1 /*
   2  * Copyright (c) 2001, 2017, 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/g1ConcurrentRefine.hpp"
  27 #include "gc/g1/g1ConcurrentRefineThread.hpp"
  28 #include "logging/log.hpp"
  29 #include "runtime/java.hpp"
  30 #include "runtime/thread.hpp"
  31 #include "utilities/debug.hpp"
  32 #include "utilities/globalDefinitions.hpp"
  33 #include "utilities/pair.hpp"
  34 #include <math.h>
  35 
  36 G1ConcurrentRefineThread* G1ConcurrentRefineThreadControl::create_refinement_thread(uint worker_id, bool initializing) {
  37   G1ConcurrentRefineThread* result = NULL;
  38   if (initializing || !InjectGCWorkerCreationFailure) {
  39     result = new G1ConcurrentRefineThread(_cr, worker_id);
  40   }
  41   if (result == NULL || result->osthread() == NULL) {
  42     log_warning(gc)("Failed to create refinement thread %u, no more %s",
  43                     worker_id,
  44                     result == NULL ? "memory" : "OS threads");
  45   }
  46   return result;
  47 }
  48 
  49 G1ConcurrentRefineThreadControl::G1ConcurrentRefineThreadControl() :
  50   _cr(NULL),
  51   _threads(NULL),
  52   _num_max_threads(0)
  53 {
  54 }
  55 
  56 G1ConcurrentRefineThreadControl::~G1ConcurrentRefineThreadControl() {
  57   for (uint i = 0; i < _num_max_threads; i++) {
  58     G1ConcurrentRefineThread* t = _threads[i];
  59     if (t != NULL) {
  60       delete t;
  61     }
  62   }
  63   FREE_C_HEAP_ARRAY(G1ConcurrentRefineThread*, _threads);
  64 }
  65 
  66 jint G1ConcurrentRefineThreadControl::initialize(G1ConcurrentRefine* cr, uint num_max_threads) {
  67   assert(cr != NULL, "G1ConcurrentRefine must not be NULL");
  68   _cr = cr;
  69   _num_max_threads = num_max_threads;
  70 
  71   _threads = NEW_C_HEAP_ARRAY_RETURN_NULL(G1ConcurrentRefineThread*, num_max_threads, mtGC);
  72   if (_threads == NULL) {
  73     vm_shutdown_during_initialization("Could not allocate thread holder array.");
  74     return JNI_ENOMEM;
  75   }
  76 
  77   for (uint i = 0; i < num_max_threads; i++) {
  78     if (UseDynamicNumberOfGCThreads && i != 0 /* Always start first thread. */) {
  79       _threads[i] = NULL;
  80     } else {
  81       _threads[i] = create_refinement_thread(i, true);
  82       if (_threads[i] == NULL) {
  83         vm_shutdown_during_initialization("Could not allocate refinement threads.");
  84         return JNI_ENOMEM;
  85       }
  86     }
  87   }
  88   return JNI_OK;
  89 }
  90 
  91 void G1ConcurrentRefineThreadControl::maybe_activate_next(uint cur_worker_id) {
  92   assert(cur_worker_id < _num_max_threads,
  93          "Activating another thread from %u not allowed since there can be at most %u",
  94          cur_worker_id, _num_max_threads);
  95   if (cur_worker_id == (_num_max_threads - 1)) {
  96     // Already the last thread, there is no more thread to activate.
  97     return;
  98   }
  99 
 100   uint worker_id = cur_worker_id + 1;
 101   G1ConcurrentRefineThread* thread_to_activate = _threads[worker_id];
 102   if (thread_to_activate == NULL) {
 103     // Still need to create the thread...
 104     _threads[worker_id] = create_refinement_thread(worker_id, false);
 105     thread_to_activate = _threads[worker_id];
 106   }
 107   if (thread_to_activate != NULL) {
 108     thread_to_activate->activate();
 109   }
 110 }
 111 
 112 void G1ConcurrentRefineThreadControl::print_on(outputStream* st) const {
 113   for (uint i = 0; i < _num_max_threads; ++i) {
 114     if (_threads[i] != NULL) {
 115       _threads[i]->print_on(st);
 116       st->cr();
 117     }
 118   }
 119 }
 120 
 121 void G1ConcurrentRefineThreadControl::worker_threads_do(ThreadClosure* tc) {
 122   for (uint i = 0; i < _num_max_threads; i++) {
 123     if (_threads[i] != NULL) {
 124       tc->do_thread(_threads[i]);
 125     }
 126   }
 127 }
 128 
 129 void G1ConcurrentRefineThreadControl::stop() {
 130   for (uint i = 0; i < _num_max_threads; i++) {
 131     if (_threads[i] != NULL) {
 132       _threads[i]->stop();
 133     }
 134   }
 135 }
 136 
 137 // Arbitrary but large limits, to simplify some of the zone calculations.
 138 // The general idea is to allow expressions like
 139 //   MIN2(x OP y, max_XXX_zone)
 140 // without needing to check for overflow in "x OP y", because the
 141 // ranges for x and y have been restricted.
 142 STATIC_ASSERT(sizeof(LP64_ONLY(jint) NOT_LP64(jshort)) <= (sizeof(size_t)/2));
 143 const size_t max_yellow_zone = LP64_ONLY(max_jint) NOT_LP64(max_jshort);
 144 const size_t max_green_zone = max_yellow_zone / 2;
 145 const size_t max_red_zone = INT_MAX; // For dcqs.set_max_completed_queue.
 146 STATIC_ASSERT(max_yellow_zone <= max_red_zone);
 147 
 148 // Range check assertions for green zone values.
 149 #define assert_zone_constraints_g(green)                        \
 150   do {                                                          \
 151     size_t azc_g_green = (green);                               \
 152     assert(azc_g_green <= max_green_zone,                       \
 153            "green exceeds max: " SIZE_FORMAT, azc_g_green);     \
 154   } while (0)
 155 
 156 // Range check assertions for green and yellow zone values.
 157 #define assert_zone_constraints_gy(green, yellow)                       \
 158   do {                                                                  \
 159     size_t azc_gy_green = (green);                                      \
 160     size_t azc_gy_yellow = (yellow);                                    \
 161     assert_zone_constraints_g(azc_gy_green);                            \
 162     assert(azc_gy_yellow <= max_yellow_zone,                            \
 163            "yellow exceeds max: " SIZE_FORMAT, azc_gy_yellow);          \
 164     assert(azc_gy_green <= azc_gy_yellow,                               \
 165            "green (" SIZE_FORMAT ") exceeds yellow (" SIZE_FORMAT ")",  \
 166            azc_gy_green, azc_gy_yellow);                                \
 167   } while (0)
 168 
 169 // Range check assertions for green, yellow, and red zone values.
 170 #define assert_zone_constraints_gyr(green, yellow, red)                 \
 171   do {                                                                  \
 172     size_t azc_gyr_green = (green);                                     \
 173     size_t azc_gyr_yellow = (yellow);                                   \
 174     size_t azc_gyr_red = (red);                                         \
 175     assert_zone_constraints_gy(azc_gyr_green, azc_gyr_yellow);          \
 176     assert(azc_gyr_red <= max_red_zone,                                 \
 177            "red exceeds max: " SIZE_FORMAT, azc_gyr_red);               \
 178     assert(azc_gyr_yellow <= azc_gyr_red,                               \
 179            "yellow (" SIZE_FORMAT ") exceeds red (" SIZE_FORMAT ")",    \
 180            azc_gyr_yellow, azc_gyr_red);                                \
 181   } while (0)
 182 
 183 // Logging tag sequence for refinement control updates.
 184 #define CTRL_TAGS gc, ergo, refine
 185 
 186 // For logging zone values, ensuring consistency of level and tags.
 187 #define LOG_ZONES(...) log_debug( CTRL_TAGS )(__VA_ARGS__)
 188 
 189 // Package for pair of refinement thread activation and deactivation
 190 // thresholds.  The activation and deactivation levels are resp. the first
 191 // and second values of the pair.
 192 typedef Pair<size_t, size_t> Thresholds;
 193 inline size_t activation_level(const Thresholds& t) { return t.first; }
 194 inline size_t deactivation_level(const Thresholds& t) { return t.second; }
 195 
 196 static Thresholds calc_thresholds(size_t green_zone,
 197                                   size_t yellow_zone,
 198                                   uint worker_i) {
 199   double yellow_size = yellow_zone - green_zone;
 200   double step = yellow_size / G1ConcurrentRefine::max_num_threads();
 201   if (worker_i == 0) {
 202     // Potentially activate worker 0 more aggressively, to keep
 203     // available buffers near green_zone value.  When yellow_size is
 204     // large we don't want to allow a full step to accumulate before
 205     // doing any processing, as that might lead to significantly more
 206     // than green_zone buffers to be processed by update_rs.
 207     step = MIN2(step, ParallelGCThreads / 2.0);
 208   }
 209   size_t activate_offset = static_cast<size_t>(ceil(step * (worker_i + 1)));
 210   size_t deactivate_offset = static_cast<size_t>(floor(step * worker_i));
 211   return Thresholds(green_zone + activate_offset,
 212                     green_zone + deactivate_offset);
 213 }
 214 
 215 G1ConcurrentRefine::G1ConcurrentRefine(size_t green_zone,
 216                                        size_t yellow_zone,
 217                                        size_t red_zone,
 218                                        size_t min_yellow_zone_size) :
 219   _thread_control(),
 220   _green_zone(green_zone),
 221   _yellow_zone(yellow_zone),
 222   _red_zone(red_zone),
 223   _min_yellow_zone_size(min_yellow_zone_size)
 224 {
 225   assert_zone_constraints_gyr(green_zone, yellow_zone, red_zone);
 226 }
 227 
 228 jint G1ConcurrentRefine::initialize() {
 229   return _thread_control.initialize(this, max_num_threads());
 230 }
 231 
 232 static size_t calc_min_yellow_zone_size() {
 233   size_t step = G1ConcRefinementThresholdStep;
 234   uint n_workers = G1ConcurrentRefine::max_num_threads();
 235   if ((max_yellow_zone / step) < n_workers) {
 236     return max_yellow_zone;
 237   } else {
 238     return step * n_workers;
 239   }
 240 }
 241 
 242 static size_t calc_init_green_zone() {
 243   size_t green = G1ConcRefinementGreenZone;
 244   if (FLAG_IS_DEFAULT(G1ConcRefinementGreenZone)) {
 245     green = ParallelGCThreads;
 246   }
 247   return MIN2(green, max_green_zone);
 248 }
 249 
 250 static size_t calc_init_yellow_zone(size_t green, size_t min_size) {
 251   size_t config = G1ConcRefinementYellowZone;
 252   size_t size = 0;
 253   if (FLAG_IS_DEFAULT(G1ConcRefinementYellowZone)) {
 254     size = green * 2;
 255   } else if (green < config) {
 256     size = config - green;
 257   }
 258   size = MAX2(size, min_size);
 259   size = MIN2(size, max_yellow_zone);
 260   return MIN2(green + size, max_yellow_zone);
 261 }
 262 
 263 static size_t calc_init_red_zone(size_t green, size_t yellow) {
 264   size_t size = yellow - green;
 265   if (!FLAG_IS_DEFAULT(G1ConcRefinementRedZone)) {
 266     size_t config = G1ConcRefinementRedZone;
 267     if (yellow < config) {
 268       size = MAX2(size, config - yellow);
 269     }
 270   }
 271   return MIN2(yellow + size, max_red_zone);
 272 }
 273 
 274 G1ConcurrentRefine* G1ConcurrentRefine::create(jint* ecode) {
 275   size_t min_yellow_zone_size = calc_min_yellow_zone_size();
 276   size_t green_zone = calc_init_green_zone();
 277   size_t yellow_zone = calc_init_yellow_zone(green_zone, min_yellow_zone_size);
 278   size_t red_zone = calc_init_red_zone(green_zone, yellow_zone);
 279 
 280   LOG_ZONES("Initial Refinement Zones: "
 281             "green: " SIZE_FORMAT ", "
 282             "yellow: " SIZE_FORMAT ", "
 283             "red: " SIZE_FORMAT ", "
 284             "min yellow size: " SIZE_FORMAT,
 285             green_zone, yellow_zone, red_zone, min_yellow_zone_size);
 286 
 287   G1ConcurrentRefine* cr = new G1ConcurrentRefine(green_zone,
 288                                                   yellow_zone,
 289                                                   red_zone,
 290                                                   min_yellow_zone_size);
 291 
 292   if (cr == NULL) {
 293     *ecode = JNI_ENOMEM;
 294     vm_shutdown_during_initialization("Could not create G1ConcurrentRefine");
 295     return NULL;
 296   }
 297 
 298   *ecode = cr->initialize();
 299   return cr;
 300 }
 301 
 302 void G1ConcurrentRefine::stop() {
 303   _thread_control.stop();
 304 }
 305 
 306 G1ConcurrentRefine::~G1ConcurrentRefine() {
 307 }
 308 
 309 void G1ConcurrentRefine::threads_do(ThreadClosure *tc) {
 310   _thread_control.worker_threads_do(tc);
 311 }
 312 
 313 uint G1ConcurrentRefine::max_num_threads() {
 314   return G1ConcRefinementThreads;
 315 }
 316 
 317 void G1ConcurrentRefine::print_threads_on(outputStream* st) const {
 318   _thread_control.print_on(st);
 319 }
 320 
 321 static size_t calc_new_green_zone(size_t green,
 322                                   double update_rs_time,
 323                                   size_t update_rs_processed_buffers,
 324                                   double goal_ms) {
 325   // Adjust green zone based on whether we're meeting the time goal.
 326   // Limit to max_green_zone.
 327   const double inc_k = 1.1, dec_k = 0.9;
 328   if (update_rs_time > goal_ms) {
 329     if (green > 0) {
 330       green = static_cast<size_t>(green * dec_k);
 331     }
 332   } else if (update_rs_time < goal_ms &&
 333              update_rs_processed_buffers > green) {
 334     green = static_cast<size_t>(MAX2(green * inc_k, green + 1.0));
 335     green = MIN2(green, max_green_zone);
 336   }
 337   return green;
 338 }
 339 
 340 static size_t calc_new_yellow_zone(size_t green, size_t min_yellow_size) {
 341   size_t size = green * 2;
 342   size = MAX2(size, min_yellow_size);
 343   return MIN2(green + size, max_yellow_zone);
 344 }
 345 
 346 static size_t calc_new_red_zone(size_t green, size_t yellow) {
 347   return MIN2(yellow + (yellow - green), max_red_zone);
 348 }
 349 
 350 void G1ConcurrentRefine::update_zones(double update_rs_time,
 351                                       size_t update_rs_processed_buffers,
 352                                       double goal_ms) {
 353   log_trace( CTRL_TAGS )("Updating Refinement Zones: "
 354                          "update_rs time: %.3fms, "
 355                          "update_rs buffers: " SIZE_FORMAT ", "
 356                          "update_rs goal time: %.3fms",
 357                          update_rs_time,
 358                          update_rs_processed_buffers,
 359                          goal_ms);
 360 
 361   _green_zone = calc_new_green_zone(_green_zone,
 362                                     update_rs_time,
 363                                     update_rs_processed_buffers,
 364                                     goal_ms);
 365   _yellow_zone = calc_new_yellow_zone(_green_zone, _min_yellow_zone_size);
 366   _red_zone = calc_new_red_zone(_green_zone, _yellow_zone);
 367 
 368   assert_zone_constraints_gyr(_green_zone, _yellow_zone, _red_zone);
 369   LOG_ZONES("Updated Refinement Zones: "
 370             "green: " SIZE_FORMAT ", "
 371             "yellow: " SIZE_FORMAT ", "
 372             "red: " SIZE_FORMAT,
 373             _green_zone, _yellow_zone, _red_zone);
 374 }
 375 
 376 void G1ConcurrentRefine::adjust(double update_rs_time,
 377                                 size_t update_rs_processed_buffers,
 378                                 double goal_ms) {
 379   DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
 380 
 381   if (G1UseAdaptiveConcRefinement) {
 382     update_zones(update_rs_time, update_rs_processed_buffers, goal_ms);
 383 
 384     // Change the barrier params
 385     if (max_num_threads() == 0) {
 386       // Disable dcqs notification when there are no threads to notify.
 387       dcqs.set_process_completed_threshold(INT_MAX);
 388     } else {
 389       // Worker 0 is the primary; wakeup is via dcqs notification.
 390       STATIC_ASSERT(max_yellow_zone <= INT_MAX);
 391       size_t activate = activation_threshold(0);
 392       dcqs.set_process_completed_threshold((int)activate);
 393     }
 394     dcqs.set_max_completed_queue((int)red_zone());
 395   }
 396 
 397   size_t curr_queue_size = dcqs.completed_buffers_num();
 398   if (curr_queue_size >= yellow_zone()) {
 399     dcqs.set_completed_queue_padding(curr_queue_size);
 400   } else {
 401     dcqs.set_completed_queue_padding(0);
 402   }
 403   dcqs.notify_if_necessary();
 404 }
 405 
 406 size_t G1ConcurrentRefine::activation_threshold(uint worker_id) const {
 407   Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, worker_id);
 408   return activation_level(thresholds);
 409 }
 410 
 411 size_t G1ConcurrentRefine::deactivation_threshold(uint worker_id) const {
 412   Thresholds thresholds = calc_thresholds(_green_zone, _yellow_zone, worker_id);
 413   return deactivation_level(thresholds);
 414 }
 415 
 416 uint G1ConcurrentRefine::worker_id_offset() {
 417   return DirtyCardQueueSet::num_par_ids();
 418 }
 419 
 420 void G1ConcurrentRefine::maybe_activate_more_threads(uint worker_id, size_t num_cur_buffers) {
 421   if (activation_threshold(worker_id + 1) > num_cur_buffers) {
 422     _thread_control.maybe_activate_next(worker_id);
 423   }
 424 }
 425 
 426 bool G1ConcurrentRefine::do_refinement_step(uint worker_id) {
 427   DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
 428 
 429   size_t curr_buffer_num = dcqs.completed_buffers_num();
 430   // If the number of the buffers falls down into the yellow zone,
 431   // that means that the transition period after the evacuation pause has ended.
 432   // Since the value written to the DCQS is the same for all threads, there is no
 433   // need to synchronize.
 434   if (dcqs.completed_queue_padding() > 0 && curr_buffer_num <= yellow_zone()) {
 435     dcqs.set_completed_queue_padding(0);
 436   }
 437 
 438   maybe_activate_more_threads(worker_id, curr_buffer_num);
 439 
 440   // Process the next buffer, if there are enough left.
 441   return dcqs.refine_completed_buffer_concurrently(worker_id + worker_id_offset(),
 442                                                    deactivation_threshold(worker_id));
 443 }