mono/metadata/sgen-workers.c

   1 /*
   2  * Copyright 2001-2003 Ximian, Inc
   3  * Copyright 2003-2010 Novell, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining
   6  * a copy of this software and associated documentation files (the
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sublicense, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be
  14  * included in all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  17  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  18  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  19  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
  20  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  21  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  22  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 typedef struct _WorkerData WorkerData;
  26 struct _WorkerData {
  27         pthread_t thread;
  28         MonoSemType start_worker_sem;
  29         gboolean is_working;
  30         GrayQueue private_gray_queue; /* only read/written by worker thread */
  31         int shared_buffer_increment;
  32         int shared_buffer_index;
  33 };
  34
  35 static int workers_num;
  36 static WorkerData *workers_data;
  37 static WorkerData workers_gc_thread_data;
  38
  39 static int workers_num_working;
  40
  41 static GrayQueue workers_distribute_gray_queue;
  42
  43 #define WORKERS_DISTRIBUTE_GRAY_QUEUE (major.is_parallel ? &workers_distribute_gray_queue : &gray_queue)
  44
  45 /*
  46  * Must be a power of 2.  It seems that larger values don't help much.
  47  * The main reason to make this larger would be to sustain a bigger
  48  * number of worker threads.
  49  */
  50 #define WORKERS_SHARED_BUFFER_SIZE      16
  51 static GrayQueueSection *workers_shared_buffer [WORKERS_SHARED_BUFFER_SIZE];
  52 static int workers_shared_buffer_used;
  53
  54 static const int workers_primes [] = { 3, 5, 7, 11, 13, 17, 23, 29 };
  55
  56 static MonoSemType workers_done_sem;
  57
  58 static long long stat_shared_buffer_insert_tries;
  59 static long long stat_shared_buffer_insert_full;
  60 static long long stat_shared_buffer_insert_iterations;
  61 static long long stat_shared_buffer_insert_failures;
  62 static long long stat_shared_buffer_remove_tries;
  63 static long long stat_shared_buffer_remove_iterations;
  64 static long long stat_shared_buffer_remove_empty;
  65
  66 static void
  67 workers_gray_queue_share_redirect (GrayQueue *queue)
  68 {
  69         GrayQueueSection *section;
  70         WorkerData *data = queue->alloc_prepare_data;
  71         int increment = data->shared_buffer_increment;
  72
  73         while ((section = gray_object_dequeue_section (queue))) {
  74                 int i, index;
  75
  76                 HEAVY_STAT (++stat_shared_buffer_insert_tries);
  77
  78                 if (workers_shared_buffer_used == WORKERS_SHARED_BUFFER_SIZE) {
  79                         HEAVY_STAT (++stat_shared_buffer_insert_full);
  80                         gray_object_enqueue_section (queue, section);
  81                         return;
  82                 }
  83
  84                 index = data->shared_buffer_index;
  85                 for (i = 0; i < WORKERS_SHARED_BUFFER_SIZE; ++i) {
  86                         GrayQueueSection *old = workers_shared_buffer [index];
  87                         HEAVY_STAT (++stat_shared_buffer_insert_iterations);
  88                         if (!old) {
  89                                 if (SGEN_CAS_PTR ((void**)&workers_shared_buffer [index], section, NULL) == NULL) {
  90                                         SGEN_ATOMIC_ADD (workers_shared_buffer_used, 1);
  91                                         //g_print ("thread %d put section %d\n", data - workers_data, index);
  92                                         break;
  93                                 }
  94                         }
  95                         index = (index + increment) & (WORKERS_SHARED_BUFFER_SIZE - 1);
  96                 }
  97                 data->shared_buffer_index = index;
  98
  99                 if (i == WORKERS_SHARED_BUFFER_SIZE) {
 100                         /* unsuccessful */
 101                         HEAVY_STAT (++stat_shared_buffer_insert_failures);
 102                         gray_object_enqueue_section (queue, section);
 103                         return;
 104                 }
 105         }
 106 }
 107
 108 static gboolean
 109 workers_get_work (WorkerData *data)
 110 {
 111         int i, index;
 112         int increment = data->shared_buffer_increment;
 113
 114         HEAVY_STAT (++stat_shared_buffer_remove_tries);
 115
 116         index = data->shared_buffer_index;
 117         for (i = 0; i < WORKERS_SHARED_BUFFER_SIZE; ++i) {
 118                 GrayQueueSection *section;
 119
 120                 HEAVY_STAT (++stat_shared_buffer_remove_iterations);
 121
 122                 do {
 123                         section = workers_shared_buffer [index];
 124                         if (!section)
 125                                 break;
 126                 } while (SGEN_CAS_PTR ((void**)&workers_shared_buffer [index], NULL, section) != section);
 127
 128                 if (section) {
 129                         SGEN_ATOMIC_ADD (workers_shared_buffer_used, -1);
 130                         gray_object_enqueue_section (&data->private_gray_queue, section);
 131                         data->shared_buffer_index = index;
 132                         //g_print ("thread %d popped section %d\n", data - workers_data, index);
 133                         return TRUE;
 134                 }
 135
 136                 index = (index + increment) & (WORKERS_SHARED_BUFFER_SIZE - 1);
 137         }
 138
 139         HEAVY_STAT (++stat_shared_buffer_remove_empty);
 140
 141         data->shared_buffer_index = index;
 142         return FALSE;
 143 }
 144
 145 /* returns the new value */
 146 static int
 147 workers_change_num_working (int delta)
 148 {
 149         int old, new;
 150
 151         if (!major.is_parallel)
 152                 return -1;
 153
 154         do {
 155                 old = workers_num_working;
 156                 new = old + delta;
 157         } while (InterlockedCompareExchange (&workers_num_working, new, old) != old);
 158         return new;
 159 }
 160
 161 static void*
 162 workers_thread_func (void *data_untyped)
 163 {
 164         WorkerData *data = data_untyped;
 165         SgenInternalAllocator allocator;
 166
 167         memset (&allocator, 0, sizeof (allocator));
 168
 169         gray_object_queue_init_with_alloc_prepare (&data->private_gray_queue, &allocator,
 170                         workers_gray_queue_share_redirect, data);
 171
 172         for (;;) {
 173                 //g_print ("worker waiting for start %d\n", data->start_worker_sem);
 174
 175                 MONO_SEM_WAIT (&data->start_worker_sem);
 176
 177                 //g_print ("worker starting\n");
 178
 179                 for (;;) {
 180                         do {
 181                                 drain_gray_stack (&data->private_gray_queue);
 182                         } while (workers_get_work (data));
 183
 184                         /*
 185                          * FIXME: This might never terminate with
 186                          * multiple threads!
 187                          */
 188
 189                         if (workers_change_num_working (-1) == 0)
 190                                 break;
 191
 192                         /* we weren't the last one working */
 193                         //g_print ("sleeping\n");
 194                         usleep (5000);
 195                         workers_change_num_working (1);
 196                 }
 197
 198                 gray_object_queue_init (&data->private_gray_queue, &allocator);
 199
 200                 MONO_SEM_POST (&workers_done_sem);
 201
 202                 //g_print ("worker done\n");
 203         }
 204
 205         /* dummy return to make compilers happy */
 206         return NULL;
 207 }
 208
 209 static void
 210 workers_distribute_gray_queue_sections (void)
 211 {
 212         if (!major.is_parallel)
 213                 return;
 214
 215         workers_gray_queue_share_redirect (&workers_distribute_gray_queue);
 216 }
 217
 218 static void
 219 workers_init (int num_workers)
 220 {
 221         int i;
 222
 223         if (!major.is_parallel)
 224                 return;
 225
 226         //g_print ("initing %d workers\n", num_workers);
 227
 228         workers_num = num_workers;
 229         workers_data = mono_sgen_alloc_internal_dynamic (sizeof (WorkerData) * num_workers, INTERNAL_MEM_WORKER_DATA);
 230         MONO_SEM_INIT (&workers_done_sem, 0);
 231         workers_gc_thread_data.shared_buffer_increment = 1;
 232         workers_gc_thread_data.shared_buffer_index = 0;
 233         gray_object_queue_init_with_alloc_prepare (&workers_distribute_gray_queue, mono_sgen_get_unmanaged_allocator (),
 234                         workers_gray_queue_share_redirect, &workers_gc_thread_data);
 235
 236         g_assert (num_workers <= sizeof (workers_primes) / sizeof (workers_primes [0]));
 237         for (i = 0; i < workers_num; ++i) {
 238                 workers_data [i].shared_buffer_increment = workers_primes [i];
 239                 workers_data [i].shared_buffer_index = 0;
 240         }
 241
 242         mono_counters_register ("Shared buffer insert tries", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_shared_buffer_insert_tries);
 243         mono_counters_register ("Shared buffer insert full", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_shared_buffer_insert_full);
 244         mono_counters_register ("Shared buffer insert iterations", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_shared_buffer_insert_iterations);
 245         mono_counters_register ("Shared buffer insert failures", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_shared_buffer_insert_failures);
 246         mono_counters_register ("Shared buffer remove tries", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_shared_buffer_remove_tries);
 247         mono_counters_register ("Shared buffer remove iterations", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_shared_buffer_remove_iterations);
 248         mono_counters_register ("Shared buffer remove empty", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_shared_buffer_remove_empty);
 249 }
 250
 251 /* only the GC thread is allowed to start and join workers */
 252
 253 static void
 254 workers_start_worker (int index)
 255 {
 256         g_assert (index >= 0 && index < workers_num);
 257
 258         if (workers_data [index].is_working)
 259                 return;
 260
 261         if (!workers_data [index].thread) {
 262                 //g_print ("initing thread %d\n", index);
 263                 MONO_SEM_INIT (&workers_data [index].start_worker_sem, 0);
 264                 pthread_create (&workers_data [index].thread, NULL, workers_thread_func, &workers_data [index]);
 265         }
 266
 267         workers_data [index].is_working = TRUE;
 268         MONO_SEM_POST (&workers_data [index].start_worker_sem);
 269         //g_print ("posted thread start %d %d\n", index, workers_data [index].start_worker_sem);
 270 }
 271
 272 static void
 273 workers_start_all_workers (int num_additional_workers)
 274 {
 275         int i;
 276
 277         if (!major.is_parallel)
 278                 return;
 279
 280         g_assert (workers_num_working == 0);
 281         workers_num_working = workers_num + num_additional_workers;
 282
 283         for (i = 0; i < workers_num; ++i)
 284                 workers_start_worker (i);
 285 }
 286
 287 static void
 288 workers_join (void)
 289 {
 290         int i;
 291
 292         if (!major.is_parallel)
 293                 return;
 294
 295         //g_print ("joining\n");
 296         for (i = 0; i < workers_num; ++i) {
 297                 if (workers_data [i].is_working)
 298                         MONO_SEM_WAIT (&workers_done_sem);
 299         }
 300         for (i = 0; i < workers_num; ++i)
 301                 workers_data [i].is_working = FALSE;
 302         //g_print ("joined\n");
 303
 304         g_assert (workers_num_working == 0);
 305         g_assert (workers_shared_buffer_used == 0);
 306
 307         for (i = 0; i < WORKERS_SHARED_BUFFER_SIZE; ++i)
 308                 g_assert (!workers_shared_buffer [i]);
 309 }
 310
 311 gboolean
 312 mono_sgen_is_worker_thread (pthread_t thread)
 313 {
 314         int i;
 315
 316         if (!major.is_parallel)
 317                 return FALSE;
 318
 319         for (i = 0; i < workers_num; ++i) {
 320                 if (workers_data [i].thread == thread)
 321                         return TRUE;
 322         }
 323         return FALSE;
 324 }