Wed, 20 Apr 2011 19:19:30 -0700
7037276: Unnecessary double traversal of dirty card windows
Summary: Short-circuited an unnecessary double traversal of dirty card windows when iterating younger refs. Also renamed some cardtable methods for more clarity.
Reviewed-by: jmasa, stefank, poonam
1 /*
2 * Copyright (c) 2007, 2011 Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "memory/allocation.inline.hpp"
27 #include "memory/cardTableModRefBS.hpp"
28 #include "memory/cardTableRS.hpp"
29 #include "memory/sharedHeap.hpp"
30 #include "memory/space.inline.hpp"
31 #include "memory/universe.hpp"
32 #include "runtime/java.hpp"
33 #include "runtime/mutexLocker.hpp"
34 #include "runtime/virtualspace.hpp"
36 void CardTableModRefBS::non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr,
37 DirtyCardToOopClosure* dcto_cl,
38 ClearNoncleanCardWrapper* cl,
39 int n_threads) {
40 assert(n_threads > 0, "Error: expected n_threads > 0");
41 assert((n_threads == 1 && ParallelGCThreads == 0) ||
42 n_threads <= (int)ParallelGCThreads,
43 "# worker threads != # requested!");
44 // Make sure the LNC array is valid for the space.
45 jbyte** lowest_non_clean;
46 uintptr_t lowest_non_clean_base_chunk_index;
47 size_t lowest_non_clean_chunk_size;
48 get_LNC_array_for_space(sp, lowest_non_clean,
49 lowest_non_clean_base_chunk_index,
50 lowest_non_clean_chunk_size);
52 int n_strides = n_threads * StridesPerThread;
53 SequentialSubTasksDone* pst = sp->par_seq_tasks();
54 pst->set_n_threads(n_threads);
55 pst->set_n_tasks(n_strides);
57 int stride = 0;
58 while (!pst->is_task_claimed(/* reference */ stride)) {
59 process_stride(sp, mr, stride, n_strides, dcto_cl, cl,
60 lowest_non_clean,
61 lowest_non_clean_base_chunk_index,
62 lowest_non_clean_chunk_size);
63 }
64 if (pst->all_tasks_completed()) {
65 // Clear lowest_non_clean array for next time.
66 intptr_t first_chunk_index = addr_to_chunk_index(mr.start());
67 uintptr_t last_chunk_index = addr_to_chunk_index(mr.last());
68 for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) {
69 intptr_t ind = ch - lowest_non_clean_base_chunk_index;
70 assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size,
71 "Bounds error");
72 lowest_non_clean[ind] = NULL;
73 }
74 }
75 }
77 void
78 CardTableModRefBS::
79 process_stride(Space* sp,
80 MemRegion used,
81 jint stride, int n_strides,
82 DirtyCardToOopClosure* dcto_cl,
83 ClearNoncleanCardWrapper* cl,
84 jbyte** lowest_non_clean,
85 uintptr_t lowest_non_clean_base_chunk_index,
86 size_t lowest_non_clean_chunk_size) {
87 // We don't have to go downwards here; it wouldn't help anyway,
88 // because of parallelism.
90 // Find the first card address of the first chunk in the stride that is
91 // at least "bottom" of the used region.
92 jbyte* start_card = byte_for(used.start());
93 jbyte* end_card = byte_after(used.last());
94 uintptr_t start_chunk = addr_to_chunk_index(used.start());
95 uintptr_t start_chunk_stride_num = start_chunk % n_strides;
96 jbyte* chunk_card_start;
98 if ((uintptr_t)stride >= start_chunk_stride_num) {
99 chunk_card_start = (jbyte*)(start_card +
100 (stride - start_chunk_stride_num) *
101 CardsPerStrideChunk);
102 } else {
103 // Go ahead to the next chunk group boundary, then to the requested stride.
104 chunk_card_start = (jbyte*)(start_card +
105 (n_strides - start_chunk_stride_num + stride) *
106 CardsPerStrideChunk);
107 }
109 while (chunk_card_start < end_card) {
110 // We don't have to go downwards here; it wouldn't help anyway,
111 // because of parallelism. (We take care with "min_done"; see below.)
112 // Invariant: chunk_mr should be fully contained within the "used" region.
113 jbyte* chunk_card_end = chunk_card_start + CardsPerStrideChunk;
114 MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start),
115 chunk_card_end >= end_card ?
116 used.end() : addr_for(chunk_card_end));
117 assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");
118 assert(used.contains(chunk_mr), "chunk_mr should be subset of used");
120 // Process the chunk.
121 process_chunk_boundaries(sp,
122 dcto_cl,
123 chunk_mr,
124 used,
125 lowest_non_clean,
126 lowest_non_clean_base_chunk_index,
127 lowest_non_clean_chunk_size);
129 // We do not call the non_clean_card_iterate_serial() version because
130 // we want to clear the cards, and the ClearNoncleanCardWrapper closure
131 // itself does the work of finding contiguous dirty ranges of cards to
132 // process (and clear).
133 cl->do_MemRegion(chunk_mr);
135 // Find the next chunk of the stride.
136 chunk_card_start += CardsPerStrideChunk * n_strides;
137 }
138 }
140 void
141 CardTableModRefBS::
142 process_chunk_boundaries(Space* sp,
143 DirtyCardToOopClosure* dcto_cl,
144 MemRegion chunk_mr,
145 MemRegion used,
146 jbyte** lowest_non_clean,
147 uintptr_t lowest_non_clean_base_chunk_index,
148 size_t lowest_non_clean_chunk_size)
149 {
150 // We must worry about the chunk boundaries.
152 // First, set our max_to_do:
153 HeapWord* max_to_do = NULL;
154 uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start());
155 cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index;
157 if (chunk_mr.end() < used.end()) {
158 // This is not the last chunk in the used region. What is the last
159 // object?
160 HeapWord* last_block = sp->block_start(chunk_mr.end());
161 assert(last_block <= chunk_mr.end(), "In case this property changes.");
162 if (last_block == chunk_mr.end()
163 || !sp->block_is_obj(last_block)) {
164 max_to_do = chunk_mr.end();
166 } else {
167 // It is an object and starts before the end of the current chunk.
168 // last_obj_card is the card corresponding to the start of the last object
169 // in the chunk. Note that the last object may not start in
170 // the chunk.
171 jbyte* last_obj_card = byte_for(last_block);
172 if (!card_may_have_been_dirty(*last_obj_card)) {
173 // The card containing the head is not dirty. Any marks in
174 // subsequent cards still in this chunk must have been made
175 // precisely; we can cap processing at the end.
176 max_to_do = chunk_mr.end();
177 } else {
178 // The last object must be considered dirty, and extends onto the
179 // following chunk. Look for a dirty card in that chunk that will
180 // bound our processing.
181 jbyte* limit_card = NULL;
182 size_t last_block_size = sp->block_size(last_block);
183 jbyte* last_card_of_last_obj =
184 byte_for(last_block + last_block_size - 1);
185 jbyte* first_card_of_next_chunk = byte_for(chunk_mr.end());
186 // This search potentially goes a long distance looking
187 // for the next card that will be scanned. For example,
188 // an object that is an array of primitives will not
189 // have any cards covering regions interior to the array
190 // that will need to be scanned. The scan can be terminated
191 // at the last card of the next chunk. That would leave
192 // limit_card as NULL and would result in "max_to_do"
193 // being set with the LNC value or with the end
194 // of the last block.
195 jbyte* last_card_of_next_chunk = first_card_of_next_chunk +
196 CardsPerStrideChunk;
197 assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start())
198 == CardsPerStrideChunk, "last card of next chunk may be wrong");
199 jbyte* last_card_to_check = (jbyte*) MIN2(last_card_of_last_obj,
200 last_card_of_next_chunk);
201 for (jbyte* cur = first_card_of_next_chunk;
202 cur <= last_card_to_check; cur++) {
203 if (card_will_be_scanned(*cur)) {
204 limit_card = cur; break;
205 }
206 }
207 assert(0 <= cur_chunk_index+1 &&
208 cur_chunk_index+1 < lowest_non_clean_chunk_size,
209 "Bounds error.");
210 // LNC for the next chunk
211 jbyte* lnc_card = lowest_non_clean[cur_chunk_index+1];
212 if (limit_card == NULL) {
213 limit_card = lnc_card;
214 }
215 if (limit_card != NULL) {
216 if (lnc_card != NULL) {
217 limit_card = (jbyte*)MIN2((intptr_t)limit_card,
218 (intptr_t)lnc_card);
219 }
220 max_to_do = addr_for(limit_card);
221 } else {
222 max_to_do = last_block + last_block_size;
223 }
224 }
225 }
226 assert(max_to_do != NULL, "OOPS!");
227 } else {
228 max_to_do = used.end();
229 }
230 // Now we can set the closure we're using so it doesn't to beyond
231 // max_to_do.
232 dcto_cl->set_min_done(max_to_do);
233 #ifndef PRODUCT
234 dcto_cl->set_last_bottom(max_to_do);
235 #endif
237 // Now we set *our" lowest_non_clean entry.
238 // Find the object that spans our boundary, if one exists.
239 // Nothing to do on the first chunk.
240 if (chunk_mr.start() > used.start()) {
241 // first_block is the block possibly spanning the chunk start
242 HeapWord* first_block = sp->block_start(chunk_mr.start());
243 // Does the block span the start of the chunk and is it
244 // an object?
245 if (first_block < chunk_mr.start() &&
246 sp->block_is_obj(first_block)) {
247 jbyte* first_dirty_card = NULL;
248 jbyte* last_card_of_first_obj =
249 byte_for(first_block + sp->block_size(first_block) - 1);
250 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
251 jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());
252 jbyte* last_card_to_check =
253 (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,
254 (intptr_t) last_card_of_first_obj);
255 for (jbyte* cur = first_card_of_cur_chunk;
256 cur <= last_card_to_check; cur++) {
257 if (card_will_be_scanned(*cur)) {
258 first_dirty_card = cur; break;
259 }
260 }
261 if (first_dirty_card != NULL) {
262 assert(0 <= cur_chunk_index &&
263 cur_chunk_index < lowest_non_clean_chunk_size,
264 "Bounds error.");
265 lowest_non_clean[cur_chunk_index] = first_dirty_card;
266 }
267 }
268 }
269 }
271 void
272 CardTableModRefBS::
273 get_LNC_array_for_space(Space* sp,
274 jbyte**& lowest_non_clean,
275 uintptr_t& lowest_non_clean_base_chunk_index,
276 size_t& lowest_non_clean_chunk_size) {
278 int i = find_covering_region_containing(sp->bottom());
279 MemRegion covered = _covered[i];
280 size_t n_chunks = chunks_to_cover(covered);
282 // Only the first thread to obtain the lock will resize the
283 // LNC array for the covered region. Any later expansion can't affect
284 // the used_at_save_marks region.
285 // (I observed a bug in which the first thread to execute this would
286 // resize, and then it would cause "expand_and_allocates" that would
287 // Increase the number of chunks in the covered region. Then a second
288 // thread would come and execute this, see that the size didn't match,
289 // and free and allocate again. So the first thread would be using a
290 // freed "_lowest_non_clean" array.)
292 // Do a dirty read here. If we pass the conditional then take the rare
293 // event lock and do the read again in case some other thread had already
294 // succeeded and done the resize.
295 int cur_collection = Universe::heap()->total_collections();
296 if (_last_LNC_resizing_collection[i] != cur_collection) {
297 MutexLocker x(ParGCRareEvent_lock);
298 if (_last_LNC_resizing_collection[i] != cur_collection) {
299 if (_lowest_non_clean[i] == NULL ||
300 n_chunks != _lowest_non_clean_chunk_size[i]) {
302 // Should we delete the old?
303 if (_lowest_non_clean[i] != NULL) {
304 assert(n_chunks != _lowest_non_clean_chunk_size[i],
305 "logical consequence");
306 FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]);
307 _lowest_non_clean[i] = NULL;
308 }
309 // Now allocate a new one if necessary.
310 if (_lowest_non_clean[i] == NULL) {
311 _lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks);
312 _lowest_non_clean_chunk_size[i] = n_chunks;
313 _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());
314 for (int j = 0; j < (int)n_chunks; j++)
315 _lowest_non_clean[i][j] = NULL;
316 }
317 }
318 _last_LNC_resizing_collection[i] = cur_collection;
319 }
320 }
321 // In any case, now do the initialization.
322 lowest_non_clean = _lowest_non_clean[i];
323 lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];
324 lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i];
325 }