aoqi@0: /* aoqi@0: * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. aoqi@0: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. aoqi@0: * aoqi@0: * This code is free software; you can redistribute it and/or modify it aoqi@0: * under the terms of the GNU General Public License version 2 only, as aoqi@0: * published by the Free Software Foundation. aoqi@0: * aoqi@0: * This code is distributed in the hope that it will be useful, but WITHOUT aoqi@0: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or aoqi@0: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License aoqi@0: * version 2 for more details (a copy is included in the LICENSE file that aoqi@0: * accompanied this code). aoqi@0: * aoqi@0: * You should have received a copy of the GNU General Public License version aoqi@0: * 2 along with this work; if not, write to the Free Software Foundation, aoqi@0: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. aoqi@0: * aoqi@0: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA aoqi@0: * or visit www.oracle.com if you need additional information or have any aoqi@0: * questions. aoqi@0: * aoqi@0: */ aoqi@0: aoqi@0: #include "precompiled.hpp" aoqi@0: #include "gc_implementation/g1/collectionSetChooser.hpp" aoqi@0: #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" aoqi@0: #include "gc_implementation/g1/g1CollectorPolicy.hpp" aoqi@0: #include "gc_implementation/g1/g1ErgoVerbose.hpp" aoqi@0: #include "memory/space.inline.hpp" aoqi@0: aoqi@0: // Even though we don't use the GC efficiency in our heuristics as aoqi@0: // much as we used to, we still order according to GC efficiency. This aoqi@0: // will cause regions with a lot of live objects and large RSets to aoqi@0: // end up at the end of the array. Given that we might skip collecting aoqi@0: // the last few old regions, if after a few mixed GCs the remaining aoqi@0: // have reclaimable bytes under a certain threshold, the hope is that aoqi@0: // the ones we'll skip are ones with both large RSets and a lot of aoqi@0: // live objects, not the ones with just a lot of live objects if we aoqi@0: // ordered according to the amount of reclaimable bytes per region. aoqi@0: static int order_regions(HeapRegion* hr1, HeapRegion* hr2) { aoqi@0: if (hr1 == NULL) { aoqi@0: if (hr2 == NULL) { aoqi@0: return 0; aoqi@0: } else { aoqi@0: return 1; aoqi@0: } aoqi@0: } else if (hr2 == NULL) { aoqi@0: return -1; aoqi@0: } aoqi@0: aoqi@0: double gc_eff1 = hr1->gc_efficiency(); aoqi@0: double gc_eff2 = hr2->gc_efficiency(); aoqi@0: if (gc_eff1 > gc_eff2) { aoqi@0: return -1; aoqi@0: } if (gc_eff1 < gc_eff2) { aoqi@0: return 1; aoqi@0: } else { aoqi@0: return 0; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: static int order_regions(HeapRegion** hr1p, HeapRegion** hr2p) { aoqi@0: return order_regions(*hr1p, *hr2p); aoqi@0: } aoqi@0: aoqi@0: CollectionSetChooser::CollectionSetChooser() : aoqi@0: // The line below is the worst bit of C++ hackery I've ever written aoqi@0: // (Detlefs, 11/23). You should think of it as equivalent to aoqi@0: // "_regions(100, true)": initialize the growable array and inform it aoqi@0: // that it should allocate its elem array(s) on the C heap. aoqi@0: // aoqi@0: // The first argument, however, is actually a comma expression aoqi@0: // (set_allocation_type(this, C_HEAP), 100). The purpose of the aoqi@0: // set_allocation_type() call is to replace the default allocation aoqi@0: // type for embedded objects STACK_OR_EMBEDDED with C_HEAP. It will aoqi@0: // allow to pass the assert in GenericGrowableArray() which checks aoqi@0: // that a growable array object must be on C heap if elements are. aoqi@0: // aoqi@0: // Note: containing object is allocated on C heap since it is CHeapObj. aoqi@0: // aoqi@0: _regions((ResourceObj::set_allocation_type((address) &_regions, aoqi@0: ResourceObj::C_HEAP), aoqi@0: 100), true /* C_Heap */), aoqi@0: _curr_index(0), _length(0), _first_par_unreserved_idx(0), aoqi@0: _region_live_threshold_bytes(0), _remaining_reclaimable_bytes(0) { aoqi@0: _region_live_threshold_bytes = aoqi@0: HeapRegion::GrainBytes * (size_t) G1MixedGCLiveThresholdPercent / 100; aoqi@0: } aoqi@0: aoqi@0: #ifndef PRODUCT aoqi@0: void CollectionSetChooser::verify() { aoqi@0: guarantee(_length <= regions_length(), aoqi@0: err_msg("_length: %u regions length: %u", _length, regions_length())); aoqi@0: guarantee(_curr_index <= _length, aoqi@0: err_msg("_curr_index: %u _length: %u", _curr_index, _length)); aoqi@0: uint index = 0; aoqi@0: size_t sum_of_reclaimable_bytes = 0; aoqi@0: while (index < _curr_index) { aoqi@0: guarantee(regions_at(index) == NULL, aoqi@0: "all entries before _curr_index should be NULL"); aoqi@0: index += 1; aoqi@0: } aoqi@0: HeapRegion *prev = NULL; aoqi@0: while (index < _length) { aoqi@0: HeapRegion *curr = regions_at(index++); aoqi@0: guarantee(curr != NULL, "Regions in _regions array cannot be NULL"); aoqi@0: guarantee(!curr->is_young(), "should not be young!"); aoqi@0: guarantee(!curr->isHumongous(), "should not be humongous!"); aoqi@0: if (prev != NULL) { aoqi@0: guarantee(order_regions(prev, curr) != 1, aoqi@0: err_msg("GC eff prev: %1.4f GC eff curr: %1.4f", aoqi@0: prev->gc_efficiency(), curr->gc_efficiency())); aoqi@0: } aoqi@0: sum_of_reclaimable_bytes += curr->reclaimable_bytes(); aoqi@0: prev = curr; aoqi@0: } aoqi@0: guarantee(sum_of_reclaimable_bytes == _remaining_reclaimable_bytes, aoqi@0: err_msg("reclaimable bytes inconsistent, " kevinw@9327: "remaining: " SIZE_FORMAT " sum: " SIZE_FORMAT, aoqi@0: _remaining_reclaimable_bytes, sum_of_reclaimable_bytes)); aoqi@0: } aoqi@0: #endif // !PRODUCT aoqi@0: aoqi@0: void CollectionSetChooser::sort_regions() { aoqi@0: // First trim any unused portion of the top in the parallel case. aoqi@0: if (_first_par_unreserved_idx > 0) { aoqi@0: assert(_first_par_unreserved_idx <= regions_length(), aoqi@0: "Or we didn't reserved enough length"); aoqi@0: regions_trunc_to(_first_par_unreserved_idx); aoqi@0: } aoqi@0: _regions.sort(order_regions); aoqi@0: assert(_length <= regions_length(), "Requirement"); aoqi@0: #ifdef ASSERT aoqi@0: for (uint i = 0; i < _length; i++) { aoqi@0: assert(regions_at(i) != NULL, "Should be true by sorting!"); aoqi@0: } aoqi@0: #endif // ASSERT aoqi@0: if (G1PrintRegionLivenessInfo) { aoqi@0: G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Sorting"); aoqi@0: for (uint i = 0; i < _length; ++i) { aoqi@0: HeapRegion* r = regions_at(i); aoqi@0: cl.doHeapRegion(r); aoqi@0: } aoqi@0: } aoqi@0: verify(); aoqi@0: } aoqi@0: aoqi@0: aoqi@0: void CollectionSetChooser::add_region(HeapRegion* hr) { aoqi@0: assert(!hr->isHumongous(), aoqi@0: "Humongous regions shouldn't be added to the collection set"); aoqi@0: assert(!hr->is_young(), "should not be young!"); aoqi@0: _regions.append(hr); aoqi@0: _length++; aoqi@0: _remaining_reclaimable_bytes += hr->reclaimable_bytes(); aoqi@0: hr->calc_gc_efficiency(); aoqi@0: } aoqi@0: aoqi@0: void CollectionSetChooser::prepare_for_par_region_addition(uint n_regions, aoqi@0: uint chunk_size) { aoqi@0: _first_par_unreserved_idx = 0; aoqi@0: uint n_threads = (uint) ParallelGCThreads; aoqi@0: if (UseDynamicNumberOfGCThreads) { aoqi@0: assert(G1CollectedHeap::heap()->workers()->active_workers() > 0, aoqi@0: "Should have been set earlier"); aoqi@0: // This is defensive code. As the assertion above says, the number aoqi@0: // of active threads should be > 0, but in case there is some path aoqi@0: // or some improperly initialized variable with leads to no aoqi@0: // active threads, protect against that in a product build. aoqi@0: n_threads = MAX2(G1CollectedHeap::heap()->workers()->active_workers(), aoqi@0: 1U); aoqi@0: } aoqi@0: uint max_waste = n_threads * chunk_size; aoqi@0: // it should be aligned with respect to chunk_size aoqi@0: uint aligned_n_regions = (n_regions + chunk_size - 1) / chunk_size * chunk_size; aoqi@0: assert(aligned_n_regions % chunk_size == 0, "should be aligned"); aoqi@0: regions_at_put_grow(aligned_n_regions + max_waste - 1, NULL); aoqi@0: } aoqi@0: aoqi@0: uint CollectionSetChooser::claim_array_chunk(uint chunk_size) { aoqi@0: uint res = (uint) Atomic::add((jint) chunk_size, aoqi@0: (volatile jint*) &_first_par_unreserved_idx); aoqi@0: assert(regions_length() > res + chunk_size - 1, aoqi@0: "Should already have been expanded"); aoqi@0: return res - chunk_size; aoqi@0: } aoqi@0: aoqi@0: void CollectionSetChooser::set_region(uint index, HeapRegion* hr) { aoqi@0: assert(regions_at(index) == NULL, "precondition"); aoqi@0: assert(!hr->is_young(), "should not be young!"); aoqi@0: regions_at_put(index, hr); aoqi@0: hr->calc_gc_efficiency(); aoqi@0: } aoqi@0: aoqi@0: void CollectionSetChooser::update_totals(uint region_num, aoqi@0: size_t reclaimable_bytes) { aoqi@0: // Only take the lock if we actually need to update the totals. aoqi@0: if (region_num > 0) { aoqi@0: assert(reclaimable_bytes > 0, "invariant"); aoqi@0: // We could have just used atomics instead of taking the aoqi@0: // lock. However, we currently don't have an atomic add for size_t. aoqi@0: MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); aoqi@0: _length += region_num; aoqi@0: _remaining_reclaimable_bytes += reclaimable_bytes; aoqi@0: } else { aoqi@0: assert(reclaimable_bytes == 0, "invariant"); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: void CollectionSetChooser::clear() { aoqi@0: _regions.clear(); aoqi@0: _curr_index = 0; aoqi@0: _length = 0; aoqi@0: _remaining_reclaimable_bytes = 0; aoqi@0: };