ysr@777: /*
ysr@777:  * Copyright 2001-2007 Sun Microsystems, Inc.  All Rights Reserved.
ysr@777:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
ysr@777:  *
ysr@777:  * This code is free software; you can redistribute it and/or modify it
ysr@777:  * under the terms of the GNU General Public License version 2 only, as
ysr@777:  * published by the Free Software Foundation.
ysr@777:  *
ysr@777:  * This code is distributed in the hope that it will be useful, but WITHOUT
ysr@777:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
ysr@777:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
ysr@777:  * version 2 for more details (a copy is included in the LICENSE file that
ysr@777:  * accompanied this code).
ysr@777:  *
ysr@777:  * You should have received a copy of the GNU General Public License version
ysr@777:  * 2 along with this work; if not, write to the Free Software Foundation,
ysr@777:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
ysr@777:  *
ysr@777:  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
ysr@777:  * CA 95054 USA or visit www.sun.com if you need additional information or
ysr@777:  * have any questions.
ysr@777:  *
ysr@777:  */
ysr@777: 
ysr@777: #include "incls/_precompiled.incl"
ysr@777: #include "incls/_concurrentG1Refine.cpp.incl"
ysr@777: 
ysr@777: bool ConcurrentG1Refine::_enabled = false;
ysr@777: 
ysr@777: ConcurrentG1Refine::ConcurrentG1Refine() :
ysr@777:   _pya(PYA_continue), _last_pya(PYA_continue),
ysr@777:   _last_cards_during(), _first_traversal(false),
ysr@777:   _card_counts(NULL), _cur_card_count_histo(NULL), _cum_card_count_histo(NULL),
ysr@777:   _hot_cache(NULL),
ysr@777:   _def_use_cache(false), _use_cache(false),
ysr@777:   _n_periods(0), _total_cards(0), _total_travs(0)
ysr@777: {
ysr@777:   if (G1ConcRefine) {
ysr@777:     _cg1rThread = new ConcurrentG1RefineThread(this);
ysr@777:     assert(cg1rThread() != NULL, "Conc refine should have been created");
ysr@777:     assert(cg1rThread()->cg1r() == this,
ysr@777:            "Conc refine thread should refer to this");
ysr@777:   } else {
ysr@777:     _cg1rThread = NULL;
ysr@777:   }
ysr@777: }
ysr@777: 
ysr@777: void ConcurrentG1Refine::init() {
ysr@777:   if (G1ConcRSLogCacheSize > 0 || G1ConcRSCountTraversals) {
ysr@777:     G1CollectedHeap* g1h = G1CollectedHeap::heap();
ysr@777:     _n_card_counts =
ysr@777:       (unsigned) (g1h->g1_reserved_obj_bytes() >> CardTableModRefBS::card_shift);
ysr@777:     _card_counts = NEW_C_HEAP_ARRAY(unsigned char, _n_card_counts);
ysr@777:     for (size_t i = 0; i < _n_card_counts; i++) _card_counts[i] = 0;
ysr@777:     ModRefBarrierSet* bs = g1h->mr_bs();
ysr@777:     guarantee(bs->is_a(BarrierSet::CardTableModRef), "Precondition");
ysr@777:     CardTableModRefBS* ctbs = (CardTableModRefBS*)bs;
ysr@777:     _ct_bot = ctbs->byte_for_const(g1h->reserved_region().start());
ysr@777:     if (G1ConcRSCountTraversals) {
ysr@777:       _cur_card_count_histo = NEW_C_HEAP_ARRAY(unsigned, 256);
ysr@777:       _cum_card_count_histo = NEW_C_HEAP_ARRAY(unsigned, 256);
ysr@777:       for (int i = 0; i < 256; i++) {
ysr@777:         _cur_card_count_histo[i] = 0;
ysr@777:         _cum_card_count_histo[i] = 0;
ysr@777:       }
ysr@777:     }
ysr@777:   }
ysr@777:   if (G1ConcRSLogCacheSize > 0) {
ysr@777:     _def_use_cache = true;
ysr@777:     _use_cache = true;
ysr@777:     _hot_cache_size = (1 << G1ConcRSLogCacheSize);
ysr@777:     _hot_cache = NEW_C_HEAP_ARRAY(jbyte*, _hot_cache_size);
ysr@777:     _n_hot = 0;
ysr@777:     _hot_cache_idx = 0;
ysr@777:   }
ysr@777: }
ysr@777: 
ysr@777: ConcurrentG1Refine::~ConcurrentG1Refine() {
ysr@777:   if (G1ConcRSLogCacheSize > 0 || G1ConcRSCountTraversals) {
ysr@777:     assert(_card_counts != NULL, "Logic");
ysr@777:     FREE_C_HEAP_ARRAY(unsigned char, _card_counts);
ysr@777:     assert(_cur_card_count_histo != NULL, "Logic");
ysr@777:     FREE_C_HEAP_ARRAY(unsigned, _cur_card_count_histo);
ysr@777:     assert(_cum_card_count_histo != NULL, "Logic");
ysr@777:     FREE_C_HEAP_ARRAY(unsigned, _cum_card_count_histo);
ysr@777:   }
ysr@777:   if (G1ConcRSLogCacheSize > 0) {
ysr@777:     assert(_hot_cache != NULL, "Logic");
ysr@777:     FREE_C_HEAP_ARRAY(jbyte*, _hot_cache);
ysr@777:   }
ysr@777: }
ysr@777: 
ysr@777: bool ConcurrentG1Refine::refine() {
ysr@777:   G1CollectedHeap* g1h = G1CollectedHeap::heap();
ysr@777:   unsigned cards_before = g1h->g1_rem_set()->conc_refine_cards();
ysr@777:   clear_hot_cache();  // Any previous values in this are now invalid.
ysr@777:   g1h->g1_rem_set()->concurrentRefinementPass(this);
ysr@777:   _traversals++;
ysr@777:   unsigned cards_after = g1h->g1_rem_set()->conc_refine_cards();
ysr@777:   unsigned cards_during = cards_after-cards_before;
ysr@777:   // If this is the first traversal in the current enabling
ysr@777:   // and we did some cards, or if the number of cards found is decreasing
ysr@777:   // sufficiently quickly, then keep going.  Otherwise, sleep a while.
ysr@777:   bool res =
ysr@777:     (_first_traversal && cards_during > 0)
ysr@777:     ||
ysr@777:     (!_first_traversal && cards_during * 3 < _last_cards_during * 2);
ysr@777:   _last_cards_during = cards_during;
ysr@777:   _first_traversal = false;
ysr@777:   return res;
ysr@777: }
ysr@777: 
ysr@777: void ConcurrentG1Refine::enable() {
ysr@777:   MutexLocker x(G1ConcRefine_mon);
ysr@777:   if (!_enabled) {
ysr@777:     _enabled = true;
ysr@777:     _first_traversal = true; _last_cards_during = 0;
ysr@777:     G1ConcRefine_mon->notify_all();
ysr@777:   }
ysr@777: }
ysr@777: 
ysr@777: unsigned ConcurrentG1Refine::disable() {
ysr@777:   MutexLocker x(G1ConcRefine_mon);
ysr@777:   if (_enabled) {
ysr@777:     _enabled = false;
ysr@777:     return _traversals;
ysr@777:   } else {
ysr@777:     return 0;
ysr@777:   }
ysr@777: }
ysr@777: 
ysr@777: void ConcurrentG1Refine::wait_for_ConcurrentG1Refine_enabled() {
ysr@777:   G1ConcRefine_mon->lock();
ysr@777:   while (!_enabled) {
ysr@777:     G1ConcRefine_mon->wait(Mutex::_no_safepoint_check_flag);
ysr@777:   }
ysr@777:   G1ConcRefine_mon->unlock();
ysr@777:   _traversals = 0;
ysr@777: };
ysr@777: 
ysr@777: void ConcurrentG1Refine::set_pya_restart() {
ysr@777:   // If we're using the log-based RS barrier, the above will cause
ysr@777:   // in-progress traversals of completed log buffers to quit early; we will
ysr@777:   // also abandon all other buffers.
ysr@777:   if (G1RSBarrierUseQueue) {
ysr@777:     DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
ysr@777:     dcqs.abandon_logs();
ysr@777:     if (_cg1rThread->do_traversal()) {
ysr@777:       _pya = PYA_restart;
ysr@777:     } else {
ysr@777:       _cg1rThread->set_do_traversal(true);
ysr@777:       // Reset the post-yield actions.
ysr@777:       _pya = PYA_continue;
ysr@777:       _last_pya = PYA_continue;
ysr@777:     }
ysr@777:   } else {
ysr@777:     _pya = PYA_restart;
ysr@777:   }
ysr@777: }
ysr@777: 
ysr@777: void ConcurrentG1Refine::set_pya_cancel() {
ysr@777:   _pya = PYA_cancel;
ysr@777: }
ysr@777: 
ysr@777: PostYieldAction ConcurrentG1Refine::get_pya() {
ysr@777:   if (_pya != PYA_continue) {
ysr@777:     jint val = _pya;
ysr@777:     while (true) {
ysr@777:       jint val_read = Atomic::cmpxchg(PYA_continue, &_pya, val);
ysr@777:       if (val_read == val) {
ysr@777:         PostYieldAction res = (PostYieldAction)val;
ysr@777:         assert(res != PYA_continue, "Only the refine thread should reset.");
ysr@777:         _last_pya = res;
ysr@777:         return res;
ysr@777:       } else {
ysr@777:         val = val_read;
ysr@777:       }
ysr@777:     }
ysr@777:   }
ysr@777:   // QQQ WELL WHAT DO WE RETURN HERE???
ysr@777:   // make up something!
ysr@777:   return PYA_continue;
ysr@777: }
ysr@777: 
ysr@777: PostYieldAction ConcurrentG1Refine::get_last_pya() {
ysr@777:   PostYieldAction res = _last_pya;
ysr@777:   _last_pya = PYA_continue;
ysr@777:   return res;
ysr@777: }
ysr@777: 
ysr@777: bool ConcurrentG1Refine::do_traversal() {
ysr@777:   return _cg1rThread->do_traversal();
ysr@777: }
ysr@777: 
ysr@777: int ConcurrentG1Refine::add_card_count(jbyte* card_ptr) {
ysr@777:   size_t card_num = (card_ptr - _ct_bot);
ysr@777:   guarantee(0 <= card_num && card_num < _n_card_counts, "Bounds");
ysr@777:   unsigned char cnt = _card_counts[card_num];
ysr@777:   if (cnt < 255) _card_counts[card_num]++;
ysr@777:   return cnt;
ysr@777:   _total_travs++;
ysr@777: }
ysr@777: 
ysr@777: jbyte* ConcurrentG1Refine::cache_insert(jbyte* card_ptr) {
ysr@777:   int count = add_card_count(card_ptr);
ysr@777:   // Count previously unvisited cards.
ysr@777:   if (count == 0) _total_cards++;
ysr@777:   // We'll assume a traversal unless we store it in the cache.
ysr@777:   if (count < G1ConcRSHotCardLimit) {
ysr@777:     _total_travs++;
ysr@777:     return card_ptr;
ysr@777:   }
ysr@777:   // Otherwise, it's hot.
ysr@777:   jbyte* res = NULL;
ysr@777:   MutexLockerEx x(HotCardCache_lock, Mutex::_no_safepoint_check_flag);
ysr@777:   if (_n_hot == _hot_cache_size) {
ysr@777:     _total_travs++;
ysr@777:     res = _hot_cache[_hot_cache_idx];
ysr@777:     _n_hot--;
ysr@777:   }
ysr@777:   // Now _n_hot < _hot_cache_size, and we can insert at _hot_cache_idx.
ysr@777:   _hot_cache[_hot_cache_idx] = card_ptr;
ysr@777:   _hot_cache_idx++;
ysr@777:   if (_hot_cache_idx == _hot_cache_size) _hot_cache_idx = 0;
ysr@777:   _n_hot++;
ysr@777:   return res;
ysr@777: }
ysr@777: 
ysr@777: 
ysr@777: void ConcurrentG1Refine::clean_up_cache(int worker_i, G1RemSet* g1rs) {
ysr@777:   assert(!use_cache(), "cache should be disabled");
ysr@777:   int start_ind = _hot_cache_idx-1;
ysr@777:   for (int i = 0; i < _n_hot; i++) {
ysr@777:     int ind = start_ind - i;
ysr@777:     if (ind < 0) ind = ind + _hot_cache_size;
ysr@777:     jbyte* entry = _hot_cache[ind];
ysr@777:     if (entry != NULL) {
ysr@777:       g1rs->concurrentRefineOneCard(entry, worker_i);
ysr@777:     }
ysr@777:   }
ysr@777:   _n_hot = 0;
ysr@777:   _hot_cache_idx = 0;
ysr@777: }
ysr@777: 
ysr@777: void ConcurrentG1Refine::clear_and_record_card_counts() {
ysr@777:   if (G1ConcRSLogCacheSize == 0 && !G1ConcRSCountTraversals) return;
ysr@777:   _n_periods++;
ysr@777:   if (G1ConcRSCountTraversals) {
ysr@777:     for (size_t i = 0; i < _n_card_counts; i++) {
ysr@777:       unsigned char bucket = _card_counts[i];
ysr@777:       _cur_card_count_histo[bucket]++;
ysr@777:       _card_counts[i] = 0;
ysr@777:     }
ysr@777:     gclog_or_tty->print_cr("Card counts:");
ysr@777:     for (int i = 0; i < 256; i++) {
ysr@777:       if (_cur_card_count_histo[i] > 0) {
ysr@777:         gclog_or_tty->print_cr("  %3d: %9d", i, _cur_card_count_histo[i]);
ysr@777:         _cum_card_count_histo[i] += _cur_card_count_histo[i];
ysr@777:         _cur_card_count_histo[i] = 0;
ysr@777:       }
ysr@777:     }
ysr@777:   } else {
ysr@777:     assert(G1ConcRSLogCacheSize > 0, "Logic");
ysr@777:     Copy::fill_to_words((HeapWord*)(&_card_counts[0]),
ysr@777:                         _n_card_counts / HeapWordSize);
ysr@777:   }
ysr@777: }
ysr@777: 
ysr@777: void
ysr@777: ConcurrentG1Refine::
ysr@777: print_card_count_histo_range(unsigned* histo, int from, int to,
ysr@777:                              float& cum_card_pct,
ysr@777:                              float& cum_travs_pct) {
ysr@777:   unsigned cards = 0;
ysr@777:   unsigned travs = 0;
ysr@777:   guarantee(to <= 256, "Precondition");
ysr@777:   for (int i = from; i < to-1; i++) {
ysr@777:     cards += histo[i];
ysr@777:     travs += histo[i] * i;
ysr@777:   }
ysr@777:   if (to == 256) {
ysr@777:     unsigned histo_card_sum = 0;
ysr@777:     unsigned histo_trav_sum = 0;
ysr@777:     for (int i = 1; i < 255; i++) {
ysr@777:       histo_trav_sum += histo[i] * i;
ysr@777:     }
ysr@777:     cards += histo[255];
ysr@777:     // correct traversals for the last one.
ysr@777:     unsigned travs_255 = (unsigned) (_total_travs - histo_trav_sum);
ysr@777:     travs += travs_255;
ysr@777: 
ysr@777:   } else {
ysr@777:     cards += histo[to-1];
ysr@777:     travs += histo[to-1] * (to-1);
ysr@777:   }
ysr@777:   float fperiods = (float)_n_periods;
ysr@777:   float f_tot_cards = (float)_total_cards/fperiods;
ysr@777:   float f_tot_travs = (float)_total_travs/fperiods;
ysr@777:   if (cards > 0) {
ysr@777:     float fcards = (float)cards/fperiods;
ysr@777:     float ftravs = (float)travs/fperiods;
ysr@777:     if (to == 256) {
ysr@777:       gclog_or_tty->print(" %4d-       %10.2f%10.2f", from, fcards, ftravs);
ysr@777:     } else {
ysr@777:       gclog_or_tty->print(" %4d-%4d   %10.2f%10.2f", from, to-1, fcards, ftravs);
ysr@777:     }
ysr@777:     float pct_cards = fcards*100.0/f_tot_cards;
ysr@777:     cum_card_pct += pct_cards;
ysr@777:     float pct_travs = ftravs*100.0/f_tot_travs;
ysr@777:     cum_travs_pct += pct_travs;
ysr@777:     gclog_or_tty->print_cr("%10.2f%10.2f%10.2f%10.2f",
ysr@777:                   pct_cards, cum_card_pct,
ysr@777:                   pct_travs, cum_travs_pct);
ysr@777:   }
ysr@777: }
ysr@777: 
ysr@777: void ConcurrentG1Refine::print_final_card_counts() {
ysr@777:   if (!G1ConcRSCountTraversals) return;
ysr@777: 
ysr@777:   gclog_or_tty->print_cr("Did %d total traversals of %d distinct cards.",
ysr@777:                 _total_travs, _total_cards);
ysr@777:   float fperiods = (float)_n_periods;
ysr@777:   gclog_or_tty->print_cr("  This is an average of %8.2f traversals, %8.2f cards, "
ysr@777:                 "per collection.", (float)_total_travs/fperiods,
ysr@777:                 (float)_total_cards/fperiods);
ysr@777:   gclog_or_tty->print_cr("  This is an average of %8.2f traversals/distinct "
ysr@777:                 "dirty card.\n",
ysr@777:                 _total_cards > 0 ?
ysr@777:                 (float)_total_travs/(float)_total_cards : 0.0);
ysr@777: 
ysr@777: 
ysr@777:   gclog_or_tty->print_cr("Histogram:\n\n%10s   %10s%10s%10s%10s%10s%10s",
ysr@777:                 "range", "# cards", "# travs", "% cards", "(cum)",
ysr@777:                 "% travs", "(cum)");
ysr@777:   gclog_or_tty->print_cr("------------------------------------------------------------"
ysr@777:                 "-------------");
ysr@777:   float cum_cards_pct = 0.0;
ysr@777:   float cum_travs_pct = 0.0;
ysr@777:   for (int i = 1; i < 10; i++) {
ysr@777:     print_card_count_histo_range(_cum_card_count_histo, i, i+1,
ysr@777:                                  cum_cards_pct, cum_travs_pct);
ysr@777:   }
ysr@777:   for (int i = 10; i < 100; i += 10) {
ysr@777:     print_card_count_histo_range(_cum_card_count_histo, i, i+10,
ysr@777:                                  cum_cards_pct, cum_travs_pct);
ysr@777:   }
ysr@777:   print_card_count_histo_range(_cum_card_count_histo, 100, 150,
ysr@777:                                cum_cards_pct, cum_travs_pct);
ysr@777:   print_card_count_histo_range(_cum_card_count_histo, 150, 200,
ysr@777:                                cum_cards_pct, cum_travs_pct);
ysr@777:   print_card_count_histo_range(_cum_card_count_histo, 150, 255,
ysr@777:                                cum_cards_pct, cum_travs_pct);
ysr@777:   print_card_count_histo_range(_cum_card_count_histo, 255, 256,
ysr@777:                                cum_cards_pct, cum_travs_pct);
ysr@777: }