Tue, 08 Aug 2017 15:57:29 +0800
merge
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 */
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"
38 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
40 void CardTableModRefBS::non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr,
41 OopsInGenClosure* cl,
42 CardTableRS* ct,
43 int n_threads) {
44 assert(n_threads > 0, "Error: expected n_threads > 0");
45 assert((n_threads == 1 && ParallelGCThreads == 0) ||
46 n_threads <= (int)ParallelGCThreads,
47 "# worker threads != # requested!");
48 assert(!Thread::current()->is_VM_thread() || (n_threads == 1), "There is only 1 VM thread");
49 assert(UseDynamicNumberOfGCThreads ||
50 !FLAG_IS_DEFAULT(ParallelGCThreads) ||
51 n_threads == (int)ParallelGCThreads,
52 "# worker threads != # requested!");
53 // Make sure the LNC array is valid for the space.
54 jbyte** lowest_non_clean;
55 uintptr_t lowest_non_clean_base_chunk_index;
56 size_t lowest_non_clean_chunk_size;
57 get_LNC_array_for_space(sp, lowest_non_clean,
58 lowest_non_clean_base_chunk_index,
59 lowest_non_clean_chunk_size);
61 uint n_strides = n_threads * ParGCStridesPerThread;
62 SequentialSubTasksDone* pst = sp->par_seq_tasks();
63 // Sets the condition for completion of the subtask (how many threads
64 // need to finish in order to be done).
65 pst->set_n_threads(n_threads);
66 pst->set_n_tasks(n_strides);
68 uint stride = 0;
69 while (!pst->is_task_claimed(/* reference */ stride)) {
70 process_stride(sp, mr, stride, n_strides, cl, ct,
71 lowest_non_clean,
72 lowest_non_clean_base_chunk_index,
73 lowest_non_clean_chunk_size);
74 }
75 if (pst->all_tasks_completed()) {
76 // Clear lowest_non_clean array for next time.
77 intptr_t first_chunk_index = addr_to_chunk_index(mr.start());
78 uintptr_t last_chunk_index = addr_to_chunk_index(mr.last());
79 for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) {
80 intptr_t ind = ch - lowest_non_clean_base_chunk_index;
81 assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size,
82 "Bounds error");
83 lowest_non_clean[ind] = NULL;
84 }
85 }
86 }
88 void
89 CardTableModRefBS::
90 process_stride(Space* sp,
91 MemRegion used,
92 jint stride, int n_strides,
93 OopsInGenClosure* cl,
94 CardTableRS* ct,
95 jbyte** lowest_non_clean,
96 uintptr_t lowest_non_clean_base_chunk_index,
97 size_t lowest_non_clean_chunk_size) {
98 // We go from higher to lower addresses here; it wouldn't help that much
99 // because of the strided parallelism pattern used here.
101 // Find the first card address of the first chunk in the stride that is
102 // at least "bottom" of the used region.
103 jbyte* start_card = byte_for(used.start());
104 jbyte* end_card = byte_after(used.last());
105 uintptr_t start_chunk = addr_to_chunk_index(used.start());
106 uintptr_t start_chunk_stride_num = start_chunk % n_strides;
107 jbyte* chunk_card_start;
109 if ((uintptr_t)stride >= start_chunk_stride_num) {
110 chunk_card_start = (jbyte*)(start_card +
111 (stride - start_chunk_stride_num) *
112 ParGCCardsPerStrideChunk);
113 } else {
114 // Go ahead to the next chunk group boundary, then to the requested stride.
115 chunk_card_start = (jbyte*)(start_card +
116 (n_strides - start_chunk_stride_num + stride) *
117 ParGCCardsPerStrideChunk);
118 }
120 while (chunk_card_start < end_card) {
121 // Even though we go from lower to higher addresses below, the
122 // strided parallelism can interleave the actual processing of the
123 // dirty pages in various ways. For a specific chunk within this
124 // stride, we take care to avoid double scanning or missing a card
125 // by suitably initializing the "min_done" field in process_chunk_boundaries()
126 // below, together with the dirty region extension accomplished in
127 // DirtyCardToOopClosure::do_MemRegion().
128 jbyte* chunk_card_end = chunk_card_start + ParGCCardsPerStrideChunk;
129 // Invariant: chunk_mr should be fully contained within the "used" region.
130 MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start),
131 chunk_card_end >= end_card ?
132 used.end() : addr_for(chunk_card_end));
133 assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");
134 assert(used.contains(chunk_mr), "chunk_mr should be subset of used");
136 DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(),
137 cl->gen_boundary());
138 ClearNoncleanCardWrapper clear_cl(dcto_cl, ct);
141 // Process the chunk.
142 process_chunk_boundaries(sp,
143 dcto_cl,
144 chunk_mr,
145 used,
146 lowest_non_clean,
147 lowest_non_clean_base_chunk_index,
148 lowest_non_clean_chunk_size);
150 // We want the LNC array updates above in process_chunk_boundaries
151 // to be visible before any of the card table value changes as a
152 // result of the dirty card iteration below.
153 OrderAccess::storestore();
155 // We do not call the non_clean_card_iterate_serial() version because
156 // we want to clear the cards: clear_cl here does the work of finding
157 // contiguous dirty ranges of cards to process and clear.
158 clear_cl.do_MemRegion(chunk_mr);
160 // Find the next chunk of the stride.
161 chunk_card_start += ParGCCardsPerStrideChunk * n_strides;
162 }
163 }
166 // If you want a talkative process_chunk_boundaries,
167 // then #define NOISY(x) x
168 #ifdef NOISY
169 #error "Encountered a global preprocessor flag, NOISY, which might clash with local definition to follow"
170 #else
171 #define NOISY(x)
172 #endif
174 void
175 CardTableModRefBS::
176 process_chunk_boundaries(Space* sp,
177 DirtyCardToOopClosure* dcto_cl,
178 MemRegion chunk_mr,
179 MemRegion used,
180 jbyte** lowest_non_clean,
181 uintptr_t lowest_non_clean_base_chunk_index,
182 size_t lowest_non_clean_chunk_size)
183 {
184 // We must worry about non-array objects that cross chunk boundaries,
185 // because such objects are both precisely and imprecisely marked:
186 // .. if the head of such an object is dirty, the entire object
187 // needs to be scanned, under the interpretation that this
188 // was an imprecise mark
189 // .. if the head of such an object is not dirty, we can assume
190 // precise marking and it's efficient to scan just the dirty
191 // cards.
192 // In either case, each scanned reference must be scanned precisely
193 // once so as to avoid cloning of a young referent. For efficiency,
194 // our closures depend on this property and do not protect against
195 // double scans.
197 uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start());
198 cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index;
200 NOISY(tty->print_cr("===========================================================================");)
201 NOISY(tty->print_cr(" process_chunk_boundary: Called with [" PTR_FORMAT "," PTR_FORMAT ")",
202 chunk_mr.start(), chunk_mr.end());)
204 // First, set "our" lowest_non_clean entry, which would be
205 // used by the thread scanning an adjoining left chunk with
206 // a non-array object straddling the mutual boundary.
207 // Find the object that spans our boundary, if one exists.
208 // first_block is the block possibly straddling our left boundary.
209 HeapWord* first_block = sp->block_start(chunk_mr.start());
210 assert((chunk_mr.start() != used.start()) || (first_block == chunk_mr.start()),
211 "First chunk should always have a co-initial block");
212 // Does the block straddle the chunk's left boundary, and is it
213 // a non-array object?
214 if (first_block < chunk_mr.start() // first block straddles left bdry
215 && sp->block_is_obj(first_block) // first block is an object
216 && !(oop(first_block)->is_objArray() // first block is not an array (arrays are precisely dirtied)
217 || oop(first_block)->is_typeArray())) {
218 // Find our least non-clean card, so that a left neighbour
219 // does not scan an object straddling the mutual boundary
220 // too far to the right, and attempt to scan a portion of
221 // that object twice.
222 jbyte* first_dirty_card = NULL;
223 jbyte* last_card_of_first_obj =
224 byte_for(first_block + sp->block_size(first_block) - 1);
225 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
226 jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());
227 jbyte* last_card_to_check =
228 (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,
229 (intptr_t) last_card_of_first_obj);
230 // Note that this does not need to go beyond our last card
231 // if our first object completely straddles this chunk.
232 for (jbyte* cur = first_card_of_cur_chunk;
233 cur <= last_card_to_check; cur++) {
234 jbyte val = *cur;
235 if (card_will_be_scanned(val)) {
236 first_dirty_card = cur; break;
237 } else {
238 assert(!card_may_have_been_dirty(val), "Error");
239 }
240 }
241 if (first_dirty_card != NULL) {
242 NOISY(tty->print_cr(" LNC: Found a dirty card at " PTR_FORMAT " in current chunk",
243 first_dirty_card);)
244 assert(0 <= cur_chunk_index && cur_chunk_index < lowest_non_clean_chunk_size,
245 "Bounds error.");
246 assert(lowest_non_clean[cur_chunk_index] == NULL,
247 "Write exactly once : value should be stable hereafter for this round");
248 lowest_non_clean[cur_chunk_index] = first_dirty_card;
249 } NOISY(else {
250 tty->print_cr(" LNC: Found no dirty card in current chunk; leaving LNC entry NULL");
251 // In the future, we could have this thread look for a non-NULL value to copy from its
252 // right neighbour (up to the end of the first object).
253 if (last_card_of_cur_chunk < last_card_of_first_obj) {
254 tty->print_cr(" LNC: BEWARE!!! first obj straddles past right end of chunk:\n"
255 " might be efficient to get value from right neighbour?");
256 }
257 })
258 } else {
259 // In this case we can help our neighbour by just asking them
260 // to stop at our first card (even though it may not be dirty).
261 NOISY(tty->print_cr(" LNC: first block is not a non-array object; setting LNC to first card of current chunk");)
262 assert(lowest_non_clean[cur_chunk_index] == NULL, "Write once : value should be stable hereafter");
263 jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
264 lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk;
265 }
266 NOISY(tty->print_cr(" process_chunk_boundary: lowest_non_clean[" INTPTR_FORMAT "] = " PTR_FORMAT
267 " which corresponds to the heap address " PTR_FORMAT,
268 cur_chunk_index, lowest_non_clean[cur_chunk_index],
269 (lowest_non_clean[cur_chunk_index] != NULL)
270 ? addr_for(lowest_non_clean[cur_chunk_index])
271 : NULL);)
272 NOISY(tty->print_cr("---------------------------------------------------------------------------");)
274 // Next, set our own max_to_do, which will strictly/exclusively bound
275 // the highest address that we will scan past the right end of our chunk.
276 HeapWord* max_to_do = NULL;
277 if (chunk_mr.end() < used.end()) {
278 // This is not the last chunk in the used region.
279 // What is our last block? We check the first block of
280 // the next (right) chunk rather than strictly check our last block
281 // because it's potentially more efficient to do so.
282 HeapWord* const last_block = sp->block_start(chunk_mr.end());
283 assert(last_block <= chunk_mr.end(), "In case this property changes.");
284 if ((last_block == chunk_mr.end()) // our last block does not straddle boundary
285 || !sp->block_is_obj(last_block) // last_block isn't an object
286 || oop(last_block)->is_objArray() // last_block is an array (precisely marked)
287 || oop(last_block)->is_typeArray()) {
288 max_to_do = chunk_mr.end();
289 NOISY(tty->print_cr(" process_chunk_boundary: Last block on this card is not a non-array object;\n"
290 " max_to_do left at " PTR_FORMAT, max_to_do);)
291 } else {
292 assert(last_block < chunk_mr.end(), "Tautology");
293 // It is a non-array object that straddles the right boundary of this chunk.
294 // last_obj_card is the card corresponding to the start of the last object
295 // in the chunk. Note that the last object may not start in
296 // the chunk.
297 jbyte* const last_obj_card = byte_for(last_block);
298 const jbyte val = *last_obj_card;
299 if (!card_will_be_scanned(val)) {
300 assert(!card_may_have_been_dirty(val), "Error");
301 // The card containing the head is not dirty. Any marks on
302 // subsequent cards still in this chunk must have been made
303 // precisely; we can cap processing at the end of our chunk.
304 max_to_do = chunk_mr.end();
305 NOISY(tty->print_cr(" process_chunk_boundary: Head of last object on this card is not dirty;\n"
306 " max_to_do left at " PTR_FORMAT,
307 max_to_do);)
308 } else {
309 // The last object must be considered dirty, and extends onto the
310 // following chunk. Look for a dirty card in that chunk that will
311 // bound our processing.
312 jbyte* limit_card = NULL;
313 const size_t last_block_size = sp->block_size(last_block);
314 jbyte* const last_card_of_last_obj =
315 byte_for(last_block + last_block_size - 1);
316 jbyte* const first_card_of_next_chunk = byte_for(chunk_mr.end());
317 // This search potentially goes a long distance looking
318 // for the next card that will be scanned, terminating
319 // at the end of the last_block, if no earlier dirty card
320 // is found.
321 assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == ParGCCardsPerStrideChunk,
322 "last card of next chunk may be wrong");
323 for (jbyte* cur = first_card_of_next_chunk;
324 cur <= last_card_of_last_obj; cur++) {
325 const jbyte val = *cur;
326 if (card_will_be_scanned(val)) {
327 NOISY(tty->print_cr(" Found a non-clean card " PTR_FORMAT " with value 0x%x",
328 cur, (int)val);)
329 limit_card = cur; break;
330 } else {
331 assert(!card_may_have_been_dirty(val), "Error: card can't be skipped");
332 }
333 }
334 if (limit_card != NULL) {
335 max_to_do = addr_for(limit_card);
336 assert(limit_card != NULL && max_to_do != NULL, "Error");
337 NOISY(tty->print_cr(" process_chunk_boundary: Found a dirty card at " PTR_FORMAT
338 " max_to_do set at " PTR_FORMAT " which is before end of last block in chunk: "
339 PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
340 limit_card, max_to_do, last_block, last_block_size, (last_block+last_block_size));)
341 } else {
342 // The following is a pessimistic value, because it's possible
343 // that a dirty card on a subsequent chunk has been cleared by
344 // the time we get to look at it; we'll correct for that further below,
345 // using the LNC array which records the least non-clean card
346 // before cards were cleared in a particular chunk.
347 limit_card = last_card_of_last_obj;
348 max_to_do = last_block + last_block_size;
349 assert(limit_card != NULL && max_to_do != NULL, "Error");
350 NOISY(tty->print_cr(" process_chunk_boundary: Found no dirty card before end of last block in chunk\n"
351 " Setting limit_card to " PTR_FORMAT
352 " and max_to_do " PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
353 limit_card, last_block, last_block_size, max_to_do);)
354 }
355 assert(0 < cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size,
356 "Bounds error.");
357 // It is possible that a dirty card for the last object may have been
358 // cleared before we had a chance to examine it. In that case, the value
359 // will have been logged in the LNC for that chunk.
360 // We need to examine as many chunks to the right as this object
361 // covers. However, we need to bound this checking to the largest
362 // entry in the LNC array: this is because the heap may expand
363 // after the LNC array has been created but before we reach this point,
364 // and the last block in our chunk may have been expanded to include
365 // the expansion delta (and possibly subsequently allocated from, so
366 // it wouldn't be sufficient to check whether that last block was
367 // or was not an object at this point).
368 uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1)
369 - lowest_non_clean_base_chunk_index;
370 const uintptr_t last_chunk_index = addr_to_chunk_index(used.last())
371 - lowest_non_clean_base_chunk_index;
372 if (last_chunk_index_to_check > last_chunk_index) {
373 assert(last_block + last_block_size > used.end(),
374 err_msg("Inconsistency detected: last_block [" PTR_FORMAT "," PTR_FORMAT "]"
375 " does not exceed used.end() = " PTR_FORMAT ","
376 " yet last_chunk_index_to_check " INTPTR_FORMAT
377 " exceeds last_chunk_index " INTPTR_FORMAT,
378 last_block, last_block + last_block_size,
379 used.end(),
380 last_chunk_index_to_check, last_chunk_index));
381 assert(sp->used_region().end() > used.end(),
382 err_msg("Expansion did not happen: "
383 "[" PTR_FORMAT "," PTR_FORMAT ") -> [" PTR_FORMAT "," PTR_FORMAT ")",
384 sp->used_region().start(), sp->used_region().end(), used.start(), used.end()));
385 NOISY(tty->print_cr(" process_chunk_boundary: heap expanded; explicitly bounding last_chunk");)
386 last_chunk_index_to_check = last_chunk_index;
387 }
388 for (uintptr_t lnc_index = cur_chunk_index + 1;
389 lnc_index <= last_chunk_index_to_check;
390 lnc_index++) {
391 jbyte* lnc_card = lowest_non_clean[lnc_index];
392 if (lnc_card != NULL) {
393 // we can stop at the first non-NULL entry we find
394 if (lnc_card <= limit_card) {
395 NOISY(tty->print_cr(" process_chunk_boundary: LNC card " PTR_FORMAT " is lower than limit_card " PTR_FORMAT,
396 " max_to_do will be lowered to " PTR_FORMAT " from " PTR_FORMAT,
397 lnc_card, limit_card, addr_for(lnc_card), max_to_do);)
398 limit_card = lnc_card;
399 max_to_do = addr_for(limit_card);
400 assert(limit_card != NULL && max_to_do != NULL, "Error");
401 }
402 // In any case, we break now
403 break;
404 } // else continue to look for a non-NULL entry if any
405 }
406 assert(limit_card != NULL && max_to_do != NULL, "Error");
407 }
408 assert(max_to_do != NULL, "OOPS 1 !");
409 }
410 assert(max_to_do != NULL, "OOPS 2!");
411 } else {
412 max_to_do = used.end();
413 NOISY(tty->print_cr(" process_chunk_boundary: Last chunk of this space;\n"
414 " max_to_do left at " PTR_FORMAT,
415 max_to_do);)
416 }
417 assert(max_to_do != NULL, "OOPS 3!");
418 // Now we can set the closure we're using so it doesn't to beyond
419 // max_to_do.
420 dcto_cl->set_min_done(max_to_do);
421 #ifndef PRODUCT
422 dcto_cl->set_last_bottom(max_to_do);
423 #endif
424 NOISY(tty->print_cr("===========================================================================\n");)
425 }
427 #undef NOISY
429 void
430 CardTableModRefBS::
431 get_LNC_array_for_space(Space* sp,
432 jbyte**& lowest_non_clean,
433 uintptr_t& lowest_non_clean_base_chunk_index,
434 size_t& lowest_non_clean_chunk_size) {
436 int i = find_covering_region_containing(sp->bottom());
437 MemRegion covered = _covered[i];
438 size_t n_chunks = chunks_to_cover(covered);
440 // Only the first thread to obtain the lock will resize the
441 // LNC array for the covered region. Any later expansion can't affect
442 // the used_at_save_marks region.
443 // (I observed a bug in which the first thread to execute this would
444 // resize, and then it would cause "expand_and_allocate" that would
445 // increase the number of chunks in the covered region. Then a second
446 // thread would come and execute this, see that the size didn't match,
447 // and free and allocate again. So the first thread would be using a
448 // freed "_lowest_non_clean" array.)
450 // Do a dirty read here. If we pass the conditional then take the rare
451 // event lock and do the read again in case some other thread had already
452 // succeeded and done the resize.
453 int cur_collection = Universe::heap()->total_collections();
454 if (_last_LNC_resizing_collection[i] != cur_collection) {
455 MutexLocker x(ParGCRareEvent_lock);
456 if (_last_LNC_resizing_collection[i] != cur_collection) {
457 if (_lowest_non_clean[i] == NULL ||
458 n_chunks != _lowest_non_clean_chunk_size[i]) {
460 // Should we delete the old?
461 if (_lowest_non_clean[i] != NULL) {
462 assert(n_chunks != _lowest_non_clean_chunk_size[i],
463 "logical consequence");
464 FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i], mtGC);
465 _lowest_non_clean[i] = NULL;
466 }
467 // Now allocate a new one if necessary.
468 if (_lowest_non_clean[i] == NULL) {
469 _lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks, mtGC);
470 _lowest_non_clean_chunk_size[i] = n_chunks;
471 _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());
472 for (int j = 0; j < (int)n_chunks; j++)
473 _lowest_non_clean[i][j] = NULL;
474 }
475 }
476 _last_LNC_resizing_collection[i] = cur_collection;
477 }
478 }
479 // In any case, now do the initialization.
480 lowest_non_clean = _lowest_non_clean[i];
481 lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];
482 lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i];
483 }