1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/g1BarrierSet.hpp" 27 #include "gc/g1/g1ConcurrentRefine.hpp" 28 #include "gc/g1/g1ConcurrentRefineThread.hpp" 29 #include "logging/log.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "runtime/java.hpp" 32 #include "runtime/thread.hpp" 33 #include "utilities/debug.hpp" 34 #include "utilities/globalDefinitions.hpp" 35 #include "utilities/pair.hpp" 36 #include <math.h> 37 38 G1ConcurrentRefineThread* G1ConcurrentRefineThreadControl::create_refinement_thread(uint worker_id, bool initializing) { 39 G1ConcurrentRefineThread* result = NULL; 40 if (initializing || !InjectGCWorkerCreationFailure) { 41 result = new G1ConcurrentRefineThread(_cr, worker_id); 42 } 43 if (result == NULL || result->osthread() == NULL) { 44 log_warning(gc)("Failed to create refinement thread %u, no more %s", 45 worker_id, 46 result == NULL ? "memory" : "OS threads"); 47 } 48 return result; 49 } 50 51 G1ConcurrentRefineThreadControl::G1ConcurrentRefineThreadControl() : 52 _cr(NULL), 53 _threads(NULL), 54 _num_max_threads(0) 55 { 56 } 57 58 G1ConcurrentRefineThreadControl::~G1ConcurrentRefineThreadControl() { 59 for (uint i = 0; i < _num_max_threads; i++) { 60 G1ConcurrentRefineThread* t = _threads[i]; 61 if (t != NULL) { 62 delete t; 63 } 64 } 65 FREE_C_HEAP_ARRAY(G1ConcurrentRefineThread*, _threads); 66 } 67 68 jint G1ConcurrentRefineThreadControl::initialize(G1ConcurrentRefine* cr, uint num_max_threads) { 69 assert(cr != NULL, "G1ConcurrentRefine must not be NULL"); 70 _cr = cr; 71 _num_max_threads = num_max_threads; 72 73 _threads = NEW_C_HEAP_ARRAY_RETURN_NULL(G1ConcurrentRefineThread*, num_max_threads, mtGC); 74 if (_threads == NULL) { 75 vm_shutdown_during_initialization("Could not allocate thread holder array."); 76 return JNI_ENOMEM; 77 } 78 79 for (uint i = 0; i < num_max_threads; i++) { 80 if (UseDynamicNumberOfGCThreads && i != 0 /* Always start first thread. */) { 81 _threads[i] = NULL; 82 } else { 83 _threads[i] = create_refinement_thread(i, true); 84 if (_threads[i] == NULL) { 85 vm_shutdown_during_initialization("Could not allocate refinement threads."); 86 return JNI_ENOMEM; 87 } 88 } 89 } 90 return JNI_OK; 91 } 92 93 void G1ConcurrentRefineThreadControl::maybe_activate_next(uint cur_worker_id) { 94 assert(cur_worker_id < _num_max_threads, 95 "Activating another thread from %u not allowed since there can be at most %u", 96 cur_worker_id, _num_max_threads); 97 if (cur_worker_id == (_num_max_threads - 1)) { 98 // Already the last thread, there is no more thread to activate. 99 return; 100 } 101 102 uint worker_id = cur_worker_id + 1; 103 G1ConcurrentRefineThread* thread_to_activate = _threads[worker_id]; 104 if (thread_to_activate == NULL) { 105 // Still need to create the thread... 106 _threads[worker_id] = create_refinement_thread(worker_id, false); 107 thread_to_activate = _threads[worker_id]; 108 } 109 if (thread_to_activate != NULL && !thread_to_activate->is_active()) { 110 thread_to_activate->activate(); 111 } 112 } 113 114 void G1ConcurrentRefineThreadControl::print_on(outputStream* st) const { 115 for (uint i = 0; i < _num_max_threads; ++i) { 116 if (_threads[i] != NULL) { 117 _threads[i]->print_on(st); 118 st->cr(); 119 } 120 } 121 } 122 123 void G1ConcurrentRefineThreadControl::worker_threads_do(ThreadClosure* tc) { 124 for (uint i = 0; i < _num_max_threads; i++) { 125 if (_threads[i] != NULL) { 126 tc->do_thread(_threads[i]); 127 } 128 } 129 } 130 131 void G1ConcurrentRefineThreadControl::stop() { 132 for (uint i = 0; i < _num_max_threads; i++) { 133 if (_threads[i] != NULL) { 134 _threads[i]->stop(); 135 } 136 } 137 } 138 139 // Arbitrary but large limits, to simplify some of the zone calculations. 140 // The general idea is to allow expressions like 141 // MIN2(x OP y, max_XXX_zone) 142 // without needing to check for overflow in "x OP y", because the 143 // ranges for x and y have been restricted. 144 STATIC_ASSERT(sizeof(LP64_ONLY(jint) NOT_LP64(jshort)) <= (sizeof(size_t)/2)); 145 const size_t max_yellow_zone = LP64_ONLY(max_jint) NOT_LP64(max_jshort); 146 const size_t max_green_zone = max_yellow_zone / 2; 147 const size_t max_red_zone = INT_MAX; // For dcqs.set_max_completed_buffers. 148 STATIC_ASSERT(max_yellow_zone <= max_red_zone); 149 150 // Range check assertions for green zone values. 151 #define assert_zone_constraints_g(green) \ 152 do { \ 153 size_t azc_g_green = (green); \ 154 assert(azc_g_green <= max_green_zone, \ 155 "green exceeds max: " SIZE_FORMAT, azc_g_green); \ 156 } while (0) 157 158 // Range check assertions for green and yellow zone values. 159 #define assert_zone_constraints_gy(green, yellow) \ 160 do { \ 161 size_t azc_gy_green = (green); \ 162 size_t azc_gy_yellow = (yellow); \ 163 assert_zone_constraints_g(azc_gy_green); \ 164 assert(azc_gy_yellow <= max_yellow_zone, \ 165 "yellow exceeds max: " SIZE_FORMAT, azc_gy_yellow); \ 166 assert(azc_gy_green <= azc_gy_yellow, \ 167 "green (" SIZE_FORMAT ") exceeds yellow (" SIZE_FORMAT ")", \ 168 azc_gy_green, azc_gy_yellow); \ 169 } while (0) 170 171 // Range check assertions for green, yellow, and red zone values. 172 #define assert_zone_constraints_gyr(green, yellow, red) \ 173 do { \ 174 size_t azc_gyr_green = (green); \ 175 size_t azc_gyr_yellow = (yellow); \ 176 size_t azc_gyr_red = (red); \ 177 assert_zone_constraints_gy(azc_gyr_green, azc_gyr_yellow); \ 178 assert(azc_gyr_red <= max_red_zone, \ 179 "red exceeds max: " SIZE_FORMAT, azc_gyr_red); \ 180 assert(azc_gyr_yellow <= azc_gyr_red, \ 181 "yellow (" SIZE_FORMAT ") exceeds red (" SIZE_FORMAT ")", \ 182 azc_gyr_yellow, azc_gyr_red); \ 183 } while (0) 184 185 // Logging tag sequence for refinement control updates. 186 #define CTRL_TAGS gc, ergo, refine 187 188 // For logging zone values, ensuring consistency of level and tags. 189 #define LOG_ZONES(...) log_debug( CTRL_TAGS )(__VA_ARGS__) 190 191 // Package for pair of refinement thread activation and deactivation 192 // thresholds. The activation and deactivation levels are resp. the first 193 // and second values of the pair. 194 typedef Pair<size_t, size_t> Thresholds; 195 inline size_t activation_level(const Thresholds& t) { return t.first; } 196 inline size_t deactivation_level(const Thresholds& t) { return t.second; } 197 198 static Thresholds calc_thresholds(size_t green_zone, 199 size_t yellow_zone, 200 uint worker_i) { 201 double yellow_size = yellow_zone - green_zone; 202 double step = yellow_size / G1ConcurrentRefine::max_num_threads(); 203 if (worker_i == 0) { 204 // Potentially activate worker 0 more aggressively, to keep 205 // available buffers near green_zone value. When yellow_size is 206 // large we don't want to allow a full step to accumulate before 207 // doing any processing, as that might lead to significantly more 208 // than green_zone buffers to be processed by update_rs. 209 step = MIN2(step, ParallelGCThreads / 2.0); 210 } 211 size_t activate_offset = static_cast<size_t>(ceil(step * (worker_i + 1))); 212 size_t deactivate_offset = static_cast<size_t>(floor(step * worker_i)); 213 return Thresholds(green_zone + activate_offset, 214 green_zone + deactivate_offset); 215 } 216 217 G1ConcurrentRefine::G1ConcurrentRefine(size_t green_zone, 218 size_t yellow_zone, 219 size_t red_zone, 220 size_t min_yellow_zone_size) : 221 _thread_control(), 222 _green_zone(green_zone), 223 _yellow_zone(yellow_zone), 224 _red_zone(red_zone), 225 _min_yellow_zone_size(min_yellow_zone_size) 226 { 227 assert_zone_constraints_gyr(green_zone, yellow_zone, red_zone); 228 } 229 230 jint G1ConcurrentRefine::initialize() { 231 return _thread_control.initialize(this, max_num_threads()); 232 } 233 234 static size_t calc_min_yellow_zone_size() { 235 size_t step = G1ConcRefinementThresholdStep; 236 uint n_workers = G1ConcurrentRefine::max_num_threads(); 237 if ((max_yellow_zone / step) < n_workers) { 238 return max_yellow_zone; 239 } else { 240 return step * n_workers; 241 } 242 } 243 244 static size_t calc_init_green_zone() { 245 size_t green = G1ConcRefinementGreenZone; 246 if (FLAG_IS_DEFAULT(G1ConcRefinementGreenZone)) { 247 green = ParallelGCThreads; 248 } 249 return MIN2(green, max_green_zone); 250 } 251 252 static size_t calc_init_yellow_zone(size_t green, size_t min_size) { 253 size_t config = G1ConcRefinementYellowZone; 254 size_t size = 0; 255 if (FLAG_IS_DEFAULT(G1ConcRefinementYellowZone)) { 256 size = green * 2; 257 } else if (green < config) { 258 size = config - green; 259 } 260 size = MAX2(size, min_size); 261 size = MIN2(size, max_yellow_zone); 262 return MIN2(green + size, max_yellow_zone); 263 } 264 265 static size_t calc_init_red_zone(size_t green, size_t yellow) { 266 size_t size = yellow - green; 267 if (!FLAG_IS_DEFAULT(G1ConcRefinementRedZone)) { 268 size_t config = G1ConcRefinementRedZone; 269 if (yellow < config) { 270 size = MAX2(size, config - yellow); 271 } 272 } 273 return MIN2(yellow + size, max_red_zone); 274 } 275 276 G1ConcurrentRefine* G1ConcurrentRefine::create(jint* ecode) { 277 size_t min_yellow_zone_size = calc_min_yellow_zone_size(); 278 size_t green_zone = calc_init_green_zone(); 279 size_t yellow_zone = calc_init_yellow_zone(green_zone, min_yellow_zone_size); 280 size_t red_zone = calc_init_red_zone(green_zone, yellow_zone); 281 282 LOG_ZONES("Initial Refinement Zones: " 283 "green: " SIZE_FORMAT ", " 284 "yellow: " SIZE_FORMAT ", " 285 "red: " SIZE_FORMAT ", " 286 "min yellow size: " SIZE_FORMAT, 287 green_zone, yellow_zone, red_zone, min_yellow_zone_size); 288 289 G1ConcurrentRefine* cr = new G1ConcurrentRefine(green_zone, 290 yellow_zone, 291 red_zone, 292 min_yellow_zone_size); 293 294 if (cr == NULL) { 295 *ecode = JNI_ENOMEM; 296 vm_shutdown_during_initialization("Could not create G1ConcurrentRefine"); 297 return NULL; 298 } 299 300 *ecode = cr->initialize(); 301 return cr; 302 } 303 304 void G1ConcurrentRefine::stop() { 305 _thread_control.stop(); 306 } 307 308 G1ConcurrentRefine::~G1ConcurrentRefine() { 309 } 310 311 void G1ConcurrentRefine::threads_do(ThreadClosure *tc) { 312 _thread_control.worker_threads_do(tc); 313 } 314 315 uint G1ConcurrentRefine::max_num_threads() { 316 return G1ConcRefinementThreads; 317 } 318 319 void G1ConcurrentRefine::print_threads_on(outputStream* st) const { 320 _thread_control.print_on(st); 321 } 322 323 static size_t calc_new_green_zone(size_t green, 324 double update_rs_time, 325 size_t update_rs_processed_buffers, 326 double goal_ms) { 327 // Adjust green zone based on whether we're meeting the time goal. 328 // Limit to max_green_zone. 329 const double inc_k = 1.1, dec_k = 0.9; 330 if (update_rs_time > goal_ms) { 331 if (green > 0) { 332 green = static_cast<size_t>(green * dec_k); 333 } 334 } else if (update_rs_time < goal_ms && 335 update_rs_processed_buffers > green) { 336 green = static_cast<size_t>(MAX2(green * inc_k, green + 1.0)); 337 green = MIN2(green, max_green_zone); 338 } 339 return green; 340 } 341 342 static size_t calc_new_yellow_zone(size_t green, size_t min_yellow_size) { 343 size_t size = green * 2; 344 size = MAX2(size, min_yellow_size); 345 return MIN2(green + size, max_yellow_zone); 346 } 347 348 static size_t calc_new_red_zone(size_t green, size_t yellow) { 349 return MIN2(yellow + (yellow - green), max_red_zone); 350 } 351 352 void G1ConcurrentRefine::update_zones(double update_rs_time, 353 size_t update_rs_processed_buffers, 354 double goal_ms) { 355 log_trace( CTRL_TAGS )("Updating Refinement Zones: " 356 "update_rs time: %.3fms, " 357 "update_rs buffers: " SIZE_FORMAT ", " 358 "update_rs goal time: %.3fms", 359 update_rs_time, 360 update_rs_processed_buffers, 361 goal_ms); 362 363 _green_zone = calc_new_green_zone(_green_zone, 364 update_rs_time, 365 update_rs_processed_buffers, 366 goal_ms); 367 _yellow_zone = calc_new_yellow_zone(_green_zone, _min_yellow_zone_size); 368 _red_zone = calc_new_red_zone(_green_zone, _yellow_zone); 369 370 assert_zone_constraints_gyr(_green_zone, _yellow_zone, _red_zone); 371 LOG_ZONES("Updated Refinement Zones: " 372 "green: " SIZE_FORMAT ", " 373 "yellow: " SIZE_FORMAT ", " 374 "red: " SIZE_FORMAT, 375 _green_zone, _yellow_zone, _red_zone); 376 } 377 378 void G1ConcurrentRefine::adjust(double update_rs_time, 379 size_t update_rs_processed_buffers, 380 double goal_ms) { 381 DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set(); 382 383 if (G1UseAdaptiveConcRefinement) { 384 update_zones(update_rs_time, update_rs_processed_buffers, goal_ms); 385 386 // Change the barrier params 387 if (max_num_threads() == 0) { 388 // Disable dcqs notification when there are no threads to notify. 389 dcqs.set_process_completed_buffers_threshold(DirtyCardQueueSet::ProcessCompletedBuffersThresholdNever); 390 } else { 391 // Worker 0 is the primary; wakeup is via dcqs notification. 392 STATIC_ASSERT(max_yellow_zone <= INT_MAX); 393 size_t activate = activation_threshold(0); 394 dcqs.set_process_completed_buffers_threshold(activate); 395 } 396 dcqs.set_max_completed_buffers(red_zone()); 397 } 398 399 size_t curr_queue_size = dcqs.completed_buffers_num(); 400 if ((dcqs.max_completed_buffers() > 0) && 401 (curr_queue_size >= yellow_zone())) { 402 dcqs.set_completed_buffers_padding(curr_queue_size); 403 } else { 404 dcqs.set_completed_buffers_padding(0); 405 } 406 dcqs.notify_if_necessary(); 407 } 408 409 size_t G1ConcurrentRefine::activation_threshold(uint worker_id) const { 410 Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, worker_id); 411 return activation_level(thresholds); 412 } 413 414 size_t G1ConcurrentRefine::deactivation_threshold(uint worker_id) const { 415 Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, worker_id); 416 return deactivation_level(thresholds); 417 } 418 419 uint G1ConcurrentRefine::worker_id_offset() { 420 return DirtyCardQueueSet::num_par_ids(); 421 } 422 423 void G1ConcurrentRefine::maybe_activate_more_threads(uint worker_id, size_t num_cur_buffers) { 424 if (num_cur_buffers > activation_threshold(worker_id + 1)) { 425 _thread_control.maybe_activate_next(worker_id); 426 } 427 } 428 429 bool G1ConcurrentRefine::do_refinement_step(uint worker_id) { 430 DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set(); 431 432 size_t curr_buffer_num = dcqs.completed_buffers_num(); 433 // If the number of the buffers falls down into the yellow zone, 434 // that means that the transition period after the evacuation pause has ended. 435 // Since the value written to the DCQS is the same for all threads, there is no 436 // need to synchronize. 437 if (dcqs.completed_buffers_padding() > 0 && curr_buffer_num <= yellow_zone()) { 438 dcqs.set_completed_buffers_padding(0); 439 } 440 441 maybe_activate_more_threads(worker_id, curr_buffer_num); 442 443 // Process the next buffer, if there are enough left. 444 return dcqs.refine_completed_buffer_concurrently(worker_id + worker_id_offset(), 445 deactivation_threshold(worker_id)); 446 }