1 /*
2 * Copyright (c) 2007, 2014, 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/orderAccess.inline.hpp"
36 #include "runtime/virtualspace.hpp"
37 #include "runtime/vmThread.hpp"
38
39 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
40
41 void CardTableModRefBS::non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr,
42 OopsInGenClosure* cl,
43 CardTableRS* ct,
44 int n_threads) {
45 assert(n_threads > 0, "Error: expected n_threads > 0");
46 assert((n_threads == 1 && ParallelGCThreads == 0) ||
47 n_threads <= (int)ParallelGCThreads,
48 "# worker threads != # requested!");
49 assert(!Thread::current()->is_VM_thread() || (n_threads == 1), "There is only 1 VM thread");
50 assert(UseDynamicNumberOfGCThreads ||
51 !FLAG_IS_DEFAULT(ParallelGCThreads) ||
52 n_threads == (int)ParallelGCThreads,
53 "# worker threads != # requested!");
54 // Make sure the LNC array is valid for the space.
55 jbyte** lowest_non_clean;
56 uintptr_t lowest_non_clean_base_chunk_index;
57 size_t lowest_non_clean_chunk_size;
58 get_LNC_array_for_space(sp, lowest_non_clean,
59 lowest_non_clean_base_chunk_index,
60 lowest_non_clean_chunk_size);
61
62 uint n_strides = n_threads * ParGCStridesPerThread;
63 SequentialSubTasksDone* pst = sp->par_seq_tasks();
64 // Sets the condition for completion of the subtask (how many threads
65 // need to finish in order to be done).
66 pst->set_n_threads(n_threads);
67 pst->set_n_tasks(n_strides);
68
69 uint stride = 0;
70 while (!pst->is_task_claimed(/* reference */ stride)) {
71 process_stride(sp, mr, stride, n_strides, cl, ct,
72 lowest_non_clean,
73 lowest_non_clean_base_chunk_index,
74 lowest_non_clean_chunk_size);
75 }
76 if (pst->all_tasks_completed()) {
77 // Clear lowest_non_clean array for next time.
78 intptr_t first_chunk_index = addr_to_chunk_index(mr.start());
79 uintptr_t last_chunk_index = addr_to_chunk_index(mr.last());
80 for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) {
81 intptr_t ind = ch - lowest_non_clean_base_chunk_index;
82 assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size,
83 "Bounds error");
84 lowest_non_clean[ind] = NULL;
85 }
86 }
87 }
88
89 void
90 CardTableModRefBS::
91 process_stride(Space* sp,
92 MemRegion used,
93 jint stride, int n_strides,
94 OopsInGenClosure* cl,
95 CardTableRS* ct,
96 jbyte** lowest_non_clean,
97 uintptr_t lowest_non_clean_base_chunk_index,
98 size_t lowest_non_clean_chunk_size) {
99 // We go from higher to lower addresses here; it wouldn't help that much
100 // because of the strided parallelism pattern used here.
101
102 // Find the first card address of the first chunk in the stride that is
103 // at least "bottom" of the used region.
104 jbyte* start_card = byte_for(used.start());
105 jbyte* end_card = byte_after(used.last());
106 uintptr_t start_chunk = addr_to_chunk_index(used.start());
107 uintptr_t start_chunk_stride_num = start_chunk % n_strides;
108 jbyte* chunk_card_start;
109
110 if ((uintptr_t)stride >= start_chunk_stride_num) {
111 chunk_card_start = (jbyte*)(start_card +
112 (stride - start_chunk_stride_num) *
113 ParGCCardsPerStrideChunk);
114 } else {
115 // Go ahead to the next chunk group boundary, then to the requested stride.
116 chunk_card_start = (jbyte*)(start_card +
117 (n_strides - start_chunk_stride_num + stride) *
118 ParGCCardsPerStrideChunk);
119 }
120
121 while (chunk_card_start < end_card) {
122 // Even though we go from lower to higher addresses below, the
123 // strided parallelism can interleave the actual processing of the
124 // dirty pages in various ways. For a specific chunk within this
125 // stride, we take care to avoid double scanning or missing a card
126 // by suitably initializing the "min_done" field in process_chunk_boundaries()
127 // below, together with the dirty region extension accomplished in
128 // DirtyCardToOopClosure::do_MemRegion().
129 jbyte* chunk_card_end = chunk_card_start + ParGCCardsPerStrideChunk;
130 // Invariant: chunk_mr should be fully contained within the "used" region.
131 MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start),
132 chunk_card_end >= end_card ?
133 used.end() : addr_for(chunk_card_end));
134 assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");
135 assert(used.contains(chunk_mr), "chunk_mr should be subset of used");
136
137 DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(),
138 cl->gen_boundary());
139 ClearNoncleanCardWrapper clear_cl(dcto_cl, ct);
140
141
142 // Process the chunk.
143 process_chunk_boundaries(sp,
144 dcto_cl,
145 chunk_mr,
146 used,
147 lowest_non_clean,
148 lowest_non_clean_base_chunk_index,
149 lowest_non_clean_chunk_size);
150
151 // We want the LNC array updates above in process_chunk_boundaries
152 // to be visible before any of the card table value changes as a
153 // result of the dirty card iteration below.
154 OrderAccess::storestore();
155
156 // We do not call the non_clean_card_iterate_serial() version because
157 // we want to clear the cards: clear_cl here does the work of finding
158 // contiguous dirty ranges of cards to process and clear.
159 clear_cl.do_MemRegion(chunk_mr);
160
161 // Find the next chunk of the stride.
162 chunk_card_start += ParGCCardsPerStrideChunk * n_strides;
163 }
164 }
165
166
167 // If you want a talkative process_chunk_boundaries,
168 // then #define NOISY(x) x
169 #ifdef NOISY
170 #error "Encountered a global preprocessor flag, NOISY, which might clash with local definition to follow"
171 #else
172 #define NOISY(x)
173 #endif
174
175 void
176 CardTableModRefBS::
177 process_chunk_boundaries(Space* sp,
178 DirtyCardToOopClosure* dcto_cl,
179 MemRegion chunk_mr,
180 MemRegion used,
181 jbyte** lowest_non_clean,
182 uintptr_t lowest_non_clean_base_chunk_index,
183 size_t lowest_non_clean_chunk_size)
184 {
185 // We must worry about non-array objects that cross chunk boundaries,
186 // because such objects are both precisely and imprecisely marked:
187 // .. if the head of such an object is dirty, the entire object
188 // needs to be scanned, under the interpretation that this
189 // was an imprecise mark
190 // .. if the head of such an object is not dirty, we can assume
191 // precise marking and it's efficient to scan just the dirty
192 // cards.
193 // In either case, each scanned reference must be scanned precisely
194 // once so as to avoid cloning of a young referent. For efficiency,
195 // our closures depend on this property and do not protect against
196 // double scans.
197
198 uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start());
199 cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index;
200
201 NOISY(tty->print_cr("===========================================================================");)
202 NOISY(tty->print_cr(" process_chunk_boundary: Called with [" PTR_FORMAT "," PTR_FORMAT ")",
203 chunk_mr.start(), chunk_mr.end());)
204
205 // First, set "our" lowest_non_clean entry, which would be
206 // used by the thread scanning an adjoining left chunk with
207 // a non-array object straddling the mutual boundary.
208 // Find the object that spans our boundary, if one exists.
209 // first_block is the block possibly straddling our left boundary.
210 HeapWord* first_block = sp->block_start(chunk_mr.start());
211 assert((chunk_mr.start() != used.start()) || (first_block == chunk_mr.start()),
212 "First chunk should always have a co-initial block");
213 // Does the block straddle the chunk's left boundary, and is it
214 // a non-array object?
215 if (first_block < chunk_mr.start() // first block straddles left bdry
216 && sp->block_is_obj(first_block) // first block is an object
217 && !(oop(first_block)->is_objArray() // first block is not an array (arrays are precisely dirtied)
218 || oop(first_block)->is_typeArray())) {
219 // Find our least non-clean card, so that a left neighbour
220 // does not scan an object straddling the mutual boundary
221 // too far to the right, and attempt to scan a portion of
222 // that object twice.
223 jbyte* first_dirty_card = NULL;
224 jbyte* last_card_of_first_obj =
225 byte_for(first_block + sp->block_size(first_block) - 1);
226 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
227 jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());
228 jbyte* last_card_to_check =
229 (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,
230 (intptr_t) last_card_of_first_obj);
231 // Note that this does not need to go beyond our last card
232 // if our first object completely straddles this chunk.
233 for (jbyte* cur = first_card_of_cur_chunk;
234 cur <= last_card_to_check; cur++) {
235 jbyte val = *cur;
236 if (card_will_be_scanned(val)) {
237 first_dirty_card = cur; break;
238 } else {
239 assert(!card_may_have_been_dirty(val), "Error");
240 }
241 }
242 if (first_dirty_card != NULL) {
243 NOISY(tty->print_cr(" LNC: Found a dirty card at " PTR_FORMAT " in current chunk",
244 first_dirty_card);)
245 assert(0 <= cur_chunk_index && cur_chunk_index < lowest_non_clean_chunk_size,
246 "Bounds error.");
247 assert(lowest_non_clean[cur_chunk_index] == NULL,
248 "Write exactly once : value should be stable hereafter for this round");
249 lowest_non_clean[cur_chunk_index] = first_dirty_card;
250 } NOISY(else {
251 tty->print_cr(" LNC: Found no dirty card in current chunk; leaving LNC entry NULL");
252 // In the future, we could have this thread look for a non-NULL value to copy from its
253 // right neighbour (up to the end of the first object).
254 if (last_card_of_cur_chunk < last_card_of_first_obj) {
255 tty->print_cr(" LNC: BEWARE!!! first obj straddles past right end of chunk:\n"
256 " might be efficient to get value from right neighbour?");
257 }
258 })
259 } else {
260 // In this case we can help our neighbour by just asking them
261 // to stop at our first card (even though it may not be dirty).
262 NOISY(tty->print_cr(" LNC: first block is not a non-array object; setting LNC to first card of current chunk");)
263 assert(lowest_non_clean[cur_chunk_index] == NULL, "Write once : value should be stable hereafter");
264 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
265 lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk;
266 }
267 NOISY(tty->print_cr(" process_chunk_boundary: lowest_non_clean[" INTPTR_FORMAT "] = " PTR_FORMAT
268 " which corresponds to the heap address " PTR_FORMAT,
269 cur_chunk_index, lowest_non_clean[cur_chunk_index],
270 (lowest_non_clean[cur_chunk_index] != NULL)
271 ? addr_for(lowest_non_clean[cur_chunk_index])
272 : NULL);)
273 NOISY(tty->print_cr("---------------------------------------------------------------------------");)
274
275 // Next, set our own max_to_do, which will strictly/exclusively bound
276 // the highest address that we will scan past the right end of our chunk.
277 HeapWord* max_to_do = NULL;
278 if (chunk_mr.end() < used.end()) {
279 // This is not the last chunk in the used region.
280 // What is our last block? We check the first block of
281 // the next (right) chunk rather than strictly check our last block
282 // because it's potentially more efficient to do so.
283 HeapWord* const last_block = sp->block_start(chunk_mr.end());
284 assert(last_block <= chunk_mr.end(), "In case this property changes.");
285 if ((last_block == chunk_mr.end()) // our last block does not straddle boundary
286 || !sp->block_is_obj(last_block) // last_block isn't an object
287 || oop(last_block)->is_objArray() // last_block is an array (precisely marked)
288 || oop(last_block)->is_typeArray()) {
289 max_to_do = chunk_mr.end();
290 NOISY(tty->print_cr(" process_chunk_boundary: Last block on this card is not a non-array object;\n"
291 " max_to_do left at " PTR_FORMAT, max_to_do);)
292 } else {
293 assert(last_block < chunk_mr.end(), "Tautology");
294 // It is a non-array object that straddles the right boundary of this chunk.
295 // last_obj_card is the card corresponding to the start of the last object
296 // in the chunk. Note that the last object may not start in
297 // the chunk.
298 jbyte* const last_obj_card = byte_for(last_block);
299 const jbyte val = *last_obj_card;
300 if (!card_will_be_scanned(val)) {
301 assert(!card_may_have_been_dirty(val), "Error");
302 // The card containing the head is not dirty. Any marks on
303 // subsequent cards still in this chunk must have been made
304 // precisely; we can cap processing at the end of our chunk.
305 max_to_do = chunk_mr.end();
306 NOISY(tty->print_cr(" process_chunk_boundary: Head of last object on this card is not dirty;\n"
307 " max_to_do left at " PTR_FORMAT,
308 max_to_do);)
309 } else {
310 // The last object must be considered dirty, and extends onto the
311 // following chunk. Look for a dirty card in that chunk that will
312 // bound our processing.
313 jbyte* limit_card = NULL;
314 const size_t last_block_size = sp->block_size(last_block);
315 jbyte* const last_card_of_last_obj =
316 byte_for(last_block + last_block_size - 1);
317 jbyte* const first_card_of_next_chunk = byte_for(chunk_mr.end());
318 // This search potentially goes a long distance looking
319 // for the next card that will be scanned, terminating
320 // at the end of the last_block, if no earlier dirty card
321 // is found.
322 assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == ParGCCardsPerStrideChunk,
323 "last card of next chunk may be wrong");
324 for (jbyte* cur = first_card_of_next_chunk;
325 cur <= last_card_of_last_obj; cur++) {
326 const jbyte val = *cur;
327 if (card_will_be_scanned(val)) {
328 NOISY(tty->print_cr(" Found a non-clean card " PTR_FORMAT " with value 0x%x",
329 cur, (int)val);)
330 limit_card = cur; break;
331 } else {
332 assert(!card_may_have_been_dirty(val), "Error: card can't be skipped");
333 }
334 }
335 if (limit_card != NULL) {
336 max_to_do = addr_for(limit_card);
337 assert(limit_card != NULL && max_to_do != NULL, "Error");
338 NOISY(tty->print_cr(" process_chunk_boundary: Found a dirty card at " PTR_FORMAT
339 " max_to_do set at " PTR_FORMAT " which is before end of last block in chunk: "
340 PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
341 limit_card, max_to_do, last_block, last_block_size, (last_block+last_block_size));)
342 } else {
343 // The following is a pessimistic value, because it's possible
344 // that a dirty card on a subsequent chunk has been cleared by
345 // the time we get to look at it; we'll correct for that further below,
346 // using the LNC array which records the least non-clean card
347 // before cards were cleared in a particular chunk.
348 limit_card = last_card_of_last_obj;
349 max_to_do = last_block + last_block_size;
350 assert(limit_card != NULL && max_to_do != NULL, "Error");
351 NOISY(tty->print_cr(" process_chunk_boundary: Found no dirty card before end of last block in chunk\n"
352 " Setting limit_card to " PTR_FORMAT
353 " and max_to_do " PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
354 limit_card, last_block, last_block_size, max_to_do);)
355 }
356 assert(0 < cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size,
357 "Bounds error.");
358 // It is possible that a dirty card for the last object may have been
359 // cleared before we had a chance to examine it. In that case, the value
360 // will have been logged in the LNC for that chunk.
361 // We need to examine as many chunks to the right as this object
362 // covers. However, we need to bound this checking to the largest
363 // entry in the LNC array: this is because the heap may expand
364 // after the LNC array has been created but before we reach this point,
365 // and the last block in our chunk may have been expanded to include
366 // the expansion delta (and possibly subsequently allocated from, so
367 // it wouldn't be sufficient to check whether that last block was
368 // or was not an object at this point).
369 uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1)
370 - lowest_non_clean_base_chunk_index;
371 const uintptr_t last_chunk_index = addr_to_chunk_index(used.last())
372 - lowest_non_clean_base_chunk_index;
373 if (last_chunk_index_to_check > last_chunk_index) {
374 assert(last_block + last_block_size > used.end(),
375 err_msg("Inconsistency detected: last_block [" PTR_FORMAT "," PTR_FORMAT "]"
376 " does not exceed used.end() = " PTR_FORMAT ","
377 " yet last_chunk_index_to_check " INTPTR_FORMAT
378 " exceeds last_chunk_index " INTPTR_FORMAT,
379 last_block, last_block + last_block_size,
380 used.end(),
381 last_chunk_index_to_check, last_chunk_index));
382 assert(sp->used_region().end() > used.end(),
383 err_msg("Expansion did not happen: "
384 "[" PTR_FORMAT "," PTR_FORMAT ") -> [" PTR_FORMAT "," PTR_FORMAT ")",
385 sp->used_region().start(), sp->used_region().end(), used.start(), used.end()));
386 NOISY(tty->print_cr(" process_chunk_boundary: heap expanded; explicitly bounding last_chunk");)
387 last_chunk_index_to_check = last_chunk_index;
388 }
389 for (uintptr_t lnc_index = cur_chunk_index + 1;
390 lnc_index <= last_chunk_index_to_check;
391 lnc_index++) {
392 jbyte* lnc_card = lowest_non_clean[lnc_index];
393 if (lnc_card != NULL) {
394 // we can stop at the first non-NULL entry we find
395 if (lnc_card <= limit_card) {
396 NOISY(tty->print_cr(" process_chunk_boundary: LNC card " PTR_FORMAT " is lower than limit_card " PTR_FORMAT,
397 " max_to_do will be lowered to " PTR_FORMAT " from " PTR_FORMAT,
398 lnc_card, limit_card, addr_for(lnc_card), max_to_do);)
399 limit_card = lnc_card;
400 max_to_do = addr_for(limit_card);
401 assert(limit_card != NULL && max_to_do != NULL, "Error");
402 }
403 // In any case, we break now
404 break;
405 } // else continue to look for a non-NULL entry if any
406 }
407 assert(limit_card != NULL && max_to_do != NULL, "Error");
408 }
409 assert(max_to_do != NULL, "OOPS 1 !");
410 }
411 assert(max_to_do != NULL, "OOPS 2!");
412 } else {
413 max_to_do = used.end();
414 NOISY(tty->print_cr(" process_chunk_boundary: Last chunk of this space;\n"
415 " max_to_do left at " PTR_FORMAT,
416 max_to_do);)
417 }
418 assert(max_to_do != NULL, "OOPS 3!");
419 // Now we can set the closure we're using so it doesn't to beyond
420 // max_to_do.
421 dcto_cl->set_min_done(max_to_do);
422 #ifndef PRODUCT
423 dcto_cl->set_last_bottom(max_to_do);
424 #endif
425 NOISY(tty->print_cr("===========================================================================\n");)
426 }
427
428 #undef NOISY
429
430 void
431 CardTableModRefBS::
432 get_LNC_array_for_space(Space* sp,
433 jbyte**& lowest_non_clean,
434 uintptr_t& lowest_non_clean_base_chunk_index,
435 size_t& lowest_non_clean_chunk_size) {
436
437 int i = find_covering_region_containing(sp->bottom());
438 MemRegion covered = _covered[i];
439 size_t n_chunks = chunks_to_cover(covered);
440
441 // Only the first thread to obtain the lock will resize the
442 // LNC array for the covered region. Any later expansion can't affect
443 // the used_at_save_marks region.
444 // (I observed a bug in which the first thread to execute this would
445 // resize, and then it would cause "expand_and_allocate" that would
446 // increase the number of chunks in the covered region. Then a second
447 // thread would come and execute this, see that the size didn't match,
448 // and free and allocate again. So the first thread would be using a
449 // freed "_lowest_non_clean" array.)
450
451 // Do a dirty read here. If we pass the conditional then take the rare
452 // event lock and do the read again in case some other thread had already
453 // succeeded and done the resize.
454 int cur_collection = Universe::heap()->total_collections();
455 // Updated _last_LNC_resizing_collection[i] must not be visible before
456 // _lowest_non_clean and friends are visible. Therefore use acquire/release
457 // to guarantee this on non TSO architecures.
458 if (OrderAccess::load_acquire(&_last_LNC_resizing_collection[i]) != cur_collection) {
459 MutexLocker x(ParGCRareEvent_lock);
460 // This load_acquire is here for clarity only. The MutexLocker already fences.
461 if (OrderAccess::load_acquire(&_last_LNC_resizing_collection[i]) != cur_collection) {
462 if (_lowest_non_clean[i] == NULL ||
463 n_chunks != _lowest_non_clean_chunk_size[i]) {
464
465 // Should we delete the old?
466 if (_lowest_non_clean[i] != NULL) {
467 assert(n_chunks != _lowest_non_clean_chunk_size[i],
468 "logical consequence");
469 FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i], mtGC);
470 _lowest_non_clean[i] = NULL;
471 }
472 // Now allocate a new one if necessary.
473 if (_lowest_non_clean[i] == NULL) {
474 _lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks, mtGC);
475 _lowest_non_clean_chunk_size[i] = n_chunks;
476 _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());
477 for (int j = 0; j < (int)n_chunks; j++)
478 _lowest_non_clean[i][j] = NULL;
479 }
480 }
481 // Make sure this gets visible only after _lowest_non_clean* was initialized
482 OrderAccess::release_store(&_last_LNC_resizing_collection[i], cur_collection);
483 }
484 }
485 // In any case, now do the initialization.
486 lowest_non_clean = _lowest_non_clean[i];
487 lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];
488 lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i];
489 }