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

src/share/vm/gc_implementation/parNew/parCardTableModRefBS.cpp@5cfd8d19e546 (annotated)

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

Mon, 26 Jan 2009 12:47:21 -0800

author

ysr

date

Mon, 26 Jan 2009 12:47:21 -0800

changeset 969

5cfd8d19e546

parent 435

a61af66fc99e

child 1280

df6caf649ff7

permissions

-rw-r--r--

6786503: Overflow list performance can be improved
Summary: Avoid overflow list walk in CMS & ParNew when it is unnecessary. Fix a couple of correctness issues, including a C-heap leak, in ParNew at the intersection of promotion failure, work queue overflow and object array chunking. Add stress testing option and related assertion checking.
Reviewed-by: jmasa

duke@435	1	/*
duke@435	2	* Copyright (c) 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/_parCardTableModRefBS.cpp.incl"
duke@435	27
duke@435	28	void CardTableModRefBS::par_non_clean_card_iterate_work(Space* sp, MemRegion mr,
duke@435	29	DirtyCardToOopClosure* dcto_cl,
duke@435	30	MemRegionClosure* cl,
duke@435	31	bool clear,
duke@435	32	int n_threads) {
duke@435	33	if (n_threads > 0) {
duke@435	34	assert(n_threads == (int)ParallelGCThreads, "# worker threads != # requested!");
duke@435	35
duke@435	36	// Make sure the LNC array is valid for the space.
duke@435	37	jbyte** lowest_non_clean;
duke@435	38	uintptr_t lowest_non_clean_base_chunk_index;
duke@435	39	size_t lowest_non_clean_chunk_size;
duke@435	40	get_LNC_array_for_space(sp, lowest_non_clean,
duke@435	41	lowest_non_clean_base_chunk_index,
duke@435	42	lowest_non_clean_chunk_size);
duke@435	43
duke@435	44	int n_strides = n_threads * StridesPerThread;
duke@435	45	SequentialSubTasksDone* pst = sp->par_seq_tasks();
duke@435	46	pst->set_par_threads(n_threads);
duke@435	47	pst->set_n_tasks(n_strides);
duke@435	48
duke@435	49	int stride = 0;
duke@435	50	while (!pst->is_task_claimed(/* reference */ stride)) {
duke@435	51	process_stride(sp, mr, stride, n_strides, dcto_cl, cl, clear,
duke@435	52	lowest_non_clean,
duke@435	53	lowest_non_clean_base_chunk_index,
duke@435	54	lowest_non_clean_chunk_size);
duke@435	55	}
duke@435	56	if (pst->all_tasks_completed()) {
duke@435	57	// Clear lowest_non_clean array for next time.
duke@435	58	intptr_t first_chunk_index = addr_to_chunk_index(mr.start());
duke@435	59	uintptr_t last_chunk_index = addr_to_chunk_index(mr.last());
duke@435	60	for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) {
duke@435	61	intptr_t ind = ch - lowest_non_clean_base_chunk_index;
duke@435	62	assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size,
duke@435	63	"Bounds error");
duke@435	64	lowest_non_clean[ind] = NULL;
duke@435	65	}
duke@435	66	}
duke@435	67	}
duke@435	68	}
duke@435	69
duke@435	70	void
duke@435	71	CardTableModRefBS::
duke@435	72	process_stride(Space* sp,
duke@435	73	MemRegion used,
duke@435	74	jint stride, int n_strides,
duke@435	75	DirtyCardToOopClosure* dcto_cl,
duke@435	76	MemRegionClosure* cl,
duke@435	77	bool clear,
duke@435	78	jbyte** lowest_non_clean,
duke@435	79	uintptr_t lowest_non_clean_base_chunk_index,
duke@435	80	size_t lowest_non_clean_chunk_size) {
duke@435	81	// We don't have to go downwards here; it wouldn't help anyway,
duke@435	82	// because of parallelism.
duke@435	83
duke@435	84	// Find the first card address of the first chunk in the stride that is
duke@435	85	// at least "bottom" of the used region.
duke@435	86	jbyte* start_card = byte_for(used.start());
duke@435	87	jbyte* end_card = byte_after(used.last());
duke@435	88	uintptr_t start_chunk = addr_to_chunk_index(used.start());
duke@435	89	uintptr_t start_chunk_stride_num = start_chunk % n_strides;
duke@435	90	jbyte* chunk_card_start;
duke@435	91
duke@435	92	if ((uintptr_t)stride >= start_chunk_stride_num) {
duke@435	93	chunk_card_start = (jbyte*)(start_card +
duke@435	94	(stride - start_chunk_stride_num) *
duke@435	95	CardsPerStrideChunk);
duke@435	96	} else {
duke@435	97	// Go ahead to the next chunk group boundary, then to the requested stride.
duke@435	98	chunk_card_start = (jbyte*)(start_card +
duke@435	99	(n_strides - start_chunk_stride_num + stride) *
duke@435	100	CardsPerStrideChunk);
duke@435	101	}
duke@435	102
duke@435	103	while (chunk_card_start < end_card) {
duke@435	104	// We don't have to go downwards here; it wouldn't help anyway,
duke@435	105	// because of parallelism. (We take care with "min_done"; see below.)
duke@435	106	// Invariant: chunk_mr should be fully contained within the "used" region.
duke@435	107	jbyte* chunk_card_end = chunk_card_start + CardsPerStrideChunk;
duke@435	108	MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start),
duke@435	109	chunk_card_end >= end_card ?
duke@435	110	used.end() : addr_for(chunk_card_end));
duke@435	111	assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");
duke@435	112	assert(used.contains(chunk_mr), "chunk_mr should be subset of used");
duke@435	113
duke@435	114	// Process the chunk.
duke@435	115	process_chunk_boundaries(sp,
duke@435	116	dcto_cl,
duke@435	117	chunk_mr,
duke@435	118	used,
duke@435	119	lowest_non_clean,
duke@435	120	lowest_non_clean_base_chunk_index,
duke@435	121	lowest_non_clean_chunk_size);
duke@435	122
duke@435	123	non_clean_card_iterate_work(chunk_mr, cl, clear);
duke@435	124
duke@435	125	// Find the next chunk of the stride.
duke@435	126	chunk_card_start += CardsPerStrideChunk * n_strides;
duke@435	127	}
duke@435	128	}
duke@435	129
duke@435	130	void
duke@435	131	CardTableModRefBS::
duke@435	132	process_chunk_boundaries(Space* sp,
duke@435	133	DirtyCardToOopClosure* dcto_cl,
duke@435	134	MemRegion chunk_mr,
duke@435	135	MemRegion used,
duke@435	136	jbyte** lowest_non_clean,
duke@435	137	uintptr_t lowest_non_clean_base_chunk_index,
duke@435	138	size_t lowest_non_clean_chunk_size)
duke@435	139	{
duke@435	140	// We must worry about the chunk boundaries.
duke@435	141
duke@435	142	// First, set our max_to_do:
duke@435	143	HeapWord* max_to_do = NULL;
duke@435	144	uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start());
duke@435	145	cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index;
duke@435	146
duke@435	147	if (chunk_mr.end() < used.end()) {
duke@435	148	// This is not the last chunk in the used region. What is the last
duke@435	149	// object?
duke@435	150	HeapWord* last_block = sp->block_start(chunk_mr.end());
duke@435	151	assert(last_block <= chunk_mr.end(), "In case this property changes.");
duke@435	152	if (last_block == chunk_mr.end()
duke@435	153	|| !sp->block_is_obj(last_block)) {
duke@435	154	max_to_do = chunk_mr.end();
duke@435	155
duke@435	156	} else {
duke@435	157	// It is an object and starts before the end of the current chunk.
duke@435	158	// last_obj_card is the card corresponding to the start of the last object
duke@435	159	// in the chunk. Note that the last object may not start in
duke@435	160	// the chunk.
duke@435	161	jbyte* last_obj_card = byte_for(last_block);
duke@435	162	if (!card_may_have_been_dirty(*last_obj_card)) {
duke@435	163	// The card containing the head is not dirty. Any marks in
duke@435	164	// subsequent cards still in this chunk must have been made
duke@435	165	// precisely; we can cap processing at the end.
duke@435	166	max_to_do = chunk_mr.end();
duke@435	167	} else {
duke@435	168	// The last object must be considered dirty, and extends onto the
duke@435	169	// following chunk. Look for a dirty card in that chunk that will
duke@435	170	// bound our processing.
duke@435	171	jbyte* limit_card = NULL;
duke@435	172	size_t last_block_size = sp->block_size(last_block);
duke@435	173	jbyte* last_card_of_last_obj =
duke@435	174	byte_for(last_block + last_block_size - 1);
duke@435	175	jbyte* first_card_of_next_chunk = byte_for(chunk_mr.end());
duke@435	176	// This search potentially goes a long distance looking
duke@435	177	// for the next card that will be scanned. For example,
duke@435	178	// an object that is an array of primitives will not
duke@435	179	// have any cards covering regions interior to the array
duke@435	180	// that will need to be scanned. The scan can be terminated
duke@435	181	// at the last card of the next chunk. That would leave
duke@435	182	// limit_card as NULL and would result in "max_to_do"
duke@435	183	// being set with the LNC value or with the end
duke@435	184	// of the last block.
duke@435	185	jbyte* last_card_of_next_chunk = first_card_of_next_chunk +
duke@435	186	CardsPerStrideChunk;
duke@435	187	assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start())
duke@435	188	== CardsPerStrideChunk, "last card of next chunk may be wrong");
duke@435	189	jbyte* last_card_to_check = (jbyte*) MIN2(last_card_of_last_obj,
duke@435	190	last_card_of_next_chunk);
duke@435	191	for (jbyte* cur = first_card_of_next_chunk;
duke@435	192	cur <= last_card_to_check; cur++) {
duke@435	193	if (card_will_be_scanned(*cur)) {
duke@435	194	limit_card = cur; break;
duke@435	195	}
duke@435	196	}
duke@435	197	assert(0 <= cur_chunk_index+1 &&
duke@435	198	cur_chunk_index+1 < lowest_non_clean_chunk_size,
duke@435	199	"Bounds error.");
duke@435	200	// LNC for the next chunk
duke@435	201	jbyte* lnc_card = lowest_non_clean[cur_chunk_index+1];
duke@435	202	if (limit_card == NULL) {
duke@435	203	limit_card = lnc_card;
duke@435	204	}
duke@435	205	if (limit_card != NULL) {
duke@435	206	if (lnc_card != NULL) {
duke@435	207	limit_card = (jbyte*)MIN2((intptr_t)limit_card,
duke@435	208	(intptr_t)lnc_card);
duke@435	209	}
duke@435	210	max_to_do = addr_for(limit_card);
duke@435	211	} else {
duke@435	212	max_to_do = last_block + last_block_size;
duke@435	213	}
duke@435	214	}
duke@435	215	}
duke@435	216	assert(max_to_do != NULL, "OOPS!");
duke@435	217	} else {
duke@435	218	max_to_do = used.end();
duke@435	219	}
duke@435	220	// Now we can set the closure we're using so it doesn't to beyond
duke@435	221	// max_to_do.
duke@435	222	dcto_cl->set_min_done(max_to_do);
duke@435	223	#ifndef PRODUCT
duke@435	224	dcto_cl->set_last_bottom(max_to_do);
duke@435	225	#endif
duke@435	226
duke@435	227	// Now we set *our" lowest_non_clean entry.
duke@435	228	// Find the object that spans our boundary, if one exists.
duke@435	229	// Nothing to do on the first chunk.
duke@435	230	if (chunk_mr.start() > used.start()) {
duke@435	231	// first_block is the block possibly spanning the chunk start
duke@435	232	HeapWord* first_block = sp->block_start(chunk_mr.start());
duke@435	233	// Does the block span the start of the chunk and is it
duke@435	234	// an object?
duke@435	235	if (first_block < chunk_mr.start() &&
duke@435	236	sp->block_is_obj(first_block)) {
duke@435	237	jbyte* first_dirty_card = NULL;
duke@435	238	jbyte* last_card_of_first_obj =
duke@435	239	byte_for(first_block + sp->block_size(first_block) - 1);
duke@435	240	jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
duke@435	241	jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());
duke@435	242	jbyte* last_card_to_check =
duke@435	243	(jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,
duke@435	244	(intptr_t) last_card_of_first_obj);
duke@435	245	for (jbyte* cur = first_card_of_cur_chunk;
duke@435	246	cur <= last_card_to_check; cur++) {
duke@435	247	if (card_will_be_scanned(*cur)) {
duke@435	248	first_dirty_card = cur; break;
duke@435	249	}
duke@435	250	}
duke@435	251	if (first_dirty_card != NULL) {
duke@435	252	assert(0 <= cur_chunk_index &&
duke@435	253	cur_chunk_index < lowest_non_clean_chunk_size,
duke@435	254	"Bounds error.");
duke@435	255	lowest_non_clean[cur_chunk_index] = first_dirty_card;
duke@435	256	}
duke@435	257	}
duke@435	258	}
duke@435	259	}
duke@435	260
duke@435	261	void
duke@435	262	CardTableModRefBS::
duke@435	263	get_LNC_array_for_space(Space* sp,
duke@435	264	jbyte**& lowest_non_clean,
duke@435	265	uintptr_t& lowest_non_clean_base_chunk_index,
duke@435	266	size_t& lowest_non_clean_chunk_size) {
duke@435	267
duke@435	268	int i = find_covering_region_containing(sp->bottom());
duke@435	269	MemRegion covered = _covered[i];
duke@435	270	size_t n_chunks = chunks_to_cover(covered);
duke@435	271
duke@435	272	// Only the first thread to obtain the lock will resize the
duke@435	273	// LNC array for the covered region. Any later expansion can't affect
duke@435	274	// the used_at_save_marks region.
duke@435	275	// (I observed a bug in which the first thread to execute this would
duke@435	276	// resize, and then it would cause "expand_and_allocates" that would
duke@435	277	// Increase the number of chunks in the covered region. Then a second
duke@435	278	// thread would come and execute this, see that the size didn't match,
duke@435	279	// and free and allocate again. So the first thread would be using a
duke@435	280	// freed "_lowest_non_clean" array.)
duke@435	281
duke@435	282	// Do a dirty read here. If we pass the conditional then take the rare
duke@435	283	// event lock and do the read again in case some other thread had already
duke@435	284	// succeeded and done the resize.
duke@435	285	int cur_collection = Universe::heap()->total_collections();
duke@435	286	if (_last_LNC_resizing_collection[i] != cur_collection) {
duke@435	287	MutexLocker x(ParGCRareEvent_lock);
duke@435	288	if (_last_LNC_resizing_collection[i] != cur_collection) {
duke@435	289	if (_lowest_non_clean[i] == NULL ||
duke@435	290	n_chunks != _lowest_non_clean_chunk_size[i]) {
duke@435	291
duke@435	292	// Should we delete the old?
duke@435	293	if (_lowest_non_clean[i] != NULL) {
duke@435	294	assert(n_chunks != _lowest_non_clean_chunk_size[i],
duke@435	295	"logical consequence");
duke@435	296	FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]);
duke@435	297	_lowest_non_clean[i] = NULL;
duke@435	298	}
duke@435	299	// Now allocate a new one if necessary.
duke@435	300	if (_lowest_non_clean[i] == NULL) {
duke@435	301	_lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks);
duke@435	302	_lowest_non_clean_chunk_size[i] = n_chunks;
duke@435	303	_lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());
duke@435	304	for (int j = 0; j < (int)n_chunks; j++)
duke@435	305	_lowest_non_clean[i][j] = NULL;
duke@435	306	}
duke@435	307	}
duke@435	308	_last_LNC_resizing_collection[i] = cur_collection;
duke@435	309	}
duke@435	310	}
duke@435	311	// In any case, now do the initialization.
duke@435	312	lowest_non_clean = _lowest_non_clean[i];
duke@435	313	lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];
duke@435	314	lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i];
duke@435	315	}