• file
• revisions
• annotate
• diff
• comparison
• raw

src/share/vm/gc_implementation/parNew/parNewGeneration.cpp@feeb96a45707 (annotated)

src/share/vm/gc_implementation/parNew/parNewGeneration.cpp

Wed, 28 May 2008 21:06:24 -0700

author

coleenp

date

Wed, 28 May 2008 21:06:24 -0700

changeset 602

feeb96a45707

parent 548

ba764ed4b6f2

child 631

d1605aabd0a1

child 698

12eea04c8b06

permissions

-rw-r--r--

6696264: assert("narrow oop can never be zero") for GCBasher & ParNewGC
Summary: decouple set_klass() with zeroing the gap when compressed.
Reviewed-by: kvn, ysr, jrose

duke@435	1	/*
duke@435	2	* Copyright 2001-2007 Sun Microsystems, Inc. All Rights Reserved.
duke@435	3	* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435	4	*
duke@435	5	* This code is free software; you can redistribute it and/or modify it
duke@435	6	* under the terms of the GNU General Public License version 2 only, as
duke@435	7	* published by the Free Software Foundation.
duke@435	8	*
duke@435	9	* This code is distributed in the hope that it will be useful, but WITHOUT
duke@435	10	* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435	11	* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@435	12	* version 2 for more details (a copy is included in the LICENSE file that
duke@435	13	* accompanied this code).
duke@435	14	*
duke@435	15	* You should have received a copy of the GNU General Public License version
duke@435	16	* 2 along with this work; if not, write to the Free Software Foundation,
duke@435	17	* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435	18	*
duke@435	19	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
duke@435	20	* CA 95054 USA or visit www.sun.com if you need additional information or
duke@435	21	* have any questions.
duke@435	22	*
duke@435	23	*/
duke@435	24
duke@435	25	# include "incls/_precompiled.incl"
duke@435	26	# include "incls/_parNewGeneration.cpp.incl"
duke@435	27
duke@435	28	#ifdef _MSC_VER
duke@435	29	#pragma warning( push )
duke@435	30	#pragma warning( disable:4355 ) // 'this' : used in base member initializer list
duke@435	31	#endif
duke@435	32	ParScanThreadState::ParScanThreadState(Space* to_space_,
duke@435	33	ParNewGeneration* gen_,
duke@435	34	Generation* old_gen_,
duke@435	35	int thread_num_,
duke@435	36	ObjToScanQueueSet* work_queue_set_,
duke@435	37	size_t desired_plab_sz_,
duke@435	38	ParallelTaskTerminator& term_) :
duke@435	39	_to_space(to_space_), _old_gen(old_gen_), _thread_num(thread_num_),
duke@435	40	_work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
duke@435	41	_ageTable(false), // false ==> not the global age table, no perf data.
duke@435	42	_to_space_alloc_buffer(desired_plab_sz_),
duke@435	43	_to_space_closure(gen_, this), _old_gen_closure(gen_, this),
duke@435	44	_to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
duke@435	45	_older_gen_closure(gen_, this),
duke@435	46	_evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
duke@435	47	&_to_space_root_closure, gen_, &_old_gen_root_closure,
duke@435	48	work_queue_set_, &term_),
duke@435	49	_is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
duke@435	50	_keep_alive_closure(&_scan_weak_ref_closure),
duke@435	51	_pushes(0), _pops(0), _steals(0), _steal_attempts(0), _term_attempts(0),
duke@435	52	_strong_roots_time(0.0), _term_time(0.0)
duke@435	53	{
duke@435	54	_survivor_chunk_array =
duke@435	55	(ChunkArray*) old_gen()->get_data_recorder(thread_num());
duke@435	56	_hash_seed = 17; // Might want to take time-based random value.
duke@435	57	_start = os::elapsedTime();
duke@435	58	_old_gen_closure.set_generation(old_gen_);
duke@435	59	_old_gen_root_closure.set_generation(old_gen_);
duke@435	60	}
duke@435	61	#ifdef _MSC_VER
duke@435	62	#pragma warning( pop )
duke@435	63	#endif
duke@435	64
duke@435	65	void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
duke@435	66	size_t plab_word_size) {
duke@435	67	ChunkArray* sca = survivor_chunk_array();
duke@435	68	if (sca != NULL) {
duke@435	69	// A non-null SCA implies that we want the PLAB data recorded.
duke@435	70	sca->record_sample(plab_start, plab_word_size);
duke@435	71	}
duke@435	72	}
duke@435	73
duke@435	74	bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
duke@435	75	return new_obj->is_objArray() &&
duke@435	76	arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
duke@435	77	new_obj != old_obj;
duke@435	78	}
duke@435	79
duke@435	80	void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
duke@435	81	assert(old->is_objArray(), "must be obj array");
duke@435	82	assert(old->is_forwarded(), "must be forwarded");
duke@435	83	assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
duke@435	84	assert(!_old_gen->is_in(old), "must be in young generation.");
duke@435	85
duke@435	86	objArrayOop obj = objArrayOop(old->forwardee());
duke@435	87	// Process ParGCArrayScanChunk elements now
duke@435	88	// and push the remainder back onto queue
duke@435	89	int start = arrayOop(old)->length();
duke@435	90	int end = obj->length();
duke@435	91	int remainder = end - start;
duke@435	92	assert(start <= end, "just checking");
duke@435	93	if (remainder > 2 * ParGCArrayScanChunk) {
duke@435	94	// Test above combines last partial chunk with a full chunk
duke@435	95	end = start + ParGCArrayScanChunk;
duke@435	96	arrayOop(old)->set_length(end);
duke@435	97	// Push remainder.
duke@435	98	bool ok = work_queue()->push(old);
duke@435	99	assert(ok, "just popped, push must be okay");
duke@435	100	note_push();
duke@435	101	} else {
duke@435	102	// Restore length so that it can be used if there
duke@435	103	// is a promotion failure and forwarding pointers
duke@435	104	// must be removed.
duke@435	105	arrayOop(old)->set_length(end);
duke@435	106	}
coleenp@548	107
duke@435	108	// process our set of indices (include header in first chunk)
coleenp@548	109	// should make sure end is even (aligned to HeapWord in case of compressed oops)
duke@435	110	if ((HeapWord *)obj < young_old_boundary()) {
duke@435	111	// object is in to_space
coleenp@548	112	obj->oop_iterate_range(&_to_space_closure, start, end);
duke@435	113	} else {
duke@435	114	// object is in old generation
coleenp@548	115	obj->oop_iterate_range(&_old_gen_closure, start, end);
duke@435	116	}
duke@435	117	}
duke@435	118
duke@435	119
duke@435	120	void ParScanThreadState::trim_queues(int max_size) {
duke@435	121	ObjToScanQueue* queue = work_queue();
duke@435	122	while (queue->size() > (juint)max_size) {
duke@435	123	oop obj_to_scan;
duke@435	124	if (queue->pop_local(obj_to_scan)) {
duke@435	125	note_pop();
duke@435	126
duke@435	127	if ((HeapWord *)obj_to_scan < young_old_boundary()) {
duke@435	128	if (obj_to_scan->is_objArray() &&
duke@435	129	obj_to_scan->is_forwarded() &&
duke@435	130	obj_to_scan->forwardee() != obj_to_scan) {
duke@435	131	scan_partial_array_and_push_remainder(obj_to_scan);
duke@435	132	} else {
duke@435	133	// object is in to_space
duke@435	134	obj_to_scan->oop_iterate(&_to_space_closure);
duke@435	135	}
duke@435	136	} else {
duke@435	137	// object is in old generation
duke@435	138	obj_to_scan->oop_iterate(&_old_gen_closure);
duke@435	139	}
duke@435	140	}
duke@435	141	}
duke@435	142	}
duke@435	143
duke@435	144	HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
duke@435	145
duke@435	146	// Otherwise, if the object is small enough, try to reallocate the
duke@435	147	// buffer.
duke@435	148	HeapWord* obj = NULL;
duke@435	149	if (!_to_space_full) {
duke@435	150	ParGCAllocBuffer* const plab = to_space_alloc_buffer();
duke@435	151	Space* const sp = to_space();
duke@435	152	if (word_sz * 100 <
duke@435	153	ParallelGCBufferWastePct * plab->word_sz()) {
duke@435	154	// Is small enough; abandon this buffer and start a new one.
duke@435	155	plab->retire(false, false);
duke@435	156	size_t buf_size = plab->word_sz();
duke@435	157	HeapWord* buf_space = sp->par_allocate(buf_size);
duke@435	158	if (buf_space == NULL) {
duke@435	159	const size_t min_bytes =
duke@435	160	ParGCAllocBuffer::min_size() << LogHeapWordSize;
duke@435	161	size_t free_bytes = sp->free();
duke@435	162	while(buf_space == NULL && free_bytes >= min_bytes) {
duke@435	163	buf_size = free_bytes >> LogHeapWordSize;
duke@435	164	assert(buf_size == (size_t)align_object_size(buf_size),
duke@435	165	"Invariant");
duke@435	166	buf_space = sp->par_allocate(buf_size);
duke@435	167	free_bytes = sp->free();
duke@435	168	}
duke@435	169	}
duke@435	170	if (buf_space != NULL) {
duke@435	171	plab->set_word_size(buf_size);
duke@435	172	plab->set_buf(buf_space);
duke@435	173	record_survivor_plab(buf_space, buf_size);
duke@435	174	obj = plab->allocate(word_sz);
duke@435	175	// Note that we cannot compare buf_size < word_sz below
duke@435	176	// because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
duke@435	177	assert(obj != NULL || plab->words_remaining() < word_sz,
duke@435	178	"Else should have been able to allocate");
duke@435	179	// It's conceivable that we may be able to use the
duke@435	180	// buffer we just grabbed for subsequent small requests
duke@435	181	// even if not for this one.
duke@435	182	} else {
duke@435	183	// We're used up.
duke@435	184	_to_space_full = true;
duke@435	185	}
duke@435	186
duke@435	187	} else {
duke@435	188	// Too large; allocate the object individually.
duke@435	189	obj = sp->par_allocate(word_sz);
duke@435	190	}
duke@435	191	}
duke@435	192	return obj;
duke@435	193	}
duke@435	194
duke@435	195
duke@435	196	void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
duke@435	197	size_t word_sz) {
duke@435	198	// Is the alloc in the current alloc buffer?
duke@435	199	if (to_space_alloc_buffer()->contains(obj)) {
duke@435	200	assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
duke@435	201	"Should contain whole object.");
duke@435	202	to_space_alloc_buffer()->undo_allocation(obj, word_sz);
duke@435	203	} else {
duke@435	204	SharedHeap::fill_region_with_object(MemRegion(obj, word_sz));
duke@435	205	}
duke@435	206	}
duke@435	207
duke@435	208	class ParScanThreadStateSet: private ResourceArray {
duke@435	209	public:
duke@435	210	// Initializes states for the specified number of threads;
duke@435	211	ParScanThreadStateSet(int num_threads,
duke@435	212	Space& to_space,
duke@435	213	ParNewGeneration& gen,
duke@435	214	Generation& old_gen,
duke@435	215	ObjToScanQueueSet& queue_set,
duke@435	216	size_t desired_plab_sz,
duke@435	217	ParallelTaskTerminator& term);
duke@435	218	inline ParScanThreadState& thread_sate(int i);
duke@435	219	int pushes() { return _pushes; }
duke@435	220	int pops() { return _pops; }
duke@435	221	int steals() { return _steals; }
duke@435	222	void reset();
duke@435	223	void flush();
duke@435	224	private:
duke@435	225	ParallelTaskTerminator& _term;
duke@435	226	ParNewGeneration& _gen;
duke@435	227	Generation& _next_gen;
duke@435	228	// staticstics
duke@435	229	int _pushes;
duke@435	230	int _pops;
duke@435	231	int _steals;
duke@435	232	};
duke@435	233
duke@435	234
duke@435	235	ParScanThreadStateSet::ParScanThreadStateSet(
duke@435	236	int num_threads, Space& to_space, ParNewGeneration& gen,
duke@435	237	Generation& old_gen, ObjToScanQueueSet& queue_set,
duke@435	238	size_t desired_plab_sz, ParallelTaskTerminator& term)
duke@435	239	: ResourceArray(sizeof(ParScanThreadState), num_threads),
duke@435	240	_gen(gen), _next_gen(old_gen), _term(term),
duke@435	241	_pushes(0), _pops(0), _steals(0)
duke@435	242	{
duke@435	243	assert(num_threads > 0, "sanity check!");
duke@435	244	// Initialize states.
duke@435	245	for (int i = 0; i < num_threads; ++i) {
duke@435	246	new ((ParScanThreadState*)_data + i)
duke@435	247	ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
duke@435	248	desired_plab_sz, term);
duke@435	249	}
duke@435	250	}
duke@435	251
duke@435	252	inline ParScanThreadState& ParScanThreadStateSet::thread_sate(int i)
duke@435	253	{
duke@435	254	assert(i >= 0 && i < length(), "sanity check!");
duke@435	255	return ((ParScanThreadState*)_data)[i];
duke@435	256	}
duke@435	257
duke@435	258
duke@435	259	void ParScanThreadStateSet::reset()
duke@435	260	{
duke@435	261	_term.reset_for_reuse();
duke@435	262	}
duke@435	263
duke@435	264	void ParScanThreadStateSet::flush()
duke@435	265	{
duke@435	266	for (int i = 0; i < length(); ++i) {
duke@435	267	ParScanThreadState& par_scan_state = thread_sate(i);
duke@435	268
duke@435	269	// Flush stats related to To-space PLAB activity and
duke@435	270	// retire the last buffer.
duke@435	271	par_scan_state.to_space_alloc_buffer()->
duke@435	272	flush_stats_and_retire(_gen.plab_stats(),
duke@435	273	false /* !retain */);
duke@435	274
duke@435	275	// Every thread has its own age table. We need to merge
duke@435	276	// them all into one.
duke@435	277	ageTable *local_table = par_scan_state.age_table();
duke@435	278	_gen.age_table()->merge(local_table);
duke@435	279
duke@435	280	// Inform old gen that we're done.
duke@435	281	_next_gen.par_promote_alloc_done(i);
duke@435	282	_next_gen.par_oop_since_save_marks_iterate_done(i);
duke@435	283
duke@435	284	// Flush stats related to work queue activity (push/pop/steal)
duke@435	285	// This could conceivably become a bottleneck; if so, we'll put the
duke@435	286	// stat's gathering under the flag.
duke@435	287	if (PAR_STATS_ENABLED) {
duke@435	288	_pushes += par_scan_state.pushes();
duke@435	289	_pops += par_scan_state.pops();
duke@435	290	_steals += par_scan_state.steals();
duke@435	291	if (ParallelGCVerbose) {
duke@435	292	gclog_or_tty->print("Thread %d complete:\n"
duke@435	293	" Pushes: %7d Pops: %7d Steals %7d (in %d attempts)\n",
duke@435	294	i, par_scan_state.pushes(), par_scan_state.pops(),
duke@435	295	par_scan_state.steals(), par_scan_state.steal_attempts());
duke@435	296	if (par_scan_state.overflow_pushes() > 0 ||
duke@435	297	par_scan_state.overflow_refills() > 0) {
duke@435	298	gclog_or_tty->print(" Overflow pushes: %7d "
duke@435	299	"Overflow refills: %7d for %d objs.\n",
duke@435	300	par_scan_state.overflow_pushes(),
duke@435	301	par_scan_state.overflow_refills(),
duke@435	302	par_scan_state.overflow_refill_objs());
duke@435	303	}
duke@435	304
duke@435	305	double elapsed = par_scan_state.elapsed();
duke@435	306	double strong_roots = par_scan_state.strong_roots_time();
duke@435	307	double term = par_scan_state.term_time();
duke@435	308	gclog_or_tty->print(
duke@435	309	" Elapsed: %7.2f ms.\n"
duke@435	310	" Strong roots: %7.2f ms (%6.2f%%)\n"
duke@435	311	" Termination: %7.2f ms (%6.2f%%) (in %d entries)\n",
duke@435	312	elapsed * 1000.0,
duke@435	313	strong_roots * 1000.0, (strong_roots*100.0/elapsed),
duke@435	314	term * 1000.0, (term*100.0/elapsed),
duke@435	315	par_scan_state.term_attempts());
duke@435	316	}
duke@435	317	}
duke@435	318	}
duke@435	319	}
duke@435	320
duke@435	321	ParScanClosure::ParScanClosure(ParNewGeneration* g,
duke@435	322	ParScanThreadState* par_scan_state) :
duke@435	323	OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g)
duke@435	324	{
duke@435	325	assert(_g->level() == 0, "Optimized for youngest generation");
duke@435	326	_boundary = _g->reserved().end();
duke@435	327	}
duke@435	328
coleenp@548	329	void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
coleenp@548	330	void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
coleenp@548	331
coleenp@548	332	void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
coleenp@548	333	void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
coleenp@548	334
coleenp@548	335	void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
coleenp@548	336	void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
coleenp@548	337
coleenp@548	338	void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
coleenp@548	339	void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
coleenp@548	340
duke@435	341	ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
duke@435	342	ParScanThreadState* par_scan_state)
duke@435	343	: ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
coleenp@548	344	{}
coleenp@548	345
coleenp@548	346	void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
coleenp@548	347	void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
duke@435	348
duke@435	349	#ifdef WIN32
duke@435	350	#pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
duke@435	351	#endif
duke@435	352
duke@435	353	ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
duke@435	354	ParScanThreadState* par_scan_state_,
duke@435	355	ParScanWithoutBarrierClosure* to_space_closure_,
duke@435	356	ParScanWithBarrierClosure* old_gen_closure_,
duke@435	357	ParRootScanWithoutBarrierClosure* to_space_root_closure_,
duke@435	358	ParNewGeneration* par_gen_,
duke@435	359	ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
duke@435	360	ObjToScanQueueSet* task_queues_,
duke@435	361	ParallelTaskTerminator* terminator_) :
duke@435	362
duke@435	363	_par_scan_state(par_scan_state_),
duke@435	364	_to_space_closure(to_space_closure_),
duke@435	365	_old_gen_closure(old_gen_closure_),
duke@435	366	_to_space_root_closure(to_space_root_closure_),
duke@435	367	_old_gen_root_closure(old_gen_root_closure_),
duke@435	368	_par_gen(par_gen_),
duke@435	369	_task_queues(task_queues_),
duke@435	370	_terminator(terminator_)
duke@435	371	{}
duke@435	372
duke@435	373	void ParEvacuateFollowersClosure::do_void() {
duke@435	374	ObjToScanQueue* work_q = par_scan_state()->work_queue();
duke@435	375
duke@435	376	while (true) {
duke@435	377
duke@435	378	// Scan to-space and old-gen objs until we run out of both.
duke@435	379	oop obj_to_scan;
duke@435	380	par_scan_state()->trim_queues(0);
duke@435	381
duke@435	382	// We have no local work, attempt to steal from other threads.
duke@435	383
duke@435	384	// attempt to steal work from promoted.
duke@435	385	par_scan_state()->note_steal_attempt();
duke@435	386	if (task_queues()->steal(par_scan_state()->thread_num(),
duke@435	387	par_scan_state()->hash_seed(),
duke@435	388	obj_to_scan)) {
duke@435	389	par_scan_state()->note_steal();
duke@435	390	bool res = work_q->push(obj_to_scan);
duke@435	391	assert(res, "Empty queue should have room for a push.");
duke@435	392
duke@435	393	par_scan_state()->note_push();
duke@435	394	// if successful, goto Start.
duke@435	395	continue;
duke@435	396
duke@435	397	// try global overflow list.
duke@435	398	} else if (par_gen()->take_from_overflow_list(par_scan_state())) {
duke@435	399	continue;
duke@435	400	}
duke@435	401
duke@435	402	// Otherwise, offer termination.
duke@435	403	par_scan_state()->start_term_time();
duke@435	404	if (terminator()->offer_termination()) break;
duke@435	405	par_scan_state()->end_term_time();
duke@435	406	}
duke@435	407	// Finish the last termination pause.
duke@435	408	par_scan_state()->end_term_time();
duke@435	409	}
duke@435	410
duke@435	411	ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen,
duke@435	412	HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
duke@435	413	AbstractGangTask("ParNewGeneration collection"),
duke@435	414	_gen(gen), _next_gen(next_gen),
duke@435	415	_young_old_boundary(young_old_boundary),
duke@435	416	_state_set(state_set)
duke@435	417	{}
duke@435	418
duke@435	419	void ParNewGenTask::work(int i) {
duke@435	420	GenCollectedHeap* gch = GenCollectedHeap::heap();
duke@435	421	// Since this is being done in a separate thread, need new resource
duke@435	422	// and handle marks.
duke@435	423	ResourceMark rm;
duke@435	424	HandleMark hm;
duke@435	425	// We would need multiple old-gen queues otherwise.
duke@435	426	guarantee(gch->n_gens() == 2,
duke@435	427	"Par young collection currently only works with one older gen.");
duke@435	428
duke@435	429	Generation* old_gen = gch->next_gen(_gen);
duke@435	430
duke@435	431	ParScanThreadState& par_scan_state = _state_set->thread_sate(i);
duke@435	432	par_scan_state.set_young_old_boundary(_young_old_boundary);
duke@435	433
duke@435	434	par_scan_state.start_strong_roots();
duke@435	435	gch->gen_process_strong_roots(_gen->level(),
duke@435	436	true, // Process younger gens, if any,
duke@435	437	// as strong roots.
duke@435	438	false,// not collecting perm generation.
duke@435	439	SharedHeap::SO_AllClasses,
duke@435	440	&par_scan_state.older_gen_closure(),
duke@435	441	&par_scan_state.to_space_root_closure());
duke@435	442	par_scan_state.end_strong_roots();
duke@435	443
duke@435	444	// "evacuate followers".
duke@435	445	par_scan_state.evacuate_followers_closure().do_void();
duke@435	446	}
duke@435	447
duke@435	448	#ifdef _MSC_VER
duke@435	449	#pragma warning( push )
duke@435	450	#pragma warning( disable:4355 ) // 'this' : used in base member initializer list
duke@435	451	#endif
duke@435	452	ParNewGeneration::
duke@435	453	ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
duke@435	454	: DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
duke@435	455	_overflow_list(NULL),
duke@435	456	_is_alive_closure(this),
duke@435	457	_plab_stats(YoungPLABSize, PLABWeight)
duke@435	458	{
duke@435	459	_task_queues = new ObjToScanQueueSet(ParallelGCThreads);
duke@435	460	guarantee(_task_queues != NULL, "task_queues allocation failure.");
duke@435	461
duke@435	462	for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
duke@435	463	ObjToScanQueuePadded *q_padded = new ObjToScanQueuePadded();
duke@435	464	guarantee(q_padded != NULL, "work_queue Allocation failure.");
duke@435	465
duke@435	466	_task_queues->register_queue(i1, &q_padded->work_queue);
duke@435	467	}
duke@435	468
duke@435	469	for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
duke@435	470	_task_queues->queue(i2)->initialize();
duke@435	471
duke@435	472	if (UsePerfData) {
duke@435	473	EXCEPTION_MARK;
duke@435	474	ResourceMark rm;
duke@435	475
duke@435	476	const char* cname =
duke@435	477	PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
duke@435	478	PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
duke@435	479	ParallelGCThreads, CHECK);
duke@435	480	}
duke@435	481	}
duke@435	482	#ifdef _MSC_VER
duke@435	483	#pragma warning( pop )
duke@435	484	#endif
duke@435	485
duke@435	486	// ParNewGeneration::
duke@435	487	ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
duke@435	488	DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
duke@435	489
coleenp@548	490	template <class T>
coleenp@548	491	void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
coleenp@548	492	#ifdef ASSERT
coleenp@548	493	{
coleenp@548	494	assert(!oopDesc::is_null(*p), "expected non-null ref");
coleenp@548	495	oop obj = oopDesc::load_decode_heap_oop_not_null(p);
coleenp@548	496	// We never expect to see a null reference being processed
coleenp@548	497	// as a weak reference.
coleenp@548	498	assert(obj->is_oop(), "expected an oop while scanning weak refs");
coleenp@548	499	}
coleenp@548	500	#endif // ASSERT
duke@435	501
duke@435	502	_par_cl->do_oop_nv(p);
duke@435	503
duke@435	504	if (Universe::heap()->is_in_reserved(p)) {
coleenp@548	505	oop obj = oopDesc::load_decode_heap_oop_not_null(p);
coleenp@548	506	_rs->write_ref_field_gc_par(p, obj);
duke@435	507	}
duke@435	508	}
duke@435	509
coleenp@548	510	void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
coleenp@548	511	void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
coleenp@548	512
duke@435	513	// ParNewGeneration::
duke@435	514	KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
duke@435	515	DefNewGeneration::KeepAliveClosure(cl) {}
duke@435	516
coleenp@548	517	template <class T>
coleenp@548	518	void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
coleenp@548	519	#ifdef ASSERT
coleenp@548	520	{
coleenp@548	521	assert(!oopDesc::is_null(*p), "expected non-null ref");
coleenp@548	522	oop obj = oopDesc::load_decode_heap_oop_not_null(p);
coleenp@548	523	// We never expect to see a null reference being processed
coleenp@548	524	// as a weak reference.
coleenp@548	525	assert(obj->is_oop(), "expected an oop while scanning weak refs");
coleenp@548	526	}
coleenp@548	527	#endif // ASSERT
duke@435	528
duke@435	529	_cl->do_oop_nv(p);
duke@435	530
duke@435	531	if (Universe::heap()->is_in_reserved(p)) {
coleenp@548	532	oop obj = oopDesc::load_decode_heap_oop_not_null(p);
coleenp@548	533	_rs->write_ref_field_gc_par(p, obj);
duke@435	534	}
duke@435	535	}
duke@435	536
coleenp@548	537	void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
coleenp@548	538	void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
coleenp@548	539
coleenp@548	540	template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
coleenp@548	541	T heap_oop = oopDesc::load_heap_oop(p);
coleenp@548	542	if (!oopDesc::is_null(heap_oop)) {
coleenp@548	543	oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
duke@435	544	if ((HeapWord*)obj < _boundary) {
duke@435	545	assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
coleenp@548	546	oop new_obj = obj->is_forwarded()
coleenp@548	547	? obj->forwardee()
coleenp@548	548	: _g->DefNewGeneration::copy_to_survivor_space(obj);
coleenp@548	549	oopDesc::encode_store_heap_oop_not_null(p, new_obj);
duke@435	550	}
duke@435	551	if (_gc_barrier) {
duke@435	552	// If p points to a younger generation, mark the card.
duke@435	553	if ((HeapWord*)obj < _gen_boundary) {
duke@435	554	_rs->write_ref_field_gc_par(p, obj);
duke@435	555	}
duke@435	556	}
duke@435	557	}
duke@435	558	}
duke@435	559
coleenp@548	560	void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
coleenp@548	561	void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
coleenp@548	562
duke@435	563	class ParNewRefProcTaskProxy: public AbstractGangTask {
duke@435	564	typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
duke@435	565	public:
duke@435	566	ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen,
duke@435	567	Generation& next_gen,
duke@435	568	HeapWord* young_old_boundary,
duke@435	569	ParScanThreadStateSet& state_set);
duke@435	570
duke@435	571	private:
duke@435	572	virtual void work(int i);
duke@435	573
duke@435	574	private:
duke@435	575	ParNewGeneration& _gen;
duke@435	576	ProcessTask& _task;
duke@435	577	Generation& _next_gen;
duke@435	578	HeapWord* _young_old_boundary;
duke@435	579	ParScanThreadStateSet& _state_set;
duke@435	580	};
duke@435	581
duke@435	582	ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(
duke@435	583	ProcessTask& task, ParNewGeneration& gen,
duke@435	584	Generation& next_gen,
duke@435	585	HeapWord* young_old_boundary,
duke@435	586	ParScanThreadStateSet& state_set)
duke@435	587	: AbstractGangTask("ParNewGeneration parallel reference processing"),
duke@435	588	_gen(gen),
duke@435	589	_task(task),
duke@435	590	_next_gen(next_gen),
duke@435	591	_young_old_boundary(young_old_boundary),
duke@435	592	_state_set(state_set)
duke@435	593	{
duke@435	594	}
duke@435	595
duke@435	596	void ParNewRefProcTaskProxy::work(int i)
duke@435	597	{
duke@435	598	ResourceMark rm;
duke@435	599	HandleMark hm;
duke@435	600	ParScanThreadState& par_scan_state = _state_set.thread_sate(i);
duke@435	601	par_scan_state.set_young_old_boundary(_young_old_boundary);
duke@435	602	_task.work(i, par_scan_state.is_alive_closure(),
duke@435	603	par_scan_state.keep_alive_closure(),
duke@435	604	par_scan_state.evacuate_followers_closure());
duke@435	605	}
duke@435	606
duke@435	607	class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
duke@435	608	typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
duke@435	609	EnqueueTask& _task;
duke@435	610
duke@435	611	public:
duke@435	612	ParNewRefEnqueueTaskProxy(EnqueueTask& task)
duke@435	613	: AbstractGangTask("ParNewGeneration parallel reference enqueue"),
duke@435	614	_task(task)
duke@435	615	{ }
duke@435	616
duke@435	617	virtual void work(int i)
duke@435	618	{
duke@435	619	_task.work(i);
duke@435	620	}
duke@435	621	};
duke@435	622
duke@435	623
duke@435	624	void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
duke@435	625	{
duke@435	626	GenCollectedHeap* gch = GenCollectedHeap::heap();
duke@435	627	assert(gch->kind() == CollectedHeap::GenCollectedHeap,
duke@435	628	"not a generational heap");
duke@435	629	WorkGang* workers = gch->workers();
duke@435	630	assert(workers != NULL, "Need parallel worker threads.");
duke@435	631	ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
duke@435	632	_generation.reserved().end(), _state_set);
duke@435	633	workers->run_task(&rp_task);
duke@435	634	_state_set.reset();
duke@435	635	}
duke@435	636
duke@435	637	void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
duke@435	638	{
duke@435	639	GenCollectedHeap* gch = GenCollectedHeap::heap();
duke@435	640	WorkGang* workers = gch->workers();
duke@435	641	assert(workers != NULL, "Need parallel worker threads.");
duke@435	642	ParNewRefEnqueueTaskProxy enq_task(task);
duke@435	643	workers->run_task(&enq_task);
duke@435	644	}
duke@435	645
duke@435	646	void ParNewRefProcTaskExecutor::set_single_threaded_mode()
duke@435	647	{
duke@435	648	_state_set.flush();
duke@435	649	GenCollectedHeap* gch = GenCollectedHeap::heap();
duke@435	650	gch->set_par_threads(0); // 0 ==> non-parallel.
duke@435	651	gch->save_marks();
duke@435	652	}
duke@435	653
duke@435	654	ScanClosureWithParBarrier::
duke@435	655	ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
duke@435	656	ScanClosure(g, gc_barrier) {}
duke@435	657
duke@435	658	EvacuateFollowersClosureGeneral::
duke@435	659	EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
duke@435	660	OopsInGenClosure* cur,
duke@435	661	OopsInGenClosure* older) :
duke@435	662	_gch(gch), _level(level),
duke@435	663	_scan_cur_or_nonheap(cur), _scan_older(older)
duke@435	664	{}
duke@435	665
duke@435	666	void EvacuateFollowersClosureGeneral::do_void() {
duke@435	667	do {
duke@435	668	// Beware: this call will lead to closure applications via virtual
duke@435	669	// calls.
duke@435	670	_gch->oop_since_save_marks_iterate(_level,
duke@435	671	_scan_cur_or_nonheap,
duke@435	672	_scan_older);
duke@435	673	} while (!_gch->no_allocs_since_save_marks(_level));
duke@435	674	}
duke@435	675
duke@435	676
duke@435	677	bool ParNewGeneration::_avoid_promotion_undo = false;
duke@435	678
duke@435	679	void ParNewGeneration::adjust_desired_tenuring_threshold() {
duke@435	680	// Set the desired survivor size to half the real survivor space
duke@435	681	_tenuring_threshold =
duke@435	682	age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize);
duke@435	683	}
duke@435	684
duke@435	685	// A Generation that does parallel young-gen collection.
duke@435	686
duke@435	687	void ParNewGeneration::collect(bool full,
duke@435	688	bool clear_all_soft_refs,
duke@435	689	size_t size,
duke@435	690	bool is_tlab) {
duke@435	691	assert(full || size > 0, "otherwise we don't want to collect");
duke@435	692	GenCollectedHeap* gch = GenCollectedHeap::heap();
duke@435	693	assert(gch->kind() == CollectedHeap::GenCollectedHeap,
duke@435	694	"not a CMS generational heap");
duke@435	695	AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
duke@435	696	WorkGang* workers = gch->workers();
duke@435	697	_next_gen = gch->next_gen(this);
duke@435	698	assert(_next_gen != NULL,
duke@435	699	"This must be the youngest gen, and not the only gen");
duke@435	700	assert(gch->n_gens() == 2,
duke@435	701	"Par collection currently only works with single older gen.");
duke@435	702	// Do we have to avoid promotion_undo?
duke@435	703	if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) {
duke@435	704	set_avoid_promotion_undo(true);
duke@435	705	}
duke@435	706
duke@435	707	// If the next generation is too full to accomodate worst-case promotion
duke@435	708	// from this generation, pass on collection; let the next generation
duke@435	709	// do it.
duke@435	710	if (!collection_attempt_is_safe()) {
duke@435	711	gch->set_incremental_collection_will_fail();
duke@435	712	return;
duke@435	713	}
duke@435	714	assert(to()->is_empty(), "Else not collection_attempt_is_safe");
duke@435	715
duke@435	716	init_assuming_no_promotion_failure();
duke@435	717
duke@435	718	if (UseAdaptiveSizePolicy) {
duke@435	719	set_survivor_overflow(false);
duke@435	720	size_policy->minor_collection_begin();
duke@435	721	}
duke@435	722
duke@435	723	TraceTime t1("GC", PrintGC && !PrintGCDetails, true, gclog_or_tty);
duke@435	724	// Capture heap used before collection (for printing).
duke@435	725	size_t gch_prev_used = gch->used();
duke@435	726
duke@435	727	SpecializationStats::clear();
duke@435	728
duke@435	729	age_table()->clear();
duke@435	730	to()->clear();
duke@435	731
duke@435	732	gch->save_marks();
duke@435	733	assert(workers != NULL, "Need parallel worker threads.");
duke@435	734	ParallelTaskTerminator _term(workers->total_workers(), task_queues());
duke@435	735	ParScanThreadStateSet thread_state_set(workers->total_workers(),
duke@435	736	*to(), *this, *_next_gen, *task_queues(),
duke@435	737	desired_plab_sz(), _term);
duke@435	738
duke@435	739	ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set);
duke@435	740	int n_workers = workers->total_workers();
duke@435	741	gch->set_par_threads(n_workers);
duke@435	742	gch->change_strong_roots_parity();
duke@435	743	gch->rem_set()->prepare_for_younger_refs_iterate(true);
duke@435	744	// It turns out that even when we're using 1 thread, doing the work in a
duke@435	745	// separate thread causes wide variance in run times. We can't help this
duke@435	746	// in the multi-threaded case, but we special-case n=1 here to get
duke@435	747	// repeatable measurements of the 1-thread overhead of the parallel code.
duke@435	748	if (n_workers > 1) {
duke@435	749	workers->run_task(&tsk);
duke@435	750	} else {
duke@435	751	tsk.work(0);
duke@435	752	}
duke@435	753	thread_state_set.reset();
duke@435	754
duke@435	755	if (PAR_STATS_ENABLED && ParallelGCVerbose) {
duke@435	756	gclog_or_tty->print("Thread totals:\n"
duke@435	757	" Pushes: %7d Pops: %7d Steals %7d (sum = %7d).\n",
duke@435	758	thread_state_set.pushes(), thread_state_set.pops(),
duke@435	759	thread_state_set.steals(),
duke@435	760	thread_state_set.pops()+thread_state_set.steals());
duke@435	761	}
duke@435	762	assert(thread_state_set.pushes() == thread_state_set.pops() + thread_state_set.steals(),
duke@435	763	"Or else the queues are leaky.");
duke@435	764
duke@435	765	// For now, process discovered weak refs sequentially.
duke@435	766	#ifdef COMPILER2
duke@435	767	ReferencePolicy *soft_ref_policy = new LRUMaxHeapPolicy();
duke@435	768	#else
duke@435	769	ReferencePolicy *soft_ref_policy = new LRUCurrentHeapPolicy();
duke@435	770	#endif // COMPILER2
duke@435	771
duke@435	772	// Process (weak) reference objects found during scavenge.
duke@435	773	IsAliveClosure is_alive(this);
duke@435	774	ScanWeakRefClosure scan_weak_ref(this);
duke@435	775	KeepAliveClosure keep_alive(&scan_weak_ref);
duke@435	776	ScanClosure scan_without_gc_barrier(this, false);
duke@435	777	ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
duke@435	778	set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
duke@435	779	EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
duke@435	780	&scan_without_gc_barrier, &scan_with_gc_barrier);
duke@435	781	if (ref_processor()->processing_is_mt()) {
duke@435	782	ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
duke@435	783	ref_processor()->process_discovered_references(
duke@435	784	soft_ref_policy, &is_alive, &keep_alive, &evacuate_followers,
duke@435	785	&task_executor);
duke@435	786	} else {
duke@435	787	thread_state_set.flush();
duke@435	788	gch->set_par_threads(0); // 0 ==> non-parallel.
duke@435	789	gch->save_marks();
duke@435	790	ref_processor()->process_discovered_references(
duke@435	791	soft_ref_policy, &is_alive, &keep_alive, &evacuate_followers,
duke@435	792	NULL);
duke@435	793	}
duke@435	794	if (!promotion_failed()) {
duke@435	795	// Swap the survivor spaces.
duke@435	796	eden()->clear();
duke@435	797	from()->clear();
duke@435	798	swap_spaces();
duke@435	799
duke@435	800	assert(to()->is_empty(), "to space should be empty now");
duke@435	801	} else {
duke@435	802	assert(HandlePromotionFailure,
duke@435	803	"Should only be here if promotion failure handling is on");
duke@435	804	if (_promo_failure_scan_stack != NULL) {
duke@435	805	// Can be non-null because of reference processing.
duke@435	806	// Free stack with its elements.
duke@435	807	delete _promo_failure_scan_stack;
duke@435	808	_promo_failure_scan_stack = NULL;
duke@435	809	}
duke@435	810	remove_forwarding_pointers();
duke@435	811	if (PrintGCDetails) {
duke@435	812	gclog_or_tty->print(" (promotion failed)");
duke@435	813	}
duke@435	814	// All the spaces are in play for mark-sweep.
duke@435	815	swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
duke@435	816	from()->set_next_compaction_space(to());
duke@435	817	gch->set_incremental_collection_will_fail();
jmasa@441	818
jmasa@441	819	// Reset the PromotionFailureALot counters.
jmasa@441	820	NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
duke@435	821	}
duke@435	822	// set new iteration safe limit for the survivor spaces
duke@435	823	from()->set_concurrent_iteration_safe_limit(from()->top());
duke@435	824	to()->set_concurrent_iteration_safe_limit(to()->top());
duke@435	825
duke@435	826	adjust_desired_tenuring_threshold();
duke@435	827	if (ResizePLAB) {
duke@435	828	plab_stats()->adjust_desired_plab_sz();
duke@435	829	}
duke@435	830
duke@435	831	if (PrintGC && !PrintGCDetails) {
duke@435	832	gch->print_heap_change(gch_prev_used);
duke@435	833	}
duke@435	834
duke@435	835	if (UseAdaptiveSizePolicy) {
duke@435	836	size_policy->minor_collection_end(gch->gc_cause());
duke@435	837	size_policy->avg_survived()->sample(from()->used());
duke@435	838	}
duke@435	839
duke@435	840	update_time_of_last_gc(os::javaTimeMillis());
duke@435	841
duke@435	842	SpecializationStats::print();
duke@435	843
duke@435	844	ref_processor()->set_enqueuing_is_done(true);
duke@435	845	if (ref_processor()->processing_is_mt()) {
duke@435	846	ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
duke@435	847	ref_processor()->enqueue_discovered_references(&task_executor);
duke@435	848	} else {
duke@435	849	ref_processor()->enqueue_discovered_references(NULL);
duke@435	850	}
duke@435	851	ref_processor()->verify_no_references_recorded();
duke@435	852	}
duke@435	853
duke@435	854	static int sum;
duke@435	855	void ParNewGeneration::waste_some_time() {
duke@435	856	for (int i = 0; i < 100; i++) {
duke@435	857	sum += i;
duke@435	858	}
duke@435	859	}
duke@435	860
duke@435	861	static const oop ClaimedForwardPtr = oop(0x4);
duke@435	862
duke@435	863	// Because of concurrency, there are times where an object for which
duke@435	864	// "is_forwarded()" is true contains an "interim" forwarding pointer
duke@435	865	// value. Such a value will soon be overwritten with a real value.
duke@435	866	// This method requires "obj" to have a forwarding pointer, and waits, if
duke@435	867	// necessary for a real one to be inserted, and returns it.
duke@435	868
duke@435	869	oop ParNewGeneration::real_forwardee(oop obj) {
duke@435	870	oop forward_ptr = obj->forwardee();
duke@435	871	if (forward_ptr != ClaimedForwardPtr) {
duke@435	872	return forward_ptr;
duke@435	873	} else {
duke@435	874	return real_forwardee_slow(obj);
duke@435	875	}
duke@435	876	}
duke@435	877
duke@435	878	oop ParNewGeneration::real_forwardee_slow(oop obj) {
duke@435	879	// Spin-read if it is claimed but not yet written by another thread.
duke@435	880	oop forward_ptr = obj->forwardee();
duke@435	881	while (forward_ptr == ClaimedForwardPtr) {
duke@435	882	waste_some_time();
duke@435	883	assert(obj->is_forwarded(), "precondition");
duke@435	884	forward_ptr = obj->forwardee();
duke@435	885	}
duke@435	886	return forward_ptr;
duke@435	887	}
duke@435	888
duke@435	889	#ifdef ASSERT
duke@435	890	bool ParNewGeneration::is_legal_forward_ptr(oop p) {
duke@435	891	return
duke@435	892	(_avoid_promotion_undo && p == ClaimedForwardPtr)
duke@435	893	|| Universe::heap()->is_in_reserved(p);
duke@435	894	}
duke@435	895	#endif
duke@435	896
duke@435	897	void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
duke@435	898	if ((m != markOopDesc::prototype()) &&
duke@435	899	(!UseBiasedLocking || (m != markOopDesc::biased_locking_prototype()))) {
duke@435	900	MutexLocker ml(ParGCRareEvent_lock);
duke@435	901	DefNewGeneration::preserve_mark_if_necessary(obj, m);
duke@435	902	}
duke@435	903	}
duke@435	904
duke@435	905	// Multiple GC threads may try to promote an object. If the object
duke@435	906	// is successfully promoted, a forwarding pointer will be installed in
duke@435	907	// the object in the young generation. This method claims the right
duke@435	908	// to install the forwarding pointer before it copies the object,
duke@435	909	// thus avoiding the need to undo the copy as in
duke@435	910	// copy_to_survivor_space_avoiding_with_undo.
duke@435	911
duke@435	912	oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
duke@435	913	ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
duke@435	914	// In the sequential version, this assert also says that the object is
duke@435	915	// not forwarded. That might not be the case here. It is the case that
duke@435	916	// the caller observed it to be not forwarded at some time in the past.
duke@435	917	assert(is_in_reserved(old), "shouldn't be scavenging this oop");
duke@435	918
duke@435	919	// The sequential code read "old->age()" below. That doesn't work here,
duke@435	920	// since the age is in the mark word, and that might be overwritten with
duke@435	921	// a forwarding pointer by a parallel thread. So we must save the mark
duke@435	922	// word in a local and then analyze it.
duke@435	923	oopDesc dummyOld;
duke@435	924	dummyOld.set_mark(m);
duke@435	925	assert(!dummyOld.is_forwarded(),
duke@435	926	"should not be called with forwarding pointer mark word.");
duke@435	927
duke@435	928	oop new_obj = NULL;
duke@435	929	oop forward_ptr;
duke@435	930
duke@435	931	// Try allocating obj in to-space (unless too old)
duke@435	932	if (dummyOld.age() < tenuring_threshold()) {
duke@435	933	new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
duke@435	934	if (new_obj == NULL) {
duke@435	935	set_survivor_overflow(true);
duke@435	936	}
duke@435	937	}
duke@435	938
duke@435	939	if (new_obj == NULL) {
duke@435	940	// Either to-space is full or we decided to promote
duke@435	941	// try allocating obj tenured
duke@435	942
duke@435	943	// Attempt to install a null forwarding pointer (atomically),
duke@435	944	// to claim the right to install the real forwarding pointer.
duke@435	945	forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
duke@435	946	if (forward_ptr != NULL) {
duke@435	947	// someone else beat us to it.
duke@435	948	return real_forwardee(old);
duke@435	949	}
duke@435	950
duke@435	951	new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
duke@435	952	old, m, sz);
duke@435	953
duke@435	954	if (new_obj == NULL) {
duke@435	955	if (!HandlePromotionFailure) {
duke@435	956	// A failed promotion likely means the MaxLiveObjectEvacuationRatio flag
duke@435	957	// is incorrectly set. In any case, its seriously wrong to be here!
duke@435	958	vm_exit_out_of_memory(sz*wordSize, "promotion");
duke@435	959	}
duke@435	960	// promotion failed, forward to self
duke@435	961	_promotion_failed = true;
duke@435	962	new_obj = old;
duke@435	963
duke@435	964	preserve_mark_if_necessary(old, m);
duke@435	965	}
duke@435	966
duke@435	967	old->forward_to(new_obj);
duke@435	968	forward_ptr = NULL;
duke@435	969	} else {
duke@435	970	// Is in to-space; do copying ourselves.
duke@435	971	Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
duke@435	972	forward_ptr = old->forward_to_atomic(new_obj);
duke@435	973	// Restore the mark word copied above.
duke@435	974	new_obj->set_mark(m);
duke@435	975	// Increment age if obj still in new generation
duke@435	976	new_obj->incr_age();
duke@435	977	par_scan_state->age_table()->add(new_obj, sz);
duke@435	978	}
duke@435	979	assert(new_obj != NULL, "just checking");
duke@435	980
duke@435	981	if (forward_ptr == NULL) {
duke@435	982	oop obj_to_push = new_obj;
duke@435	983	if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
duke@435	984	// Length field used as index of next element to be scanned.
duke@435	985	// Real length can be obtained from real_forwardee()
duke@435	986	arrayOop(old)->set_length(0);
duke@435	987	obj_to_push = old;
duke@435	988	assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
duke@435	989	"push forwarded object");
duke@435	990	}
duke@435	991	// Push it on one of the queues of to-be-scanned objects.
duke@435	992	if (!par_scan_state->work_queue()->push(obj_to_push)) {
duke@435	993	// Add stats for overflow pushes.
duke@435	994	if (Verbose && PrintGCDetails) {
duke@435	995	gclog_or_tty->print("queue overflow!\n");
duke@435	996	}
duke@435	997	push_on_overflow_list(old);
duke@435	998	par_scan_state->note_overflow_push();
duke@435	999	}
duke@435	1000	par_scan_state->note_push();
duke@435	1001
duke@435	1002	return new_obj;
duke@435	1003	}
duke@435	1004
duke@435	1005	// Oops. Someone beat us to it. Undo the allocation. Where did we
duke@435	1006	// allocate it?
duke@435	1007	if (is_in_reserved(new_obj)) {
duke@435	1008	// Must be in to_space.
duke@435	1009	assert(to()->is_in_reserved(new_obj), "Checking");
duke@435	1010	if (forward_ptr == ClaimedForwardPtr) {
duke@435	1011	// Wait to get the real forwarding pointer value.
duke@435	1012	forward_ptr = real_forwardee(old);
duke@435	1013	}
duke@435	1014	par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
duke@435	1015	}
duke@435	1016
duke@435	1017	return forward_ptr;
duke@435	1018	}
duke@435	1019
duke@435	1020
duke@435	1021	// Multiple GC threads may try to promote the same object. If two
duke@435	1022	// or more GC threads copy the object, only one wins the race to install
duke@435	1023	// the forwarding pointer. The other threads have to undo their copy.
duke@435	1024
duke@435	1025	oop ParNewGeneration::copy_to_survivor_space_with_undo(
duke@435	1026	ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
duke@435	1027
duke@435	1028	// In the sequential version, this assert also says that the object is
duke@435	1029	// not forwarded. That might not be the case here. It is the case that
duke@435	1030	// the caller observed it to be not forwarded at some time in the past.
duke@435	1031	assert(is_in_reserved(old), "shouldn't be scavenging this oop");
duke@435	1032
duke@435	1033	// The sequential code read "old->age()" below. That doesn't work here,
duke@435	1034	// since the age is in the mark word, and that might be overwritten with
duke@435	1035	// a forwarding pointer by a parallel thread. So we must save the mark
duke@435	1036	// word here, install it in a local oopDesc, and then analyze it.
duke@435	1037	oopDesc dummyOld;
duke@435	1038	dummyOld.set_mark(m);
duke@435	1039	assert(!dummyOld.is_forwarded(),
duke@435	1040	"should not be called with forwarding pointer mark word.");
duke@435	1041
duke@435	1042	bool failed_to_promote = false;
duke@435	1043	oop new_obj = NULL;
duke@435	1044	oop forward_ptr;
duke@435	1045
duke@435	1046	// Try allocating obj in to-space (unless too old)
duke@435	1047	if (dummyOld.age() < tenuring_threshold()) {
duke@435	1048	new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
duke@435	1049	if (new_obj == NULL) {
duke@435	1050	set_survivor_overflow(true);
duke@435	1051	}
duke@435	1052	}
duke@435	1053
duke@435	1054	if (new_obj == NULL) {
duke@435	1055	// Either to-space is full or we decided to promote
duke@435	1056	// try allocating obj tenured
duke@435	1057	new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
duke@435	1058	old, m, sz);
duke@435	1059
duke@435	1060	if (new_obj == NULL) {
duke@435	1061	if (!HandlePromotionFailure) {
duke@435	1062	// A failed promotion likely means the MaxLiveObjectEvacuationRatio
duke@435	1063	// flag is incorrectly set. In any case, its seriously wrong to be
duke@435	1064	// here!
duke@435	1065	vm_exit_out_of_memory(sz*wordSize, "promotion");
duke@435	1066	}
duke@435	1067	// promotion failed, forward to self
duke@435	1068	forward_ptr = old->forward_to_atomic(old);
duke@435	1069	new_obj = old;
duke@435	1070
duke@435	1071	if (forward_ptr != NULL) {
duke@435	1072	return forward_ptr; // someone else succeeded
duke@435	1073	}
duke@435	1074
duke@435	1075	_promotion_failed = true;
duke@435	1076	failed_to_promote = true;
duke@435	1077
duke@435	1078	preserve_mark_if_necessary(old, m);
duke@435	1079	}
duke@435	1080	} else {
duke@435	1081	// Is in to-space; do copying ourselves.
duke@435	1082	Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
duke@435	1083	// Restore the mark word copied above.
duke@435	1084	new_obj->set_mark(m);
duke@435	1085	// Increment age if new_obj still in new generation
duke@435	1086	new_obj->incr_age();
duke@435	1087	par_scan_state->age_table()->add(new_obj, sz);
duke@435	1088	}
duke@435	1089	assert(new_obj != NULL, "just checking");
duke@435	1090
duke@435	1091	// Now attempt to install the forwarding pointer (atomically).
duke@435	1092	// We have to copy the mark word before overwriting with forwarding
duke@435	1093	// ptr, so we can restore it below in the copy.
duke@435	1094	if (!failed_to_promote) {
duke@435	1095	forward_ptr = old->forward_to_atomic(new_obj);
duke@435	1096	}
duke@435	1097
duke@435	1098	if (forward_ptr == NULL) {
duke@435	1099	oop obj_to_push = new_obj;
duke@435	1100	if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
duke@435	1101	// Length field used as index of next element to be scanned.
duke@435	1102	// Real length can be obtained from real_forwardee()
duke@435	1103	arrayOop(old)->set_length(0);
duke@435	1104	obj_to_push = old;
duke@435	1105	assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
duke@435	1106	"push forwarded object");
duke@435	1107	}
duke@435	1108	// Push it on one of the queues of to-be-scanned objects.
duke@435	1109	if (!par_scan_state->work_queue()->push(obj_to_push)) {
duke@435	1110	// Add stats for overflow pushes.
duke@435	1111	push_on_overflow_list(old);
duke@435	1112	par_scan_state->note_overflow_push();
duke@435	1113	}
duke@435	1114	par_scan_state->note_push();
duke@435	1115
duke@435	1116	return new_obj;
duke@435	1117	}
duke@435	1118
duke@435	1119	// Oops. Someone beat us to it. Undo the allocation. Where did we
duke@435	1120	// allocate it?
duke@435	1121	if (is_in_reserved(new_obj)) {
duke@435	1122	// Must be in to_space.
duke@435	1123	assert(to()->is_in_reserved(new_obj), "Checking");
duke@435	1124	par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
duke@435	1125	} else {
duke@435	1126	assert(!_avoid_promotion_undo, "Should not be here if avoiding.");
duke@435	1127	_next_gen->par_promote_alloc_undo(par_scan_state->thread_num(),
duke@435	1128	(HeapWord*)new_obj, sz);
duke@435	1129	}
duke@435	1130
duke@435	1131	return forward_ptr;
duke@435	1132	}
duke@435	1133
duke@435	1134	void ParNewGeneration::push_on_overflow_list(oop from_space_obj) {
duke@435	1135	oop cur_overflow_list = _overflow_list;
duke@435	1136	// if the object has been forwarded to itself, then we cannot
duke@435	1137	// use the klass pointer for the linked list. Instead we have
duke@435	1138	// to allocate an oopDesc in the C-Heap and use that for the linked list.
duke@435	1139	if (from_space_obj->forwardee() == from_space_obj) {
duke@435	1140	oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1);
duke@435	1141	listhead->forward_to(from_space_obj);
duke@435	1142	from_space_obj = listhead;
duke@435	1143	}
duke@435	1144	while (true) {
duke@435	1145	from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
duke@435	1146	oop observed_overflow_list =
duke@435	1147	(oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
duke@435	1148	if (observed_overflow_list == cur_overflow_list) break;
duke@435	1149	// Otherwise...
duke@435	1150	cur_overflow_list = observed_overflow_list;
duke@435	1151	}
duke@435	1152	}
duke@435	1153
duke@435	1154	bool
duke@435	1155	ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
duke@435	1156	ObjToScanQueue* work_q = par_scan_state->work_queue();
duke@435	1157	// How many to take?
duke@435	1158	int objsFromOverflow = MIN2(work_q->max_elems()/4,
duke@435	1159	(juint)ParGCDesiredObjsFromOverflowList);
duke@435	1160
duke@435	1161	if (_overflow_list == NULL) return false;
duke@435	1162
duke@435	1163	// Otherwise, there was something there; try claiming the list.
duke@435	1164	oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
duke@435	1165
duke@435	1166	if (prefix == NULL) {
duke@435	1167	return false;
duke@435	1168	}
duke@435	1169	// Trim off a prefix of at most objsFromOverflow items
duke@435	1170	int i = 1;
duke@435	1171	oop cur = prefix;
coleenp@602	1172	while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
duke@435	1173	i++; cur = oop(cur->klass());
duke@435	1174	}
duke@435	1175
duke@435	1176	// Reattach remaining (suffix) to overflow list
coleenp@602	1177	if (cur->klass_or_null() != NULL) {
duke@435	1178	oop suffix = oop(cur->klass());
duke@435	1179	cur->set_klass_to_list_ptr(NULL);
duke@435	1180
duke@435	1181	// Find last item of suffix list
duke@435	1182	oop last = suffix;
coleenp@602	1183	while (last->klass_or_null() != NULL) {
duke@435	1184	last = oop(last->klass());
duke@435	1185	}
duke@435	1186	// Atomically prepend suffix to current overflow list
duke@435	1187	oop cur_overflow_list = _overflow_list;
duke@435	1188	while (true) {
duke@435	1189	last->set_klass_to_list_ptr(cur_overflow_list);
duke@435	1190	oop observed_overflow_list =
duke@435	1191	(oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
duke@435	1192	if (observed_overflow_list == cur_overflow_list) break;
duke@435	1193	// Otherwise...
duke@435	1194	cur_overflow_list = observed_overflow_list;
duke@435	1195	}
duke@435	1196	}
duke@435	1197
duke@435	1198	// Push objects on prefix list onto this thread's work queue
duke@435	1199	assert(cur != NULL, "program logic");
duke@435	1200	cur = prefix;
duke@435	1201	int n = 0;
duke@435	1202	while (cur != NULL) {
duke@435	1203	oop obj_to_push = cur->forwardee();
duke@435	1204	oop next = oop(cur->klass());
duke@435	1205	cur->set_klass(obj_to_push->klass());
duke@435	1206	if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
duke@435	1207	obj_to_push = cur;
duke@435	1208	assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
duke@435	1209	}
duke@435	1210	work_q->push(obj_to_push);
duke@435	1211	cur = next;
duke@435	1212	n++;
duke@435	1213	}
duke@435	1214	par_scan_state->note_overflow_refill(n);
duke@435	1215	return true;
duke@435	1216	}
duke@435	1217
duke@435	1218	void ParNewGeneration::ref_processor_init()
duke@435	1219	{
duke@435	1220	if (_ref_processor == NULL) {
duke@435	1221	// Allocate and initialize a reference processor
duke@435	1222	_ref_processor = ReferenceProcessor::create_ref_processor(
duke@435	1223	_reserved, // span
duke@435	1224	refs_discovery_is_atomic(), // atomic_discovery
duke@435	1225	refs_discovery_is_mt(), // mt_discovery
duke@435	1226	NULL, // is_alive_non_header
duke@435	1227	ParallelGCThreads,
duke@435	1228	ParallelRefProcEnabled);
duke@435	1229	}
duke@435	1230	}
duke@435	1231
duke@435	1232	const char* ParNewGeneration::name() const {
duke@435	1233	return "par new generation";
duke@435	1234	}