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/concurrentG1Refine.hpp"
  27 #include "gc/g1/g1CollectedHeap.inline.hpp"
  28 #include "gc/g1/g1GCPhaseTimes.hpp"
  29 #include "gc/g1/g1StringDedup.hpp"
  30 #include "gc/g1/workerDataArray.inline.hpp"
  31 #include "memory/allocation.hpp"
  32 #include "logging/log.hpp"
  33 #include "runtime/timer.hpp"
  34 #include "runtime/os.hpp"
  35 
  36 // Helper class for avoiding interleaved logging
  37 class LineBuffer: public StackObj {
  38 
  39 private:
  40   static const int BUFFER_LEN = 1024;
  41   static const int INDENT_CHARS = 3;
  42   char _buffer[BUFFER_LEN];
  43   int _indent_level;
  44   int _cur;
  45 
  46   void vappend(const char* format, va_list ap)  ATTRIBUTE_PRINTF(2, 0) {
  47     int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap);
  48     if (res != -1) {
  49       _cur += res;
  50     } else {
  51       DEBUG_ONLY(warning("buffer too small in LineBuffer");)
  52       _buffer[BUFFER_LEN -1] = 0;
  53       _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again
  54     }
  55   }
  56 
  57 public:
  58   explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) {
  59     for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) {
  60       _buffer[_cur] = ' ';
  61     }
  62   }
  63 
  64 #ifndef PRODUCT
  65   ~LineBuffer() {
  66     assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?");
  67   }
  68 #endif
  69 
  70   void append(const char* format, ...)  ATTRIBUTE_PRINTF(2, 3) {
  71     va_list ap;
  72     va_start(ap, format);
  73     vappend(format, ap);
  74     va_end(ap);
  75   }
  76 
  77   const char* to_string() {
  78     _cur = _indent_level * INDENT_CHARS;
  79     return _buffer;
  80   }
  81 };
  82 
  83 static const char* Indents[4] = {"", "   ", "      ", "         "};
  84 
  85 G1GCPhaseTimes::G1GCPhaseTimes(uint max_gc_threads) :
  86   _max_gc_threads(max_gc_threads)
  87 {
  88   assert(max_gc_threads > 0, "Must have some GC threads");
  89 
  90   _gc_par_phases[GCWorkerStart] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Start:", false, 2);
  91   _gc_par_phases[ExtRootScan] = new WorkerDataArray<double>(max_gc_threads, "Ext Root Scanning:", true, 2);
  92 
  93   // Root scanning phases
  94   _gc_par_phases[ThreadRoots] = new WorkerDataArray<double>(max_gc_threads, "Thread Roots:", true, 3);
  95   _gc_par_phases[StringTableRoots] = new WorkerDataArray<double>(max_gc_threads, "StringTable Roots:", true, 3);
  96   _gc_par_phases[UniverseRoots] = new WorkerDataArray<double>(max_gc_threads, "Universe Roots:", true, 3);
  97   _gc_par_phases[JNIRoots] = new WorkerDataArray<double>(max_gc_threads, "JNI Handles Roots:", true, 3);
  98   _gc_par_phases[ObjectSynchronizerRoots] = new WorkerDataArray<double>(max_gc_threads, "ObjectSynchronizer Roots:", true, 3);
  99   _gc_par_phases[FlatProfilerRoots] = new WorkerDataArray<double>(max_gc_threads, "FlatProfiler Roots:", true, 3);
 100   _gc_par_phases[ManagementRoots] = new WorkerDataArray<double>(max_gc_threads, "Management Roots:", true, 3);
 101   _gc_par_phases[SystemDictionaryRoots] = new WorkerDataArray<double>(max_gc_threads, "SystemDictionary Roots:", true, 3);
 102   _gc_par_phases[CLDGRoots] = new WorkerDataArray<double>(max_gc_threads, "CLDG Roots:", true, 3);
 103   _gc_par_phases[JVMTIRoots] = new WorkerDataArray<double>(max_gc_threads, "JVMTI Roots:", true, 3);
 104   _gc_par_phases[CMRefRoots] = new WorkerDataArray<double>(max_gc_threads, "CM RefProcessor Roots:", true, 3);
 105   _gc_par_phases[WaitForStrongCLD] = new WorkerDataArray<double>(max_gc_threads, "Wait For Strong CLD:", true, 3);
 106   _gc_par_phases[WeakCLDRoots] = new WorkerDataArray<double>(max_gc_threads, "Weak CLD Roots:", true, 3);
 107   _gc_par_phases[SATBFiltering] = new WorkerDataArray<double>(max_gc_threads, "SATB Filtering:", true, 3);
 108 
 109   _gc_par_phases[UpdateRS] = new WorkerDataArray<double>(max_gc_threads, "Update RS:", true, 2);
 110   _gc_par_phases[ScanHCC] = new WorkerDataArray<double>(max_gc_threads, "Scan HCC:", true, 3);
 111   _gc_par_phases[ScanHCC]->set_enabled(ConcurrentG1Refine::hot_card_cache_enabled());
 112   _gc_par_phases[ScanRS] = new WorkerDataArray<double>(max_gc_threads, "Scan RS:", true, 2);
 113   _gc_par_phases[CodeRoots] = new WorkerDataArray<double>(max_gc_threads, "Code Root Scanning:", true, 2);
 114   _gc_par_phases[ObjCopy] = new WorkerDataArray<double>(max_gc_threads, "Object Copy:", true, 2);
 115   _gc_par_phases[Termination] = new WorkerDataArray<double>(max_gc_threads, "Termination:", true, 2);
 116   _gc_par_phases[GCWorkerTotal] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Total:", true, 2);
 117   _gc_par_phases[GCWorkerEnd] = new WorkerDataArray<double>(max_gc_threads, "GC Worker End:", false, 2);
 118   _gc_par_phases[Other] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Other:", true, 2);



 119 
 120   _update_rs_processed_buffers = new WorkerDataArray<size_t>(max_gc_threads, "Processed Buffers:", true, 3);
 121   _gc_par_phases[UpdateRS]->link_thread_work_items(_update_rs_processed_buffers);
 122 
 123   _termination_attempts = new WorkerDataArray<size_t>(max_gc_threads, "Termination Attempts:", true, 3);
 124   _gc_par_phases[Termination]->link_thread_work_items(_termination_attempts);
 125 
 126   _gc_par_phases[StringDedupQueueFixup] = new WorkerDataArray<double>(max_gc_threads, "Queue Fixup:", true, 2);
 127   _gc_par_phases[StringDedupTableFixup] = new WorkerDataArray<double>(max_gc_threads, "Table Fixup:", true, 2);





 128 
 129   _gc_par_phases[RedirtyCards] = new WorkerDataArray<double>(max_gc_threads, "Parallel Redirty:", true, 3);
 130   _redirtied_cards = new WorkerDataArray<size_t>(max_gc_threads, "Redirtied Cards:", true, 3);
 131   _gc_par_phases[RedirtyCards]->link_thread_work_items(_redirtied_cards);
 132 }
 133 
 134 void G1GCPhaseTimes::note_gc_start(uint active_gc_threads) {
 135   assert(active_gc_threads > 0, "The number of threads must be > 0");
 136   assert(active_gc_threads <= _max_gc_threads, "The number of active threads must be <= the max number of threads");
 137   _gc_start_counter = os::elapsed_counter();
 138   _active_gc_threads = active_gc_threads;
 139   _cur_expand_heap_time_ms = 0.0;
 140   _external_accounted_time_ms = 0.0;
 141 
 142   for (int i = 0; i < GCParPhasesSentinel; i++) {

 143     _gc_par_phases[i]->reset();
 144   }
 145 
 146   _gc_par_phases[StringDedupQueueFixup]->set_enabled(G1StringDedup::is_enabled());
 147   _gc_par_phases[StringDedupTableFixup]->set_enabled(G1StringDedup::is_enabled());
 148 }
 149 
 150 void G1GCPhaseTimes::note_gc_end() {
 151   _gc_pause_time_ms = TimeHelper::counter_to_millis(os::elapsed_counter() - _gc_start_counter);
 152   for (uint i = 0; i < _active_gc_threads; i++) {
 153     double worker_time = _gc_par_phases[GCWorkerEnd]->get(i) - _gc_par_phases[GCWorkerStart]->get(i);
 154     record_time_secs(GCWorkerTotal, i , worker_time);
 155 
 156     double worker_known_time =
 157         _gc_par_phases[ExtRootScan]->get(i) +
 158         _gc_par_phases[SATBFiltering]->get(i) +
 159         _gc_par_phases[UpdateRS]->get(i) +
 160         _gc_par_phases[ScanRS]->get(i) +
 161         _gc_par_phases[CodeRoots]->get(i) +
 162         _gc_par_phases[ObjCopy]->get(i) +
 163         _gc_par_phases[Termination]->get(i);
 164 
 165     record_time_secs(Other, i, worker_time - worker_known_time);
 166   }
 167 
 168   for (int i = 0; i < GCParPhasesSentinel; i++) {

 169     _gc_par_phases[i]->verify(_active_gc_threads);
 170   }
 171 }
 172 
 173 void G1GCPhaseTimes::print_stats(const char* indent, const char* str, double value) {
 174   log_debug(gc, phases)("%s%s: %.1lf ms", indent, str, value);
 175 }
 176 
 177 double G1GCPhaseTimes::accounted_time_ms() {
 178     // First subtract any externally accounted time
 179     double misc_time_ms = _external_accounted_time_ms;
 180 
 181     // Subtract the root region scanning wait time. It's initialized to
 182     // zero at the start of the pause.
 183     misc_time_ms += _root_region_scan_wait_time_ms;
 184 
 185     misc_time_ms += _cur_collection_par_time_ms;
 186 
 187     // Now subtract the time taken to fix up roots in generated code
 188     misc_time_ms += _cur_collection_code_root_fixup_time_ms;
 189 
 190     // Strong code root purge time
 191     misc_time_ms += _cur_strong_code_root_purge_time_ms;
 192 
 193     if (G1StringDedup::is_enabled()) {
 194       // String dedup fixup time
 195       misc_time_ms += _cur_string_dedup_fixup_time_ms;
 196     }
 197 
 198     // Subtract the time taken to clean the card table from the
 199     // current value of "other time"
 200     misc_time_ms += _cur_clear_ct_time_ms;
 201 
 202     // Remove expand heap time from "other time"
 203     misc_time_ms += _cur_expand_heap_time_ms;
 204 
 205     return misc_time_ms;
 206 }
 207 
 208 // record the time a phase took in seconds
 209 void G1GCPhaseTimes::record_time_secs(GCParPhases phase, uint worker_i, double secs) {
 210   _gc_par_phases[phase]->set(worker_i, secs);
 211 }
 212 
 213 // add a number of seconds to a phase
 214 void G1GCPhaseTimes::add_time_secs(GCParPhases phase, uint worker_i, double secs) {
 215   _gc_par_phases[phase]->add(worker_i, secs);
 216 }
 217 
 218 void G1GCPhaseTimes::record_thread_work_item(GCParPhases phase, uint worker_i, size_t count) {
 219   _gc_par_phases[phase]->set_thread_work_item(worker_i, count);
 220 }
 221 
 222 // return the average time for a phase in milliseconds
 223 double G1GCPhaseTimes::average_time_ms(GCParPhases phase) {
 224   return _gc_par_phases[phase]->average(_active_gc_threads) * 1000.0;
 225 }
 226 
 227 double G1GCPhaseTimes::get_time_ms(GCParPhases phase, uint worker_i) {
 228   return _gc_par_phases[phase]->get(worker_i) * 1000.0;
 229 }
 230 
 231 double G1GCPhaseTimes::sum_time_ms(GCParPhases phase) {
 232   return _gc_par_phases[phase]->sum(_active_gc_threads) * 1000.0;
 233 }
 234 
 235 double G1GCPhaseTimes::min_time_ms(GCParPhases phase) {
 236   return _gc_par_phases[phase]->minimum(_active_gc_threads) * 1000.0;
 237 }
 238 
 239 double G1GCPhaseTimes::max_time_ms(GCParPhases phase) {
 240   return _gc_par_phases[phase]->maximum(_active_gc_threads) * 1000.0;
 241 }
 242 
 243 size_t G1GCPhaseTimes::get_thread_work_item(GCParPhases phase, uint worker_i) {
 244   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
 245   return _gc_par_phases[phase]->thread_work_items()->get(worker_i);
 246 }
 247 
 248 size_t G1GCPhaseTimes::sum_thread_work_items(GCParPhases phase) {
 249   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
 250   return _gc_par_phases[phase]->thread_work_items()->sum(_active_gc_threads);
 251 }
 252 
 253 double G1GCPhaseTimes::average_thread_work_items(GCParPhases phase) {
 254   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
 255   return _gc_par_phases[phase]->thread_work_items()->average(_active_gc_threads);
 256 }
 257 
 258 size_t G1GCPhaseTimes::min_thread_work_items(GCParPhases phase) {
 259   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
 260   return _gc_par_phases[phase]->thread_work_items()->minimum(_active_gc_threads);
 261 }
 262 
 263 size_t G1GCPhaseTimes::max_thread_work_items(GCParPhases phase) {
 264   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
 265   return _gc_par_phases[phase]->thread_work_items()->maximum(_active_gc_threads);
 266 }
 267 
 268 class G1GCParPhasePrinter : public StackObj {
 269   G1GCPhaseTimes* _phase_times;
 270  public:
 271   G1GCParPhasePrinter(G1GCPhaseTimes* phase_times) : _phase_times(phase_times) {}
 272 
 273   void print(G1GCPhaseTimes::GCParPhases phase_id) {
 274     WorkerDataArray<double>* phase = _phase_times->_gc_par_phases[phase_id];
 275 
 276     if (phase->_length == 1) {
 277       print_single_length(phase_id, phase);
 278     } else {
 279       print_multi_length(phase_id, phase);
 280     }
 281   }

 282 




 283 
 284  private:
 285   void print_single_length(G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<double>* phase) {
 286     // No need for min, max, average and sum for only one worker
 287     log_debug(gc, phases)("%s%s:  %.1lf", Indents[phase->_indent_level], phase->_title, _phase_times->get_time_ms(phase_id, 0));
 288 
 289     WorkerDataArray<size_t>* work_items = phase->_thread_work_items;
 290     if (work_items != NULL) {
 291       log_debug(gc, phases)("%s%s:  " SIZE_FORMAT, Indents[work_items->_indent_level], work_items->_title, _phase_times->sum_thread_work_items(phase_id));
 292     }

 293   }

 294 
 295   void print_time_values(const char* indent, G1GCPhaseTimes::GCParPhases phase_id) {
 296     if (log_is_enabled(Trace, gc)) {
 297       LineBuffer buf(0);
 298       uint active_length = _phase_times->_active_gc_threads;
 299       for (uint i = 0; i < active_length; ++i) {
 300         buf.append(" %4.1lf", _phase_times->get_time_ms(phase_id, i));
 301       }
 302       const char* line = buf.to_string();
 303       log_trace(gc, phases)("%s%-25s%s", indent, "", line);
 304     }
 305   }

 306 
 307   void print_count_values(const char* indent, G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<size_t>* thread_work_items) {
 308     if (log_is_enabled(Trace, gc)) {
 309       LineBuffer buf(0);
 310       uint active_length = _phase_times->_active_gc_threads;
 311       for (uint i = 0; i < active_length; ++i) {
 312         buf.append("  " SIZE_FORMAT, _phase_times->get_thread_work_item(phase_id, i));
 313       }
 314       const char* line = buf.to_string();
 315       log_trace(gc, phases)("%s%-25s%s", indent, "", line);
 316     }
 317   }

 318 
 319   void print_thread_work_items(G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<size_t>* thread_work_items) {
 320     const char* indent = Indents[thread_work_items->_indent_level];
 321 
 322     assert(thread_work_items->_print_sum, "%s does not have print sum true even though it is a count", thread_work_items->_title);

 323 
 324     log_debug(gc, phases)("%s%-25s Min: " SIZE_FORMAT ", Avg: %4.1lf, Max: " SIZE_FORMAT ", Diff: " SIZE_FORMAT ", Sum: " SIZE_FORMAT,
 325         indent, thread_work_items->_title,
 326         _phase_times->min_thread_work_items(phase_id), _phase_times->average_thread_work_items(phase_id), _phase_times->max_thread_work_items(phase_id),
 327         _phase_times->max_thread_work_items(phase_id) - _phase_times->min_thread_work_items(phase_id), _phase_times->sum_thread_work_items(phase_id));
 328 
 329     print_count_values(indent, phase_id, thread_work_items);
 330   }
 331 
 332   void print_multi_length(G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<double>* phase) {
 333     const char* indent = Indents[phase->_indent_level];
 334 
 335     if (phase->_print_sum) {
 336       log_debug(gc, phases)("%s%-25s Min: %4.1lf, Avg: %4.1lf, Max: %4.1lf, Diff: %4.1lf, Sum: %4.1lf",
 337           indent, phase->_title,
 338           _phase_times->min_time_ms(phase_id), _phase_times->average_time_ms(phase_id), _phase_times->max_time_ms(phase_id),
 339           _phase_times->max_time_ms(phase_id) - _phase_times->min_time_ms(phase_id), _phase_times->sum_time_ms(phase_id));
 340     } else {
 341       log_debug(gc, phases)("%s%-25s Min: %4.1lf, Avg: %4.1lf, Max: %4.1lf, Diff: %4.1lf",
 342           indent, phase->_title,
 343           _phase_times->min_time_ms(phase_id), _phase_times->average_time_ms(phase_id), _phase_times->max_time_ms(phase_id),
 344           _phase_times->max_time_ms(phase_id) - _phase_times->min_time_ms(phase_id));
 345     }
 346 
 347     print_time_values(indent, phase_id);

 348 
 349     if (phase->_thread_work_items != NULL) {
 350       print_thread_work_items(phase_id, phase->_thread_work_items);
 351     }
 352   }
 353 };
 354 
 355 void G1GCPhaseTimes::print() {
 356   note_gc_end();
 357 
 358   G1GCParPhasePrinter par_phase_printer(this);
 359 
 360   if (_root_region_scan_wait_time_ms > 0.0) {
 361     print_stats(Indents[1], "Root Region Scan Waiting", _root_region_scan_wait_time_ms);
 362   }
 363 
 364   print_stats(Indents[1], "Parallel Time", _cur_collection_par_time_ms);
 365   for (int i = 0; i <= GCMainParPhasesLast; i++) {
 366     par_phase_printer.print((GCParPhases) i);
 367   }

















 368 
 369   print_stats(Indents[1], "Code Root Fixup", _cur_collection_code_root_fixup_time_ms);
 370   print_stats(Indents[1], "Code Root Purge", _cur_strong_code_root_purge_time_ms);
 371   if (G1StringDedup::is_enabled()) {
 372     print_stats(Indents[1], "String Dedup Fixup", _cur_string_dedup_fixup_time_ms);
 373     for (int i = StringDedupPhasesFirst; i <= StringDedupPhasesLast; i++) {
 374       par_phase_printer.print((GCParPhases) i);
 375     }
 376   }
 377   print_stats(Indents[1], "Clear CT", _cur_clear_ct_time_ms);
 378   print_stats(Indents[1], "Expand Heap After Collection", _cur_expand_heap_time_ms);
 379   double misc_time_ms = _gc_pause_time_ms - accounted_time_ms();
 380   print_stats(Indents[1], "Other", misc_time_ms);




 381   if (_cur_verify_before_time_ms > 0.0) {
 382     print_stats(Indents[2], "Verify Before", _cur_verify_before_time_ms);
 383   }
 384   if (G1CollectedHeap::heap()->evacuation_failed()) {
 385     double evac_fail_handling = _cur_evac_fail_recalc_used + _cur_evac_fail_remove_self_forwards +
 386       _cur_evac_fail_restore_remsets;
 387     print_stats(Indents[2], "Evacuation Failure", evac_fail_handling);
 388     log_trace(gc, phases)("%sRecalculate Used: %.1lf ms", Indents[3], _cur_evac_fail_recalc_used);
 389     log_trace(gc, phases)("%sRemove Self Forwards: %.1lf ms", Indents[3], _cur_evac_fail_remove_self_forwards);
 390     log_trace(gc, phases)("%sRestore RemSet: %.1lf ms", Indents[3], _cur_evac_fail_restore_remsets);
 391   }
 392   print_stats(Indents[2], "Choose CSet",
 393     (_recorded_young_cset_choice_time_ms +
 394     _recorded_non_young_cset_choice_time_ms));
 395   print_stats(Indents[2], "Ref Proc", _cur_ref_proc_time_ms);
 396   print_stats(Indents[2], "Ref Enq", _cur_ref_enq_time_ms);
 397   print_stats(Indents[2], "Redirty Cards", _recorded_redirty_logged_cards_time_ms);
 398   par_phase_printer.print(RedirtyCards);
 399   if (G1EagerReclaimHumongousObjects) {
 400     print_stats(Indents[2], "Humongous Register", _cur_fast_reclaim_humongous_register_time_ms);
 401 
 402     log_trace(gc, phases)("%sHumongous Total: " SIZE_FORMAT, Indents[3], _cur_fast_reclaim_humongous_total);
 403     log_trace(gc, phases)("%sHumongous Candidate: " SIZE_FORMAT, Indents[3], _cur_fast_reclaim_humongous_candidates);
 404     print_stats(Indents[2], "Humongous Reclaim", _cur_fast_reclaim_humongous_time_ms);
 405     log_trace(gc, phases)("%sHumongous Reclaimed: " SIZE_FORMAT, Indents[3], _cur_fast_reclaim_humongous_reclaimed);
 406   }
 407   print_stats(Indents[2], "Free CSet",
 408     (_recorded_young_free_cset_time_ms +
 409     _recorded_non_young_free_cset_time_ms));
 410   log_trace(gc, phases)("%sYoung Free CSet: %.1lf ms", Indents[3], _recorded_young_free_cset_time_ms);
 411   log_trace(gc, phases)("%sNon-Young Free CSet: %.1lf ms", Indents[3], _recorded_non_young_free_cset_time_ms);
 412   if (_cur_verify_after_time_ms > 0.0) {
 413     print_stats(Indents[2], "Verify After", _cur_verify_after_time_ms);
 414   }
 415 }
 416 
 417 G1GCParPhaseTimesTracker::G1GCParPhaseTimesTracker(G1GCPhaseTimes* phase_times, G1GCPhaseTimes::GCParPhases phase, uint worker_id) :
 418     _phase_times(phase_times), _phase(phase), _worker_id(worker_id) {
 419   if (_phase_times != NULL) {
 420     _start_time = os::elapsedTime();
 421   }
 422 }
 423 
 424 G1GCParPhaseTimesTracker::~G1GCParPhaseTimesTracker() {
 425   if (_phase_times != NULL) {
 426     _phase_times->record_time_secs(_phase, _worker_id, os::elapsedTime() - _start_time);
 427   }
 428 }
 429 

  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/concurrentG1Refine.hpp"
  27 #include "gc/g1/g1CollectedHeap.inline.hpp"
  28 #include "gc/g1/g1GCPhaseTimes.hpp"
  29 #include "gc/g1/g1StringDedup.hpp"
  30 #include "gc/g1/workerDataArray.inline.hpp"
  31 #include "memory/resourceArea.hpp"
  32 #include "logging/log.hpp"
  33 #include "runtime/timer.hpp"
  34 #include "runtime/os.hpp"
  35 
  36 static const char* Indents[5] = {"", "  ", "    ", "     ", "       "};















































  37 
  38 G1GCPhaseTimes::G1GCPhaseTimes(uint max_gc_threads) :
  39   _max_gc_threads(max_gc_threads)
  40 {
  41   assert(max_gc_threads > 0, "Must have some GC threads");
  42 
  43   _gc_par_phases[GCWorkerStart] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Start (ms):");
  44   _gc_par_phases[ExtRootScan] = new WorkerDataArray<double>(max_gc_threads, "Ext Root Scanning (ms):");
  45 
  46   // Root scanning phases
  47   _gc_par_phases[ThreadRoots] = new WorkerDataArray<double>(max_gc_threads, "Thread Roots (ms):");
  48   _gc_par_phases[StringTableRoots] = new WorkerDataArray<double>(max_gc_threads, "StringTable Roots (ms):");
  49   _gc_par_phases[UniverseRoots] = new WorkerDataArray<double>(max_gc_threads, "Universe Roots (ms):");
  50   _gc_par_phases[JNIRoots] = new WorkerDataArray<double>(max_gc_threads, "JNI Handles Roots (ms):");
  51   _gc_par_phases[ObjectSynchronizerRoots] = new WorkerDataArray<double>(max_gc_threads, "ObjectSynchronizer Roots (ms):");
  52   _gc_par_phases[FlatProfilerRoots] = new WorkerDataArray<double>(max_gc_threads, "FlatProfiler Roots (ms):");
  53   _gc_par_phases[ManagementRoots] = new WorkerDataArray<double>(max_gc_threads, "Management Roots (ms):");
  54   _gc_par_phases[SystemDictionaryRoots] = new WorkerDataArray<double>(max_gc_threads, "SystemDictionary Roots (ms):");
  55   _gc_par_phases[CLDGRoots] = new WorkerDataArray<double>(max_gc_threads, "CLDG Roots (ms):");
  56   _gc_par_phases[JVMTIRoots] = new WorkerDataArray<double>(max_gc_threads, "JVMTI Roots (ms):");
  57   _gc_par_phases[CMRefRoots] = new WorkerDataArray<double>(max_gc_threads, "CM RefProcessor Roots (ms):");
  58   _gc_par_phases[WaitForStrongCLD] = new WorkerDataArray<double>(max_gc_threads, "Wait For Strong CLD (ms):");
  59   _gc_par_phases[WeakCLDRoots] = new WorkerDataArray<double>(max_gc_threads, "Weak CLD Roots (ms):");
  60   _gc_par_phases[SATBFiltering] = new WorkerDataArray<double>(max_gc_threads, "SATB Filtering (ms):");
  61 
  62   _gc_par_phases[UpdateRS] = new WorkerDataArray<double>(max_gc_threads, "Update RS (ms):");
  63   if (ConcurrentG1Refine::hot_card_cache_enabled()) {
  64     _gc_par_phases[ScanHCC] = new WorkerDataArray<double>(max_gc_threads, "Scan HCC (ms):");
  65   } else {
  66     _gc_par_phases[ScanHCC] = NULL;
  67   }
  68   _gc_par_phases[ScanRS] = new WorkerDataArray<double>(max_gc_threads, "Scan RS (ms):");
  69   _gc_par_phases[CodeRoots] = new WorkerDataArray<double>(max_gc_threads, "Code Root Scanning (ms):");
  70   _gc_par_phases[ObjCopy] = new WorkerDataArray<double>(max_gc_threads, "Object Copy (ms):");
  71   _gc_par_phases[Termination] = new WorkerDataArray<double>(max_gc_threads, "Termination (ms):");
  72   _gc_par_phases[GCWorkerTotal] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Total (ms):");
  73   _gc_par_phases[GCWorkerEnd] = new WorkerDataArray<double>(max_gc_threads, "GC Worker End (ms):");
  74   _gc_par_phases[Other] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Other (ms):");
  75 
  76   _update_rs_processed_buffers = new WorkerDataArray<size_t>(max_gc_threads, "Processed Buffers:");
  77   _gc_par_phases[UpdateRS]->link_thread_work_items(_update_rs_processed_buffers);
  78 
  79   _termination_attempts = new WorkerDataArray<size_t>(max_gc_threads, "Termination Attempts:");
  80   _gc_par_phases[Termination]->link_thread_work_items(_termination_attempts);
  81 
  82   if (UseStringDeduplication) {
  83     _gc_par_phases[StringDedupQueueFixup] = new WorkerDataArray<double>(max_gc_threads, "Queue Fixup (ms):");
  84     _gc_par_phases[StringDedupTableFixup] = new WorkerDataArray<double>(max_gc_threads, "Table Fixup (ms):");
  85   } else {
  86     _gc_par_phases[StringDedupQueueFixup] = NULL;
  87     _gc_par_phases[StringDedupTableFixup] = NULL;
  88   }
  89 
  90   _gc_par_phases[RedirtyCards] = new WorkerDataArray<double>(max_gc_threads, "Parallel Redirty:  (ms)");
  91   _redirtied_cards = new WorkerDataArray<size_t>(max_gc_threads, "Redirtied Cards:");
  92   _gc_par_phases[RedirtyCards]->link_thread_work_items(_redirtied_cards);
  93 }
  94 
  95 void G1GCPhaseTimes::note_gc_start(uint active_gc_threads) {
  96   assert(active_gc_threads > 0, "The number of threads must be > 0");
  97   assert(active_gc_threads <= _max_gc_threads, "The number of active threads must be <= the max number of threads");
  98   _gc_start_counter = os::elapsed_counter();
  99   _active_gc_threads = active_gc_threads;
 100   _cur_expand_heap_time_ms = 0.0;
 101   _external_accounted_time_ms = 0.0;
 102 
 103   for (int i = 0; i < GCParPhasesSentinel; i++) {
 104     if (_gc_par_phases[i] != NULL) {
 105       _gc_par_phases[i]->reset();
 106     }
 107   }


 108 }
 109 
 110 void G1GCPhaseTimes::note_gc_end() {
 111   _gc_pause_time_ms = TimeHelper::counter_to_millis(os::elapsed_counter() - _gc_start_counter);
 112   for (uint i = 0; i < _active_gc_threads; i++) {
 113     double worker_time = _gc_par_phases[GCWorkerEnd]->get(i) - _gc_par_phases[GCWorkerStart]->get(i);
 114     record_time_secs(GCWorkerTotal, i , worker_time);
 115 
 116     double worker_known_time =
 117         _gc_par_phases[ExtRootScan]->get(i) +
 118         _gc_par_phases[SATBFiltering]->get(i) +
 119         _gc_par_phases[UpdateRS]->get(i) +
 120         _gc_par_phases[ScanRS]->get(i) +
 121         _gc_par_phases[CodeRoots]->get(i) +
 122         _gc_par_phases[ObjCopy]->get(i) +
 123         _gc_par_phases[Termination]->get(i);
 124 
 125     record_time_secs(Other, i, worker_time - worker_known_time);
 126   }
 127 
 128   for (int i = 0; i < GCParPhasesSentinel; i++) {
 129     if (_gc_par_phases[i] != NULL) {
 130       _gc_par_phases[i]->verify(_active_gc_threads);
 131     }

























 132   }









 133 }
 134 
 135 // record the time a phase took in seconds
 136 void G1GCPhaseTimes::record_time_secs(GCParPhases phase, uint worker_i, double secs) {
 137   _gc_par_phases[phase]->set(worker_i, secs);
 138 }
 139 
 140 // add a number of seconds to a phase
 141 void G1GCPhaseTimes::add_time_secs(GCParPhases phase, uint worker_i, double secs) {
 142   _gc_par_phases[phase]->add(worker_i, secs);
 143 }
 144 
 145 void G1GCPhaseTimes::record_thread_work_item(GCParPhases phase, uint worker_i, size_t count) {
 146   _gc_par_phases[phase]->set_thread_work_item(worker_i, count);
 147 }
 148 
 149 // return the average time for a phase in milliseconds
 150 double G1GCPhaseTimes::average_time_ms(GCParPhases phase) {
 151   return _gc_par_phases[phase]->average(_active_gc_threads) * 1000.0;
 152 }
 153 





















 154 size_t G1GCPhaseTimes::sum_thread_work_items(GCParPhases phase) {
 155   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
 156   return _gc_par_phases[phase]->thread_work_items()->sum(_active_gc_threads);
 157 }
 158 
 159 template <class T>
 160 void G1GCPhaseTimes::details(T* phase, const char* indent) {
 161   LogHandle(gc, phases, task) log;
 162   if (log.is_level(LogLevel::Trace)) {
 163     outputStream* trace_out = log.trace_stream();
 164     trace_out->print("%s", indent);
 165     phase->print_details_on(trace_out, _active_gc_threads);





















 166   }
 167 }
 168 
 169 void G1GCPhaseTimes::log_phase(WorkerDataArray<double>* phase, uint indent, outputStream* out, bool print_sum) {
 170   out->print("%s", Indents[indent]);
 171   phase->print_summary_on(out, _active_gc_threads, print_sum);
 172   details(phase, Indents[indent]);
 173 
 174   WorkerDataArray<size_t>* work_items = phase->thread_work_items();





 175   if (work_items != NULL) {
 176     out->print("%s", Indents[indent + 1]);
 177     work_items->print_summary_on(out, _active_gc_threads, true);
 178     details(work_items, Indents[indent + 1]);
 179   }
 180 }
 181 
 182 void G1GCPhaseTimes::debug_phase(WorkerDataArray<double>* phase) {
 183   LogHandle(gc, phases) log;
 184   if (log.is_level(LogLevel::Debug)) {
 185     ResourceMark rm;
 186     log_phase(phase, 2, log.debug_stream(), true);





 187   }
 188 }
 189 
 190 void G1GCPhaseTimes::trace_phase(WorkerDataArray<double>* phase, bool print_sum) {
 191   LogHandle(gc, phases) log;
 192   if (log.is_level(LogLevel::Trace)) {
 193     ResourceMark rm;
 194     log_phase(phase, 3, log.trace_stream(), print_sum);





 195   }
 196 }
 197 
 198 #define PHASE_DOUBLE_FORMAT "%s%s: %.1lfms"
 199 #define PHASE_SIZE_FORMAT "%s%s: " SIZE_FORMAT
 200 
 201 #define info_line(str, value) \
 202   log_info(gc, phases)(PHASE_DOUBLE_FORMAT, Indents[1], str, value);
 203 
 204 #define debug_line(str, value) \
 205   log_debug(gc, phases)(PHASE_DOUBLE_FORMAT, Indents[2], str, value);


 206 
 207 #define trace_line(str, value) \
 208   log_trace(gc, phases)(PHASE_DOUBLE_FORMAT, Indents[3], str, value);
 209 
 210 #define trace_line_sz(str, value) \
 211   log_trace(gc, phases)(PHASE_SIZE_FORMAT, Indents[3], str, value);












 212 
 213 #define trace_line_ms(str, value) \
 214   log_trace(gc, phases)(PHASE_SIZE_FORMAT, Indents[3], str, value);
 215 
 216 #define info_line_and_account(str, value) \
 217   info_line(str, value);                  \
 218   accounted_time_ms += value;


 219 
 220 void G1GCPhaseTimes::print() {
 221   note_gc_end();
 222 
 223   double accounted_time_ms = _external_accounted_time_ms;

 224   if (_root_region_scan_wait_time_ms > 0.0) {
 225     info_line_and_account("Root Region Scan Waiting", _root_region_scan_wait_time_ms);
 226   }
 227 
 228   info_line_and_account("Evacuate Collection Set", _cur_collection_par_time_ms);
 229   trace_phase(_gc_par_phases[GCWorkerStart], false);
 230   debug_phase(_gc_par_phases[ExtRootScan]);
 231   for (int i = ThreadRoots; i <= SATBFiltering; i++) {
 232     trace_phase(_gc_par_phases[i]);
 233   }
 234   debug_phase(_gc_par_phases[UpdateRS]);
 235   if (ConcurrentG1Refine::hot_card_cache_enabled()) {
 236     trace_phase(_gc_par_phases[ScanHCC]);
 237   }
 238   debug_phase(_gc_par_phases[ScanRS]);
 239   debug_phase(_gc_par_phases[CodeRoots]);
 240   debug_phase(_gc_par_phases[ObjCopy]);
 241   debug_phase(_gc_par_phases[Termination]);
 242   debug_phase(_gc_par_phases[Other]);
 243   debug_phase(_gc_par_phases[GCWorkerTotal]);
 244   trace_phase(_gc_par_phases[GCWorkerEnd], false);
 245 
 246   info_line_and_account("Code Roots", _cur_collection_code_root_fixup_time_ms + _cur_strong_code_root_purge_time_ms);
 247   debug_line("Code Roots Fixup", _cur_collection_code_root_fixup_time_ms);
 248   debug_line("Code Roots Purge", _cur_strong_code_root_purge_time_ms);
 249 


 250   if (G1StringDedup::is_enabled()) {
 251     info_line_and_account("String Dedup Fixup", _cur_string_dedup_fixup_time_ms);
 252     debug_phase(_gc_par_phases[StringDedupQueueFixup]);
 253     debug_phase(_gc_par_phases[StringDedupTableFixup]);
 254   }
 255   info_line_and_account("Clear Card Table", _cur_clear_ct_time_ms);
 256   info_line_and_account("Expand Heap After Collection", _cur_expand_heap_time_ms);
 257 
 258   double free_cset_time = _recorded_young_free_cset_time_ms + _recorded_non_young_free_cset_time_ms;
 259   info_line_and_account("Free Collection Set", free_cset_time);
 260   debug_line("Young Free Collection Set", _recorded_young_free_cset_time_ms);
 261   debug_line("Non-Young Free Collection Set", _recorded_non_young_free_cset_time_ms);
 262 
 263   info_line("Other", _gc_pause_time_ms - accounted_time_ms);
 264   if (_cur_verify_before_time_ms > 0.0) {
 265     debug_line("Verify Before", _cur_verify_before_time_ms);
 266   }
 267   if (G1CollectedHeap::heap()->evacuation_failed()) {
 268     double evac_fail_handling = _cur_evac_fail_recalc_used + _cur_evac_fail_remove_self_forwards +
 269       _cur_evac_fail_restore_remsets;
 270     debug_line("Evacuation Failure", evac_fail_handling);
 271     trace_line("Recalculate Used", _cur_evac_fail_recalc_used);
 272     trace_line("Remove Self Forwards",_cur_evac_fail_remove_self_forwards);
 273     trace_line("Restore RemSet", _cur_evac_fail_restore_remsets);
 274   }
 275   debug_line("Choose CSet", (_recorded_young_cset_choice_time_ms + _recorded_non_young_cset_choice_time_ms));
 276   debug_line("Ref Proc", _cur_ref_proc_time_ms);
 277   debug_line("Ref Enq", _cur_ref_enq_time_ms);
 278   debug_line("Redirty Cards", _recorded_redirty_logged_cards_time_ms);
 279   trace_phase(_gc_par_phases[RedirtyCards]);


 280   if (G1EagerReclaimHumongousObjects) {
 281     debug_line("Humongous Register", _cur_fast_reclaim_humongous_register_time_ms);
 282     trace_line_sz("Humongous Total", _cur_fast_reclaim_humongous_total);
 283     trace_line_sz("Humongous Candidate", _cur_fast_reclaim_humongous_candidates);
 284     debug_line("Humongous Reclaim", _cur_fast_reclaim_humongous_time_ms);
 285     trace_line_sz("Humongous Reclaimed", _cur_fast_reclaim_humongous_reclaimed);
 286   }






 287   if (_cur_verify_after_time_ms > 0.0) {
 288     debug_line("Verify After", _cur_verify_after_time_ms);
 289   }
 290 }
 291 
 292 G1GCParPhaseTimesTracker::G1GCParPhaseTimesTracker(G1GCPhaseTimes* phase_times, G1GCPhaseTimes::GCParPhases phase, uint worker_id) :
 293     _phase_times(phase_times), _phase(phase), _worker_id(worker_id) {
 294   if (_phase_times != NULL) {
 295     _start_time = os::elapsedTime();
 296   }
 297 }
 298 
 299 G1GCParPhaseTimesTracker::~G1GCParPhaseTimesTracker() {
 300   if (_phase_times != NULL) {
 301     _phase_times->record_time_secs(_phase, _worker_id, os::elapsedTime() - _start_time);
 302   }
 303 }
 304