duke@435: /* johnc@3538: * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * 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. duke@435: * duke@435: */ duke@435: stefank@2314: #include "precompiled.hpp" stefank@2314: #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp" stefank@2314: #include "gc_implementation/parNew/parNewGeneration.hpp" stefank@2314: #include "gc_implementation/parNew/parOopClosures.inline.hpp" stefank@2314: #include "gc_implementation/shared/adaptiveSizePolicy.hpp" stefank@2314: #include "gc_implementation/shared/ageTable.hpp" johnc@3982: #include "gc_implementation/shared/parGCAllocBuffer.hpp" stefank@2314: #include "gc_implementation/shared/spaceDecorator.hpp" stefank@2314: #include "memory/defNewGeneration.inline.hpp" stefank@2314: #include "memory/genCollectedHeap.hpp" stefank@2314: #include "memory/genOopClosures.inline.hpp" stefank@2314: #include "memory/generation.hpp" stefank@2314: #include "memory/generation.inline.hpp" stefank@2314: #include "memory/referencePolicy.hpp" stefank@2314: #include "memory/resourceArea.hpp" stefank@2314: #include "memory/sharedHeap.hpp" stefank@2314: #include "memory/space.hpp" stefank@2314: #include "oops/objArrayOop.hpp" stefank@2314: #include "oops/oop.inline.hpp" stefank@2314: #include "oops/oop.pcgc.inline.hpp" stefank@2314: #include "runtime/handles.hpp" stefank@2314: #include "runtime/handles.inline.hpp" stefank@2314: #include "runtime/java.hpp" stefank@2314: #include "runtime/thread.hpp" stefank@2314: #include "utilities/copy.hpp" stefank@2314: #include "utilities/globalDefinitions.hpp" stefank@2314: #include "utilities/workgroup.hpp" duke@435: duke@435: #ifdef _MSC_VER duke@435: #pragma warning( push ) duke@435: #pragma warning( disable:4355 ) // 'this' : used in base member initializer list duke@435: #endif duke@435: ParScanThreadState::ParScanThreadState(Space* to_space_, duke@435: ParNewGeneration* gen_, duke@435: Generation* old_gen_, duke@435: int thread_num_, duke@435: ObjToScanQueueSet* work_queue_set_, zgu@3900: Stack* overflow_stacks_, duke@435: size_t desired_plab_sz_, duke@435: ParallelTaskTerminator& term_) : ysr@1114: _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_), duke@435: _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false), jcoomes@2191: _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL), duke@435: _ageTable(false), // false ==> not the global age table, no perf data. duke@435: _to_space_alloc_buffer(desired_plab_sz_), duke@435: _to_space_closure(gen_, this), _old_gen_closure(gen_, this), duke@435: _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this), duke@435: _older_gen_closure(gen_, this), duke@435: _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, duke@435: &_to_space_root_closure, gen_, &_old_gen_root_closure, duke@435: work_queue_set_, &term_), duke@435: _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this), duke@435: _keep_alive_closure(&_scan_weak_ref_closure), ysr@1580: _promotion_failure_size(0), duke@435: _strong_roots_time(0.0), _term_time(0.0) duke@435: { jcoomes@2065: #if TASKQUEUE_STATS jcoomes@2065: _term_attempts = 0; jcoomes@2065: _overflow_refills = 0; jcoomes@2065: _overflow_refill_objs = 0; jcoomes@2065: #endif // TASKQUEUE_STATS jcoomes@2065: duke@435: _survivor_chunk_array = duke@435: (ChunkArray*) old_gen()->get_data_recorder(thread_num()); duke@435: _hash_seed = 17; // Might want to take time-based random value. duke@435: _start = os::elapsedTime(); duke@435: _old_gen_closure.set_generation(old_gen_); duke@435: _old_gen_root_closure.set_generation(old_gen_); duke@435: } duke@435: #ifdef _MSC_VER duke@435: #pragma warning( pop ) duke@435: #endif duke@435: duke@435: void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, duke@435: size_t plab_word_size) { duke@435: ChunkArray* sca = survivor_chunk_array(); duke@435: if (sca != NULL) { duke@435: // A non-null SCA implies that we want the PLAB data recorded. duke@435: sca->record_sample(plab_start, plab_word_size); duke@435: } duke@435: } duke@435: duke@435: bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { duke@435: return new_obj->is_objArray() && duke@435: arrayOop(new_obj)->length() > ParGCArrayScanChunk && duke@435: new_obj != old_obj; duke@435: } duke@435: duke@435: void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { duke@435: assert(old->is_objArray(), "must be obj array"); duke@435: assert(old->is_forwarded(), "must be forwarded"); duke@435: assert(Universe::heap()->is_in_reserved(old), "must be in heap."); ysr@1114: assert(!old_gen()->is_in(old), "must be in young generation."); duke@435: duke@435: objArrayOop obj = objArrayOop(old->forwardee()); duke@435: // Process ParGCArrayScanChunk elements now duke@435: // and push the remainder back onto queue duke@435: int start = arrayOop(old)->length(); duke@435: int end = obj->length(); duke@435: int remainder = end - start; duke@435: assert(start <= end, "just checking"); duke@435: if (remainder > 2 * ParGCArrayScanChunk) { duke@435: // Test above combines last partial chunk with a full chunk duke@435: end = start + ParGCArrayScanChunk; duke@435: arrayOop(old)->set_length(end); duke@435: // Push remainder. duke@435: bool ok = work_queue()->push(old); duke@435: assert(ok, "just popped, push must be okay"); duke@435: } else { duke@435: // Restore length so that it can be used if there duke@435: // is a promotion failure and forwarding pointers duke@435: // must be removed. duke@435: arrayOop(old)->set_length(end); duke@435: } coleenp@548: duke@435: // process our set of indices (include header in first chunk) coleenp@548: // should make sure end is even (aligned to HeapWord in case of compressed oops) duke@435: if ((HeapWord *)obj < young_old_boundary()) { duke@435: // object is in to_space coleenp@548: obj->oop_iterate_range(&_to_space_closure, start, end); duke@435: } else { duke@435: // object is in old generation coleenp@548: obj->oop_iterate_range(&_old_gen_closure, start, end); duke@435: } duke@435: } duke@435: duke@435: duke@435: void ParScanThreadState::trim_queues(int max_size) { duke@435: ObjToScanQueue* queue = work_queue(); ysr@1114: do { ysr@1114: while (queue->size() > (juint)max_size) { ysr@1114: oop obj_to_scan; ysr@1114: if (queue->pop_local(obj_to_scan)) { ysr@1114: if ((HeapWord *)obj_to_scan < young_old_boundary()) { ysr@1114: if (obj_to_scan->is_objArray() && ysr@1114: obj_to_scan->is_forwarded() && ysr@1114: obj_to_scan->forwardee() != obj_to_scan) { ysr@1114: scan_partial_array_and_push_remainder(obj_to_scan); ysr@1114: } else { ysr@1114: // object is in to_space ysr@1114: obj_to_scan->oop_iterate(&_to_space_closure); ysr@1114: } duke@435: } else { ysr@1114: // object is in old generation ysr@1114: obj_to_scan->oop_iterate(&_old_gen_closure); duke@435: } duke@435: } duke@435: } ysr@1114: // For the case of compressed oops, we have a private, non-shared ysr@1114: // overflow stack, so we eagerly drain it so as to more evenly ysr@1114: // distribute load early. Note: this may be good to do in ysr@1114: // general rather than delay for the final stealing phase. ysr@1114: // If applicable, we'll transfer a set of objects over to our ysr@1114: // work queue, allowing them to be stolen and draining our ysr@1114: // private overflow stack. ysr@1114: } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this)); ysr@1114: } ysr@1114: ysr@1114: bool ParScanThreadState::take_from_overflow_stack() { ysr@1130: assert(ParGCUseLocalOverflow, "Else should not call"); ysr@1114: assert(young_gen()->overflow_list() == NULL, "Error"); ysr@1114: ObjToScanQueue* queue = work_queue(); zgu@3900: Stack* const of_stack = overflow_stack(); jcoomes@2191: const size_t num_overflow_elems = of_stack->size(); jcoomes@2191: const size_t space_available = queue->max_elems() - queue->size(); jcoomes@2191: const size_t num_take_elems = MIN3(space_available / 4, jcoomes@2191: ParGCDesiredObjsFromOverflowList, jcoomes@2191: num_overflow_elems); ysr@1114: // Transfer the most recent num_take_elems from the overflow ysr@1114: // stack to our work queue. ysr@1114: for (size_t i = 0; i != num_take_elems; i++) { ysr@1114: oop cur = of_stack->pop(); ysr@1114: oop obj_to_push = cur->forwardee(); ysr@1114: assert(Universe::heap()->is_in_reserved(cur), "Should be in heap"); ysr@1114: assert(!old_gen()->is_in_reserved(cur), "Should be in young gen"); ysr@1114: assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap"); ysr@1114: if (should_be_partially_scanned(obj_to_push, cur)) { ysr@1114: assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); ysr@1114: obj_to_push = cur; ysr@1114: } ysr@1114: bool ok = queue->push(obj_to_push); ysr@1114: assert(ok, "Should have succeeded"); duke@435: } ysr@1114: assert(young_gen()->overflow_list() == NULL, "Error"); ysr@1114: return num_take_elems > 0; // was something transferred? ysr@1114: } ysr@1114: ysr@1114: void ParScanThreadState::push_on_overflow_stack(oop p) { ysr@1130: assert(ParGCUseLocalOverflow, "Else should not call"); ysr@1114: overflow_stack()->push(p); ysr@1114: assert(young_gen()->overflow_list() == NULL, "Error"); duke@435: } duke@435: duke@435: HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { duke@435: duke@435: // Otherwise, if the object is small enough, try to reallocate the duke@435: // buffer. duke@435: HeapWord* obj = NULL; duke@435: if (!_to_space_full) { duke@435: ParGCAllocBuffer* const plab = to_space_alloc_buffer(); duke@435: Space* const sp = to_space(); duke@435: if (word_sz * 100 < duke@435: ParallelGCBufferWastePct * plab->word_sz()) { duke@435: // Is small enough; abandon this buffer and start a new one. duke@435: plab->retire(false, false); duke@435: size_t buf_size = plab->word_sz(); duke@435: HeapWord* buf_space = sp->par_allocate(buf_size); duke@435: if (buf_space == NULL) { duke@435: const size_t min_bytes = duke@435: ParGCAllocBuffer::min_size() << LogHeapWordSize; duke@435: size_t free_bytes = sp->free(); duke@435: while(buf_space == NULL && free_bytes >= min_bytes) { duke@435: buf_size = free_bytes >> LogHeapWordSize; duke@435: assert(buf_size == (size_t)align_object_size(buf_size), duke@435: "Invariant"); duke@435: buf_space = sp->par_allocate(buf_size); duke@435: free_bytes = sp->free(); duke@435: } duke@435: } duke@435: if (buf_space != NULL) { duke@435: plab->set_word_size(buf_size); duke@435: plab->set_buf(buf_space); duke@435: record_survivor_plab(buf_space, buf_size); duke@435: obj = plab->allocate(word_sz); duke@435: // Note that we cannot compare buf_size < word_sz below duke@435: // because of AlignmentReserve (see ParGCAllocBuffer::allocate()). duke@435: assert(obj != NULL || plab->words_remaining() < word_sz, duke@435: "Else should have been able to allocate"); duke@435: // It's conceivable that we may be able to use the duke@435: // buffer we just grabbed for subsequent small requests duke@435: // even if not for this one. duke@435: } else { duke@435: // We're used up. duke@435: _to_space_full = true; duke@435: } duke@435: duke@435: } else { duke@435: // Too large; allocate the object individually. duke@435: obj = sp->par_allocate(word_sz); duke@435: } duke@435: } duke@435: return obj; duke@435: } duke@435: duke@435: duke@435: void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, duke@435: size_t word_sz) { duke@435: // Is the alloc in the current alloc buffer? duke@435: if (to_space_alloc_buffer()->contains(obj)) { duke@435: assert(to_space_alloc_buffer()->contains(obj + word_sz - 1), duke@435: "Should contain whole object."); duke@435: to_space_alloc_buffer()->undo_allocation(obj, word_sz); duke@435: } else { jcoomes@916: CollectedHeap::fill_with_object(obj, word_sz); duke@435: } duke@435: } duke@435: ysr@1580: void ParScanThreadState::print_and_clear_promotion_failure_size() { ysr@1580: if (_promotion_failure_size != 0) { ysr@1580: if (PrintPromotionFailure) { ysr@1580: gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ", ysr@1580: _thread_num, _promotion_failure_size); ysr@1580: } ysr@1580: _promotion_failure_size = 0; ysr@1580: } ysr@1580: } ysr@1580: duke@435: class ParScanThreadStateSet: private ResourceArray { duke@435: public: duke@435: // Initializes states for the specified number of threads; duke@435: ParScanThreadStateSet(int num_threads, duke@435: Space& to_space, duke@435: ParNewGeneration& gen, duke@435: Generation& old_gen, duke@435: ObjToScanQueueSet& queue_set, zgu@3900: Stack* overflow_stacks_, duke@435: size_t desired_plab_sz, duke@435: ParallelTaskTerminator& term); jcoomes@2065: jcoomes@2065: ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } jcoomes@2065: ysr@1580: inline ParScanThreadState& thread_state(int i); jcoomes@2065: jmasa@3294: void reset(int active_workers, bool promotion_failed); duke@435: void flush(); jcoomes@2065: jcoomes@2065: #if TASKQUEUE_STATS jcoomes@2065: static void jcoomes@2065: print_termination_stats_hdr(outputStream* const st = gclog_or_tty); jcoomes@2065: void print_termination_stats(outputStream* const st = gclog_or_tty); jcoomes@2065: static void jcoomes@2065: print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty); jcoomes@2065: void print_taskqueue_stats(outputStream* const st = gclog_or_tty); jcoomes@2065: void reset_stats(); jcoomes@2065: #endif // TASKQUEUE_STATS jcoomes@2065: duke@435: private: duke@435: ParallelTaskTerminator& _term; duke@435: ParNewGeneration& _gen; duke@435: Generation& _next_gen; jmasa@3294: public: jmasa@3294: bool is_valid(int id) const { return id < length(); } jmasa@3294: ParallelTaskTerminator* terminator() { return &_term; } duke@435: }; duke@435: duke@435: duke@435: ParScanThreadStateSet::ParScanThreadStateSet( duke@435: int num_threads, Space& to_space, ParNewGeneration& gen, duke@435: Generation& old_gen, ObjToScanQueueSet& queue_set, zgu@3900: Stack* overflow_stacks, duke@435: size_t desired_plab_sz, ParallelTaskTerminator& term) duke@435: : ResourceArray(sizeof(ParScanThreadState), num_threads), jcoomes@2065: _gen(gen), _next_gen(old_gen), _term(term) duke@435: { duke@435: assert(num_threads > 0, "sanity check!"); jcoomes@2191: assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), jcoomes@2191: "overflow_stack allocation mismatch"); duke@435: // Initialize states. duke@435: for (int i = 0; i < num_threads; ++i) { duke@435: new ((ParScanThreadState*)_data + i) duke@435: ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, jcoomes@2191: overflow_stacks, desired_plab_sz, term); duke@435: } duke@435: } duke@435: ysr@1580: inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) duke@435: { duke@435: assert(i >= 0 && i < length(), "sanity check!"); duke@435: return ((ParScanThreadState*)_data)[i]; duke@435: } duke@435: duke@435: jmasa@3294: void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed) duke@435: { jmasa@3294: _term.reset_for_reuse(active_threads); ysr@1580: if (promotion_failed) { ysr@1580: for (int i = 0; i < length(); ++i) { ysr@1580: thread_state(i).print_and_clear_promotion_failure_size(); ysr@1580: } ysr@1580: } duke@435: } duke@435: jcoomes@2065: #if TASKQUEUE_STATS jcoomes@2065: void jcoomes@2065: ParScanThreadState::reset_stats() jcoomes@2065: { jcoomes@2065: taskqueue_stats().reset(); jcoomes@2065: _term_attempts = 0; jcoomes@2065: _overflow_refills = 0; jcoomes@2065: _overflow_refill_objs = 0; jcoomes@2065: } jcoomes@2065: jcoomes@2065: void ParScanThreadStateSet::reset_stats() jcoomes@2065: { jcoomes@2065: for (int i = 0; i < length(); ++i) { jcoomes@2065: thread_state(i).reset_stats(); jcoomes@2065: } jcoomes@2065: } jcoomes@2065: jcoomes@2065: void jcoomes@2065: ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) jcoomes@2065: { jcoomes@2065: st->print_raw_cr("GC Termination Stats"); jcoomes@2065: st->print_raw_cr(" elapsed --strong roots-- " jcoomes@2065: "-------termination-------"); jcoomes@2065: st->print_raw_cr("thr ms ms % " jcoomes@2065: " ms % attempts"); jcoomes@2065: st->print_raw_cr("--- --------- --------- ------ " jcoomes@2065: "--------- ------ --------"); jcoomes@2065: } jcoomes@2065: jcoomes@2065: void ParScanThreadStateSet::print_termination_stats(outputStream* const st) jcoomes@2065: { jcoomes@2065: print_termination_stats_hdr(st); jcoomes@2065: jcoomes@2065: for (int i = 0; i < length(); ++i) { jcoomes@2065: const ParScanThreadState & pss = thread_state(i); jcoomes@2065: const double elapsed_ms = pss.elapsed_time() * 1000.0; jcoomes@2065: const double s_roots_ms = pss.strong_roots_time() * 1000.0; jcoomes@2065: const double term_ms = pss.term_time() * 1000.0; jcoomes@2065: st->print_cr("%3d %9.2f %9.2f %6.2f " jcoomes@2065: "%9.2f %6.2f " SIZE_FORMAT_W(8), jcoomes@2065: i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, jcoomes@2065: term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); jcoomes@2065: } jcoomes@2065: } jcoomes@2065: jcoomes@2065: // Print stats related to work queue activity. jcoomes@2065: void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) jcoomes@2065: { jcoomes@2065: st->print_raw_cr("GC Task Stats"); jcoomes@2065: st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); jcoomes@2065: st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); jcoomes@2065: } jcoomes@2065: jcoomes@2065: void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) jcoomes@2065: { jcoomes@2065: print_taskqueue_stats_hdr(st); jcoomes@2065: jcoomes@2065: TaskQueueStats totals; jcoomes@2065: for (int i = 0; i < length(); ++i) { jcoomes@2065: const ParScanThreadState & pss = thread_state(i); jcoomes@2065: const TaskQueueStats & stats = pss.taskqueue_stats(); jcoomes@2065: st->print("%3d ", i); stats.print(st); st->cr(); jcoomes@2065: totals += stats; jcoomes@2065: jcoomes@2065: if (pss.overflow_refills() > 0) { jcoomes@2065: st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " jcoomes@2065: SIZE_FORMAT_W(10) " overflow objects", jcoomes@2065: pss.overflow_refills(), pss.overflow_refill_objs()); jcoomes@2065: } jcoomes@2065: } jcoomes@2065: st->print("tot "); totals.print(st); st->cr(); jcoomes@2065: jcoomes@2065: DEBUG_ONLY(totals.verify()); jcoomes@2065: } jcoomes@2065: #endif // TASKQUEUE_STATS jcoomes@2065: duke@435: void ParScanThreadStateSet::flush() duke@435: { ysr@1580: // Work in this loop should be kept as lightweight as ysr@1580: // possible since this might otherwise become a bottleneck ysr@1580: // to scaling. Should we add heavy-weight work into this ysr@1580: // loop, consider parallelizing the loop into the worker threads. duke@435: for (int i = 0; i < length(); ++i) { ysr@1580: ParScanThreadState& par_scan_state = thread_state(i); duke@435: duke@435: // Flush stats related to To-space PLAB activity and duke@435: // retire the last buffer. duke@435: par_scan_state.to_space_alloc_buffer()-> duke@435: flush_stats_and_retire(_gen.plab_stats(), johnc@3982: true /* end_of_gc */, johnc@3982: false /* retain */); duke@435: duke@435: // Every thread has its own age table. We need to merge duke@435: // them all into one. duke@435: ageTable *local_table = par_scan_state.age_table(); duke@435: _gen.age_table()->merge(local_table); duke@435: duke@435: // Inform old gen that we're done. duke@435: _next_gen.par_promote_alloc_done(i); duke@435: _next_gen.par_oop_since_save_marks_iterate_done(i); jcoomes@2065: } duke@435: ysr@1580: if (UseConcMarkSweepGC && ParallelGCThreads > 0) { ysr@1580: // We need to call this even when ResizeOldPLAB is disabled ysr@1580: // so as to avoid breaking some asserts. While we may be able ysr@1580: // to avoid this by reorganizing the code a bit, I am loathe ysr@1580: // to do that unless we find cases where ergo leads to bad ysr@1580: // performance. ysr@1580: CFLS_LAB::compute_desired_plab_size(); ysr@1580: } duke@435: } duke@435: duke@435: ParScanClosure::ParScanClosure(ParNewGeneration* g, duke@435: ParScanThreadState* par_scan_state) : coleenp@4037: OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) duke@435: { duke@435: assert(_g->level() == 0, "Optimized for youngest generation"); duke@435: _boundary = _g->reserved().end(); duke@435: } duke@435: coleenp@548: void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); } coleenp@548: void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); } coleenp@548: coleenp@548: void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); } coleenp@548: void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); } coleenp@548: coleenp@548: void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } coleenp@548: void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } coleenp@548: coleenp@548: void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } coleenp@548: void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } coleenp@548: duke@435: ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, duke@435: ParScanThreadState* par_scan_state) duke@435: : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) coleenp@548: {} coleenp@548: coleenp@548: void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); } coleenp@548: void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } duke@435: duke@435: #ifdef WIN32 duke@435: #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ duke@435: #endif duke@435: duke@435: ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( duke@435: ParScanThreadState* par_scan_state_, duke@435: ParScanWithoutBarrierClosure* to_space_closure_, duke@435: ParScanWithBarrierClosure* old_gen_closure_, duke@435: ParRootScanWithoutBarrierClosure* to_space_root_closure_, duke@435: ParNewGeneration* par_gen_, duke@435: ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, duke@435: ObjToScanQueueSet* task_queues_, duke@435: ParallelTaskTerminator* terminator_) : duke@435: duke@435: _par_scan_state(par_scan_state_), duke@435: _to_space_closure(to_space_closure_), duke@435: _old_gen_closure(old_gen_closure_), duke@435: _to_space_root_closure(to_space_root_closure_), duke@435: _old_gen_root_closure(old_gen_root_closure_), duke@435: _par_gen(par_gen_), duke@435: _task_queues(task_queues_), duke@435: _terminator(terminator_) duke@435: {} duke@435: duke@435: void ParEvacuateFollowersClosure::do_void() { duke@435: ObjToScanQueue* work_q = par_scan_state()->work_queue(); duke@435: duke@435: while (true) { duke@435: duke@435: // Scan to-space and old-gen objs until we run out of both. duke@435: oop obj_to_scan; duke@435: par_scan_state()->trim_queues(0); duke@435: duke@435: // We have no local work, attempt to steal from other threads. duke@435: duke@435: // attempt to steal work from promoted. duke@435: if (task_queues()->steal(par_scan_state()->thread_num(), duke@435: par_scan_state()->hash_seed(), duke@435: obj_to_scan)) { duke@435: bool res = work_q->push(obj_to_scan); duke@435: assert(res, "Empty queue should have room for a push."); duke@435: duke@435: // if successful, goto Start. duke@435: continue; duke@435: duke@435: // try global overflow list. duke@435: } else if (par_gen()->take_from_overflow_list(par_scan_state())) { duke@435: continue; duke@435: } duke@435: duke@435: // Otherwise, offer termination. duke@435: par_scan_state()->start_term_time(); duke@435: if (terminator()->offer_termination()) break; duke@435: par_scan_state()->end_term_time(); duke@435: } ysr@969: assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, ysr@969: "Broken overflow list?"); duke@435: // Finish the last termination pause. duke@435: par_scan_state()->end_term_time(); duke@435: } duke@435: duke@435: ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen, duke@435: HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) : duke@435: AbstractGangTask("ParNewGeneration collection"), duke@435: _gen(gen), _next_gen(next_gen), duke@435: _young_old_boundary(young_old_boundary), duke@435: _state_set(state_set) duke@435: {} duke@435: jmasa@3294: // Reset the terminator for the given number of jmasa@3294: // active threads. jmasa@3294: void ParNewGenTask::set_for_termination(int active_workers) { jmasa@3294: _state_set->reset(active_workers, _gen->promotion_failed()); jmasa@3294: // Should the heap be passed in? There's only 1 for now so jmasa@3294: // grab it instead. jmasa@3294: GenCollectedHeap* gch = GenCollectedHeap::heap(); jmasa@3294: gch->set_n_termination(active_workers); jmasa@3294: } jmasa@3294: jmasa@3294: // The "i" passed to this method is the part of the work for jmasa@3294: // this thread. It is not the worker ID. The "i" is derived jmasa@3294: // from _started_workers which is incremented in internal_note_start() jmasa@3294: // called in GangWorker loop() and which is called under the jmasa@3294: // which is called under the protection of the gang monitor and is jmasa@3294: // called after a task is started. So "i" is based on jmasa@3294: // first-come-first-served. jmasa@3294: jmasa@3357: void ParNewGenTask::work(uint worker_id) { duke@435: GenCollectedHeap* gch = GenCollectedHeap::heap(); duke@435: // Since this is being done in a separate thread, need new resource duke@435: // and handle marks. duke@435: ResourceMark rm; duke@435: HandleMark hm; duke@435: // We would need multiple old-gen queues otherwise. ysr@1114: assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen."); duke@435: duke@435: Generation* old_gen = gch->next_gen(_gen); duke@435: jmasa@3357: ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); jmasa@3357: assert(_state_set->is_valid(worker_id), "Should not have been called"); jmasa@3294: duke@435: par_scan_state.set_young_old_boundary(_young_old_boundary); duke@435: coleenp@4037: KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(), coleenp@4037: gch->rem_set()->klass_rem_set()); coleenp@4037: coleenp@4037: int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; coleenp@4037: duke@435: par_scan_state.start_strong_roots(); duke@435: gch->gen_process_strong_roots(_gen->level(), jrose@1424: true, // Process younger gens, if any, jrose@1424: // as strong roots. jrose@1424: false, // no scope; this is parallel code coleenp@4037: true, // is scavenging coleenp@4037: SharedHeap::ScanningOption(so), jrose@1424: &par_scan_state.to_space_root_closure(), jrose@1424: true, // walk *all* scavengable nmethods coleenp@4037: &par_scan_state.older_gen_closure(), coleenp@4037: &klass_scan_closure); duke@435: par_scan_state.end_strong_roots(); duke@435: duke@435: // "evacuate followers". duke@435: par_scan_state.evacuate_followers_closure().do_void(); duke@435: } duke@435: duke@435: #ifdef _MSC_VER duke@435: #pragma warning( push ) duke@435: #pragma warning( disable:4355 ) // 'this' : used in base member initializer list duke@435: #endif duke@435: ParNewGeneration:: duke@435: ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level) duke@435: : DefNewGeneration(rs, initial_byte_size, level, "PCopy"), duke@435: _overflow_list(NULL), duke@435: _is_alive_closure(this), duke@435: _plab_stats(YoungPLABSize, PLABWeight) duke@435: { ysr@969: NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) ysr@969: NOT_PRODUCT(_num_par_pushes = 0;) duke@435: _task_queues = new ObjToScanQueueSet(ParallelGCThreads); duke@435: guarantee(_task_queues != NULL, "task_queues allocation failure."); duke@435: duke@435: for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { jcoomes@2020: ObjToScanQueue *q = new ObjToScanQueue(); jcoomes@2020: guarantee(q != NULL, "work_queue Allocation failure."); jcoomes@2020: _task_queues->register_queue(i1, q); duke@435: } duke@435: duke@435: for (uint i2 = 0; i2 < ParallelGCThreads; i2++) duke@435: _task_queues->queue(i2)->initialize(); duke@435: jcoomes@2191: _overflow_stacks = NULL; jcoomes@2191: if (ParGCUseLocalOverflow) { zgu@3900: zgu@3900: // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal zgu@3900: // with ',' zgu@3900: typedef Stack GCOopStack; zgu@3900: zgu@3900: _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); jcoomes@2191: for (size_t i = 0; i < ParallelGCThreads; ++i) { zgu@3900: new (_overflow_stacks + i) Stack(); ysr@1130: } ysr@1130: } ysr@1130: duke@435: if (UsePerfData) { duke@435: EXCEPTION_MARK; duke@435: ResourceMark rm; duke@435: duke@435: const char* cname = duke@435: PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); duke@435: PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, duke@435: ParallelGCThreads, CHECK); duke@435: } duke@435: } duke@435: #ifdef _MSC_VER duke@435: #pragma warning( pop ) duke@435: #endif duke@435: duke@435: // ParNewGeneration:: duke@435: ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : duke@435: DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} duke@435: coleenp@548: template coleenp@548: void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { coleenp@548: #ifdef ASSERT coleenp@548: { coleenp@548: assert(!oopDesc::is_null(*p), "expected non-null ref"); coleenp@548: oop obj = oopDesc::load_decode_heap_oop_not_null(p); coleenp@548: // We never expect to see a null reference being processed coleenp@548: // as a weak reference. coleenp@548: assert(obj->is_oop(), "expected an oop while scanning weak refs"); coleenp@548: } coleenp@548: #endif // ASSERT duke@435: duke@435: _par_cl->do_oop_nv(p); duke@435: duke@435: if (Universe::heap()->is_in_reserved(p)) { coleenp@548: oop obj = oopDesc::load_decode_heap_oop_not_null(p); coleenp@548: _rs->write_ref_field_gc_par(p, obj); duke@435: } duke@435: } duke@435: coleenp@548: void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } coleenp@548: void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } coleenp@548: duke@435: // ParNewGeneration:: duke@435: KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : duke@435: DefNewGeneration::KeepAliveClosure(cl) {} duke@435: coleenp@548: template coleenp@548: void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { coleenp@548: #ifdef ASSERT coleenp@548: { coleenp@548: assert(!oopDesc::is_null(*p), "expected non-null ref"); coleenp@548: oop obj = oopDesc::load_decode_heap_oop_not_null(p); coleenp@548: // We never expect to see a null reference being processed coleenp@548: // as a weak reference. coleenp@548: assert(obj->is_oop(), "expected an oop while scanning weak refs"); coleenp@548: } coleenp@548: #endif // ASSERT duke@435: duke@435: _cl->do_oop_nv(p); duke@435: duke@435: if (Universe::heap()->is_in_reserved(p)) { coleenp@548: oop obj = oopDesc::load_decode_heap_oop_not_null(p); coleenp@548: _rs->write_ref_field_gc_par(p, obj); duke@435: } duke@435: } duke@435: coleenp@548: void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } coleenp@548: void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } coleenp@548: coleenp@548: template void ScanClosureWithParBarrier::do_oop_work(T* p) { coleenp@548: T heap_oop = oopDesc::load_heap_oop(p); coleenp@548: if (!oopDesc::is_null(heap_oop)) { coleenp@548: oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); duke@435: if ((HeapWord*)obj < _boundary) { duke@435: assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); coleenp@548: oop new_obj = obj->is_forwarded() coleenp@548: ? obj->forwardee() coleenp@548: : _g->DefNewGeneration::copy_to_survivor_space(obj); coleenp@548: oopDesc::encode_store_heap_oop_not_null(p, new_obj); duke@435: } duke@435: if (_gc_barrier) { duke@435: // If p points to a younger generation, mark the card. duke@435: if ((HeapWord*)obj < _gen_boundary) { duke@435: _rs->write_ref_field_gc_par(p, obj); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: coleenp@548: void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } coleenp@548: void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } coleenp@548: duke@435: class ParNewRefProcTaskProxy: public AbstractGangTask { duke@435: typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; duke@435: public: duke@435: ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen, duke@435: Generation& next_gen, duke@435: HeapWord* young_old_boundary, duke@435: ParScanThreadStateSet& state_set); duke@435: duke@435: private: jmasa@3357: virtual void work(uint worker_id); jmasa@3294: virtual void set_for_termination(int active_workers) { jmasa@3294: _state_set.terminator()->reset_for_reuse(active_workers); jmasa@3294: } duke@435: private: duke@435: ParNewGeneration& _gen; duke@435: ProcessTask& _task; duke@435: Generation& _next_gen; duke@435: HeapWord* _young_old_boundary; duke@435: ParScanThreadStateSet& _state_set; duke@435: }; duke@435: duke@435: ParNewRefProcTaskProxy::ParNewRefProcTaskProxy( duke@435: ProcessTask& task, ParNewGeneration& gen, duke@435: Generation& next_gen, duke@435: HeapWord* young_old_boundary, duke@435: ParScanThreadStateSet& state_set) duke@435: : AbstractGangTask("ParNewGeneration parallel reference processing"), duke@435: _gen(gen), duke@435: _task(task), duke@435: _next_gen(next_gen), duke@435: _young_old_boundary(young_old_boundary), duke@435: _state_set(state_set) duke@435: { duke@435: } duke@435: jmasa@3357: void ParNewRefProcTaskProxy::work(uint worker_id) duke@435: { duke@435: ResourceMark rm; duke@435: HandleMark hm; jmasa@3357: ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); duke@435: par_scan_state.set_young_old_boundary(_young_old_boundary); jmasa@3357: _task.work(worker_id, par_scan_state.is_alive_closure(), duke@435: par_scan_state.keep_alive_closure(), duke@435: par_scan_state.evacuate_followers_closure()); duke@435: } duke@435: duke@435: class ParNewRefEnqueueTaskProxy: public AbstractGangTask { duke@435: typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; duke@435: EnqueueTask& _task; duke@435: duke@435: public: duke@435: ParNewRefEnqueueTaskProxy(EnqueueTask& task) duke@435: : AbstractGangTask("ParNewGeneration parallel reference enqueue"), duke@435: _task(task) duke@435: { } duke@435: jmasa@3357: virtual void work(uint worker_id) duke@435: { jmasa@3357: _task.work(worker_id); duke@435: } duke@435: }; duke@435: duke@435: duke@435: void ParNewRefProcTaskExecutor::execute(ProcessTask& task) duke@435: { duke@435: GenCollectedHeap* gch = GenCollectedHeap::heap(); duke@435: assert(gch->kind() == CollectedHeap::GenCollectedHeap, duke@435: "not a generational heap"); jmasa@3294: FlexibleWorkGang* workers = gch->workers(); duke@435: assert(workers != NULL, "Need parallel worker threads."); jmasa@3294: _state_set.reset(workers->active_workers(), _generation.promotion_failed()); duke@435: ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(), duke@435: _generation.reserved().end(), _state_set); duke@435: workers->run_task(&rp_task); jmasa@3294: _state_set.reset(0 /* bad value in debug if not reset */, jmasa@3294: _generation.promotion_failed()); duke@435: } duke@435: duke@435: void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) duke@435: { duke@435: GenCollectedHeap* gch = GenCollectedHeap::heap(); jmasa@3294: FlexibleWorkGang* workers = gch->workers(); duke@435: assert(workers != NULL, "Need parallel worker threads."); duke@435: ParNewRefEnqueueTaskProxy enq_task(task); duke@435: workers->run_task(&enq_task); duke@435: } duke@435: duke@435: void ParNewRefProcTaskExecutor::set_single_threaded_mode() duke@435: { duke@435: _state_set.flush(); duke@435: GenCollectedHeap* gch = GenCollectedHeap::heap(); duke@435: gch->set_par_threads(0); // 0 ==> non-parallel. duke@435: gch->save_marks(); duke@435: } duke@435: duke@435: ScanClosureWithParBarrier:: duke@435: ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : duke@435: ScanClosure(g, gc_barrier) {} duke@435: duke@435: EvacuateFollowersClosureGeneral:: duke@435: EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level, duke@435: OopsInGenClosure* cur, duke@435: OopsInGenClosure* older) : duke@435: _gch(gch), _level(level), duke@435: _scan_cur_or_nonheap(cur), _scan_older(older) duke@435: {} duke@435: duke@435: void EvacuateFollowersClosureGeneral::do_void() { duke@435: do { duke@435: // Beware: this call will lead to closure applications via virtual duke@435: // calls. duke@435: _gch->oop_since_save_marks_iterate(_level, duke@435: _scan_cur_or_nonheap, duke@435: _scan_older); duke@435: } while (!_gch->no_allocs_since_save_marks(_level)); duke@435: } duke@435: duke@435: duke@435: bool ParNewGeneration::_avoid_promotion_undo = false; duke@435: duke@435: void ParNewGeneration::adjust_desired_tenuring_threshold() { duke@435: // Set the desired survivor size to half the real survivor space duke@435: _tenuring_threshold = duke@435: age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize); duke@435: } duke@435: duke@435: // A Generation that does parallel young-gen collection. duke@435: duke@435: void ParNewGeneration::collect(bool full, duke@435: bool clear_all_soft_refs, duke@435: size_t size, duke@435: bool is_tlab) { duke@435: assert(full || size > 0, "otherwise we don't want to collect"); duke@435: GenCollectedHeap* gch = GenCollectedHeap::heap(); duke@435: assert(gch->kind() == CollectedHeap::GenCollectedHeap, duke@435: "not a CMS generational heap"); duke@435: AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); jmasa@3294: FlexibleWorkGang* workers = gch->workers(); jmasa@3294: assert(workers != NULL, "Need workgang for parallel work"); jmasa@3294: int active_workers = jmasa@3294: AdaptiveSizePolicy::calc_active_workers(workers->total_workers(), jmasa@3294: workers->active_workers(), jmasa@3294: Threads::number_of_non_daemon_threads()); jmasa@3294: workers->set_active_workers(active_workers); duke@435: _next_gen = gch->next_gen(this); duke@435: assert(_next_gen != NULL, duke@435: "This must be the youngest gen, and not the only gen"); duke@435: assert(gch->n_gens() == 2, duke@435: "Par collection currently only works with single older gen."); duke@435: // Do we have to avoid promotion_undo? duke@435: if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) { duke@435: set_avoid_promotion_undo(true); duke@435: } duke@435: duke@435: // If the next generation is too full to accomodate worst-case promotion duke@435: // from this generation, pass on collection; let the next generation duke@435: // do it. duke@435: if (!collection_attempt_is_safe()) { ysr@2243: gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one duke@435: return; duke@435: } duke@435: assert(to()->is_empty(), "Else not collection_attempt_is_safe"); duke@435: duke@435: init_assuming_no_promotion_failure(); duke@435: duke@435: if (UseAdaptiveSizePolicy) { duke@435: set_survivor_overflow(false); duke@435: size_policy->minor_collection_begin(); duke@435: } duke@435: brutisso@3767: TraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, gclog_or_tty); duke@435: // Capture heap used before collection (for printing). duke@435: size_t gch_prev_used = gch->used(); duke@435: duke@435: SpecializationStats::clear(); duke@435: duke@435: age_table()->clear(); jmasa@698: to()->clear(SpaceDecorator::Mangle); duke@435: duke@435: gch->save_marks(); duke@435: assert(workers != NULL, "Need parallel worker threads."); jmasa@3294: int n_workers = active_workers; jmasa@3294: jmasa@3294: // Set the correct parallelism (number of queues) in the reference processor jmasa@3294: ref_processor()->set_active_mt_degree(n_workers); jmasa@3294: jmasa@3294: // Always set the terminator for the active number of workers jmasa@3294: // because only those workers go through the termination protocol. jmasa@3294: ParallelTaskTerminator _term(n_workers, task_queues()); jmasa@3294: ParScanThreadStateSet thread_state_set(workers->active_workers(), duke@435: *to(), *this, *_next_gen, *task_queues(), ysr@1130: _overflow_stacks, desired_plab_sz(), _term); duke@435: duke@435: ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set); duke@435: gch->set_par_threads(n_workers); duke@435: gch->rem_set()->prepare_for_younger_refs_iterate(true); duke@435: // It turns out that even when we're using 1 thread, doing the work in a duke@435: // separate thread causes wide variance in run times. We can't help this duke@435: // in the multi-threaded case, but we special-case n=1 here to get duke@435: // repeatable measurements of the 1-thread overhead of the parallel code. duke@435: if (n_workers > 1) { jrose@1424: GenCollectedHeap::StrongRootsScope srs(gch); duke@435: workers->run_task(&tsk); duke@435: } else { jrose@1424: GenCollectedHeap::StrongRootsScope srs(gch); duke@435: tsk.work(0); duke@435: } jmasa@3294: thread_state_set.reset(0 /* Bad value in debug if not reset */, jmasa@3294: promotion_failed()); duke@435: duke@435: // Process (weak) reference objects found during scavenge. ysr@888: ReferenceProcessor* rp = ref_processor(); duke@435: IsAliveClosure is_alive(this); duke@435: ScanWeakRefClosure scan_weak_ref(this); duke@435: KeepAliveClosure keep_alive(&scan_weak_ref); duke@435: ScanClosure scan_without_gc_barrier(this, false); duke@435: ScanClosureWithParBarrier scan_with_gc_barrier(this, true); duke@435: set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); duke@435: EvacuateFollowersClosureGeneral evacuate_followers(gch, _level, duke@435: &scan_without_gc_barrier, &scan_with_gc_barrier); ysr@892: rp->setup_policy(clear_all_soft_refs); jmasa@3294: // Can the mt_degree be set later (at run_task() time would be best)? jmasa@3294: rp->set_active_mt_degree(active_workers); ysr@888: if (rp->processing_is_mt()) { duke@435: ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); ysr@888: rp->process_discovered_references(&is_alive, &keep_alive, ysr@888: &evacuate_followers, &task_executor); duke@435: } else { duke@435: thread_state_set.flush(); duke@435: gch->set_par_threads(0); // 0 ==> non-parallel. duke@435: gch->save_marks(); ysr@888: rp->process_discovered_references(&is_alive, &keep_alive, ysr@888: &evacuate_followers, NULL); duke@435: } duke@435: if (!promotion_failed()) { duke@435: // Swap the survivor spaces. jmasa@698: eden()->clear(SpaceDecorator::Mangle); jmasa@698: from()->clear(SpaceDecorator::Mangle); jmasa@698: if (ZapUnusedHeapArea) { jmasa@698: // This is now done here because of the piece-meal mangling which jmasa@698: // can check for valid mangling at intermediate points in the jmasa@698: // collection(s). When a minor collection fails to collect jmasa@698: // sufficient space resizing of the young generation can occur jmasa@698: // an redistribute the spaces in the young generation. Mangle jmasa@698: // here so that unzapped regions don't get distributed to jmasa@698: // other spaces. jmasa@698: to()->mangle_unused_area(); jmasa@698: } duke@435: swap_spaces(); duke@435: jmasa@1822: // A successful scavenge should restart the GC time limit count which is jmasa@1822: // for full GC's. jmasa@1822: size_policy->reset_gc_overhead_limit_count(); jmasa@1822: duke@435: assert(to()->is_empty(), "to space should be empty now"); duke@435: } else { jcoomes@2191: assert(_promo_failure_scan_stack.is_empty(), "post condition"); jcoomes@2191: _promo_failure_scan_stack.clear(true); // Clear cached segments. jcoomes@2191: duke@435: remove_forwarding_pointers(); duke@435: if (PrintGCDetails) { duke@435: gclog_or_tty->print(" (promotion failed)"); duke@435: } duke@435: // All the spaces are in play for mark-sweep. duke@435: swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. duke@435: from()->set_next_compaction_space(to()); ysr@2243: gch->set_incremental_collection_failed(); ysr@1580: // Inform the next generation that a promotion failure occurred. ysr@1580: _next_gen->promotion_failure_occurred(); jmasa@441: jmasa@441: // Reset the PromotionFailureALot counters. jmasa@441: NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();) duke@435: } duke@435: // set new iteration safe limit for the survivor spaces duke@435: from()->set_concurrent_iteration_safe_limit(from()->top()); duke@435: to()->set_concurrent_iteration_safe_limit(to()->top()); duke@435: duke@435: adjust_desired_tenuring_threshold(); duke@435: if (ResizePLAB) { johnc@4130: plab_stats()->adjust_desired_plab_sz(n_workers); duke@435: } duke@435: duke@435: if (PrintGC && !PrintGCDetails) { duke@435: gch->print_heap_change(gch_prev_used); duke@435: } duke@435: jcoomes@2067: if (PrintGCDetails && ParallelGCVerbose) { jcoomes@2067: TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats()); jcoomes@2067: TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats()); jcoomes@2067: } jcoomes@2065: duke@435: if (UseAdaptiveSizePolicy) { duke@435: size_policy->minor_collection_end(gch->gc_cause()); duke@435: size_policy->avg_survived()->sample(from()->used()); duke@435: } duke@435: johnc@3538: // We need to use a monotonically non-deccreasing time in ms johnc@3538: // or we will see time-warp warnings and os::javaTimeMillis() johnc@3538: // does not guarantee monotonicity. johnc@3538: jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; johnc@3538: update_time_of_last_gc(now); duke@435: duke@435: SpecializationStats::print(); duke@435: ysr@888: rp->set_enqueuing_is_done(true); ysr@888: if (rp->processing_is_mt()) { duke@435: ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); ysr@888: rp->enqueue_discovered_references(&task_executor); duke@435: } else { ysr@888: rp->enqueue_discovered_references(NULL); duke@435: } ysr@888: rp->verify_no_references_recorded(); duke@435: } duke@435: duke@435: static int sum; duke@435: void ParNewGeneration::waste_some_time() { duke@435: for (int i = 0; i < 100; i++) { duke@435: sum += i; duke@435: } duke@435: } duke@435: duke@435: static const oop ClaimedForwardPtr = oop(0x4); duke@435: duke@435: // Because of concurrency, there are times where an object for which duke@435: // "is_forwarded()" is true contains an "interim" forwarding pointer duke@435: // value. Such a value will soon be overwritten with a real value. duke@435: // This method requires "obj" to have a forwarding pointer, and waits, if duke@435: // necessary for a real one to be inserted, and returns it. duke@435: duke@435: oop ParNewGeneration::real_forwardee(oop obj) { duke@435: oop forward_ptr = obj->forwardee(); duke@435: if (forward_ptr != ClaimedForwardPtr) { duke@435: return forward_ptr; duke@435: } else { duke@435: return real_forwardee_slow(obj); duke@435: } duke@435: } duke@435: duke@435: oop ParNewGeneration::real_forwardee_slow(oop obj) { duke@435: // Spin-read if it is claimed but not yet written by another thread. duke@435: oop forward_ptr = obj->forwardee(); duke@435: while (forward_ptr == ClaimedForwardPtr) { duke@435: waste_some_time(); duke@435: assert(obj->is_forwarded(), "precondition"); duke@435: forward_ptr = obj->forwardee(); duke@435: } duke@435: return forward_ptr; duke@435: } duke@435: duke@435: #ifdef ASSERT duke@435: bool ParNewGeneration::is_legal_forward_ptr(oop p) { duke@435: return duke@435: (_avoid_promotion_undo && p == ClaimedForwardPtr) duke@435: || Universe::heap()->is_in_reserved(p); duke@435: } duke@435: #endif duke@435: duke@435: void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { ysr@2380: if (m->must_be_preserved_for_promotion_failure(obj)) { ysr@2380: // We should really have separate per-worker stacks, rather ysr@2380: // than use locking of a common pair of stacks. duke@435: MutexLocker ml(ParGCRareEvent_lock); ysr@2380: preserve_mark(obj, m); duke@435: } duke@435: } duke@435: duke@435: // Multiple GC threads may try to promote an object. If the object duke@435: // is successfully promoted, a forwarding pointer will be installed in duke@435: // the object in the young generation. This method claims the right duke@435: // to install the forwarding pointer before it copies the object, duke@435: // thus avoiding the need to undo the copy as in duke@435: // copy_to_survivor_space_avoiding_with_undo. duke@435: duke@435: oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo( duke@435: ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { duke@435: // In the sequential version, this assert also says that the object is duke@435: // not forwarded. That might not be the case here. It is the case that duke@435: // the caller observed it to be not forwarded at some time in the past. duke@435: assert(is_in_reserved(old), "shouldn't be scavenging this oop"); duke@435: duke@435: // The sequential code read "old->age()" below. That doesn't work here, duke@435: // since the age is in the mark word, and that might be overwritten with duke@435: // a forwarding pointer by a parallel thread. So we must save the mark duke@435: // word in a local and then analyze it. duke@435: oopDesc dummyOld; duke@435: dummyOld.set_mark(m); duke@435: assert(!dummyOld.is_forwarded(), duke@435: "should not be called with forwarding pointer mark word."); duke@435: duke@435: oop new_obj = NULL; duke@435: oop forward_ptr; duke@435: duke@435: // Try allocating obj in to-space (unless too old) duke@435: if (dummyOld.age() < tenuring_threshold()) { duke@435: new_obj = (oop)par_scan_state->alloc_in_to_space(sz); duke@435: if (new_obj == NULL) { duke@435: set_survivor_overflow(true); duke@435: } duke@435: } duke@435: duke@435: if (new_obj == NULL) { duke@435: // Either to-space is full or we decided to promote duke@435: // try allocating obj tenured duke@435: duke@435: // Attempt to install a null forwarding pointer (atomically), duke@435: // to claim the right to install the real forwarding pointer. duke@435: forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); duke@435: if (forward_ptr != NULL) { duke@435: // someone else beat us to it. duke@435: return real_forwardee(old); duke@435: } duke@435: duke@435: new_obj = _next_gen->par_promote(par_scan_state->thread_num(), duke@435: old, m, sz); duke@435: duke@435: if (new_obj == NULL) { duke@435: // promotion failed, forward to self duke@435: _promotion_failed = true; duke@435: new_obj = old; duke@435: duke@435: preserve_mark_if_necessary(old, m); ysr@1580: // Log the size of the maiden promotion failure ysr@1580: par_scan_state->log_promotion_failure(sz); duke@435: } duke@435: duke@435: old->forward_to(new_obj); duke@435: forward_ptr = NULL; duke@435: } else { duke@435: // Is in to-space; do copying ourselves. duke@435: Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); duke@435: forward_ptr = old->forward_to_atomic(new_obj); duke@435: // Restore the mark word copied above. duke@435: new_obj->set_mark(m); duke@435: // Increment age if obj still in new generation duke@435: new_obj->incr_age(); duke@435: par_scan_state->age_table()->add(new_obj, sz); duke@435: } duke@435: assert(new_obj != NULL, "just checking"); duke@435: coleenp@4037: #ifndef PRODUCT coleenp@4037: // This code must come after the CAS test, or it will print incorrect coleenp@4037: // information. coleenp@4037: if (TraceScavenge) { coleenp@4037: gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", coleenp@4037: is_in_reserved(new_obj) ? "copying" : "tenuring", coleenp@4037: new_obj->klass()->internal_name(), old, new_obj, new_obj->size()); coleenp@4037: } coleenp@4037: #endif coleenp@4037: duke@435: if (forward_ptr == NULL) { duke@435: oop obj_to_push = new_obj; duke@435: if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { duke@435: // Length field used as index of next element to be scanned. duke@435: // Real length can be obtained from real_forwardee() duke@435: arrayOop(old)->set_length(0); duke@435: obj_to_push = old; duke@435: assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, duke@435: "push forwarded object"); duke@435: } duke@435: // Push it on one of the queues of to-be-scanned objects. ysr@969: bool simulate_overflow = false; ysr@969: NOT_PRODUCT( ysr@969: if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { ysr@969: // simulate a stack overflow ysr@969: simulate_overflow = true; ysr@969: } ysr@969: ) ysr@969: if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { duke@435: // Add stats for overflow pushes. duke@435: if (Verbose && PrintGCDetails) { duke@435: gclog_or_tty->print("queue overflow!\n"); duke@435: } ysr@969: push_on_overflow_list(old, par_scan_state); jcoomes@2065: TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); duke@435: } duke@435: duke@435: return new_obj; duke@435: } duke@435: duke@435: // Oops. Someone beat us to it. Undo the allocation. Where did we duke@435: // allocate it? duke@435: if (is_in_reserved(new_obj)) { duke@435: // Must be in to_space. duke@435: assert(to()->is_in_reserved(new_obj), "Checking"); duke@435: if (forward_ptr == ClaimedForwardPtr) { duke@435: // Wait to get the real forwarding pointer value. duke@435: forward_ptr = real_forwardee(old); duke@435: } duke@435: par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); duke@435: } duke@435: duke@435: return forward_ptr; duke@435: } duke@435: duke@435: duke@435: // Multiple GC threads may try to promote the same object. If two duke@435: // or more GC threads copy the object, only one wins the race to install duke@435: // the forwarding pointer. The other threads have to undo their copy. duke@435: duke@435: oop ParNewGeneration::copy_to_survivor_space_with_undo( duke@435: ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { duke@435: duke@435: // In the sequential version, this assert also says that the object is duke@435: // not forwarded. That might not be the case here. It is the case that duke@435: // the caller observed it to be not forwarded at some time in the past. duke@435: assert(is_in_reserved(old), "shouldn't be scavenging this oop"); duke@435: duke@435: // The sequential code read "old->age()" below. That doesn't work here, duke@435: // since the age is in the mark word, and that might be overwritten with duke@435: // a forwarding pointer by a parallel thread. So we must save the mark duke@435: // word here, install it in a local oopDesc, and then analyze it. duke@435: oopDesc dummyOld; duke@435: dummyOld.set_mark(m); duke@435: assert(!dummyOld.is_forwarded(), duke@435: "should not be called with forwarding pointer mark word."); duke@435: duke@435: bool failed_to_promote = false; duke@435: oop new_obj = NULL; duke@435: oop forward_ptr; duke@435: duke@435: // Try allocating obj in to-space (unless too old) duke@435: if (dummyOld.age() < tenuring_threshold()) { duke@435: new_obj = (oop)par_scan_state->alloc_in_to_space(sz); duke@435: if (new_obj == NULL) { duke@435: set_survivor_overflow(true); duke@435: } duke@435: } duke@435: duke@435: if (new_obj == NULL) { duke@435: // Either to-space is full or we decided to promote duke@435: // try allocating obj tenured duke@435: new_obj = _next_gen->par_promote(par_scan_state->thread_num(), duke@435: old, m, sz); duke@435: duke@435: if (new_obj == NULL) { duke@435: // promotion failed, forward to self duke@435: forward_ptr = old->forward_to_atomic(old); duke@435: new_obj = old; duke@435: duke@435: if (forward_ptr != NULL) { duke@435: return forward_ptr; // someone else succeeded duke@435: } duke@435: duke@435: _promotion_failed = true; duke@435: failed_to_promote = true; duke@435: duke@435: preserve_mark_if_necessary(old, m); ysr@1580: // Log the size of the maiden promotion failure ysr@1580: par_scan_state->log_promotion_failure(sz); duke@435: } duke@435: } else { duke@435: // Is in to-space; do copying ourselves. duke@435: Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); duke@435: // Restore the mark word copied above. duke@435: new_obj->set_mark(m); duke@435: // Increment age if new_obj still in new generation duke@435: new_obj->incr_age(); duke@435: par_scan_state->age_table()->add(new_obj, sz); duke@435: } duke@435: assert(new_obj != NULL, "just checking"); duke@435: coleenp@4037: #ifndef PRODUCT coleenp@4037: // This code must come after the CAS test, or it will print incorrect coleenp@4037: // information. coleenp@4037: if (TraceScavenge) { coleenp@4037: gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", coleenp@4037: is_in_reserved(new_obj) ? "copying" : "tenuring", coleenp@4037: new_obj->klass()->internal_name(), old, new_obj, new_obj->size()); coleenp@4037: } coleenp@4037: #endif coleenp@4037: duke@435: // Now attempt to install the forwarding pointer (atomically). duke@435: // We have to copy the mark word before overwriting with forwarding duke@435: // ptr, so we can restore it below in the copy. duke@435: if (!failed_to_promote) { duke@435: forward_ptr = old->forward_to_atomic(new_obj); duke@435: } duke@435: duke@435: if (forward_ptr == NULL) { duke@435: oop obj_to_push = new_obj; duke@435: if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { duke@435: // Length field used as index of next element to be scanned. duke@435: // Real length can be obtained from real_forwardee() duke@435: arrayOop(old)->set_length(0); duke@435: obj_to_push = old; duke@435: assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, duke@435: "push forwarded object"); duke@435: } duke@435: // Push it on one of the queues of to-be-scanned objects. ysr@969: bool simulate_overflow = false; ysr@969: NOT_PRODUCT( ysr@969: if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { ysr@969: // simulate a stack overflow ysr@969: simulate_overflow = true; ysr@969: } ysr@969: ) ysr@969: if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { duke@435: // Add stats for overflow pushes. ysr@969: push_on_overflow_list(old, par_scan_state); jcoomes@2065: TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); duke@435: } duke@435: duke@435: return new_obj; duke@435: } duke@435: duke@435: // Oops. Someone beat us to it. Undo the allocation. Where did we duke@435: // allocate it? duke@435: if (is_in_reserved(new_obj)) { duke@435: // Must be in to_space. duke@435: assert(to()->is_in_reserved(new_obj), "Checking"); duke@435: par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); duke@435: } else { duke@435: assert(!_avoid_promotion_undo, "Should not be here if avoiding."); duke@435: _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(), duke@435: (HeapWord*)new_obj, sz); duke@435: } duke@435: duke@435: return forward_ptr; duke@435: } duke@435: ysr@969: #ifndef PRODUCT ysr@969: // It's OK to call this multi-threaded; the worst thing ysr@969: // that can happen is that we'll get a bunch of closely ysr@969: // spaced simulated oveflows, but that's OK, in fact ysr@969: // probably good as it would exercise the overflow code ysr@969: // under contention. ysr@969: bool ParNewGeneration::should_simulate_overflow() { ysr@969: if (_overflow_counter-- <= 0) { // just being defensive ysr@969: _overflow_counter = ParGCWorkQueueOverflowInterval; ysr@969: return true; ysr@969: } else { ysr@969: return false; ysr@969: } ysr@969: } ysr@969: #endif ysr@969: ysr@1114: // In case we are using compressed oops, we need to be careful. ysr@1114: // If the object being pushed is an object array, then its length ysr@1114: // field keeps track of the "grey boundary" at which the next ysr@1114: // incremental scan will be done (see ParGCArrayScanChunk). ysr@1114: // When using compressed oops, this length field is kept in the ysr@1114: // lower 32 bits of the erstwhile klass word and cannot be used ysr@1114: // for the overflow chaining pointer (OCP below). As such the OCP ysr@1114: // would itself need to be compressed into the top 32-bits in this ysr@1114: // case. Unfortunately, see below, in the event that we have a ysr@1114: // promotion failure, the node to be pushed on the list can be ysr@1114: // outside of the Java heap, so the heap-based pointer compression ysr@1114: // would not work (we would have potential aliasing between C-heap ysr@1114: // and Java-heap pointers). For this reason, when using compressed ysr@1114: // oops, we simply use a worker-thread-local, non-shared overflow ysr@1114: // list in the form of a growable array, with a slightly different ysr@1114: // overflow stack draining strategy. If/when we start using fat ysr@1114: // stacks here, we can go back to using (fat) pointer chains ysr@1114: // (although some performance comparisons would be useful since ysr@1114: // single global lists have their own performance disadvantages ysr@1114: // as we were made painfully aware not long ago, see 6786503). ysr@969: #define BUSY (oop(0x1aff1aff)) ysr@969: void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { ysr@1114: assert(is_in_reserved(from_space_obj), "Should be from this generation"); ysr@1130: if (ParGCUseLocalOverflow) { ysr@1114: // In the case of compressed oops, we use a private, not-shared ysr@1114: // overflow stack. ysr@1114: par_scan_state->push_on_overflow_stack(from_space_obj); ysr@1114: } else { ysr@1130: assert(!UseCompressedOops, "Error"); ysr@1114: // if the object has been forwarded to itself, then we cannot ysr@1114: // use the klass pointer for the linked list. Instead we have ysr@1114: // to allocate an oopDesc in the C-Heap and use that for the linked list. ysr@1114: // XXX This is horribly inefficient when a promotion failure occurs ysr@1114: // and should be fixed. XXX FIX ME !!! ysr@969: #ifndef PRODUCT ysr@1114: Atomic::inc_ptr(&_num_par_pushes); ysr@1114: assert(_num_par_pushes > 0, "Tautology"); ysr@969: #endif ysr@1114: if (from_space_obj->forwardee() == from_space_obj) { zgu@3900: oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC); ysr@1114: listhead->forward_to(from_space_obj); ysr@1114: from_space_obj = listhead; ysr@1114: } ysr@1114: oop observed_overflow_list = _overflow_list; ysr@1114: oop cur_overflow_list; ysr@1114: do { ysr@1114: cur_overflow_list = observed_overflow_list; ysr@1114: if (cur_overflow_list != BUSY) { ysr@1114: from_space_obj->set_klass_to_list_ptr(cur_overflow_list); ysr@1114: } else { ysr@1114: from_space_obj->set_klass_to_list_ptr(NULL); ysr@1114: } ysr@1114: observed_overflow_list = ysr@1114: (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list); ysr@1114: } while (cur_overflow_list != observed_overflow_list); duke@435: } duke@435: } duke@435: ysr@1114: bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { ysr@1114: bool res; ysr@1114: ysr@1130: if (ParGCUseLocalOverflow) { ysr@1114: res = par_scan_state->take_from_overflow_stack(); ysr@1114: } else { ysr@1130: assert(!UseCompressedOops, "Error"); ysr@1114: res = take_from_overflow_list_work(par_scan_state); ysr@1114: } ysr@1114: return res; ysr@1114: } ysr@1114: ysr@1114: ysr@969: // *NOTE*: The overflow list manipulation code here and ysr@969: // in CMSCollector:: are very similar in shape, ysr@969: // except that in the CMS case we thread the objects ysr@969: // directly into the list via their mark word, and do ysr@969: // not need to deal with special cases below related ysr@969: // to chunking of object arrays and promotion failure ysr@969: // handling. ysr@969: // CR 6797058 has been filed to attempt consolidation of ysr@969: // the common code. ysr@969: // Because of the common code, if you make any changes in ysr@969: // the code below, please check the CMS version to see if ysr@969: // similar changes might be needed. ysr@969: // See CMSCollector::par_take_from_overflow_list() for ysr@969: // more extensive documentation comments. ysr@1114: bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { duke@435: ObjToScanQueue* work_q = par_scan_state->work_queue(); duke@435: // How many to take? ysr@1114: size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, ysr@969: (size_t)ParGCDesiredObjsFromOverflowList); duke@435: jcoomes@2191: assert(!UseCompressedOops, "Error"); ysr@1114: assert(par_scan_state->overflow_stack() == NULL, "Error"); duke@435: if (_overflow_list == NULL) return false; duke@435: duke@435: // Otherwise, there was something there; try claiming the list. ysr@969: oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); ysr@969: // Trim off a prefix of at most objsFromOverflow items ysr@969: Thread* tid = Thread::current(); ysr@969: size_t spin_count = (size_t)ParallelGCThreads; ysr@969: size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); ysr@969: for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { ysr@969: // someone grabbed it before we did ... ysr@969: // ... we spin for a short while... ysr@969: os::sleep(tid, sleep_time_millis, false); ysr@969: if (_overflow_list == NULL) { ysr@969: // nothing left to take ysr@969: return false; ysr@969: } else if (_overflow_list != BUSY) { ysr@969: // try and grab the prefix ysr@969: prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); ysr@969: } duke@435: } ysr@969: if (prefix == NULL || prefix == BUSY) { ysr@969: // Nothing to take or waited long enough ysr@969: if (prefix == NULL) { ysr@969: // Write back the NULL in case we overwrote it with BUSY above ysr@969: // and it is still the same value. ysr@969: (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); ysr@969: } ysr@969: return false; ysr@969: } ysr@969: assert(prefix != NULL && prefix != BUSY, "Error"); ysr@969: size_t i = 1; duke@435: oop cur = prefix; coleenp@602: while (i < objsFromOverflow && cur->klass_or_null() != NULL) { coleenp@4037: i++; cur = cur->list_ptr_from_klass(); duke@435: } duke@435: duke@435: // Reattach remaining (suffix) to overflow list ysr@969: if (cur->klass_or_null() == NULL) { ysr@969: // Write back the NULL in lieu of the BUSY we wrote ysr@969: // above and it is still the same value. ysr@969: if (_overflow_list == BUSY) { ysr@969: (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); duke@435: } ysr@969: } else { coleenp@4037: assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error"); coleenp@4037: oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list ysr@969: cur->set_klass_to_list_ptr(NULL); // break off suffix ysr@969: // It's possible that the list is still in the empty(busy) state ysr@969: // we left it in a short while ago; in that case we may be ysr@969: // able to place back the suffix. ysr@969: oop observed_overflow_list = _overflow_list; ysr@969: oop cur_overflow_list = observed_overflow_list; ysr@969: bool attached = false; ysr@969: while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { ysr@969: observed_overflow_list = ysr@969: (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); ysr@969: if (cur_overflow_list == observed_overflow_list) { ysr@969: attached = true; ysr@969: break; ysr@969: } else cur_overflow_list = observed_overflow_list; ysr@969: } ysr@969: if (!attached) { ysr@969: // Too bad, someone else got in in between; we'll need to do a splice. ysr@969: // Find the last item of suffix list ysr@969: oop last = suffix; ysr@969: while (last->klass_or_null() != NULL) { coleenp@4037: last = last->list_ptr_from_klass(); ysr@969: } ysr@969: // Atomically prepend suffix to current overflow list ysr@969: observed_overflow_list = _overflow_list; ysr@969: do { ysr@969: cur_overflow_list = observed_overflow_list; ysr@969: if (cur_overflow_list != BUSY) { ysr@969: // Do the splice ... ysr@969: last->set_klass_to_list_ptr(cur_overflow_list); ysr@969: } else { // cur_overflow_list == BUSY ysr@969: last->set_klass_to_list_ptr(NULL); ysr@969: } ysr@969: observed_overflow_list = ysr@969: (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); ysr@969: } while (cur_overflow_list != observed_overflow_list); duke@435: } duke@435: } duke@435: duke@435: // Push objects on prefix list onto this thread's work queue ysr@969: assert(prefix != NULL && prefix != BUSY, "program logic"); duke@435: cur = prefix; ysr@969: ssize_t n = 0; duke@435: while (cur != NULL) { duke@435: oop obj_to_push = cur->forwardee(); coleenp@4037: oop next = cur->list_ptr_from_klass(); duke@435: cur->set_klass(obj_to_push->klass()); ysr@969: // This may be an array object that is self-forwarded. In that case, the list pointer ysr@969: // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. ysr@969: if (!is_in_reserved(cur)) { ysr@969: // This can become a scaling bottleneck when there is work queue overflow coincident ysr@969: // with promotion failure. ysr@969: oopDesc* f = cur; zgu@3900: FREE_C_HEAP_ARRAY(oopDesc, f, mtGC); ysr@969: } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { ysr@969: assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); duke@435: obj_to_push = cur; duke@435: } ysr@969: bool ok = work_q->push(obj_to_push); ysr@969: assert(ok, "Should have succeeded"); duke@435: cur = next; duke@435: n++; duke@435: } jcoomes@2065: TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); ysr@969: #ifndef PRODUCT ysr@969: assert(_num_par_pushes >= n, "Too many pops?"); ysr@969: Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); ysr@969: #endif duke@435: return true; duke@435: } ysr@969: #undef BUSY duke@435: duke@435: void ParNewGeneration::ref_processor_init() duke@435: { duke@435: if (_ref_processor == NULL) { duke@435: // Allocate and initialize a reference processor ysr@2651: _ref_processor = ysr@2651: new ReferenceProcessor(_reserved, // span ysr@2651: ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing ysr@2651: (int) ParallelGCThreads, // mt processing degree ysr@2651: refs_discovery_is_mt(), // mt discovery ysr@2651: (int) ParallelGCThreads, // mt discovery degree ysr@2651: refs_discovery_is_atomic(), // atomic_discovery ysr@2651: NULL, // is_alive_non_header ysr@2651: false); // write barrier for next field updates duke@435: } duke@435: } duke@435: duke@435: const char* ParNewGeneration::name() const { duke@435: return "par new generation"; duke@435: } jmasa@2188: jmasa@2188: bool ParNewGeneration::in_use() { jmasa@2188: return UseParNewGC && ParallelGCThreads > 0; jmasa@2188: }