ysr@777: /*
brutisso@3923:  * Copyright (c) 2001, 2012, Oracle and/or its affiliates. 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:  *
trims@1907:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907:  * or visit www.oracle.com if you need additional information or have any
trims@1907:  * questions.
ysr@777:  *
ysr@777:  */
ysr@777: 
stefank@2314: #include "precompiled.hpp"
stefank@2314: #include "gc_implementation/g1/concurrentG1Refine.hpp"
stefank@2314: #include "gc_implementation/g1/concurrentG1RefineThread.hpp"
stefank@2314: #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
stefank@2314: #include "gc_implementation/g1/g1CollectorPolicy.hpp"
brutisso@3923: #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
stefank@2314: #include "gc_implementation/g1/g1RemSet.hpp"
stefank@2314: #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
stefank@2314: #include "memory/space.inline.hpp"
stefank@2314: #include "runtime/atomic.hpp"
johnc@2713: #include "runtime/java.hpp"
stefank@2314: #include "utilities/copy.hpp"
ysr@777: 
johnc@1325: // Possible sizes for the card counts cache: odd primes that roughly double in size.
johnc@1325: // (See jvmtiTagMap.cpp).
johnc@2713: 
johnc@2713: #define MAX_SIZE ((size_t) -1)
johnc@2713: 
johnc@2713: size_t ConcurrentG1Refine::_cc_cache_sizes[] = {
johnc@2713:           16381,    32771,    76831,    150001,   307261,
johnc@2713:          614563,  1228891,  2457733,   4915219,  9830479,
johnc@2713:        19660831, 39321619, 78643219, 157286461,  MAX_SIZE
johnc@1325:   };
johnc@1325: 
ysr@777: ConcurrentG1Refine::ConcurrentG1Refine() :
johnc@1325:   _card_counts(NULL), _card_epochs(NULL),
johnc@2713:   _n_card_counts(0), _max_cards(0), _max_n_card_counts(0),
johnc@1325:   _cache_size_index(0), _expand_card_counts(false),
ysr@777:   _hot_cache(NULL),
ysr@777:   _def_use_cache(false), _use_cache(false),
tonyp@2716:   // We initialize the epochs of the array to 0. By initializing
tonyp@2716:   // _n_periods to 1 and not 0 we automatically invalidate all the
tonyp@2716:   // entries on the array. Otherwise we might accidentally think that
tonyp@2716:   // we claimed a card that was in fact never set (see CR7033292).
tonyp@2716:   _n_periods(1),
iveresov@1229:   _threads(NULL), _n_threads(0)
ysr@777: {
iveresov@1546: 
iveresov@1546:   // Ergomonically select initial concurrent refinement parameters
tonyp@1717:   if (FLAG_IS_DEFAULT(G1ConcRefinementGreenZone)) {
tonyp@1717:     FLAG_SET_DEFAULT(G1ConcRefinementGreenZone, MAX2<int>(ParallelGCThreads, 1));
iveresov@1546:   }
tonyp@1717:   set_green_zone(G1ConcRefinementGreenZone);
iveresov@1546: 
tonyp@1717:   if (FLAG_IS_DEFAULT(G1ConcRefinementYellowZone)) {
tonyp@1717:     FLAG_SET_DEFAULT(G1ConcRefinementYellowZone, green_zone() * 3);
iveresov@1546:   }
tonyp@1717:   set_yellow_zone(MAX2<int>(G1ConcRefinementYellowZone, green_zone()));
iveresov@1546: 
tonyp@1717:   if (FLAG_IS_DEFAULT(G1ConcRefinementRedZone)) {
tonyp@1717:     FLAG_SET_DEFAULT(G1ConcRefinementRedZone, yellow_zone() * 2);
iveresov@1546:   }
tonyp@1717:   set_red_zone(MAX2<int>(G1ConcRefinementRedZone, yellow_zone()));
iveresov@1546:   _n_worker_threads = thread_num();
iveresov@1546:   // We need one extra thread to do the young gen rset size sampling.
iveresov@1546:   _n_threads = _n_worker_threads + 1;
iveresov@1546:   reset_threshold_step();
iveresov@1546: 
zgu@3900:   _threads = NEW_C_HEAP_ARRAY(ConcurrentG1RefineThread*, _n_threads, mtGC);
iveresov@1546:   int worker_id_offset = (int)DirtyCardQueueSet::num_par_ids();
iveresov@1546:   ConcurrentG1RefineThread *next = NULL;
iveresov@1546:   for (int i = _n_threads - 1; i >= 0; i--) {
iveresov@1546:     ConcurrentG1RefineThread* t = new ConcurrentG1RefineThread(this, next, worker_id_offset, i);
iveresov@1546:     assert(t != NULL, "Conc refine should have been created");
iveresov@1546:     assert(t->cg1r() == this, "Conc refine thread should refer to this");
iveresov@1546:     _threads[i] = t;
iveresov@1546:     next = t;
ysr@777:   }
ysr@777: }
ysr@777: 
iveresov@1546: void ConcurrentG1Refine::reset_threshold_step() {
tonyp@1717:   if (FLAG_IS_DEFAULT(G1ConcRefinementThresholdStep)) {
iveresov@1546:     _thread_threshold_step = (yellow_zone() - green_zone()) / (worker_thread_num() + 1);
iveresov@1546:   } else {
tonyp@1717:     _thread_threshold_step = G1ConcRefinementThresholdStep;
iveresov@1230:   }
iveresov@1546: }
iveresov@1546: 
iveresov@1546: int ConcurrentG1Refine::thread_num() {
tonyp@1717:   return MAX2<int>((G1ConcRefinementThreads > 0) ? G1ConcRefinementThreads : ParallelGCThreads, 1);
iveresov@1230: }
iveresov@1230: 
ysr@777: void ConcurrentG1Refine::init() {
johnc@1325:   if (G1ConcRSLogCacheSize > 0) {
johnc@1325:     _g1h = G1CollectedHeap::heap();
johnc@2713: 
johnc@2713:     _max_cards = _g1h->max_capacity() >> CardTableModRefBS::card_shift;
johnc@2713:     _max_n_card_counts = _max_cards * G1MaxHotCardCountSizePercent / 100;
johnc@1325: 
johnc@1325:     size_t max_card_num = ((size_t)1 << (sizeof(unsigned)*BitsPerByte-1)) - 1;
johnc@2713:     guarantee(_max_cards < max_card_num, "card_num representation");
johnc@1325: 
johnc@2713:     // We need _n_card_counts to be less than _max_n_card_counts here
johnc@2713:     // so that the expansion call (below) actually allocates the
johnc@2713:     // _counts and _epochs arrays.
johnc@2713:     assert(_n_card_counts == 0, "pre-condition");
johnc@2713:     assert(_max_n_card_counts > 0, "pre-condition");
johnc@2713: 
johnc@2713:     // Find the index into cache size array that is of a size that's
johnc@2713:     // large enough to hold desired_sz.
johnc@2713:     size_t desired_sz = _max_cards / InitialCacheFraction;
johnc@2713:     int desired_sz_index = 0;
johnc@2713:     while (_cc_cache_sizes[desired_sz_index] < desired_sz) {
johnc@2713:       desired_sz_index += 1;
johnc@2713:       assert(desired_sz_index <  MAX_CC_CACHE_INDEX, "invariant");
johnc@1325:     }
johnc@2713:     assert(desired_sz_index <  MAX_CC_CACHE_INDEX, "invariant");
johnc@1325: 
johnc@2713:     // If the desired_sz value is between two sizes then
johnc@2713:     // _cc_cache_sizes[desired_sz_index-1] < desired_sz <= _cc_cache_sizes[desired_sz_index]
johnc@2713:     // we will start with the lower size in the optimistic expectation that
johnc@2713:     // we will not need to expand up. Note desired_sz_index could also be 0.
johnc@2713:     if (desired_sz_index > 0 &&
johnc@2713:         _cc_cache_sizes[desired_sz_index] > desired_sz) {
johnc@2713:       desired_sz_index -= 1;
johnc@2713:     }
johnc@1325: 
johnc@2713:     if (!expand_card_count_cache(desired_sz_index)) {
johnc@2713:       // Allocation was unsuccessful - exit
johnc@2713:       vm_exit_during_initialization("Could not reserve enough space for card count cache");
johnc@2713:     }
johnc@2713:     assert(_n_card_counts > 0, "post-condition");
johnc@2713:     assert(_cache_size_index == desired_sz_index, "post-condition");
johnc@1325: 
johnc@1325:     Copy::fill_to_bytes(&_card_counts[0],
johnc@1325:                         _n_card_counts * sizeof(CardCountCacheEntry));
johnc@1325:     Copy::fill_to_bytes(&_card_epochs[0], _n_card_counts * sizeof(CardEpochCacheEntry));
johnc@1325: 
johnc@1325:     ModRefBarrierSet* bs = _g1h->mr_bs();
ysr@777:     guarantee(bs->is_a(BarrierSet::CardTableModRef), "Precondition");
johnc@1325:     _ct_bs = (CardTableModRefBS*)bs;
johnc@1325:     _ct_bot = _ct_bs->byte_for_const(_g1h->reserved_region().start());
johnc@1325: 
ysr@777:     _def_use_cache = true;
ysr@777:     _use_cache = true;
ysr@777:     _hot_cache_size = (1 << G1ConcRSLogCacheSize);
zgu@3900:     _hot_cache = NEW_C_HEAP_ARRAY(jbyte*, _hot_cache_size, mtGC);
ysr@777:     _n_hot = 0;
ysr@777:     _hot_cache_idx = 0;
johnc@1324: 
johnc@1324:     // For refining the cards in the hot cache in parallel
johnc@1324:     int n_workers = (ParallelGCThreads > 0 ?
johnc@1325:                         _g1h->workers()->total_workers() : 1);
johnc@1324:     _hot_cache_par_chunk_size = MAX2(1, _hot_cache_size / n_workers);
johnc@1324:     _hot_cache_par_claimed_idx = 0;
ysr@777:   }
ysr@777: }
ysr@777: 
iveresov@1229: void ConcurrentG1Refine::stop() {
iveresov@1229:   if (_threads != NULL) {
iveresov@1229:     for (int i = 0; i < _n_threads; i++) {
iveresov@1229:       _threads[i]->stop();
iveresov@1229:     }
iveresov@1229:   }
iveresov@1229: }
iveresov@1229: 
iveresov@1546: void ConcurrentG1Refine::reinitialize_threads() {
iveresov@1546:   reset_threshold_step();
iveresov@1546:   if (_threads != NULL) {
iveresov@1546:     for (int i = 0; i < _n_threads; i++) {
iveresov@1546:       _threads[i]->initialize();
iveresov@1546:     }
iveresov@1546:   }
iveresov@1546: }
iveresov@1546: 
ysr@777: ConcurrentG1Refine::~ConcurrentG1Refine() {
johnc@1325:   if (G1ConcRSLogCacheSize > 0) {
johnc@2713:     // Please see the comment in allocate_card_count_cache
johnc@2713:     // for why we call os::malloc() and os::free() directly.
ysr@777:     assert(_card_counts != NULL, "Logic");
zgu@3900:     os::free(_card_counts, mtGC);
johnc@1325:     assert(_card_epochs != NULL, "Logic");
zgu@3900:     os::free(_card_epochs, mtGC);
johnc@2713: 
ysr@777:     assert(_hot_cache != NULL, "Logic");
zgu@3900:     FREE_C_HEAP_ARRAY(jbyte*, _hot_cache, mtGC);
ysr@777:   }
iveresov@1229:   if (_threads != NULL) {
iveresov@1229:     for (int i = 0; i < _n_threads; i++) {
iveresov@1229:       delete _threads[i];
iveresov@1229:     }
zgu@3900:     FREE_C_HEAP_ARRAY(ConcurrentG1RefineThread*, _threads, mtGC);
ysr@777:   }
ysr@777: }
ysr@777: 
iveresov@1229: void ConcurrentG1Refine::threads_do(ThreadClosure *tc) {
iveresov@1229:   if (_threads != NULL) {
iveresov@1229:     for (int i = 0; i < _n_threads; i++) {
iveresov@1229:       tc->do_thread(_threads[i]);
iveresov@1229:     }
ysr@777:   }
ysr@777: }
ysr@777: 
johnc@1325: bool ConcurrentG1Refine::is_young_card(jbyte* card_ptr) {
johnc@1325:   HeapWord* start = _ct_bs->addr_for(card_ptr);
johnc@1325:   HeapRegion* r = _g1h->heap_region_containing(start);
johnc@1325:   if (r != NULL && r->is_young()) {
johnc@1325:     return true;
johnc@1325:   }
johnc@1325:   // This card is not associated with a heap region
johnc@1325:   // so can't be young.
johnc@1325:   return false;
ysr@777: }
ysr@777: 
johnc@1325: jbyte* ConcurrentG1Refine::add_card_count(jbyte* card_ptr, int* count, bool* defer) {
johnc@1325:   unsigned new_card_num = ptr_2_card_num(card_ptr);
johnc@1325:   unsigned bucket = hash(new_card_num);
johnc@1325:   assert(0 <= bucket && bucket < _n_card_counts, "Bounds");
johnc@1325: 
johnc@1325:   CardCountCacheEntry* count_ptr = &_card_counts[bucket];
johnc@1325:   CardEpochCacheEntry* epoch_ptr = &_card_epochs[bucket];
johnc@1325: 
johnc@1325:   // We have to construct a new entry if we haven't updated the counts
johnc@1325:   // during the current period, or if the count was updated for a
johnc@1325:   // different card number.
johnc@1325:   unsigned int new_epoch = (unsigned int) _n_periods;
johnc@1325:   julong new_epoch_entry = make_epoch_entry(new_card_num, new_epoch);
johnc@1325: 
johnc@1325:   while (true) {
johnc@1325:     // Fetch the previous epoch value
johnc@1325:     julong prev_epoch_entry = epoch_ptr->_value;
johnc@1325:     julong cas_res;
johnc@1325: 
johnc@1325:     if (extract_epoch(prev_epoch_entry) != new_epoch) {
johnc@1325:       // This entry has not yet been updated during this period.
johnc@1325:       // Note: we update the epoch value atomically to ensure
johnc@1325:       // that there is only one winner that updates the cached
johnc@1325:       // card_ptr value even though all the refine threads share
johnc@1325:       // the same epoch value.
johnc@1325: 
johnc@1325:       cas_res = (julong) Atomic::cmpxchg((jlong) new_epoch_entry,
johnc@1325:                                          (volatile jlong*)&epoch_ptr->_value,
johnc@1325:                                          (jlong) prev_epoch_entry);
johnc@1325: 
johnc@1325:       if (cas_res == prev_epoch_entry) {
johnc@1325:         // We have successfully won the race to update the
johnc@1325:         // epoch and card_num value. Make it look like the
johnc@1325:         // count and eviction count were previously cleared.
johnc@1325:         count_ptr->_count = 1;
johnc@1325:         count_ptr->_evict_count = 0;
johnc@1325:         *count = 0;
johnc@1325:         // We can defer the processing of card_ptr
johnc@1325:         *defer = true;
johnc@1325:         return card_ptr;
johnc@1325:       }
johnc@1325:       // We did not win the race to update the epoch field, so some other
johnc@1325:       // thread must have done it. The value that gets returned by CAS
johnc@1325:       // should be the new epoch value.
johnc@1325:       assert(extract_epoch(cas_res) == new_epoch, "unexpected epoch");
johnc@1325:       // We could 'continue' here or just re-read the previous epoch value
johnc@1325:       prev_epoch_entry = epoch_ptr->_value;
johnc@1325:     }
johnc@1325: 
johnc@1325:     // The epoch entry for card_ptr has been updated during this period.
johnc@1325:     unsigned old_card_num = extract_card_num(prev_epoch_entry);
johnc@1325: 
johnc@1325:     // The card count that will be returned to caller
johnc@1325:     *count = count_ptr->_count;
johnc@1325: 
johnc@1325:     // Are we updating the count for the same card?
johnc@1325:     if (new_card_num == old_card_num) {
johnc@1325:       // Same card - just update the count. We could have more than one
johnc@1325:       // thread racing to update count for the current card. It should be
johnc@1325:       // OK not to use a CAS as the only penalty should be some missed
johnc@1325:       // increments of the count which delays identifying the card as "hot".
johnc@1325: 
johnc@1325:       if (*count < max_jubyte) count_ptr->_count++;
johnc@1325:       // We can defer the processing of card_ptr
johnc@1325:       *defer = true;
johnc@1325:       return card_ptr;
johnc@1325:     }
johnc@1325: 
johnc@1325:     // Different card - evict old card info
johnc@1325:     if (count_ptr->_evict_count < max_jubyte) count_ptr->_evict_count++;
johnc@1325:     if (count_ptr->_evict_count > G1CardCountCacheExpandThreshold) {
johnc@1325:       // Trigger a resize the next time we clear
johnc@1325:       _expand_card_counts = true;
johnc@1325:     }
johnc@1325: 
johnc@1325:     cas_res = (julong) Atomic::cmpxchg((jlong) new_epoch_entry,
johnc@1325:                                        (volatile jlong*)&epoch_ptr->_value,
johnc@1325:                                        (jlong) prev_epoch_entry);
johnc@1325: 
johnc@1325:     if (cas_res == prev_epoch_entry) {
johnc@1325:       // We successfully updated the card num value in the epoch entry
johnc@1325:       count_ptr->_count = 0; // initialize counter for new card num
johnc@2021:       jbyte* old_card_ptr = card_num_2_ptr(old_card_num);
johnc@1325: 
johnc@1325:       // Even though the region containg the card at old_card_num was not
johnc@1325:       // in the young list when old_card_num was recorded in the epoch
johnc@1325:       // cache it could have been added to the free list and subsequently
johnc@2021:       // added to the young list in the intervening time. See CR 6817995.
johnc@2021:       // We do not deal with this case here - it will be handled in
johnc@2021:       // HeapRegion::oops_on_card_seq_iterate_careful after it has been
johnc@2021:       // determined that the region containing the card has been allocated
johnc@2021:       // to, and it's safe to check the young type of the region.
johnc@1325: 
johnc@1325:       // We do not want to defer processing of card_ptr in this case
johnc@1325:       // (we need to refine old_card_ptr and card_ptr)
johnc@1325:       *defer = false;
johnc@1325:       return old_card_ptr;
johnc@1325:     }
johnc@1325:     // Someone else beat us - try again.
johnc@1325:   }
johnc@1325: }
johnc@1325: 
johnc@1325: jbyte* ConcurrentG1Refine::cache_insert(jbyte* card_ptr, bool* defer) {
johnc@1325:   int count;
johnc@1325:   jbyte* cached_ptr = add_card_count(card_ptr, &count, defer);
johnc@1325:   assert(cached_ptr != NULL, "bad cached card ptr");
johnc@1681: 
johnc@2021:   // We've just inserted a card pointer into the card count cache
johnc@2021:   // and got back the card that we just inserted or (evicted) the
johnc@2021:   // previous contents of that count slot.
johnc@1681: 
johnc@2021:   // The card we got back could be in a young region. When the
johnc@2021:   // returned card (if evicted) was originally inserted, we had
johnc@2021:   // determined that its containing region was not young. However
johnc@2021:   // it is possible for the region to be freed during a cleanup
johnc@2021:   // pause, then reallocated and tagged as young which will result
johnc@2021:   // in the returned card residing in a young region.
johnc@2021:   //
johnc@2021:   // We do not deal with this case here - the change from non-young
johnc@2021:   // to young could be observed at any time - it will be handled in
johnc@2021:   // HeapRegion::oops_on_card_seq_iterate_careful after it has been
johnc@2021:   // determined that the region containing the card has been allocated
johnc@2021:   // to.
johnc@1325: 
johnc@1325:   // The card pointer we obtained from card count cache is not hot
johnc@1325:   // so do not store it in the cache; return it for immediate
johnc@1325:   // refining.
ysr@777:   if (count < G1ConcRSHotCardLimit) {
johnc@1325:     return cached_ptr;
ysr@777:   }
johnc@1325: 
johnc@2021:   // Otherwise, the pointer we got from the _card_counts cache is 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:     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.
johnc@1325:   _hot_cache[_hot_cache_idx] = cached_ptr;
ysr@777:   _hot_cache_idx++;
ysr@777:   if (_hot_cache_idx == _hot_cache_size) _hot_cache_idx = 0;
ysr@777:   _n_hot++;
johnc@1325: 
johnc@2021:   // The card obtained from the hot card cache could be in a young
johnc@2021:   // region. See above on how this can happen.
johnc@1325: 
ysr@777:   return res;
ysr@777: }
ysr@777: 
johnc@2060: void ConcurrentG1Refine::clean_up_cache(int worker_i,
johnc@2060:                                         G1RemSet* g1rs,
johnc@2060:                                         DirtyCardQueue* into_cset_dcq) {
ysr@777:   assert(!use_cache(), "cache should be disabled");
johnc@1324:   int start_idx;
johnc@1324: 
johnc@1324:   while ((start_idx = _hot_cache_par_claimed_idx) < _n_hot) { // read once
johnc@1324:     int end_idx = start_idx + _hot_cache_par_chunk_size;
johnc@1324: 
johnc@1324:     if (start_idx ==
johnc@1324:         Atomic::cmpxchg(end_idx, &_hot_cache_par_claimed_idx, start_idx)) {
johnc@1324:       // The current worker has successfully claimed the chunk [start_idx..end_idx)
johnc@1324:       end_idx = MIN2(end_idx, _n_hot);
johnc@1324:       for (int i = start_idx; i < end_idx; i++) {
johnc@1324:         jbyte* entry = _hot_cache[i];
johnc@1324:         if (entry != NULL) {
johnc@2060:           if (g1rs->concurrentRefineOneCard(entry, worker_i, true)) {
johnc@2060:             // 'entry' contains references that point into the current
johnc@2060:             // collection set. We need to record 'entry' in the DCQS
johnc@2060:             // that's used for that purpose.
johnc@2060:             //
johnc@2060:             // The only time we care about recording cards that contain
johnc@2060:             // references that point into the collection set is during
johnc@2060:             // RSet updating while within an evacuation pause.
johnc@2060:             // In this case worker_i should be the id of a GC worker thread
johnc@2060:             assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
brutisso@2646:             assert(worker_i < (int) (ParallelGCThreads == 0 ? 1 : ParallelGCThreads), "incorrect worker id");
johnc@2060:             into_cset_dcq->enqueue(entry);
johnc@2060:           }
johnc@1324:         }
johnc@1324:       }
ysr@777:     }
ysr@777:   }
ysr@777: }
ysr@777: 
johnc@2713: // The arrays used to hold the card counts and the epochs must have
johnc@2713: // a 1:1 correspondence. Hence they are allocated and freed together
johnc@2713: // Returns true if the allocations of both the counts and epochs
johnc@2713: // were successful; false otherwise.
johnc@2713: bool ConcurrentG1Refine::allocate_card_count_cache(size_t n,
johnc@2713:                                                    CardCountCacheEntry** counts,
johnc@2713:                                                    CardEpochCacheEntry** epochs) {
johnc@2713:   // We call the allocation/free routines directly for the counts
johnc@2713:   // and epochs arrays. The NEW_C_HEAP_ARRAY/FREE_C_HEAP_ARRAY
johnc@2713:   // macros call AllocateHeap and FreeHeap respectively.
johnc@2713:   // AllocateHeap will call vm_exit_out_of_memory in the event
johnc@2713:   // of an allocation failure and abort the JVM. With the
johnc@2713:   // _counts/epochs arrays we only need to abort the JVM if the
johnc@2713:   // initial allocation of these arrays fails.
johnc@2713:   //
johnc@2713:   // Additionally AllocateHeap/FreeHeap do some tracing of
johnc@2713:   // allocate/free calls so calling one without calling the
johnc@2713:   // other can cause inconsistencies in the tracing. So we
johnc@2713:   // call neither.
johnc@2713: 
johnc@2713:   assert(*counts == NULL, "out param");
johnc@2713:   assert(*epochs == NULL, "out param");
johnc@2713: 
johnc@2713:   size_t counts_size = n * sizeof(CardCountCacheEntry);
johnc@2713:   size_t epochs_size = n * sizeof(CardEpochCacheEntry);
johnc@2713: 
zgu@3900:   *counts = (CardCountCacheEntry*) os::malloc(counts_size, mtGC);
johnc@2713:   if (*counts == NULL) {
johnc@2713:     // allocation was unsuccessful
johnc@2713:     return false;
johnc@2713:   }
johnc@2713: 
zgu@3900:   *epochs = (CardEpochCacheEntry*) os::malloc(epochs_size, mtGC);
johnc@2713:   if (*epochs == NULL) {
johnc@2713:     // allocation was unsuccessful - free counts array
johnc@2713:     assert(*counts != NULL, "must be");
zgu@3900:     os::free(*counts, mtGC);
johnc@2713:     *counts = NULL;
johnc@2713:     return false;
johnc@2713:   }
johnc@2713: 
johnc@2713:   // We successfully allocated both counts and epochs
johnc@2713:   return true;
johnc@2713: }
johnc@2713: 
johnc@2713: // Returns true if the card counts/epochs cache was
johnc@2713: // successfully expanded; false otherwise.
johnc@2713: bool ConcurrentG1Refine::expand_card_count_cache(int cache_size_idx) {
johnc@2713:   // Can we expand the card count and epoch tables?
johnc@1325:   if (_n_card_counts < _max_n_card_counts) {
johnc@2713:     assert(cache_size_idx >= 0 && cache_size_idx  < MAX_CC_CACHE_INDEX, "oob");
johnc@1325: 
johnc@2713:     size_t cache_size = _cc_cache_sizes[cache_size_idx];
johnc@1325:     // Make sure we don't go bigger than we will ever need
johnc@2713:     cache_size = MIN2(cache_size, _max_n_card_counts);
johnc@1325: 
johnc@2713:     // Should we expand the card count and card epoch tables?
johnc@2713:     if (cache_size > _n_card_counts) {
johnc@2713:       // We have been asked to allocate new, larger, arrays for
johnc@2713:       // the card counts and the epochs. Attempt the allocation
johnc@2713:       // of both before we free the existing arrays in case
johnc@2713:       // the allocation is unsuccessful...
johnc@2713:       CardCountCacheEntry* counts = NULL;
johnc@2713:       CardEpochCacheEntry* epochs = NULL;
johnc@2713: 
johnc@2713:       if (allocate_card_count_cache(cache_size, &counts, &epochs)) {
johnc@2713:         // Allocation was successful.
johnc@2713:         // We can just free the old arrays; we're
johnc@2713:         // not interested in preserving the contents
zgu@3900:         if (_card_counts != NULL) os::free(_card_counts, mtGC);
zgu@3900:         if (_card_epochs != NULL) os::free(_card_epochs, mtGC);
johnc@2713: 
johnc@2713:         // Cache the size of the arrays and the index that got us there.
johnc@2713:         _n_card_counts = cache_size;
johnc@2713:         _cache_size_index = cache_size_idx;
johnc@2713: 
johnc@2713:         _card_counts = counts;
johnc@2713:         _card_epochs = epochs;
johnc@2713: 
johnc@2713:         // We successfully allocated/expanded the caches.
johnc@2713:         return true;
johnc@2713:       }
ysr@777:     }
ysr@777:   }
johnc@2713: 
johnc@2713:   // We did not successfully expand the caches.
johnc@2713:   return false;
ysr@777: }
ysr@777: 
johnc@1325: void ConcurrentG1Refine::clear_and_record_card_counts() {
brutisso@3923:   if (G1ConcRSLogCacheSize == 0) {
brutisso@3923:     return;
brutisso@3923:   }
johnc@1325: 
johnc@1325:   double start = os::elapsedTime();
johnc@1325: 
johnc@1325:   if (_expand_card_counts) {
johnc@2713:     int new_idx = _cache_size_index + 1;
johnc@2713: 
johnc@2713:     if (expand_card_count_cache(new_idx)) {
johnc@2713:       // Allocation was successful and  _n_card_counts has
johnc@2713:       // been updated to the new size. We only need to clear
johnc@2713:       // the epochs so we don't read a bogus epoch value
johnc@2713:       // when inserting a card into the hot card cache.
johnc@2713:       Copy::fill_to_bytes(&_card_epochs[0], _n_card_counts * sizeof(CardEpochCacheEntry));
johnc@2713:     }
johnc@1325:     _expand_card_counts = false;
ysr@777:   }
ysr@777: 
johnc@1325:   int this_epoch = (int) _n_periods;
johnc@1325:   assert((this_epoch+1) <= max_jint, "to many periods");
johnc@1325:   // Update epoch
johnc@1325:   _n_periods++;
brutisso@3923:   double cc_clear_time_ms = (os::elapsedTime() - start) * 1000;
brutisso@3923:   _g1h->g1_policy()->phase_times()->record_cc_clear_time_ms(cc_clear_time_ms);
ysr@777: }
tonyp@1454: 
tonyp@1454: void ConcurrentG1Refine::print_worker_threads_on(outputStream* st) const {
tonyp@1454:   for (int i = 0; i < _n_threads; ++i) {
tonyp@1454:     _threads[i]->print_on(st);
tonyp@1454:     st->cr();
tonyp@1454:   }
tonyp@1454: }