1 /*
2 * Copyright (c) 2007, 2012, 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 */
24
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"
36 #include "runtime/vmThread.hpp"
37
38 void CardTableModRefBS::non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr,
39 OopsInGenClosure* cl,
40 CardTableRS* ct,
41 int n_threads) {
42 assert(n_threads > 0, "Error: expected n_threads > 0");
43 assert((n_threads == 1 && ParallelGCThreads == 0) ||
44 n_threads <= (int)ParallelGCThreads,
45 "# worker threads != # requested!");
46 assert(!Thread::current()->is_VM_thread() || (n_threads == 1), "There is only 1 VM thread");
47 assert(UseDynamicNumberOfGCThreads ||
48 !FLAG_IS_DEFAULT(ParallelGCThreads) ||
49 n_threads == (int)ParallelGCThreads,
50 "# worker threads != # requested!");
51 // Make sure the LNC array is valid for the space.
52 jbyte** lowest_non_clean;
53 uintptr_t lowest_non_clean_base_chunk_index;
54 size_t lowest_non_clean_chunk_size;
55 get_LNC_array_for_space(sp, lowest_non_clean,
56 lowest_non_clean_base_chunk_index,
57 lowest_non_clean_chunk_size);
58
59 uint n_strides = n_threads * ParGCStridesPerThread;
60 SequentialSubTasksDone* pst = sp->par_seq_tasks();
61 // Sets the condition for completion of the subtask (how many threads
62 // need to finish in order to be done).
63 pst->set_n_threads(n_threads);
64 pst->set_n_tasks(n_strides);
65
66 uint stride = 0;
67 while (!pst->is_task_claimed(/* reference */ stride)) {
68 process_stride(sp, mr, stride, n_strides, cl, ct,
69 lowest_non_clean,
70 lowest_non_clean_base_chunk_index,
71 lowest_non_clean_chunk_size);
72 }
73 if (pst->all_tasks_completed()) {
74 // Clear lowest_non_clean array for next time.
75 intptr_t first_chunk_index = addr_to_chunk_index(mr.start());
76 uintptr_t last_chunk_index = addr_to_chunk_index(mr.last());
77 for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) {
78 intptr_t ind = ch - lowest_non_clean_base_chunk_index;
79 assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size,
80 "Bounds error");
81 lowest_non_clean[ind] = NULL;
82 }
83 }
84 }
85
86 void
87 CardTableModRefBS::
88 process_stride(Space* sp,
89 MemRegion used,
90 jint stride, int n_strides,
91 OopsInGenClosure* cl,
92 CardTableRS* ct,
93 jbyte** lowest_non_clean,
94 uintptr_t lowest_non_clean_base_chunk_index,
95 size_t lowest_non_clean_chunk_size) {
96 // We go from higher to lower addresses here; it wouldn't help that much
97 // because of the strided parallelism pattern used here.
98
99 // Find the first card address of the first chunk in the stride that is
100 // at least "bottom" of the used region.
101 jbyte* start_card = byte_for(used.start());
102 jbyte* end_card = byte_after(used.last());
103 uintptr_t start_chunk = addr_to_chunk_index(used.start());
104 uintptr_t start_chunk_stride_num = start_chunk % n_strides;
105 jbyte* chunk_card_start;
106
107 if ((uintptr_t)stride >= start_chunk_stride_num) {
108 chunk_card_start = (jbyte*)(start_card +
109 (stride - start_chunk_stride_num) *
110 ParGCCardsPerStrideChunk);
111 } else {
112 // Go ahead to the next chunk group boundary, then to the requested stride.
113 chunk_card_start = (jbyte*)(start_card +
114 (n_strides - start_chunk_stride_num + stride) *
115 ParGCCardsPerStrideChunk);
116 }
117
118 while (chunk_card_start < end_card) {
119 // Even though we go from lower to higher addresses below, the
120 // strided parallelism can interleave the actual processing of the
121 // dirty pages in various ways. For a specific chunk within this
122 // stride, we take care to avoid double scanning or missing a card
123 // by suitably initializing the "min_done" field in process_chunk_boundaries()
124 // below, together with the dirty region extension accomplished in
125 // DirtyCardToOopClosure::do_MemRegion().
126 jbyte* chunk_card_end = chunk_card_start + ParGCCardsPerStrideChunk;
127 // Invariant: chunk_mr should be fully contained within the "used" region.
128 MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start),
129 chunk_card_end >= end_card ?
130 used.end() : addr_for(chunk_card_end));
131 assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");
132 assert(used.contains(chunk_mr), "chunk_mr should be subset of used");
133
134 DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(),
135 cl->gen_boundary());
136 ClearNoncleanCardWrapper clear_cl(dcto_cl, ct);
137
138
139 // Process the chunk.
140 process_chunk_boundaries(sp,
141 dcto_cl,
142 chunk_mr,
143 used,
144 lowest_non_clean,
145 lowest_non_clean_base_chunk_index,
146 lowest_non_clean_chunk_size);
147
148 // We want the LNC array updates above in process_chunk_boundaries
149 // to be visible before any of the card table value changes as a
150 // result of the dirty card iteration below.
151 OrderAccess::storestore();
152
153 // We do not call the non_clean_card_iterate_serial() version because
154 // we want to clear the cards: clear_cl here does the work of finding
155 // contiguous dirty ranges of cards to process and clear.
156 clear_cl.do_MemRegion(chunk_mr);
157
158 // Find the next chunk of the stride.
159 chunk_card_start += ParGCCardsPerStrideChunk * n_strides;
160 }
161 }
162
163
164 // If you want a talkative process_chunk_boundaries,
165 // then #define NOISY(x) x
166 #ifdef NOISY
167 #error "Encountered a global preprocessor flag, NOISY, which might clash with local definition to follow"
168 #else
169 #define NOISY(x)
170 #endif
171
172 void
173 CardTableModRefBS::
174 process_chunk_boundaries(Space* sp,
175 DirtyCardToOopClosure* dcto_cl,
176 MemRegion chunk_mr,
177 MemRegion used,
178 jbyte** lowest_non_clean,
179 uintptr_t lowest_non_clean_base_chunk_index,
180 size_t lowest_non_clean_chunk_size)
181 {
182 // We must worry about non-array objects that cross chunk boundaries,
183 // because such objects are both precisely and imprecisely marked:
184 // .. if the head of such an object is dirty, the entire object
185 // needs to be scanned, under the interpretation that this
186 // was an imprecise mark
187 // .. if the head of such an object is not dirty, we can assume
188 // precise marking and it's efficient to scan just the dirty
189 // cards.
190 // In either case, each scanned reference must be scanned precisely
191 // once so as to avoid cloning of a young referent. For efficiency,
192 // our closures depend on this property and do not protect against
193 // double scans.
194
195 uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start());
196 cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index;
197
198 NOISY(tty->print_cr("===========================================================================");)
199 NOISY(tty->print_cr(" process_chunk_boundary: Called with [" PTR_FORMAT "," PTR_FORMAT ")",
200 chunk_mr.start(), chunk_mr.end());)
201
202 // First, set "our" lowest_non_clean entry, which would be
203 // used by the thread scanning an adjoining left chunk with
204 // a non-array object straddling the mutual boundary.
205 // Find the object that spans our boundary, if one exists.
206 // first_block is the block possibly straddling our left boundary.
207 HeapWord* first_block = sp->block_start(chunk_mr.start());
208 assert((chunk_mr.start() != used.start()) || (first_block == chunk_mr.start()),
209 "First chunk should always have a co-initial block");
210 // Does the block straddle the chunk's left boundary, and is it
211 // a non-array object?
212 if (first_block < chunk_mr.start() // first block straddles left bdry
213 && sp->block_is_obj(first_block) // first block is an object
214 && !(oop(first_block)->is_objArray() // first block is not an array (arrays are precisely dirtied)
215 || oop(first_block)->is_typeArray())) {
216 // Find our least non-clean card, so that a left neighbour
217 // does not scan an object straddling the mutual boundary
218 // too far to the right, and attempt to scan a portion of
219 // that object twice.
220 jbyte* first_dirty_card = NULL;
221 jbyte* last_card_of_first_obj =
222 byte_for(first_block + sp->block_size(first_block) - 1);
223 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
224 jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());
225 jbyte* last_card_to_check =
226 (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,
227 (intptr_t) last_card_of_first_obj);
228 // Note that this does not need to go beyond our last card
229 // if our first object completely straddles this chunk.
230 for (jbyte* cur = first_card_of_cur_chunk;
231 cur <= last_card_to_check; cur++) {
232 jbyte val = *cur;
233 if (card_will_be_scanned(val)) {
234 first_dirty_card = cur; break;
235 } else {
236 assert(!card_may_have_been_dirty(val), "Error");
237 }
238 }
239 if (first_dirty_card != NULL) {
240 NOISY(tty->print_cr(" LNC: Found a dirty card at " PTR_FORMAT " in current chunk",
241 first_dirty_card);)
242 assert(0 <= cur_chunk_index && cur_chunk_index < lowest_non_clean_chunk_size,
243 "Bounds error.");
244 assert(lowest_non_clean[cur_chunk_index] == NULL,
245 "Write exactly once : value should be stable hereafter for this round");
246 lowest_non_clean[cur_chunk_index] = first_dirty_card;
247 } NOISY(else {
248 tty->print_cr(" LNC: Found no dirty card in current chunk; leaving LNC entry NULL");
249 // In the future, we could have this thread look for a non-NULL value to copy from its
250 // right neighbour (up to the end of the first object).
251 if (last_card_of_cur_chunk < last_card_of_first_obj) {
252 tty->print_cr(" LNC: BEWARE!!! first obj straddles past right end of chunk:\n"
253 " might be efficient to get value from right neighbour?");
254 }
255 })
256 } else {
257 // In this case we can help our neighbour by just asking them
258 // to stop at our first card (even though it may not be dirty).
259 NOISY(tty->print_cr(" LNC: first block is not a non-array object; setting LNC to first card of current chunk");)
260 assert(lowest_non_clean[cur_chunk_index] == NULL, "Write once : value should be stable hereafter");
261 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
262 lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk;
263 }
264 NOISY(tty->print_cr(" process_chunk_boundary: lowest_non_clean[" INTPTR_FORMAT "] = " PTR_FORMAT
265 " which corresponds to the heap address " PTR_FORMAT,
266 cur_chunk_index, lowest_non_clean[cur_chunk_index],
267 (lowest_non_clean[cur_chunk_index] != NULL)
268 ? addr_for(lowest_non_clean[cur_chunk_index])
269 : NULL);)
270 NOISY(tty->print_cr("---------------------------------------------------------------------------");)
271
272 // Next, set our own max_to_do, which will strictly/exclusively bound
273 // the highest address that we will scan past the right end of our chunk.
274 HeapWord* max_to_do = NULL;
275 if (chunk_mr.end() < used.end()) {
276 // This is not the last chunk in the used region.
277 // What is our last block? We check the first block of
278 // the next (right) chunk rather than strictly check our last block
279 // because it's potentially more efficient to do so.
280 HeapWord* const last_block = sp->block_start(chunk_mr.end());
281 assert(last_block <= chunk_mr.end(), "In case this property changes.");
282 if ((last_block == chunk_mr.end()) // our last block does not straddle boundary
283 || !sp->block_is_obj(last_block) // last_block isn't an object
284 || oop(last_block)->is_objArray() // last_block is an array (precisely marked)
285 || oop(last_block)->is_typeArray()) {
286 max_to_do = chunk_mr.end();
287 NOISY(tty->print_cr(" process_chunk_boundary: Last block on this card is not a non-array object;\n"
288 " max_to_do left at " PTR_FORMAT, max_to_do);)
289 } else {
290 assert(last_block < chunk_mr.end(), "Tautology");
291 // It is a non-array object that straddles the right boundary of this chunk.
292 // last_obj_card is the card corresponding to the start of the last object
293 // in the chunk. Note that the last object may not start in
294 // the chunk.
295 jbyte* const last_obj_card = byte_for(last_block);
296 const jbyte val = *last_obj_card;
297 if (!card_will_be_scanned(val)) {
298 assert(!card_may_have_been_dirty(val), "Error");
299 // The card containing the head is not dirty. Any marks on
300 // subsequent cards still in this chunk must have been made
301 // precisely; we can cap processing at the end of our chunk.
302 max_to_do = chunk_mr.end();
303 NOISY(tty->print_cr(" process_chunk_boundary: Head of last object on this card is not dirty;\n"
304 " max_to_do left at " PTR_FORMAT,
305 max_to_do);)
306 } else {
307 // The last object must be considered dirty, and extends onto the
308 // following chunk. Look for a dirty card in that chunk that will
309 // bound our processing.
310 jbyte* limit_card = NULL;
311 const size_t last_block_size = sp->block_size(last_block);
312 jbyte* const last_card_of_last_obj =
313 byte_for(last_block + last_block_size - 1);
314 jbyte* const first_card_of_next_chunk = byte_for(chunk_mr.end());
315 // This search potentially goes a long distance looking
316 // for the next card that will be scanned, terminating
317 // at the end of the last_block, if no earlier dirty card
318 // is found.
319 assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == ParGCCardsPerStrideChunk,
320 "last card of next chunk may be wrong");
321 for (jbyte* cur = first_card_of_next_chunk;
322 cur <= last_card_of_last_obj; cur++) {
323 const jbyte val = *cur;
324 if (card_will_be_scanned(val)) {
325 NOISY(tty->print_cr(" Found a non-clean card " PTR_FORMAT " with value 0x%x",
326 cur, (int)val);)
327 limit_card = cur; break;
328 } else {
329 assert(!card_may_have_been_dirty(val), "Error: card can't be skipped");
330 }
331 }
332 if (limit_card != NULL) {
333 max_to_do = addr_for(limit_card);
334 assert(limit_card != NULL && max_to_do != NULL, "Error");
335 NOISY(tty->print_cr(" process_chunk_boundary: Found a dirty card at " PTR_FORMAT
336 " max_to_do set at " PTR_FORMAT " which is before end of last block in chunk: "
337 PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
338 limit_card, max_to_do, last_block, last_block_size, (last_block+last_block_size));)
339 } else {
340 // The following is a pessimistic value, because it's possible
341 // that a dirty card on a subsequent chunk has been cleared by
342 // the time we get to look at it; we'll correct for that further below,
343 // using the LNC array which records the least non-clean card
344 // before cards were cleared in a particular chunk.
345 limit_card = last_card_of_last_obj;
346 max_to_do = last_block + last_block_size;
347 assert(limit_card != NULL && max_to_do != NULL, "Error");
348 NOISY(tty->print_cr(" process_chunk_boundary: Found no dirty card before end of last block in chunk\n"
349 " Setting limit_card to " PTR_FORMAT
350 " and max_to_do " PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
351 limit_card, last_block, last_block_size, max_to_do);)
352 }
353 assert(0 < cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size,
354 "Bounds error.");
355 // It is possible that a dirty card for the last object may have been
356 // cleared before we had a chance to examine it. In that case, the value
357 // will have been logged in the LNC for that chunk.
358 // We need to examine as many chunks to the right as this object
359 // covers. However, we need to bound this checking to the largest
360 // entry in the LNC array: this is because the heap may expand
361 // after the LNC array has been created but before we reach this point,
362 // and the last block in our chunk may have been expanded to include
363 // the expansion delta (and possibly subsequently allocated from, so
364 // it wouldn't be sufficient to check whether that last block was
365 // or was not an object at this point).
366 uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1)
367 - lowest_non_clean_base_chunk_index;
368 const uintptr_t last_chunk_index = addr_to_chunk_index(used.last())
369 - lowest_non_clean_base_chunk_index;
370 if (last_chunk_index_to_check > last_chunk_index) {
371 assert(last_block + last_block_size > used.end(),
372 err_msg("Inconsistency detected: last_block [" PTR_FORMAT "," PTR_FORMAT "]"
373 " does not exceed used.end() = " PTR_FORMAT ","
374 " yet last_chunk_index_to_check " INTPTR_FORMAT
375 " exceeds last_chunk_index " INTPTR_FORMAT,
376 last_block, last_block + last_block_size,
377 used.end(),
378 last_chunk_index_to_check, last_chunk_index));
379 assert(sp->used_region().end() > used.end(),
380 err_msg("Expansion did not happen: "
381 "[" PTR_FORMAT "," PTR_FORMAT ") -> [" PTR_FORMAT "," PTR_FORMAT ")",
382 sp->used_region().start(), sp->used_region().end(), used.start(), used.end()));
383 NOISY(tty->print_cr(" process_chunk_boundary: heap expanded; explicitly bounding last_chunk");)
384 last_chunk_index_to_check = last_chunk_index;
385 }
386 for (uintptr_t lnc_index = cur_chunk_index + 1;
387 lnc_index <= last_chunk_index_to_check;
388 lnc_index++) {
389 jbyte* lnc_card = lowest_non_clean[lnc_index];
390 if (lnc_card != NULL) {
391 // we can stop at the first non-NULL entry we find
392 if (lnc_card <= limit_card) {
393 NOISY(tty->print_cr(" process_chunk_boundary: LNC card " PTR_FORMAT " is lower than limit_card " PTR_FORMAT,
394 " max_to_do will be lowered to " PTR_FORMAT " from " PTR_FORMAT,
395 lnc_card, limit_card, addr_for(lnc_card), max_to_do);)
396 limit_card = lnc_card;
397 max_to_do = addr_for(limit_card);
398 assert(limit_card != NULL && max_to_do != NULL, "Error");
399 }
400 // In any case, we break now
401 break;
402 } // else continue to look for a non-NULL entry if any
403 }
404 assert(limit_card != NULL && max_to_do != NULL, "Error");
405 }
406 assert(max_to_do != NULL, "OOPS 1 !");
407 }
408 assert(max_to_do != NULL, "OOPS 2!");
409 } else {
410 max_to_do = used.end();
411 NOISY(tty->print_cr(" process_chunk_boundary: Last chunk of this space;\n"
412 " max_to_do left at " PTR_FORMAT,
413 max_to_do);)
414 }
415 assert(max_to_do != NULL, "OOPS 3!");
416 // Now we can set the closure we're using so it doesn't to beyond
417 // max_to_do.
418 dcto_cl->set_min_done(max_to_do);
419 #ifndef PRODUCT
420 dcto_cl->set_last_bottom(max_to_do);
421 #endif
422 NOISY(tty->print_cr("===========================================================================\n");)
423 }
424
425 #undef NOISY
426
427 void
428 CardTableModRefBS::
429 get_LNC_array_for_space(Space* sp,
430 jbyte**& lowest_non_clean,
431 uintptr_t& lowest_non_clean_base_chunk_index,
432 size_t& lowest_non_clean_chunk_size) {
433
434 int i = find_covering_region_containing(sp->bottom());
435 MemRegion covered = _covered[i];
436 size_t n_chunks = chunks_to_cover(covered);
437
438 // Only the first thread to obtain the lock will resize the
439 // LNC array for the covered region. Any later expansion can't affect
440 // the used_at_save_marks region.
441 // (I observed a bug in which the first thread to execute this would
442 // resize, and then it would cause "expand_and_allocate" that would
443 // increase the number of chunks in the covered region. Then a second
444 // thread would come and execute this, see that the size didn't match,
445 // and free and allocate again. So the first thread would be using a
446 // freed "_lowest_non_clean" array.)
447
448 // Do a dirty read here. If we pass the conditional then take the rare
449 // event lock and do the read again in case some other thread had already
450 // succeeded and done the resize.
451 int cur_collection = Universe::heap()->total_collections();
452 if (_last_LNC_resizing_collection[i] != cur_collection) {
453 MutexLocker x(ParGCRareEvent_lock);
454 if (_last_LNC_resizing_collection[i] != cur_collection) {
455 if (_lowest_non_clean[i] == NULL ||
456 n_chunks != _lowest_non_clean_chunk_size[i]) {
457
458 // Should we delete the old?
459 if (_lowest_non_clean[i] != NULL) {
460 assert(n_chunks != _lowest_non_clean_chunk_size[i],
461 "logical consequence");
462 FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i], mtGC);
463 _lowest_non_clean[i] = NULL;
464 }
465 // Now allocate a new one if necessary.
466 if (_lowest_non_clean[i] == NULL) {
467 _lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks, mtGC);
468 _lowest_non_clean_chunk_size[i] = n_chunks;
469 _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());
470 for (int j = 0; j < (int)n_chunks; j++)
471 _lowest_non_clean[i][j] = NULL;
472 }
473 }
474 _last_LNC_resizing_collection[i] = cur_collection;
475 }
476 }
477 // In any case, now do the initialization.
478 lowest_non_clean = _lowest_non_clean[i];
479 lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];
480 lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i];
481 }