MS_BLOCK_SIZE - 1))) +#define MS_BLOCK_DATA_FOR_OBJ(o) ((char*)((mword)(o) & ~(mword)(ms_block_size - 1))) typedef struct { MSBlockInfo info; @@ -128,19 +135,28 @@ typedef struct { //casting to int is fine since blocks are 32k #define MS_CALC_MARK_BIT(w,b,o) do { \ int i = ((int)((char*)(o) - MS_BLOCK_DATA_FOR_OBJ ((o)))) >> SGEN_ALLOC_ALIGN_BITS; \ - if (sizeof (mword) == 4) { \ - (w) = i >> 5; \ - (b) = i & 31; \ - } else { \ - (w) = i >> 6; \ - (b) = i & 63; \ - } \ + (w) = i >> 5; \ + (b) = i & 31; \ } while (0) #define MS_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] & (ONE_P << (b))) #define MS_SET_MARK_BIT(bl,w,b) ((bl)->mark_words [(w)] |= (ONE_P << (b))) +#define MS_SET_MARK_BIT_PAR(bl,w,b,first) do { \ + guint32 tmp_mark_word = (bl)->mark_words [(w)]; \ + guint32 old_mark_word; \ + first = FALSE; \ + while (!(tmp_mark_word & (ONE_P << (b)))) { \ + old_mark_word = tmp_mark_word; \ + tmp_mark_word = InterlockedCompareExchange ((volatile gint32*)&(bl)->mark_words [w], old_mark_word | (ONE_P << (b)), old_mark_word); \ + if (tmp_mark_word == old_mark_word) { \ + first = TRUE; \ + break; \ + } \ + } \ + } while (0) -#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)) + +#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)) #define MS_BLOCK_OBJ_SIZE_FACTOR (pow (2.0, 1.0 / 3)) @@ -177,6 +193,9 @@ static volatile int sweep_state = SWEEP_STATE_SWEPT; static gboolean concurrent_mark; static gboolean concurrent_sweep = TRUE; +SgenThreadPool sweep_pool_inst; +SgenThreadPool *sweep_pool; + #define BLOCK_IS_TAGGED_HAS_REFERENCES(bl) SGEN_POINTER_IS_TAGGED_1 ((bl)) #define BLOCK_TAG_HAS_REFERENCES(bl) SGEN_POINTER_TAG_1 ((bl)) @@ -188,7 +207,7 @@ static gboolean concurrent_sweep = TRUE; #define BLOCK_TAG(bl) ((bl)->has_references ? BLOCK_TAG_HAS_REFERENCES ((bl)) : (bl)) /* all allocated blocks in the system */ -static SgenArrayList allocated_blocks = SGEN_ARRAY_LIST_INIT (NULL, NULL, NULL, INTERNAL_MEM_PIN_QUEUE); +static SgenArrayList allocated_blocks = SGEN_ARRAY_LIST_INIT (NULL, sgen_array_list_default_is_slot_set, sgen_array_list_default_cas_setter, INTERNAL_MEM_PIN_QUEUE); /* non-allocated block free-list */ static void *empty_blocks = NULL; @@ -235,17 +254,16 @@ static volatile size_t num_major_sections = 0; * thread only ever adds blocks to the free list, so the ABA problem can't occur. */ static MSBlockInfo * volatile *free_block_lists [MS_BLOCK_TYPE_MAX]; +static MonoNativeTlsKey worker_block_free_list_key; static guint64 stat_major_blocks_alloced = 0; static guint64 stat_major_blocks_freed = 0; static guint64 stat_major_blocks_lazy_swept = 0; -#if SIZEOF_VOID_P != 8 static guint64 stat_major_blocks_freed_ideal = 0; static guint64 stat_major_blocks_freed_less_ideal = 0; static guint64 stat_major_blocks_freed_individual = 0; static guint64 stat_major_blocks_alloced_less_ideal = 0; -#endif #ifdef SGEN_COUNT_NUMBER_OF_MAJOR_OBJECTS_MARKED static guint64 num_major_objects_marked = 0; @@ -286,6 +304,7 @@ ms_find_block_obj_size_index (size_t size) #define FREE_BLOCKS_FROM(lists,p,r) (lists [((p) ? MS_BLOCK_FLAG_PINNED : 0) | ((r) ? MS_BLOCK_FLAG_REFS : 0)]) #define FREE_BLOCKS(p,r) (FREE_BLOCKS_FROM (free_block_lists, (p), (r))) +#define FREE_BLOCKS_LOCAL(p,r) (FREE_BLOCKS_FROM (((MSBlockInfo***)mono_native_tls_get_value (worker_block_free_list_key)), (p), (r))) #define MS_BLOCK_OBJ_SIZE_INDEX(s) \ (((s)+7)>>3 < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES ? \ @@ -293,13 +312,13 @@ ms_find_block_obj_size_index (size_t size) ms_find_block_obj_size_index ((s))) static void* -major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits) +major_alloc_heap (mword nursery_size, mword nursery_align) { char *start; if (nursery_align) - start = (char *)sgen_alloc_os_memory_aligned (nursery_size, nursery_align, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery"); + start = (char *)sgen_alloc_os_memory_aligned (nursery_size, nursery_align, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery", MONO_MEM_ACCOUNT_SGEN_NURSERY); else - start = (char *)sgen_alloc_os_memory (nursery_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery"); + start = (char *)sgen_alloc_os_memory (nursery_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), "nursery", MONO_MEM_ACCOUNT_SGEN_NURSERY); return start; } @@ -307,7 +326,7 @@ major_alloc_heap (mword nursery_size, mword nursery_align, int the_nursery_bits) static void update_heap_boundaries_for_block (MSBlockInfo *block) { - sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE); + sgen_update_heap_boundaries ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), (mword)MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size); } /* @@ -329,9 +348,9 @@ ms_get_empty_block (void) */ int alloc_num = MS_BLOCK_ALLOC_NUM; for (;;) { - p = (char *)sgen_alloc_os_memory_aligned (MS_BLOCK_SIZE * alloc_num, MS_BLOCK_SIZE, + p = (char *)sgen_alloc_os_memory_aligned (ms_block_size * alloc_num, ms_block_size, (SgenAllocFlags)(SGEN_ALLOC_HEAP | SGEN_ALLOC_ACTIVATE), - alloc_num == 1 ? "major heap section" : NULL); + alloc_num == 1 ? "major heap section" : NULL, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP); if (p) break; alloc_num >>= 1; @@ -348,7 +367,7 @@ ms_get_empty_block (void) empty = empty_blocks; *(void**)block = empty; } while (SGEN_CAS_PTR ((gpointer*)&empty_blocks, block, empty) != empty); - p += MS_BLOCK_SIZE; + p += ms_block_size; } SGEN_ATOMIC_ADD_P (num_empty_blocks, alloc_num); @@ -372,7 +391,7 @@ ms_get_empty_block (void) *(void**)block = NULL; - g_assert (!((mword)block & (MS_BLOCK_SIZE - 1))); + g_assert (!((mword)block & (ms_block_size - 1))); return block; } @@ -387,10 +406,10 @@ ms_free_block (MSBlockInfo *info) void *empty; char *block = MS_BLOCK_FOR_BLOCK_INFO (info); - sgen_memgov_release_space (MS_BLOCK_SIZE, SPACE_MAJOR); + sgen_memgov_release_space (ms_block_size, SPACE_MAJOR); if (info->cardtable_mod_union) - sgen_card_table_free_mod_union (info->cardtable_mod_union, block, MS_BLOCK_SIZE); - memset (block, 0, MS_BLOCK_SIZE); + sgen_card_table_free_mod_union (info->cardtable_mod_union, block, ms_block_size); + memset (block, 0, ms_block_size); do { empty = empty_blocks; @@ -399,7 +418,7 @@ ms_free_block (MSBlockInfo *info) SGEN_ATOMIC_ADD_P (num_empty_blocks, 1); - binary_protocol_block_free (block, MS_BLOCK_SIZE); + binary_protocol_block_free (block, ms_block_size); } static gboolean @@ -514,7 +533,7 @@ ms_alloc_block (int size_index, gboolean pinned, gboolean has_references) char *obj_start; int i; - if (!sgen_memgov_try_alloc_space (MS_BLOCK_SIZE, SPACE_MAJOR)) + if (!sgen_memgov_try_alloc_space (ms_block_size, SPACE_MAJOR)) return FALSE; info = (MSBlockInfo*)ms_get_empty_block (); @@ -540,7 +559,7 @@ ms_alloc_block (int size_index, gboolean pinned, gboolean has_references) update_heap_boundaries_for_block (info); - binary_protocol_block_alloc (info, MS_BLOCK_SIZE); + binary_protocol_block_alloc (info, ms_block_size); /* build free list */ obj_start = MS_BLOCK_FOR_BLOCK_INFO (info) + MS_BLOCK_SKIP; @@ -568,7 +587,7 @@ ptr_is_in_major_block (char *ptr, char **start, gboolean *pinned) MSBlockInfo *block; FOREACH_BLOCK_NO_LOCK (block) { - if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) { + if (ptr >= MS_BLOCK_FOR_BLOCK_INFO (block) && ptr <= MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size) { int count = MS_BLOCK_FREE / block->obj_size; int i; @@ -684,6 +703,55 @@ major_alloc_object (GCVTable vtable, size_t size, gboolean has_references) return alloc_obj (vtable, size, FALSE, has_references); } +/* + * This can only be called by sgen workers. While this is called we assume + * that no other thread is accessing the block free lists. The world should + * be stopped and the gc thread should be waiting for workers to finish. + */ +static GCObject* +major_alloc_object_par (GCVTable vtable, size_t size, gboolean has_references) +{ + int size_index = MS_BLOCK_OBJ_SIZE_INDEX (size); + MSBlockInfo * volatile * free_blocks = FREE_BLOCKS (FALSE, has_references); + MSBlockInfo **free_blocks_local = FREE_BLOCKS_LOCAL (FALSE, has_references); + void *obj; + + if (free_blocks_local [size_index]) { +get_slot: + obj = unlink_slot_from_free_list_uncontested (free_blocks_local, size_index); + } else { + MSBlockInfo *block; +get_block: + block = free_blocks [size_index]; + if (!block) { + if (G_UNLIKELY (!ms_alloc_block (size_index, FALSE, has_references))) + return NULL; + goto get_block; + } else { + MSBlockInfo *next_free = block->next_free; + /* + * Once a block is removed from the main list, it cannot return on the list until + * all the workers are finished and sweep is starting. This means we don't need + * to account for ABA problems. + */ + if (SGEN_CAS_PTR ((volatile gpointer *)&free_blocks [size_index], next_free, block) != block) + goto get_block; + block->next_free = free_blocks_local [size_index]; + free_blocks_local [size_index] = block; + + goto get_slot; + } + } + + /* FIXME: assumes object layout */ + *(GCVTable*)obj = vtable; + + /* FIXME is it worth CAS-ing here */ + total_allocated_major += block_obj_sizes [size_index]; + + return (GCObject *)obj; +} + /* * We're not freeing the block if it's empty. We leave that work for * the next major collection. @@ -859,7 +927,7 @@ major_finish_sweep_checking (void) wait: job = sweep_job; if (job) - sgen_thread_pool_job_wait (job); + sgen_thread_pool_job_wait (sweep_pool, job); SGEN_ASSERT (0, !sweep_job, "Why did the sweep job not null itself?"); SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "How is the sweep job done but we're not swept?"); } @@ -872,6 +940,7 @@ major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc call gboolean pinned = flags & ITERATE_OBJECTS_PINNED; MSBlockInfo *block; + /* No actual sweeping will take place if we are in the middle of a major collection. */ major_finish_sweep_checking (); FOREACH_BLOCK_NO_LOCK (block) { int count = MS_BLOCK_FREE / block->obj_size; @@ -881,26 +950,13 @@ major_iterate_objects (IterateObjectsFlags flags, IterateObjectCallbackFunc call continue; if (!block->pinned && !non_pinned) continue; - if (sweep && lazy_sweep) { + if (sweep && lazy_sweep && !block_is_swept_or_marking (block)) { sweep_block (block); SGEN_ASSERT (6, block->state == BLOCK_STATE_SWEPT, "Block must be swept after sweeping"); } for (i = 0; i < count; ++i) { void **obj = (void**) MS_BLOCK_OBJ (block, i); - /* - * We've finished sweep checking, but if we're sweeping lazily and - * the flags don't require us to sweep, the block might still need - * sweeping. In that case, we need to consult the mark bits to tell - * us whether an object slot is live. - */ - if (!block_is_swept_or_marking (block)) { - int word, bit; - SGEN_ASSERT (6, !sweep && block->state == BLOCK_STATE_NEED_SWEEPING, "Has sweeping not finished?"); - MS_CALC_MARK_BIT (word, bit, obj); - if (!MS_MARK_BIT (block, word, bit)) - continue; - } if (MS_OBJ_ALLOCED (obj, block)) callback ((GCObject*)obj, block->obj_size, data); } @@ -916,7 +972,7 @@ major_is_valid_object (char *object) int idx; char *obj; - if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= object)) + if ((MS_BLOCK_FOR_BLOCK_INFO (block) > object) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size) <= object)) continue; idx = MS_BLOCK_OBJ_INDEX (object, block); @@ -943,7 +999,7 @@ major_describe_pointer (char *ptr) int w, b; gboolean marked; - if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE) <= ptr)) + if ((MS_BLOCK_FOR_BLOCK_INFO (block) > ptr) || ((MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size) <= ptr)) continue; SGEN_LOG (0, "major-ptr (block %p sz %d pin %d ref %d)\n", @@ -1044,13 +1100,13 @@ get_cardtable_mod_union_for_block (MSBlockInfo *block, gboolean allocate) return mod_union; else if (!allocate) return NULL; - mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE); + mod_union = sgen_card_table_alloc_mod_union (MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size); other = (guint8 *)SGEN_CAS_PTR ((gpointer*)&block->cardtable_mod_union, mod_union, NULL); if (!other) { SGEN_ASSERT (0, block->cardtable_mod_union == mod_union, "Why did CAS not replace?"); return mod_union; } - sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE); + sgen_card_table_free_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size); return other; } @@ -1098,12 +1154,27 @@ major_block_is_evacuating (MSBlockInfo *block) if (!MS_MARK_BIT ((block), __word, __bit)) { \ MS_SET_MARK_BIT ((block), __word, __bit); \ if (sgen_gc_descr_has_references (desc)) \ - GRAY_OBJECT_ENQUEUE ((queue), (obj), (desc)); \ + GRAY_OBJECT_ENQUEUE_SERIAL ((queue), (obj), (desc)); \ + binary_protocol_mark ((obj), (gpointer)SGEN_LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \ + INC_NUM_MAJOR_OBJECTS_MARKED (); \ + } \ + } while (0) +#define MS_MARK_OBJECT_AND_ENQUEUE_PAR(obj,desc,block,queue) do { \ + int __word, __bit; \ + gboolean first; \ + MS_CALC_MARK_BIT (__word, __bit, (obj)); \ + SGEN_ASSERT (9, MS_OBJ_ALLOCED ((obj), (block)), "object %p not allocated", obj); \ + MS_SET_MARK_BIT_PAR ((block), __word, __bit, first); \ + if (first) { \ + if (sgen_gc_descr_has_references (desc)) \ + GRAY_OBJECT_ENQUEUE_PARALLEL ((queue), (obj), (desc)); \ binary_protocol_mark ((obj), (gpointer)SGEN_LOAD_VTABLE ((obj)), sgen_safe_object_get_size ((obj))); \ INC_NUM_MAJOR_OBJECTS_MARKED (); \ } \ } while (0) + + static void pin_major_object (GCObject *obj, SgenGrayQueue *queue) { @@ -1117,6 +1188,7 @@ pin_major_object (GCObject *obj, SgenGrayQueue *queue) MS_MARK_OBJECT_AND_ENQUEUE (obj, sgen_obj_get_descriptor (obj), block, queue); } +#define COPY_OR_MARK_PARALLEL #include "sgen-major-copy-object.h" static long long @@ -1167,6 +1239,12 @@ static guint64 stat_drain_loops; #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_no_evacuation #include "sgen-marksweep-drain-gray-stack.h" +#define COPY_OR_MARK_PARALLEL +#define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_par_no_evacuation +#define SCAN_OBJECT_FUNCTION_NAME major_scan_object_par_no_evacuation +#define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_par_no_evacuation +#include "sgen-marksweep-drain-gray-stack.h" + #define COPY_OR_MARK_WITH_EVACUATION #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_with_evacuation #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_with_evacuation @@ -1175,12 +1253,28 @@ static guint64 stat_drain_loops; #define SCAN_PTR_FIELD_FUNCTION_NAME major_scan_ptr_field_with_evacuation #include "sgen-marksweep-drain-gray-stack.h" +#define COPY_OR_MARK_PARALLEL +#define COPY_OR_MARK_WITH_EVACUATION +#define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_par_with_evacuation +#define SCAN_OBJECT_FUNCTION_NAME major_scan_object_par_with_evacuation +#define SCAN_VTYPE_FUNCTION_NAME major_scan_vtype_par_with_evacuation +#define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_par_with_evacuation +#define SCAN_PTR_FIELD_FUNCTION_NAME major_scan_ptr_field_par_with_evacuation +#include "sgen-marksweep-drain-gray-stack.h" + #define COPY_OR_MARK_CONCURRENT #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_concurrent_no_evacuation #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_concurrent_no_evacuation #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_concurrent_no_evacuation #include "sgen-marksweep-drain-gray-stack.h" +#define COPY_OR_MARK_PARALLEL +#define COPY_OR_MARK_CONCURRENT +#define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_concurrent_par_no_evacuation +#define SCAN_OBJECT_FUNCTION_NAME major_scan_object_concurrent_par_no_evacuation +#define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_concurrent_par_no_evacuation +#include "sgen-marksweep-drain-gray-stack.h" + #define COPY_OR_MARK_CONCURRENT_WITH_EVACUATION #define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_concurrent_with_evacuation #define SCAN_OBJECT_FUNCTION_NAME major_scan_object_concurrent_with_evacuation @@ -1189,6 +1283,15 @@ static guint64 stat_drain_loops; #define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_concurrent_with_evacuation #include "sgen-marksweep-drain-gray-stack.h" +#define COPY_OR_MARK_PARALLEL +#define COPY_OR_MARK_CONCURRENT_WITH_EVACUATION +#define COPY_OR_MARK_FUNCTION_NAME major_copy_or_mark_object_concurrent_par_with_evacuation +#define SCAN_OBJECT_FUNCTION_NAME major_scan_object_concurrent_par_with_evacuation +#define SCAN_VTYPE_FUNCTION_NAME major_scan_vtype_concurrent_par_with_evacuation +#define SCAN_PTR_FIELD_FUNCTION_NAME major_scan_ptr_field_concurrent_par_with_evacuation +#define DRAIN_GRAY_STACK_FUNCTION_NAME drain_gray_stack_concurrent_par_with_evacuation +#include "sgen-marksweep-drain-gray-stack.h" + static inline gboolean major_is_evacuating (void) { @@ -1211,6 +1314,15 @@ drain_gray_stack (SgenGrayQueue *queue) return drain_gray_stack_no_evacuation (queue); } +static gboolean +drain_gray_stack_par (SgenGrayQueue *queue) +{ + if (major_is_evacuating ()) + return drain_gray_stack_par_with_evacuation (queue); + else + return drain_gray_stack_par_no_evacuation (queue); +} + static gboolean drain_gray_stack_concurrent (SgenGrayQueue *queue) { @@ -1220,6 +1332,15 @@ drain_gray_stack_concurrent (SgenGrayQueue *queue) return drain_gray_stack_concurrent_no_evacuation (queue); } +static gboolean +drain_gray_stack_concurrent_par (SgenGrayQueue *queue) +{ + if (major_is_evacuating ()) + return drain_gray_stack_concurrent_par_with_evacuation (queue); + else + return drain_gray_stack_concurrent_par_no_evacuation (queue); +} + static void major_copy_or_mark_object_canonical (GCObject **ptr, SgenGrayQueue *queue) { @@ -1232,12 +1353,24 @@ major_copy_or_mark_object_concurrent_canonical (GCObject **ptr, SgenGrayQueue *q major_copy_or_mark_object_concurrent_with_evacuation (ptr, *ptr, queue); } +static void +major_copy_or_mark_object_concurrent_par_canonical (GCObject **ptr, SgenGrayQueue *queue) +{ + major_copy_or_mark_object_concurrent_par_with_evacuation (ptr, *ptr, queue); +} + static void major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQueue *queue) { major_copy_or_mark_object_with_evacuation (ptr, *ptr, queue); } +static void +major_copy_or_mark_object_concurrent_par_finish_canonical (GCObject **ptr, SgenGrayQueue *queue) +{ + major_copy_or_mark_object_par_with_evacuation (ptr, *ptr, queue); +} + static void mark_pinned_objects_in_block (MSBlockInfo *block, size_t first_entry, size_t last_entry, SgenGrayQueue *queue) { @@ -1276,10 +1409,10 @@ static inline void sweep_block_for_size (MSBlockInfo *block, int count, int obj_size) { int obj_index; + void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, 0, obj_size); - for (obj_index = 0; obj_index < count; ++obj_index) { + for (obj_index = 0; obj_index < count; ++obj_index, obj = (void*)((mword)obj + obj_size)) { int word, bit; - void *obj = MS_BLOCK_OBJ_FOR_SIZE (block, obj_index, obj_size); MS_CALC_MARK_BIT (word, bit, obj); if (MS_MARK_BIT (block, word, bit)) { @@ -1308,7 +1441,7 @@ try_set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state gint32 old_state = SGEN_CAS (&block->state, new_state, expected_state); gboolean success = old_state == expected_state; if (success) - binary_protocol_block_set_state (block, MS_BLOCK_SIZE, old_state, new_state); + binary_protocol_block_set_state (block, ms_block_size, old_state, new_state); return success; } @@ -1317,7 +1450,7 @@ set_block_state (MSBlockInfo *block, gint32 new_state, gint32 expected_state) { SGEN_ASSERT (6, block->state == expected_state, "Block state incorrect before set"); block->state = new_state; - binary_protocol_block_set_state (block, MS_BLOCK_SIZE, expected_state, new_state); + binary_protocol_block_set_state (block, ms_block_size, expected_state, new_state); } /* @@ -1370,7 +1503,7 @@ sweep_block (MSBlockInfo *block) } /* reset mark bits */ - memset (block->mark_words, 0, sizeof (mword) * MS_NUM_MARK_WORDS); + memset (block->mark_words, 0, sizeof (guint32) * MS_NUM_MARK_WORDS); /* Reverse free list so that it's in address order */ reversed = NULL; @@ -1416,6 +1549,40 @@ static size_t *sweep_num_blocks; static volatile size_t num_major_sections_before_sweep; static volatile size_t num_major_sections_freed_in_sweep; +static void +sgen_worker_clear_free_block_lists (WorkerData *worker) +{ + int i, j; + + if (!worker->free_block_lists) + return; + + for (i = 0; i < MS_BLOCK_TYPE_MAX; i++) { + for (j = 0; j < num_block_obj_sizes; j++) { + ((MSBlockInfo***) worker->free_block_lists) [i][j] = NULL; + } + } +} + +static void +sgen_worker_clear_free_block_lists_evac (WorkerData *worker) +{ + int i, j; + + if (!worker->free_block_lists) + return; + + for (i = 0; i < MS_BLOCK_TYPE_MAX; i++) { + for (j = 0; j < num_block_obj_sizes; j++) { + if (((MSBlockInfo***) worker->free_block_lists) [i][j]) + SGEN_ASSERT (0, !((MSBlockInfo***) worker->free_block_lists) [i][j]->next_free, "Why do we have linked free blocks on the workers"); + + if (evacuate_block_obj_sizes [j]) + ((MSBlockInfo***) worker->free_block_lists) [i][j] = NULL; + } + } +} + static void sweep_start (void) { @@ -1432,7 +1599,7 @@ sweep_start (void) free_blocks [j] = NULL; } - sgen_array_list_remove_nulls (&allocated_blocks); + sgen_workers_foreach (sgen_worker_clear_free_block_lists); } static void sweep_finish (void); @@ -1574,6 +1741,7 @@ ensure_block_is_checked_for_sweeping (guint32 block_index, gboolean wait, gboole ms_free_block (block); SGEN_ATOMIC_ADD_P (num_major_sections, -1); + SGEN_ATOMIC_ADD_P (num_major_sections_freed_in_sweep, 1); tagged_block = NULL; } @@ -1620,13 +1788,8 @@ sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job) * cooperate with the sweep thread to finish sweeping, and they will traverse from * low to high, to avoid constantly colliding on the same blocks. */ - for (block_index = num_blocks; block_index-- > 0;) { - /* - * The block might have been freed by another thread doing some checking - * work. - */ - if (!ensure_block_is_checked_for_sweeping (block_index, TRUE, NULL)) - ++num_major_sections_freed_in_sweep; + for (block_index = allocated_blocks.next_slot; block_index-- > 0;) { + ensure_block_is_checked_for_sweeping (block_index, TRUE, NULL); } while (!try_set_sweep_state (SWEEP_STATE_COMPACTING, SWEEP_STATE_SWEEPING)) { @@ -1652,7 +1815,7 @@ sweep_job_func (void *thread_data_untyped, SgenThreadPoolJob *job) */ if (concurrent_sweep && lazy_sweep) { sweep_blocks_job = sgen_thread_pool_job_alloc ("sweep_blocks", sweep_blocks_job_func, sizeof (SgenThreadPoolJob)); - sgen_thread_pool_job_enqueue (sweep_blocks_job); + sgen_thread_pool_job_enqueue (sweep_pool, sweep_blocks_job); } sweep_finish (); @@ -1695,15 +1858,13 @@ major_sweep (void) sweep_start (); - SGEN_ASSERT (0, num_major_sections == allocated_blocks.next_slot, "We don't know how many blocks we have?"); - num_major_sections_before_sweep = num_major_sections; num_major_sections_freed_in_sweep = 0; SGEN_ASSERT (0, !sweep_job, "We haven't finished the last sweep?"); if (concurrent_sweep) { sweep_job = sgen_thread_pool_job_alloc ("sweep", sweep_job_func, sizeof (SgenThreadPoolJob)); - sgen_thread_pool_job_enqueue (sweep_job); + sgen_thread_pool_job_enqueue (sweep_pool, sweep_job); } else { sweep_job_func (NULL, NULL); } @@ -1905,6 +2066,9 @@ major_start_major_collection (void) sgen_evacuation_freelist_blocks (&free_block_lists [MS_BLOCK_FLAG_REFS][i], i); } + /* We expect workers to have very few blocks on the freelist, just evacuate them */ + sgen_workers_foreach (sgen_worker_clear_free_block_lists_evac); + if (lazy_sweep && concurrent_sweep) { /* * sweep_blocks_job is created before sweep_finish, which we wait for above @@ -1913,7 +2077,7 @@ major_start_major_collection (void) */ SgenThreadPoolJob *job = sweep_blocks_job; if (job) - sgen_thread_pool_job_wait (job); + sgen_thread_pool_job_wait (sweep_pool, job); } if (lazy_sweep && !concurrent_sweep) @@ -1953,7 +2117,12 @@ major_finish_major_collection (ScannedObjectCounts *counts) #endif } -#if SIZEOF_VOID_P != 8 +static SgenThreadPool* +major_get_sweep_pool (void) +{ + return sweep_pool; +} + static int compare_pointers (const void *va, const void *vb) { char *a = *(char**)va, *b = *(char**)vb; @@ -1963,17 +2132,13 @@ compare_pointers (const void *va, const void *vb) { return 1; return 0; } -#endif /* * This is called with sweep completed and the world stopped. */ static void -major_free_swept_blocks (size_t allowance) +major_free_swept_blocks (size_t section_reserve) { - /* FIXME: This is probably too much. It's assuming all objects are small. */ - size_t section_reserve = allowance / MS_BLOCK_SIZE; - SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks"); #ifdef TARGET_WIN32 @@ -1984,7 +2149,6 @@ major_free_swept_blocks (size_t allowance) return; #endif -#if SIZEOF_VOID_P != 8 { int i, num_empty_blocks_orig, num_blocks, arr_length; void *block; @@ -2043,7 +2207,7 @@ major_free_swept_blocks (size_t allowance) SGEN_ASSERT (6, first >= 0 && d > first, "algorithm is wrong"); - if ((char*)block != ((char*)empty_block_arr [d-1]) + MS_BLOCK_SIZE) { + if ((char*)block != ((char*)empty_block_arr [d-1]) + ms_block_size) { first = d; continue; } @@ -2057,7 +2221,7 @@ major_free_swept_blocks (size_t allowance) * we're iterating. */ int j; - sgen_free_os_memory (empty_block_arr [first], MS_BLOCK_SIZE * num_blocks, SGEN_ALLOC_HEAP); + sgen_free_os_memory (empty_block_arr [first], ms_block_size * num_blocks, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP); for (j = first; j <= d; ++j) empty_block_arr [j] = NULL; dest = first; @@ -2104,11 +2268,10 @@ major_free_swept_blocks (size_t allowance) */ if (num_empty_blocks <= num_major_sections) return; -#endif while (num_empty_blocks > section_reserve) { void *next = *(void**)empty_blocks; - sgen_free_os_memory (empty_blocks, MS_BLOCK_SIZE, SGEN_ALLOC_HEAP); + sgen_free_os_memory (empty_blocks, ms_block_size, SGEN_ALLOC_HEAP, MONO_MEM_ACCOUNT_SGEN_MARKSWEEP); empty_blocks = next; /* * Needs not be atomic because this is running @@ -2117,9 +2280,7 @@ major_free_swept_blocks (size_t allowance) --num_empty_blocks; ++stat_major_blocks_freed; -#if SIZEOF_VOID_P != 8 ++stat_major_blocks_freed_individual; -#endif } } @@ -2131,7 +2292,7 @@ major_pin_objects (SgenGrayQueue *queue) FOREACH_BLOCK_NO_LOCK (block) { size_t first_entry, last_entry; SGEN_ASSERT (6, block_is_swept_or_marking (block), "All blocks must be swept when we're pinning."); - sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SIZE, + sgen_find_optimized_pin_queue_area (MS_BLOCK_FOR_BLOCK_INFO (block) + MS_BLOCK_SKIP, MS_BLOCK_FOR_BLOCK_INFO (block) + ms_block_size, &first_entry, &last_entry); mark_pinned_objects_in_block (block, first_entry, last_entry, queue); } END_FOREACH_BLOCK_NO_LOCK; @@ -2187,7 +2348,7 @@ static size_t get_bytes_survived_last_sweep (void) { SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Can only query unswept sections after sweep"); - return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * MS_BLOCK_SIZE; + return (num_major_sections_before_sweep - num_major_sections_freed_in_sweep) * ms_block_size; } static gboolean @@ -2241,7 +2402,7 @@ major_iterate_block_ranges (sgen_cardtable_block_callback callback) FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) { if (has_references) - callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE); + callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size); } END_FOREACH_BLOCK_NO_LOCK; } @@ -2254,7 +2415,7 @@ major_iterate_live_block_ranges (sgen_cardtable_block_callback callback) major_finish_sweep_checking (); FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) { if (has_references) - callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE); + callback ((mword)MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size); } END_FOREACH_BLOCK_NO_LOCK; } @@ -2275,7 +2436,7 @@ static guint8* initial_skip_card (guint8 *card_data) { mword *cards = (mword*)card_data; - mword card; + mword card = 0; int i; for (i = 0; i < CARD_WORDS_PER_BLOCK; ++i) { card = cards [i]; @@ -2303,17 +2464,22 @@ initial_skip_card (guint8 *card_data) #define MS_BLOCK_OBJ_INDEX_FAST(o,b,os) (((char*)(o) - ((b) + MS_BLOCK_SKIP)) / (os)) #define MS_BLOCK_OBJ_FAST(b,os,i) ((b) + MS_BLOCK_SKIP + (os) * (i)) -#define MS_OBJ_ALLOCED_FAST(o,b) (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + MS_BLOCK_SIZE)) +#define MS_OBJ_ALLOCED_FAST(o,b) (*(void**)(o) && (*(char**)(o) < (b) || *(char**)(o) >= (b) + ms_block_size)) static void scan_card_table_for_block (MSBlockInfo *block, CardTableScanType scan_type, ScanCopyContext ctx) { SgenGrayQueue *queue = ctx.queue; ScanObjectFunc scan_func = ctx.ops->scan_object; + /* + * FIXME: On systems with very large pages, we allocate fairly large + * arrays on the stack here. This shouldn't be a problem once block + * size is no longer required to be a multiple of the system page size. + */ #ifndef SGEN_HAVE_OVERLAPPING_CARDS - guint8 cards_copy [CARDS_PER_BLOCK]; + guint8 *cards_copy = alloca (sizeof (guint8) * CARDS_PER_BLOCK); #endif - guint8 cards_preclean [CARDS_PER_BLOCK]; + guint8 *cards_preclean = alloca (sizeof (guint8) * CARDS_PER_BLOCK); gboolean small_objects; int block_obj_size; char *block_start; @@ -2441,7 +2607,7 @@ scan_card_table_for_block (MSBlockInfo *block, CardTableScanType scan_type, Scan } static void -major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx) +major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx, int job_index, int job_split_count) { MSBlockInfo *block; gboolean has_references, was_sweeping, skip_scan; @@ -2455,8 +2621,10 @@ major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx) binary_protocol_major_card_table_scan_start (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION); FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) { + if (__index % job_split_count != job_index) + continue; #ifdef PREFETCH_CARDS - int prefetch_index = __index + 6; + int prefetch_index = __index + 6 * job_split_count; if (prefetch_index < allocated_blocks.next_slot) { MSBlockInfo *prefetch_block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, prefetch_index)); PREFETCH_READ (prefetch_block); @@ -2554,7 +2722,7 @@ update_cardtable_mod_union (void) if (has_dirty_cards) { size_t num_cards; guint8 *mod_union = get_cardtable_mod_union_for_block (block, TRUE); - sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), MS_BLOCK_SIZE, &num_cards); + sgen_card_table_update_mod_union (mod_union, MS_BLOCK_FOR_BLOCK_INFO (block), ms_block_size, &num_cards); SGEN_ASSERT (6, num_cards == CARDS_PER_BLOCK, "Number of cards calculation is wrong"); } } END_FOREACH_BLOCK_NO_LOCK; @@ -2566,15 +2734,41 @@ static void post_param_init (SgenMajorCollector *collector) { collector->sweeps_lazily = lazy_sweep; - collector->needs_thread_pool = concurrent_mark || concurrent_sweep; } +/* We are guaranteed to be called by the worker in question */ static void -sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent) +sgen_worker_init_callback (gpointer worker_untyped) { int i; + WorkerData *worker = (WorkerData*) worker_untyped; + MSBlockInfo ***worker_free_blocks = (MSBlockInfo ***) sgen_alloc_internal_dynamic (sizeof (MSBlockInfo**) * MS_BLOCK_TYPE_MAX, INTERNAL_MEM_MS_TABLES, TRUE); - sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, sizeof (MSBlockInfo)); + for (i = 0; i < MS_BLOCK_TYPE_MAX; i++) + worker_free_blocks [i] = (MSBlockInfo **) sgen_alloc_internal_dynamic (sizeof (MSBlockInfo*) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE); + + worker->free_block_lists = worker_free_blocks; + + mono_native_tls_set_value (worker_block_free_list_key, worker_free_blocks); +} + +static void +thread_pool_init_func (void *data_untyped) +{ + sgen_client_thread_register_worker (); +} + +static void +sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent, gboolean is_parallel) +{ + int i; + + ms_block_size = mono_pagesize (); + + if (ms_block_size < MS_BLOCK_SIZE_MIN) + ms_block_size = MS_BLOCK_SIZE_MIN; + + sgen_register_fixed_internal_mem_type (INTERNAL_MEM_MS_BLOCK_INFO, SIZEOF_MS_BLOCK_INFO); num_block_obj_sizes = ms_calculate_block_obj_sizes (MS_BLOCK_OBJ_SIZE_FACTOR, NULL); block_obj_sizes = (int *)sgen_alloc_internal_dynamic (sizeof (int) * num_block_obj_sizes, INTERNAL_MEM_MS_TABLES, TRUE); @@ -2605,21 +2799,25 @@ sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurr for (i = 0; i < MS_NUM_FAST_BLOCK_OBJ_SIZE_INDEXES * 8; ++i) g_assert (MS_BLOCK_OBJ_SIZE_INDEX (i) == ms_find_block_obj_size_index (i)); + /* We can do this because we always init the minor before the major */ + if (is_parallel || sgen_get_minor_collector ()->is_parallel) { + mono_native_tls_alloc (&worker_block_free_list_key, NULL); + collector->worker_init_cb = sgen_worker_init_callback; + } + mono_counters_register ("# major blocks allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced); mono_counters_register ("# major blocks freed", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed); mono_counters_register ("# major blocks lazy swept", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_lazy_swept); -#if SIZEOF_VOID_P != 8 mono_counters_register ("# major blocks freed ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_ideal); mono_counters_register ("# major blocks freed less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_less_ideal); mono_counters_register ("# major blocks freed individually", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_freed_individual); mono_counters_register ("# major blocks allocated less ideally", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_major_blocks_alloced_less_ideal); -#endif - collector->section_size = MAJOR_SECTION_SIZE; + collector->section_size = ms_block_size; concurrent_mark = is_concurrent; collector->is_concurrent = is_concurrent; - collector->needs_thread_pool = is_concurrent || concurrent_sweep; + collector->is_parallel = is_parallel; collector->get_and_reset_num_major_objects_marked = major_get_and_reset_num_major_objects_marked; collector->supports_cardtable = TRUE; @@ -2629,6 +2827,7 @@ sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurr collector->alloc_degraded = major_alloc_degraded; collector->alloc_object = major_alloc_object; + collector->alloc_object_par = major_alloc_object_par; collector->free_pinned_object = free_pinned_object; collector->iterate_objects = major_iterate_objects; collector->free_non_pinned_object = major_free_non_pinned_object; @@ -2664,9 +2863,11 @@ sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurr collector->is_valid_object = major_is_valid_object; collector->describe_pointer = major_describe_pointer; collector->count_cards = major_count_cards; + collector->get_sweep_pool = major_get_sweep_pool; collector->major_ops_serial.copy_or_mark_object = major_copy_or_mark_object_canonical; collector->major_ops_serial.scan_object = major_scan_object_with_evacuation; + collector->major_ops_serial.scan_ptr_field = major_scan_ptr_field_with_evacuation; collector->major_ops_serial.drain_gray_stack = drain_gray_stack; if (is_concurrent) { collector->major_ops_concurrent_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_canonical; @@ -2680,6 +2881,20 @@ sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurr collector->major_ops_concurrent_finish.scan_vtype = major_scan_vtype_with_evacuation; collector->major_ops_concurrent_finish.scan_ptr_field = major_scan_ptr_field_with_evacuation; collector->major_ops_concurrent_finish.drain_gray_stack = drain_gray_stack; + + if (is_parallel) { + collector->major_ops_conc_par_start.copy_or_mark_object = major_copy_or_mark_object_concurrent_par_canonical; + collector->major_ops_conc_par_start.scan_object = major_scan_object_concurrent_par_with_evacuation; + collector->major_ops_conc_par_start.scan_vtype = major_scan_vtype_concurrent_par_with_evacuation; + collector->major_ops_conc_par_start.scan_ptr_field = major_scan_ptr_field_concurrent_par_with_evacuation; + collector->major_ops_conc_par_start.drain_gray_stack = drain_gray_stack_concurrent_par; + + collector->major_ops_conc_par_finish.copy_or_mark_object = major_copy_or_mark_object_concurrent_par_finish_canonical; + collector->major_ops_conc_par_finish.scan_object = major_scan_object_par_with_evacuation; + collector->major_ops_conc_par_finish.scan_vtype = major_scan_vtype_par_with_evacuation; + collector->major_ops_conc_par_finish.scan_ptr_field = major_scan_ptr_field_par_with_evacuation; + collector->major_ops_conc_par_finish.drain_gray_stack = drain_gray_stack_par; + } } #ifdef HEAVY_STATISTICS @@ -2707,19 +2922,31 @@ sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurr SGEN_ASSERT (0, SGEN_MAX_SMALL_OBJ_SIZE <= MS_BLOCK_FREE / 2, "MAX_SMALL_OBJ_SIZE must be at most MS_BLOCK_FREE / 2"); /*cardtable requires major pages to be 8 cards aligned*/ - g_assert ((MS_BLOCK_SIZE % (8 * CARD_SIZE_IN_BYTES)) == 0); + g_assert ((ms_block_size % (8 * CARD_SIZE_IN_BYTES)) == 0); + + if (concurrent_sweep) { + SgenThreadPool **thread_datas = &sweep_pool; + sweep_pool = &sweep_pool_inst; + sgen_thread_pool_init (sweep_pool, 1, thread_pool_init_func, NULL, NULL, NULL, (SgenThreadPoolData**)&thread_datas); + } } void sgen_marksweep_init (SgenMajorCollector *collector) { - sgen_marksweep_init_internal (collector, FALSE); + sgen_marksweep_init_internal (collector, FALSE, FALSE); } void sgen_marksweep_conc_init (SgenMajorCollector *collector) { - sgen_marksweep_init_internal (collector, TRUE); + sgen_marksweep_init_internal (collector, TRUE, FALSE); +} + +void +sgen_marksweep_conc_par_init (SgenMajorCollector *collector) +{ + sgen_marksweep_init_internal (collector, TRUE, TRUE); } #endif