[sgen] Blocks on the freelist don't need to have been swept
[mono.git] / mono / sgen / sgen-marksweep.c
index d6c4b66785b3eaf58dac2f7d29a6cfce1c439e5f..450adcf3443a5d504f1dc843b29293ce0c1a95d3 100644 (file)
@@ -1,5 +1,6 @@
-/*
- * sgen-marksweep.c: The Mark & Sweep major collector.
+/**
+ * \file
+ * The Mark & Sweep major collector.
  *
  * Author:
  *     Mark Probst <mark.probst@gmail.com>
@@ -134,6 +135,20 @@ typedef struct {
 
 #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))
 
@@ -230,6 +245,7 @@ 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;
@@ -279,6 +295,7 @@ 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 ?      \
@@ -677,6 +694,55 @@ 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.
@@ -1079,12 +1145,27 @@ 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)
 {
@@ -1098,6 +1179,7 @@ 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
@@ -1148,6 +1230,12 @@ 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
@@ -1156,6 +1244,15 @@ 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
@@ -1208,6 +1305,15 @@ 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)
 {
@@ -1250,6 +1356,12 @@ major_copy_or_mark_object_concurrent_finish_canonical (GCObject **ptr, SgenGrayQ
        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)
 {
@@ -1428,6 +1540,21 @@ 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
 sweep_start (void)
 {
@@ -1443,6 +1570,8 @@ sweep_start (void)
                for (j = 0; j < num_block_obj_sizes; ++j)
                        free_blocks [j] = NULL;
        }
+
+       sgen_workers_foreach (sgen_worker_clear_free_block_lists);
 }
 
 static void sweep_finish (void);
@@ -2436,7 +2565,7 @@ 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;
@@ -2450,8 +2579,10 @@ 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);
@@ -2564,6 +2695,22 @@ post_param_init (SgenMajorCollector *collector)
        collector->needs_thread_pool = concurrent_mark || concurrent_sweep;
 }
 
+/* We are guaranteed to be called by the worker in question */
+static void
+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);
+
+       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
 sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurrent, gboolean is_parallel)
 {
@@ -2623,6 +2770,7 @@ 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;
@@ -2661,6 +2809,7 @@ sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurr
 
        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;
@@ -2682,8 +2831,15 @@ sgen_marksweep_init_internal (SgenMajorCollector *collector, gboolean is_concurr
                        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;
 
-                       /* FIXME use parallel obj ops */
-                       collector->major_ops_conc_par_finish = collector->major_ops_concurrent_finish;
+                       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;
+
+                       collector->worker_init_cb = sgen_worker_init_callback;
+
+                       mono_native_tls_alloc (&worker_block_free_list_key, NULL);
                }
        }