Tue, 29 Apr 2014 15:17:27 +0200
8042195: Introduce umbrella header orderAccess.inline.hpp.
Reviewed-by: dholmes, kvn, stefank, twisti
1 /*
2 * Copyright (c) 2001, 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 "gc_implementation/concurrentMarkSweep/cmsLockVerifier.hpp"
27 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
28 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
29 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
30 #include "gc_implementation/shared/liveRange.hpp"
31 #include "gc_implementation/shared/spaceDecorator.hpp"
32 #include "gc_interface/collectedHeap.inline.hpp"
33 #include "memory/allocation.inline.hpp"
34 #include "memory/blockOffsetTable.inline.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "memory/universe.inline.hpp"
37 #include "oops/oop.inline.hpp"
38 #include "runtime/globals.hpp"
39 #include "runtime/handles.inline.hpp"
40 #include "runtime/init.hpp"
41 #include "runtime/java.hpp"
42 #include "runtime/orderAccess.inline.hpp"
43 #include "runtime/vmThread.hpp"
44 #include "utilities/copy.hpp"
46 /////////////////////////////////////////////////////////////////////////
47 //// CompactibleFreeListSpace
48 /////////////////////////////////////////////////////////////////////////
50 // highest ranked free list lock rank
51 int CompactibleFreeListSpace::_lockRank = Mutex::leaf + 3;
53 // Defaults are 0 so things will break badly if incorrectly initialized.
54 size_t CompactibleFreeListSpace::IndexSetStart = 0;
55 size_t CompactibleFreeListSpace::IndexSetStride = 0;
57 size_t MinChunkSize = 0;
59 void CompactibleFreeListSpace::set_cms_values() {
60 // Set CMS global values
61 assert(MinChunkSize == 0, "already set");
63 // MinChunkSize should be a multiple of MinObjAlignment and be large enough
64 // for chunks to contain a FreeChunk.
65 size_t min_chunk_size_in_bytes = align_size_up(sizeof(FreeChunk), MinObjAlignmentInBytes);
66 MinChunkSize = min_chunk_size_in_bytes / BytesPerWord;
68 assert(IndexSetStart == 0 && IndexSetStride == 0, "already set");
69 IndexSetStart = MinChunkSize;
70 IndexSetStride = MinObjAlignment;
71 }
73 // Constructor
74 CompactibleFreeListSpace::CompactibleFreeListSpace(BlockOffsetSharedArray* bs,
75 MemRegion mr, bool use_adaptive_freelists,
76 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
77 _dictionaryChoice(dictionaryChoice),
78 _adaptive_freelists(use_adaptive_freelists),
79 _bt(bs, mr),
80 // free list locks are in the range of values taken by _lockRank
81 // This range currently is [_leaf+2, _leaf+3]
82 // Note: this requires that CFLspace c'tors
83 // are called serially in the order in which the locks are
84 // are acquired in the program text. This is true today.
85 _freelistLock(_lockRank--, "CompactibleFreeListSpace._lock", true),
86 _parDictionaryAllocLock(Mutex::leaf - 1, // == rank(ExpandHeap_lock) - 1
87 "CompactibleFreeListSpace._dict_par_lock", true),
88 _rescan_task_size(CardTableModRefBS::card_size_in_words * BitsPerWord *
89 CMSRescanMultiple),
90 _marking_task_size(CardTableModRefBS::card_size_in_words * BitsPerWord *
91 CMSConcMarkMultiple),
92 _collector(NULL)
93 {
94 assert(sizeof(FreeChunk) / BytesPerWord <= MinChunkSize,
95 "FreeChunk is larger than expected");
96 _bt.set_space(this);
97 initialize(mr, SpaceDecorator::Clear, SpaceDecorator::Mangle);
98 // We have all of "mr", all of which we place in the dictionary
99 // as one big chunk. We'll need to decide here which of several
100 // possible alternative dictionary implementations to use. For
101 // now the choice is easy, since we have only one working
102 // implementation, namely, the simple binary tree (splaying
103 // temporarily disabled).
104 switch (dictionaryChoice) {
105 case FreeBlockDictionary<FreeChunk>::dictionaryBinaryTree:
106 _dictionary = new AFLBinaryTreeDictionary(mr);
107 break;
108 case FreeBlockDictionary<FreeChunk>::dictionarySplayTree:
109 case FreeBlockDictionary<FreeChunk>::dictionarySkipList:
110 default:
111 warning("dictionaryChoice: selected option not understood; using"
112 " default BinaryTreeDictionary implementation instead.");
113 }
114 assert(_dictionary != NULL, "CMS dictionary initialization");
115 // The indexed free lists are initially all empty and are lazily
116 // filled in on demand. Initialize the array elements to NULL.
117 initializeIndexedFreeListArray();
119 // Not using adaptive free lists assumes that allocation is first
120 // from the linAB's. Also a cms perm gen which can be compacted
121 // has to have the klass's klassKlass allocated at a lower
122 // address in the heap than the klass so that the klassKlass is
123 // moved to its new location before the klass is moved.
124 // Set the _refillSize for the linear allocation blocks
125 if (!use_adaptive_freelists) {
126 FreeChunk* fc = _dictionary->get_chunk(mr.word_size(),
127 FreeBlockDictionary<FreeChunk>::atLeast);
128 // The small linAB initially has all the space and will allocate
129 // a chunk of any size.
130 HeapWord* addr = (HeapWord*) fc;
131 _smallLinearAllocBlock.set(addr, fc->size() ,
132 1024*SmallForLinearAlloc, fc->size());
133 // Note that _unallocated_block is not updated here.
134 // Allocations from the linear allocation block should
135 // update it.
136 } else {
137 _smallLinearAllocBlock.set(0, 0, 1024*SmallForLinearAlloc,
138 SmallForLinearAlloc);
139 }
140 // CMSIndexedFreeListReplenish should be at least 1
141 CMSIndexedFreeListReplenish = MAX2((uintx)1, CMSIndexedFreeListReplenish);
142 _promoInfo.setSpace(this);
143 if (UseCMSBestFit) {
144 _fitStrategy = FreeBlockBestFitFirst;
145 } else {
146 _fitStrategy = FreeBlockStrategyNone;
147 }
148 check_free_list_consistency();
150 // Initialize locks for parallel case.
152 if (CollectedHeap::use_parallel_gc_threads()) {
153 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
154 _indexedFreeListParLocks[i] = new Mutex(Mutex::leaf - 1, // == ExpandHeap_lock - 1
155 "a freelist par lock",
156 true);
157 DEBUG_ONLY(
158 _indexedFreeList[i].set_protecting_lock(_indexedFreeListParLocks[i]);
159 )
160 }
161 _dictionary->set_par_lock(&_parDictionaryAllocLock);
162 }
163 }
165 // Like CompactibleSpace forward() but always calls cross_threshold() to
166 // update the block offset table. Removed initialize_threshold call because
167 // CFLS does not use a block offset array for contiguous spaces.
168 HeapWord* CompactibleFreeListSpace::forward(oop q, size_t size,
169 CompactPoint* cp, HeapWord* compact_top) {
170 // q is alive
171 // First check if we should switch compaction space
172 assert(this == cp->space, "'this' should be current compaction space.");
173 size_t compaction_max_size = pointer_delta(end(), compact_top);
174 assert(adjustObjectSize(size) == cp->space->adjust_object_size_v(size),
175 "virtual adjustObjectSize_v() method is not correct");
176 size_t adjusted_size = adjustObjectSize(size);
177 assert(compaction_max_size >= MinChunkSize || compaction_max_size == 0,
178 "no small fragments allowed");
179 assert(minimum_free_block_size() == MinChunkSize,
180 "for de-virtualized reference below");
181 // Can't leave a nonzero size, residual fragment smaller than MinChunkSize
182 if (adjusted_size + MinChunkSize > compaction_max_size &&
183 adjusted_size != compaction_max_size) {
184 do {
185 // switch to next compaction space
186 cp->space->set_compaction_top(compact_top);
187 cp->space = cp->space->next_compaction_space();
188 if (cp->space == NULL) {
189 cp->gen = GenCollectedHeap::heap()->prev_gen(cp->gen);
190 assert(cp->gen != NULL, "compaction must succeed");
191 cp->space = cp->gen->first_compaction_space();
192 assert(cp->space != NULL, "generation must have a first compaction space");
193 }
194 compact_top = cp->space->bottom();
195 cp->space->set_compaction_top(compact_top);
196 // The correct adjusted_size may not be the same as that for this method
197 // (i.e., cp->space may no longer be "this" so adjust the size again.
198 // Use the virtual method which is not used above to save the virtual
199 // dispatch.
200 adjusted_size = cp->space->adjust_object_size_v(size);
201 compaction_max_size = pointer_delta(cp->space->end(), compact_top);
202 assert(cp->space->minimum_free_block_size() == 0, "just checking");
203 } while (adjusted_size > compaction_max_size);
204 }
206 // store the forwarding pointer into the mark word
207 if ((HeapWord*)q != compact_top) {
208 q->forward_to(oop(compact_top));
209 assert(q->is_gc_marked(), "encoding the pointer should preserve the mark");
210 } else {
211 // if the object isn't moving we can just set the mark to the default
212 // mark and handle it specially later on.
213 q->init_mark();
214 assert(q->forwardee() == NULL, "should be forwarded to NULL");
215 }
217 compact_top += adjusted_size;
219 // we need to update the offset table so that the beginnings of objects can be
220 // found during scavenge. Note that we are updating the offset table based on
221 // where the object will be once the compaction phase finishes.
223 // Always call cross_threshold(). A contiguous space can only call it when
224 // the compaction_top exceeds the current threshold but not for an
225 // non-contiguous space.
226 cp->threshold =
227 cp->space->cross_threshold(compact_top - adjusted_size, compact_top);
228 return compact_top;
229 }
231 // A modified copy of OffsetTableContigSpace::cross_threshold() with _offsets -> _bt
232 // and use of single_block instead of alloc_block. The name here is not really
233 // appropriate - maybe a more general name could be invented for both the
234 // contiguous and noncontiguous spaces.
236 HeapWord* CompactibleFreeListSpace::cross_threshold(HeapWord* start, HeapWord* the_end) {
237 _bt.single_block(start, the_end);
238 return end();
239 }
241 // Initialize them to NULL.
242 void CompactibleFreeListSpace::initializeIndexedFreeListArray() {
243 for (size_t i = 0; i < IndexSetSize; i++) {
244 // Note that on platforms where objects are double word aligned,
245 // the odd array elements are not used. It is convenient, however,
246 // to map directly from the object size to the array element.
247 _indexedFreeList[i].reset(IndexSetSize);
248 _indexedFreeList[i].set_size(i);
249 assert(_indexedFreeList[i].count() == 0, "reset check failed");
250 assert(_indexedFreeList[i].head() == NULL, "reset check failed");
251 assert(_indexedFreeList[i].tail() == NULL, "reset check failed");
252 assert(_indexedFreeList[i].hint() == IndexSetSize, "reset check failed");
253 }
254 }
256 void CompactibleFreeListSpace::resetIndexedFreeListArray() {
257 for (size_t i = 1; i < IndexSetSize; i++) {
258 assert(_indexedFreeList[i].size() == (size_t) i,
259 "Indexed free list sizes are incorrect");
260 _indexedFreeList[i].reset(IndexSetSize);
261 assert(_indexedFreeList[i].count() == 0, "reset check failed");
262 assert(_indexedFreeList[i].head() == NULL, "reset check failed");
263 assert(_indexedFreeList[i].tail() == NULL, "reset check failed");
264 assert(_indexedFreeList[i].hint() == IndexSetSize, "reset check failed");
265 }
266 }
268 void CompactibleFreeListSpace::reset(MemRegion mr) {
269 resetIndexedFreeListArray();
270 dictionary()->reset();
271 if (BlockOffsetArrayUseUnallocatedBlock) {
272 assert(end() == mr.end(), "We are compacting to the bottom of CMS gen");
273 // Everything's allocated until proven otherwise.
274 _bt.set_unallocated_block(end());
275 }
276 if (!mr.is_empty()) {
277 assert(mr.word_size() >= MinChunkSize, "Chunk size is too small");
278 _bt.single_block(mr.start(), mr.word_size());
279 FreeChunk* fc = (FreeChunk*) mr.start();
280 fc->set_size(mr.word_size());
281 if (mr.word_size() >= IndexSetSize ) {
282 returnChunkToDictionary(fc);
283 } else {
284 _bt.verify_not_unallocated((HeapWord*)fc, fc->size());
285 _indexedFreeList[mr.word_size()].return_chunk_at_head(fc);
286 }
287 coalBirth(mr.word_size());
288 }
289 _promoInfo.reset();
290 _smallLinearAllocBlock._ptr = NULL;
291 _smallLinearAllocBlock._word_size = 0;
292 }
294 void CompactibleFreeListSpace::reset_after_compaction() {
295 // Reset the space to the new reality - one free chunk.
296 MemRegion mr(compaction_top(), end());
297 reset(mr);
298 // Now refill the linear allocation block(s) if possible.
299 if (_adaptive_freelists) {
300 refillLinearAllocBlocksIfNeeded();
301 } else {
302 // Place as much of mr in the linAB as we can get,
303 // provided it was big enough to go into the dictionary.
304 FreeChunk* fc = dictionary()->find_largest_dict();
305 if (fc != NULL) {
306 assert(fc->size() == mr.word_size(),
307 "Why was the chunk broken up?");
308 removeChunkFromDictionary(fc);
309 HeapWord* addr = (HeapWord*) fc;
310 _smallLinearAllocBlock.set(addr, fc->size() ,
311 1024*SmallForLinearAlloc, fc->size());
312 // Note that _unallocated_block is not updated here.
313 }
314 }
315 }
317 // Walks the entire dictionary, returning a coterminal
318 // chunk, if it exists. Use with caution since it involves
319 // a potentially complete walk of a potentially large tree.
320 FreeChunk* CompactibleFreeListSpace::find_chunk_at_end() {
322 assert_lock_strong(&_freelistLock);
324 return dictionary()->find_chunk_ends_at(end());
325 }
328 #ifndef PRODUCT
329 void CompactibleFreeListSpace::initializeIndexedFreeListArrayReturnedBytes() {
330 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
331 _indexedFreeList[i].allocation_stats()->set_returned_bytes(0);
332 }
333 }
335 size_t CompactibleFreeListSpace::sumIndexedFreeListArrayReturnedBytes() {
336 size_t sum = 0;
337 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
338 sum += _indexedFreeList[i].allocation_stats()->returned_bytes();
339 }
340 return sum;
341 }
343 size_t CompactibleFreeListSpace::totalCountInIndexedFreeLists() const {
344 size_t count = 0;
345 for (size_t i = IndexSetStart; i < IndexSetSize; i++) {
346 debug_only(
347 ssize_t total_list_count = 0;
348 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL;
349 fc = fc->next()) {
350 total_list_count++;
351 }
352 assert(total_list_count == _indexedFreeList[i].count(),
353 "Count in list is incorrect");
354 )
355 count += _indexedFreeList[i].count();
356 }
357 return count;
358 }
360 size_t CompactibleFreeListSpace::totalCount() {
361 size_t num = totalCountInIndexedFreeLists();
362 num += dictionary()->total_count();
363 if (_smallLinearAllocBlock._word_size != 0) {
364 num++;
365 }
366 return num;
367 }
368 #endif
370 bool CompactibleFreeListSpace::is_free_block(const HeapWord* p) const {
371 FreeChunk* fc = (FreeChunk*) p;
372 return fc->is_free();
373 }
375 size_t CompactibleFreeListSpace::used() const {
376 return capacity() - free();
377 }
379 size_t CompactibleFreeListSpace::free() const {
380 // "MT-safe, but not MT-precise"(TM), if you will: i.e.
381 // if you do this while the structures are in flux you
382 // may get an approximate answer only; for instance
383 // because there is concurrent allocation either
384 // directly by mutators or for promotion during a GC.
385 // It's "MT-safe", however, in the sense that you are guaranteed
386 // not to crash and burn, for instance, because of walking
387 // pointers that could disappear as you were walking them.
388 // The approximation is because the various components
389 // that are read below are not read atomically (and
390 // further the computation of totalSizeInIndexedFreeLists()
391 // is itself a non-atomic computation. The normal use of
392 // this is during a resize operation at the end of GC
393 // and at that time you are guaranteed to get the
394 // correct actual value. However, for instance, this is
395 // also read completely asynchronously by the "perf-sampler"
396 // that supports jvmstat, and you are apt to see the values
397 // flicker in such cases.
398 assert(_dictionary != NULL, "No _dictionary?");
399 return (_dictionary->total_chunk_size(DEBUG_ONLY(freelistLock())) +
400 totalSizeInIndexedFreeLists() +
401 _smallLinearAllocBlock._word_size) * HeapWordSize;
402 }
404 size_t CompactibleFreeListSpace::max_alloc_in_words() const {
405 assert(_dictionary != NULL, "No _dictionary?");
406 assert_locked();
407 size_t res = _dictionary->max_chunk_size();
408 res = MAX2(res, MIN2(_smallLinearAllocBlock._word_size,
409 (size_t) SmallForLinearAlloc - 1));
410 // XXX the following could potentially be pretty slow;
411 // should one, pesimally for the rare cases when res
412 // caclulated above is less than IndexSetSize,
413 // just return res calculated above? My reasoning was that
414 // those cases will be so rare that the extra time spent doesn't
415 // really matter....
416 // Note: do not change the loop test i >= res + IndexSetStride
417 // to i > res below, because i is unsigned and res may be zero.
418 for (size_t i = IndexSetSize - 1; i >= res + IndexSetStride;
419 i -= IndexSetStride) {
420 if (_indexedFreeList[i].head() != NULL) {
421 assert(_indexedFreeList[i].count() != 0, "Inconsistent FreeList");
422 return i;
423 }
424 }
425 return res;
426 }
428 void LinearAllocBlock::print_on(outputStream* st) const {
429 st->print_cr(" LinearAllocBlock: ptr = " PTR_FORMAT ", word_size = " SIZE_FORMAT
430 ", refillsize = " SIZE_FORMAT ", allocation_size_limit = " SIZE_FORMAT,
431 p2i(_ptr), _word_size, _refillSize, _allocation_size_limit);
432 }
434 void CompactibleFreeListSpace::print_on(outputStream* st) const {
435 st->print_cr("COMPACTIBLE FREELIST SPACE");
436 st->print_cr(" Space:");
437 Space::print_on(st);
439 st->print_cr("promoInfo:");
440 _promoInfo.print_on(st);
442 st->print_cr("_smallLinearAllocBlock");
443 _smallLinearAllocBlock.print_on(st);
445 // dump_memory_block(_smallLinearAllocBlock->_ptr, 128);
447 st->print_cr(" _fitStrategy = %s, _adaptive_freelists = %s",
448 _fitStrategy?"true":"false", _adaptive_freelists?"true":"false");
449 }
451 void CompactibleFreeListSpace::print_indexed_free_lists(outputStream* st)
452 const {
453 reportIndexedFreeListStatistics();
454 gclog_or_tty->print_cr("Layout of Indexed Freelists");
455 gclog_or_tty->print_cr("---------------------------");
456 AdaptiveFreeList<FreeChunk>::print_labels_on(st, "size");
457 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
458 _indexedFreeList[i].print_on(gclog_or_tty);
459 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL;
460 fc = fc->next()) {
461 gclog_or_tty->print_cr("\t[" PTR_FORMAT "," PTR_FORMAT ") %s",
462 p2i(fc), p2i((HeapWord*)fc + i),
463 fc->cantCoalesce() ? "\t CC" : "");
464 }
465 }
466 }
468 void CompactibleFreeListSpace::print_promo_info_blocks(outputStream* st)
469 const {
470 _promoInfo.print_on(st);
471 }
473 void CompactibleFreeListSpace::print_dictionary_free_lists(outputStream* st)
474 const {
475 _dictionary->report_statistics();
476 st->print_cr("Layout of Freelists in Tree");
477 st->print_cr("---------------------------");
478 _dictionary->print_free_lists(st);
479 }
481 class BlkPrintingClosure: public BlkClosure {
482 const CMSCollector* _collector;
483 const CompactibleFreeListSpace* _sp;
484 const CMSBitMap* _live_bit_map;
485 const bool _post_remark;
486 outputStream* _st;
487 public:
488 BlkPrintingClosure(const CMSCollector* collector,
489 const CompactibleFreeListSpace* sp,
490 const CMSBitMap* live_bit_map,
491 outputStream* st):
492 _collector(collector),
493 _sp(sp),
494 _live_bit_map(live_bit_map),
495 _post_remark(collector->abstract_state() > CMSCollector::FinalMarking),
496 _st(st) { }
497 size_t do_blk(HeapWord* addr);
498 };
500 size_t BlkPrintingClosure::do_blk(HeapWord* addr) {
501 size_t sz = _sp->block_size_no_stall(addr, _collector);
502 assert(sz != 0, "Should always be able to compute a size");
503 if (_sp->block_is_obj(addr)) {
504 const bool dead = _post_remark && !_live_bit_map->isMarked(addr);
505 _st->print_cr(PTR_FORMAT ": %s object of size " SIZE_FORMAT "%s",
506 p2i(addr),
507 dead ? "dead" : "live",
508 sz,
509 (!dead && CMSPrintObjectsInDump) ? ":" : ".");
510 if (CMSPrintObjectsInDump && !dead) {
511 oop(addr)->print_on(_st);
512 _st->print_cr("--------------------------------------");
513 }
514 } else { // free block
515 _st->print_cr(PTR_FORMAT ": free block of size " SIZE_FORMAT "%s",
516 p2i(addr), sz, CMSPrintChunksInDump ? ":" : ".");
517 if (CMSPrintChunksInDump) {
518 ((FreeChunk*)addr)->print_on(_st);
519 _st->print_cr("--------------------------------------");
520 }
521 }
522 return sz;
523 }
525 void CompactibleFreeListSpace::dump_at_safepoint_with_locks(CMSCollector* c,
526 outputStream* st) {
527 st->print_cr("\n=========================");
528 st->print_cr("Block layout in CMS Heap:");
529 st->print_cr("=========================");
530 BlkPrintingClosure bpcl(c, this, c->markBitMap(), st);
531 blk_iterate(&bpcl);
533 st->print_cr("\n=======================================");
534 st->print_cr("Order & Layout of Promotion Info Blocks");
535 st->print_cr("=======================================");
536 print_promo_info_blocks(st);
538 st->print_cr("\n===========================");
539 st->print_cr("Order of Indexed Free Lists");
540 st->print_cr("=========================");
541 print_indexed_free_lists(st);
543 st->print_cr("\n=================================");
544 st->print_cr("Order of Free Lists in Dictionary");
545 st->print_cr("=================================");
546 print_dictionary_free_lists(st);
547 }
550 void CompactibleFreeListSpace::reportFreeListStatistics() const {
551 assert_lock_strong(&_freelistLock);
552 assert(PrintFLSStatistics != 0, "Reporting error");
553 _dictionary->report_statistics();
554 if (PrintFLSStatistics > 1) {
555 reportIndexedFreeListStatistics();
556 size_t total_size = totalSizeInIndexedFreeLists() +
557 _dictionary->total_chunk_size(DEBUG_ONLY(freelistLock()));
558 gclog_or_tty->print(" free=" SIZE_FORMAT " frag=%1.4f\n", total_size, flsFrag());
559 }
560 }
562 void CompactibleFreeListSpace::reportIndexedFreeListStatistics() const {
563 assert_lock_strong(&_freelistLock);
564 gclog_or_tty->print("Statistics for IndexedFreeLists:\n"
565 "--------------------------------\n");
566 size_t total_size = totalSizeInIndexedFreeLists();
567 size_t free_blocks = numFreeBlocksInIndexedFreeLists();
568 gclog_or_tty->print("Total Free Space: " SIZE_FORMAT "\n", total_size);
569 gclog_or_tty->print("Max Chunk Size: " SIZE_FORMAT "\n", maxChunkSizeInIndexedFreeLists());
570 gclog_or_tty->print("Number of Blocks: " SIZE_FORMAT "\n", free_blocks);
571 if (free_blocks != 0) {
572 gclog_or_tty->print("Av. Block Size: " SIZE_FORMAT "\n", total_size/free_blocks);
573 }
574 }
576 size_t CompactibleFreeListSpace::numFreeBlocksInIndexedFreeLists() const {
577 size_t res = 0;
578 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
579 debug_only(
580 ssize_t recount = 0;
581 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL;
582 fc = fc->next()) {
583 recount += 1;
584 }
585 assert(recount == _indexedFreeList[i].count(),
586 "Incorrect count in list");
587 )
588 res += _indexedFreeList[i].count();
589 }
590 return res;
591 }
593 size_t CompactibleFreeListSpace::maxChunkSizeInIndexedFreeLists() const {
594 for (size_t i = IndexSetSize - 1; i != 0; i -= IndexSetStride) {
595 if (_indexedFreeList[i].head() != NULL) {
596 assert(_indexedFreeList[i].count() != 0, "Inconsistent FreeList");
597 return (size_t)i;
598 }
599 }
600 return 0;
601 }
603 void CompactibleFreeListSpace::set_end(HeapWord* value) {
604 HeapWord* prevEnd = end();
605 assert(prevEnd != value, "unnecessary set_end call");
606 assert(prevEnd == NULL || !BlockOffsetArrayUseUnallocatedBlock || value >= unallocated_block(),
607 "New end is below unallocated block");
608 _end = value;
609 if (prevEnd != NULL) {
610 // Resize the underlying block offset table.
611 _bt.resize(pointer_delta(value, bottom()));
612 if (value <= prevEnd) {
613 assert(!BlockOffsetArrayUseUnallocatedBlock || value >= unallocated_block(),
614 "New end is below unallocated block");
615 } else {
616 // Now, take this new chunk and add it to the free blocks.
617 // Note that the BOT has not yet been updated for this block.
618 size_t newFcSize = pointer_delta(value, prevEnd);
619 // XXX This is REALLY UGLY and should be fixed up. XXX
620 if (!_adaptive_freelists && _smallLinearAllocBlock._ptr == NULL) {
621 // Mark the boundary of the new block in BOT
622 _bt.mark_block(prevEnd, value);
623 // put it all in the linAB
624 if (ParallelGCThreads == 0) {
625 _smallLinearAllocBlock._ptr = prevEnd;
626 _smallLinearAllocBlock._word_size = newFcSize;
627 repairLinearAllocBlock(&_smallLinearAllocBlock);
628 } else { // ParallelGCThreads > 0
629 MutexLockerEx x(parDictionaryAllocLock(),
630 Mutex::_no_safepoint_check_flag);
631 _smallLinearAllocBlock._ptr = prevEnd;
632 _smallLinearAllocBlock._word_size = newFcSize;
633 repairLinearAllocBlock(&_smallLinearAllocBlock);
634 }
635 // Births of chunks put into a LinAB are not recorded. Births
636 // of chunks as they are allocated out of a LinAB are.
637 } else {
638 // Add the block to the free lists, if possible coalescing it
639 // with the last free block, and update the BOT and census data.
640 addChunkToFreeListsAtEndRecordingStats(prevEnd, newFcSize);
641 }
642 }
643 }
644 }
646 class FreeListSpace_DCTOC : public Filtering_DCTOC {
647 CompactibleFreeListSpace* _cfls;
648 CMSCollector* _collector;
649 protected:
650 // Override.
651 #define walk_mem_region_with_cl_DECL(ClosureType) \
652 virtual void walk_mem_region_with_cl(MemRegion mr, \
653 HeapWord* bottom, HeapWord* top, \
654 ClosureType* cl); \
655 void walk_mem_region_with_cl_par(MemRegion mr, \
656 HeapWord* bottom, HeapWord* top, \
657 ClosureType* cl); \
658 void walk_mem_region_with_cl_nopar(MemRegion mr, \
659 HeapWord* bottom, HeapWord* top, \
660 ClosureType* cl)
661 walk_mem_region_with_cl_DECL(ExtendedOopClosure);
662 walk_mem_region_with_cl_DECL(FilteringClosure);
664 public:
665 FreeListSpace_DCTOC(CompactibleFreeListSpace* sp,
666 CMSCollector* collector,
667 ExtendedOopClosure* cl,
668 CardTableModRefBS::PrecisionStyle precision,
669 HeapWord* boundary) :
670 Filtering_DCTOC(sp, cl, precision, boundary),
671 _cfls(sp), _collector(collector) {}
672 };
674 // We de-virtualize the block-related calls below, since we know that our
675 // space is a CompactibleFreeListSpace.
677 #define FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(ClosureType) \
678 void FreeListSpace_DCTOC::walk_mem_region_with_cl(MemRegion mr, \
679 HeapWord* bottom, \
680 HeapWord* top, \
681 ClosureType* cl) { \
682 bool is_par = SharedHeap::heap()->n_par_threads() > 0; \
683 if (is_par) { \
684 assert(SharedHeap::heap()->n_par_threads() == \
685 SharedHeap::heap()->workers()->active_workers(), "Mismatch"); \
686 walk_mem_region_with_cl_par(mr, bottom, top, cl); \
687 } else { \
688 walk_mem_region_with_cl_nopar(mr, bottom, top, cl); \
689 } \
690 } \
691 void FreeListSpace_DCTOC::walk_mem_region_with_cl_par(MemRegion mr, \
692 HeapWord* bottom, \
693 HeapWord* top, \
694 ClosureType* cl) { \
695 /* Skip parts that are before "mr", in case "block_start" sent us \
696 back too far. */ \
697 HeapWord* mr_start = mr.start(); \
698 size_t bot_size = _cfls->CompactibleFreeListSpace::block_size(bottom); \
699 HeapWord* next = bottom + bot_size; \
700 while (next < mr_start) { \
701 bottom = next; \
702 bot_size = _cfls->CompactibleFreeListSpace::block_size(bottom); \
703 next = bottom + bot_size; \
704 } \
705 \
706 while (bottom < top) { \
707 if (_cfls->CompactibleFreeListSpace::block_is_obj(bottom) && \
708 !_cfls->CompactibleFreeListSpace::obj_allocated_since_save_marks( \
709 oop(bottom)) && \
710 !_collector->CMSCollector::is_dead_obj(oop(bottom))) { \
711 size_t word_sz = oop(bottom)->oop_iterate(cl, mr); \
712 bottom += _cfls->adjustObjectSize(word_sz); \
713 } else { \
714 bottom += _cfls->CompactibleFreeListSpace::block_size(bottom); \
715 } \
716 } \
717 } \
718 void FreeListSpace_DCTOC::walk_mem_region_with_cl_nopar(MemRegion mr, \
719 HeapWord* bottom, \
720 HeapWord* top, \
721 ClosureType* cl) { \
722 /* Skip parts that are before "mr", in case "block_start" sent us \
723 back too far. */ \
724 HeapWord* mr_start = mr.start(); \
725 size_t bot_size = _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \
726 HeapWord* next = bottom + bot_size; \
727 while (next < mr_start) { \
728 bottom = next; \
729 bot_size = _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \
730 next = bottom + bot_size; \
731 } \
732 \
733 while (bottom < top) { \
734 if (_cfls->CompactibleFreeListSpace::block_is_obj_nopar(bottom) && \
735 !_cfls->CompactibleFreeListSpace::obj_allocated_since_save_marks( \
736 oop(bottom)) && \
737 !_collector->CMSCollector::is_dead_obj(oop(bottom))) { \
738 size_t word_sz = oop(bottom)->oop_iterate(cl, mr); \
739 bottom += _cfls->adjustObjectSize(word_sz); \
740 } else { \
741 bottom += _cfls->CompactibleFreeListSpace::block_size_nopar(bottom); \
742 } \
743 } \
744 }
746 // (There are only two of these, rather than N, because the split is due
747 // only to the introduction of the FilteringClosure, a local part of the
748 // impl of this abstraction.)
749 FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(ExtendedOopClosure)
750 FreeListSpace_DCTOC__walk_mem_region_with_cl_DEFN(FilteringClosure)
752 DirtyCardToOopClosure*
753 CompactibleFreeListSpace::new_dcto_cl(ExtendedOopClosure* cl,
754 CardTableModRefBS::PrecisionStyle precision,
755 HeapWord* boundary) {
756 return new FreeListSpace_DCTOC(this, _collector, cl, precision, boundary);
757 }
760 // Note on locking for the space iteration functions:
761 // since the collector's iteration activities are concurrent with
762 // allocation activities by mutators, absent a suitable mutual exclusion
763 // mechanism the iterators may go awry. For instace a block being iterated
764 // may suddenly be allocated or divided up and part of it allocated and
765 // so on.
767 // Apply the given closure to each block in the space.
768 void CompactibleFreeListSpace::blk_iterate_careful(BlkClosureCareful* cl) {
769 assert_lock_strong(freelistLock());
770 HeapWord *cur, *limit;
771 for (cur = bottom(), limit = end(); cur < limit;
772 cur += cl->do_blk_careful(cur));
773 }
775 // Apply the given closure to each block in the space.
776 void CompactibleFreeListSpace::blk_iterate(BlkClosure* cl) {
777 assert_lock_strong(freelistLock());
778 HeapWord *cur, *limit;
779 for (cur = bottom(), limit = end(); cur < limit;
780 cur += cl->do_blk(cur));
781 }
783 // Apply the given closure to each oop in the space.
784 void CompactibleFreeListSpace::oop_iterate(ExtendedOopClosure* cl) {
785 assert_lock_strong(freelistLock());
786 HeapWord *cur, *limit;
787 size_t curSize;
788 for (cur = bottom(), limit = end(); cur < limit;
789 cur += curSize) {
790 curSize = block_size(cur);
791 if (block_is_obj(cur)) {
792 oop(cur)->oop_iterate(cl);
793 }
794 }
795 }
797 // Apply the given closure to each oop in the space \intersect memory region.
798 void CompactibleFreeListSpace::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
799 assert_lock_strong(freelistLock());
800 if (is_empty()) {
801 return;
802 }
803 MemRegion cur = MemRegion(bottom(), end());
804 mr = mr.intersection(cur);
805 if (mr.is_empty()) {
806 return;
807 }
808 if (mr.equals(cur)) {
809 oop_iterate(cl);
810 return;
811 }
812 assert(mr.end() <= end(), "just took an intersection above");
813 HeapWord* obj_addr = block_start(mr.start());
814 HeapWord* t = mr.end();
816 SpaceMemRegionOopsIterClosure smr_blk(cl, mr);
817 if (block_is_obj(obj_addr)) {
818 // Handle first object specially.
819 oop obj = oop(obj_addr);
820 obj_addr += adjustObjectSize(obj->oop_iterate(&smr_blk));
821 } else {
822 FreeChunk* fc = (FreeChunk*)obj_addr;
823 obj_addr += fc->size();
824 }
825 while (obj_addr < t) {
826 HeapWord* obj = obj_addr;
827 obj_addr += block_size(obj_addr);
828 // If "obj_addr" is not greater than top, then the
829 // entire object "obj" is within the region.
830 if (obj_addr <= t) {
831 if (block_is_obj(obj)) {
832 oop(obj)->oop_iterate(cl);
833 }
834 } else {
835 // "obj" extends beyond end of region
836 if (block_is_obj(obj)) {
837 oop(obj)->oop_iterate(&smr_blk);
838 }
839 break;
840 }
841 }
842 }
844 // NOTE: In the following methods, in order to safely be able to
845 // apply the closure to an object, we need to be sure that the
846 // object has been initialized. We are guaranteed that an object
847 // is initialized if we are holding the Heap_lock with the
848 // world stopped.
849 void CompactibleFreeListSpace::verify_objects_initialized() const {
850 if (is_init_completed()) {
851 assert_locked_or_safepoint(Heap_lock);
852 if (Universe::is_fully_initialized()) {
853 guarantee(SafepointSynchronize::is_at_safepoint(),
854 "Required for objects to be initialized");
855 }
856 } // else make a concession at vm start-up
857 }
859 // Apply the given closure to each object in the space
860 void CompactibleFreeListSpace::object_iterate(ObjectClosure* blk) {
861 assert_lock_strong(freelistLock());
862 NOT_PRODUCT(verify_objects_initialized());
863 HeapWord *cur, *limit;
864 size_t curSize;
865 for (cur = bottom(), limit = end(); cur < limit;
866 cur += curSize) {
867 curSize = block_size(cur);
868 if (block_is_obj(cur)) {
869 blk->do_object(oop(cur));
870 }
871 }
872 }
874 // Apply the given closure to each live object in the space
875 // The usage of CompactibleFreeListSpace
876 // by the ConcurrentMarkSweepGeneration for concurrent GC's allows
877 // objects in the space with references to objects that are no longer
878 // valid. For example, an object may reference another object
879 // that has already been sweep up (collected). This method uses
880 // obj_is_alive() to determine whether it is safe to apply the closure to
881 // an object. See obj_is_alive() for details on how liveness of an
882 // object is decided.
884 void CompactibleFreeListSpace::safe_object_iterate(ObjectClosure* blk) {
885 assert_lock_strong(freelistLock());
886 NOT_PRODUCT(verify_objects_initialized());
887 HeapWord *cur, *limit;
888 size_t curSize;
889 for (cur = bottom(), limit = end(); cur < limit;
890 cur += curSize) {
891 curSize = block_size(cur);
892 if (block_is_obj(cur) && obj_is_alive(cur)) {
893 blk->do_object(oop(cur));
894 }
895 }
896 }
898 void CompactibleFreeListSpace::object_iterate_mem(MemRegion mr,
899 UpwardsObjectClosure* cl) {
900 assert_locked(freelistLock());
901 NOT_PRODUCT(verify_objects_initialized());
902 Space::object_iterate_mem(mr, cl);
903 }
905 // Callers of this iterator beware: The closure application should
906 // be robust in the face of uninitialized objects and should (always)
907 // return a correct size so that the next addr + size below gives us a
908 // valid block boundary. [See for instance,
909 // ScanMarkedObjectsAgainCarefullyClosure::do_object_careful()
910 // in ConcurrentMarkSweepGeneration.cpp.]
911 HeapWord*
912 CompactibleFreeListSpace::object_iterate_careful(ObjectClosureCareful* cl) {
913 assert_lock_strong(freelistLock());
914 HeapWord *addr, *last;
915 size_t size;
916 for (addr = bottom(), last = end();
917 addr < last; addr += size) {
918 FreeChunk* fc = (FreeChunk*)addr;
919 if (fc->is_free()) {
920 // Since we hold the free list lock, which protects direct
921 // allocation in this generation by mutators, a free object
922 // will remain free throughout this iteration code.
923 size = fc->size();
924 } else {
925 // Note that the object need not necessarily be initialized,
926 // because (for instance) the free list lock does NOT protect
927 // object initialization. The closure application below must
928 // therefore be correct in the face of uninitialized objects.
929 size = cl->do_object_careful(oop(addr));
930 if (size == 0) {
931 // An unparsable object found. Signal early termination.
932 return addr;
933 }
934 }
935 }
936 return NULL;
937 }
939 // Callers of this iterator beware: The closure application should
940 // be robust in the face of uninitialized objects and should (always)
941 // return a correct size so that the next addr + size below gives us a
942 // valid block boundary. [See for instance,
943 // ScanMarkedObjectsAgainCarefullyClosure::do_object_careful()
944 // in ConcurrentMarkSweepGeneration.cpp.]
945 HeapWord*
946 CompactibleFreeListSpace::object_iterate_careful_m(MemRegion mr,
947 ObjectClosureCareful* cl) {
948 assert_lock_strong(freelistLock());
949 // Can't use used_region() below because it may not necessarily
950 // be the same as [bottom(),end()); although we could
951 // use [used_region().start(),round_to(used_region().end(),CardSize)),
952 // that appears too cumbersome, so we just do the simpler check
953 // in the assertion below.
954 assert(!mr.is_empty() && MemRegion(bottom(),end()).contains(mr),
955 "mr should be non-empty and within used space");
956 HeapWord *addr, *end;
957 size_t size;
958 for (addr = block_start_careful(mr.start()), end = mr.end();
959 addr < end; addr += size) {
960 FreeChunk* fc = (FreeChunk*)addr;
961 if (fc->is_free()) {
962 // Since we hold the free list lock, which protects direct
963 // allocation in this generation by mutators, a free object
964 // will remain free throughout this iteration code.
965 size = fc->size();
966 } else {
967 // Note that the object need not necessarily be initialized,
968 // because (for instance) the free list lock does NOT protect
969 // object initialization. The closure application below must
970 // therefore be correct in the face of uninitialized objects.
971 size = cl->do_object_careful_m(oop(addr), mr);
972 if (size == 0) {
973 // An unparsable object found. Signal early termination.
974 return addr;
975 }
976 }
977 }
978 return NULL;
979 }
982 HeapWord* CompactibleFreeListSpace::block_start_const(const void* p) const {
983 NOT_PRODUCT(verify_objects_initialized());
984 return _bt.block_start(p);
985 }
987 HeapWord* CompactibleFreeListSpace::block_start_careful(const void* p) const {
988 return _bt.block_start_careful(p);
989 }
991 size_t CompactibleFreeListSpace::block_size(const HeapWord* p) const {
992 NOT_PRODUCT(verify_objects_initialized());
993 // This must be volatile, or else there is a danger that the compiler
994 // will compile the code below into a sometimes-infinite loop, by keeping
995 // the value read the first time in a register.
996 while (true) {
997 // We must do this until we get a consistent view of the object.
998 if (FreeChunk::indicatesFreeChunk(p)) {
999 volatile FreeChunk* fc = (volatile FreeChunk*)p;
1000 size_t res = fc->size();
1002 // Bugfix for systems with weak memory model (PPC64/IA64). The
1003 // block's free bit was set and we have read the size of the
1004 // block. Acquire and check the free bit again. If the block is
1005 // still free, the read size is correct.
1006 OrderAccess::acquire();
1008 // If the object is still a free chunk, return the size, else it
1009 // has been allocated so try again.
1010 if (FreeChunk::indicatesFreeChunk(p)) {
1011 assert(res != 0, "Block size should not be 0");
1012 return res;
1013 }
1014 } else {
1015 // must read from what 'p' points to in each loop.
1016 Klass* k = ((volatile oopDesc*)p)->klass_or_null();
1017 if (k != NULL) {
1018 assert(k->is_klass(), "Should really be klass oop.");
1019 oop o = (oop)p;
1020 assert(o->is_oop(true /* ignore mark word */), "Should be an oop.");
1022 // Bugfix for systems with weak memory model (PPC64/IA64).
1023 // The object o may be an array. Acquire to make sure that the array
1024 // size (third word) is consistent.
1025 OrderAccess::acquire();
1027 size_t res = o->size_given_klass(k);
1028 res = adjustObjectSize(res);
1029 assert(res != 0, "Block size should not be 0");
1030 return res;
1031 }
1032 }
1033 }
1034 }
1036 // TODO: Now that is_parsable is gone, we should combine these two functions.
1037 // A variant of the above that uses the Printezis bits for
1038 // unparsable but allocated objects. This avoids any possible
1039 // stalls waiting for mutators to initialize objects, and is
1040 // thus potentially faster than the variant above. However,
1041 // this variant may return a zero size for a block that is
1042 // under mutation and for which a consistent size cannot be
1043 // inferred without stalling; see CMSCollector::block_size_if_printezis_bits().
1044 size_t CompactibleFreeListSpace::block_size_no_stall(HeapWord* p,
1045 const CMSCollector* c)
1046 const {
1047 assert(MemRegion(bottom(), end()).contains(p), "p not in space");
1048 // This must be volatile, or else there is a danger that the compiler
1049 // will compile the code below into a sometimes-infinite loop, by keeping
1050 // the value read the first time in a register.
1051 DEBUG_ONLY(uint loops = 0;)
1052 while (true) {
1053 // We must do this until we get a consistent view of the object.
1054 if (FreeChunk::indicatesFreeChunk(p)) {
1055 volatile FreeChunk* fc = (volatile FreeChunk*)p;
1056 size_t res = fc->size();
1058 // Bugfix for systems with weak memory model (PPC64/IA64). The
1059 // free bit of the block was set and we have read the size of
1060 // the block. Acquire and check the free bit again. If the
1061 // block is still free, the read size is correct.
1062 OrderAccess::acquire();
1064 if (FreeChunk::indicatesFreeChunk(p)) {
1065 assert(res != 0, "Block size should not be 0");
1066 assert(loops == 0, "Should be 0");
1067 return res;
1068 }
1069 } else {
1070 // must read from what 'p' points to in each loop.
1071 Klass* k = ((volatile oopDesc*)p)->klass_or_null();
1072 // We trust the size of any object that has a non-NULL
1073 // klass and (for those in the perm gen) is parsable
1074 // -- irrespective of its conc_safe-ty.
1075 if (k != NULL) {
1076 assert(k->is_klass(), "Should really be klass oop.");
1077 oop o = (oop)p;
1078 assert(o->is_oop(), "Should be an oop");
1080 // Bugfix for systems with weak memory model (PPC64/IA64).
1081 // The object o may be an array. Acquire to make sure that the array
1082 // size (third word) is consistent.
1083 OrderAccess::acquire();
1085 size_t res = o->size_given_klass(k);
1086 res = adjustObjectSize(res);
1087 assert(res != 0, "Block size should not be 0");
1088 return res;
1089 } else {
1090 // May return 0 if P-bits not present.
1091 return c->block_size_if_printezis_bits(p);
1092 }
1093 }
1094 assert(loops == 0, "Can loop at most once");
1095 DEBUG_ONLY(loops++;)
1096 }
1097 }
1099 size_t CompactibleFreeListSpace::block_size_nopar(const HeapWord* p) const {
1100 NOT_PRODUCT(verify_objects_initialized());
1101 assert(MemRegion(bottom(), end()).contains(p), "p not in space");
1102 FreeChunk* fc = (FreeChunk*)p;
1103 if (fc->is_free()) {
1104 return fc->size();
1105 } else {
1106 // Ignore mark word because this may be a recently promoted
1107 // object whose mark word is used to chain together grey
1108 // objects (the last one would have a null value).
1109 assert(oop(p)->is_oop(true), "Should be an oop");
1110 return adjustObjectSize(oop(p)->size());
1111 }
1112 }
1114 // This implementation assumes that the property of "being an object" is
1115 // stable. But being a free chunk may not be (because of parallel
1116 // promotion.)
1117 bool CompactibleFreeListSpace::block_is_obj(const HeapWord* p) const {
1118 FreeChunk* fc = (FreeChunk*)p;
1119 assert(is_in_reserved(p), "Should be in space");
1120 // When doing a mark-sweep-compact of the CMS generation, this
1121 // assertion may fail because prepare_for_compaction() uses
1122 // space that is garbage to maintain information on ranges of
1123 // live objects so that these live ranges can be moved as a whole.
1124 // Comment out this assertion until that problem can be solved
1125 // (i.e., that the block start calculation may look at objects
1126 // at address below "p" in finding the object that contains "p"
1127 // and those objects (if garbage) may have been modified to hold
1128 // live range information.
1129 // assert(CollectedHeap::use_parallel_gc_threads() || _bt.block_start(p) == p,
1130 // "Should be a block boundary");
1131 if (FreeChunk::indicatesFreeChunk(p)) return false;
1132 Klass* k = oop(p)->klass_or_null();
1133 if (k != NULL) {
1134 // Ignore mark word because it may have been used to
1135 // chain together promoted objects (the last one
1136 // would have a null value).
1137 assert(oop(p)->is_oop(true), "Should be an oop");
1138 return true;
1139 } else {
1140 return false; // Was not an object at the start of collection.
1141 }
1142 }
1144 // Check if the object is alive. This fact is checked either by consulting
1145 // the main marking bitmap in the sweeping phase or, if it's a permanent
1146 // generation and we're not in the sweeping phase, by checking the
1147 // perm_gen_verify_bit_map where we store the "deadness" information if
1148 // we did not sweep the perm gen in the most recent previous GC cycle.
1149 bool CompactibleFreeListSpace::obj_is_alive(const HeapWord* p) const {
1150 assert(SafepointSynchronize::is_at_safepoint() || !is_init_completed(),
1151 "Else races are possible");
1152 assert(block_is_obj(p), "The address should point to an object");
1154 // If we're sweeping, we use object liveness information from the main bit map
1155 // for both perm gen and old gen.
1156 // We don't need to lock the bitmap (live_map or dead_map below), because
1157 // EITHER we are in the middle of the sweeping phase, and the
1158 // main marking bit map (live_map below) is locked,
1159 // OR we're in other phases and perm_gen_verify_bit_map (dead_map below)
1160 // is stable, because it's mutated only in the sweeping phase.
1161 // NOTE: This method is also used by jmap where, if class unloading is
1162 // off, the results can return "false" for legitimate perm objects,
1163 // when we are not in the midst of a sweeping phase, which can result
1164 // in jmap not reporting certain perm gen objects. This will be moot
1165 // if/when the perm gen goes away in the future.
1166 if (_collector->abstract_state() == CMSCollector::Sweeping) {
1167 CMSBitMap* live_map = _collector->markBitMap();
1168 return live_map->par_isMarked((HeapWord*) p);
1169 }
1170 return true;
1171 }
1173 bool CompactibleFreeListSpace::block_is_obj_nopar(const HeapWord* p) const {
1174 FreeChunk* fc = (FreeChunk*)p;
1175 assert(is_in_reserved(p), "Should be in space");
1176 assert(_bt.block_start(p) == p, "Should be a block boundary");
1177 if (!fc->is_free()) {
1178 // Ignore mark word because it may have been used to
1179 // chain together promoted objects (the last one
1180 // would have a null value).
1181 assert(oop(p)->is_oop(true), "Should be an oop");
1182 return true;
1183 }
1184 return false;
1185 }
1187 // "MT-safe but not guaranteed MT-precise" (TM); you may get an
1188 // approximate answer if you don't hold the freelistlock when you call this.
1189 size_t CompactibleFreeListSpace::totalSizeInIndexedFreeLists() const {
1190 size_t size = 0;
1191 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
1192 debug_only(
1193 // We may be calling here without the lock in which case we
1194 // won't do this modest sanity check.
1195 if (freelistLock()->owned_by_self()) {
1196 size_t total_list_size = 0;
1197 for (FreeChunk* fc = _indexedFreeList[i].head(); fc != NULL;
1198 fc = fc->next()) {
1199 total_list_size += i;
1200 }
1201 assert(total_list_size == i * _indexedFreeList[i].count(),
1202 "Count in list is incorrect");
1203 }
1204 )
1205 size += i * _indexedFreeList[i].count();
1206 }
1207 return size;
1208 }
1210 HeapWord* CompactibleFreeListSpace::par_allocate(size_t size) {
1211 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
1212 return allocate(size);
1213 }
1215 HeapWord*
1216 CompactibleFreeListSpace::getChunkFromSmallLinearAllocBlockRemainder(size_t size) {
1217 return getChunkFromLinearAllocBlockRemainder(&_smallLinearAllocBlock, size);
1218 }
1220 HeapWord* CompactibleFreeListSpace::allocate(size_t size) {
1221 assert_lock_strong(freelistLock());
1222 HeapWord* res = NULL;
1223 assert(size == adjustObjectSize(size),
1224 "use adjustObjectSize() before calling into allocate()");
1226 if (_adaptive_freelists) {
1227 res = allocate_adaptive_freelists(size);
1228 } else { // non-adaptive free lists
1229 res = allocate_non_adaptive_freelists(size);
1230 }
1232 if (res != NULL) {
1233 // check that res does lie in this space!
1234 assert(is_in_reserved(res), "Not in this space!");
1235 assert(is_aligned((void*)res), "alignment check");
1237 FreeChunk* fc = (FreeChunk*)res;
1238 fc->markNotFree();
1239 assert(!fc->is_free(), "shouldn't be marked free");
1240 assert(oop(fc)->klass_or_null() == NULL, "should look uninitialized");
1241 // Verify that the block offset table shows this to
1242 // be a single block, but not one which is unallocated.
1243 _bt.verify_single_block(res, size);
1244 _bt.verify_not_unallocated(res, size);
1245 // mangle a just allocated object with a distinct pattern.
1246 debug_only(fc->mangleAllocated(size));
1247 }
1249 return res;
1250 }
1252 HeapWord* CompactibleFreeListSpace::allocate_non_adaptive_freelists(size_t size) {
1253 HeapWord* res = NULL;
1254 // try and use linear allocation for smaller blocks
1255 if (size < _smallLinearAllocBlock._allocation_size_limit) {
1256 // if successful, the following also adjusts block offset table
1257 res = getChunkFromSmallLinearAllocBlock(size);
1258 }
1259 // Else triage to indexed lists for smaller sizes
1260 if (res == NULL) {
1261 if (size < SmallForDictionary) {
1262 res = (HeapWord*) getChunkFromIndexedFreeList(size);
1263 } else {
1264 // else get it from the big dictionary; if even this doesn't
1265 // work we are out of luck.
1266 res = (HeapWord*)getChunkFromDictionaryExact(size);
1267 }
1268 }
1270 return res;
1271 }
1273 HeapWord* CompactibleFreeListSpace::allocate_adaptive_freelists(size_t size) {
1274 assert_lock_strong(freelistLock());
1275 HeapWord* res = NULL;
1276 assert(size == adjustObjectSize(size),
1277 "use adjustObjectSize() before calling into allocate()");
1279 // Strategy
1280 // if small
1281 // exact size from small object indexed list if small
1282 // small or large linear allocation block (linAB) as appropriate
1283 // take from lists of greater sized chunks
1284 // else
1285 // dictionary
1286 // small or large linear allocation block if it has the space
1287 // Try allocating exact size from indexTable first
1288 if (size < IndexSetSize) {
1289 res = (HeapWord*) getChunkFromIndexedFreeList(size);
1290 if(res != NULL) {
1291 assert(res != (HeapWord*)_indexedFreeList[size].head(),
1292 "Not removed from free list");
1293 // no block offset table adjustment is necessary on blocks in
1294 // the indexed lists.
1296 // Try allocating from the small LinAB
1297 } else if (size < _smallLinearAllocBlock._allocation_size_limit &&
1298 (res = getChunkFromSmallLinearAllocBlock(size)) != NULL) {
1299 // if successful, the above also adjusts block offset table
1300 // Note that this call will refill the LinAB to
1301 // satisfy the request. This is different that
1302 // evm.
1303 // Don't record chunk off a LinAB? smallSplitBirth(size);
1304 } else {
1305 // Raid the exact free lists larger than size, even if they are not
1306 // overpopulated.
1307 res = (HeapWord*) getChunkFromGreater(size);
1308 }
1309 } else {
1310 // Big objects get allocated directly from the dictionary.
1311 res = (HeapWord*) getChunkFromDictionaryExact(size);
1312 if (res == NULL) {
1313 // Try hard not to fail since an allocation failure will likely
1314 // trigger a synchronous GC. Try to get the space from the
1315 // allocation blocks.
1316 res = getChunkFromSmallLinearAllocBlockRemainder(size);
1317 }
1318 }
1320 return res;
1321 }
1323 // A worst-case estimate of the space required (in HeapWords) to expand the heap
1324 // when promoting obj.
1325 size_t CompactibleFreeListSpace::expansionSpaceRequired(size_t obj_size) const {
1326 // Depending on the object size, expansion may require refilling either a
1327 // bigLAB or a smallLAB plus refilling a PromotionInfo object. MinChunkSize
1328 // is added because the dictionary may over-allocate to avoid fragmentation.
1329 size_t space = obj_size;
1330 if (!_adaptive_freelists) {
1331 space = MAX2(space, _smallLinearAllocBlock._refillSize);
1332 }
1333 space += _promoInfo.refillSize() + 2 * MinChunkSize;
1334 return space;
1335 }
1337 FreeChunk* CompactibleFreeListSpace::getChunkFromGreater(size_t numWords) {
1338 FreeChunk* ret;
1340 assert(numWords >= MinChunkSize, "Size is less than minimum");
1341 assert(linearAllocationWouldFail() || bestFitFirst(),
1342 "Should not be here");
1344 size_t i;
1345 size_t currSize = numWords + MinChunkSize;
1346 assert(currSize % MinObjAlignment == 0, "currSize should be aligned");
1347 for (i = currSize; i < IndexSetSize; i += IndexSetStride) {
1348 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[i];
1349 if (fl->head()) {
1350 ret = getFromListGreater(fl, numWords);
1351 assert(ret == NULL || ret->is_free(), "Should be returning a free chunk");
1352 return ret;
1353 }
1354 }
1356 currSize = MAX2((size_t)SmallForDictionary,
1357 (size_t)(numWords + MinChunkSize));
1359 /* Try to get a chunk that satisfies request, while avoiding
1360 fragmentation that can't be handled. */
1361 {
1362 ret = dictionary()->get_chunk(currSize);
1363 if (ret != NULL) {
1364 assert(ret->size() - numWords >= MinChunkSize,
1365 "Chunk is too small");
1366 _bt.allocated((HeapWord*)ret, ret->size());
1367 /* Carve returned chunk. */
1368 (void) splitChunkAndReturnRemainder(ret, numWords);
1369 /* Label this as no longer a free chunk. */
1370 assert(ret->is_free(), "This chunk should be free");
1371 ret->link_prev(NULL);
1372 }
1373 assert(ret == NULL || ret->is_free(), "Should be returning a free chunk");
1374 return ret;
1375 }
1376 ShouldNotReachHere();
1377 }
1379 bool CompactibleFreeListSpace::verifyChunkInIndexedFreeLists(FreeChunk* fc) const {
1380 assert(fc->size() < IndexSetSize, "Size of chunk is too large");
1381 return _indexedFreeList[fc->size()].verify_chunk_in_free_list(fc);
1382 }
1384 bool CompactibleFreeListSpace::verify_chunk_is_linear_alloc_block(FreeChunk* fc) const {
1385 assert((_smallLinearAllocBlock._ptr != (HeapWord*)fc) ||
1386 (_smallLinearAllocBlock._word_size == fc->size()),
1387 "Linear allocation block shows incorrect size");
1388 return ((_smallLinearAllocBlock._ptr == (HeapWord*)fc) &&
1389 (_smallLinearAllocBlock._word_size == fc->size()));
1390 }
1392 // Check if the purported free chunk is present either as a linear
1393 // allocation block, the size-indexed table of (smaller) free blocks,
1394 // or the larger free blocks kept in the binary tree dictionary.
1395 bool CompactibleFreeListSpace::verify_chunk_in_free_list(FreeChunk* fc) const {
1396 if (verify_chunk_is_linear_alloc_block(fc)) {
1397 return true;
1398 } else if (fc->size() < IndexSetSize) {
1399 return verifyChunkInIndexedFreeLists(fc);
1400 } else {
1401 return dictionary()->verify_chunk_in_free_list(fc);
1402 }
1403 }
1405 #ifndef PRODUCT
1406 void CompactibleFreeListSpace::assert_locked() const {
1407 CMSLockVerifier::assert_locked(freelistLock(), parDictionaryAllocLock());
1408 }
1410 void CompactibleFreeListSpace::assert_locked(const Mutex* lock) const {
1411 CMSLockVerifier::assert_locked(lock);
1412 }
1413 #endif
1415 FreeChunk* CompactibleFreeListSpace::allocateScratch(size_t size) {
1416 // In the parallel case, the main thread holds the free list lock
1417 // on behalf the parallel threads.
1418 FreeChunk* fc;
1419 {
1420 // If GC is parallel, this might be called by several threads.
1421 // This should be rare enough that the locking overhead won't affect
1422 // the sequential code.
1423 MutexLockerEx x(parDictionaryAllocLock(),
1424 Mutex::_no_safepoint_check_flag);
1425 fc = getChunkFromDictionary(size);
1426 }
1427 if (fc != NULL) {
1428 fc->dontCoalesce();
1429 assert(fc->is_free(), "Should be free, but not coalescable");
1430 // Verify that the block offset table shows this to
1431 // be a single block, but not one which is unallocated.
1432 _bt.verify_single_block((HeapWord*)fc, fc->size());
1433 _bt.verify_not_unallocated((HeapWord*)fc, fc->size());
1434 }
1435 return fc;
1436 }
1438 oop CompactibleFreeListSpace::promote(oop obj, size_t obj_size) {
1439 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1440 assert_locked();
1442 // if we are tracking promotions, then first ensure space for
1443 // promotion (including spooling space for saving header if necessary).
1444 // then allocate and copy, then track promoted info if needed.
1445 // When tracking (see PromotionInfo::track()), the mark word may
1446 // be displaced and in this case restoration of the mark word
1447 // occurs in the (oop_since_save_marks_)iterate phase.
1448 if (_promoInfo.tracking() && !_promoInfo.ensure_spooling_space()) {
1449 return NULL;
1450 }
1451 // Call the allocate(size_t, bool) form directly to avoid the
1452 // additional call through the allocate(size_t) form. Having
1453 // the compile inline the call is problematic because allocate(size_t)
1454 // is a virtual method.
1455 HeapWord* res = allocate(adjustObjectSize(obj_size));
1456 if (res != NULL) {
1457 Copy::aligned_disjoint_words((HeapWord*)obj, res, obj_size);
1458 // if we should be tracking promotions, do so.
1459 if (_promoInfo.tracking()) {
1460 _promoInfo.track((PromotedObject*)res);
1461 }
1462 }
1463 return oop(res);
1464 }
1466 HeapWord*
1467 CompactibleFreeListSpace::getChunkFromSmallLinearAllocBlock(size_t size) {
1468 assert_locked();
1469 assert(size >= MinChunkSize, "minimum chunk size");
1470 assert(size < _smallLinearAllocBlock._allocation_size_limit,
1471 "maximum from smallLinearAllocBlock");
1472 return getChunkFromLinearAllocBlock(&_smallLinearAllocBlock, size);
1473 }
1475 HeapWord*
1476 CompactibleFreeListSpace::getChunkFromLinearAllocBlock(LinearAllocBlock *blk,
1477 size_t size) {
1478 assert_locked();
1479 assert(size >= MinChunkSize, "too small");
1480 HeapWord* res = NULL;
1481 // Try to do linear allocation from blk, making sure that
1482 if (blk->_word_size == 0) {
1483 // We have probably been unable to fill this either in the prologue or
1484 // when it was exhausted at the last linear allocation. Bail out until
1485 // next time.
1486 assert(blk->_ptr == NULL, "consistency check");
1487 return NULL;
1488 }
1489 assert(blk->_word_size != 0 && blk->_ptr != NULL, "consistency check");
1490 res = getChunkFromLinearAllocBlockRemainder(blk, size);
1491 if (res != NULL) return res;
1493 // about to exhaust this linear allocation block
1494 if (blk->_word_size == size) { // exactly satisfied
1495 res = blk->_ptr;
1496 _bt.allocated(res, blk->_word_size);
1497 } else if (size + MinChunkSize <= blk->_refillSize) {
1498 size_t sz = blk->_word_size;
1499 // Update _unallocated_block if the size is such that chunk would be
1500 // returned to the indexed free list. All other chunks in the indexed
1501 // free lists are allocated from the dictionary so that _unallocated_block
1502 // has already been adjusted for them. Do it here so that the cost
1503 // for all chunks added back to the indexed free lists.
1504 if (sz < SmallForDictionary) {
1505 _bt.allocated(blk->_ptr, sz);
1506 }
1507 // Return the chunk that isn't big enough, and then refill below.
1508 addChunkToFreeLists(blk->_ptr, sz);
1509 split_birth(sz);
1510 // Don't keep statistics on adding back chunk from a LinAB.
1511 } else {
1512 // A refilled block would not satisfy the request.
1513 return NULL;
1514 }
1516 blk->_ptr = NULL; blk->_word_size = 0;
1517 refillLinearAllocBlock(blk);
1518 assert(blk->_ptr == NULL || blk->_word_size >= size + MinChunkSize,
1519 "block was replenished");
1520 if (res != NULL) {
1521 split_birth(size);
1522 repairLinearAllocBlock(blk);
1523 } else if (blk->_ptr != NULL) {
1524 res = blk->_ptr;
1525 size_t blk_size = blk->_word_size;
1526 blk->_word_size -= size;
1527 blk->_ptr += size;
1528 split_birth(size);
1529 repairLinearAllocBlock(blk);
1530 // Update BOT last so that other (parallel) GC threads see a consistent
1531 // view of the BOT and free blocks.
1532 // Above must occur before BOT is updated below.
1533 OrderAccess::storestore();
1534 _bt.split_block(res, blk_size, size); // adjust block offset table
1535 }
1536 return res;
1537 }
1539 HeapWord* CompactibleFreeListSpace::getChunkFromLinearAllocBlockRemainder(
1540 LinearAllocBlock* blk,
1541 size_t size) {
1542 assert_locked();
1543 assert(size >= MinChunkSize, "too small");
1545 HeapWord* res = NULL;
1546 // This is the common case. Keep it simple.
1547 if (blk->_word_size >= size + MinChunkSize) {
1548 assert(blk->_ptr != NULL, "consistency check");
1549 res = blk->_ptr;
1550 // Note that the BOT is up-to-date for the linAB before allocation. It
1551 // indicates the start of the linAB. The split_block() updates the
1552 // BOT for the linAB after the allocation (indicates the start of the
1553 // next chunk to be allocated).
1554 size_t blk_size = blk->_word_size;
1555 blk->_word_size -= size;
1556 blk->_ptr += size;
1557 split_birth(size);
1558 repairLinearAllocBlock(blk);
1559 // Update BOT last so that other (parallel) GC threads see a consistent
1560 // view of the BOT and free blocks.
1561 // Above must occur before BOT is updated below.
1562 OrderAccess::storestore();
1563 _bt.split_block(res, blk_size, size); // adjust block offset table
1564 _bt.allocated(res, size);
1565 }
1566 return res;
1567 }
1569 FreeChunk*
1570 CompactibleFreeListSpace::getChunkFromIndexedFreeList(size_t size) {
1571 assert_locked();
1572 assert(size < SmallForDictionary, "just checking");
1573 FreeChunk* res;
1574 res = _indexedFreeList[size].get_chunk_at_head();
1575 if (res == NULL) {
1576 res = getChunkFromIndexedFreeListHelper(size);
1577 }
1578 _bt.verify_not_unallocated((HeapWord*) res, size);
1579 assert(res == NULL || res->size() == size, "Incorrect block size");
1580 return res;
1581 }
1583 FreeChunk*
1584 CompactibleFreeListSpace::getChunkFromIndexedFreeListHelper(size_t size,
1585 bool replenish) {
1586 assert_locked();
1587 FreeChunk* fc = NULL;
1588 if (size < SmallForDictionary) {
1589 assert(_indexedFreeList[size].head() == NULL ||
1590 _indexedFreeList[size].surplus() <= 0,
1591 "List for this size should be empty or under populated");
1592 // Try best fit in exact lists before replenishing the list
1593 if (!bestFitFirst() || (fc = bestFitSmall(size)) == NULL) {
1594 // Replenish list.
1595 //
1596 // Things tried that failed.
1597 // Tried allocating out of the two LinAB's first before
1598 // replenishing lists.
1599 // Tried small linAB of size 256 (size in indexed list)
1600 // and replenishing indexed lists from the small linAB.
1601 //
1602 FreeChunk* newFc = NULL;
1603 const size_t replenish_size = CMSIndexedFreeListReplenish * size;
1604 if (replenish_size < SmallForDictionary) {
1605 // Do not replenish from an underpopulated size.
1606 if (_indexedFreeList[replenish_size].surplus() > 0 &&
1607 _indexedFreeList[replenish_size].head() != NULL) {
1608 newFc = _indexedFreeList[replenish_size].get_chunk_at_head();
1609 } else if (bestFitFirst()) {
1610 newFc = bestFitSmall(replenish_size);
1611 }
1612 }
1613 if (newFc == NULL && replenish_size > size) {
1614 assert(CMSIndexedFreeListReplenish > 1, "ctl pt invariant");
1615 newFc = getChunkFromIndexedFreeListHelper(replenish_size, false);
1616 }
1617 // Note: The stats update re split-death of block obtained above
1618 // will be recorded below precisely when we know we are going to
1619 // be actually splitting it into more than one pieces below.
1620 if (newFc != NULL) {
1621 if (replenish || CMSReplenishIntermediate) {
1622 // Replenish this list and return one block to caller.
1623 size_t i;
1624 FreeChunk *curFc, *nextFc;
1625 size_t num_blk = newFc->size() / size;
1626 assert(num_blk >= 1, "Smaller than requested?");
1627 assert(newFc->size() % size == 0, "Should be integral multiple of request");
1628 if (num_blk > 1) {
1629 // we are sure we will be splitting the block just obtained
1630 // into multiple pieces; record the split-death of the original
1631 splitDeath(replenish_size);
1632 }
1633 // carve up and link blocks 0, ..., num_blk - 2
1634 // The last chunk is not added to the lists but is returned as the
1635 // free chunk.
1636 for (curFc = newFc, nextFc = (FreeChunk*)((HeapWord*)curFc + size),
1637 i = 0;
1638 i < (num_blk - 1);
1639 curFc = nextFc, nextFc = (FreeChunk*)((HeapWord*)nextFc + size),
1640 i++) {
1641 curFc->set_size(size);
1642 // Don't record this as a return in order to try and
1643 // determine the "returns" from a GC.
1644 _bt.verify_not_unallocated((HeapWord*) fc, size);
1645 _indexedFreeList[size].return_chunk_at_tail(curFc, false);
1646 _bt.mark_block((HeapWord*)curFc, size);
1647 split_birth(size);
1648 // Don't record the initial population of the indexed list
1649 // as a split birth.
1650 }
1652 // check that the arithmetic was OK above
1653 assert((HeapWord*)nextFc == (HeapWord*)newFc + num_blk*size,
1654 "inconsistency in carving newFc");
1655 curFc->set_size(size);
1656 _bt.mark_block((HeapWord*)curFc, size);
1657 split_birth(size);
1658 fc = curFc;
1659 } else {
1660 // Return entire block to caller
1661 fc = newFc;
1662 }
1663 }
1664 }
1665 } else {
1666 // Get a free chunk from the free chunk dictionary to be returned to
1667 // replenish the indexed free list.
1668 fc = getChunkFromDictionaryExact(size);
1669 }
1670 // assert(fc == NULL || fc->is_free(), "Should be returning a free chunk");
1671 return fc;
1672 }
1674 FreeChunk*
1675 CompactibleFreeListSpace::getChunkFromDictionary(size_t size) {
1676 assert_locked();
1677 FreeChunk* fc = _dictionary->get_chunk(size,
1678 FreeBlockDictionary<FreeChunk>::atLeast);
1679 if (fc == NULL) {
1680 return NULL;
1681 }
1682 _bt.allocated((HeapWord*)fc, fc->size());
1683 if (fc->size() >= size + MinChunkSize) {
1684 fc = splitChunkAndReturnRemainder(fc, size);
1685 }
1686 assert(fc->size() >= size, "chunk too small");
1687 assert(fc->size() < size + MinChunkSize, "chunk too big");
1688 _bt.verify_single_block((HeapWord*)fc, fc->size());
1689 return fc;
1690 }
1692 FreeChunk*
1693 CompactibleFreeListSpace::getChunkFromDictionaryExact(size_t size) {
1694 assert_locked();
1695 FreeChunk* fc = _dictionary->get_chunk(size,
1696 FreeBlockDictionary<FreeChunk>::atLeast);
1697 if (fc == NULL) {
1698 return fc;
1699 }
1700 _bt.allocated((HeapWord*)fc, fc->size());
1701 if (fc->size() == size) {
1702 _bt.verify_single_block((HeapWord*)fc, size);
1703 return fc;
1704 }
1705 assert(fc->size() > size, "get_chunk() guarantee");
1706 if (fc->size() < size + MinChunkSize) {
1707 // Return the chunk to the dictionary and go get a bigger one.
1708 returnChunkToDictionary(fc);
1709 fc = _dictionary->get_chunk(size + MinChunkSize,
1710 FreeBlockDictionary<FreeChunk>::atLeast);
1711 if (fc == NULL) {
1712 return NULL;
1713 }
1714 _bt.allocated((HeapWord*)fc, fc->size());
1715 }
1716 assert(fc->size() >= size + MinChunkSize, "tautology");
1717 fc = splitChunkAndReturnRemainder(fc, size);
1718 assert(fc->size() == size, "chunk is wrong size");
1719 _bt.verify_single_block((HeapWord*)fc, size);
1720 return fc;
1721 }
1723 void
1724 CompactibleFreeListSpace::returnChunkToDictionary(FreeChunk* chunk) {
1725 assert_locked();
1727 size_t size = chunk->size();
1728 _bt.verify_single_block((HeapWord*)chunk, size);
1729 // adjust _unallocated_block downward, as necessary
1730 _bt.freed((HeapWord*)chunk, size);
1731 _dictionary->return_chunk(chunk);
1732 #ifndef PRODUCT
1733 if (CMSCollector::abstract_state() != CMSCollector::Sweeping) {
1734 TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >* tc = TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >::as_TreeChunk(chunk);
1735 TreeList<FreeChunk, AdaptiveFreeList<FreeChunk> >* tl = tc->list();
1736 tl->verify_stats();
1737 }
1738 #endif // PRODUCT
1739 }
1741 void
1742 CompactibleFreeListSpace::returnChunkToFreeList(FreeChunk* fc) {
1743 assert_locked();
1744 size_t size = fc->size();
1745 _bt.verify_single_block((HeapWord*) fc, size);
1746 _bt.verify_not_unallocated((HeapWord*) fc, size);
1747 if (_adaptive_freelists) {
1748 _indexedFreeList[size].return_chunk_at_tail(fc);
1749 } else {
1750 _indexedFreeList[size].return_chunk_at_head(fc);
1751 }
1752 #ifndef PRODUCT
1753 if (CMSCollector::abstract_state() != CMSCollector::Sweeping) {
1754 _indexedFreeList[size].verify_stats();
1755 }
1756 #endif // PRODUCT
1757 }
1759 // Add chunk to end of last block -- if it's the largest
1760 // block -- and update BOT and census data. We would
1761 // of course have preferred to coalesce it with the
1762 // last block, but it's currently less expensive to find the
1763 // largest block than it is to find the last.
1764 void
1765 CompactibleFreeListSpace::addChunkToFreeListsAtEndRecordingStats(
1766 HeapWord* chunk, size_t size) {
1767 // check that the chunk does lie in this space!
1768 assert(chunk != NULL && is_in_reserved(chunk), "Not in this space!");
1769 // One of the parallel gc task threads may be here
1770 // whilst others are allocating.
1771 Mutex* lock = NULL;
1772 if (ParallelGCThreads != 0) {
1773 lock = &_parDictionaryAllocLock;
1774 }
1775 FreeChunk* ec;
1776 {
1777 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
1778 ec = dictionary()->find_largest_dict(); // get largest block
1779 if (ec != NULL && ec->end() == (uintptr_t*) chunk) {
1780 // It's a coterminal block - we can coalesce.
1781 size_t old_size = ec->size();
1782 coalDeath(old_size);
1783 removeChunkFromDictionary(ec);
1784 size += old_size;
1785 } else {
1786 ec = (FreeChunk*)chunk;
1787 }
1788 }
1789 ec->set_size(size);
1790 debug_only(ec->mangleFreed(size));
1791 if (size < SmallForDictionary && ParallelGCThreads != 0) {
1792 lock = _indexedFreeListParLocks[size];
1793 }
1794 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
1795 addChunkAndRepairOffsetTable((HeapWord*)ec, size, true);
1796 // record the birth under the lock since the recording involves
1797 // manipulation of the list on which the chunk lives and
1798 // if the chunk is allocated and is the last on the list,
1799 // the list can go away.
1800 coalBirth(size);
1801 }
1803 void
1804 CompactibleFreeListSpace::addChunkToFreeLists(HeapWord* chunk,
1805 size_t size) {
1806 // check that the chunk does lie in this space!
1807 assert(chunk != NULL && is_in_reserved(chunk), "Not in this space!");
1808 assert_locked();
1809 _bt.verify_single_block(chunk, size);
1811 FreeChunk* fc = (FreeChunk*) chunk;
1812 fc->set_size(size);
1813 debug_only(fc->mangleFreed(size));
1814 if (size < SmallForDictionary) {
1815 returnChunkToFreeList(fc);
1816 } else {
1817 returnChunkToDictionary(fc);
1818 }
1819 }
1821 void
1822 CompactibleFreeListSpace::addChunkAndRepairOffsetTable(HeapWord* chunk,
1823 size_t size, bool coalesced) {
1824 assert_locked();
1825 assert(chunk != NULL, "null chunk");
1826 if (coalesced) {
1827 // repair BOT
1828 _bt.single_block(chunk, size);
1829 }
1830 addChunkToFreeLists(chunk, size);
1831 }
1833 // We _must_ find the purported chunk on our free lists;
1834 // we assert if we don't.
1835 void
1836 CompactibleFreeListSpace::removeFreeChunkFromFreeLists(FreeChunk* fc) {
1837 size_t size = fc->size();
1838 assert_locked();
1839 debug_only(verifyFreeLists());
1840 if (size < SmallForDictionary) {
1841 removeChunkFromIndexedFreeList(fc);
1842 } else {
1843 removeChunkFromDictionary(fc);
1844 }
1845 _bt.verify_single_block((HeapWord*)fc, size);
1846 debug_only(verifyFreeLists());
1847 }
1849 void
1850 CompactibleFreeListSpace::removeChunkFromDictionary(FreeChunk* fc) {
1851 size_t size = fc->size();
1852 assert_locked();
1853 assert(fc != NULL, "null chunk");
1854 _bt.verify_single_block((HeapWord*)fc, size);
1855 _dictionary->remove_chunk(fc);
1856 // adjust _unallocated_block upward, as necessary
1857 _bt.allocated((HeapWord*)fc, size);
1858 }
1860 void
1861 CompactibleFreeListSpace::removeChunkFromIndexedFreeList(FreeChunk* fc) {
1862 assert_locked();
1863 size_t size = fc->size();
1864 _bt.verify_single_block((HeapWord*)fc, size);
1865 NOT_PRODUCT(
1866 if (FLSVerifyIndexTable) {
1867 verifyIndexedFreeList(size);
1868 }
1869 )
1870 _indexedFreeList[size].remove_chunk(fc);
1871 NOT_PRODUCT(
1872 if (FLSVerifyIndexTable) {
1873 verifyIndexedFreeList(size);
1874 }
1875 )
1876 }
1878 FreeChunk* CompactibleFreeListSpace::bestFitSmall(size_t numWords) {
1879 /* A hint is the next larger size that has a surplus.
1880 Start search at a size large enough to guarantee that
1881 the excess is >= MIN_CHUNK. */
1882 size_t start = align_object_size(numWords + MinChunkSize);
1883 if (start < IndexSetSize) {
1884 AdaptiveFreeList<FreeChunk>* it = _indexedFreeList;
1885 size_t hint = _indexedFreeList[start].hint();
1886 while (hint < IndexSetSize) {
1887 assert(hint % MinObjAlignment == 0, "hint should be aligned");
1888 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[hint];
1889 if (fl->surplus() > 0 && fl->head() != NULL) {
1890 // Found a list with surplus, reset original hint
1891 // and split out a free chunk which is returned.
1892 _indexedFreeList[start].set_hint(hint);
1893 FreeChunk* res = getFromListGreater(fl, numWords);
1894 assert(res == NULL || res->is_free(),
1895 "Should be returning a free chunk");
1896 return res;
1897 }
1898 hint = fl->hint(); /* keep looking */
1899 }
1900 /* None found. */
1901 it[start].set_hint(IndexSetSize);
1902 }
1903 return NULL;
1904 }
1906 /* Requires fl->size >= numWords + MinChunkSize */
1907 FreeChunk* CompactibleFreeListSpace::getFromListGreater(AdaptiveFreeList<FreeChunk>* fl,
1908 size_t numWords) {
1909 FreeChunk *curr = fl->head();
1910 size_t oldNumWords = curr->size();
1911 assert(numWords >= MinChunkSize, "Word size is too small");
1912 assert(curr != NULL, "List is empty");
1913 assert(oldNumWords >= numWords + MinChunkSize,
1914 "Size of chunks in the list is too small");
1916 fl->remove_chunk(curr);
1917 // recorded indirectly by splitChunkAndReturnRemainder -
1918 // smallSplit(oldNumWords, numWords);
1919 FreeChunk* new_chunk = splitChunkAndReturnRemainder(curr, numWords);
1920 // Does anything have to be done for the remainder in terms of
1921 // fixing the card table?
1922 assert(new_chunk == NULL || new_chunk->is_free(),
1923 "Should be returning a free chunk");
1924 return new_chunk;
1925 }
1927 FreeChunk*
1928 CompactibleFreeListSpace::splitChunkAndReturnRemainder(FreeChunk* chunk,
1929 size_t new_size) {
1930 assert_locked();
1931 size_t size = chunk->size();
1932 assert(size > new_size, "Split from a smaller block?");
1933 assert(is_aligned(chunk), "alignment problem");
1934 assert(size == adjustObjectSize(size), "alignment problem");
1935 size_t rem_size = size - new_size;
1936 assert(rem_size == adjustObjectSize(rem_size), "alignment problem");
1937 assert(rem_size >= MinChunkSize, "Free chunk smaller than minimum");
1938 FreeChunk* ffc = (FreeChunk*)((HeapWord*)chunk + new_size);
1939 assert(is_aligned(ffc), "alignment problem");
1940 ffc->set_size(rem_size);
1941 ffc->link_next(NULL);
1942 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads.
1943 // Above must occur before BOT is updated below.
1944 // adjust block offset table
1945 OrderAccess::storestore();
1946 assert(chunk->is_free() && ffc->is_free(), "Error");
1947 _bt.split_block((HeapWord*)chunk, chunk->size(), new_size);
1948 if (rem_size < SmallForDictionary) {
1949 bool is_par = (SharedHeap::heap()->n_par_threads() > 0);
1950 if (is_par) _indexedFreeListParLocks[rem_size]->lock();
1951 assert(!is_par ||
1952 (SharedHeap::heap()->n_par_threads() ==
1953 SharedHeap::heap()->workers()->active_workers()), "Mismatch");
1954 returnChunkToFreeList(ffc);
1955 split(size, rem_size);
1956 if (is_par) _indexedFreeListParLocks[rem_size]->unlock();
1957 } else {
1958 returnChunkToDictionary(ffc);
1959 split(size ,rem_size);
1960 }
1961 chunk->set_size(new_size);
1962 return chunk;
1963 }
1965 void
1966 CompactibleFreeListSpace::sweep_completed() {
1967 // Now that space is probably plentiful, refill linear
1968 // allocation blocks as needed.
1969 refillLinearAllocBlocksIfNeeded();
1970 }
1972 void
1973 CompactibleFreeListSpace::gc_prologue() {
1974 assert_locked();
1975 if (PrintFLSStatistics != 0) {
1976 gclog_or_tty->print("Before GC:\n");
1977 reportFreeListStatistics();
1978 }
1979 refillLinearAllocBlocksIfNeeded();
1980 }
1982 void
1983 CompactibleFreeListSpace::gc_epilogue() {
1984 assert_locked();
1985 if (PrintGCDetails && Verbose && !_adaptive_freelists) {
1986 if (_smallLinearAllocBlock._word_size == 0)
1987 warning("CompactibleFreeListSpace(epilogue):: Linear allocation failure");
1988 }
1989 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency");
1990 _promoInfo.stopTrackingPromotions();
1991 repairLinearAllocationBlocks();
1992 // Print Space's stats
1993 if (PrintFLSStatistics != 0) {
1994 gclog_or_tty->print("After GC:\n");
1995 reportFreeListStatistics();
1996 }
1997 }
1999 // Iteration support, mostly delegated from a CMS generation
2001 void CompactibleFreeListSpace::save_marks() {
2002 assert(Thread::current()->is_VM_thread(),
2003 "Global variable should only be set when single-threaded");
2004 // Mark the "end" of the used space at the time of this call;
2005 // note, however, that promoted objects from this point
2006 // on are tracked in the _promoInfo below.
2007 set_saved_mark_word(unallocated_block());
2008 #ifdef ASSERT
2009 // Check the sanity of save_marks() etc.
2010 MemRegion ur = used_region();
2011 MemRegion urasm = used_region_at_save_marks();
2012 assert(ur.contains(urasm),
2013 err_msg(" Error at save_marks(): [" PTR_FORMAT "," PTR_FORMAT ")"
2014 " should contain [" PTR_FORMAT "," PTR_FORMAT ")",
2015 p2i(ur.start()), p2i(ur.end()), p2i(urasm.start()), p2i(urasm.end())));
2016 #endif
2017 // inform allocator that promotions should be tracked.
2018 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency");
2019 _promoInfo.startTrackingPromotions();
2020 }
2022 bool CompactibleFreeListSpace::no_allocs_since_save_marks() {
2023 assert(_promoInfo.tracking(), "No preceding save_marks?");
2024 assert(SharedHeap::heap()->n_par_threads() == 0,
2025 "Shouldn't be called if using parallel gc.");
2026 return _promoInfo.noPromotions();
2027 }
2029 #define CFLS_OOP_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2030 \
2031 void CompactibleFreeListSpace:: \
2032 oop_since_save_marks_iterate##nv_suffix(OopClosureType* blk) { \
2033 assert(SharedHeap::heap()->n_par_threads() == 0, \
2034 "Shouldn't be called (yet) during parallel part of gc."); \
2035 _promoInfo.promoted_oops_iterate##nv_suffix(blk); \
2036 /* \
2037 * This also restores any displaced headers and removes the elements from \
2038 * the iteration set as they are processed, so that we have a clean slate \
2039 * at the end of the iteration. Note, thus, that if new objects are \
2040 * promoted as a result of the iteration they are iterated over as well. \
2041 */ \
2042 assert(_promoInfo.noPromotions(), "_promoInfo inconsistency"); \
2043 }
2045 ALL_SINCE_SAVE_MARKS_CLOSURES(CFLS_OOP_SINCE_SAVE_MARKS_DEFN)
2047 bool CompactibleFreeListSpace::linearAllocationWouldFail() const {
2048 return _smallLinearAllocBlock._word_size == 0;
2049 }
2051 void CompactibleFreeListSpace::repairLinearAllocationBlocks() {
2052 // Fix up linear allocation blocks to look like free blocks
2053 repairLinearAllocBlock(&_smallLinearAllocBlock);
2054 }
2056 void CompactibleFreeListSpace::repairLinearAllocBlock(LinearAllocBlock* blk) {
2057 assert_locked();
2058 if (blk->_ptr != NULL) {
2059 assert(blk->_word_size != 0 && blk->_word_size >= MinChunkSize,
2060 "Minimum block size requirement");
2061 FreeChunk* fc = (FreeChunk*)(blk->_ptr);
2062 fc->set_size(blk->_word_size);
2063 fc->link_prev(NULL); // mark as free
2064 fc->dontCoalesce();
2065 assert(fc->is_free(), "just marked it free");
2066 assert(fc->cantCoalesce(), "just marked it uncoalescable");
2067 }
2068 }
2070 void CompactibleFreeListSpace::refillLinearAllocBlocksIfNeeded() {
2071 assert_locked();
2072 if (_smallLinearAllocBlock._ptr == NULL) {
2073 assert(_smallLinearAllocBlock._word_size == 0,
2074 "Size of linAB should be zero if the ptr is NULL");
2075 // Reset the linAB refill and allocation size limit.
2076 _smallLinearAllocBlock.set(0, 0, 1024*SmallForLinearAlloc, SmallForLinearAlloc);
2077 }
2078 refillLinearAllocBlockIfNeeded(&_smallLinearAllocBlock);
2079 }
2081 void
2082 CompactibleFreeListSpace::refillLinearAllocBlockIfNeeded(LinearAllocBlock* blk) {
2083 assert_locked();
2084 assert((blk->_ptr == NULL && blk->_word_size == 0) ||
2085 (blk->_ptr != NULL && blk->_word_size >= MinChunkSize),
2086 "blk invariant");
2087 if (blk->_ptr == NULL) {
2088 refillLinearAllocBlock(blk);
2089 }
2090 if (PrintMiscellaneous && Verbose) {
2091 if (blk->_word_size == 0) {
2092 warning("CompactibleFreeListSpace(prologue):: Linear allocation failure");
2093 }
2094 }
2095 }
2097 void
2098 CompactibleFreeListSpace::refillLinearAllocBlock(LinearAllocBlock* blk) {
2099 assert_locked();
2100 assert(blk->_word_size == 0 && blk->_ptr == NULL,
2101 "linear allocation block should be empty");
2102 FreeChunk* fc;
2103 if (blk->_refillSize < SmallForDictionary &&
2104 (fc = getChunkFromIndexedFreeList(blk->_refillSize)) != NULL) {
2105 // A linAB's strategy might be to use small sizes to reduce
2106 // fragmentation but still get the benefits of allocation from a
2107 // linAB.
2108 } else {
2109 fc = getChunkFromDictionary(blk->_refillSize);
2110 }
2111 if (fc != NULL) {
2112 blk->_ptr = (HeapWord*)fc;
2113 blk->_word_size = fc->size();
2114 fc->dontCoalesce(); // to prevent sweeper from sweeping us up
2115 }
2116 }
2118 // Support for concurrent collection policy decisions.
2119 bool CompactibleFreeListSpace::should_concurrent_collect() const {
2120 // In the future we might want to add in frgamentation stats --
2121 // including erosion of the "mountain" into this decision as well.
2122 return !adaptive_freelists() && linearAllocationWouldFail();
2123 }
2125 // Support for compaction
2127 void CompactibleFreeListSpace::prepare_for_compaction(CompactPoint* cp) {
2128 SCAN_AND_FORWARD(cp,end,block_is_obj,block_size);
2129 // prepare_for_compaction() uses the space between live objects
2130 // so that later phase can skip dead space quickly. So verification
2131 // of the free lists doesn't work after.
2132 }
2134 #define obj_size(q) adjustObjectSize(oop(q)->size())
2135 #define adjust_obj_size(s) adjustObjectSize(s)
2137 void CompactibleFreeListSpace::adjust_pointers() {
2138 // In other versions of adjust_pointers(), a bail out
2139 // based on the amount of live data in the generation
2140 // (i.e., if 0, bail out) may be used.
2141 // Cannot test used() == 0 here because the free lists have already
2142 // been mangled by the compaction.
2144 SCAN_AND_ADJUST_POINTERS(adjust_obj_size);
2145 // See note about verification in prepare_for_compaction().
2146 }
2148 void CompactibleFreeListSpace::compact() {
2149 SCAN_AND_COMPACT(obj_size);
2150 }
2152 // fragmentation_metric = 1 - [sum of (fbs**2) / (sum of fbs)**2]
2153 // where fbs is free block sizes
2154 double CompactibleFreeListSpace::flsFrag() const {
2155 size_t itabFree = totalSizeInIndexedFreeLists();
2156 double frag = 0.0;
2157 size_t i;
2159 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
2160 double sz = i;
2161 frag += _indexedFreeList[i].count() * (sz * sz);
2162 }
2164 double totFree = itabFree +
2165 _dictionary->total_chunk_size(DEBUG_ONLY(freelistLock()));
2166 if (totFree > 0) {
2167 frag = ((frag + _dictionary->sum_of_squared_block_sizes()) /
2168 (totFree * totFree));
2169 frag = (double)1.0 - frag;
2170 } else {
2171 assert(frag == 0.0, "Follows from totFree == 0");
2172 }
2173 return frag;
2174 }
2176 void CompactibleFreeListSpace::beginSweepFLCensus(
2177 float inter_sweep_current,
2178 float inter_sweep_estimate,
2179 float intra_sweep_estimate) {
2180 assert_locked();
2181 size_t i;
2182 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
2183 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[i];
2184 if (PrintFLSStatistics > 1) {
2185 gclog_or_tty->print("size[" SIZE_FORMAT "] : ", i);
2186 }
2187 fl->compute_desired(inter_sweep_current, inter_sweep_estimate, intra_sweep_estimate);
2188 fl->set_coal_desired((ssize_t)((double)fl->desired() * CMSSmallCoalSurplusPercent));
2189 fl->set_before_sweep(fl->count());
2190 fl->set_bfr_surp(fl->surplus());
2191 }
2192 _dictionary->begin_sweep_dict_census(CMSLargeCoalSurplusPercent,
2193 inter_sweep_current,
2194 inter_sweep_estimate,
2195 intra_sweep_estimate);
2196 }
2198 void CompactibleFreeListSpace::setFLSurplus() {
2199 assert_locked();
2200 size_t i;
2201 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
2202 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i];
2203 fl->set_surplus(fl->count() -
2204 (ssize_t)((double)fl->desired() * CMSSmallSplitSurplusPercent));
2205 }
2206 }
2208 void CompactibleFreeListSpace::setFLHints() {
2209 assert_locked();
2210 size_t i;
2211 size_t h = IndexSetSize;
2212 for (i = IndexSetSize - 1; i != 0; i -= IndexSetStride) {
2213 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i];
2214 fl->set_hint(h);
2215 if (fl->surplus() > 0) {
2216 h = i;
2217 }
2218 }
2219 }
2221 void CompactibleFreeListSpace::clearFLCensus() {
2222 assert_locked();
2223 size_t i;
2224 for (i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
2225 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i];
2226 fl->set_prev_sweep(fl->count());
2227 fl->set_coal_births(0);
2228 fl->set_coal_deaths(0);
2229 fl->set_split_births(0);
2230 fl->set_split_deaths(0);
2231 }
2232 }
2234 void CompactibleFreeListSpace::endSweepFLCensus(size_t sweep_count) {
2235 if (PrintFLSStatistics > 0) {
2236 HeapWord* largestAddr = (HeapWord*) dictionary()->find_largest_dict();
2237 gclog_or_tty->print_cr("CMS: Large block " PTR_FORMAT,
2238 p2i(largestAddr));
2239 }
2240 setFLSurplus();
2241 setFLHints();
2242 if (PrintGC && PrintFLSCensus > 0) {
2243 printFLCensus(sweep_count);
2244 }
2245 clearFLCensus();
2246 assert_locked();
2247 _dictionary->end_sweep_dict_census(CMSLargeSplitSurplusPercent);
2248 }
2250 bool CompactibleFreeListSpace::coalOverPopulated(size_t size) {
2251 if (size < SmallForDictionary) {
2252 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size];
2253 return (fl->coal_desired() < 0) ||
2254 ((int)fl->count() > fl->coal_desired());
2255 } else {
2256 return dictionary()->coal_dict_over_populated(size);
2257 }
2258 }
2260 void CompactibleFreeListSpace::smallCoalBirth(size_t size) {
2261 assert(size < SmallForDictionary, "Size too large for indexed list");
2262 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size];
2263 fl->increment_coal_births();
2264 fl->increment_surplus();
2265 }
2267 void CompactibleFreeListSpace::smallCoalDeath(size_t size) {
2268 assert(size < SmallForDictionary, "Size too large for indexed list");
2269 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size];
2270 fl->increment_coal_deaths();
2271 fl->decrement_surplus();
2272 }
2274 void CompactibleFreeListSpace::coalBirth(size_t size) {
2275 if (size < SmallForDictionary) {
2276 smallCoalBirth(size);
2277 } else {
2278 dictionary()->dict_census_update(size,
2279 false /* split */,
2280 true /* birth */);
2281 }
2282 }
2284 void CompactibleFreeListSpace::coalDeath(size_t size) {
2285 if(size < SmallForDictionary) {
2286 smallCoalDeath(size);
2287 } else {
2288 dictionary()->dict_census_update(size,
2289 false /* split */,
2290 false /* birth */);
2291 }
2292 }
2294 void CompactibleFreeListSpace::smallSplitBirth(size_t size) {
2295 assert(size < SmallForDictionary, "Size too large for indexed list");
2296 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size];
2297 fl->increment_split_births();
2298 fl->increment_surplus();
2299 }
2301 void CompactibleFreeListSpace::smallSplitDeath(size_t size) {
2302 assert(size < SmallForDictionary, "Size too large for indexed list");
2303 AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[size];
2304 fl->increment_split_deaths();
2305 fl->decrement_surplus();
2306 }
2308 void CompactibleFreeListSpace::split_birth(size_t size) {
2309 if (size < SmallForDictionary) {
2310 smallSplitBirth(size);
2311 } else {
2312 dictionary()->dict_census_update(size,
2313 true /* split */,
2314 true /* birth */);
2315 }
2316 }
2318 void CompactibleFreeListSpace::splitDeath(size_t size) {
2319 if (size < SmallForDictionary) {
2320 smallSplitDeath(size);
2321 } else {
2322 dictionary()->dict_census_update(size,
2323 true /* split */,
2324 false /* birth */);
2325 }
2326 }
2328 void CompactibleFreeListSpace::split(size_t from, size_t to1) {
2329 size_t to2 = from - to1;
2330 splitDeath(from);
2331 split_birth(to1);
2332 split_birth(to2);
2333 }
2335 void CompactibleFreeListSpace::print() const {
2336 print_on(tty);
2337 }
2339 void CompactibleFreeListSpace::prepare_for_verify() {
2340 assert_locked();
2341 repairLinearAllocationBlocks();
2342 // Verify that the SpoolBlocks look like free blocks of
2343 // appropriate sizes... To be done ...
2344 }
2346 class VerifyAllBlksClosure: public BlkClosure {
2347 private:
2348 const CompactibleFreeListSpace* _sp;
2349 const MemRegion _span;
2350 HeapWord* _last_addr;
2351 size_t _last_size;
2352 bool _last_was_obj;
2353 bool _last_was_live;
2355 public:
2356 VerifyAllBlksClosure(const CompactibleFreeListSpace* sp,
2357 MemRegion span) : _sp(sp), _span(span),
2358 _last_addr(NULL), _last_size(0),
2359 _last_was_obj(false), _last_was_live(false) { }
2361 virtual size_t do_blk(HeapWord* addr) {
2362 size_t res;
2363 bool was_obj = false;
2364 bool was_live = false;
2365 if (_sp->block_is_obj(addr)) {
2366 was_obj = true;
2367 oop p = oop(addr);
2368 guarantee(p->is_oop(), "Should be an oop");
2369 res = _sp->adjustObjectSize(p->size());
2370 if (_sp->obj_is_alive(addr)) {
2371 was_live = true;
2372 p->verify();
2373 }
2374 } else {
2375 FreeChunk* fc = (FreeChunk*)addr;
2376 res = fc->size();
2377 if (FLSVerifyLists && !fc->cantCoalesce()) {
2378 guarantee(_sp->verify_chunk_in_free_list(fc),
2379 "Chunk should be on a free list");
2380 }
2381 }
2382 if (res == 0) {
2383 gclog_or_tty->print_cr("Livelock: no rank reduction!");
2384 gclog_or_tty->print_cr(
2385 " Current: addr = " PTR_FORMAT ", size = " SIZE_FORMAT ", obj = %s, live = %s \n"
2386 " Previous: addr = " PTR_FORMAT ", size = " SIZE_FORMAT ", obj = %s, live = %s \n",
2387 p2i(addr), res, was_obj ?"true":"false", was_live ?"true":"false",
2388 p2i(_last_addr), _last_size, _last_was_obj?"true":"false", _last_was_live?"true":"false");
2389 _sp->print_on(gclog_or_tty);
2390 guarantee(false, "Seppuku!");
2391 }
2392 _last_addr = addr;
2393 _last_size = res;
2394 _last_was_obj = was_obj;
2395 _last_was_live = was_live;
2396 return res;
2397 }
2398 };
2400 class VerifyAllOopsClosure: public OopClosure {
2401 private:
2402 const CMSCollector* _collector;
2403 const CompactibleFreeListSpace* _sp;
2404 const MemRegion _span;
2405 const bool _past_remark;
2406 const CMSBitMap* _bit_map;
2408 protected:
2409 void do_oop(void* p, oop obj) {
2410 if (_span.contains(obj)) { // the interior oop points into CMS heap
2411 if (!_span.contains(p)) { // reference from outside CMS heap
2412 // Should be a valid object; the first disjunct below allows
2413 // us to sidestep an assertion in block_is_obj() that insists
2414 // that p be in _sp. Note that several generations (and spaces)
2415 // are spanned by _span (CMS heap) above.
2416 guarantee(!_sp->is_in_reserved(obj) ||
2417 _sp->block_is_obj((HeapWord*)obj),
2418 "Should be an object");
2419 guarantee(obj->is_oop(), "Should be an oop");
2420 obj->verify();
2421 if (_past_remark) {
2422 // Remark has been completed, the object should be marked
2423 _bit_map->isMarked((HeapWord*)obj);
2424 }
2425 } else { // reference within CMS heap
2426 if (_past_remark) {
2427 // Remark has been completed -- so the referent should have
2428 // been marked, if referring object is.
2429 if (_bit_map->isMarked(_collector->block_start(p))) {
2430 guarantee(_bit_map->isMarked((HeapWord*)obj), "Marking error?");
2431 }
2432 }
2433 }
2434 } else if (_sp->is_in_reserved(p)) {
2435 // the reference is from FLS, and points out of FLS
2436 guarantee(obj->is_oop(), "Should be an oop");
2437 obj->verify();
2438 }
2439 }
2441 template <class T> void do_oop_work(T* p) {
2442 T heap_oop = oopDesc::load_heap_oop(p);
2443 if (!oopDesc::is_null(heap_oop)) {
2444 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
2445 do_oop(p, obj);
2446 }
2447 }
2449 public:
2450 VerifyAllOopsClosure(const CMSCollector* collector,
2451 const CompactibleFreeListSpace* sp, MemRegion span,
2452 bool past_remark, CMSBitMap* bit_map) :
2453 _collector(collector), _sp(sp), _span(span),
2454 _past_remark(past_remark), _bit_map(bit_map) { }
2456 virtual void do_oop(oop* p) { VerifyAllOopsClosure::do_oop_work(p); }
2457 virtual void do_oop(narrowOop* p) { VerifyAllOopsClosure::do_oop_work(p); }
2458 };
2460 void CompactibleFreeListSpace::verify() const {
2461 assert_lock_strong(&_freelistLock);
2462 verify_objects_initialized();
2463 MemRegion span = _collector->_span;
2464 bool past_remark = (_collector->abstract_state() ==
2465 CMSCollector::Sweeping);
2467 ResourceMark rm;
2468 HandleMark hm;
2470 // Check integrity of CFL data structures
2471 _promoInfo.verify();
2472 _dictionary->verify();
2473 if (FLSVerifyIndexTable) {
2474 verifyIndexedFreeLists();
2475 }
2476 // Check integrity of all objects and free blocks in space
2477 {
2478 VerifyAllBlksClosure cl(this, span);
2479 ((CompactibleFreeListSpace*)this)->blk_iterate(&cl); // cast off const
2480 }
2481 // Check that all references in the heap to FLS
2482 // are to valid objects in FLS or that references in
2483 // FLS are to valid objects elsewhere in the heap
2484 if (FLSVerifyAllHeapReferences)
2485 {
2486 VerifyAllOopsClosure cl(_collector, this, span, past_remark,
2487 _collector->markBitMap());
2488 CollectedHeap* ch = Universe::heap();
2490 // Iterate over all oops in the heap. Uses the _no_header version
2491 // since we are not interested in following the klass pointers.
2492 ch->oop_iterate_no_header(&cl);
2493 }
2495 if (VerifyObjectStartArray) {
2496 // Verify the block offset table
2497 _bt.verify();
2498 }
2499 }
2501 #ifndef PRODUCT
2502 void CompactibleFreeListSpace::verifyFreeLists() const {
2503 if (FLSVerifyLists) {
2504 _dictionary->verify();
2505 verifyIndexedFreeLists();
2506 } else {
2507 if (FLSVerifyDictionary) {
2508 _dictionary->verify();
2509 }
2510 if (FLSVerifyIndexTable) {
2511 verifyIndexedFreeLists();
2512 }
2513 }
2514 }
2515 #endif
2517 void CompactibleFreeListSpace::verifyIndexedFreeLists() const {
2518 size_t i = 0;
2519 for (; i < IndexSetStart; i++) {
2520 guarantee(_indexedFreeList[i].head() == NULL, "should be NULL");
2521 }
2522 for (; i < IndexSetSize; i++) {
2523 verifyIndexedFreeList(i);
2524 }
2525 }
2527 void CompactibleFreeListSpace::verifyIndexedFreeList(size_t size) const {
2528 FreeChunk* fc = _indexedFreeList[size].head();
2529 FreeChunk* tail = _indexedFreeList[size].tail();
2530 size_t num = _indexedFreeList[size].count();
2531 size_t n = 0;
2532 guarantee(((size >= IndexSetStart) && (size % IndexSetStride == 0)) || fc == NULL,
2533 "Slot should have been empty");
2534 for (; fc != NULL; fc = fc->next(), n++) {
2535 guarantee(fc->size() == size, "Size inconsistency");
2536 guarantee(fc->is_free(), "!free?");
2537 guarantee(fc->next() == NULL || fc->next()->prev() == fc, "Broken list");
2538 guarantee((fc->next() == NULL) == (fc == tail), "Incorrect tail");
2539 }
2540 guarantee(n == num, "Incorrect count");
2541 }
2543 #ifndef PRODUCT
2544 void CompactibleFreeListSpace::check_free_list_consistency() const {
2545 assert((TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >::min_size() <= IndexSetSize),
2546 "Some sizes can't be allocated without recourse to"
2547 " linear allocation buffers");
2548 assert((TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >::min_size()*HeapWordSize == sizeof(TreeChunk<FreeChunk, AdaptiveFreeList<FreeChunk> >)),
2549 "else MIN_TREE_CHUNK_SIZE is wrong");
2550 assert(IndexSetStart != 0, "IndexSetStart not initialized");
2551 assert(IndexSetStride != 0, "IndexSetStride not initialized");
2552 }
2553 #endif
2555 void CompactibleFreeListSpace::printFLCensus(size_t sweep_count) const {
2556 assert_lock_strong(&_freelistLock);
2557 AdaptiveFreeList<FreeChunk> total;
2558 gclog_or_tty->print("end sweep# " SIZE_FORMAT "\n", sweep_count);
2559 AdaptiveFreeList<FreeChunk>::print_labels_on(gclog_or_tty, "size");
2560 size_t total_free = 0;
2561 for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
2562 const AdaptiveFreeList<FreeChunk> *fl = &_indexedFreeList[i];
2563 total_free += fl->count() * fl->size();
2564 if (i % (40*IndexSetStride) == 0) {
2565 AdaptiveFreeList<FreeChunk>::print_labels_on(gclog_or_tty, "size");
2566 }
2567 fl->print_on(gclog_or_tty);
2568 total.set_bfr_surp( total.bfr_surp() + fl->bfr_surp() );
2569 total.set_surplus( total.surplus() + fl->surplus() );
2570 total.set_desired( total.desired() + fl->desired() );
2571 total.set_prev_sweep( total.prev_sweep() + fl->prev_sweep() );
2572 total.set_before_sweep(total.before_sweep() + fl->before_sweep());
2573 total.set_count( total.count() + fl->count() );
2574 total.set_coal_births( total.coal_births() + fl->coal_births() );
2575 total.set_coal_deaths( total.coal_deaths() + fl->coal_deaths() );
2576 total.set_split_births(total.split_births() + fl->split_births());
2577 total.set_split_deaths(total.split_deaths() + fl->split_deaths());
2578 }
2579 total.print_on(gclog_or_tty, "TOTAL");
2580 gclog_or_tty->print_cr("Total free in indexed lists "
2581 SIZE_FORMAT " words", total_free);
2582 gclog_or_tty->print("growth: %8.5f deficit: %8.5f\n",
2583 (double)(total.split_births()+total.coal_births()-total.split_deaths()-total.coal_deaths())/
2584 (total.prev_sweep() != 0 ? (double)total.prev_sweep() : 1.0),
2585 (double)(total.desired() - total.count())/(total.desired() != 0 ? (double)total.desired() : 1.0));
2586 _dictionary->print_dict_census();
2587 }
2589 ///////////////////////////////////////////////////////////////////////////
2590 // CFLS_LAB
2591 ///////////////////////////////////////////////////////////////////////////
2593 #define VECTOR_257(x) \
2594 /* 1 2 3 4 5 6 7 8 9 1x 11 12 13 14 15 16 17 18 19 2x 21 22 23 24 25 26 27 28 29 3x 31 32 */ \
2595 { x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \
2596 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \
2597 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \
2598 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \
2599 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \
2600 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \
2601 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \
2602 x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, x, \
2603 x }
2605 // Initialize with default setting of CMSParPromoteBlocksToClaim, _not_
2606 // OldPLABSize, whose static default is different; if overridden at the
2607 // command-line, this will get reinitialized via a call to
2608 // modify_initialization() below.
2609 AdaptiveWeightedAverage CFLS_LAB::_blocks_to_claim[] =
2610 VECTOR_257(AdaptiveWeightedAverage(OldPLABWeight, (float)CMSParPromoteBlocksToClaim));
2611 size_t CFLS_LAB::_global_num_blocks[] = VECTOR_257(0);
2612 uint CFLS_LAB::_global_num_workers[] = VECTOR_257(0);
2614 CFLS_LAB::CFLS_LAB(CompactibleFreeListSpace* cfls) :
2615 _cfls(cfls)
2616 {
2617 assert(CompactibleFreeListSpace::IndexSetSize == 257, "Modify VECTOR_257() macro above");
2618 for (size_t i = CompactibleFreeListSpace::IndexSetStart;
2619 i < CompactibleFreeListSpace::IndexSetSize;
2620 i += CompactibleFreeListSpace::IndexSetStride) {
2621 _indexedFreeList[i].set_size(i);
2622 _num_blocks[i] = 0;
2623 }
2624 }
2626 static bool _CFLS_LAB_modified = false;
2628 void CFLS_LAB::modify_initialization(size_t n, unsigned wt) {
2629 assert(!_CFLS_LAB_modified, "Call only once");
2630 _CFLS_LAB_modified = true;
2631 for (size_t i = CompactibleFreeListSpace::IndexSetStart;
2632 i < CompactibleFreeListSpace::IndexSetSize;
2633 i += CompactibleFreeListSpace::IndexSetStride) {
2634 _blocks_to_claim[i].modify(n, wt, true /* force */);
2635 }
2636 }
2638 HeapWord* CFLS_LAB::alloc(size_t word_sz) {
2639 FreeChunk* res;
2640 assert(word_sz == _cfls->adjustObjectSize(word_sz), "Error");
2641 if (word_sz >= CompactibleFreeListSpace::IndexSetSize) {
2642 // This locking manages sync with other large object allocations.
2643 MutexLockerEx x(_cfls->parDictionaryAllocLock(),
2644 Mutex::_no_safepoint_check_flag);
2645 res = _cfls->getChunkFromDictionaryExact(word_sz);
2646 if (res == NULL) return NULL;
2647 } else {
2648 AdaptiveFreeList<FreeChunk>* fl = &_indexedFreeList[word_sz];
2649 if (fl->count() == 0) {
2650 // Attempt to refill this local free list.
2651 get_from_global_pool(word_sz, fl);
2652 // If it didn't work, give up.
2653 if (fl->count() == 0) return NULL;
2654 }
2655 res = fl->get_chunk_at_head();
2656 assert(res != NULL, "Why was count non-zero?");
2657 }
2658 res->markNotFree();
2659 assert(!res->is_free(), "shouldn't be marked free");
2660 assert(oop(res)->klass_or_null() == NULL, "should look uninitialized");
2661 // mangle a just allocated object with a distinct pattern.
2662 debug_only(res->mangleAllocated(word_sz));
2663 return (HeapWord*)res;
2664 }
2666 // Get a chunk of blocks of the right size and update related
2667 // book-keeping stats
2668 void CFLS_LAB::get_from_global_pool(size_t word_sz, AdaptiveFreeList<FreeChunk>* fl) {
2669 // Get the #blocks we want to claim
2670 size_t n_blks = (size_t)_blocks_to_claim[word_sz].average();
2671 assert(n_blks > 0, "Error");
2672 assert(ResizePLAB || n_blks == OldPLABSize, "Error");
2673 // In some cases, when the application has a phase change,
2674 // there may be a sudden and sharp shift in the object survival
2675 // profile, and updating the counts at the end of a scavenge
2676 // may not be quick enough, giving rise to large scavenge pauses
2677 // during these phase changes. It is beneficial to detect such
2678 // changes on-the-fly during a scavenge and avoid such a phase-change
2679 // pothole. The following code is a heuristic attempt to do that.
2680 // It is protected by a product flag until we have gained
2681 // enough experience with this heuristic and fine-tuned its behaviour.
2682 // WARNING: This might increase fragmentation if we overreact to
2683 // small spikes, so some kind of historical smoothing based on
2684 // previous experience with the greater reactivity might be useful.
2685 // Lacking sufficient experience, CMSOldPLABResizeQuicker is disabled by
2686 // default.
2687 if (ResizeOldPLAB && CMSOldPLABResizeQuicker) {
2688 size_t multiple = _num_blocks[word_sz]/(CMSOldPLABToleranceFactor*CMSOldPLABNumRefills*n_blks);
2689 n_blks += CMSOldPLABReactivityFactor*multiple*n_blks;
2690 n_blks = MIN2(n_blks, CMSOldPLABMax);
2691 }
2692 assert(n_blks > 0, "Error");
2693 _cfls->par_get_chunk_of_blocks(word_sz, n_blks, fl);
2694 // Update stats table entry for this block size
2695 _num_blocks[word_sz] += fl->count();
2696 }
2698 void CFLS_LAB::compute_desired_plab_size() {
2699 for (size_t i = CompactibleFreeListSpace::IndexSetStart;
2700 i < CompactibleFreeListSpace::IndexSetSize;
2701 i += CompactibleFreeListSpace::IndexSetStride) {
2702 assert((_global_num_workers[i] == 0) == (_global_num_blocks[i] == 0),
2703 "Counter inconsistency");
2704 if (_global_num_workers[i] > 0) {
2705 // Need to smooth wrt historical average
2706 if (ResizeOldPLAB) {
2707 _blocks_to_claim[i].sample(
2708 MAX2((size_t)CMSOldPLABMin,
2709 MIN2((size_t)CMSOldPLABMax,
2710 _global_num_blocks[i]/(_global_num_workers[i]*CMSOldPLABNumRefills))));
2711 }
2712 // Reset counters for next round
2713 _global_num_workers[i] = 0;
2714 _global_num_blocks[i] = 0;
2715 if (PrintOldPLAB) {
2716 gclog_or_tty->print_cr("[" SIZE_FORMAT "]: " SIZE_FORMAT, i, (size_t)_blocks_to_claim[i].average());
2717 }
2718 }
2719 }
2720 }
2722 // If this is changed in the future to allow parallel
2723 // access, one would need to take the FL locks and,
2724 // depending on how it is used, stagger access from
2725 // parallel threads to reduce contention.
2726 void CFLS_LAB::retire(int tid) {
2727 // We run this single threaded with the world stopped;
2728 // so no need for locks and such.
2729 NOT_PRODUCT(Thread* t = Thread::current();)
2730 assert(Thread::current()->is_VM_thread(), "Error");
2731 for (size_t i = CompactibleFreeListSpace::IndexSetStart;
2732 i < CompactibleFreeListSpace::IndexSetSize;
2733 i += CompactibleFreeListSpace::IndexSetStride) {
2734 assert(_num_blocks[i] >= (size_t)_indexedFreeList[i].count(),
2735 "Can't retire more than what we obtained");
2736 if (_num_blocks[i] > 0) {
2737 size_t num_retire = _indexedFreeList[i].count();
2738 assert(_num_blocks[i] > num_retire, "Should have used at least one");
2739 {
2740 // MutexLockerEx x(_cfls->_indexedFreeListParLocks[i],
2741 // Mutex::_no_safepoint_check_flag);
2743 // Update globals stats for num_blocks used
2744 _global_num_blocks[i] += (_num_blocks[i] - num_retire);
2745 _global_num_workers[i]++;
2746 assert(_global_num_workers[i] <= ParallelGCThreads, "Too big");
2747 if (num_retire > 0) {
2748 _cfls->_indexedFreeList[i].prepend(&_indexedFreeList[i]);
2749 // Reset this list.
2750 _indexedFreeList[i] = AdaptiveFreeList<FreeChunk>();
2751 _indexedFreeList[i].set_size(i);
2752 }
2753 }
2754 if (PrintOldPLAB) {
2755 gclog_or_tty->print_cr("%d[" SIZE_FORMAT "]: " SIZE_FORMAT "/" SIZE_FORMAT "/" SIZE_FORMAT,
2756 tid, i, num_retire, _num_blocks[i], (size_t)_blocks_to_claim[i].average());
2757 }
2758 // Reset stats for next round
2759 _num_blocks[i] = 0;
2760 }
2761 }
2762 }
2764 void CompactibleFreeListSpace:: par_get_chunk_of_blocks(size_t word_sz, size_t n, AdaptiveFreeList<FreeChunk>* fl) {
2765 assert(fl->count() == 0, "Precondition.");
2766 assert(word_sz < CompactibleFreeListSpace::IndexSetSize,
2767 "Precondition");
2769 // We'll try all multiples of word_sz in the indexed set, starting with
2770 // word_sz itself and, if CMSSplitIndexedFreeListBlocks, try larger multiples,
2771 // then try getting a big chunk and splitting it.
2772 {
2773 bool found;
2774 int k;
2775 size_t cur_sz;
2776 for (k = 1, cur_sz = k * word_sz, found = false;
2777 (cur_sz < CompactibleFreeListSpace::IndexSetSize) &&
2778 (CMSSplitIndexedFreeListBlocks || k <= 1);
2779 k++, cur_sz = k * word_sz) {
2780 AdaptiveFreeList<FreeChunk> fl_for_cur_sz; // Empty.
2781 fl_for_cur_sz.set_size(cur_sz);
2782 {
2783 MutexLockerEx x(_indexedFreeListParLocks[cur_sz],
2784 Mutex::_no_safepoint_check_flag);
2785 AdaptiveFreeList<FreeChunk>* gfl = &_indexedFreeList[cur_sz];
2786 if (gfl->count() != 0) {
2787 // nn is the number of chunks of size cur_sz that
2788 // we'd need to split k-ways each, in order to create
2789 // "n" chunks of size word_sz each.
2790 const size_t nn = MAX2(n/k, (size_t)1);
2791 gfl->getFirstNChunksFromList(nn, &fl_for_cur_sz);
2792 found = true;
2793 if (k > 1) {
2794 // Update split death stats for the cur_sz-size blocks list:
2795 // we increment the split death count by the number of blocks
2796 // we just took from the cur_sz-size blocks list and which
2797 // we will be splitting below.
2798 ssize_t deaths = gfl->split_deaths() +
2799 fl_for_cur_sz.count();
2800 gfl->set_split_deaths(deaths);
2801 }
2802 }
2803 }
2804 // Now transfer fl_for_cur_sz to fl. Common case, we hope, is k = 1.
2805 if (found) {
2806 if (k == 1) {
2807 fl->prepend(&fl_for_cur_sz);
2808 } else {
2809 // Divide each block on fl_for_cur_sz up k ways.
2810 FreeChunk* fc;
2811 while ((fc = fl_for_cur_sz.get_chunk_at_head()) != NULL) {
2812 // Must do this in reverse order, so that anybody attempting to
2813 // access the main chunk sees it as a single free block until we
2814 // change it.
2815 size_t fc_size = fc->size();
2816 assert(fc->is_free(), "Error");
2817 for (int i = k-1; i >= 0; i--) {
2818 FreeChunk* ffc = (FreeChunk*)((HeapWord*)fc + i * word_sz);
2819 assert((i != 0) ||
2820 ((fc == ffc) && ffc->is_free() &&
2821 (ffc->size() == k*word_sz) && (fc_size == word_sz)),
2822 "Counting error");
2823 ffc->set_size(word_sz);
2824 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads.
2825 ffc->link_next(NULL);
2826 // Above must occur before BOT is updated below.
2827 OrderAccess::storestore();
2828 // splitting from the right, fc_size == i * word_sz
2829 _bt.mark_block((HeapWord*)ffc, word_sz, true /* reducing */);
2830 fc_size -= word_sz;
2831 assert(fc_size == i*word_sz, "Error");
2832 _bt.verify_not_unallocated((HeapWord*)ffc, word_sz);
2833 _bt.verify_single_block((HeapWord*)fc, fc_size);
2834 _bt.verify_single_block((HeapWord*)ffc, word_sz);
2835 // Push this on "fl".
2836 fl->return_chunk_at_head(ffc);
2837 }
2838 // TRAP
2839 assert(fl->tail()->next() == NULL, "List invariant.");
2840 }
2841 }
2842 // Update birth stats for this block size.
2843 size_t num = fl->count();
2844 MutexLockerEx x(_indexedFreeListParLocks[word_sz],
2845 Mutex::_no_safepoint_check_flag);
2846 ssize_t births = _indexedFreeList[word_sz].split_births() + num;
2847 _indexedFreeList[word_sz].set_split_births(births);
2848 return;
2849 }
2850 }
2851 }
2852 // Otherwise, we'll split a block from the dictionary.
2853 FreeChunk* fc = NULL;
2854 FreeChunk* rem_fc = NULL;
2855 size_t rem;
2856 {
2857 MutexLockerEx x(parDictionaryAllocLock(),
2858 Mutex::_no_safepoint_check_flag);
2859 while (n > 0) {
2860 fc = dictionary()->get_chunk(MAX2(n * word_sz, _dictionary->min_size()),
2861 FreeBlockDictionary<FreeChunk>::atLeast);
2862 if (fc != NULL) {
2863 _bt.allocated((HeapWord*)fc, fc->size(), true /* reducing */); // update _unallocated_blk
2864 dictionary()->dict_census_update(fc->size(),
2865 true /*split*/,
2866 false /*birth*/);
2867 break;
2868 } else {
2869 n--;
2870 }
2871 }
2872 if (fc == NULL) return;
2873 // Otherwise, split up that block.
2874 assert((ssize_t)n >= 1, "Control point invariant");
2875 assert(fc->is_free(), "Error: should be a free block");
2876 _bt.verify_single_block((HeapWord*)fc, fc->size());
2877 const size_t nn = fc->size() / word_sz;
2878 n = MIN2(nn, n);
2879 assert((ssize_t)n >= 1, "Control point invariant");
2880 rem = fc->size() - n * word_sz;
2881 // If there is a remainder, and it's too small, allocate one fewer.
2882 if (rem > 0 && rem < MinChunkSize) {
2883 n--; rem += word_sz;
2884 }
2885 // Note that at this point we may have n == 0.
2886 assert((ssize_t)n >= 0, "Control point invariant");
2888 // If n is 0, the chunk fc that was found is not large
2889 // enough to leave a viable remainder. We are unable to
2890 // allocate even one block. Return fc to the
2891 // dictionary and return, leaving "fl" empty.
2892 if (n == 0) {
2893 returnChunkToDictionary(fc);
2894 assert(fl->count() == 0, "We never allocated any blocks");
2895 return;
2896 }
2898 // First return the remainder, if any.
2899 // Note that we hold the lock until we decide if we're going to give
2900 // back the remainder to the dictionary, since a concurrent allocation
2901 // may otherwise see the heap as empty. (We're willing to take that
2902 // hit if the block is a small block.)
2903 if (rem > 0) {
2904 size_t prefix_size = n * word_sz;
2905 rem_fc = (FreeChunk*)((HeapWord*)fc + prefix_size);
2906 rem_fc->set_size(rem);
2907 rem_fc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads.
2908 rem_fc->link_next(NULL);
2909 // Above must occur before BOT is updated below.
2910 assert((ssize_t)n > 0 && prefix_size > 0 && rem_fc > fc, "Error");
2911 OrderAccess::storestore();
2912 _bt.split_block((HeapWord*)fc, fc->size(), prefix_size);
2913 assert(fc->is_free(), "Error");
2914 fc->set_size(prefix_size);
2915 if (rem >= IndexSetSize) {
2916 returnChunkToDictionary(rem_fc);
2917 dictionary()->dict_census_update(rem, true /*split*/, true /*birth*/);
2918 rem_fc = NULL;
2919 }
2920 // Otherwise, return it to the small list below.
2921 }
2922 }
2923 if (rem_fc != NULL) {
2924 MutexLockerEx x(_indexedFreeListParLocks[rem],
2925 Mutex::_no_safepoint_check_flag);
2926 _bt.verify_not_unallocated((HeapWord*)rem_fc, rem_fc->size());
2927 _indexedFreeList[rem].return_chunk_at_head(rem_fc);
2928 smallSplitBirth(rem);
2929 }
2930 assert((ssize_t)n > 0 && fc != NULL, "Consistency");
2931 // Now do the splitting up.
2932 // Must do this in reverse order, so that anybody attempting to
2933 // access the main chunk sees it as a single free block until we
2934 // change it.
2935 size_t fc_size = n * word_sz;
2936 // All but first chunk in this loop
2937 for (ssize_t i = n-1; i > 0; i--) {
2938 FreeChunk* ffc = (FreeChunk*)((HeapWord*)fc + i * word_sz);
2939 ffc->set_size(word_sz);
2940 ffc->link_prev(NULL); // Mark as a free block for other (parallel) GC threads.
2941 ffc->link_next(NULL);
2942 // Above must occur before BOT is updated below.
2943 OrderAccess::storestore();
2944 // splitting from the right, fc_size == (n - i + 1) * wordsize
2945 _bt.mark_block((HeapWord*)ffc, word_sz, true /* reducing */);
2946 fc_size -= word_sz;
2947 _bt.verify_not_unallocated((HeapWord*)ffc, ffc->size());
2948 _bt.verify_single_block((HeapWord*)ffc, ffc->size());
2949 _bt.verify_single_block((HeapWord*)fc, fc_size);
2950 // Push this on "fl".
2951 fl->return_chunk_at_head(ffc);
2952 }
2953 // First chunk
2954 assert(fc->is_free() && fc->size() == n*word_sz, "Error: should still be a free block");
2955 // The blocks above should show their new sizes before the first block below
2956 fc->set_size(word_sz);
2957 fc->link_prev(NULL); // idempotent wrt free-ness, see assert above
2958 fc->link_next(NULL);
2959 _bt.verify_not_unallocated((HeapWord*)fc, fc->size());
2960 _bt.verify_single_block((HeapWord*)fc, fc->size());
2961 fl->return_chunk_at_head(fc);
2963 assert((ssize_t)n > 0 && (ssize_t)n == fl->count(), "Incorrect number of blocks");
2964 {
2965 // Update the stats for this block size.
2966 MutexLockerEx x(_indexedFreeListParLocks[word_sz],
2967 Mutex::_no_safepoint_check_flag);
2968 const ssize_t births = _indexedFreeList[word_sz].split_births() + n;
2969 _indexedFreeList[word_sz].set_split_births(births);
2970 // ssize_t new_surplus = _indexedFreeList[word_sz].surplus() + n;
2971 // _indexedFreeList[word_sz].set_surplus(new_surplus);
2972 }
2974 // TRAP
2975 assert(fl->tail()->next() == NULL, "List invariant.");
2976 }
2978 // Set up the space's par_seq_tasks structure for work claiming
2979 // for parallel rescan. See CMSParRemarkTask where this is currently used.
2980 // XXX Need to suitably abstract and generalize this and the next
2981 // method into one.
2982 void
2983 CompactibleFreeListSpace::
2984 initialize_sequential_subtasks_for_rescan(int n_threads) {
2985 // The "size" of each task is fixed according to rescan_task_size.
2986 assert(n_threads > 0, "Unexpected n_threads argument");
2987 const size_t task_size = rescan_task_size();
2988 size_t n_tasks = (used_region().word_size() + task_size - 1)/task_size;
2989 assert((n_tasks == 0) == used_region().is_empty(), "n_tasks incorrect");
2990 assert(n_tasks == 0 ||
2991 ((used_region().start() + (n_tasks - 1)*task_size < used_region().end()) &&
2992 (used_region().start() + n_tasks*task_size >= used_region().end())),
2993 "n_tasks calculation incorrect");
2994 SequentialSubTasksDone* pst = conc_par_seq_tasks();
2995 assert(!pst->valid(), "Clobbering existing data?");
2996 // Sets the condition for completion of the subtask (how many threads
2997 // need to finish in order to be done).
2998 pst->set_n_threads(n_threads);
2999 pst->set_n_tasks((int)n_tasks);
3000 }
3002 // Set up the space's par_seq_tasks structure for work claiming
3003 // for parallel concurrent marking. See CMSConcMarkTask where this is currently used.
3004 void
3005 CompactibleFreeListSpace::
3006 initialize_sequential_subtasks_for_marking(int n_threads,
3007 HeapWord* low) {
3008 // The "size" of each task is fixed according to rescan_task_size.
3009 assert(n_threads > 0, "Unexpected n_threads argument");
3010 const size_t task_size = marking_task_size();
3011 assert(task_size > CardTableModRefBS::card_size_in_words &&
3012 (task_size % CardTableModRefBS::card_size_in_words == 0),
3013 "Otherwise arithmetic below would be incorrect");
3014 MemRegion span = _gen->reserved();
3015 if (low != NULL) {
3016 if (span.contains(low)) {
3017 // Align low down to a card boundary so that
3018 // we can use block_offset_careful() on span boundaries.
3019 HeapWord* aligned_low = (HeapWord*)align_size_down((uintptr_t)low,
3020 CardTableModRefBS::card_size);
3021 // Clip span prefix at aligned_low
3022 span = span.intersection(MemRegion(aligned_low, span.end()));
3023 } else if (low > span.end()) {
3024 span = MemRegion(low, low); // Null region
3025 } // else use entire span
3026 }
3027 assert(span.is_empty() ||
3028 ((uintptr_t)span.start() % CardTableModRefBS::card_size == 0),
3029 "span should start at a card boundary");
3030 size_t n_tasks = (span.word_size() + task_size - 1)/task_size;
3031 assert((n_tasks == 0) == span.is_empty(), "Inconsistency");
3032 assert(n_tasks == 0 ||
3033 ((span.start() + (n_tasks - 1)*task_size < span.end()) &&
3034 (span.start() + n_tasks*task_size >= span.end())),
3035 "n_tasks calculation incorrect");
3036 SequentialSubTasksDone* pst = conc_par_seq_tasks();
3037 assert(!pst->valid(), "Clobbering existing data?");
3038 // Sets the condition for completion of the subtask (how many threads
3039 // need to finish in order to be done).
3040 pst->set_n_threads(n_threads);
3041 pst->set_n_tasks((int)n_tasks);
3042 }