Tue, 10 May 2011 00:33:21 -0700
6883834: ParNew: assert(!_g->to()->is_in_reserved(obj),"Scanning field twice?") with LargeObjects tests
Summary: Fixed process_chunk_boundaries(), used for parallel card scanning when using ParNew/CMS, so as to prevent double-scanning, or worse, non-scanning of imprecisely marked objects exceeding parallel chunk size. Made some sizing parameters for parallel card scanning diagnostic, disabled ParallelGCRetainPLAB, and elaborated and clarified some comments.
Reviewed-by: stefank, johnc
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 "oops/oop.inline.hpp"
33 #include "runtime/java.hpp"
34 #include "runtime/mutexLocker.hpp"
35 #include "runtime/virtualspace.hpp"
37 void CardTableModRefBS::non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr,
38 OopsInGenClosure* cl,
39 CardTableRS* ct,
40 int n_threads) {
41 assert(n_threads > 0, "Error: expected n_threads > 0");
42 assert((n_threads == 1 && ParallelGCThreads == 0) ||
43 n_threads <= (int)ParallelGCThreads,
44 "# worker threads != # requested!");
45 // Make sure the LNC array is valid for the space.
46 jbyte** lowest_non_clean;
47 uintptr_t lowest_non_clean_base_chunk_index;
48 size_t lowest_non_clean_chunk_size;
49 get_LNC_array_for_space(sp, lowest_non_clean,
50 lowest_non_clean_base_chunk_index,
51 lowest_non_clean_chunk_size);
53 int n_strides = n_threads * ParGCStridesPerThread;
54 SequentialSubTasksDone* pst = sp->par_seq_tasks();
55 pst->set_n_threads(n_threads);
56 pst->set_n_tasks(n_strides);
58 int stride = 0;
59 while (!pst->is_task_claimed(/* reference */ stride)) {
60 process_stride(sp, mr, stride, n_strides, cl, ct,
61 lowest_non_clean,
62 lowest_non_clean_base_chunk_index,
63 lowest_non_clean_chunk_size);
64 }
65 if (pst->all_tasks_completed()) {
66 // Clear lowest_non_clean array for next time.
67 intptr_t first_chunk_index = addr_to_chunk_index(mr.start());
68 uintptr_t last_chunk_index = addr_to_chunk_index(mr.last());
69 for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) {
70 intptr_t ind = ch - lowest_non_clean_base_chunk_index;
71 assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size,
72 "Bounds error");
73 lowest_non_clean[ind] = NULL;
74 }
75 }
76 }
78 void
79 CardTableModRefBS::
80 process_stride(Space* sp,
81 MemRegion used,
82 jint stride, int n_strides,
83 OopsInGenClosure* cl,
84 CardTableRS* ct,
85 jbyte** lowest_non_clean,
86 uintptr_t lowest_non_clean_base_chunk_index,
87 size_t lowest_non_clean_chunk_size) {
88 // We go from higher to lower addresses here; it wouldn't help that much
89 // because of the strided parallelism pattern used here.
91 // Find the first card address of the first chunk in the stride that is
92 // at least "bottom" of the used region.
93 jbyte* start_card = byte_for(used.start());
94 jbyte* end_card = byte_after(used.last());
95 uintptr_t start_chunk = addr_to_chunk_index(used.start());
96 uintptr_t start_chunk_stride_num = start_chunk % n_strides;
97 jbyte* chunk_card_start;
99 if ((uintptr_t)stride >= start_chunk_stride_num) {
100 chunk_card_start = (jbyte*)(start_card +
101 (stride - start_chunk_stride_num) *
102 ParGCCardsPerStrideChunk);
103 } else {
104 // Go ahead to the next chunk group boundary, then to the requested stride.
105 chunk_card_start = (jbyte*)(start_card +
106 (n_strides - start_chunk_stride_num + stride) *
107 ParGCCardsPerStrideChunk);
108 }
110 while (chunk_card_start < end_card) {
111 // Even though we go from lower to higher addresses below, the
112 // strided parallelism can interleave the actual processing of the
113 // dirty pages in various ways. For a specific chunk within this
114 // stride, we take care to avoid double scanning or missing a card
115 // by suitably initializing the "min_done" field in process_chunk_boundaries()
116 // below, together with the dirty region extension accomplished in
117 // DirtyCardToOopClosure::do_MemRegion().
118 jbyte* chunk_card_end = chunk_card_start + ParGCCardsPerStrideChunk;
119 // Invariant: chunk_mr should be fully contained within the "used" region.
120 MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start),
121 chunk_card_end >= end_card ?
122 used.end() : addr_for(chunk_card_end));
123 assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");
124 assert(used.contains(chunk_mr), "chunk_mr should be subset of used");
126 DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(),
127 cl->gen_boundary());
128 ClearNoncleanCardWrapper clear_cl(dcto_cl, ct);
131 // Process the chunk.
132 process_chunk_boundaries(sp,
133 dcto_cl,
134 chunk_mr,
135 used,
136 lowest_non_clean,
137 lowest_non_clean_base_chunk_index,
138 lowest_non_clean_chunk_size);
140 // We want the LNC array updates above in process_chunk_boundaries
141 // to be visible before any of the card table value changes as a
142 // result of the dirty card iteration below.
143 OrderAccess::storestore();
145 // We do not call the non_clean_card_iterate_serial() version because
146 // we want to clear the cards: clear_cl here does the work of finding
147 // contiguous dirty ranges of cards to process and clear.
148 clear_cl.do_MemRegion(chunk_mr);
150 // Find the next chunk of the stride.
151 chunk_card_start += ParGCCardsPerStrideChunk * n_strides;
152 }
153 }
156 // If you want a talkative process_chunk_boundaries,
157 // then #define NOISY(x) x
158 #ifdef NOISY
159 #error "Encountered a global preprocessor flag, NOISY, which might clash with local definition to follow"
160 #else
161 #define NOISY(x)
162 #endif
164 void
165 CardTableModRefBS::
166 process_chunk_boundaries(Space* sp,
167 DirtyCardToOopClosure* dcto_cl,
168 MemRegion chunk_mr,
169 MemRegion used,
170 jbyte** lowest_non_clean,
171 uintptr_t lowest_non_clean_base_chunk_index,
172 size_t lowest_non_clean_chunk_size)
173 {
174 // We must worry about non-array objects that cross chunk boundaries,
175 // because such objects are both precisely and imprecisely marked:
176 // .. if the head of such an object is dirty, the entire object
177 // needs to be scanned, under the interpretation that this
178 // was an imprecise mark
179 // .. if the head of such an object is not dirty, we can assume
180 // precise marking and it's efficient to scan just the dirty
181 // cards.
182 // In either case, each scanned reference must be scanned precisely
183 // once so as to avoid cloning of a young referent. For efficiency,
184 // our closures depend on this property and do not protect against
185 // double scans.
187 uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start());
188 cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index;
190 NOISY(tty->print_cr("===========================================================================");)
191 NOISY(tty->print_cr(" process_chunk_boundary: Called with [" PTR_FORMAT "," PTR_FORMAT ")",
192 chunk_mr.start(), chunk_mr.end());)
194 // First, set "our" lowest_non_clean entry, which would be
195 // used by the thread scanning an adjoining left chunk with
196 // a non-array object straddling the mutual boundary.
197 // Find the object that spans our boundary, if one exists.
198 // first_block is the block possibly straddling our left boundary.
199 HeapWord* first_block = sp->block_start(chunk_mr.start());
200 assert((chunk_mr.start() != used.start()) || (first_block == chunk_mr.start()),
201 "First chunk should always have a co-initial block");
202 // Does the block straddle the chunk's left boundary, and is it
203 // a non-array object?
204 if (first_block < chunk_mr.start() // first block straddles left bdry
205 && sp->block_is_obj(first_block) // first block is an object
206 && !(oop(first_block)->is_objArray() // first block is not an array (arrays are precisely dirtied)
207 || oop(first_block)->is_typeArray())) {
208 // Find our least non-clean card, so that a left neighbour
209 // does not scan an object straddling the mutual boundary
210 // too far to the right, and attempt to scan a portion of
211 // that object twice.
212 jbyte* first_dirty_card = NULL;
213 jbyte* last_card_of_first_obj =
214 byte_for(first_block + sp->block_size(first_block) - 1);
215 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
216 jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());
217 jbyte* last_card_to_check =
218 (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,
219 (intptr_t) last_card_of_first_obj);
220 // Note that this does not need to go beyond our last card
221 // if our first object completely straddles this chunk.
222 for (jbyte* cur = first_card_of_cur_chunk;
223 cur <= last_card_to_check; cur++) {
224 jbyte val = *cur;
225 if (card_will_be_scanned(val)) {
226 first_dirty_card = cur; break;
227 } else {
228 assert(!card_may_have_been_dirty(val), "Error");
229 }
230 }
231 if (first_dirty_card != NULL) {
232 NOISY(tty->print_cr(" LNC: Found a dirty card at " PTR_FORMAT " in current chunk",
233 first_dirty_card);)
234 assert(0 <= cur_chunk_index && cur_chunk_index < lowest_non_clean_chunk_size,
235 "Bounds error.");
236 assert(lowest_non_clean[cur_chunk_index] == NULL,
237 "Write exactly once : value should be stable hereafter for this round");
238 lowest_non_clean[cur_chunk_index] = first_dirty_card;
239 } NOISY(else {
240 tty->print_cr(" LNC: Found no dirty card in current chunk; leaving LNC entry NULL");
241 // In the future, we could have this thread look for a non-NULL value to copy from its
242 // right neighbour (up to the end of the first object).
243 if (last_card_of_cur_chunk < last_card_of_first_obj) {
244 tty->print_cr(" LNC: BEWARE!!! first obj straddles past right end of chunk:\n"
245 " might be efficient to get value from right neighbour?");
246 }
247 })
248 } else {
249 // In this case we can help our neighbour by just asking them
250 // to stop at our first card (even though it may not be dirty).
251 NOISY(tty->print_cr(" LNC: first block is not a non-array object; setting LNC to first card of current chunk");)
252 assert(lowest_non_clean[cur_chunk_index] == NULL, "Write once : value should be stable hereafter");
253 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
254 lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk;
255 }
256 NOISY(tty->print_cr(" process_chunk_boundary: lowest_non_clean[" INTPTR_FORMAT "] = " PTR_FORMAT
257 " which corresponds to the heap address " PTR_FORMAT,
258 cur_chunk_index, lowest_non_clean[cur_chunk_index],
259 (lowest_non_clean[cur_chunk_index] != NULL)
260 ? addr_for(lowest_non_clean[cur_chunk_index])
261 : NULL);)
262 NOISY(tty->print_cr("---------------------------------------------------------------------------");)
264 // Next, set our own max_to_do, which will strictly/exclusively bound
265 // the highest address that we will scan past the right end of our chunk.
266 HeapWord* max_to_do = NULL;
267 if (chunk_mr.end() < used.end()) {
268 // This is not the last chunk in the used region.
269 // What is our last block? We check the first block of
270 // the next (right) chunk rather than strictly check our last block
271 // because it's potentially more efficient to do so.
272 HeapWord* const last_block = sp->block_start(chunk_mr.end());
273 assert(last_block <= chunk_mr.end(), "In case this property changes.");
274 if ((last_block == chunk_mr.end()) // our last block does not straddle boundary
275 || !sp->block_is_obj(last_block) // last_block isn't an object
276 || oop(last_block)->is_objArray() // last_block is an array (precisely marked)
277 || oop(last_block)->is_typeArray()) {
278 max_to_do = chunk_mr.end();
279 NOISY(tty->print_cr(" process_chunk_boundary: Last block on this card is not a non-array object;\n"
280 " max_to_do left at " PTR_FORMAT, max_to_do);)
281 } else {
282 assert(last_block < chunk_mr.end(), "Tautology");
283 // It is a non-array object that straddles the right boundary of this chunk.
284 // last_obj_card is the card corresponding to the start of the last object
285 // in the chunk. Note that the last object may not start in
286 // the chunk.
287 jbyte* const last_obj_card = byte_for(last_block);
288 const jbyte val = *last_obj_card;
289 if (!card_will_be_scanned(val)) {
290 assert(!card_may_have_been_dirty(val), "Error");
291 // The card containing the head is not dirty. Any marks on
292 // subsequent cards still in this chunk must have been made
293 // precisely; we can cap processing at the end of our chunk.
294 max_to_do = chunk_mr.end();
295 NOISY(tty->print_cr(" process_chunk_boundary: Head of last object on this card is not dirty;\n"
296 " max_to_do left at " PTR_FORMAT,
297 max_to_do);)
298 } else {
299 // The last object must be considered dirty, and extends onto the
300 // following chunk. Look for a dirty card in that chunk that will
301 // bound our processing.
302 jbyte* limit_card = NULL;
303 const size_t last_block_size = sp->block_size(last_block);
304 jbyte* const last_card_of_last_obj =
305 byte_for(last_block + last_block_size - 1);
306 jbyte* const first_card_of_next_chunk = byte_for(chunk_mr.end());
307 // This search potentially goes a long distance looking
308 // for the next card that will be scanned, terminating
309 // at the end of the last_block, if no earlier dirty card
310 // is found.
311 assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == ParGCCardsPerStrideChunk,
312 "last card of next chunk may be wrong");
313 for (jbyte* cur = first_card_of_next_chunk;
314 cur <= last_card_of_last_obj; cur++) {
315 const jbyte val = *cur;
316 if (card_will_be_scanned(val)) {
317 NOISY(tty->print_cr(" Found a non-clean card " PTR_FORMAT " with value 0x%x",
318 cur, (int)val);)
319 limit_card = cur; break;
320 } else {
321 assert(!card_may_have_been_dirty(val), "Error: card can't be skipped");
322 }
323 }
324 if (limit_card != NULL) {
325 max_to_do = addr_for(limit_card);
326 assert(limit_card != NULL && max_to_do != NULL, "Error");
327 NOISY(tty->print_cr(" process_chunk_boundary: Found a dirty card at " PTR_FORMAT
328 " max_to_do set at " PTR_FORMAT " which is before end of last block in chunk: "
329 PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
330 limit_card, max_to_do, last_block, last_block_size, (last_block+last_block_size));)
331 } else {
332 // The following is a pessimistic value, because it's possible
333 // that a dirty card on a subsequent chunk has been cleared by
334 // the time we get to look at it; we'll correct for that further below,
335 // using the LNC array which records the least non-clean card
336 // before cards were cleared in a particular chunk.
337 limit_card = last_card_of_last_obj;
338 max_to_do = last_block + last_block_size;
339 assert(limit_card != NULL && max_to_do != NULL, "Error");
340 NOISY(tty->print_cr(" process_chunk_boundary: Found no dirty card before end of last block in chunk\n"
341 " Setting limit_card to " PTR_FORMAT
342 " and max_to_do " PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
343 limit_card, last_block, last_block_size, max_to_do);)
344 }
345 assert(0 < cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size,
346 "Bounds error.");
347 // It is possible that a dirty card for the last object may have been
348 // cleared before we had a chance to examine it. In that case, the value
349 // will have been logged in the LNC for that chunk.
350 // We need to examine as many chunks to the right as this object
351 // covers.
352 const uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1)
353 - lowest_non_clean_base_chunk_index;
354 DEBUG_ONLY(const uintptr_t last_chunk_index = addr_to_chunk_index(used.end())
355 - lowest_non_clean_base_chunk_index;)
356 assert(last_chunk_index_to_check <= last_chunk_index,
357 err_msg("Out of bounds: last_chunk_index_to_check " INTPTR_FORMAT
358 " exceeds last_chunk_index " INTPTR_FORMAT,
359 last_chunk_index_to_check, last_chunk_index));
360 for (uintptr_t lnc_index = cur_chunk_index + 1;
361 lnc_index <= last_chunk_index_to_check;
362 lnc_index++) {
363 jbyte* lnc_card = lowest_non_clean[lnc_index];
364 if (lnc_card != NULL) {
365 // we can stop at the first non-NULL entry we find
366 if (lnc_card <= limit_card) {
367 NOISY(tty->print_cr(" process_chunk_boundary: LNC card " PTR_FORMAT " is lower than limit_card " PTR_FORMAT,
368 " max_to_do will be lowered to " PTR_FORMAT " from " PTR_FORMAT,
369 lnc_card, limit_card, addr_for(lnc_card), max_to_do);)
370 limit_card = lnc_card;
371 max_to_do = addr_for(limit_card);
372 assert(limit_card != NULL && max_to_do != NULL, "Error");
373 }
374 // In any case, we break now
375 break;
376 } // else continue to look for a non-NULL entry if any
377 }
378 assert(limit_card != NULL && max_to_do != NULL, "Error");
379 }
380 assert(max_to_do != NULL, "OOPS 1 !");
381 }
382 assert(max_to_do != NULL, "OOPS 2!");
383 } else {
384 max_to_do = used.end();
385 NOISY(tty->print_cr(" process_chunk_boundary: Last chunk of this space;\n"
386 " max_to_do left at " PTR_FORMAT,
387 max_to_do);)
388 }
389 assert(max_to_do != NULL, "OOPS 3!");
390 // Now we can set the closure we're using so it doesn't to beyond
391 // max_to_do.
392 dcto_cl->set_min_done(max_to_do);
393 #ifndef PRODUCT
394 dcto_cl->set_last_bottom(max_to_do);
395 #endif
396 NOISY(tty->print_cr("===========================================================================\n");)
397 }
399 #undef NOISY
401 void
402 CardTableModRefBS::
403 get_LNC_array_for_space(Space* sp,
404 jbyte**& lowest_non_clean,
405 uintptr_t& lowest_non_clean_base_chunk_index,
406 size_t& lowest_non_clean_chunk_size) {
408 int i = find_covering_region_containing(sp->bottom());
409 MemRegion covered = _covered[i];
410 size_t n_chunks = chunks_to_cover(covered);
412 // Only the first thread to obtain the lock will resize the
413 // LNC array for the covered region. Any later expansion can't affect
414 // the used_at_save_marks region.
415 // (I observed a bug in which the first thread to execute this would
416 // resize, and then it would cause "expand_and_allocate" that would
417 // increase the number of chunks in the covered region. Then a second
418 // thread would come and execute this, see that the size didn't match,
419 // and free and allocate again. So the first thread would be using a
420 // freed "_lowest_non_clean" array.)
422 // Do a dirty read here. If we pass the conditional then take the rare
423 // event lock and do the read again in case some other thread had already
424 // succeeded and done the resize.
425 int cur_collection = Universe::heap()->total_collections();
426 if (_last_LNC_resizing_collection[i] != cur_collection) {
427 MutexLocker x(ParGCRareEvent_lock);
428 if (_last_LNC_resizing_collection[i] != cur_collection) {
429 if (_lowest_non_clean[i] == NULL ||
430 n_chunks != _lowest_non_clean_chunk_size[i]) {
432 // Should we delete the old?
433 if (_lowest_non_clean[i] != NULL) {
434 assert(n_chunks != _lowest_non_clean_chunk_size[i],
435 "logical consequence");
436 FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]);
437 _lowest_non_clean[i] = NULL;
438 }
439 // Now allocate a new one if necessary.
440 if (_lowest_non_clean[i] == NULL) {
441 _lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks);
442 _lowest_non_clean_chunk_size[i] = n_chunks;
443 _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());
444 for (int j = 0; j < (int)n_chunks; j++)
445 _lowest_non_clean[i][j] = NULL;
446 }
447 }
448 _last_LNC_resizing_collection[i] = cur_collection;
449 }
450 }
451 // In any case, now do the initialization.
452 lowest_non_clean = _lowest_non_clean[i];
453 lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];
454 lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i];
455 }