brutisso@3923: /*
tschatzl@7893:  * Copyright (c) 2013, 2014, Oracle and/or its affiliates. All rights reserved.
brutisso@3923:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
brutisso@3923:  *
brutisso@3923:  * This code is free software; you can redistribute it and/or modify it
brutisso@3923:  * under the terms of the GNU General Public License version 2 only, as
brutisso@3923:  * published by the Free Software Foundation.
brutisso@3923:  *
brutisso@3923:  * This code is distributed in the hope that it will be useful, but WITHOUT
brutisso@3923:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
brutisso@3923:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
brutisso@3923:  * version 2 for more details (a copy is included in the LICENSE file that
brutisso@3923:  * accompanied this code).
brutisso@3923:  *
brutisso@3923:  * You should have received a copy of the GNU General Public License version
brutisso@3923:  * 2 along with this work; if not, write to the Free Software Foundation,
brutisso@3923:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
brutisso@3923:  *
brutisso@3923:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
brutisso@3923:  * or visit www.oracle.com if you need additional information or have any
brutisso@3923:  * questions.
brutisso@3923:  *
brutisso@3923:  */
brutisso@3923: 
brutisso@3923: #include "precompiled.hpp"
brutisso@3923: #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
brutisso@3923: #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
brutisso@3923: #include "gc_implementation/g1/g1Log.hpp"
pliden@6413: #include "gc_implementation/g1/g1StringDedup.hpp"
brutisso@7658: #include "memory/allocation.hpp"
brutisso@7658: #include "runtime/os.hpp"
brutisso@3923: 
brutisso@3923: // Helper class for avoiding interleaved logging
brutisso@3923: class LineBuffer: public StackObj {
brutisso@3923: 
brutisso@3923: private:
brutisso@3923:   static const int BUFFER_LEN = 1024;
brutisso@3923:   static const int INDENT_CHARS = 3;
brutisso@3923:   char _buffer[BUFFER_LEN];
brutisso@3923:   int _indent_level;
brutisso@3923:   int _cur;
brutisso@3923: 
drchase@6680:   void vappend(const char* format, va_list ap)  ATTRIBUTE_PRINTF(2, 0) {
brutisso@3923:     int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap);
brutisso@3923:     if (res != -1) {
brutisso@3923:       _cur += res;
brutisso@3923:     } else {
brutisso@3923:       DEBUG_ONLY(warning("buffer too small in LineBuffer");)
brutisso@3923:       _buffer[BUFFER_LEN -1] = 0;
brutisso@3923:       _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again
brutisso@3923:     }
brutisso@3923:   }
brutisso@3923: 
brutisso@3923: public:
brutisso@3923:   explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) {
brutisso@3923:     for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) {
brutisso@3923:       _buffer[_cur] = ' ';
brutisso@3923:     }
brutisso@3923:   }
brutisso@3923: 
brutisso@3923: #ifndef PRODUCT
brutisso@3923:   ~LineBuffer() {
brutisso@3923:     assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?");
brutisso@3923:   }
brutisso@3923: #endif
brutisso@3923: 
drchase@6680:   void append(const char* format, ...)  ATTRIBUTE_PRINTF(2, 3) {
brutisso@3923:     va_list ap;
brutisso@3923:     va_start(ap, format);
brutisso@3923:     vappend(format, ap);
brutisso@3923:     va_end(ap);
brutisso@3923:   }
brutisso@3923: 
brutisso@7658:   void print_cr() {
brutisso@7658:     gclog_or_tty->print_cr("%s", _buffer);
brutisso@7658:     _cur = _indent_level * INDENT_CHARS;
brutisso@7658:   }
brutisso@7658: 
drchase@6680:   void append_and_print_cr(const char* format, ...)  ATTRIBUTE_PRINTF(2, 3) {
brutisso@3923:     va_list ap;
brutisso@3923:     va_start(ap, format);
brutisso@3923:     vappend(format, ap);
brutisso@3923:     va_end(ap);
brutisso@7658:     print_cr();
brutisso@3923:   }
brutisso@3923: };
brutisso@3923: 
brutisso@4015: template <class T>
brutisso@7658: class WorkerDataArray  : public CHeapObj<mtGC> {
brutisso@7658:   friend class G1GCParPhasePrinter;
brutisso@7658:   T*          _data;
brutisso@7658:   uint        _length;
brutisso@7658:   const char* _title;
brutisso@7658:   bool        _print_sum;
brutisso@7658:   int         _log_level;
brutisso@7658:   uint        _indent_level;
brutisso@7658:   bool        _enabled;
brutisso@7658: 
brutisso@7658:   WorkerDataArray<size_t>* _thread_work_items;
brutisso@7658: 
brutisso@7658:   NOT_PRODUCT(T uninitialized();)
brutisso@7658: 
brutisso@7658:   // We are caching the sum and average to only have to calculate them once.
brutisso@7658:   // This is not done in an MT-safe way. It is intended to allow single
brutisso@7658:   // threaded code to call sum() and average() multiple times in any order
brutisso@7658:   // without having to worry about the cost.
brutisso@7658:   bool   _has_new_data;
brutisso@7658:   T      _sum;
brutisso@7658:   T      _min;
brutisso@7658:   T      _max;
brutisso@7658:   double _average;
brutisso@7658: 
brutisso@7658:  public:
brutisso@7658:   WorkerDataArray(uint length, const char* title, bool print_sum, int log_level, uint indent_level) :
brutisso@7658:     _title(title), _length(0), _print_sum(print_sum), _log_level(log_level), _indent_level(indent_level),
brutisso@7658:     _has_new_data(true), _thread_work_items(NULL), _enabled(true) {
brutisso@7658:     assert(length > 0, "Must have some workers to store data for");
brutisso@7658:     _length = length;
brutisso@7658:     _data = NEW_C_HEAP_ARRAY(T, _length, mtGC);
brutisso@4015:   }
brutisso@4015: 
brutisso@7658:   ~WorkerDataArray() {
brutisso@7658:     FREE_C_HEAP_ARRAY(T, _data, mtGC);
brutisso@4015:   }
brutisso@4015: 
brutisso@7658:   void link_thread_work_items(WorkerDataArray<size_t>* thread_work_items) {
brutisso@7658:     _thread_work_items = thread_work_items;
brutisso@4015:   }
brutisso@4015: 
brutisso@7658:   WorkerDataArray<size_t>* thread_work_items() { return _thread_work_items; }
brutisso@7658: 
brutisso@7658:   void set(uint worker_i, T value) {
brutisso@7658:     assert(worker_i < _length, err_msg("Worker %d is greater than max: %d", worker_i, _length));
brutisso@7658:     assert(_data[worker_i] == WorkerDataArray<T>::uninitialized(), err_msg("Overwriting data for worker %d in %s", worker_i, _title));
brutisso@7658:     _data[worker_i] = value;
brutisso@7658:     _has_new_data = true;
brutisso@4015:   }
brutisso@7658: 
brutisso@7658:   void set_thread_work_item(uint worker_i, size_t value) {
brutisso@7658:     assert(_thread_work_items != NULL, "No sub count");
brutisso@7658:     _thread_work_items->set(worker_i, value);
brutisso@7658:   }
brutisso@7658: 
brutisso@7658:   T get(uint worker_i) {
brutisso@7658:     assert(worker_i < _length, err_msg("Worker %d is greater than max: %d", worker_i, _length));
brutisso@7658:     assert(_data[worker_i] != WorkerDataArray<T>::uninitialized(), err_msg("No data added for worker %d", worker_i));
brutisso@7658:     return _data[worker_i];
brutisso@7658:   }
brutisso@7658: 
brutisso@7658:   void add(uint worker_i, T value) {
brutisso@7658:     assert(worker_i < _length, err_msg("Worker %d is greater than max: %d", worker_i, _length));
brutisso@7658:     assert(_data[worker_i] != WorkerDataArray<T>::uninitialized(), err_msg("No data to add to for worker %d", worker_i));
brutisso@7658:     _data[worker_i] += value;
brutisso@7658:     _has_new_data = true;
brutisso@7658:   }
brutisso@7658: 
dbuck@8310:   double average(uint active_threads){
dbuck@8310:     calculate_totals(active_threads);
brutisso@7658:     return _average;
brutisso@7658:   }
brutisso@7658: 
dbuck@8310:   T sum(uint active_threads) {
dbuck@8310:     calculate_totals(active_threads);
brutisso@7658:     return _sum;
brutisso@7658:   }
brutisso@7658: 
dbuck@8310:   T minimum(uint active_threads) {
dbuck@8310:     calculate_totals(active_threads);
brutisso@7658:     return _min;
brutisso@7658:   }
brutisso@7658: 
dbuck@8310:   T maximum(uint active_threads) {
dbuck@8310:     calculate_totals(active_threads);
brutisso@7658:     return _max;
brutisso@7658:   }
brutisso@7658: 
brutisso@7658:   void reset() PRODUCT_RETURN;
dbuck@8310:   void verify(uint active_threads) PRODUCT_RETURN;
brutisso@7658: 
brutisso@7658:   void set_enabled(bool enabled) { _enabled = enabled; }
brutisso@7658: 
brutisso@7658:   int log_level() { return _log_level;  }
brutisso@7658: 
brutisso@7658:  private:
brutisso@7658: 
dbuck@8310:   void calculate_totals(uint active_threads){
brutisso@7658:     if (!_has_new_data) {
brutisso@7658:       return;
brutisso@7658:     }
brutisso@7658: 
brutisso@7658:     _sum = (T)0;
brutisso@7658:     _min = _data[0];
brutisso@7658:     _max = _min;
dbuck@8310:     assert(active_threads <= _length, "Wrong number of active threads");
dbuck@8310:     for (uint i = 0; i < active_threads; ++i) {
brutisso@7658:       T val = _data[i];
brutisso@7658:       _sum += val;
brutisso@7658:       _min = MIN2(_min, val);
brutisso@7658:       _max = MAX2(_max, val);
brutisso@7658:     }
dbuck@8310:     _average = (double)_sum / (double)active_threads;
brutisso@7658:     _has_new_data = false;
brutisso@7658:   }
brutisso@7658: };
brutisso@7658: 
brutisso@4015: 
jmasa@4097: #ifndef PRODUCT
brutisso@4015: 
brutisso@7658: template <>
brutisso@7658: size_t WorkerDataArray<size_t>::uninitialized() {
brutisso@7658:   return (size_t)-1;
brutisso@7658: }
brutisso@7658: 
brutisso@7658: template <>
brutisso@7658: double WorkerDataArray<double>::uninitialized() {
brutisso@7658:   return -1.0;
brutisso@7658: }
brutisso@4475: 
brutisso@4015: template <class T>
brutisso@4015: void WorkerDataArray<T>::reset() {
brutisso@4015:   for (uint i = 0; i < _length; i++) {
brutisso@7658:     _data[i] = WorkerDataArray<T>::uninitialized();
brutisso@7658:   }
brutisso@7658:   if (_thread_work_items != NULL) {
brutisso@7658:     _thread_work_items->reset();
brutisso@4015:   }
brutisso@4015: }
brutisso@4015: 
brutisso@4015: template <class T>
dbuck@8310: void WorkerDataArray<T>::verify(uint active_threads) {
brutisso@7658:   if (!_enabled) {
brutisso@7658:     return;
brutisso@7658:   }
brutisso@7658: 
dbuck@8310:   assert(active_threads <= _length, "Wrong number of active threads");
dbuck@8310:   for (uint i = 0; i < active_threads; i++) {
brutisso@7658:     assert(_data[i] != WorkerDataArray<T>::uninitialized(),
brutisso@7658:         err_msg("Invalid data for worker %u in '%s'", i, _title));
brutisso@7658:   }
brutisso@7658:   if (_thread_work_items != NULL) {
dbuck@8310:     _thread_work_items->verify(active_threads);
brutisso@4015:   }
brutisso@4015: }
brutisso@4015: 
brutisso@4015: #endif
brutisso@4015: 
brutisso@3923: G1GCPhaseTimes::G1GCPhaseTimes(uint max_gc_threads) :
brutisso@7658:   _max_gc_threads(max_gc_threads)
brutisso@3923: {
brutisso@3923:   assert(max_gc_threads > 0, "Must have some GC threads");
brutisso@7658: 
brutisso@7658:   _gc_par_phases[GCWorkerStart] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Start (ms)", false, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[ExtRootScan] = new WorkerDataArray<double>(max_gc_threads, "Ext Root Scanning (ms)", true, G1Log::LevelFiner, 2);
brutisso@7660: 
brutisso@7660:   // Root scanning phases
brutisso@7660:   _gc_par_phases[ThreadRoots] = new WorkerDataArray<double>(max_gc_threads, "Thread Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[StringTableRoots] = new WorkerDataArray<double>(max_gc_threads, "StringTable Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[UniverseRoots] = new WorkerDataArray<double>(max_gc_threads, "Universe Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[JNIRoots] = new WorkerDataArray<double>(max_gc_threads, "JNI Handles Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[ObjectSynchronizerRoots] = new WorkerDataArray<double>(max_gc_threads, "ObjectSynchronizer Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[FlatProfilerRoots] = new WorkerDataArray<double>(max_gc_threads, "FlatProfiler Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[ManagementRoots] = new WorkerDataArray<double>(max_gc_threads, "Management Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[SystemDictionaryRoots] = new WorkerDataArray<double>(max_gc_threads, "SystemDictionary Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[CLDGRoots] = new WorkerDataArray<double>(max_gc_threads, "CLDG Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[JVMTIRoots] = new WorkerDataArray<double>(max_gc_threads, "JVMTI Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[CodeCacheRoots] = new WorkerDataArray<double>(max_gc_threads, "CodeCache Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[CMRefRoots] = new WorkerDataArray<double>(max_gc_threads, "CM RefProcessor Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[WaitForStrongCLD] = new WorkerDataArray<double>(max_gc_threads, "Wait For Strong CLD (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[WeakCLDRoots] = new WorkerDataArray<double>(max_gc_threads, "Weak CLD Roots (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660:   _gc_par_phases[SATBFiltering] = new WorkerDataArray<double>(max_gc_threads, "SATB Filtering (ms)", true, G1Log::LevelFinest, 3);
brutisso@7660: 
brutisso@7658:   _gc_par_phases[UpdateRS] = new WorkerDataArray<double>(max_gc_threads, "Update RS (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[ScanRS] = new WorkerDataArray<double>(max_gc_threads, "Scan RS (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[CodeRoots] = new WorkerDataArray<double>(max_gc_threads, "Code Root Scanning (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[ObjCopy] = new WorkerDataArray<double>(max_gc_threads, "Object Copy (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[Termination] = new WorkerDataArray<double>(max_gc_threads, "Termination (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[GCWorkerTotal] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Total (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[GCWorkerEnd] = new WorkerDataArray<double>(max_gc_threads, "GC Worker End (ms)", false, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[Other] = new WorkerDataArray<double>(max_gc_threads, "GC Worker Other (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658: 
brutisso@7658:   _update_rs_processed_buffers = new WorkerDataArray<size_t>(max_gc_threads, "Processed Buffers", true, G1Log::LevelFiner, 3);
brutisso@7658:   _gc_par_phases[UpdateRS]->link_thread_work_items(_update_rs_processed_buffers);
brutisso@7658: 
brutisso@7658:   _termination_attempts = new WorkerDataArray<size_t>(max_gc_threads, "Termination Attempts", true, G1Log::LevelFinest, 3);
brutisso@7658:   _gc_par_phases[Termination]->link_thread_work_items(_termination_attempts);
brutisso@7658: 
brutisso@7658:   _gc_par_phases[StringDedupQueueFixup] = new WorkerDataArray<double>(max_gc_threads, "Queue Fixup (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658:   _gc_par_phases[StringDedupTableFixup] = new WorkerDataArray<double>(max_gc_threads, "Table Fixup (ms)", true, G1Log::LevelFiner, 2);
brutisso@7658: 
brutisso@7658:   _gc_par_phases[RedirtyCards] = new WorkerDataArray<double>(max_gc_threads, "Parallel Redirty", true, G1Log::LevelFinest, 3);
brutisso@7658:   _redirtied_cards = new WorkerDataArray<size_t>(max_gc_threads, "Redirtied Cards", true, G1Log::LevelFinest, 3);
brutisso@7658:   _gc_par_phases[RedirtyCards]->link_thread_work_items(_redirtied_cards);
brutisso@3923: }
brutisso@3923: 
brutisso@7658: void G1GCPhaseTimes::note_gc_start(uint active_gc_threads, bool mark_in_progress) {
brutisso@3923:   assert(active_gc_threads > 0, "The number of threads must be > 0");
brutisso@7658:   assert(active_gc_threads <= _max_gc_threads, "The number of active threads must be <= the max number of threads");
brutisso@3923:   _active_gc_threads = active_gc_threads;
brutisso@3923: 
brutisso@7658:   for (int i = 0; i < GCParPhasesSentinel; i++) {
brutisso@7658:     _gc_par_phases[i]->reset();
brutisso@7658:   }
tschatzl@6930: 
brutisso@7658:   _gc_par_phases[StringDedupQueueFixup]->set_enabled(G1StringDedup::is_enabled());
brutisso@7658:   _gc_par_phases[StringDedupTableFixup]->set_enabled(G1StringDedup::is_enabled());
brutisso@3923: }
brutisso@3923: 
brutisso@4015: void G1GCPhaseTimes::note_gc_end() {
brutisso@7658:   for (uint i = 0; i < _active_gc_threads; i++) {
brutisso@7658:     double worker_time = _gc_par_phases[GCWorkerEnd]->get(i) - _gc_par_phases[GCWorkerStart]->get(i);
brutisso@7658:     record_time_secs(GCWorkerTotal, i , worker_time);
brutisso@3923: 
brutisso@7658:     double worker_known_time =
brutisso@7658:         _gc_par_phases[ExtRootScan]->get(i) +
brutisso@7658:         _gc_par_phases[SATBFiltering]->get(i) +
brutisso@7658:         _gc_par_phases[UpdateRS]->get(i) +
brutisso@7658:         _gc_par_phases[ScanRS]->get(i) +
brutisso@7658:         _gc_par_phases[CodeRoots]->get(i) +
brutisso@7658:         _gc_par_phases[ObjCopy]->get(i) +
brutisso@7658:         _gc_par_phases[Termination]->get(i);
brutisso@3923: 
brutisso@7658:     record_time_secs(Other, i, worker_time - worker_known_time);
brutisso@4475:   }
brutisso@3923: 
brutisso@7658:   for (int i = 0; i < GCParPhasesSentinel; i++) {
dbuck@8310:     _gc_par_phases[i]->verify(_active_gc_threads);
brutisso@7658:   }
pliden@6413: }
pliden@6413: 
brutisso@3923: void G1GCPhaseTimes::print_stats(int level, const char* str, double value) {
brutisso@3923:   LineBuffer(level).append_and_print_cr("[%s: %.1lf ms]", str, value);
brutisso@3923: }
brutisso@3923: 
tschatzl@7010: void G1GCPhaseTimes::print_stats(int level, const char* str, size_t value) {
kevinw@9327:   LineBuffer(level).append_and_print_cr("[%s: " SIZE_FORMAT "]", str, value);
tschatzl@7010: }
tschatzl@7010: 
vkempik@6552: void G1GCPhaseTimes::print_stats(int level, const char* str, double value, uint workers) {
brutisso@7658:   LineBuffer(level).append_and_print_cr("[%s: %.1lf ms, GC Workers: %u]", str, value, workers);
brutisso@3923: }
brutisso@3923: 
brutisso@3923: double G1GCPhaseTimes::accounted_time_ms() {
brutisso@3923:     // Subtract the root region scanning wait time. It's initialized to
brutisso@3923:     // zero at the start of the pause.
brutisso@3923:     double misc_time_ms = _root_region_scan_wait_time_ms;
brutisso@3923: 
brutisso@3923:     misc_time_ms += _cur_collection_par_time_ms;
brutisso@3923: 
brutisso@3923:     // Now subtract the time taken to fix up roots in generated code
brutisso@3923:     misc_time_ms += _cur_collection_code_root_fixup_time_ms;
brutisso@3923: 
tschatzl@6402:     // Strong code root purge time
tschatzl@6402:     misc_time_ms += _cur_strong_code_root_purge_time_ms;
tschatzl@6402: 
pliden@6413:     if (G1StringDedup::is_enabled()) {
pliden@6413:       // String dedup fixup time
pliden@6413:       misc_time_ms += _cur_string_dedup_fixup_time_ms;
pliden@6413:     }
pliden@6413: 
brutisso@3923:     // Subtract the time taken to clean the card table from the
brutisso@3923:     // current value of "other time"
brutisso@3923:     misc_time_ms += _cur_clear_ct_time_ms;
brutisso@3923: 
brutisso@3923:     return misc_time_ms;
brutisso@3923: }
brutisso@3923: 
brutisso@7658: // record the time a phase took in seconds
brutisso@7658: void G1GCPhaseTimes::record_time_secs(GCParPhases phase, uint worker_i, double secs) {
brutisso@7658:   _gc_par_phases[phase]->set(worker_i, secs);
brutisso@7658: }
brutisso@7658: 
brutisso@7658: // add a number of seconds to a phase
brutisso@7658: void G1GCPhaseTimes::add_time_secs(GCParPhases phase, uint worker_i, double secs) {
brutisso@7658:   _gc_par_phases[phase]->add(worker_i, secs);
brutisso@7658: }
brutisso@7658: 
brutisso@7658: void G1GCPhaseTimes::record_thread_work_item(GCParPhases phase, uint worker_i, size_t count) {
brutisso@7658:   _gc_par_phases[phase]->set_thread_work_item(worker_i, count);
brutisso@7658: }
brutisso@7658: 
brutisso@7658: // return the average time for a phase in milliseconds
brutisso@7658: double G1GCPhaseTimes::average_time_ms(GCParPhases phase) {
dbuck@8310:   return _gc_par_phases[phase]->average(_active_gc_threads) * 1000.0;
brutisso@7658: }
brutisso@7658: 
brutisso@7658: double G1GCPhaseTimes::get_time_ms(GCParPhases phase, uint worker_i) {
brutisso@7658:   return _gc_par_phases[phase]->get(worker_i) * 1000.0;
brutisso@7658: }
brutisso@7658: 
brutisso@7658: double G1GCPhaseTimes::sum_time_ms(GCParPhases phase) {
dbuck@8310:   return _gc_par_phases[phase]->sum(_active_gc_threads) * 1000.0;
brutisso@7658: }
brutisso@7658: 
brutisso@7658: double G1GCPhaseTimes::min_time_ms(GCParPhases phase) {
dbuck@8310:   return _gc_par_phases[phase]->minimum(_active_gc_threads) * 1000.0;
brutisso@7658: }
brutisso@7658: 
brutisso@7658: double G1GCPhaseTimes::max_time_ms(GCParPhases phase) {
dbuck@8310:   return _gc_par_phases[phase]->maximum(_active_gc_threads) * 1000.0;
brutisso@7658: }
brutisso@7658: 
brutisso@7658: size_t G1GCPhaseTimes::get_thread_work_item(GCParPhases phase, uint worker_i) {
brutisso@7658:   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
brutisso@7658:   return _gc_par_phases[phase]->thread_work_items()->get(worker_i);
brutisso@7658: }
brutisso@7658: 
brutisso@7658: size_t G1GCPhaseTimes::sum_thread_work_items(GCParPhases phase) {
brutisso@7658:   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
dbuck@8310:   return _gc_par_phases[phase]->thread_work_items()->sum(_active_gc_threads);
brutisso@7658: }
brutisso@7658: 
brutisso@7658: double G1GCPhaseTimes::average_thread_work_items(GCParPhases phase) {
brutisso@7658:   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
dbuck@8310:   return _gc_par_phases[phase]->thread_work_items()->average(_active_gc_threads);
brutisso@7658: }
brutisso@7658: 
brutisso@7658: size_t G1GCPhaseTimes::min_thread_work_items(GCParPhases phase) {
brutisso@7658:   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
dbuck@8310:   return _gc_par_phases[phase]->thread_work_items()->minimum(_active_gc_threads);
brutisso@7658: }
brutisso@7658: 
brutisso@7658: size_t G1GCPhaseTimes::max_thread_work_items(GCParPhases phase) {
brutisso@7658:   assert(_gc_par_phases[phase]->thread_work_items() != NULL, "No sub count");
dbuck@8310:   return _gc_par_phases[phase]->thread_work_items()->maximum(_active_gc_threads);
brutisso@7658: }
brutisso@7658: 
brutisso@7658: class G1GCParPhasePrinter : public StackObj {
brutisso@7658:   G1GCPhaseTimes* _phase_times;
brutisso@7658:  public:
brutisso@7658:   G1GCParPhasePrinter(G1GCPhaseTimes* phase_times) : _phase_times(phase_times) {}
brutisso@7658: 
brutisso@7658:   void print(G1GCPhaseTimes::GCParPhases phase_id) {
brutisso@7658:     WorkerDataArray<double>* phase = _phase_times->_gc_par_phases[phase_id];
brutisso@7658: 
brutisso@7658:     if (phase->_log_level > G1Log::level() || !phase->_enabled) {
brutisso@7658:       return;
brutisso@7658:     }
brutisso@7658: 
brutisso@7658:     if (phase->_length == 1) {
brutisso@7658:       print_single_length(phase_id, phase);
brutisso@7658:     } else {
brutisso@7658:       print_multi_length(phase_id, phase);
brutisso@7658:     }
brutisso@7658:   }
brutisso@7658: 
brutisso@7658:  private:
brutisso@7658: 
brutisso@7658:   void print_single_length(G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<double>* phase) {
brutisso@7658:     // No need for min, max, average and sum for only one worker
brutisso@7658:     LineBuffer buf(phase->_indent_level);
brutisso@7658:     buf.append_and_print_cr("[%s:  %.1lf]", phase->_title, _phase_times->get_time_ms(phase_id, 0));
brutisso@7658: 
brutisso@7658:     if (phase->_thread_work_items != NULL) {
brutisso@7658:       LineBuffer buf2(phase->_thread_work_items->_indent_level);
kevinw@9327:       buf2.append_and_print_cr("[%s:  " SIZE_FORMAT "]", phase->_thread_work_items->_title, _phase_times->sum_thread_work_items(phase_id));
brutisso@7658:     }
brutisso@7658:   }
brutisso@7658: 
brutisso@7658:   void print_time_values(LineBuffer& buf, G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<double>* phase) {
dbuck@8310:     uint active_length = _phase_times->_active_gc_threads;
dbuck@8310:     for (uint i = 0; i < active_length; ++i) {
brutisso@7658:       buf.append("  %.1lf", _phase_times->get_time_ms(phase_id, i));
brutisso@7658:     }
brutisso@7658:     buf.print_cr();
brutisso@7658:   }
brutisso@7658: 
brutisso@7658:   void print_count_values(LineBuffer& buf, G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<size_t>* thread_work_items) {
dbuck@8310:     uint active_length = _phase_times->_active_gc_threads;
dbuck@8310:     for (uint i = 0; i < active_length; ++i) {
brutisso@7658:       buf.append("  " SIZE_FORMAT, _phase_times->get_thread_work_item(phase_id, i));
brutisso@7658:     }
brutisso@7658:     buf.print_cr();
brutisso@7658:   }
brutisso@7658: 
brutisso@7658:   void print_thread_work_items(G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<size_t>* thread_work_items) {
brutisso@7658:     LineBuffer buf(thread_work_items->_indent_level);
brutisso@7658:     buf.append("[%s:", thread_work_items->_title);
brutisso@7658: 
brutisso@7658:     if (G1Log::finest()) {
brutisso@7658:       print_count_values(buf, phase_id, thread_work_items);
brutisso@7658:     }
brutisso@7658: 
brutisso@7658:     assert(thread_work_items->_print_sum, err_msg("%s does not have print sum true even though it is a count", thread_work_items->_title));
brutisso@7658: 
brutisso@7658:     buf.append_and_print_cr(" Min: " SIZE_FORMAT ", Avg: %.1lf, Max: " SIZE_FORMAT ", Diff: " SIZE_FORMAT ", Sum: " SIZE_FORMAT "]",
brutisso@7658:         _phase_times->min_thread_work_items(phase_id), _phase_times->average_thread_work_items(phase_id), _phase_times->max_thread_work_items(phase_id),
brutisso@7658:         _phase_times->max_thread_work_items(phase_id) - _phase_times->min_thread_work_items(phase_id), _phase_times->sum_thread_work_items(phase_id));
brutisso@7658:   }
brutisso@7658: 
brutisso@7658:   void print_multi_length(G1GCPhaseTimes::GCParPhases phase_id, WorkerDataArray<double>* phase) {
brutisso@7658:     LineBuffer buf(phase->_indent_level);
brutisso@7658:     buf.append("[%s:", phase->_title);
brutisso@7658: 
brutisso@7658:     if (G1Log::finest()) {
brutisso@7658:       print_time_values(buf, phase_id, phase);
brutisso@7658:     }
brutisso@7658: 
brutisso@7658:     buf.append(" Min: %.1lf, Avg: %.1lf, Max: %.1lf, Diff: %.1lf",
brutisso@7658:         _phase_times->min_time_ms(phase_id), _phase_times->average_time_ms(phase_id), _phase_times->max_time_ms(phase_id),
brutisso@7658:         _phase_times->max_time_ms(phase_id) - _phase_times->min_time_ms(phase_id));
brutisso@7658: 
brutisso@7658:     if (phase->_print_sum) {
brutisso@7658:       // for things like the start and end times the sum is not
brutisso@7658:       // that relevant
brutisso@7658:       buf.append(", Sum: %.1lf", _phase_times->sum_time_ms(phase_id));
brutisso@7658:     }
brutisso@7658: 
brutisso@7658:     buf.append_and_print_cr("]");
brutisso@7658: 
brutisso@7658:     if (phase->_thread_work_items != NULL) {
brutisso@7658:       print_thread_work_items(phase_id, phase->_thread_work_items);
brutisso@7658:     }
brutisso@7658:   }
brutisso@7658: };
brutisso@7658: 
brutisso@4015: void G1GCPhaseTimes::print(double pause_time_sec) {
brutisso@7658:   G1GCParPhasePrinter par_phase_printer(this);
brutisso@7658: 
brutisso@3923:   if (_root_region_scan_wait_time_ms > 0.0) {
brutisso@3923:     print_stats(1, "Root Region Scan Waiting", _root_region_scan_wait_time_ms);
brutisso@3923:   }
brutisso@7658: 
brutisso@7658:   print_stats(1, "Parallel Time", _cur_collection_par_time_ms, _active_gc_threads);
brutisso@7658:   for (int i = 0; i <= GCMainParPhasesLast; i++) {
brutisso@7658:     par_phase_printer.print((GCParPhases) i);
brutisso@3923:   }
brutisso@7658: 
brutisso@3923:   print_stats(1, "Code Root Fixup", _cur_collection_code_root_fixup_time_ms);
tschatzl@6402:   print_stats(1, "Code Root Purge", _cur_strong_code_root_purge_time_ms);
pliden@6413:   if (G1StringDedup::is_enabled()) {
pliden@6413:     print_stats(1, "String Dedup Fixup", _cur_string_dedup_fixup_time_ms, _active_gc_threads);
brutisso@7658:     for (int i = StringDedupPhasesFirst; i <= StringDedupPhasesLast; i++) {
brutisso@7658:       par_phase_printer.print((GCParPhases) i);
brutisso@7658:     }
pliden@6413:   }
brutisso@3923:   print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
brutisso@4015:   double misc_time_ms = pause_time_sec * MILLIUNITS - accounted_time_ms();
brutisso@3923:   print_stats(1, "Other", misc_time_ms);
brutisso@4015:   if (_cur_verify_before_time_ms > 0.0) {
brutisso@4015:     print_stats(2, "Verify Before", _cur_verify_before_time_ms);
brutisso@4015:   }
tschatzl@6406:   if (G1CollectedHeap::heap()->evacuation_failed()) {
tschatzl@6406:     double evac_fail_handling = _cur_evac_fail_recalc_used + _cur_evac_fail_remove_self_forwards +
tschatzl@6406:       _cur_evac_fail_restore_remsets;
tschatzl@6406:     print_stats(2, "Evacuation Failure", evac_fail_handling);
tschatzl@6406:     if (G1Log::finest()) {
tschatzl@6406:       print_stats(3, "Recalculate Used", _cur_evac_fail_recalc_used);
tschatzl@6406:       print_stats(3, "Remove Self Forwards", _cur_evac_fail_remove_self_forwards);
tschatzl@6406:       print_stats(3, "Restore RemSet", _cur_evac_fail_restore_remsets);
tschatzl@6406:     }
tschatzl@6406:   }
brutisso@3923:   print_stats(2, "Choose CSet",
brutisso@3923:     (_recorded_young_cset_choice_time_ms +
brutisso@3923:     _recorded_non_young_cset_choice_time_ms));
brutisso@3923:   print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
brutisso@3923:   print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
tschatzl@7218:   print_stats(2, "Redirty Cards", _recorded_redirty_logged_cards_time_ms);
brutisso@7658:   par_phase_printer.print(RedirtyCards);
tschatzl@7828: 
tschatzl@7828:   if (G1EagerReclaimHumongousObjects) {
tschatzl@7828:     print_stats(2, "Humongous Register", _cur_fast_reclaim_humongous_register_time_ms);
tschatzl@7010:     if (G1Log::finest()) {
tschatzl@7010:       print_stats(3, "Humongous Total", _cur_fast_reclaim_humongous_total);
tschatzl@7010:       print_stats(3, "Humongous Candidate", _cur_fast_reclaim_humongous_candidates);
tschatzl@7828:     }
tschatzl@7828:     print_stats(2, "Humongous Reclaim", _cur_fast_reclaim_humongous_time_ms);
tschatzl@7828:     if (G1Log::finest()) {
tschatzl@7010:       print_stats(3, "Humongous Reclaimed", _cur_fast_reclaim_humongous_reclaimed);
tschatzl@7010:     }
tschatzl@7010:   }
brutisso@3923:   print_stats(2, "Free CSet",
brutisso@3923:     (_recorded_young_free_cset_time_ms +
brutisso@3923:     _recorded_non_young_free_cset_time_ms));
tschatzl@6404:   if (G1Log::finest()) {
tschatzl@6404:     print_stats(3, "Young Free CSet", _recorded_young_free_cset_time_ms);
tschatzl@6404:     print_stats(3, "Non-Young Free CSet", _recorded_non_young_free_cset_time_ms);
tschatzl@6404:   }
brutisso@4015:   if (_cur_verify_after_time_ms > 0.0) {
brutisso@4015:     print_stats(2, "Verify After", _cur_verify_after_time_ms);
brutisso@4015:   }
brutisso@3923: }
brutisso@7658: 
brutisso@7658: G1GCParPhaseTimesTracker::G1GCParPhaseTimesTracker(G1GCPhaseTimes* phase_times, G1GCPhaseTimes::GCParPhases phase, uint worker_id) :
brutisso@7658:     _phase_times(phase_times), _phase(phase), _worker_id(worker_id) {
brutisso@7658:   if (_phase_times != NULL) {
brutisso@7658:     _start_time = os::elapsedTime();
brutisso@7658:   }
brutisso@7658: }
brutisso@7658: 
brutisso@7658: G1GCParPhaseTimesTracker::~G1GCParPhaseTimesTracker() {
brutisso@7658:   if (_phase_times != NULL) {
brutisso@7658:     _phase_times->record_time_secs(_phase, _worker_id, os::elapsedTime() - _start_time);
brutisso@7658:   }
brutisso@7658: }
brutisso@7658: