1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc/cms/cmsCollectorPolicy.hpp"
31 #include "gc/cms/cmsOopClosures.inline.hpp"
32 #include "gc/cms/compactibleFreeListSpace.hpp"
33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
34 #include "gc/cms/concurrentMarkSweepThread.hpp"
35 #include "gc/cms/parNewGeneration.hpp"
36 #include "gc/cms/vmCMSOperations.hpp"
37 #include "gc/serial/genMarkSweep.hpp"
38 #include "gc/serial/tenuredGeneration.hpp"
39 #include "gc/shared/adaptiveSizePolicy.hpp"
40 #include "gc/shared/cardGeneration.inline.hpp"
41 #include "gc/shared/cardTableRS.hpp"
42 #include "gc/shared/collectedHeap.inline.hpp"
43 #include "gc/shared/collectorCounters.hpp"
44 #include "gc/shared/collectorPolicy.hpp"
45 #include "gc/shared/gcLocker.inline.hpp"
46 #include "gc/shared/gcPolicyCounters.hpp"
47 #include "gc/shared/gcTimer.hpp"
48 #include "gc/shared/gcTrace.hpp"
49 #include "gc/shared/gcTraceTime.hpp"
50 #include "gc/shared/genCollectedHeap.hpp"
51 #include "gc/shared/genOopClosures.inline.hpp"
52 #include "gc/shared/isGCActiveMark.hpp"
53 #include "gc/shared/referencePolicy.hpp"
54 #include "gc/shared/strongRootsScope.hpp"
55 #include "gc/shared/taskqueue.inline.hpp"
56 #include "memory/allocation.hpp"
57 #include "memory/iterator.inline.hpp"
58 #include "memory/padded.hpp"
59 #include "memory/resourceArea.hpp"
60 #include "oops/oop.inline.hpp"
61 #include "prims/jvmtiExport.hpp"
62 #include "runtime/atomic.inline.hpp"
63 #include "runtime/globals_extension.hpp"
64 #include "runtime/handles.inline.hpp"
65 #include "runtime/java.hpp"
66 #include "runtime/orderAccess.inline.hpp"
67 #include "runtime/vmThread.hpp"
68 #include "services/memoryService.hpp"
69 #include "services/runtimeService.hpp"
70 #include "utilities/stack.inline.hpp"
71
72 // statics
73 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
74 bool CMSCollector::_full_gc_requested = false;
75 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
76
77 //////////////////////////////////////////////////////////////////
78 // In support of CMS/VM thread synchronization
79 //////////////////////////////////////////////////////////////////
80 // We split use of the CGC_lock into 2 "levels".
81 // The low-level locking is of the usual CGC_lock monitor. We introduce
82 // a higher level "token" (hereafter "CMS token") built on top of the
83 // low level monitor (hereafter "CGC lock").
84 // The token-passing protocol gives priority to the VM thread. The
85 // CMS-lock doesn't provide any fairness guarantees, but clients
86 // should ensure that it is only held for very short, bounded
87 // durations.
88 //
89 // When either of the CMS thread or the VM thread is involved in
90 // collection operations during which it does not want the other
91 // thread to interfere, it obtains the CMS token.
92 //
93 // If either thread tries to get the token while the other has
94 // it, that thread waits. However, if the VM thread and CMS thread
95 // both want the token, then the VM thread gets priority while the
96 // CMS thread waits. This ensures, for instance, that the "concurrent"
97 // phases of the CMS thread's work do not block out the VM thread
98 // for long periods of time as the CMS thread continues to hog
99 // the token. (See bug 4616232).
100 //
101 // The baton-passing functions are, however, controlled by the
102 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
103 // and here the low-level CMS lock, not the high level token,
104 // ensures mutual exclusion.
105 //
106 // Two important conditions that we have to satisfy:
107 // 1. if a thread does a low-level wait on the CMS lock, then it
108 // relinquishes the CMS token if it were holding that token
109 // when it acquired the low-level CMS lock.
110 // 2. any low-level notifications on the low-level lock
111 // should only be sent when a thread has relinquished the token.
112 //
113 // In the absence of either property, we'd have potential deadlock.
114 //
115 // We protect each of the CMS (concurrent and sequential) phases
116 // with the CMS _token_, not the CMS _lock_.
117 //
118 // The only code protected by CMS lock is the token acquisition code
119 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
120 // baton-passing code.
121 //
122 // Unfortunately, i couldn't come up with a good abstraction to factor and
123 // hide the naked CGC_lock manipulation in the baton-passing code
124 // further below. That's something we should try to do. Also, the proof
125 // of correctness of this 2-level locking scheme is far from obvious,
126 // and potentially quite slippery. We have an uneasy suspicion, for instance,
127 // that there may be a theoretical possibility of delay/starvation in the
128 // low-level lock/wait/notify scheme used for the baton-passing because of
129 // potential interference with the priority scheme embodied in the
130 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
131 // invocation further below and marked with "XXX 20011219YSR".
132 // Indeed, as we note elsewhere, this may become yet more slippery
133 // in the presence of multiple CMS and/or multiple VM threads. XXX
134
135 class CMSTokenSync: public StackObj {
136 private:
137 bool _is_cms_thread;
138 public:
139 CMSTokenSync(bool is_cms_thread):
140 _is_cms_thread(is_cms_thread) {
141 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
142 "Incorrect argument to constructor");
143 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
144 }
145
146 ~CMSTokenSync() {
147 assert(_is_cms_thread ?
148 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
149 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
150 "Incorrect state");
151 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
152 }
153 };
154
155 // Convenience class that does a CMSTokenSync, and then acquires
156 // upto three locks.
157 class CMSTokenSyncWithLocks: public CMSTokenSync {
158 private:
159 // Note: locks are acquired in textual declaration order
160 // and released in the opposite order
161 MutexLockerEx _locker1, _locker2, _locker3;
162 public:
163 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
164 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
165 CMSTokenSync(is_cms_thread),
166 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
167 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
168 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
169 { }
170 };
171
172
173 //////////////////////////////////////////////////////////////////
174 // Concurrent Mark-Sweep Generation /////////////////////////////
175 //////////////////////////////////////////////////////////////////
176
177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
178
179 // This struct contains per-thread things necessary to support parallel
180 // young-gen collection.
181 class CMSParGCThreadState: public CHeapObj<mtGC> {
182 public:
183 CFLS_LAB lab;
184 PromotionInfo promo;
185
186 // Constructor.
187 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
188 promo.setSpace(cfls);
189 }
190 };
191
192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
193 ReservedSpace rs, size_t initial_byte_size,
194 CardTableRS* ct, bool use_adaptive_freelists,
195 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
196 CardGeneration(rs, initial_byte_size, ct),
197 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
198 _did_compact(false)
199 {
200 HeapWord* bottom = (HeapWord*) _virtual_space.low();
201 HeapWord* end = (HeapWord*) _virtual_space.high();
202
203 _direct_allocated_words = 0;
204 NOT_PRODUCT(
205 _numObjectsPromoted = 0;
206 _numWordsPromoted = 0;
207 _numObjectsAllocated = 0;
208 _numWordsAllocated = 0;
209 )
210
211 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
212 use_adaptive_freelists,
213 dictionaryChoice);
214 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
215 _cmsSpace->_old_gen = this;
216
217 _gc_stats = new CMSGCStats();
218
219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
220 // offsets match. The ability to tell free chunks from objects
221 // depends on this property.
222 debug_only(
223 FreeChunk* junk = NULL;
224 assert(UseCompressedClassPointers ||
225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
226 "Offset of FreeChunk::_prev within FreeChunk must match"
227 " that of OopDesc::_klass within OopDesc");
228 )
229
230 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
231 for (uint i = 0; i < ParallelGCThreads; i++) {
232 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
233 }
234
235 _incremental_collection_failed = false;
236 // The "dilatation_factor" is the expansion that can occur on
237 // account of the fact that the minimum object size in the CMS
238 // generation may be larger than that in, say, a contiguous young
239 // generation.
240 // Ideally, in the calculation below, we'd compute the dilatation
241 // factor as: MinChunkSize/(promoting_gen's min object size)
242 // Since we do not have such a general query interface for the
243 // promoting generation, we'll instead just use the minimum
244 // object size (which today is a header's worth of space);
245 // note that all arithmetic is in units of HeapWords.
246 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
247 assert(_dilatation_factor >= 1.0, "from previous assert");
248 }
249
250
251 // The field "_initiating_occupancy" represents the occupancy percentage
252 // at which we trigger a new collection cycle. Unless explicitly specified
253 // via CMSInitiatingOccupancyFraction (argument "io" below), it
254 // is calculated by:
255 //
256 // Let "f" be MinHeapFreeRatio in
257 //
258 // _initiating_occupancy = 100-f +
259 // f * (CMSTriggerRatio/100)
260 // where CMSTriggerRatio is the argument "tr" below.
261 //
262 // That is, if we assume the heap is at its desired maximum occupancy at the
263 // end of a collection, we let CMSTriggerRatio of the (purported) free
264 // space be allocated before initiating a new collection cycle.
265 //
266 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
267 assert(io <= 100 && tr <= 100, "Check the arguments");
268 if (io >= 0) {
269 _initiating_occupancy = (double)io / 100.0;
270 } else {
271 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
272 (double)(tr * MinHeapFreeRatio) / 100.0)
273 / 100.0;
274 }
275 }
276
277 void ConcurrentMarkSweepGeneration::ref_processor_init() {
278 assert(collector() != NULL, "no collector");
279 collector()->ref_processor_init();
280 }
281
282 void CMSCollector::ref_processor_init() {
283 if (_ref_processor == NULL) {
284 // Allocate and initialize a reference processor
285 _ref_processor =
286 new ReferenceProcessor(_span, // span
287 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
288 ParallelGCThreads, // mt processing degree
289 _cmsGen->refs_discovery_is_mt(), // mt discovery
290 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
291 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
292 &_is_alive_closure); // closure for liveness info
293 // Initialize the _ref_processor field of CMSGen
294 _cmsGen->set_ref_processor(_ref_processor);
295
296 }
297 }
298
299 AdaptiveSizePolicy* CMSCollector::size_policy() {
300 GenCollectedHeap* gch = GenCollectedHeap::heap();
301 return gch->gen_policy()->size_policy();
302 }
303
304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
305
306 const char* gen_name = "old";
307 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy();
308 // Generation Counters - generation 1, 1 subspace
309 _gen_counters = new GenerationCounters(gen_name, 1, 1,
310 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
311
312 _space_counters = new GSpaceCounters(gen_name, 0,
313 _virtual_space.reserved_size(),
314 this, _gen_counters);
315 }
316
317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
318 _cms_gen(cms_gen)
319 {
320 assert(alpha <= 100, "bad value");
321 _saved_alpha = alpha;
322
323 // Initialize the alphas to the bootstrap value of 100.
324 _gc0_alpha = _cms_alpha = 100;
325
326 _cms_begin_time.update();
327 _cms_end_time.update();
328
329 _gc0_duration = 0.0;
330 _gc0_period = 0.0;
331 _gc0_promoted = 0;
332
333 _cms_duration = 0.0;
334 _cms_period = 0.0;
335 _cms_allocated = 0;
336
337 _cms_used_at_gc0_begin = 0;
338 _cms_used_at_gc0_end = 0;
339 _allow_duty_cycle_reduction = false;
340 _valid_bits = 0;
341 }
342
343 double CMSStats::cms_free_adjustment_factor(size_t free) const {
344 // TBD: CR 6909490
345 return 1.0;
346 }
347
348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
349 }
350
351 // If promotion failure handling is on use
352 // the padded average size of the promotion for each
353 // young generation collection.
354 double CMSStats::time_until_cms_gen_full() const {
355 size_t cms_free = _cms_gen->cmsSpace()->free();
356 GenCollectedHeap* gch = GenCollectedHeap::heap();
357 size_t expected_promotion = MIN2(gch->young_gen()->capacity(),
358 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
359 if (cms_free > expected_promotion) {
360 // Start a cms collection if there isn't enough space to promote
361 // for the next young collection. Use the padded average as
362 // a safety factor.
363 cms_free -= expected_promotion;
364
365 // Adjust by the safety factor.
366 double cms_free_dbl = (double)cms_free;
367 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
368 // Apply a further correction factor which tries to adjust
369 // for recent occurance of concurrent mode failures.
370 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
371 cms_free_dbl = cms_free_dbl * cms_adjustment;
372
373 if (PrintGCDetails && Verbose) {
374 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
375 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
376 cms_free, expected_promotion);
377 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
378 cms_free_dbl, cms_consumption_rate() + 1.0);
379 }
380 // Add 1 in case the consumption rate goes to zero.
381 return cms_free_dbl / (cms_consumption_rate() + 1.0);
382 }
383 return 0.0;
384 }
385
386 // Compare the duration of the cms collection to the
387 // time remaining before the cms generation is empty.
388 // Note that the time from the start of the cms collection
389 // to the start of the cms sweep (less than the total
390 // duration of the cms collection) can be used. This
391 // has been tried and some applications experienced
392 // promotion failures early in execution. This was
393 // possibly because the averages were not accurate
394 // enough at the beginning.
395 double CMSStats::time_until_cms_start() const {
396 // We add "gc0_period" to the "work" calculation
397 // below because this query is done (mostly) at the
398 // end of a scavenge, so we need to conservatively
399 // account for that much possible delay
400 // in the query so as to avoid concurrent mode failures
401 // due to starting the collection just a wee bit too
402 // late.
403 double work = cms_duration() + gc0_period();
404 double deadline = time_until_cms_gen_full();
405 // If a concurrent mode failure occurred recently, we want to be
406 // more conservative and halve our expected time_until_cms_gen_full()
407 if (work > deadline) {
408 if (Verbose && PrintGCDetails) {
409 gclog_or_tty->print(
410 " CMSCollector: collect because of anticipated promotion "
411 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
412 gc0_period(), time_until_cms_gen_full());
413 }
414 return 0.0;
415 }
416 return work - deadline;
417 }
418
419 #ifndef PRODUCT
420 void CMSStats::print_on(outputStream *st) const {
421 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
422 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
423 gc0_duration(), gc0_period(), gc0_promoted());
424 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
425 cms_duration(), cms_period(), cms_allocated());
426 st->print(",cms_since_beg=%g,cms_since_end=%g",
427 cms_time_since_begin(), cms_time_since_end());
428 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
429 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
430
431 if (valid()) {
432 st->print(",promo_rate=%g,cms_alloc_rate=%g",
433 promotion_rate(), cms_allocation_rate());
434 st->print(",cms_consumption_rate=%g,time_until_full=%g",
435 cms_consumption_rate(), time_until_cms_gen_full());
436 }
437 st->print(" ");
438 }
439 #endif // #ifndef PRODUCT
440
441 CMSCollector::CollectorState CMSCollector::_collectorState =
442 CMSCollector::Idling;
443 bool CMSCollector::_foregroundGCIsActive = false;
444 bool CMSCollector::_foregroundGCShouldWait = false;
445
446 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
447 CardTableRS* ct,
448 ConcurrentMarkSweepPolicy* cp):
449 _cmsGen(cmsGen),
450 _ct(ct),
451 _ref_processor(NULL), // will be set later
452 _conc_workers(NULL), // may be set later
453 _abort_preclean(false),
454 _start_sampling(false),
455 _between_prologue_and_epilogue(false),
456 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
457 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
458 -1 /* lock-free */, "No_lock" /* dummy */),
459 _modUnionClosurePar(&_modUnionTable),
460 // Adjust my span to cover old (cms) gen
461 _span(cmsGen->reserved()),
462 // Construct the is_alive_closure with _span & markBitMap
463 _is_alive_closure(_span, &_markBitMap),
464 _restart_addr(NULL),
465 _overflow_list(NULL),
466 _stats(cmsGen),
467 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
468 //verify that this lock should be acquired with safepoint check.
469 Monitor::_safepoint_check_sometimes)),
470 _eden_chunk_array(NULL), // may be set in ctor body
471 _eden_chunk_capacity(0), // -- ditto --
472 _eden_chunk_index(0), // -- ditto --
473 _survivor_plab_array(NULL), // -- ditto --
474 _survivor_chunk_array(NULL), // -- ditto --
475 _survivor_chunk_capacity(0), // -- ditto --
476 _survivor_chunk_index(0), // -- ditto --
477 _ser_pmc_preclean_ovflw(0),
478 _ser_kac_preclean_ovflw(0),
479 _ser_pmc_remark_ovflw(0),
480 _par_pmc_remark_ovflw(0),
481 _ser_kac_ovflw(0),
482 _par_kac_ovflw(0),
483 #ifndef PRODUCT
484 _num_par_pushes(0),
485 #endif
486 _collection_count_start(0),
487 _verifying(false),
488 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
489 _completed_initialization(false),
490 _collector_policy(cp),
491 _should_unload_classes(CMSClassUnloadingEnabled),
492 _concurrent_cycles_since_last_unload(0),
493 _roots_scanning_options(GenCollectedHeap::SO_None),
494 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
495 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
496 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
497 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
498 _cms_start_registered(false)
499 {
500 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
501 ExplicitGCInvokesConcurrent = true;
502 }
503 // Now expand the span and allocate the collection support structures
504 // (MUT, marking bit map etc.) to cover both generations subject to
505 // collection.
506
507 // For use by dirty card to oop closures.
508 _cmsGen->cmsSpace()->set_collector(this);
509
510 // Allocate MUT and marking bit map
511 {
512 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
513 if (!_markBitMap.allocate(_span)) {
514 warning("Failed to allocate CMS Bit Map");
515 return;
516 }
517 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
518 }
519 {
520 _modUnionTable.allocate(_span);
521 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
522 }
523
524 if (!_markStack.allocate(MarkStackSize)) {
525 warning("Failed to allocate CMS Marking Stack");
526 return;
527 }
528
529 // Support for multi-threaded concurrent phases
530 if (CMSConcurrentMTEnabled) {
531 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
532 // just for now
533 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
534 }
535 if (ConcGCThreads > 1) {
536 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
537 ConcGCThreads, true);
538 if (_conc_workers == NULL) {
539 warning("GC/CMS: _conc_workers allocation failure: "
540 "forcing -CMSConcurrentMTEnabled");
541 CMSConcurrentMTEnabled = false;
542 } else {
543 _conc_workers->initialize_workers();
544 }
545 } else {
546 CMSConcurrentMTEnabled = false;
547 }
548 }
549 if (!CMSConcurrentMTEnabled) {
550 ConcGCThreads = 0;
551 } else {
552 // Turn off CMSCleanOnEnter optimization temporarily for
553 // the MT case where it's not fixed yet; see 6178663.
554 CMSCleanOnEnter = false;
555 }
556 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
557 "Inconsistency");
558
559 // Parallel task queues; these are shared for the
560 // concurrent and stop-world phases of CMS, but
561 // are not shared with parallel scavenge (ParNew).
562 {
563 uint i;
564 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
565
566 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
567 || ParallelRefProcEnabled)
568 && num_queues > 0) {
569 _task_queues = new OopTaskQueueSet(num_queues);
570 if (_task_queues == NULL) {
571 warning("task_queues allocation failure.");
572 return;
573 }
574 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
575 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
576 for (i = 0; i < num_queues; i++) {
577 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
578 if (q == NULL) {
579 warning("work_queue allocation failure.");
580 return;
581 }
582 _task_queues->register_queue(i, q);
583 }
584 for (i = 0; i < num_queues; i++) {
585 _task_queues->queue(i)->initialize();
586 _hash_seed[i] = 17; // copied from ParNew
587 }
588 }
589 }
590
591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
592
593 // Clip CMSBootstrapOccupancy between 0 and 100.
594 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
595
596 // Now tell CMS generations the identity of their collector
597 ConcurrentMarkSweepGeneration::set_collector(this);
598
599 // Create & start a CMS thread for this CMS collector
600 _cmsThread = ConcurrentMarkSweepThread::start(this);
601 assert(cmsThread() != NULL, "CMS Thread should have been created");
602 assert(cmsThread()->collector() == this,
603 "CMS Thread should refer to this gen");
604 assert(CGC_lock != NULL, "Where's the CGC_lock?");
605
606 // Support for parallelizing young gen rescan
607 GenCollectedHeap* gch = GenCollectedHeap::heap();
608 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
609 _young_gen = (ParNewGeneration*)gch->young_gen();
610 if (gch->supports_inline_contig_alloc()) {
611 _top_addr = gch->top_addr();
612 _end_addr = gch->end_addr();
613 assert(_young_gen != NULL, "no _young_gen");
614 _eden_chunk_index = 0;
615 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
616 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
617 }
618
619 // Support for parallelizing survivor space rescan
620 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
621 const size_t max_plab_samples =
622 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
623
624 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
625 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
626 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
627 _survivor_chunk_capacity = max_plab_samples;
628 for (uint i = 0; i < ParallelGCThreads; i++) {
629 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
630 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
631 assert(cur->end() == 0, "Should be 0");
632 assert(cur->array() == vec, "Should be vec");
633 assert(cur->capacity() == max_plab_samples, "Error");
634 }
635 }
636
637 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
638 _gc_counters = new CollectorCounters("CMS", 1);
639 _completed_initialization = true;
640 _inter_sweep_timer.start(); // start of time
641 }
642
643 const char* ConcurrentMarkSweepGeneration::name() const {
644 return "concurrent mark-sweep generation";
645 }
646 void ConcurrentMarkSweepGeneration::update_counters() {
647 if (UsePerfData) {
648 _space_counters->update_all();
649 _gen_counters->update_all();
650 }
651 }
652
653 // this is an optimized version of update_counters(). it takes the
654 // used value as a parameter rather than computing it.
655 //
656 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
657 if (UsePerfData) {
658 _space_counters->update_used(used);
659 _space_counters->update_capacity();
660 _gen_counters->update_all();
661 }
662 }
663
664 void ConcurrentMarkSweepGeneration::print() const {
665 Generation::print();
666 cmsSpace()->print();
667 }
668
669 #ifndef PRODUCT
670 void ConcurrentMarkSweepGeneration::print_statistics() {
671 cmsSpace()->printFLCensus(0);
672 }
673 #endif
674
675 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
676 GenCollectedHeap* gch = GenCollectedHeap::heap();
677 if (PrintGCDetails) {
678 // I didn't want to change the logging when removing the level concept,
679 // but I guess this logging could say "old" or something instead of "1".
680 assert(gch->is_old_gen(this),
681 "The CMS generation should be the old generation");
682 uint level = 1;
683 if (Verbose) {
684 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]",
685 level, short_name(), s, used(), capacity());
686 } else {
687 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]",
688 level, short_name(), s, used() / K, capacity() / K);
689 }
690 }
691 if (Verbose) {
692 gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")",
693 gch->used(), gch->capacity());
694 } else {
695 gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)",
696 gch->used() / K, gch->capacity() / K);
697 }
698 }
699
700 size_t
701 ConcurrentMarkSweepGeneration::contiguous_available() const {
702 // dld proposes an improvement in precision here. If the committed
703 // part of the space ends in a free block we should add that to
704 // uncommitted size in the calculation below. Will make this
705 // change later, staying with the approximation below for the
706 // time being. -- ysr.
707 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
708 }
709
710 size_t
711 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
712 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
713 }
714
715 size_t ConcurrentMarkSweepGeneration::max_available() const {
716 return free() + _virtual_space.uncommitted_size();
717 }
718
719 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
720 size_t available = max_available();
721 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
722 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
723 if (Verbose && PrintGCDetails) {
724 gclog_or_tty->print_cr(
725 "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "),"
726 "max_promo(" SIZE_FORMAT ")",
727 res? "":" not", available, res? ">=":"<",
728 av_promo, max_promotion_in_bytes);
729 }
730 return res;
731 }
732
733 // At a promotion failure dump information on block layout in heap
734 // (cms old generation).
735 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
736 if (CMSDumpAtPromotionFailure) {
737 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
738 }
739 }
740
741 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
742 // Clear the promotion information. These pointers can be adjusted
743 // along with all the other pointers into the heap but
744 // compaction is expected to be a rare event with
745 // a heap using cms so don't do it without seeing the need.
746 for (uint i = 0; i < ParallelGCThreads; i++) {
747 _par_gc_thread_states[i]->promo.reset();
748 }
749 }
750
751 void ConcurrentMarkSweepGeneration::compute_new_size() {
752 assert_locked_or_safepoint(Heap_lock);
753
754 // If incremental collection failed, we just want to expand
755 // to the limit.
756 if (incremental_collection_failed()) {
757 clear_incremental_collection_failed();
758 grow_to_reserved();
759 return;
760 }
761
762 // The heap has been compacted but not reset yet.
763 // Any metric such as free() or used() will be incorrect.
764
765 CardGeneration::compute_new_size();
766
767 // Reset again after a possible resizing
768 if (did_compact()) {
769 cmsSpace()->reset_after_compaction();
770 }
771 }
772
773 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
774 assert_locked_or_safepoint(Heap_lock);
775
776 // If incremental collection failed, we just want to expand
777 // to the limit.
778 if (incremental_collection_failed()) {
779 clear_incremental_collection_failed();
780 grow_to_reserved();
781 return;
782 }
783
784 double free_percentage = ((double) free()) / capacity();
785 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
786 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
787
788 // compute expansion delta needed for reaching desired free percentage
789 if (free_percentage < desired_free_percentage) {
790 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
791 assert(desired_capacity >= capacity(), "invalid expansion size");
792 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
793 if (PrintGCDetails && Verbose) {
794 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
795 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
796 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
797 gclog_or_tty->print_cr(" Desired free fraction %f", desired_free_percentage);
798 gclog_or_tty->print_cr(" Maximum free fraction %f", maximum_free_percentage);
799 gclog_or_tty->print_cr(" Capacity " SIZE_FORMAT, capacity() / 1000);
800 gclog_or_tty->print_cr(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
801 GenCollectedHeap* gch = GenCollectedHeap::heap();
802 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation");
803 size_t young_size = gch->young_gen()->capacity();
804 gclog_or_tty->print_cr(" Young gen size " SIZE_FORMAT, young_size / 1000);
805 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
806 gclog_or_tty->print_cr(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
807 gclog_or_tty->print_cr(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
808 }
809 // safe if expansion fails
810 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
811 if (PrintGCDetails && Verbose) {
812 gclog_or_tty->print_cr(" Expanded free fraction %f", ((double) free()) / capacity());
813 }
814 } else {
815 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
816 assert(desired_capacity <= capacity(), "invalid expansion size");
817 size_t shrink_bytes = capacity() - desired_capacity;
818 // Don't shrink unless the delta is greater than the minimum shrink we want
819 if (shrink_bytes >= MinHeapDeltaBytes) {
820 shrink_free_list_by(shrink_bytes);
821 }
822 }
823 }
824
825 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
826 return cmsSpace()->freelistLock();
827 }
828
829 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
830 CMSSynchronousYieldRequest yr;
831 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
832 return have_lock_and_allocate(size, tlab);
833 }
834
835 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
836 bool tlab /* ignored */) {
837 assert_lock_strong(freelistLock());
838 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
839 HeapWord* res = cmsSpace()->allocate(adjustedSize);
840 // Allocate the object live (grey) if the background collector has
841 // started marking. This is necessary because the marker may
842 // have passed this address and consequently this object will
843 // not otherwise be greyed and would be incorrectly swept up.
844 // Note that if this object contains references, the writing
845 // of those references will dirty the card containing this object
846 // allowing the object to be blackened (and its references scanned)
847 // either during a preclean phase or at the final checkpoint.
848 if (res != NULL) {
849 // We may block here with an uninitialized object with
850 // its mark-bit or P-bits not yet set. Such objects need
851 // to be safely navigable by block_start().
852 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
853 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
854 collector()->direct_allocated(res, adjustedSize);
855 _direct_allocated_words += adjustedSize;
856 // allocation counters
857 NOT_PRODUCT(
858 _numObjectsAllocated++;
859 _numWordsAllocated += (int)adjustedSize;
860 )
861 }
862 return res;
863 }
864
865 // In the case of direct allocation by mutators in a generation that
866 // is being concurrently collected, the object must be allocated
867 // live (grey) if the background collector has started marking.
868 // This is necessary because the marker may
869 // have passed this address and consequently this object will
870 // not otherwise be greyed and would be incorrectly swept up.
871 // Note that if this object contains references, the writing
872 // of those references will dirty the card containing this object
873 // allowing the object to be blackened (and its references scanned)
874 // either during a preclean phase or at the final checkpoint.
875 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
876 assert(_markBitMap.covers(start, size), "Out of bounds");
877 if (_collectorState >= Marking) {
878 MutexLockerEx y(_markBitMap.lock(),
879 Mutex::_no_safepoint_check_flag);
880 // [see comments preceding SweepClosure::do_blk() below for details]
881 //
882 // Can the P-bits be deleted now? JJJ
883 //
884 // 1. need to mark the object as live so it isn't collected
885 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
886 // 3. need to mark the end of the object so marking, precleaning or sweeping
887 // can skip over uninitialized or unparsable objects. An allocated
888 // object is considered uninitialized for our purposes as long as
889 // its klass word is NULL. All old gen objects are parsable
890 // as soon as they are initialized.)
891 _markBitMap.mark(start); // object is live
892 _markBitMap.mark(start + 1); // object is potentially uninitialized?
893 _markBitMap.mark(start + size - 1);
894 // mark end of object
895 }
896 // check that oop looks uninitialized
897 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
898 }
899
900 void CMSCollector::promoted(bool par, HeapWord* start,
901 bool is_obj_array, size_t obj_size) {
902 assert(_markBitMap.covers(start), "Out of bounds");
903 // See comment in direct_allocated() about when objects should
904 // be allocated live.
905 if (_collectorState >= Marking) {
906 // we already hold the marking bit map lock, taken in
907 // the prologue
908 if (par) {
909 _markBitMap.par_mark(start);
910 } else {
911 _markBitMap.mark(start);
912 }
913 // We don't need to mark the object as uninitialized (as
914 // in direct_allocated above) because this is being done with the
915 // world stopped and the object will be initialized by the
916 // time the marking, precleaning or sweeping get to look at it.
917 // But see the code for copying objects into the CMS generation,
918 // where we need to ensure that concurrent readers of the
919 // block offset table are able to safely navigate a block that
920 // is in flux from being free to being allocated (and in
921 // transition while being copied into) and subsequently
922 // becoming a bona-fide object when the copy/promotion is complete.
923 assert(SafepointSynchronize::is_at_safepoint(),
924 "expect promotion only at safepoints");
925
926 if (_collectorState < Sweeping) {
927 // Mark the appropriate cards in the modUnionTable, so that
928 // this object gets scanned before the sweep. If this is
929 // not done, CMS generation references in the object might
930 // not get marked.
931 // For the case of arrays, which are otherwise precisely
932 // marked, we need to dirty the entire array, not just its head.
933 if (is_obj_array) {
934 // The [par_]mark_range() method expects mr.end() below to
935 // be aligned to the granularity of a bit's representation
936 // in the heap. In the case of the MUT below, that's a
937 // card size.
938 MemRegion mr(start,
939 (HeapWord*)round_to((intptr_t)(start + obj_size),
940 CardTableModRefBS::card_size /* bytes */));
941 if (par) {
942 _modUnionTable.par_mark_range(mr);
943 } else {
944 _modUnionTable.mark_range(mr);
945 }
946 } else { // not an obj array; we can just mark the head
947 if (par) {
948 _modUnionTable.par_mark(start);
949 } else {
950 _modUnionTable.mark(start);
951 }
952 }
953 }
954 }
955 }
956
957 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
958 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
959 // allocate, copy and if necessary update promoinfo --
960 // delegate to underlying space.
961 assert_lock_strong(freelistLock());
962
963 #ifndef PRODUCT
964 if (GenCollectedHeap::heap()->promotion_should_fail()) {
965 return NULL;
966 }
967 #endif // #ifndef PRODUCT
968
969 oop res = _cmsSpace->promote(obj, obj_size);
970 if (res == NULL) {
971 // expand and retry
972 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
973 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
974 // Since this is the old generation, we don't try to promote
975 // into a more senior generation.
976 res = _cmsSpace->promote(obj, obj_size);
977 }
978 if (res != NULL) {
979 // See comment in allocate() about when objects should
980 // be allocated live.
981 assert(obj->is_oop(), "Will dereference klass pointer below");
982 collector()->promoted(false, // Not parallel
983 (HeapWord*)res, obj->is_objArray(), obj_size);
984 // promotion counters
985 NOT_PRODUCT(
986 _numObjectsPromoted++;
987 _numWordsPromoted +=
988 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
989 )
990 }
991 return res;
992 }
993
994
995 // IMPORTANT: Notes on object size recognition in CMS.
996 // ---------------------------------------------------
997 // A block of storage in the CMS generation is always in
998 // one of three states. A free block (FREE), an allocated
999 // object (OBJECT) whose size() method reports the correct size,
1000 // and an intermediate state (TRANSIENT) in which its size cannot
1001 // be accurately determined.
1002 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1003 // -----------------------------------------------------
1004 // FREE: klass_word & 1 == 1; mark_word holds block size
1005 //
1006 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1007 // obj->size() computes correct size
1008 //
1009 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1010 //
1011 // STATE IDENTIFICATION: (64 bit+COOPS)
1012 // ------------------------------------
1013 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1014 //
1015 // OBJECT: klass_word installed; klass_word != 0;
1016 // obj->size() computes correct size
1017 //
1018 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1019 //
1020 //
1021 // STATE TRANSITION DIAGRAM
1022 //
1023 // mut / parnew mut / parnew
1024 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1025 // ^ |
1026 // |------------------------ DEAD <------------------------------------|
1027 // sweep mut
1028 //
1029 // While a block is in TRANSIENT state its size cannot be determined
1030 // so readers will either need to come back later or stall until
1031 // the size can be determined. Note that for the case of direct
1032 // allocation, P-bits, when available, may be used to determine the
1033 // size of an object that may not yet have been initialized.
1034
1035 // Things to support parallel young-gen collection.
1036 oop
1037 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1038 oop old, markOop m,
1039 size_t word_sz) {
1040 #ifndef PRODUCT
1041 if (GenCollectedHeap::heap()->promotion_should_fail()) {
1042 return NULL;
1043 }
1044 #endif // #ifndef PRODUCT
1045
1046 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1047 PromotionInfo* promoInfo = &ps->promo;
1048 // if we are tracking promotions, then first ensure space for
1049 // promotion (including spooling space for saving header if necessary).
1050 // then allocate and copy, then track promoted info if needed.
1051 // When tracking (see PromotionInfo::track()), the mark word may
1052 // be displaced and in this case restoration of the mark word
1053 // occurs in the (oop_since_save_marks_)iterate phase.
1054 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1055 // Out of space for allocating spooling buffers;
1056 // try expanding and allocating spooling buffers.
1057 if (!expand_and_ensure_spooling_space(promoInfo)) {
1058 return NULL;
1059 }
1060 }
1061 assert(promoInfo->has_spooling_space(), "Control point invariant");
1062 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1063 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1064 if (obj_ptr == NULL) {
1065 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1066 if (obj_ptr == NULL) {
1067 return NULL;
1068 }
1069 }
1070 oop obj = oop(obj_ptr);
1071 OrderAccess::storestore();
1072 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1073 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1074 // IMPORTANT: See note on object initialization for CMS above.
1075 // Otherwise, copy the object. Here we must be careful to insert the
1076 // klass pointer last, since this marks the block as an allocated object.
1077 // Except with compressed oops it's the mark word.
1078 HeapWord* old_ptr = (HeapWord*)old;
1079 // Restore the mark word copied above.
1080 obj->set_mark(m);
1081 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1082 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1083 OrderAccess::storestore();
1084
1085 if (UseCompressedClassPointers) {
1086 // Copy gap missed by (aligned) header size calculation below
1087 obj->set_klass_gap(old->klass_gap());
1088 }
1089 if (word_sz > (size_t)oopDesc::header_size()) {
1090 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1091 obj_ptr + oopDesc::header_size(),
1092 word_sz - oopDesc::header_size());
1093 }
1094
1095 // Now we can track the promoted object, if necessary. We take care
1096 // to delay the transition from uninitialized to full object
1097 // (i.e., insertion of klass pointer) until after, so that it
1098 // atomically becomes a promoted object.
1099 if (promoInfo->tracking()) {
1100 promoInfo->track((PromotedObject*)obj, old->klass());
1101 }
1102 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1103 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1104 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1105
1106 // Finally, install the klass pointer (this should be volatile).
1107 OrderAccess::storestore();
1108 obj->set_klass(old->klass());
1109 // We should now be able to calculate the right size for this object
1110 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1111
1112 collector()->promoted(true, // parallel
1113 obj_ptr, old->is_objArray(), word_sz);
1114
1115 NOT_PRODUCT(
1116 Atomic::inc_ptr(&_numObjectsPromoted);
1117 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1118 )
1119
1120 return obj;
1121 }
1122
1123 void
1124 ConcurrentMarkSweepGeneration::
1125 par_promote_alloc_done(int thread_num) {
1126 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1127 ps->lab.retire(thread_num);
1128 }
1129
1130 void
1131 ConcurrentMarkSweepGeneration::
1132 par_oop_since_save_marks_iterate_done(int thread_num) {
1133 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1134 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1135 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1136 }
1137
1138 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1139 size_t size,
1140 bool tlab)
1141 {
1142 // We allow a STW collection only if a full
1143 // collection was requested.
1144 return full || should_allocate(size, tlab); // FIX ME !!!
1145 // This and promotion failure handling are connected at the
1146 // hip and should be fixed by untying them.
1147 }
1148
1149 bool CMSCollector::shouldConcurrentCollect() {
1150 if (_full_gc_requested) {
1151 if (Verbose && PrintGCDetails) {
1152 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1153 " gc request (or gc_locker)");
1154 }
1155 return true;
1156 }
1157
1158 FreelistLocker x(this);
1159 // ------------------------------------------------------------------
1160 // Print out lots of information which affects the initiation of
1161 // a collection.
1162 if (PrintCMSInitiationStatistics && stats().valid()) {
1163 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1164 gclog_or_tty->stamp();
1165 gclog_or_tty->cr();
1166 stats().print_on(gclog_or_tty);
1167 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1168 stats().time_until_cms_gen_full());
1169 gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free());
1170 gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT,
1171 _cmsGen->contiguous_available());
1172 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1173 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1174 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1175 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1176 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1177 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1178 gclog_or_tty->print_cr("metadata initialized %d",
1179 MetaspaceGC::should_concurrent_collect());
1180 }
1181 // ------------------------------------------------------------------
1182
1183 // If the estimated time to complete a cms collection (cms_duration())
1184 // is less than the estimated time remaining until the cms generation
1185 // is full, start a collection.
1186 if (!UseCMSInitiatingOccupancyOnly) {
1187 if (stats().valid()) {
1188 if (stats().time_until_cms_start() == 0.0) {
1189 return true;
1190 }
1191 } else {
1192 // We want to conservatively collect somewhat early in order
1193 // to try and "bootstrap" our CMS/promotion statistics;
1194 // this branch will not fire after the first successful CMS
1195 // collection because the stats should then be valid.
1196 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1197 if (Verbose && PrintGCDetails) {
1198 gclog_or_tty->print_cr(
1199 " CMSCollector: collect for bootstrapping statistics:"
1200 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1201 _bootstrap_occupancy);
1202 }
1203 return true;
1204 }
1205 }
1206 }
1207
1208 // Otherwise, we start a collection cycle if
1209 // old gen want a collection cycle started. Each may use
1210 // an appropriate criterion for making this decision.
1211 // XXX We need to make sure that the gen expansion
1212 // criterion dovetails well with this. XXX NEED TO FIX THIS
1213 if (_cmsGen->should_concurrent_collect()) {
1214 if (Verbose && PrintGCDetails) {
1215 gclog_or_tty->print_cr("CMS old gen initiated");
1216 }
1217 return true;
1218 }
1219
1220 // We start a collection if we believe an incremental collection may fail;
1221 // this is not likely to be productive in practice because it's probably too
1222 // late anyway.
1223 GenCollectedHeap* gch = GenCollectedHeap::heap();
1224 assert(gch->collector_policy()->is_generation_policy(),
1225 "You may want to check the correctness of the following");
1226 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1227 if (Verbose && PrintGCDetails) {
1228 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1229 }
1230 return true;
1231 }
1232
1233 if (MetaspaceGC::should_concurrent_collect()) {
1234 if (Verbose && PrintGCDetails) {
1235 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1236 }
1237 return true;
1238 }
1239
1240 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1241 if (CMSTriggerInterval >= 0) {
1242 if (CMSTriggerInterval == 0) {
1243 // Trigger always
1244 return true;
1245 }
1246
1247 // Check the CMS time since begin (we do not check the stats validity
1248 // as we want to be able to trigger the first CMS cycle as well)
1249 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1250 if (Verbose && PrintGCDetails) {
1251 if (stats().valid()) {
1252 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1253 stats().cms_time_since_begin());
1254 } else {
1255 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)");
1256 }
1257 }
1258 return true;
1259 }
1260 }
1261
1262 return false;
1263 }
1264
1265 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1266
1267 // Clear _expansion_cause fields of constituent generations
1268 void CMSCollector::clear_expansion_cause() {
1269 _cmsGen->clear_expansion_cause();
1270 }
1271
1272 // We should be conservative in starting a collection cycle. To
1273 // start too eagerly runs the risk of collecting too often in the
1274 // extreme. To collect too rarely falls back on full collections,
1275 // which works, even if not optimum in terms of concurrent work.
1276 // As a work around for too eagerly collecting, use the flag
1277 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1278 // giving the user an easily understandable way of controlling the
1279 // collections.
1280 // We want to start a new collection cycle if any of the following
1281 // conditions hold:
1282 // . our current occupancy exceeds the configured initiating occupancy
1283 // for this generation, or
1284 // . we recently needed to expand this space and have not, since that
1285 // expansion, done a collection of this generation, or
1286 // . the underlying space believes that it may be a good idea to initiate
1287 // a concurrent collection (this may be based on criteria such as the
1288 // following: the space uses linear allocation and linear allocation is
1289 // going to fail, or there is believed to be excessive fragmentation in
1290 // the generation, etc... or ...
1291 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1292 // the case of the old generation; see CR 6543076):
1293 // we may be approaching a point at which allocation requests may fail because
1294 // we will be out of sufficient free space given allocation rate estimates.]
1295 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1296
1297 assert_lock_strong(freelistLock());
1298 if (occupancy() > initiating_occupancy()) {
1299 if (PrintGCDetails && Verbose) {
1300 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1301 short_name(), occupancy(), initiating_occupancy());
1302 }
1303 return true;
1304 }
1305 if (UseCMSInitiatingOccupancyOnly) {
1306 return false;
1307 }
1308 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1309 if (PrintGCDetails && Verbose) {
1310 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1311 short_name());
1312 }
1313 return true;
1314 }
1315 if (_cmsSpace->should_concurrent_collect()) {
1316 if (PrintGCDetails && Verbose) {
1317 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1318 short_name());
1319 }
1320 return true;
1321 }
1322 return false;
1323 }
1324
1325 void ConcurrentMarkSweepGeneration::collect(bool full,
1326 bool clear_all_soft_refs,
1327 size_t size,
1328 bool tlab)
1329 {
1330 collector()->collect(full, clear_all_soft_refs, size, tlab);
1331 }
1332
1333 void CMSCollector::collect(bool full,
1334 bool clear_all_soft_refs,
1335 size_t size,
1336 bool tlab)
1337 {
1338 // The following "if" branch is present for defensive reasons.
1339 // In the current uses of this interface, it can be replaced with:
1340 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1341 // But I am not placing that assert here to allow future
1342 // generality in invoking this interface.
1343 if (GC_locker::is_active()) {
1344 // A consistency test for GC_locker
1345 assert(GC_locker::needs_gc(), "Should have been set already");
1346 // Skip this foreground collection, instead
1347 // expanding the heap if necessary.
1348 // Need the free list locks for the call to free() in compute_new_size()
1349 compute_new_size();
1350 return;
1351 }
1352 acquire_control_and_collect(full, clear_all_soft_refs);
1353 }
1354
1355 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1356 GenCollectedHeap* gch = GenCollectedHeap::heap();
1357 unsigned int gc_count = gch->total_full_collections();
1358 if (gc_count == full_gc_count) {
1359 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1360 _full_gc_requested = true;
1361 _full_gc_cause = cause;
1362 CGC_lock->notify(); // nudge CMS thread
1363 } else {
1364 assert(gc_count > full_gc_count, "Error: causal loop");
1365 }
1366 }
1367
1368 bool CMSCollector::is_external_interruption() {
1369 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1370 return GCCause::is_user_requested_gc(cause) ||
1371 GCCause::is_serviceability_requested_gc(cause);
1372 }
1373
1374 void CMSCollector::report_concurrent_mode_interruption() {
1375 if (is_external_interruption()) {
1376 if (PrintGCDetails) {
1377 gclog_or_tty->print(" (concurrent mode interrupted)");
1378 }
1379 } else {
1380 if (PrintGCDetails) {
1381 gclog_or_tty->print(" (concurrent mode failure)");
1382 }
1383 _gc_tracer_cm->report_concurrent_mode_failure();
1384 }
1385 }
1386
1387
1388 // The foreground and background collectors need to coordinate in order
1389 // to make sure that they do not mutually interfere with CMS collections.
1390 // When a background collection is active,
1391 // the foreground collector may need to take over (preempt) and
1392 // synchronously complete an ongoing collection. Depending on the
1393 // frequency of the background collections and the heap usage
1394 // of the application, this preemption can be seldom or frequent.
1395 // There are only certain
1396 // points in the background collection that the "collection-baton"
1397 // can be passed to the foreground collector.
1398 //
1399 // The foreground collector will wait for the baton before
1400 // starting any part of the collection. The foreground collector
1401 // will only wait at one location.
1402 //
1403 // The background collector will yield the baton before starting a new
1404 // phase of the collection (e.g., before initial marking, marking from roots,
1405 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1406 // of the loop which switches the phases. The background collector does some
1407 // of the phases (initial mark, final re-mark) with the world stopped.
1408 // Because of locking involved in stopping the world,
1409 // the foreground collector should not block waiting for the background
1410 // collector when it is doing a stop-the-world phase. The background
1411 // collector will yield the baton at an additional point just before
1412 // it enters a stop-the-world phase. Once the world is stopped, the
1413 // background collector checks the phase of the collection. If the
1414 // phase has not changed, it proceeds with the collection. If the
1415 // phase has changed, it skips that phase of the collection. See
1416 // the comments on the use of the Heap_lock in collect_in_background().
1417 //
1418 // Variable used in baton passing.
1419 // _foregroundGCIsActive - Set to true by the foreground collector when
1420 // it wants the baton. The foreground clears it when it has finished
1421 // the collection.
1422 // _foregroundGCShouldWait - Set to true by the background collector
1423 // when it is running. The foreground collector waits while
1424 // _foregroundGCShouldWait is true.
1425 // CGC_lock - monitor used to protect access to the above variables
1426 // and to notify the foreground and background collectors.
1427 // _collectorState - current state of the CMS collection.
1428 //
1429 // The foreground collector
1430 // acquires the CGC_lock
1431 // sets _foregroundGCIsActive
1432 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1433 // various locks acquired in preparation for the collection
1434 // are released so as not to block the background collector
1435 // that is in the midst of a collection
1436 // proceeds with the collection
1437 // clears _foregroundGCIsActive
1438 // returns
1439 //
1440 // The background collector in a loop iterating on the phases of the
1441 // collection
1442 // acquires the CGC_lock
1443 // sets _foregroundGCShouldWait
1444 // if _foregroundGCIsActive is set
1445 // clears _foregroundGCShouldWait, notifies _CGC_lock
1446 // waits on _CGC_lock for _foregroundGCIsActive to become false
1447 // and exits the loop.
1448 // otherwise
1449 // proceed with that phase of the collection
1450 // if the phase is a stop-the-world phase,
1451 // yield the baton once more just before enqueueing
1452 // the stop-world CMS operation (executed by the VM thread).
1453 // returns after all phases of the collection are done
1454 //
1455
1456 void CMSCollector::acquire_control_and_collect(bool full,
1457 bool clear_all_soft_refs) {
1458 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1459 assert(!Thread::current()->is_ConcurrentGC_thread(),
1460 "shouldn't try to acquire control from self!");
1461
1462 // Start the protocol for acquiring control of the
1463 // collection from the background collector (aka CMS thread).
1464 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1465 "VM thread should have CMS token");
1466 // Remember the possibly interrupted state of an ongoing
1467 // concurrent collection
1468 CollectorState first_state = _collectorState;
1469
1470 // Signal to a possibly ongoing concurrent collection that
1471 // we want to do a foreground collection.
1472 _foregroundGCIsActive = true;
1473
1474 // release locks and wait for a notify from the background collector
1475 // releasing the locks in only necessary for phases which
1476 // do yields to improve the granularity of the collection.
1477 assert_lock_strong(bitMapLock());
1478 // We need to lock the Free list lock for the space that we are
1479 // currently collecting.
1480 assert(haveFreelistLocks(), "Must be holding free list locks");
1481 bitMapLock()->unlock();
1482 releaseFreelistLocks();
1483 {
1484 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1485 if (_foregroundGCShouldWait) {
1486 // We are going to be waiting for action for the CMS thread;
1487 // it had better not be gone (for instance at shutdown)!
1488 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1489 "CMS thread must be running");
1490 // Wait here until the background collector gives us the go-ahead
1491 ConcurrentMarkSweepThread::clear_CMS_flag(
1492 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1493 // Get a possibly blocked CMS thread going:
1494 // Note that we set _foregroundGCIsActive true above,
1495 // without protection of the CGC_lock.
1496 CGC_lock->notify();
1497 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1498 "Possible deadlock");
1499 while (_foregroundGCShouldWait) {
1500 // wait for notification
1501 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1502 // Possibility of delay/starvation here, since CMS token does
1503 // not know to give priority to VM thread? Actually, i think
1504 // there wouldn't be any delay/starvation, but the proof of
1505 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1506 }
1507 ConcurrentMarkSweepThread::set_CMS_flag(
1508 ConcurrentMarkSweepThread::CMS_vm_has_token);
1509 }
1510 }
1511 // The CMS_token is already held. Get back the other locks.
1512 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1513 "VM thread should have CMS token");
1514 getFreelistLocks();
1515 bitMapLock()->lock_without_safepoint_check();
1516 if (TraceCMSState) {
1517 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1518 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state);
1519 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1520 }
1521
1522 // Inform cms gen if this was due to partial collection failing.
1523 // The CMS gen may use this fact to determine its expansion policy.
1524 GenCollectedHeap* gch = GenCollectedHeap::heap();
1525 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1526 assert(!_cmsGen->incremental_collection_failed(),
1527 "Should have been noticed, reacted to and cleared");
1528 _cmsGen->set_incremental_collection_failed();
1529 }
1530
1531 if (first_state > Idling) {
1532 report_concurrent_mode_interruption();
1533 }
1534
1535 set_did_compact(true);
1536
1537 // If the collection is being acquired from the background
1538 // collector, there may be references on the discovered
1539 // references lists. Abandon those references, since some
1540 // of them may have become unreachable after concurrent
1541 // discovery; the STW compacting collector will redo discovery
1542 // more precisely, without being subject to floating garbage.
1543 // Leaving otherwise unreachable references in the discovered
1544 // lists would require special handling.
1545 ref_processor()->disable_discovery();
1546 ref_processor()->abandon_partial_discovery();
1547 ref_processor()->verify_no_references_recorded();
1548
1549 if (first_state > Idling) {
1550 save_heap_summary();
1551 }
1552
1553 do_compaction_work(clear_all_soft_refs);
1554
1555 // Has the GC time limit been exceeded?
1556 size_t max_eden_size = _young_gen->max_capacity() -
1557 _young_gen->to()->capacity() -
1558 _young_gen->from()->capacity();
1559 GCCause::Cause gc_cause = gch->gc_cause();
1560 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1561 _young_gen->eden()->used(),
1562 _cmsGen->max_capacity(),
1563 max_eden_size,
1564 full,
1565 gc_cause,
1566 gch->collector_policy());
1567
1568 // Reset the expansion cause, now that we just completed
1569 // a collection cycle.
1570 clear_expansion_cause();
1571 _foregroundGCIsActive = false;
1572 return;
1573 }
1574
1575 // Resize the tenured generation
1576 // after obtaining the free list locks for the
1577 // two generations.
1578 void CMSCollector::compute_new_size() {
1579 assert_locked_or_safepoint(Heap_lock);
1580 FreelistLocker z(this);
1581 MetaspaceGC::compute_new_size();
1582 _cmsGen->compute_new_size_free_list();
1583 }
1584
1585 // A work method used by the foreground collector to do
1586 // a mark-sweep-compact.
1587 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1588 GenCollectedHeap* gch = GenCollectedHeap::heap();
1589
1590 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1591 gc_timer->register_gc_start();
1592
1593 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1594 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1595
1596 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id());
1597
1598 // Temporarily widen the span of the weak reference processing to
1599 // the entire heap.
1600 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1601 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1602 // Temporarily, clear the "is_alive_non_header" field of the
1603 // reference processor.
1604 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1605 // Temporarily make reference _processing_ single threaded (non-MT).
1606 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1607 // Temporarily make refs discovery atomic
1608 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1609 // Temporarily make reference _discovery_ single threaded (non-MT)
1610 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1611
1612 ref_processor()->set_enqueuing_is_done(false);
1613 ref_processor()->enable_discovery();
1614 ref_processor()->setup_policy(clear_all_soft_refs);
1615 // If an asynchronous collection finishes, the _modUnionTable is
1616 // all clear. If we are assuming the collection from an asynchronous
1617 // collection, clear the _modUnionTable.
1618 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1619 "_modUnionTable should be clear if the baton was not passed");
1620 _modUnionTable.clear_all();
1621 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
1622 "mod union for klasses should be clear if the baton was passed");
1623 _ct->klass_rem_set()->clear_mod_union();
1624
1625 // We must adjust the allocation statistics being maintained
1626 // in the free list space. We do so by reading and clearing
1627 // the sweep timer and updating the block flux rate estimates below.
1628 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1629 if (_inter_sweep_timer.is_active()) {
1630 _inter_sweep_timer.stop();
1631 // Note that we do not use this sample to update the _inter_sweep_estimate.
1632 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1633 _inter_sweep_estimate.padded_average(),
1634 _intra_sweep_estimate.padded_average());
1635 }
1636
1637 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1638 #ifdef ASSERT
1639 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1640 size_t free_size = cms_space->free();
1641 assert(free_size ==
1642 pointer_delta(cms_space->end(), cms_space->compaction_top())
1643 * HeapWordSize,
1644 "All the free space should be compacted into one chunk at top");
1645 assert(cms_space->dictionary()->total_chunk_size(
1646 debug_only(cms_space->freelistLock())) == 0 ||
1647 cms_space->totalSizeInIndexedFreeLists() == 0,
1648 "All the free space should be in a single chunk");
1649 size_t num = cms_space->totalCount();
1650 assert((free_size == 0 && num == 0) ||
1651 (free_size > 0 && (num == 1 || num == 2)),
1652 "There should be at most 2 free chunks after compaction");
1653 #endif // ASSERT
1654 _collectorState = Resetting;
1655 assert(_restart_addr == NULL,
1656 "Should have been NULL'd before baton was passed");
1657 reset(false /* == !concurrent */);
1658 _cmsGen->reset_after_compaction();
1659 _concurrent_cycles_since_last_unload = 0;
1660
1661 // Clear any data recorded in the PLAB chunk arrays.
1662 if (_survivor_plab_array != NULL) {
1663 reset_survivor_plab_arrays();
1664 }
1665
1666 // Adjust the per-size allocation stats for the next epoch.
1667 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1668 // Restart the "inter sweep timer" for the next epoch.
1669 _inter_sweep_timer.reset();
1670 _inter_sweep_timer.start();
1671
1672 gc_timer->register_gc_end();
1673
1674 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1675
1676 // For a mark-sweep-compact, compute_new_size() will be called
1677 // in the heap's do_collection() method.
1678 }
1679
1680 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1681 ContiguousSpace* eden_space = _young_gen->eden();
1682 ContiguousSpace* from_space = _young_gen->from();
1683 ContiguousSpace* to_space = _young_gen->to();
1684 // Eden
1685 if (_eden_chunk_array != NULL) {
1686 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1687 p2i(eden_space->bottom()), p2i(eden_space->top()),
1688 p2i(eden_space->end()), eden_space->capacity());
1689 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
1690 "_eden_chunk_capacity=" SIZE_FORMAT,
1691 _eden_chunk_index, _eden_chunk_capacity);
1692 for (size_t i = 0; i < _eden_chunk_index; i++) {
1693 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1694 i, p2i(_eden_chunk_array[i]));
1695 }
1696 }
1697 // Survivor
1698 if (_survivor_chunk_array != NULL) {
1699 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1700 p2i(from_space->bottom()), p2i(from_space->top()),
1701 p2i(from_space->end()), from_space->capacity());
1702 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
1703 "_survivor_chunk_capacity=" SIZE_FORMAT,
1704 _survivor_chunk_index, _survivor_chunk_capacity);
1705 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1706 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
1707 i, p2i(_survivor_chunk_array[i]));
1708 }
1709 }
1710 }
1711
1712 void CMSCollector::getFreelistLocks() const {
1713 // Get locks for all free lists in all generations that this
1714 // collector is responsible for
1715 _cmsGen->freelistLock()->lock_without_safepoint_check();
1716 }
1717
1718 void CMSCollector::releaseFreelistLocks() const {
1719 // Release locks for all free lists in all generations that this
1720 // collector is responsible for
1721 _cmsGen->freelistLock()->unlock();
1722 }
1723
1724 bool CMSCollector::haveFreelistLocks() const {
1725 // Check locks for all free lists in all generations that this
1726 // collector is responsible for
1727 assert_lock_strong(_cmsGen->freelistLock());
1728 PRODUCT_ONLY(ShouldNotReachHere());
1729 return true;
1730 }
1731
1732 // A utility class that is used by the CMS collector to
1733 // temporarily "release" the foreground collector from its
1734 // usual obligation to wait for the background collector to
1735 // complete an ongoing phase before proceeding.
1736 class ReleaseForegroundGC: public StackObj {
1737 private:
1738 CMSCollector* _c;
1739 public:
1740 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1741 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1742 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1743 // allow a potentially blocked foreground collector to proceed
1744 _c->_foregroundGCShouldWait = false;
1745 if (_c->_foregroundGCIsActive) {
1746 CGC_lock->notify();
1747 }
1748 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1749 "Possible deadlock");
1750 }
1751
1752 ~ReleaseForegroundGC() {
1753 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1754 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1755 _c->_foregroundGCShouldWait = true;
1756 }
1757 };
1758
1759 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1760 assert(Thread::current()->is_ConcurrentGC_thread(),
1761 "A CMS asynchronous collection is only allowed on a CMS thread.");
1762
1763 GenCollectedHeap* gch = GenCollectedHeap::heap();
1764 {
1765 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1766 MutexLockerEx hl(Heap_lock, safepoint_check);
1767 FreelistLocker fll(this);
1768 MutexLockerEx x(CGC_lock, safepoint_check);
1769 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
1770 // The foreground collector is active or we're
1771 // not using asynchronous collections. Skip this
1772 // background collection.
1773 assert(!_foregroundGCShouldWait, "Should be clear");
1774 return;
1775 } else {
1776 assert(_collectorState == Idling, "Should be idling before start.");
1777 _collectorState = InitialMarking;
1778 register_gc_start(cause);
1779 // Reset the expansion cause, now that we are about to begin
1780 // a new cycle.
1781 clear_expansion_cause();
1782
1783 // Clear the MetaspaceGC flag since a concurrent collection
1784 // is starting but also clear it after the collection.
1785 MetaspaceGC::set_should_concurrent_collect(false);
1786 }
1787 // Decide if we want to enable class unloading as part of the
1788 // ensuing concurrent GC cycle.
1789 update_should_unload_classes();
1790 _full_gc_requested = false; // acks all outstanding full gc requests
1791 _full_gc_cause = GCCause::_no_gc;
1792 // Signal that we are about to start a collection
1793 gch->increment_total_full_collections(); // ... starting a collection cycle
1794 _collection_count_start = gch->total_full_collections();
1795 }
1796
1797 // Used for PrintGC
1798 size_t prev_used;
1799 if (PrintGC && Verbose) {
1800 prev_used = _cmsGen->used();
1801 }
1802
1803 // The change of the collection state is normally done at this level;
1804 // the exceptions are phases that are executed while the world is
1805 // stopped. For those phases the change of state is done while the
1806 // world is stopped. For baton passing purposes this allows the
1807 // background collector to finish the phase and change state atomically.
1808 // The foreground collector cannot wait on a phase that is done
1809 // while the world is stopped because the foreground collector already
1810 // has the world stopped and would deadlock.
1811 while (_collectorState != Idling) {
1812 if (TraceCMSState) {
1813 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
1814 p2i(Thread::current()), _collectorState);
1815 }
1816 // The foreground collector
1817 // holds the Heap_lock throughout its collection.
1818 // holds the CMS token (but not the lock)
1819 // except while it is waiting for the background collector to yield.
1820 //
1821 // The foreground collector should be blocked (not for long)
1822 // if the background collector is about to start a phase
1823 // executed with world stopped. If the background
1824 // collector has already started such a phase, the
1825 // foreground collector is blocked waiting for the
1826 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1827 // are executed in the VM thread.
1828 //
1829 // The locking order is
1830 // PendingListLock (PLL) -- if applicable (FinalMarking)
1831 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1832 // CMS token (claimed in
1833 // stop_world_and_do() -->
1834 // safepoint_synchronize() -->
1835 // CMSThread::synchronize())
1836
1837 {
1838 // Check if the FG collector wants us to yield.
1839 CMSTokenSync x(true); // is cms thread
1840 if (waitForForegroundGC()) {
1841 // We yielded to a foreground GC, nothing more to be
1842 // done this round.
1843 assert(_foregroundGCShouldWait == false, "We set it to false in "
1844 "waitForForegroundGC()");
1845 if (TraceCMSState) {
1846 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1847 " exiting collection CMS state %d",
1848 p2i(Thread::current()), _collectorState);
1849 }
1850 return;
1851 } else {
1852 // The background collector can run but check to see if the
1853 // foreground collector has done a collection while the
1854 // background collector was waiting to get the CGC_lock
1855 // above. If yes, break so that _foregroundGCShouldWait
1856 // is cleared before returning.
1857 if (_collectorState == Idling) {
1858 break;
1859 }
1860 }
1861 }
1862
1863 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1864 "should be waiting");
1865
1866 switch (_collectorState) {
1867 case InitialMarking:
1868 {
1869 ReleaseForegroundGC x(this);
1870 stats().record_cms_begin();
1871 VM_CMS_Initial_Mark initial_mark_op(this);
1872 VMThread::execute(&initial_mark_op);
1873 }
1874 // The collector state may be any legal state at this point
1875 // since the background collector may have yielded to the
1876 // foreground collector.
1877 break;
1878 case Marking:
1879 // initial marking in checkpointRootsInitialWork has been completed
1880 if (markFromRoots()) { // we were successful
1881 assert(_collectorState == Precleaning, "Collector state should "
1882 "have changed");
1883 } else {
1884 assert(_foregroundGCIsActive, "Internal state inconsistency");
1885 }
1886 break;
1887 case Precleaning:
1888 // marking from roots in markFromRoots has been completed
1889 preclean();
1890 assert(_collectorState == AbortablePreclean ||
1891 _collectorState == FinalMarking,
1892 "Collector state should have changed");
1893 break;
1894 case AbortablePreclean:
1895 abortable_preclean();
1896 assert(_collectorState == FinalMarking, "Collector state should "
1897 "have changed");
1898 break;
1899 case FinalMarking:
1900 {
1901 ReleaseForegroundGC x(this);
1902
1903 VM_CMS_Final_Remark final_remark_op(this);
1904 VMThread::execute(&final_remark_op);
1905 }
1906 assert(_foregroundGCShouldWait, "block post-condition");
1907 break;
1908 case Sweeping:
1909 // final marking in checkpointRootsFinal has been completed
1910 sweep();
1911 assert(_collectorState == Resizing, "Collector state change "
1912 "to Resizing must be done under the free_list_lock");
1913
1914 case Resizing: {
1915 // Sweeping has been completed...
1916 // At this point the background collection has completed.
1917 // Don't move the call to compute_new_size() down
1918 // into code that might be executed if the background
1919 // collection was preempted.
1920 {
1921 ReleaseForegroundGC x(this); // unblock FG collection
1922 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1923 CMSTokenSync z(true); // not strictly needed.
1924 if (_collectorState == Resizing) {
1925 compute_new_size();
1926 save_heap_summary();
1927 _collectorState = Resetting;
1928 } else {
1929 assert(_collectorState == Idling, "The state should only change"
1930 " because the foreground collector has finished the collection");
1931 }
1932 }
1933 break;
1934 }
1935 case Resetting:
1936 // CMS heap resizing has been completed
1937 reset(true);
1938 assert(_collectorState == Idling, "Collector state should "
1939 "have changed");
1940
1941 MetaspaceGC::set_should_concurrent_collect(false);
1942
1943 stats().record_cms_end();
1944 // Don't move the concurrent_phases_end() and compute_new_size()
1945 // calls to here because a preempted background collection
1946 // has it's state set to "Resetting".
1947 break;
1948 case Idling:
1949 default:
1950 ShouldNotReachHere();
1951 break;
1952 }
1953 if (TraceCMSState) {
1954 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1955 p2i(Thread::current()), _collectorState);
1956 }
1957 assert(_foregroundGCShouldWait, "block post-condition");
1958 }
1959
1960 // Should this be in gc_epilogue?
1961 collector_policy()->counters()->update_counters();
1962
1963 {
1964 // Clear _foregroundGCShouldWait and, in the event that the
1965 // foreground collector is waiting, notify it, before
1966 // returning.
1967 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1968 _foregroundGCShouldWait = false;
1969 if (_foregroundGCIsActive) {
1970 CGC_lock->notify();
1971 }
1972 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1973 "Possible deadlock");
1974 }
1975 if (TraceCMSState) {
1976 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
1977 " exiting collection CMS state %d",
1978 p2i(Thread::current()), _collectorState);
1979 }
1980 if (PrintGC && Verbose) {
1981 _cmsGen->print_heap_change(prev_used);
1982 }
1983 }
1984
1985 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1986 _cms_start_registered = true;
1987 _gc_timer_cm->register_gc_start();
1988 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1989 }
1990
1991 void CMSCollector::register_gc_end() {
1992 if (_cms_start_registered) {
1993 report_heap_summary(GCWhen::AfterGC);
1994
1995 _gc_timer_cm->register_gc_end();
1996 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1997 _cms_start_registered = false;
1998 }
1999 }
2000
2001 void CMSCollector::save_heap_summary() {
2002 GenCollectedHeap* gch = GenCollectedHeap::heap();
2003 _last_heap_summary = gch->create_heap_summary();
2004 _last_metaspace_summary = gch->create_metaspace_summary();
2005 }
2006
2007 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2008 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
2009 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
2010 }
2011
2012 bool CMSCollector::waitForForegroundGC() {
2013 bool res = false;
2014 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2015 "CMS thread should have CMS token");
2016 // Block the foreground collector until the
2017 // background collectors decides whether to
2018 // yield.
2019 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2020 _foregroundGCShouldWait = true;
2021 if (_foregroundGCIsActive) {
2022 // The background collector yields to the
2023 // foreground collector and returns a value
2024 // indicating that it has yielded. The foreground
2025 // collector can proceed.
2026 res = true;
2027 _foregroundGCShouldWait = false;
2028 ConcurrentMarkSweepThread::clear_CMS_flag(
2029 ConcurrentMarkSweepThread::CMS_cms_has_token);
2030 ConcurrentMarkSweepThread::set_CMS_flag(
2031 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2032 // Get a possibly blocked foreground thread going
2033 CGC_lock->notify();
2034 if (TraceCMSState) {
2035 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2036 p2i(Thread::current()), _collectorState);
2037 }
2038 while (_foregroundGCIsActive) {
2039 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2040 }
2041 ConcurrentMarkSweepThread::set_CMS_flag(
2042 ConcurrentMarkSweepThread::CMS_cms_has_token);
2043 ConcurrentMarkSweepThread::clear_CMS_flag(
2044 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2045 }
2046 if (TraceCMSState) {
2047 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2048 p2i(Thread::current()), _collectorState);
2049 }
2050 return res;
2051 }
2052
2053 // Because of the need to lock the free lists and other structures in
2054 // the collector, common to all the generations that the collector is
2055 // collecting, we need the gc_prologues of individual CMS generations
2056 // delegate to their collector. It may have been simpler had the
2057 // current infrastructure allowed one to call a prologue on a
2058 // collector. In the absence of that we have the generation's
2059 // prologue delegate to the collector, which delegates back
2060 // some "local" work to a worker method in the individual generations
2061 // that it's responsible for collecting, while itself doing any
2062 // work common to all generations it's responsible for. A similar
2063 // comment applies to the gc_epilogue()'s.
2064 // The role of the variable _between_prologue_and_epilogue is to
2065 // enforce the invocation protocol.
2066 void CMSCollector::gc_prologue(bool full) {
2067 // Call gc_prologue_work() for the CMSGen
2068 // we are responsible for.
2069
2070 // The following locking discipline assumes that we are only called
2071 // when the world is stopped.
2072 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2073
2074 // The CMSCollector prologue must call the gc_prologues for the
2075 // "generations" that it's responsible
2076 // for.
2077
2078 assert( Thread::current()->is_VM_thread()
2079 || ( CMSScavengeBeforeRemark
2080 && Thread::current()->is_ConcurrentGC_thread()),
2081 "Incorrect thread type for prologue execution");
2082
2083 if (_between_prologue_and_epilogue) {
2084 // We have already been invoked; this is a gc_prologue delegation
2085 // from yet another CMS generation that we are responsible for, just
2086 // ignore it since all relevant work has already been done.
2087 return;
2088 }
2089
2090 // set a bit saying prologue has been called; cleared in epilogue
2091 _between_prologue_and_epilogue = true;
2092 // Claim locks for common data structures, then call gc_prologue_work()
2093 // for each CMSGen.
2094
2095 getFreelistLocks(); // gets free list locks on constituent spaces
2096 bitMapLock()->lock_without_safepoint_check();
2097
2098 // Should call gc_prologue_work() for all cms gens we are responsible for
2099 bool duringMarking = _collectorState >= Marking
2100 && _collectorState < Sweeping;
2101
2102 // The young collections clear the modified oops state, which tells if
2103 // there are any modified oops in the class. The remark phase also needs
2104 // that information. Tell the young collection to save the union of all
2105 // modified klasses.
2106 if (duringMarking) {
2107 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2108 }
2109
2110 bool registerClosure = duringMarking;
2111
2112 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2113
2114 if (!full) {
2115 stats().record_gc0_begin();
2116 }
2117 }
2118
2119 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2120
2121 _capacity_at_prologue = capacity();
2122 _used_at_prologue = used();
2123
2124 // Delegate to CMScollector which knows how to coordinate between
2125 // this and any other CMS generations that it is responsible for
2126 // collecting.
2127 collector()->gc_prologue(full);
2128 }
2129
2130 // This is a "private" interface for use by this generation's CMSCollector.
2131 // Not to be called directly by any other entity (for instance,
2132 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2133 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2134 bool registerClosure, ModUnionClosure* modUnionClosure) {
2135 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2136 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2137 "Should be NULL");
2138 if (registerClosure) {
2139 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2140 }
2141 cmsSpace()->gc_prologue();
2142 // Clear stat counters
2143 NOT_PRODUCT(
2144 assert(_numObjectsPromoted == 0, "check");
2145 assert(_numWordsPromoted == 0, "check");
2146 if (Verbose && PrintGC) {
2147 gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, "
2148 SIZE_FORMAT " bytes concurrently",
2149 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2150 }
2151 _numObjectsAllocated = 0;
2152 _numWordsAllocated = 0;
2153 )
2154 }
2155
2156 void CMSCollector::gc_epilogue(bool full) {
2157 // The following locking discipline assumes that we are only called
2158 // when the world is stopped.
2159 assert(SafepointSynchronize::is_at_safepoint(),
2160 "world is stopped assumption");
2161
2162 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2163 // if linear allocation blocks need to be appropriately marked to allow the
2164 // the blocks to be parsable. We also check here whether we need to nudge the
2165 // CMS collector thread to start a new cycle (if it's not already active).
2166 assert( Thread::current()->is_VM_thread()
2167 || ( CMSScavengeBeforeRemark
2168 && Thread::current()->is_ConcurrentGC_thread()),
2169 "Incorrect thread type for epilogue execution");
2170
2171 if (!_between_prologue_and_epilogue) {
2172 // We have already been invoked; this is a gc_epilogue delegation
2173 // from yet another CMS generation that we are responsible for, just
2174 // ignore it since all relevant work has already been done.
2175 return;
2176 }
2177 assert(haveFreelistLocks(), "must have freelist locks");
2178 assert_lock_strong(bitMapLock());
2179
2180 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2181
2182 _cmsGen->gc_epilogue_work(full);
2183
2184 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2185 // in case sampling was not already enabled, enable it
2186 _start_sampling = true;
2187 }
2188 // reset _eden_chunk_array so sampling starts afresh
2189 _eden_chunk_index = 0;
2190
2191 size_t cms_used = _cmsGen->cmsSpace()->used();
2192
2193 // update performance counters - this uses a special version of
2194 // update_counters() that allows the utilization to be passed as a
2195 // parameter, avoiding multiple calls to used().
2196 //
2197 _cmsGen->update_counters(cms_used);
2198
2199 bitMapLock()->unlock();
2200 releaseFreelistLocks();
2201
2202 if (!CleanChunkPoolAsync) {
2203 Chunk::clean_chunk_pool();
2204 }
2205
2206 set_did_compact(false);
2207 _between_prologue_and_epilogue = false; // ready for next cycle
2208 }
2209
2210 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2211 collector()->gc_epilogue(full);
2212
2213 // Also reset promotion tracking in par gc thread states.
2214 for (uint i = 0; i < ParallelGCThreads; i++) {
2215 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2216 }
2217 }
2218
2219 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2220 assert(!incremental_collection_failed(), "Should have been cleared");
2221 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2222 cmsSpace()->gc_epilogue();
2223 // Print stat counters
2224 NOT_PRODUCT(
2225 assert(_numObjectsAllocated == 0, "check");
2226 assert(_numWordsAllocated == 0, "check");
2227 if (Verbose && PrintGC) {
2228 gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, "
2229 SIZE_FORMAT " bytes",
2230 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2231 }
2232 _numObjectsPromoted = 0;
2233 _numWordsPromoted = 0;
2234 )
2235
2236 if (PrintGC && Verbose) {
2237 // Call down the chain in contiguous_available needs the freelistLock
2238 // so print this out before releasing the freeListLock.
2239 gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ",
2240 contiguous_available());
2241 }
2242 }
2243
2244 #ifndef PRODUCT
2245 bool CMSCollector::have_cms_token() {
2246 Thread* thr = Thread::current();
2247 if (thr->is_VM_thread()) {
2248 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2249 } else if (thr->is_ConcurrentGC_thread()) {
2250 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2251 } else if (thr->is_GC_task_thread()) {
2252 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2253 ParGCRareEvent_lock->owned_by_self();
2254 }
2255 return false;
2256 }
2257 #endif
2258
2259 // Check reachability of the given heap address in CMS generation,
2260 // treating all other generations as roots.
2261 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2262 // We could "guarantee" below, rather than assert, but I'll
2263 // leave these as "asserts" so that an adventurous debugger
2264 // could try this in the product build provided some subset of
2265 // the conditions were met, provided they were interested in the
2266 // results and knew that the computation below wouldn't interfere
2267 // with other concurrent computations mutating the structures
2268 // being read or written.
2269 assert(SafepointSynchronize::is_at_safepoint(),
2270 "Else mutations in object graph will make answer suspect");
2271 assert(have_cms_token(), "Should hold cms token");
2272 assert(haveFreelistLocks(), "must hold free list locks");
2273 assert_lock_strong(bitMapLock());
2274
2275 // Clear the marking bit map array before starting, but, just
2276 // for kicks, first report if the given address is already marked
2277 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2278 _markBitMap.isMarked(addr) ? "" : " not");
2279
2280 if (verify_after_remark()) {
2281 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2282 bool result = verification_mark_bm()->isMarked(addr);
2283 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2284 result ? "IS" : "is NOT");
2285 return result;
2286 } else {
2287 gclog_or_tty->print_cr("Could not compute result");
2288 return false;
2289 }
2290 }
2291
2292
2293 void
2294 CMSCollector::print_on_error(outputStream* st) {
2295 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2296 if (collector != NULL) {
2297 CMSBitMap* bitmap = &collector->_markBitMap;
2298 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2299 bitmap->print_on_error(st, " Bits: ");
2300
2301 st->cr();
2302
2303 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2304 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2305 mut_bitmap->print_on_error(st, " Bits: ");
2306 }
2307 }
2308
2309 ////////////////////////////////////////////////////////
2310 // CMS Verification Support
2311 ////////////////////////////////////////////////////////
2312 // Following the remark phase, the following invariant
2313 // should hold -- each object in the CMS heap which is
2314 // marked in markBitMap() should be marked in the verification_mark_bm().
2315
2316 class VerifyMarkedClosure: public BitMapClosure {
2317 CMSBitMap* _marks;
2318 bool _failed;
2319
2320 public:
2321 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2322
2323 bool do_bit(size_t offset) {
2324 HeapWord* addr = _marks->offsetToHeapWord(offset);
2325 if (!_marks->isMarked(addr)) {
2326 oop(addr)->print_on(gclog_or_tty);
2327 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2328 _failed = true;
2329 }
2330 return true;
2331 }
2332
2333 bool failed() { return _failed; }
2334 };
2335
2336 bool CMSCollector::verify_after_remark(bool silent) {
2337 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2338 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2339 static bool init = false;
2340
2341 assert(SafepointSynchronize::is_at_safepoint(),
2342 "Else mutations in object graph will make answer suspect");
2343 assert(have_cms_token(),
2344 "Else there may be mutual interference in use of "
2345 " verification data structures");
2346 assert(_collectorState > Marking && _collectorState <= Sweeping,
2347 "Else marking info checked here may be obsolete");
2348 assert(haveFreelistLocks(), "must hold free list locks");
2349 assert_lock_strong(bitMapLock());
2350
2351
2352 // Allocate marking bit map if not already allocated
2353 if (!init) { // first time
2354 if (!verification_mark_bm()->allocate(_span)) {
2355 return false;
2356 }
2357 init = true;
2358 }
2359
2360 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2361
2362 // Turn off refs discovery -- so we will be tracing through refs.
2363 // This is as intended, because by this time
2364 // GC must already have cleared any refs that need to be cleared,
2365 // and traced those that need to be marked; moreover,
2366 // the marking done here is not going to interfere in any
2367 // way with the marking information used by GC.
2368 NoRefDiscovery no_discovery(ref_processor());
2369
2370 #if defined(COMPILER2) || INCLUDE_JVMCI
2371 DerivedPointerTableDeactivate dpt_deact;
2372 #endif
2373
2374 // Clear any marks from a previous round
2375 verification_mark_bm()->clear_all();
2376 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2377 verify_work_stacks_empty();
2378
2379 GenCollectedHeap* gch = GenCollectedHeap::heap();
2380 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2381 // Update the saved marks which may affect the root scans.
2382 gch->save_marks();
2383
2384 if (CMSRemarkVerifyVariant == 1) {
2385 // In this first variant of verification, we complete
2386 // all marking, then check if the new marks-vector is
2387 // a subset of the CMS marks-vector.
2388 verify_after_remark_work_1();
2389 } else if (CMSRemarkVerifyVariant == 2) {
2390 // In this second variant of verification, we flag an error
2391 // (i.e. an object reachable in the new marks-vector not reachable
2392 // in the CMS marks-vector) immediately, also indicating the
2393 // identify of an object (A) that references the unmarked object (B) --
2394 // presumably, a mutation to A failed to be picked up by preclean/remark?
2395 verify_after_remark_work_2();
2396 } else {
2397 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant",
2398 CMSRemarkVerifyVariant);
2399 }
2400 if (!silent) gclog_or_tty->print(" done] ");
2401 return true;
2402 }
2403
2404 void CMSCollector::verify_after_remark_work_1() {
2405 ResourceMark rm;
2406 HandleMark hm;
2407 GenCollectedHeap* gch = GenCollectedHeap::heap();
2408
2409 // Get a clear set of claim bits for the roots processing to work with.
2410 ClassLoaderDataGraph::clear_claimed_marks();
2411
2412 // Mark from roots one level into CMS
2413 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2414 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2415
2416 {
2417 StrongRootsScope srs(1);
2418
2419 gch->gen_process_roots(&srs,
2420 GenCollectedHeap::OldGen,
2421 true, // young gen as roots
2422 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2423 should_unload_classes(),
2424 ¬Older,
2425 NULL,
2426 NULL);
2427 }
2428
2429 // Now mark from the roots
2430 MarkFromRootsClosure markFromRootsClosure(this, _span,
2431 verification_mark_bm(), verification_mark_stack(),
2432 false /* don't yield */, true /* verifying */);
2433 assert(_restart_addr == NULL, "Expected pre-condition");
2434 verification_mark_bm()->iterate(&markFromRootsClosure);
2435 while (_restart_addr != NULL) {
2436 // Deal with stack overflow: by restarting at the indicated
2437 // address.
2438 HeapWord* ra = _restart_addr;
2439 markFromRootsClosure.reset(ra);
2440 _restart_addr = NULL;
2441 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2442 }
2443 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2444 verify_work_stacks_empty();
2445
2446 // Marking completed -- now verify that each bit marked in
2447 // verification_mark_bm() is also marked in markBitMap(); flag all
2448 // errors by printing corresponding objects.
2449 VerifyMarkedClosure vcl(markBitMap());
2450 verification_mark_bm()->iterate(&vcl);
2451 if (vcl.failed()) {
2452 gclog_or_tty->print("Verification failed");
2453 gch->print_on(gclog_or_tty);
2454 fatal("CMS: failed marking verification after remark");
2455 }
2456 }
2457
2458 class VerifyKlassOopsKlassClosure : public KlassClosure {
2459 class VerifyKlassOopsClosure : public OopClosure {
2460 CMSBitMap* _bitmap;
2461 public:
2462 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2463 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2464 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2465 } _oop_closure;
2466 public:
2467 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2468 void do_klass(Klass* k) {
2469 k->oops_do(&_oop_closure);
2470 }
2471 };
2472
2473 void CMSCollector::verify_after_remark_work_2() {
2474 ResourceMark rm;
2475 HandleMark hm;
2476 GenCollectedHeap* gch = GenCollectedHeap::heap();
2477
2478 // Get a clear set of claim bits for the roots processing to work with.
2479 ClassLoaderDataGraph::clear_claimed_marks();
2480
2481 // Mark from roots one level into CMS
2482 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2483 markBitMap());
2484 CLDToOopClosure cld_closure(¬Older, true);
2485
2486 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2487
2488 {
2489 StrongRootsScope srs(1);
2490
2491 gch->gen_process_roots(&srs,
2492 GenCollectedHeap::OldGen,
2493 true, // young gen as roots
2494 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2495 should_unload_classes(),
2496 ¬Older,
2497 NULL,
2498 &cld_closure);
2499 }
2500
2501 // Now mark from the roots
2502 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2503 verification_mark_bm(), markBitMap(), verification_mark_stack());
2504 assert(_restart_addr == NULL, "Expected pre-condition");
2505 verification_mark_bm()->iterate(&markFromRootsClosure);
2506 while (_restart_addr != NULL) {
2507 // Deal with stack overflow: by restarting at the indicated
2508 // address.
2509 HeapWord* ra = _restart_addr;
2510 markFromRootsClosure.reset(ra);
2511 _restart_addr = NULL;
2512 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2513 }
2514 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2515 verify_work_stacks_empty();
2516
2517 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
2518 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
2519
2520 // Marking completed -- now verify that each bit marked in
2521 // verification_mark_bm() is also marked in markBitMap(); flag all
2522 // errors by printing corresponding objects.
2523 VerifyMarkedClosure vcl(markBitMap());
2524 verification_mark_bm()->iterate(&vcl);
2525 assert(!vcl.failed(), "Else verification above should not have succeeded");
2526 }
2527
2528 void ConcurrentMarkSweepGeneration::save_marks() {
2529 // delegate to CMS space
2530 cmsSpace()->save_marks();
2531 for (uint i = 0; i < ParallelGCThreads; i++) {
2532 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2533 }
2534 }
2535
2536 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2537 return cmsSpace()->no_allocs_since_save_marks();
2538 }
2539
2540 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2541 \
2542 void ConcurrentMarkSweepGeneration:: \
2543 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2544 cl->set_generation(this); \
2545 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2546 cl->reset_generation(); \
2547 save_marks(); \
2548 }
2549
2550 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2551
2552 void
2553 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2554 if (freelistLock()->owned_by_self()) {
2555 Generation::oop_iterate(cl);
2556 } else {
2557 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2558 Generation::oop_iterate(cl);
2559 }
2560 }
2561
2562 void
2563 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2564 if (freelistLock()->owned_by_self()) {
2565 Generation::object_iterate(cl);
2566 } else {
2567 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2568 Generation::object_iterate(cl);
2569 }
2570 }
2571
2572 void
2573 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2574 if (freelistLock()->owned_by_self()) {
2575 Generation::safe_object_iterate(cl);
2576 } else {
2577 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2578 Generation::safe_object_iterate(cl);
2579 }
2580 }
2581
2582 void
2583 ConcurrentMarkSweepGeneration::post_compact() {
2584 }
2585
2586 void
2587 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2588 // Fix the linear allocation blocks to look like free blocks.
2589
2590 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2591 // are not called when the heap is verified during universe initialization and
2592 // at vm shutdown.
2593 if (freelistLock()->owned_by_self()) {
2594 cmsSpace()->prepare_for_verify();
2595 } else {
2596 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2597 cmsSpace()->prepare_for_verify();
2598 }
2599 }
2600
2601 void
2602 ConcurrentMarkSweepGeneration::verify() {
2603 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2604 // are not called when the heap is verified during universe initialization and
2605 // at vm shutdown.
2606 if (freelistLock()->owned_by_self()) {
2607 cmsSpace()->verify();
2608 } else {
2609 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2610 cmsSpace()->verify();
2611 }
2612 }
2613
2614 void CMSCollector::verify() {
2615 _cmsGen->verify();
2616 }
2617
2618 #ifndef PRODUCT
2619 bool CMSCollector::overflow_list_is_empty() const {
2620 assert(_num_par_pushes >= 0, "Inconsistency");
2621 if (_overflow_list == NULL) {
2622 assert(_num_par_pushes == 0, "Inconsistency");
2623 }
2624 return _overflow_list == NULL;
2625 }
2626
2627 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2628 // merely consolidate assertion checks that appear to occur together frequently.
2629 void CMSCollector::verify_work_stacks_empty() const {
2630 assert(_markStack.isEmpty(), "Marking stack should be empty");
2631 assert(overflow_list_is_empty(), "Overflow list should be empty");
2632 }
2633
2634 void CMSCollector::verify_overflow_empty() const {
2635 assert(overflow_list_is_empty(), "Overflow list should be empty");
2636 assert(no_preserved_marks(), "No preserved marks");
2637 }
2638 #endif // PRODUCT
2639
2640 // Decide if we want to enable class unloading as part of the
2641 // ensuing concurrent GC cycle. We will collect and
2642 // unload classes if it's the case that:
2643 // (1) an explicit gc request has been made and the flag
2644 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
2645 // (2) (a) class unloading is enabled at the command line, and
2646 // (b) old gen is getting really full
2647 // NOTE: Provided there is no change in the state of the heap between
2648 // calls to this method, it should have idempotent results. Moreover,
2649 // its results should be monotonically increasing (i.e. going from 0 to 1,
2650 // but not 1 to 0) between successive calls between which the heap was
2651 // not collected. For the implementation below, it must thus rely on
2652 // the property that concurrent_cycles_since_last_unload()
2653 // will not decrease unless a collection cycle happened and that
2654 // _cmsGen->is_too_full() are
2655 // themselves also monotonic in that sense. See check_monotonicity()
2656 // below.
2657 void CMSCollector::update_should_unload_classes() {
2658 _should_unload_classes = false;
2659 // Condition 1 above
2660 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
2661 _should_unload_classes = true;
2662 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
2663 // Disjuncts 2.b.(i,ii,iii) above
2664 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2665 CMSClassUnloadingMaxInterval)
2666 || _cmsGen->is_too_full();
2667 }
2668 }
2669
2670 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2671 bool res = should_concurrent_collect();
2672 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2673 return res;
2674 }
2675
2676 void CMSCollector::setup_cms_unloading_and_verification_state() {
2677 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2678 || VerifyBeforeExit;
2679 const int rso = GenCollectedHeap::SO_AllCodeCache;
2680
2681 // We set the proper root for this CMS cycle here.
2682 if (should_unload_classes()) { // Should unload classes this cycle
2683 remove_root_scanning_option(rso); // Shrink the root set appropriately
2684 set_verifying(should_verify); // Set verification state for this cycle
2685 return; // Nothing else needs to be done at this time
2686 }
2687
2688 // Not unloading classes this cycle
2689 assert(!should_unload_classes(), "Inconsistency!");
2690
2691 // If we are not unloading classes then add SO_AllCodeCache to root
2692 // scanning options.
2693 add_root_scanning_option(rso);
2694
2695 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2696 set_verifying(true);
2697 } else if (verifying() && !should_verify) {
2698 // We were verifying, but some verification flags got disabled.
2699 set_verifying(false);
2700 // Exclude symbols, strings and code cache elements from root scanning to
2701 // reduce IM and RM pauses.
2702 remove_root_scanning_option(rso);
2703 }
2704 }
2705
2706
2707 #ifndef PRODUCT
2708 HeapWord* CMSCollector::block_start(const void* p) const {
2709 const HeapWord* addr = (HeapWord*)p;
2710 if (_span.contains(p)) {
2711 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2712 return _cmsGen->cmsSpace()->block_start(p);
2713 }
2714 }
2715 return NULL;
2716 }
2717 #endif
2718
2719 HeapWord*
2720 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2721 bool tlab,
2722 bool parallel) {
2723 CMSSynchronousYieldRequest yr;
2724 assert(!tlab, "Can't deal with TLAB allocation");
2725 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2726 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2727 if (GCExpandToAllocateDelayMillis > 0) {
2728 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2729 }
2730 return have_lock_and_allocate(word_size, tlab);
2731 }
2732
2733 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2734 size_t bytes,
2735 size_t expand_bytes,
2736 CMSExpansionCause::Cause cause)
2737 {
2738
2739 bool success = expand(bytes, expand_bytes);
2740
2741 // remember why we expanded; this information is used
2742 // by shouldConcurrentCollect() when making decisions on whether to start
2743 // a new CMS cycle.
2744 if (success) {
2745 set_expansion_cause(cause);
2746 if (PrintGCDetails && Verbose) {
2747 gclog_or_tty->print_cr("Expanded CMS gen for %s",
2748 CMSExpansionCause::to_string(cause));
2749 }
2750 }
2751 }
2752
2753 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2754 HeapWord* res = NULL;
2755 MutexLocker x(ParGCRareEvent_lock);
2756 while (true) {
2757 // Expansion by some other thread might make alloc OK now:
2758 res = ps->lab.alloc(word_sz);
2759 if (res != NULL) return res;
2760 // If there's not enough expansion space available, give up.
2761 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2762 return NULL;
2763 }
2764 // Otherwise, we try expansion.
2765 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2766 // Now go around the loop and try alloc again;
2767 // A competing par_promote might beat us to the expansion space,
2768 // so we may go around the loop again if promotion fails again.
2769 if (GCExpandToAllocateDelayMillis > 0) {
2770 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2771 }
2772 }
2773 }
2774
2775
2776 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2777 PromotionInfo* promo) {
2778 MutexLocker x(ParGCRareEvent_lock);
2779 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2780 while (true) {
2781 // Expansion by some other thread might make alloc OK now:
2782 if (promo->ensure_spooling_space()) {
2783 assert(promo->has_spooling_space(),
2784 "Post-condition of successful ensure_spooling_space()");
2785 return true;
2786 }
2787 // If there's not enough expansion space available, give up.
2788 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2789 return false;
2790 }
2791 // Otherwise, we try expansion.
2792 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2793 // Now go around the loop and try alloc again;
2794 // A competing allocation might beat us to the expansion space,
2795 // so we may go around the loop again if allocation fails again.
2796 if (GCExpandToAllocateDelayMillis > 0) {
2797 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2798 }
2799 }
2800 }
2801
2802 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2803 // Only shrink if a compaction was done so that all the free space
2804 // in the generation is in a contiguous block at the end.
2805 if (did_compact()) {
2806 CardGeneration::shrink(bytes);
2807 }
2808 }
2809
2810 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2811 assert_locked_or_safepoint(Heap_lock);
2812 }
2813
2814 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2815 assert_locked_or_safepoint(Heap_lock);
2816 assert_lock_strong(freelistLock());
2817 if (PrintGCDetails && Verbose) {
2818 warning("Shrinking of CMS not yet implemented");
2819 }
2820 return;
2821 }
2822
2823
2824 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2825 // phases.
2826 class CMSPhaseAccounting: public StackObj {
2827 public:
2828 CMSPhaseAccounting(CMSCollector *collector,
2829 const char *phase,
2830 const GCId gc_id,
2831 bool print_cr = true);
2832 ~CMSPhaseAccounting();
2833
2834 private:
2835 CMSCollector *_collector;
2836 const char *_phase;
2837 elapsedTimer _wallclock;
2838 bool _print_cr;
2839 const GCId _gc_id;
2840
2841 public:
2842 // Not MT-safe; so do not pass around these StackObj's
2843 // where they may be accessed by other threads.
2844 jlong wallclock_millis() {
2845 assert(_wallclock.is_active(), "Wall clock should not stop");
2846 _wallclock.stop(); // to record time
2847 jlong ret = _wallclock.milliseconds();
2848 _wallclock.start(); // restart
2849 return ret;
2850 }
2851 };
2852
2853 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2854 const char *phase,
2855 const GCId gc_id,
2856 bool print_cr) :
2857 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) {
2858
2859 if (PrintCMSStatistics != 0) {
2860 _collector->resetYields();
2861 }
2862 if (PrintGCDetails) {
2863 gclog_or_tty->gclog_stamp(_gc_id);
2864 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
2865 _collector->cmsGen()->short_name(), _phase);
2866 }
2867 _collector->resetTimer();
2868 _wallclock.start();
2869 _collector->startTimer();
2870 }
2871
2872 CMSPhaseAccounting::~CMSPhaseAccounting() {
2873 assert(_wallclock.is_active(), "Wall clock should not have stopped");
2874 _collector->stopTimer();
2875 _wallclock.stop();
2876 if (PrintGCDetails) {
2877 gclog_or_tty->gclog_stamp(_gc_id);
2878 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
2879 _collector->cmsGen()->short_name(),
2880 _phase, _collector->timerValue(), _wallclock.seconds());
2881 if (_print_cr) {
2882 gclog_or_tty->cr();
2883 }
2884 if (PrintCMSStatistics != 0) {
2885 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
2886 _collector->yields());
2887 }
2888 }
2889 }
2890
2891 // CMS work
2892
2893 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2894 class CMSParMarkTask : public AbstractGangTask {
2895 protected:
2896 CMSCollector* _collector;
2897 uint _n_workers;
2898 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2899 AbstractGangTask(name),
2900 _collector(collector),
2901 _n_workers(n_workers) {}
2902 // Work method in support of parallel rescan ... of young gen spaces
2903 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
2904 ContiguousSpace* space,
2905 HeapWord** chunk_array, size_t chunk_top);
2906 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
2907 };
2908
2909 // Parallel initial mark task
2910 class CMSParInitialMarkTask: public CMSParMarkTask {
2911 StrongRootsScope* _strong_roots_scope;
2912 public:
2913 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2914 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2915 _strong_roots_scope(strong_roots_scope) {}
2916 void work(uint worker_id);
2917 };
2918
2919 // Checkpoint the roots into this generation from outside
2920 // this generation. [Note this initial checkpoint need only
2921 // be approximate -- we'll do a catch up phase subsequently.]
2922 void CMSCollector::checkpointRootsInitial() {
2923 assert(_collectorState == InitialMarking, "Wrong collector state");
2924 check_correct_thread_executing();
2925 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
2926
2927 save_heap_summary();
2928 report_heap_summary(GCWhen::BeforeGC);
2929
2930 ReferenceProcessor* rp = ref_processor();
2931 assert(_restart_addr == NULL, "Control point invariant");
2932 {
2933 // acquire locks for subsequent manipulations
2934 MutexLockerEx x(bitMapLock(),
2935 Mutex::_no_safepoint_check_flag);
2936 checkpointRootsInitialWork();
2937 // enable ("weak") refs discovery
2938 rp->enable_discovery();
2939 _collectorState = Marking;
2940 }
2941 }
2942
2943 void CMSCollector::checkpointRootsInitialWork() {
2944 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2945 assert(_collectorState == InitialMarking, "just checking");
2946
2947 // Already have locks.
2948 assert_lock_strong(bitMapLock());
2949 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2950
2951 // Setup the verification and class unloading state for this
2952 // CMS collection cycle.
2953 setup_cms_unloading_and_verification_state();
2954
2955 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
2956 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
2957
2958 // Reset all the PLAB chunk arrays if necessary.
2959 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2960 reset_survivor_plab_arrays();
2961 }
2962
2963 ResourceMark rm;
2964 HandleMark hm;
2965
2966 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2967 GenCollectedHeap* gch = GenCollectedHeap::heap();
2968
2969 verify_work_stacks_empty();
2970 verify_overflow_empty();
2971
2972 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2973 // Update the saved marks which may affect the root scans.
2974 gch->save_marks();
2975
2976 // weak reference processing has not started yet.
2977 ref_processor()->set_enqueuing_is_done(false);
2978
2979 // Need to remember all newly created CLDs,
2980 // so that we can guarantee that the remark finds them.
2981 ClassLoaderDataGraph::remember_new_clds(true);
2982
2983 // Whenever a CLD is found, it will be claimed before proceeding to mark
2984 // the klasses. The claimed marks need to be cleared before marking starts.
2985 ClassLoaderDataGraph::clear_claimed_marks();
2986
2987 if (CMSPrintEdenSurvivorChunks) {
2988 print_eden_and_survivor_chunk_arrays();
2989 }
2990
2991 {
2992 #if defined(COMPILER2) || INCLUDE_JVMCI
2993 DerivedPointerTableDeactivate dpt_deact;
2994 #endif
2995 if (CMSParallelInitialMarkEnabled) {
2996 // The parallel version.
2997 WorkGang* workers = gch->workers();
2998 assert(workers != NULL, "Need parallel worker threads.");
2999 uint n_workers = workers->active_workers();
3000
3001 StrongRootsScope srs(n_workers);
3002
3003 CMSParInitialMarkTask tsk(this, &srs, n_workers);
3004 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3005 if (n_workers > 1) {
3006 workers->run_task(&tsk);
3007 } else {
3008 tsk.work(0);
3009 }
3010 } else {
3011 // The serial version.
3012 CLDToOopClosure cld_closure(¬Older, true);
3013 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3014
3015 StrongRootsScope srs(1);
3016
3017 gch->gen_process_roots(&srs,
3018 GenCollectedHeap::OldGen,
3019 true, // young gen as roots
3020 GenCollectedHeap::ScanningOption(roots_scanning_options()),
3021 should_unload_classes(),
3022 ¬Older,
3023 NULL,
3024 &cld_closure);
3025 }
3026 }
3027
3028 // Clear mod-union table; it will be dirtied in the prologue of
3029 // CMS generation per each young generation collection.
3030
3031 assert(_modUnionTable.isAllClear(),
3032 "Was cleared in most recent final checkpoint phase"
3033 " or no bits are set in the gc_prologue before the start of the next "
3034 "subsequent marking phase.");
3035
3036 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3037
3038 // Save the end of the used_region of the constituent generations
3039 // to be used to limit the extent of sweep in each generation.
3040 save_sweep_limits();
3041 verify_overflow_empty();
3042 }
3043
3044 bool CMSCollector::markFromRoots() {
3045 // we might be tempted to assert that:
3046 // assert(!SafepointSynchronize::is_at_safepoint(),
3047 // "inconsistent argument?");
3048 // However that wouldn't be right, because it's possible that
3049 // a safepoint is indeed in progress as a young generation
3050 // stop-the-world GC happens even as we mark in this generation.
3051 assert(_collectorState == Marking, "inconsistent state?");
3052 check_correct_thread_executing();
3053 verify_overflow_empty();
3054
3055 // Weak ref discovery note: We may be discovering weak
3056 // refs in this generation concurrent (but interleaved) with
3057 // weak ref discovery by the young generation collector.
3058
3059 CMSTokenSyncWithLocks ts(true, bitMapLock());
3060 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3061 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3062 bool res = markFromRootsWork();
3063 if (res) {
3064 _collectorState = Precleaning;
3065 } else { // We failed and a foreground collection wants to take over
3066 assert(_foregroundGCIsActive, "internal state inconsistency");
3067 assert(_restart_addr == NULL, "foreground will restart from scratch");
3068 if (PrintGCDetails) {
3069 gclog_or_tty->print_cr("bailing out to foreground collection");
3070 }
3071 }
3072 verify_overflow_empty();
3073 return res;
3074 }
3075
3076 bool CMSCollector::markFromRootsWork() {
3077 // iterate over marked bits in bit map, doing a full scan and mark
3078 // from these roots using the following algorithm:
3079 // . if oop is to the right of the current scan pointer,
3080 // mark corresponding bit (we'll process it later)
3081 // . else (oop is to left of current scan pointer)
3082 // push oop on marking stack
3083 // . drain the marking stack
3084
3085 // Note that when we do a marking step we need to hold the
3086 // bit map lock -- recall that direct allocation (by mutators)
3087 // and promotion (by the young generation collector) is also
3088 // marking the bit map. [the so-called allocate live policy.]
3089 // Because the implementation of bit map marking is not
3090 // robust wrt simultaneous marking of bits in the same word,
3091 // we need to make sure that there is no such interference
3092 // between concurrent such updates.
3093
3094 // already have locks
3095 assert_lock_strong(bitMapLock());
3096
3097 verify_work_stacks_empty();
3098 verify_overflow_empty();
3099 bool result = false;
3100 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3101 result = do_marking_mt();
3102 } else {
3103 result = do_marking_st();
3104 }
3105 return result;
3106 }
3107
3108 // Forward decl
3109 class CMSConcMarkingTask;
3110
3111 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3112 CMSCollector* _collector;
3113 CMSConcMarkingTask* _task;
3114 public:
3115 virtual void yield();
3116
3117 // "n_threads" is the number of threads to be terminated.
3118 // "queue_set" is a set of work queues of other threads.
3119 // "collector" is the CMS collector associated with this task terminator.
3120 // "yield" indicates whether we need the gang as a whole to yield.
3121 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3122 ParallelTaskTerminator(n_threads, queue_set),
3123 _collector(collector) { }
3124
3125 void set_task(CMSConcMarkingTask* task) {
3126 _task = task;
3127 }
3128 };
3129
3130 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3131 CMSConcMarkingTask* _task;
3132 public:
3133 bool should_exit_termination();
3134 void set_task(CMSConcMarkingTask* task) {
3135 _task = task;
3136 }
3137 };
3138
3139 // MT Concurrent Marking Task
3140 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3141 CMSCollector* _collector;
3142 uint _n_workers; // requested/desired # workers
3143 bool _result;
3144 CompactibleFreeListSpace* _cms_space;
3145 char _pad_front[64]; // padding to ...
3146 HeapWord* _global_finger; // ... avoid sharing cache line
3147 char _pad_back[64];
3148 HeapWord* _restart_addr;
3149
3150 // Exposed here for yielding support
3151 Mutex* const _bit_map_lock;
3152
3153 // The per thread work queues, available here for stealing
3154 OopTaskQueueSet* _task_queues;
3155
3156 // Termination (and yielding) support
3157 CMSConcMarkingTerminator _term;
3158 CMSConcMarkingTerminatorTerminator _term_term;
3159
3160 public:
3161 CMSConcMarkingTask(CMSCollector* collector,
3162 CompactibleFreeListSpace* cms_space,
3163 YieldingFlexibleWorkGang* workers,
3164 OopTaskQueueSet* task_queues):
3165 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3166 _collector(collector),
3167 _cms_space(cms_space),
3168 _n_workers(0), _result(true),
3169 _task_queues(task_queues),
3170 _term(_n_workers, task_queues, _collector),
3171 _bit_map_lock(collector->bitMapLock())
3172 {
3173 _requested_size = _n_workers;
3174 _term.set_task(this);
3175 _term_term.set_task(this);
3176 _restart_addr = _global_finger = _cms_space->bottom();
3177 }
3178
3179
3180 OopTaskQueueSet* task_queues() { return _task_queues; }
3181
3182 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3183
3184 HeapWord** global_finger_addr() { return &_global_finger; }
3185
3186 CMSConcMarkingTerminator* terminator() { return &_term; }
3187
3188 virtual void set_for_termination(uint active_workers) {
3189 terminator()->reset_for_reuse(active_workers);
3190 }
3191
3192 void work(uint worker_id);
3193 bool should_yield() {
3194 return ConcurrentMarkSweepThread::should_yield()
3195 && !_collector->foregroundGCIsActive();
3196 }
3197
3198 virtual void coordinator_yield(); // stuff done by coordinator
3199 bool result() { return _result; }
3200
3201 void reset(HeapWord* ra) {
3202 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3203 _restart_addr = _global_finger = ra;
3204 _term.reset_for_reuse();
3205 }
3206
3207 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3208 OopTaskQueue* work_q);
3209
3210 private:
3211 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3212 void do_work_steal(int i);
3213 void bump_global_finger(HeapWord* f);
3214 };
3215
3216 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3217 assert(_task != NULL, "Error");
3218 return _task->yielding();
3219 // Note that we do not need the disjunct || _task->should_yield() above
3220 // because we want terminating threads to yield only if the task
3221 // is already in the midst of yielding, which happens only after at least one
3222 // thread has yielded.
3223 }
3224
3225 void CMSConcMarkingTerminator::yield() {
3226 if (_task->should_yield()) {
3227 _task->yield();
3228 } else {
3229 ParallelTaskTerminator::yield();
3230 }
3231 }
3232
3233 ////////////////////////////////////////////////////////////////
3234 // Concurrent Marking Algorithm Sketch
3235 ////////////////////////////////////////////////////////////////
3236 // Until all tasks exhausted (both spaces):
3237 // -- claim next available chunk
3238 // -- bump global finger via CAS
3239 // -- find first object that starts in this chunk
3240 // and start scanning bitmap from that position
3241 // -- scan marked objects for oops
3242 // -- CAS-mark target, and if successful:
3243 // . if target oop is above global finger (volatile read)
3244 // nothing to do
3245 // . if target oop is in chunk and above local finger
3246 // then nothing to do
3247 // . else push on work-queue
3248 // -- Deal with possible overflow issues:
3249 // . local work-queue overflow causes stuff to be pushed on
3250 // global (common) overflow queue
3251 // . always first empty local work queue
3252 // . then get a batch of oops from global work queue if any
3253 // . then do work stealing
3254 // -- When all tasks claimed (both spaces)
3255 // and local work queue empty,
3256 // then in a loop do:
3257 // . check global overflow stack; steal a batch of oops and trace
3258 // . try to steal from other threads oif GOS is empty
3259 // . if neither is available, offer termination
3260 // -- Terminate and return result
3261 //
3262 void CMSConcMarkingTask::work(uint worker_id) {
3263 elapsedTimer _timer;
3264 ResourceMark rm;
3265 HandleMark hm;
3266
3267 DEBUG_ONLY(_collector->verify_overflow_empty();)
3268
3269 // Before we begin work, our work queue should be empty
3270 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3271 // Scan the bitmap covering _cms_space, tracing through grey objects.
3272 _timer.start();
3273 do_scan_and_mark(worker_id, _cms_space);
3274 _timer.stop();
3275 if (PrintCMSStatistics != 0) {
3276 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3277 worker_id, _timer.seconds());
3278 // XXX: need xxx/xxx type of notation, two timers
3279 }
3280
3281 // ... do work stealing
3282 _timer.reset();
3283 _timer.start();
3284 do_work_steal(worker_id);
3285 _timer.stop();
3286 if (PrintCMSStatistics != 0) {
3287 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3288 worker_id, _timer.seconds());
3289 // XXX: need xxx/xxx type of notation, two timers
3290 }
3291 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3292 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3293 // Note that under the current task protocol, the
3294 // following assertion is true even of the spaces
3295 // expanded since the completion of the concurrent
3296 // marking. XXX This will likely change under a strict
3297 // ABORT semantics.
3298 // After perm removal the comparison was changed to
3299 // greater than or equal to from strictly greater than.
3300 // Before perm removal the highest address sweep would
3301 // have been at the end of perm gen but now is at the
3302 // end of the tenured gen.
3303 assert(_global_finger >= _cms_space->end(),
3304 "All tasks have been completed");
3305 DEBUG_ONLY(_collector->verify_overflow_empty();)
3306 }
3307
3308 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3309 HeapWord* read = _global_finger;
3310 HeapWord* cur = read;
3311 while (f > read) {
3312 cur = read;
3313 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3314 if (cur == read) {
3315 // our cas succeeded
3316 assert(_global_finger >= f, "protocol consistency");
3317 break;
3318 }
3319 }
3320 }
3321
3322 // This is really inefficient, and should be redone by
3323 // using (not yet available) block-read and -write interfaces to the
3324 // stack and the work_queue. XXX FIX ME !!!
3325 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3326 OopTaskQueue* work_q) {
3327 // Fast lock-free check
3328 if (ovflw_stk->length() == 0) {
3329 return false;
3330 }
3331 assert(work_q->size() == 0, "Shouldn't steal");
3332 MutexLockerEx ml(ovflw_stk->par_lock(),
3333 Mutex::_no_safepoint_check_flag);
3334 // Grab up to 1/4 the size of the work queue
3335 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3336 (size_t)ParGCDesiredObjsFromOverflowList);
3337 num = MIN2(num, ovflw_stk->length());
3338 for (int i = (int) num; i > 0; i--) {
3339 oop cur = ovflw_stk->pop();
3340 assert(cur != NULL, "Counted wrong?");
3341 work_q->push(cur);
3342 }
3343 return num > 0;
3344 }
3345
3346 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3347 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3348 int n_tasks = pst->n_tasks();
3349 // We allow that there may be no tasks to do here because
3350 // we are restarting after a stack overflow.
3351 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3352 uint nth_task = 0;
3353
3354 HeapWord* aligned_start = sp->bottom();
3355 if (sp->used_region().contains(_restart_addr)) {
3356 // Align down to a card boundary for the start of 0th task
3357 // for this space.
3358 aligned_start =
3359 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3360 CardTableModRefBS::card_size);
3361 }
3362
3363 size_t chunk_size = sp->marking_task_size();
3364 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3365 // Having claimed the nth task in this space,
3366 // compute the chunk that it corresponds to:
3367 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3368 aligned_start + (nth_task+1)*chunk_size);
3369 // Try and bump the global finger via a CAS;
3370 // note that we need to do the global finger bump
3371 // _before_ taking the intersection below, because
3372 // the task corresponding to that region will be
3373 // deemed done even if the used_region() expands
3374 // because of allocation -- as it almost certainly will
3375 // during start-up while the threads yield in the
3376 // closure below.
3377 HeapWord* finger = span.end();
3378 bump_global_finger(finger); // atomically
3379 // There are null tasks here corresponding to chunks
3380 // beyond the "top" address of the space.
3381 span = span.intersection(sp->used_region());
3382 if (!span.is_empty()) { // Non-null task
3383 HeapWord* prev_obj;
3384 assert(!span.contains(_restart_addr) || nth_task == 0,
3385 "Inconsistency");
3386 if (nth_task == 0) {
3387 // For the 0th task, we'll not need to compute a block_start.
3388 if (span.contains(_restart_addr)) {
3389 // In the case of a restart because of stack overflow,
3390 // we might additionally skip a chunk prefix.
3391 prev_obj = _restart_addr;
3392 } else {
3393 prev_obj = span.start();
3394 }
3395 } else {
3396 // We want to skip the first object because
3397 // the protocol is to scan any object in its entirety
3398 // that _starts_ in this span; a fortiori, any
3399 // object starting in an earlier span is scanned
3400 // as part of an earlier claimed task.
3401 // Below we use the "careful" version of block_start
3402 // so we do not try to navigate uninitialized objects.
3403 prev_obj = sp->block_start_careful(span.start());
3404 // Below we use a variant of block_size that uses the
3405 // Printezis bits to avoid waiting for allocated
3406 // objects to become initialized/parsable.
3407 while (prev_obj < span.start()) {
3408 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3409 if (sz > 0) {
3410 prev_obj += sz;
3411 } else {
3412 // In this case we may end up doing a bit of redundant
3413 // scanning, but that appears unavoidable, short of
3414 // locking the free list locks; see bug 6324141.
3415 break;
3416 }
3417 }
3418 }
3419 if (prev_obj < span.end()) {
3420 MemRegion my_span = MemRegion(prev_obj, span.end());
3421 // Do the marking work within a non-empty span --
3422 // the last argument to the constructor indicates whether the
3423 // iteration should be incremental with periodic yields.
3424 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3425 &_collector->_markBitMap,
3426 work_queue(i),
3427 &_collector->_markStack);
3428 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3429 } // else nothing to do for this task
3430 } // else nothing to do for this task
3431 }
3432 // We'd be tempted to assert here that since there are no
3433 // more tasks left to claim in this space, the global_finger
3434 // must exceed space->top() and a fortiori space->end(). However,
3435 // that would not quite be correct because the bumping of
3436 // global_finger occurs strictly after the claiming of a task,
3437 // so by the time we reach here the global finger may not yet
3438 // have been bumped up by the thread that claimed the last
3439 // task.
3440 pst->all_tasks_completed();
3441 }
3442
3443 class Par_ConcMarkingClosure: public MetadataAwareOopClosure {
3444 private:
3445 CMSCollector* _collector;
3446 CMSConcMarkingTask* _task;
3447 MemRegion _span;
3448 CMSBitMap* _bit_map;
3449 CMSMarkStack* _overflow_stack;
3450 OopTaskQueue* _work_queue;
3451 protected:
3452 DO_OOP_WORK_DEFN
3453 public:
3454 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3455 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3456 MetadataAwareOopClosure(collector->ref_processor()),
3457 _collector(collector),
3458 _task(task),
3459 _span(collector->_span),
3460 _work_queue(work_queue),
3461 _bit_map(bit_map),
3462 _overflow_stack(overflow_stack)
3463 { }
3464 virtual void do_oop(oop* p);
3465 virtual void do_oop(narrowOop* p);
3466
3467 void trim_queue(size_t max);
3468 void handle_stack_overflow(HeapWord* lost);
3469 void do_yield_check() {
3470 if (_task->should_yield()) {
3471 _task->yield();
3472 }
3473 }
3474 };
3475
3476 // Grey object scanning during work stealing phase --
3477 // the salient assumption here is that any references
3478 // that are in these stolen objects being scanned must
3479 // already have been initialized (else they would not have
3480 // been published), so we do not need to check for
3481 // uninitialized objects before pushing here.
3482 void Par_ConcMarkingClosure::do_oop(oop obj) {
3483 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
3484 HeapWord* addr = (HeapWord*)obj;
3485 // Check if oop points into the CMS generation
3486 // and is not marked
3487 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3488 // a white object ...
3489 // If we manage to "claim" the object, by being the
3490 // first thread to mark it, then we push it on our
3491 // marking stack
3492 if (_bit_map->par_mark(addr)) { // ... now grey
3493 // push on work queue (grey set)
3494 bool simulate_overflow = false;
3495 NOT_PRODUCT(
3496 if (CMSMarkStackOverflowALot &&
3497 _collector->simulate_overflow()) {
3498 // simulate a stack overflow
3499 simulate_overflow = true;
3500 }
3501 )
3502 if (simulate_overflow ||
3503 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3504 // stack overflow
3505 if (PrintCMSStatistics != 0) {
3506 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3507 SIZE_FORMAT, _overflow_stack->capacity());
3508 }
3509 // We cannot assert that the overflow stack is full because
3510 // it may have been emptied since.
3511 assert(simulate_overflow ||
3512 _work_queue->size() == _work_queue->max_elems(),
3513 "Else push should have succeeded");
3514 handle_stack_overflow(addr);
3515 }
3516 } // Else, some other thread got there first
3517 do_yield_check();
3518 }
3519 }
3520
3521 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3522 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3523
3524 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3525 while (_work_queue->size() > max) {
3526 oop new_oop;
3527 if (_work_queue->pop_local(new_oop)) {
3528 assert(new_oop->is_oop(), "Should be an oop");
3529 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3530 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3531 new_oop->oop_iterate(this); // do_oop() above
3532 do_yield_check();
3533 }
3534 }
3535 }
3536
3537 // Upon stack overflow, we discard (part of) the stack,
3538 // remembering the least address amongst those discarded
3539 // in CMSCollector's _restart_address.
3540 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3541 // We need to do this under a mutex to prevent other
3542 // workers from interfering with the work done below.
3543 MutexLockerEx ml(_overflow_stack->par_lock(),
3544 Mutex::_no_safepoint_check_flag);
3545 // Remember the least grey address discarded
3546 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3547 _collector->lower_restart_addr(ra);
3548 _overflow_stack->reset(); // discard stack contents
3549 _overflow_stack->expand(); // expand the stack if possible
3550 }
3551
3552
3553 void CMSConcMarkingTask::do_work_steal(int i) {
3554 OopTaskQueue* work_q = work_queue(i);
3555 oop obj_to_scan;
3556 CMSBitMap* bm = &(_collector->_markBitMap);
3557 CMSMarkStack* ovflw = &(_collector->_markStack);
3558 int* seed = _collector->hash_seed(i);
3559 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3560 while (true) {
3561 cl.trim_queue(0);
3562 assert(work_q->size() == 0, "Should have been emptied above");
3563 if (get_work_from_overflow_stack(ovflw, work_q)) {
3564 // Can't assert below because the work obtained from the
3565 // overflow stack may already have been stolen from us.
3566 // assert(work_q->size() > 0, "Work from overflow stack");
3567 continue;
3568 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3569 assert(obj_to_scan->is_oop(), "Should be an oop");
3570 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3571 obj_to_scan->oop_iterate(&cl);
3572 } else if (terminator()->offer_termination(&_term_term)) {
3573 assert(work_q->size() == 0, "Impossible!");
3574 break;
3575 } else if (yielding() || should_yield()) {
3576 yield();
3577 }
3578 }
3579 }
3580
3581 // This is run by the CMS (coordinator) thread.
3582 void CMSConcMarkingTask::coordinator_yield() {
3583 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3584 "CMS thread should hold CMS token");
3585 // First give up the locks, then yield, then re-lock
3586 // We should probably use a constructor/destructor idiom to
3587 // do this unlock/lock or modify the MutexUnlocker class to
3588 // serve our purpose. XXX
3589 assert_lock_strong(_bit_map_lock);
3590 _bit_map_lock->unlock();
3591 ConcurrentMarkSweepThread::desynchronize(true);
3592 _collector->stopTimer();
3593 if (PrintCMSStatistics != 0) {
3594 _collector->incrementYields();
3595 }
3596
3597 // It is possible for whichever thread initiated the yield request
3598 // not to get a chance to wake up and take the bitmap lock between
3599 // this thread releasing it and reacquiring it. So, while the
3600 // should_yield() flag is on, let's sleep for a bit to give the
3601 // other thread a chance to wake up. The limit imposed on the number
3602 // of iterations is defensive, to avoid any unforseen circumstances
3603 // putting us into an infinite loop. Since it's always been this
3604 // (coordinator_yield()) method that was observed to cause the
3605 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3606 // which is by default non-zero. For the other seven methods that
3607 // also perform the yield operation, as are using a different
3608 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3609 // can enable the sleeping for those methods too, if necessary.
3610 // See 6442774.
3611 //
3612 // We really need to reconsider the synchronization between the GC
3613 // thread and the yield-requesting threads in the future and we
3614 // should really use wait/notify, which is the recommended
3615 // way of doing this type of interaction. Additionally, we should
3616 // consolidate the eight methods that do the yield operation and they
3617 // are almost identical into one for better maintainability and
3618 // readability. See 6445193.
3619 //
3620 // Tony 2006.06.29
3621 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3622 ConcurrentMarkSweepThread::should_yield() &&
3623 !CMSCollector::foregroundGCIsActive(); ++i) {
3624 os::sleep(Thread::current(), 1, false);
3625 }
3626
3627 ConcurrentMarkSweepThread::synchronize(true);
3628 _bit_map_lock->lock_without_safepoint_check();
3629 _collector->startTimer();
3630 }
3631
3632 bool CMSCollector::do_marking_mt() {
3633 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3634 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3635 conc_workers()->active_workers(),
3636 Threads::number_of_non_daemon_threads());
3637 conc_workers()->set_active_workers(num_workers);
3638
3639 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3640
3641 CMSConcMarkingTask tsk(this,
3642 cms_space,
3643 conc_workers(),
3644 task_queues());
3645
3646 // Since the actual number of workers we get may be different
3647 // from the number we requested above, do we need to do anything different
3648 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3649 // class?? XXX
3650 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3651
3652 // Refs discovery is already non-atomic.
3653 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3654 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3655 conc_workers()->start_task(&tsk);
3656 while (tsk.yielded()) {
3657 tsk.coordinator_yield();
3658 conc_workers()->continue_task(&tsk);
3659 }
3660 // If the task was aborted, _restart_addr will be non-NULL
3661 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3662 while (_restart_addr != NULL) {
3663 // XXX For now we do not make use of ABORTED state and have not
3664 // yet implemented the right abort semantics (even in the original
3665 // single-threaded CMS case). That needs some more investigation
3666 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3667 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3668 // If _restart_addr is non-NULL, a marking stack overflow
3669 // occurred; we need to do a fresh marking iteration from the
3670 // indicated restart address.
3671 if (_foregroundGCIsActive) {
3672 // We may be running into repeated stack overflows, having
3673 // reached the limit of the stack size, while making very
3674 // slow forward progress. It may be best to bail out and
3675 // let the foreground collector do its job.
3676 // Clear _restart_addr, so that foreground GC
3677 // works from scratch. This avoids the headache of
3678 // a "rescan" which would otherwise be needed because
3679 // of the dirty mod union table & card table.
3680 _restart_addr = NULL;
3681 return false;
3682 }
3683 // Adjust the task to restart from _restart_addr
3684 tsk.reset(_restart_addr);
3685 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3686 _restart_addr);
3687 _restart_addr = NULL;
3688 // Get the workers going again
3689 conc_workers()->start_task(&tsk);
3690 while (tsk.yielded()) {
3691 tsk.coordinator_yield();
3692 conc_workers()->continue_task(&tsk);
3693 }
3694 }
3695 assert(tsk.completed(), "Inconsistency");
3696 assert(tsk.result() == true, "Inconsistency");
3697 return true;
3698 }
3699
3700 bool CMSCollector::do_marking_st() {
3701 ResourceMark rm;
3702 HandleMark hm;
3703
3704 // Temporarily make refs discovery single threaded (non-MT)
3705 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3706 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3707 &_markStack, CMSYield);
3708 // the last argument to iterate indicates whether the iteration
3709 // should be incremental with periodic yields.
3710 _markBitMap.iterate(&markFromRootsClosure);
3711 // If _restart_addr is non-NULL, a marking stack overflow
3712 // occurred; we need to do a fresh iteration from the
3713 // indicated restart address.
3714 while (_restart_addr != NULL) {
3715 if (_foregroundGCIsActive) {
3716 // We may be running into repeated stack overflows, having
3717 // reached the limit of the stack size, while making very
3718 // slow forward progress. It may be best to bail out and
3719 // let the foreground collector do its job.
3720 // Clear _restart_addr, so that foreground GC
3721 // works from scratch. This avoids the headache of
3722 // a "rescan" which would otherwise be needed because
3723 // of the dirty mod union table & card table.
3724 _restart_addr = NULL;
3725 return false; // indicating failure to complete marking
3726 }
3727 // Deal with stack overflow:
3728 // we restart marking from _restart_addr
3729 HeapWord* ra = _restart_addr;
3730 markFromRootsClosure.reset(ra);
3731 _restart_addr = NULL;
3732 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3733 }
3734 return true;
3735 }
3736
3737 void CMSCollector::preclean() {
3738 check_correct_thread_executing();
3739 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3740 verify_work_stacks_empty();
3741 verify_overflow_empty();
3742 _abort_preclean = false;
3743 if (CMSPrecleaningEnabled) {
3744 if (!CMSEdenChunksRecordAlways) {
3745 _eden_chunk_index = 0;
3746 }
3747 size_t used = get_eden_used();
3748 size_t capacity = get_eden_capacity();
3749 // Don't start sampling unless we will get sufficiently
3750 // many samples.
3751 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
3752 * CMSScheduleRemarkEdenPenetration)) {
3753 _start_sampling = true;
3754 } else {
3755 _start_sampling = false;
3756 }
3757 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3758 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3759 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3760 }
3761 CMSTokenSync x(true); // is cms thread
3762 if (CMSPrecleaningEnabled) {
3763 sample_eden();
3764 _collectorState = AbortablePreclean;
3765 } else {
3766 _collectorState = FinalMarking;
3767 }
3768 verify_work_stacks_empty();
3769 verify_overflow_empty();
3770 }
3771
3772 // Try and schedule the remark such that young gen
3773 // occupancy is CMSScheduleRemarkEdenPenetration %.
3774 void CMSCollector::abortable_preclean() {
3775 check_correct_thread_executing();
3776 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3777 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3778
3779 // If Eden's current occupancy is below this threshold,
3780 // immediately schedule the remark; else preclean
3781 // past the next scavenge in an effort to
3782 // schedule the pause as described above. By choosing
3783 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3784 // we will never do an actual abortable preclean cycle.
3785 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3786 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3787 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails);
3788 // We need more smarts in the abortable preclean
3789 // loop below to deal with cases where allocation
3790 // in young gen is very very slow, and our precleaning
3791 // is running a losing race against a horde of
3792 // mutators intent on flooding us with CMS updates
3793 // (dirty cards).
3794 // One, admittedly dumb, strategy is to give up
3795 // after a certain number of abortable precleaning loops
3796 // or after a certain maximum time. We want to make
3797 // this smarter in the next iteration.
3798 // XXX FIX ME!!! YSR
3799 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3800 while (!(should_abort_preclean() ||
3801 ConcurrentMarkSweepThread::should_terminate())) {
3802 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3803 cumworkdone += workdone;
3804 loops++;
3805 // Voluntarily terminate abortable preclean phase if we have
3806 // been at it for too long.
3807 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3808 loops >= CMSMaxAbortablePrecleanLoops) {
3809 if (PrintGCDetails) {
3810 gclog_or_tty->print(" CMS: abort preclean due to loops ");
3811 }
3812 break;
3813 }
3814 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3815 if (PrintGCDetails) {
3816 gclog_or_tty->print(" CMS: abort preclean due to time ");
3817 }
3818 break;
3819 }
3820 // If we are doing little work each iteration, we should
3821 // take a short break.
3822 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3823 // Sleep for some time, waiting for work to accumulate
3824 stopTimer();
3825 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3826 startTimer();
3827 waited++;
3828 }
3829 }
3830 if (PrintCMSStatistics > 0) {
3831 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3832 loops, waited, cumworkdone);
3833 }
3834 }
3835 CMSTokenSync x(true); // is cms thread
3836 if (_collectorState != Idling) {
3837 assert(_collectorState == AbortablePreclean,
3838 "Spontaneous state transition?");
3839 _collectorState = FinalMarking;
3840 } // Else, a foreground collection completed this CMS cycle.
3841 return;
3842 }
3843
3844 // Respond to an Eden sampling opportunity
3845 void CMSCollector::sample_eden() {
3846 // Make sure a young gc cannot sneak in between our
3847 // reading and recording of a sample.
3848 assert(Thread::current()->is_ConcurrentGC_thread(),
3849 "Only the cms thread may collect Eden samples");
3850 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3851 "Should collect samples while holding CMS token");
3852 if (!_start_sampling) {
3853 return;
3854 }
3855 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3856 // is populated by the young generation.
3857 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3858 if (_eden_chunk_index < _eden_chunk_capacity) {
3859 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3860 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3861 "Unexpected state of Eden");
3862 // We'd like to check that what we just sampled is an oop-start address;
3863 // however, we cannot do that here since the object may not yet have been
3864 // initialized. So we'll instead do the check when we _use_ this sample
3865 // later.
3866 if (_eden_chunk_index == 0 ||
3867 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3868 _eden_chunk_array[_eden_chunk_index-1])
3869 >= CMSSamplingGrain)) {
3870 _eden_chunk_index++; // commit sample
3871 }
3872 }
3873 }
3874 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3875 size_t used = get_eden_used();
3876 size_t capacity = get_eden_capacity();
3877 assert(used <= capacity, "Unexpected state of Eden");
3878 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3879 _abort_preclean = true;
3880 }
3881 }
3882 }
3883
3884
3885 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3886 assert(_collectorState == Precleaning ||
3887 _collectorState == AbortablePreclean, "incorrect state");
3888 ResourceMark rm;
3889 HandleMark hm;
3890
3891 // Precleaning is currently not MT but the reference processor
3892 // may be set for MT. Disable it temporarily here.
3893 ReferenceProcessor* rp = ref_processor();
3894 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3895
3896 // Do one pass of scrubbing the discovered reference lists
3897 // to remove any reference objects with strongly-reachable
3898 // referents.
3899 if (clean_refs) {
3900 CMSPrecleanRefsYieldClosure yield_cl(this);
3901 assert(rp->span().equals(_span), "Spans should be equal");
3902 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3903 &_markStack, true /* preclean */);
3904 CMSDrainMarkingStackClosure complete_trace(this,
3905 _span, &_markBitMap, &_markStack,
3906 &keep_alive, true /* preclean */);
3907
3908 // We don't want this step to interfere with a young
3909 // collection because we don't want to take CPU
3910 // or memory bandwidth away from the young GC threads
3911 // (which may be as many as there are CPUs).
3912 // Note that we don't need to protect ourselves from
3913 // interference with mutators because they can't
3914 // manipulate the discovered reference lists nor affect
3915 // the computed reachability of the referents, the
3916 // only properties manipulated by the precleaning
3917 // of these reference lists.
3918 stopTimer();
3919 CMSTokenSyncWithLocks x(true /* is cms thread */,
3920 bitMapLock());
3921 startTimer();
3922 sample_eden();
3923
3924 // The following will yield to allow foreground
3925 // collection to proceed promptly. XXX YSR:
3926 // The code in this method may need further
3927 // tweaking for better performance and some restructuring
3928 // for cleaner interfaces.
3929 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3930 rp->preclean_discovered_references(
3931 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3932 gc_timer, _gc_tracer_cm->gc_id());
3933 }
3934
3935 if (clean_survivor) { // preclean the active survivor space(s)
3936 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3937 &_markBitMap, &_modUnionTable,
3938 &_markStack, true /* precleaning phase */);
3939 stopTimer();
3940 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3941 bitMapLock());
3942 startTimer();
3943 unsigned int before_count =
3944 GenCollectedHeap::heap()->total_collections();
3945 SurvivorSpacePrecleanClosure
3946 sss_cl(this, _span, &_markBitMap, &_markStack,
3947 &pam_cl, before_count, CMSYield);
3948 _young_gen->from()->object_iterate_careful(&sss_cl);
3949 _young_gen->to()->object_iterate_careful(&sss_cl);
3950 }
3951 MarkRefsIntoAndScanClosure
3952 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3953 &_markStack, this, CMSYield,
3954 true /* precleaning phase */);
3955 // CAUTION: The following closure has persistent state that may need to
3956 // be reset upon a decrease in the sequence of addresses it
3957 // processes.
3958 ScanMarkedObjectsAgainCarefullyClosure
3959 smoac_cl(this, _span,
3960 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3961
3962 // Preclean dirty cards in ModUnionTable and CardTable using
3963 // appropriate convergence criterion;
3964 // repeat CMSPrecleanIter times unless we find that
3965 // we are losing.
3966 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3967 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3968 "Bad convergence multiplier");
3969 assert(CMSPrecleanThreshold >= 100,
3970 "Unreasonably low CMSPrecleanThreshold");
3971
3972 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3973 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3974 numIter < CMSPrecleanIter;
3975 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3976 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3977 if (Verbose && PrintGCDetails) {
3978 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3979 }
3980 // Either there are very few dirty cards, so re-mark
3981 // pause will be small anyway, or our pre-cleaning isn't
3982 // that much faster than the rate at which cards are being
3983 // dirtied, so we might as well stop and re-mark since
3984 // precleaning won't improve our re-mark time by much.
3985 if (curNumCards <= CMSPrecleanThreshold ||
3986 (numIter > 0 &&
3987 (curNumCards * CMSPrecleanDenominator >
3988 lastNumCards * CMSPrecleanNumerator))) {
3989 numIter++;
3990 cumNumCards += curNumCards;
3991 break;
3992 }
3993 }
3994
3995 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
3996
3997 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3998 cumNumCards += curNumCards;
3999 if (PrintGCDetails && PrintCMSStatistics != 0) {
4000 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
4001 curNumCards, cumNumCards, numIter);
4002 }
4003 return cumNumCards; // as a measure of useful work done
4004 }
4005
4006 // PRECLEANING NOTES:
4007 // Precleaning involves:
4008 // . reading the bits of the modUnionTable and clearing the set bits.
4009 // . For the cards corresponding to the set bits, we scan the
4010 // objects on those cards. This means we need the free_list_lock
4011 // so that we can safely iterate over the CMS space when scanning
4012 // for oops.
4013 // . When we scan the objects, we'll be both reading and setting
4014 // marks in the marking bit map, so we'll need the marking bit map.
4015 // . For protecting _collector_state transitions, we take the CGC_lock.
4016 // Note that any races in the reading of of card table entries by the
4017 // CMS thread on the one hand and the clearing of those entries by the
4018 // VM thread or the setting of those entries by the mutator threads on the
4019 // other are quite benign. However, for efficiency it makes sense to keep
4020 // the VM thread from racing with the CMS thread while the latter is
4021 // dirty card info to the modUnionTable. We therefore also use the
4022 // CGC_lock to protect the reading of the card table and the mod union
4023 // table by the CM thread.
4024 // . We run concurrently with mutator updates, so scanning
4025 // needs to be done carefully -- we should not try to scan
4026 // potentially uninitialized objects.
4027 //
4028 // Locking strategy: While holding the CGC_lock, we scan over and
4029 // reset a maximal dirty range of the mod union / card tables, then lock
4030 // the free_list_lock and bitmap lock to do a full marking, then
4031 // release these locks; and repeat the cycle. This allows for a
4032 // certain amount of fairness in the sharing of these locks between
4033 // the CMS collector on the one hand, and the VM thread and the
4034 // mutators on the other.
4035
4036 // NOTE: preclean_mod_union_table() and preclean_card_table()
4037 // further below are largely identical; if you need to modify
4038 // one of these methods, please check the other method too.
4039
4040 size_t CMSCollector::preclean_mod_union_table(
4041 ConcurrentMarkSweepGeneration* old_gen,
4042 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4043 verify_work_stacks_empty();
4044 verify_overflow_empty();
4045
4046 // strategy: starting with the first card, accumulate contiguous
4047 // ranges of dirty cards; clear these cards, then scan the region
4048 // covered by these cards.
4049
4050 // Since all of the MUT is committed ahead, we can just use
4051 // that, in case the generations expand while we are precleaning.
4052 // It might also be fine to just use the committed part of the
4053 // generation, but we might potentially miss cards when the
4054 // generation is rapidly expanding while we are in the midst
4055 // of precleaning.
4056 HeapWord* startAddr = old_gen->reserved().start();
4057 HeapWord* endAddr = old_gen->reserved().end();
4058
4059 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4060
4061 size_t numDirtyCards, cumNumDirtyCards;
4062 HeapWord *nextAddr, *lastAddr;
4063 for (cumNumDirtyCards = numDirtyCards = 0,
4064 nextAddr = lastAddr = startAddr;
4065 nextAddr < endAddr;
4066 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4067
4068 ResourceMark rm;
4069 HandleMark hm;
4070
4071 MemRegion dirtyRegion;
4072 {
4073 stopTimer();
4074 // Potential yield point
4075 CMSTokenSync ts(true);
4076 startTimer();
4077 sample_eden();
4078 // Get dirty region starting at nextOffset (inclusive),
4079 // simultaneously clearing it.
4080 dirtyRegion =
4081 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4082 assert(dirtyRegion.start() >= nextAddr,
4083 "returned region inconsistent?");
4084 }
4085 // Remember where the next search should begin.
4086 // The returned region (if non-empty) is a right open interval,
4087 // so lastOffset is obtained from the right end of that
4088 // interval.
4089 lastAddr = dirtyRegion.end();
4090 // Should do something more transparent and less hacky XXX
4091 numDirtyCards =
4092 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4093
4094 // We'll scan the cards in the dirty region (with periodic
4095 // yields for foreground GC as needed).
4096 if (!dirtyRegion.is_empty()) {
4097 assert(numDirtyCards > 0, "consistency check");
4098 HeapWord* stop_point = NULL;
4099 stopTimer();
4100 // Potential yield point
4101 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
4102 bitMapLock());
4103 startTimer();
4104 {
4105 verify_work_stacks_empty();
4106 verify_overflow_empty();
4107 sample_eden();
4108 stop_point =
4109 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4110 }
4111 if (stop_point != NULL) {
4112 // The careful iteration stopped early either because it found an
4113 // uninitialized object, or because we were in the midst of an
4114 // "abortable preclean", which should now be aborted. Redirty
4115 // the bits corresponding to the partially-scanned or unscanned
4116 // cards. We'll either restart at the next block boundary or
4117 // abort the preclean.
4118 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4119 "Should only be AbortablePreclean.");
4120 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4121 if (should_abort_preclean()) {
4122 break; // out of preclean loop
4123 } else {
4124 // Compute the next address at which preclean should pick up;
4125 // might need bitMapLock in order to read P-bits.
4126 lastAddr = next_card_start_after_block(stop_point);
4127 }
4128 }
4129 } else {
4130 assert(lastAddr == endAddr, "consistency check");
4131 assert(numDirtyCards == 0, "consistency check");
4132 break;
4133 }
4134 }
4135 verify_work_stacks_empty();
4136 verify_overflow_empty();
4137 return cumNumDirtyCards;
4138 }
4139
4140 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4141 // below are largely identical; if you need to modify
4142 // one of these methods, please check the other method too.
4143
4144 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4145 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4146 // strategy: it's similar to precleamModUnionTable above, in that
4147 // we accumulate contiguous ranges of dirty cards, mark these cards
4148 // precleaned, then scan the region covered by these cards.
4149 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4150 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4151
4152 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4153
4154 size_t numDirtyCards, cumNumDirtyCards;
4155 HeapWord *lastAddr, *nextAddr;
4156
4157 for (cumNumDirtyCards = numDirtyCards = 0,
4158 nextAddr = lastAddr = startAddr;
4159 nextAddr < endAddr;
4160 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4161
4162 ResourceMark rm;
4163 HandleMark hm;
4164
4165 MemRegion dirtyRegion;
4166 {
4167 // See comments in "Precleaning notes" above on why we
4168 // do this locking. XXX Could the locking overheads be
4169 // too high when dirty cards are sparse? [I don't think so.]
4170 stopTimer();
4171 CMSTokenSync x(true); // is cms thread
4172 startTimer();
4173 sample_eden();
4174 // Get and clear dirty region from card table
4175 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4176 MemRegion(nextAddr, endAddr),
4177 true,
4178 CardTableModRefBS::precleaned_card_val());
4179
4180 assert(dirtyRegion.start() >= nextAddr,
4181 "returned region inconsistent?");
4182 }
4183 lastAddr = dirtyRegion.end();
4184 numDirtyCards =
4185 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4186
4187 if (!dirtyRegion.is_empty()) {
4188 stopTimer();
4189 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4190 startTimer();
4191 sample_eden();
4192 verify_work_stacks_empty();
4193 verify_overflow_empty();
4194 HeapWord* stop_point =
4195 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4196 if (stop_point != NULL) {
4197 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4198 "Should only be AbortablePreclean.");
4199 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4200 if (should_abort_preclean()) {
4201 break; // out of preclean loop
4202 } else {
4203 // Compute the next address at which preclean should pick up.
4204 lastAddr = next_card_start_after_block(stop_point);
4205 }
4206 }
4207 } else {
4208 break;
4209 }
4210 }
4211 verify_work_stacks_empty();
4212 verify_overflow_empty();
4213 return cumNumDirtyCards;
4214 }
4215
4216 class PrecleanKlassClosure : public KlassClosure {
4217 KlassToOopClosure _cm_klass_closure;
4218 public:
4219 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4220 void do_klass(Klass* k) {
4221 if (k->has_accumulated_modified_oops()) {
4222 k->clear_accumulated_modified_oops();
4223
4224 _cm_klass_closure.do_klass(k);
4225 }
4226 }
4227 };
4228
4229 // The freelist lock is needed to prevent asserts, is it really needed?
4230 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4231
4232 cl->set_freelistLock(freelistLock);
4233
4234 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4235
4236 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4237 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4238 PrecleanKlassClosure preclean_klass_closure(cl);
4239 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
4240
4241 verify_work_stacks_empty();
4242 verify_overflow_empty();
4243 }
4244
4245 void CMSCollector::checkpointRootsFinal() {
4246 assert(_collectorState == FinalMarking, "incorrect state transition?");
4247 check_correct_thread_executing();
4248 // world is stopped at this checkpoint
4249 assert(SafepointSynchronize::is_at_safepoint(),
4250 "world should be stopped");
4251 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4252
4253 verify_work_stacks_empty();
4254 verify_overflow_empty();
4255
4256 if (PrintGCDetails) {
4257 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]",
4258 _young_gen->used() / K,
4259 _young_gen->capacity() / K);
4260 }
4261 {
4262 if (CMSScavengeBeforeRemark) {
4263 GenCollectedHeap* gch = GenCollectedHeap::heap();
4264 // Temporarily set flag to false, GCH->do_collection will
4265 // expect it to be false and set to true
4266 FlagSetting fl(gch->_is_gc_active, false);
4267 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4268 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4269 gch->do_collection(true, // full (i.e. force, see below)
4270 false, // !clear_all_soft_refs
4271 0, // size
4272 false, // is_tlab
4273 GenCollectedHeap::YoungGen // type
4274 );
4275 }
4276 FreelistLocker x(this);
4277 MutexLockerEx y(bitMapLock(),
4278 Mutex::_no_safepoint_check_flag);
4279 checkpointRootsFinalWork();
4280 }
4281 verify_work_stacks_empty();
4282 verify_overflow_empty();
4283 }
4284
4285 void CMSCollector::checkpointRootsFinalWork() {
4286 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4287
4288 assert(haveFreelistLocks(), "must have free list locks");
4289 assert_lock_strong(bitMapLock());
4290
4291 ResourceMark rm;
4292 HandleMark hm;
4293
4294 GenCollectedHeap* gch = GenCollectedHeap::heap();
4295
4296 if (should_unload_classes()) {
4297 CodeCache::gc_prologue();
4298 }
4299 assert(haveFreelistLocks(), "must have free list locks");
4300 assert_lock_strong(bitMapLock());
4301
4302 // We might assume that we need not fill TLAB's when
4303 // CMSScavengeBeforeRemark is set, because we may have just done
4304 // a scavenge which would have filled all TLAB's -- and besides
4305 // Eden would be empty. This however may not always be the case --
4306 // for instance although we asked for a scavenge, it may not have
4307 // happened because of a JNI critical section. We probably need
4308 // a policy for deciding whether we can in that case wait until
4309 // the critical section releases and then do the remark following
4310 // the scavenge, and skip it here. In the absence of that policy,
4311 // or of an indication of whether the scavenge did indeed occur,
4312 // we cannot rely on TLAB's having been filled and must do
4313 // so here just in case a scavenge did not happen.
4314 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4315 // Update the saved marks which may affect the root scans.
4316 gch->save_marks();
4317
4318 if (CMSPrintEdenSurvivorChunks) {
4319 print_eden_and_survivor_chunk_arrays();
4320 }
4321
4322 {
4323 #if defined(COMPILER2) || INCLUDE_JVMCI
4324 DerivedPointerTableDeactivate dpt_deact;
4325 #endif
4326
4327 // Note on the role of the mod union table:
4328 // Since the marker in "markFromRoots" marks concurrently with
4329 // mutators, it is possible for some reachable objects not to have been
4330 // scanned. For instance, an only reference to an object A was
4331 // placed in object B after the marker scanned B. Unless B is rescanned,
4332 // A would be collected. Such updates to references in marked objects
4333 // are detected via the mod union table which is the set of all cards
4334 // dirtied since the first checkpoint in this GC cycle and prior to
4335 // the most recent young generation GC, minus those cleaned up by the
4336 // concurrent precleaning.
4337 if (CMSParallelRemarkEnabled) {
4338 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
4339 do_remark_parallel();
4340 } else {
4341 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4342 _gc_timer_cm, _gc_tracer_cm->gc_id());
4343 do_remark_non_parallel();
4344 }
4345 }
4346 verify_work_stacks_empty();
4347 verify_overflow_empty();
4348
4349 {
4350 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());)
4351 refProcessingWork();
4352 }
4353 verify_work_stacks_empty();
4354 verify_overflow_empty();
4355
4356 if (should_unload_classes()) {
4357 CodeCache::gc_epilogue();
4358 }
4359 JvmtiExport::gc_epilogue();
4360
4361 // If we encountered any (marking stack / work queue) overflow
4362 // events during the current CMS cycle, take appropriate
4363 // remedial measures, where possible, so as to try and avoid
4364 // recurrence of that condition.
4365 assert(_markStack.isEmpty(), "No grey objects");
4366 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4367 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4368 if (ser_ovflw > 0) {
4369 if (PrintCMSStatistics != 0) {
4370 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4371 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT
4372 ", kac_preclean=" SIZE_FORMAT ")",
4373 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4374 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4375 }
4376 _markStack.expand();
4377 _ser_pmc_remark_ovflw = 0;
4378 _ser_pmc_preclean_ovflw = 0;
4379 _ser_kac_preclean_ovflw = 0;
4380 _ser_kac_ovflw = 0;
4381 }
4382 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4383 if (PrintCMSStatistics != 0) {
4384 gclog_or_tty->print_cr("Work queue overflow (benign) "
4385 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4386 _par_pmc_remark_ovflw, _par_kac_ovflw);
4387 }
4388 _par_pmc_remark_ovflw = 0;
4389 _par_kac_ovflw = 0;
4390 }
4391 if (PrintCMSStatistics != 0) {
4392 if (_markStack._hit_limit > 0) {
4393 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4394 _markStack._hit_limit);
4395 }
4396 if (_markStack._failed_double > 0) {
4397 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT "),"
4398 " current capacity " SIZE_FORMAT,
4399 _markStack._failed_double,
4400 _markStack.capacity());
4401 }
4402 }
4403 _markStack._hit_limit = 0;
4404 _markStack._failed_double = 0;
4405
4406 if ((VerifyAfterGC || VerifyDuringGC) &&
4407 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4408 verify_after_remark();
4409 }
4410
4411 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4412
4413 // Change under the freelistLocks.
4414 _collectorState = Sweeping;
4415 // Call isAllClear() under bitMapLock
4416 assert(_modUnionTable.isAllClear(),
4417 "Should be clear by end of the final marking");
4418 assert(_ct->klass_rem_set()->mod_union_is_clear(),
4419 "Should be clear by end of the final marking");
4420 }
4421
4422 void CMSParInitialMarkTask::work(uint worker_id) {
4423 elapsedTimer _timer;
4424 ResourceMark rm;
4425 HandleMark hm;
4426
4427 // ---------- scan from roots --------------
4428 _timer.start();
4429 GenCollectedHeap* gch = GenCollectedHeap::heap();
4430 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4431
4432 // ---------- young gen roots --------------
4433 {
4434 work_on_young_gen_roots(worker_id, &par_mri_cl);
4435 _timer.stop();
4436 if (PrintCMSStatistics != 0) {
4437 gclog_or_tty->print_cr(
4438 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
4439 worker_id, _timer.seconds());
4440 }
4441 }
4442
4443 // ---------- remaining roots --------------
4444 _timer.reset();
4445 _timer.start();
4446
4447 CLDToOopClosure cld_closure(&par_mri_cl, true);
4448
4449 gch->gen_process_roots(_strong_roots_scope,
4450 GenCollectedHeap::OldGen,
4451 false, // yg was scanned above
4452 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4453 _collector->should_unload_classes(),
4454 &par_mri_cl,
4455 NULL,
4456 &cld_closure);
4457 assert(_collector->should_unload_classes()
4458 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4459 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4460 _timer.stop();
4461 if (PrintCMSStatistics != 0) {
4462 gclog_or_tty->print_cr(
4463 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
4464 worker_id, _timer.seconds());
4465 }
4466 }
4467
4468 // Parallel remark task
4469 class CMSParRemarkTask: public CMSParMarkTask {
4470 CompactibleFreeListSpace* _cms_space;
4471
4472 // The per-thread work queues, available here for stealing.
4473 OopTaskQueueSet* _task_queues;
4474 ParallelTaskTerminator _term;
4475 StrongRootsScope* _strong_roots_scope;
4476
4477 public:
4478 // A value of 0 passed to n_workers will cause the number of
4479 // workers to be taken from the active workers in the work gang.
4480 CMSParRemarkTask(CMSCollector* collector,
4481 CompactibleFreeListSpace* cms_space,
4482 uint n_workers, WorkGang* workers,
4483 OopTaskQueueSet* task_queues,
4484 StrongRootsScope* strong_roots_scope):
4485 CMSParMarkTask("Rescan roots and grey objects in parallel",
4486 collector, n_workers),
4487 _cms_space(cms_space),
4488 _task_queues(task_queues),
4489 _term(n_workers, task_queues),
4490 _strong_roots_scope(strong_roots_scope) { }
4491
4492 OopTaskQueueSet* task_queues() { return _task_queues; }
4493
4494 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4495
4496 ParallelTaskTerminator* terminator() { return &_term; }
4497 uint n_workers() { return _n_workers; }
4498
4499 void work(uint worker_id);
4500
4501 private:
4502 // ... of dirty cards in old space
4503 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4504 Par_MarkRefsIntoAndScanClosure* cl);
4505
4506 // ... work stealing for the above
4507 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4508 };
4509
4510 class RemarkKlassClosure : public KlassClosure {
4511 KlassToOopClosure _cm_klass_closure;
4512 public:
4513 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4514 void do_klass(Klass* k) {
4515 // Check if we have modified any oops in the Klass during the concurrent marking.
4516 if (k->has_accumulated_modified_oops()) {
4517 k->clear_accumulated_modified_oops();
4518
4519 // We could have transfered the current modified marks to the accumulated marks,
4520 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4521 } else if (k->has_modified_oops()) {
4522 // Don't clear anything, this info is needed by the next young collection.
4523 } else {
4524 // No modified oops in the Klass.
4525 return;
4526 }
4527
4528 // The klass has modified fields, need to scan the klass.
4529 _cm_klass_closure.do_klass(k);
4530 }
4531 };
4532
4533 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
4534 ParNewGeneration* young_gen = _collector->_young_gen;
4535 ContiguousSpace* eden_space = young_gen->eden();
4536 ContiguousSpace* from_space = young_gen->from();
4537 ContiguousSpace* to_space = young_gen->to();
4538
4539 HeapWord** eca = _collector->_eden_chunk_array;
4540 size_t ect = _collector->_eden_chunk_index;
4541 HeapWord** sca = _collector->_survivor_chunk_array;
4542 size_t sct = _collector->_survivor_chunk_index;
4543
4544 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4545 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4546
4547 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
4548 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
4549 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
4550 }
4551
4552 // work_queue(i) is passed to the closure
4553 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
4554 // also is passed to do_dirty_card_rescan_tasks() and to
4555 // do_work_steal() to select the i-th task_queue.
4556
4557 void CMSParRemarkTask::work(uint worker_id) {
4558 elapsedTimer _timer;
4559 ResourceMark rm;
4560 HandleMark hm;
4561
4562 // ---------- rescan from roots --------------
4563 _timer.start();
4564 GenCollectedHeap* gch = GenCollectedHeap::heap();
4565 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4566 _collector->_span, _collector->ref_processor(),
4567 &(_collector->_markBitMap),
4568 work_queue(worker_id));
4569
4570 // Rescan young gen roots first since these are likely
4571 // coarsely partitioned and may, on that account, constitute
4572 // the critical path; thus, it's best to start off that
4573 // work first.
4574 // ---------- young gen roots --------------
4575 {
4576 work_on_young_gen_roots(worker_id, &par_mrias_cl);
4577 _timer.stop();
4578 if (PrintCMSStatistics != 0) {
4579 gclog_or_tty->print_cr(
4580 "Finished young gen rescan work in %dth thread: %3.3f sec",
4581 worker_id, _timer.seconds());
4582 }
4583 }
4584
4585 // ---------- remaining roots --------------
4586 _timer.reset();
4587 _timer.start();
4588 gch->gen_process_roots(_strong_roots_scope,
4589 GenCollectedHeap::OldGen,
4590 false, // yg was scanned above
4591 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4592 _collector->should_unload_classes(),
4593 &par_mrias_cl,
4594 NULL,
4595 NULL); // The dirty klasses will be handled below
4596
4597 assert(_collector->should_unload_classes()
4598 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4599 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4600 _timer.stop();
4601 if (PrintCMSStatistics != 0) {
4602 gclog_or_tty->print_cr(
4603 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4604 worker_id, _timer.seconds());
4605 }
4606
4607 // ---------- unhandled CLD scanning ----------
4608 if (worker_id == 0) { // Single threaded at the moment.
4609 _timer.reset();
4610 _timer.start();
4611
4612 // Scan all new class loader data objects and new dependencies that were
4613 // introduced during concurrent marking.
4614 ResourceMark rm;
4615 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4616 for (int i = 0; i < array->length(); i++) {
4617 par_mrias_cl.do_cld_nv(array->at(i));
4618 }
4619
4620 // We don't need to keep track of new CLDs anymore.
4621 ClassLoaderDataGraph::remember_new_clds(false);
4622
4623 _timer.stop();
4624 if (PrintCMSStatistics != 0) {
4625 gclog_or_tty->print_cr(
4626 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
4627 worker_id, _timer.seconds());
4628 }
4629 }
4630
4631 // ---------- dirty klass scanning ----------
4632 if (worker_id == 0) { // Single threaded at the moment.
4633 _timer.reset();
4634 _timer.start();
4635
4636 // Scan all classes that was dirtied during the concurrent marking phase.
4637 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
4638 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
4639
4640 _timer.stop();
4641 if (PrintCMSStatistics != 0) {
4642 gclog_or_tty->print_cr(
4643 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
4644 worker_id, _timer.seconds());
4645 }
4646 }
4647
4648 // We might have added oops to ClassLoaderData::_handles during the
4649 // concurrent marking phase. These oops point to newly allocated objects
4650 // that are guaranteed to be kept alive. Either by the direct allocation
4651 // code, or when the young collector processes the roots. Hence,
4652 // we don't have to revisit the _handles block during the remark phase.
4653
4654 // ---------- rescan dirty cards ------------
4655 _timer.reset();
4656 _timer.start();
4657
4658 // Do the rescan tasks for each of the two spaces
4659 // (cms_space) in turn.
4660 // "worker_id" is passed to select the task_queue for "worker_id"
4661 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4662 _timer.stop();
4663 if (PrintCMSStatistics != 0) {
4664 gclog_or_tty->print_cr(
4665 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4666 worker_id, _timer.seconds());
4667 }
4668
4669 // ---------- steal work from other threads ...
4670 // ---------- ... and drain overflow list.
4671 _timer.reset();
4672 _timer.start();
4673 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4674 _timer.stop();
4675 if (PrintCMSStatistics != 0) {
4676 gclog_or_tty->print_cr(
4677 "Finished work stealing in %dth thread: %3.3f sec",
4678 worker_id, _timer.seconds());
4679 }
4680 }
4681
4682 // Note that parameter "i" is not used.
4683 void
4684 CMSParMarkTask::do_young_space_rescan(uint worker_id,
4685 OopsInGenClosure* cl, ContiguousSpace* space,
4686 HeapWord** chunk_array, size_t chunk_top) {
4687 // Until all tasks completed:
4688 // . claim an unclaimed task
4689 // . compute region boundaries corresponding to task claimed
4690 // using chunk_array
4691 // . par_oop_iterate(cl) over that region
4692
4693 ResourceMark rm;
4694 HandleMark hm;
4695
4696 SequentialSubTasksDone* pst = space->par_seq_tasks();
4697
4698 uint nth_task = 0;
4699 uint n_tasks = pst->n_tasks();
4700
4701 if (n_tasks > 0) {
4702 assert(pst->valid(), "Uninitialized use?");
4703 HeapWord *start, *end;
4704 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4705 // We claimed task # nth_task; compute its boundaries.
4706 if (chunk_top == 0) { // no samples were taken
4707 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4708 start = space->bottom();
4709 end = space->top();
4710 } else if (nth_task == 0) {
4711 start = space->bottom();
4712 end = chunk_array[nth_task];
4713 } else if (nth_task < (uint)chunk_top) {
4714 assert(nth_task >= 1, "Control point invariant");
4715 start = chunk_array[nth_task - 1];
4716 end = chunk_array[nth_task];
4717 } else {
4718 assert(nth_task == (uint)chunk_top, "Control point invariant");
4719 start = chunk_array[chunk_top - 1];
4720 end = space->top();
4721 }
4722 MemRegion mr(start, end);
4723 // Verify that mr is in space
4724 assert(mr.is_empty() || space->used_region().contains(mr),
4725 "Should be in space");
4726 // Verify that "start" is an object boundary
4727 assert(mr.is_empty() || oop(mr.start())->is_oop(),
4728 "Should be an oop");
4729 space->par_oop_iterate(mr, cl);
4730 }
4731 pst->all_tasks_completed();
4732 }
4733 }
4734
4735 void
4736 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4737 CompactibleFreeListSpace* sp, int i,
4738 Par_MarkRefsIntoAndScanClosure* cl) {
4739 // Until all tasks completed:
4740 // . claim an unclaimed task
4741 // . compute region boundaries corresponding to task claimed
4742 // . transfer dirty bits ct->mut for that region
4743 // . apply rescanclosure to dirty mut bits for that region
4744
4745 ResourceMark rm;
4746 HandleMark hm;
4747
4748 OopTaskQueue* work_q = work_queue(i);
4749 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4750 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4751 // CAUTION: This closure has state that persists across calls to
4752 // the work method dirty_range_iterate_clear() in that it has
4753 // embedded in it a (subtype of) UpwardsObjectClosure. The
4754 // use of that state in the embedded UpwardsObjectClosure instance
4755 // assumes that the cards are always iterated (even if in parallel
4756 // by several threads) in monotonically increasing order per each
4757 // thread. This is true of the implementation below which picks
4758 // card ranges (chunks) in monotonically increasing order globally
4759 // and, a-fortiori, in monotonically increasing order per thread
4760 // (the latter order being a subsequence of the former).
4761 // If the work code below is ever reorganized into a more chaotic
4762 // work-partitioning form than the current "sequential tasks"
4763 // paradigm, the use of that persistent state will have to be
4764 // revisited and modified appropriately. See also related
4765 // bug 4756801 work on which should examine this code to make
4766 // sure that the changes there do not run counter to the
4767 // assumptions made here and necessary for correctness and
4768 // efficiency. Note also that this code might yield inefficient
4769 // behavior in the case of very large objects that span one or
4770 // more work chunks. Such objects would potentially be scanned
4771 // several times redundantly. Work on 4756801 should try and
4772 // address that performance anomaly if at all possible. XXX
4773 MemRegion full_span = _collector->_span;
4774 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4775 MarkFromDirtyCardsClosure
4776 greyRescanClosure(_collector, full_span, // entire span of interest
4777 sp, bm, work_q, cl);
4778
4779 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4780 assert(pst->valid(), "Uninitialized use?");
4781 uint nth_task = 0;
4782 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
4783 MemRegion span = sp->used_region();
4784 HeapWord* start_addr = span.start();
4785 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
4786 alignment);
4787 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4788 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
4789 start_addr, "Check alignment");
4790 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
4791 chunk_size, "Check alignment");
4792
4793 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4794 // Having claimed the nth_task, compute corresponding mem-region,
4795 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4796 // The alignment restriction ensures that we do not need any
4797 // synchronization with other gang-workers while setting or
4798 // clearing bits in thus chunk of the MUT.
4799 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4800 start_addr + (nth_task+1)*chunk_size);
4801 // The last chunk's end might be way beyond end of the
4802 // used region. In that case pull back appropriately.
4803 if (this_span.end() > end_addr) {
4804 this_span.set_end(end_addr);
4805 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4806 }
4807 // Iterate over the dirty cards covering this chunk, marking them
4808 // precleaned, and setting the corresponding bits in the mod union
4809 // table. Since we have been careful to partition at Card and MUT-word
4810 // boundaries no synchronization is needed between parallel threads.
4811 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
4812 &modUnionClosure);
4813
4814 // Having transferred these marks into the modUnionTable,
4815 // rescan the marked objects on the dirty cards in the modUnionTable.
4816 // Even if this is at a synchronous collection, the initial marking
4817 // may have been done during an asynchronous collection so there
4818 // may be dirty bits in the mod-union table.
4819 _collector->_modUnionTable.dirty_range_iterate_clear(
4820 this_span, &greyRescanClosure);
4821 _collector->_modUnionTable.verifyNoOneBitsInRange(
4822 this_span.start(),
4823 this_span.end());
4824 }
4825 pst->all_tasks_completed(); // declare that i am done
4826 }
4827
4828 // . see if we can share work_queues with ParNew? XXX
4829 void
4830 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
4831 int* seed) {
4832 OopTaskQueue* work_q = work_queue(i);
4833 NOT_PRODUCT(int num_steals = 0;)
4834 oop obj_to_scan;
4835 CMSBitMap* bm = &(_collector->_markBitMap);
4836
4837 while (true) {
4838 // Completely finish any left over work from (an) earlier round(s)
4839 cl->trim_queue(0);
4840 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4841 (size_t)ParGCDesiredObjsFromOverflowList);
4842 // Now check if there's any work in the overflow list
4843 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4844 // only affects the number of attempts made to get work from the
4845 // overflow list and does not affect the number of workers. Just
4846 // pass ParallelGCThreads so this behavior is unchanged.
4847 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4848 work_q,
4849 ParallelGCThreads)) {
4850 // found something in global overflow list;
4851 // not yet ready to go stealing work from others.
4852 // We'd like to assert(work_q->size() != 0, ...)
4853 // because we just took work from the overflow list,
4854 // but of course we can't since all of that could have
4855 // been already stolen from us.
4856 // "He giveth and He taketh away."
4857 continue;
4858 }
4859 // Verify that we have no work before we resort to stealing
4860 assert(work_q->size() == 0, "Have work, shouldn't steal");
4861 // Try to steal from other queues that have work
4862 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4863 NOT_PRODUCT(num_steals++;)
4864 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
4865 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4866 // Do scanning work
4867 obj_to_scan->oop_iterate(cl);
4868 // Loop around, finish this work, and try to steal some more
4869 } else if (terminator()->offer_termination()) {
4870 break; // nirvana from the infinite cycle
4871 }
4872 }
4873 NOT_PRODUCT(
4874 if (PrintCMSStatistics != 0) {
4875 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
4876 }
4877 )
4878 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4879 "Else our work is not yet done");
4880 }
4881
4882 // Record object boundaries in _eden_chunk_array by sampling the eden
4883 // top in the slow-path eden object allocation code path and record
4884 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4885 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4886 // sampling in sample_eden() that activates during the part of the
4887 // preclean phase.
4888 void CMSCollector::sample_eden_chunk() {
4889 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4890 if (_eden_chunk_lock->try_lock()) {
4891 // Record a sample. This is the critical section. The contents
4892 // of the _eden_chunk_array have to be non-decreasing in the
4893 // address order.
4894 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4895 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4896 "Unexpected state of Eden");
4897 if (_eden_chunk_index == 0 ||
4898 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4899 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4900 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4901 _eden_chunk_index++; // commit sample
4902 }
4903 _eden_chunk_lock->unlock();
4904 }
4905 }
4906 }
4907
4908 // Return a thread-local PLAB recording array, as appropriate.
4909 void* CMSCollector::get_data_recorder(int thr_num) {
4910 if (_survivor_plab_array != NULL &&
4911 (CMSPLABRecordAlways ||
4912 (_collectorState > Marking && _collectorState < FinalMarking))) {
4913 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4914 ChunkArray* ca = &_survivor_plab_array[thr_num];
4915 ca->reset(); // clear it so that fresh data is recorded
4916 return (void*) ca;
4917 } else {
4918 return NULL;
4919 }
4920 }
4921
4922 // Reset all the thread-local PLAB recording arrays
4923 void CMSCollector::reset_survivor_plab_arrays() {
4924 for (uint i = 0; i < ParallelGCThreads; i++) {
4925 _survivor_plab_array[i].reset();
4926 }
4927 }
4928
4929 // Merge the per-thread plab arrays into the global survivor chunk
4930 // array which will provide the partitioning of the survivor space
4931 // for CMS initial scan and rescan.
4932 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4933 int no_of_gc_threads) {
4934 assert(_survivor_plab_array != NULL, "Error");
4935 assert(_survivor_chunk_array != NULL, "Error");
4936 assert(_collectorState == FinalMarking ||
4937 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4938 for (int j = 0; j < no_of_gc_threads; j++) {
4939 _cursor[j] = 0;
4940 }
4941 HeapWord* top = surv->top();
4942 size_t i;
4943 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4944 HeapWord* min_val = top; // Higher than any PLAB address
4945 uint min_tid = 0; // position of min_val this round
4946 for (int j = 0; j < no_of_gc_threads; j++) {
4947 ChunkArray* cur_sca = &_survivor_plab_array[j];
4948 if (_cursor[j] == cur_sca->end()) {
4949 continue;
4950 }
4951 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4952 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4953 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4954 if (cur_val < min_val) {
4955 min_tid = j;
4956 min_val = cur_val;
4957 } else {
4958 assert(cur_val < top, "All recorded addresses should be less");
4959 }
4960 }
4961 // At this point min_val and min_tid are respectively
4962 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4963 // and the thread (j) that witnesses that address.
4964 // We record this address in the _survivor_chunk_array[i]
4965 // and increment _cursor[min_tid] prior to the next round i.
4966 if (min_val == top) {
4967 break;
4968 }
4969 _survivor_chunk_array[i] = min_val;
4970 _cursor[min_tid]++;
4971 }
4972 // We are all done; record the size of the _survivor_chunk_array
4973 _survivor_chunk_index = i; // exclusive: [0, i)
4974 if (PrintCMSStatistics > 0) {
4975 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4976 }
4977 // Verify that we used up all the recorded entries
4978 #ifdef ASSERT
4979 size_t total = 0;
4980 for (int j = 0; j < no_of_gc_threads; j++) {
4981 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4982 total += _cursor[j];
4983 }
4984 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4985 // Check that the merged array is in sorted order
4986 if (total > 0) {
4987 for (size_t i = 0; i < total - 1; i++) {
4988 if (PrintCMSStatistics > 0) {
4989 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4990 i, p2i(_survivor_chunk_array[i]));
4991 }
4992 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4993 "Not sorted");
4994 }
4995 }
4996 #endif // ASSERT
4997 }
4998
4999 // Set up the space's par_seq_tasks structure for work claiming
5000 // for parallel initial scan and rescan of young gen.
5001 // See ParRescanTask where this is currently used.
5002 void
5003 CMSCollector::
5004 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5005 assert(n_threads > 0, "Unexpected n_threads argument");
5006
5007 // Eden space
5008 if (!_young_gen->eden()->is_empty()) {
5009 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
5010 assert(!pst->valid(), "Clobbering existing data?");
5011 // Each valid entry in [0, _eden_chunk_index) represents a task.
5012 size_t n_tasks = _eden_chunk_index + 1;
5013 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5014 // Sets the condition for completion of the subtask (how many threads
5015 // need to finish in order to be done).
5016 pst->set_n_threads(n_threads);
5017 pst->set_n_tasks((int)n_tasks);
5018 }
5019
5020 // Merge the survivor plab arrays into _survivor_chunk_array
5021 if (_survivor_plab_array != NULL) {
5022 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
5023 } else {
5024 assert(_survivor_chunk_index == 0, "Error");
5025 }
5026
5027 // To space
5028 {
5029 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
5030 assert(!pst->valid(), "Clobbering existing data?");
5031 // Sets the condition for completion of the subtask (how many threads
5032 // need to finish in order to be done).
5033 pst->set_n_threads(n_threads);
5034 pst->set_n_tasks(1);
5035 assert(pst->valid(), "Error");
5036 }
5037
5038 // From space
5039 {
5040 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
5041 assert(!pst->valid(), "Clobbering existing data?");
5042 size_t n_tasks = _survivor_chunk_index + 1;
5043 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5044 // Sets the condition for completion of the subtask (how many threads
5045 // need to finish in order to be done).
5046 pst->set_n_threads(n_threads);
5047 pst->set_n_tasks((int)n_tasks);
5048 assert(pst->valid(), "Error");
5049 }
5050 }
5051
5052 // Parallel version of remark
5053 void CMSCollector::do_remark_parallel() {
5054 GenCollectedHeap* gch = GenCollectedHeap::heap();
5055 WorkGang* workers = gch->workers();
5056 assert(workers != NULL, "Need parallel worker threads.");
5057 // Choose to use the number of GC workers most recently set
5058 // into "active_workers".
5059 uint n_workers = workers->active_workers();
5060
5061 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5062
5063 StrongRootsScope srs(n_workers);
5064
5065 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
5066
5067 // We won't be iterating over the cards in the card table updating
5068 // the younger_gen cards, so we shouldn't call the following else
5069 // the verification code as well as subsequent younger_refs_iterate
5070 // code would get confused. XXX
5071 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5072
5073 // The young gen rescan work will not be done as part of
5074 // process_roots (which currently doesn't know how to
5075 // parallelize such a scan), but rather will be broken up into
5076 // a set of parallel tasks (via the sampling that the [abortable]
5077 // preclean phase did of eden, plus the [two] tasks of
5078 // scanning the [two] survivor spaces. Further fine-grain
5079 // parallelization of the scanning of the survivor spaces
5080 // themselves, and of precleaning of the young gen itself
5081 // is deferred to the future.
5082 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5083
5084 // The dirty card rescan work is broken up into a "sequence"
5085 // of parallel tasks (per constituent space) that are dynamically
5086 // claimed by the parallel threads.
5087 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5088
5089 // It turns out that even when we're using 1 thread, doing the work in a
5090 // separate thread causes wide variance in run times. We can't help this
5091 // in the multi-threaded case, but we special-case n=1 here to get
5092 // repeatable measurements of the 1-thread overhead of the parallel code.
5093 if (n_workers > 1) {
5094 // Make refs discovery MT-safe, if it isn't already: it may not
5095 // necessarily be so, since it's possible that we are doing
5096 // ST marking.
5097 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5098 workers->run_task(&tsk);
5099 } else {
5100 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5101 tsk.work(0);
5102 }
5103
5104 // restore, single-threaded for now, any preserved marks
5105 // as a result of work_q overflow
5106 restore_preserved_marks_if_any();
5107 }
5108
5109 // Non-parallel version of remark
5110 void CMSCollector::do_remark_non_parallel() {
5111 ResourceMark rm;
5112 HandleMark hm;
5113 GenCollectedHeap* gch = GenCollectedHeap::heap();
5114 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5115
5116 MarkRefsIntoAndScanClosure
5117 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5118 &_markStack, this,
5119 false /* should_yield */, false /* not precleaning */);
5120 MarkFromDirtyCardsClosure
5121 markFromDirtyCardsClosure(this, _span,
5122 NULL, // space is set further below
5123 &_markBitMap, &_markStack, &mrias_cl);
5124 {
5125 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5126 // Iterate over the dirty cards, setting the corresponding bits in the
5127 // mod union table.
5128 {
5129 ModUnionClosure modUnionClosure(&_modUnionTable);
5130 _ct->ct_bs()->dirty_card_iterate(
5131 _cmsGen->used_region(),
5132 &modUnionClosure);
5133 }
5134 // Having transferred these marks into the modUnionTable, we just need
5135 // to rescan the marked objects on the dirty cards in the modUnionTable.
5136 // The initial marking may have been done during an asynchronous
5137 // collection so there may be dirty bits in the mod-union table.
5138 const int alignment =
5139 CardTableModRefBS::card_size * BitsPerWord;
5140 {
5141 // ... First handle dirty cards in CMS gen
5142 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5143 MemRegion ur = _cmsGen->used_region();
5144 HeapWord* lb = ur.start();
5145 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5146 MemRegion cms_span(lb, ub);
5147 _modUnionTable.dirty_range_iterate_clear(cms_span,
5148 &markFromDirtyCardsClosure);
5149 verify_work_stacks_empty();
5150 if (PrintCMSStatistics != 0) {
5151 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ",
5152 markFromDirtyCardsClosure.num_dirty_cards());
5153 }
5154 }
5155 }
5156 if (VerifyDuringGC &&
5157 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5158 HandleMark hm; // Discard invalid handles created during verification
5159 Universe::verify();
5160 }
5161 {
5162 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5163
5164 verify_work_stacks_empty();
5165
5166 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5167 StrongRootsScope srs(1);
5168
5169 gch->gen_process_roots(&srs,
5170 GenCollectedHeap::OldGen,
5171 true, // young gen as roots
5172 GenCollectedHeap::ScanningOption(roots_scanning_options()),
5173 should_unload_classes(),
5174 &mrias_cl,
5175 NULL,
5176 NULL); // The dirty klasses will be handled below
5177
5178 assert(should_unload_classes()
5179 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
5180 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5181 }
5182
5183 {
5184 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5185
5186 verify_work_stacks_empty();
5187
5188 // Scan all class loader data objects that might have been introduced
5189 // during concurrent marking.
5190 ResourceMark rm;
5191 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5192 for (int i = 0; i < array->length(); i++) {
5193 mrias_cl.do_cld_nv(array->at(i));
5194 }
5195
5196 // We don't need to keep track of new CLDs anymore.
5197 ClassLoaderDataGraph::remember_new_clds(false);
5198
5199 verify_work_stacks_empty();
5200 }
5201
5202 {
5203 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5204
5205 verify_work_stacks_empty();
5206
5207 RemarkKlassClosure remark_klass_closure(&mrias_cl);
5208 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5209
5210 verify_work_stacks_empty();
5211 }
5212
5213 // We might have added oops to ClassLoaderData::_handles during the
5214 // concurrent marking phase. These oops point to newly allocated objects
5215 // that are guaranteed to be kept alive. Either by the direct allocation
5216 // code, or when the young collector processes the roots. Hence,
5217 // we don't have to revisit the _handles block during the remark phase.
5218
5219 verify_work_stacks_empty();
5220 // Restore evacuated mark words, if any, used for overflow list links
5221 if (!CMSOverflowEarlyRestoration) {
5222 restore_preserved_marks_if_any();
5223 }
5224 verify_overflow_empty();
5225 }
5226
5227 ////////////////////////////////////////////////////////
5228 // Parallel Reference Processing Task Proxy Class
5229 ////////////////////////////////////////////////////////
5230 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5231 OopTaskQueueSet* _queues;
5232 ParallelTaskTerminator _terminator;
5233 public:
5234 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5235 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5236 ParallelTaskTerminator* terminator() { return &_terminator; }
5237 OopTaskQueueSet* queues() { return _queues; }
5238 };
5239
5240 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5241 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5242 CMSCollector* _collector;
5243 CMSBitMap* _mark_bit_map;
5244 const MemRegion _span;
5245 ProcessTask& _task;
5246
5247 public:
5248 CMSRefProcTaskProxy(ProcessTask& task,
5249 CMSCollector* collector,
5250 const MemRegion& span,
5251 CMSBitMap* mark_bit_map,
5252 AbstractWorkGang* workers,
5253 OopTaskQueueSet* task_queues):
5254 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5255 task_queues,
5256 workers->active_workers()),
5257 _task(task),
5258 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5259 {
5260 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5261 "Inconsistency in _span");
5262 }
5263
5264 OopTaskQueueSet* task_queues() { return queues(); }
5265
5266 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5267
5268 void do_work_steal(int i,
5269 CMSParDrainMarkingStackClosure* drain,
5270 CMSParKeepAliveClosure* keep_alive,
5271 int* seed);
5272
5273 virtual void work(uint worker_id);
5274 };
5275
5276 void CMSRefProcTaskProxy::work(uint worker_id) {
5277 ResourceMark rm;
5278 HandleMark hm;
5279 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5280 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5281 _mark_bit_map,
5282 work_queue(worker_id));
5283 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5284 _mark_bit_map,
5285 work_queue(worker_id));
5286 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5287 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5288 if (_task.marks_oops_alive()) {
5289 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5290 _collector->hash_seed(worker_id));
5291 }
5292 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5293 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5294 }
5295
5296 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5297 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5298 EnqueueTask& _task;
5299
5300 public:
5301 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5302 : AbstractGangTask("Enqueue reference objects in parallel"),
5303 _task(task)
5304 { }
5305
5306 virtual void work(uint worker_id)
5307 {
5308 _task.work(worker_id);
5309 }
5310 };
5311
5312 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5313 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5314 _span(span),
5315 _bit_map(bit_map),
5316 _work_queue(work_queue),
5317 _mark_and_push(collector, span, bit_map, work_queue),
5318 _low_water_mark(MIN2((work_queue->max_elems()/4),
5319 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5320 { }
5321
5322 // . see if we can share work_queues with ParNew? XXX
5323 void CMSRefProcTaskProxy::do_work_steal(int i,
5324 CMSParDrainMarkingStackClosure* drain,
5325 CMSParKeepAliveClosure* keep_alive,
5326 int* seed) {
5327 OopTaskQueue* work_q = work_queue(i);
5328 NOT_PRODUCT(int num_steals = 0;)
5329 oop obj_to_scan;
5330
5331 while (true) {
5332 // Completely finish any left over work from (an) earlier round(s)
5333 drain->trim_queue(0);
5334 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5335 (size_t)ParGCDesiredObjsFromOverflowList);
5336 // Now check if there's any work in the overflow list
5337 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5338 // only affects the number of attempts made to get work from the
5339 // overflow list and does not affect the number of workers. Just
5340 // pass ParallelGCThreads so this behavior is unchanged.
5341 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5342 work_q,
5343 ParallelGCThreads)) {
5344 // Found something in global overflow list;
5345 // not yet ready to go stealing work from others.
5346 // We'd like to assert(work_q->size() != 0, ...)
5347 // because we just took work from the overflow list,
5348 // but of course we can't, since all of that might have
5349 // been already stolen from us.
5350 continue;
5351 }
5352 // Verify that we have no work before we resort to stealing
5353 assert(work_q->size() == 0, "Have work, shouldn't steal");
5354 // Try to steal from other queues that have work
5355 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5356 NOT_PRODUCT(num_steals++;)
5357 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5358 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5359 // Do scanning work
5360 obj_to_scan->oop_iterate(keep_alive);
5361 // Loop around, finish this work, and try to steal some more
5362 } else if (terminator()->offer_termination()) {
5363 break; // nirvana from the infinite cycle
5364 }
5365 }
5366 NOT_PRODUCT(
5367 if (PrintCMSStatistics != 0) {
5368 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5369 }
5370 )
5371 }
5372
5373 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5374 {
5375 GenCollectedHeap* gch = GenCollectedHeap::heap();
5376 WorkGang* workers = gch->workers();
5377 assert(workers != NULL, "Need parallel worker threads.");
5378 CMSRefProcTaskProxy rp_task(task, &_collector,
5379 _collector.ref_processor()->span(),
5380 _collector.markBitMap(),
5381 workers, _collector.task_queues());
5382 workers->run_task(&rp_task);
5383 }
5384
5385 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5386 {
5387
5388 GenCollectedHeap* gch = GenCollectedHeap::heap();
5389 WorkGang* workers = gch->workers();
5390 assert(workers != NULL, "Need parallel worker threads.");
5391 CMSRefEnqueueTaskProxy enq_task(task);
5392 workers->run_task(&enq_task);
5393 }
5394
5395 void CMSCollector::refProcessingWork() {
5396 ResourceMark rm;
5397 HandleMark hm;
5398
5399 ReferenceProcessor* rp = ref_processor();
5400 assert(rp->span().equals(_span), "Spans should be equal");
5401 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5402 // Process weak references.
5403 rp->setup_policy(false);
5404 verify_work_stacks_empty();
5405
5406 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5407 &_markStack, false /* !preclean */);
5408 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5409 _span, &_markBitMap, &_markStack,
5410 &cmsKeepAliveClosure, false /* !preclean */);
5411 {
5412 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5413
5414 ReferenceProcessorStats stats;
5415 if (rp->processing_is_mt()) {
5416 // Set the degree of MT here. If the discovery is done MT, there
5417 // may have been a different number of threads doing the discovery
5418 // and a different number of discovered lists may have Ref objects.
5419 // That is OK as long as the Reference lists are balanced (see
5420 // balance_all_queues() and balance_queues()).
5421 GenCollectedHeap* gch = GenCollectedHeap::heap();
5422 uint active_workers = ParallelGCThreads;
5423 WorkGang* workers = gch->workers();
5424 if (workers != NULL) {
5425 active_workers = workers->active_workers();
5426 // The expectation is that active_workers will have already
5427 // been set to a reasonable value. If it has not been set,
5428 // investigate.
5429 assert(active_workers > 0, "Should have been set during scavenge");
5430 }
5431 rp->set_active_mt_degree(active_workers);
5432 CMSRefProcTaskExecutor task_executor(*this);
5433 stats = rp->process_discovered_references(&_is_alive_closure,
5434 &cmsKeepAliveClosure,
5435 &cmsDrainMarkingStackClosure,
5436 &task_executor,
5437 _gc_timer_cm,
5438 _gc_tracer_cm->gc_id());
5439 } else {
5440 stats = rp->process_discovered_references(&_is_alive_closure,
5441 &cmsKeepAliveClosure,
5442 &cmsDrainMarkingStackClosure,
5443 NULL,
5444 _gc_timer_cm,
5445 _gc_tracer_cm->gc_id());
5446 }
5447 _gc_tracer_cm->report_gc_reference_stats(stats);
5448
5449 }
5450
5451 // This is the point where the entire marking should have completed.
5452 verify_work_stacks_empty();
5453
5454 if (should_unload_classes()) {
5455 {
5456 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5457
5458 // Unload classes and purge the SystemDictionary.
5459 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5460
5461 // Unload nmethods.
5462 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5463
5464 // Prune dead klasses from subklass/sibling/implementor lists.
5465 Klass::clean_weak_klass_links(&_is_alive_closure);
5466 }
5467
5468 {
5469 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5470 // Clean up unreferenced symbols in symbol table.
5471 SymbolTable::unlink();
5472 }
5473
5474 {
5475 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());
5476 // Delete entries for dead interned strings.
5477 StringTable::unlink(&_is_alive_closure);
5478 }
5479 }
5480
5481
5482 // Restore any preserved marks as a result of mark stack or
5483 // work queue overflow
5484 restore_preserved_marks_if_any(); // done single-threaded for now
5485
5486 rp->set_enqueuing_is_done(true);
5487 if (rp->processing_is_mt()) {
5488 rp->balance_all_queues();
5489 CMSRefProcTaskExecutor task_executor(*this);
5490 rp->enqueue_discovered_references(&task_executor);
5491 } else {
5492 rp->enqueue_discovered_references(NULL);
5493 }
5494 rp->verify_no_references_recorded();
5495 assert(!rp->discovery_enabled(), "should have been disabled");
5496 }
5497
5498 #ifndef PRODUCT
5499 void CMSCollector::check_correct_thread_executing() {
5500 Thread* t = Thread::current();
5501 // Only the VM thread or the CMS thread should be here.
5502 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5503 "Unexpected thread type");
5504 // If this is the vm thread, the foreground process
5505 // should not be waiting. Note that _foregroundGCIsActive is
5506 // true while the foreground collector is waiting.
5507 if (_foregroundGCShouldWait) {
5508 // We cannot be the VM thread
5509 assert(t->is_ConcurrentGC_thread(),
5510 "Should be CMS thread");
5511 } else {
5512 // We can be the CMS thread only if we are in a stop-world
5513 // phase of CMS collection.
5514 if (t->is_ConcurrentGC_thread()) {
5515 assert(_collectorState == InitialMarking ||
5516 _collectorState == FinalMarking,
5517 "Should be a stop-world phase");
5518 // The CMS thread should be holding the CMS_token.
5519 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5520 "Potential interference with concurrently "
5521 "executing VM thread");
5522 }
5523 }
5524 }
5525 #endif
5526
5527 void CMSCollector::sweep() {
5528 assert(_collectorState == Sweeping, "just checking");
5529 check_correct_thread_executing();
5530 verify_work_stacks_empty();
5531 verify_overflow_empty();
5532 increment_sweep_count();
5533 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5534
5535 _inter_sweep_timer.stop();
5536 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5537
5538 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5539 _intra_sweep_timer.reset();
5540 _intra_sweep_timer.start();
5541 {
5542 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5543 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5544 // First sweep the old gen
5545 {
5546 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5547 bitMapLock());
5548 sweepWork(_cmsGen);
5549 }
5550
5551 // Update Universe::_heap_*_at_gc figures.
5552 // We need all the free list locks to make the abstract state
5553 // transition from Sweeping to Resetting. See detailed note
5554 // further below.
5555 {
5556 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5557 // Update heap occupancy information which is used as
5558 // input to soft ref clearing policy at the next gc.
5559 Universe::update_heap_info_at_gc();
5560 _collectorState = Resizing;
5561 }
5562 }
5563 verify_work_stacks_empty();
5564 verify_overflow_empty();
5565
5566 if (should_unload_classes()) {
5567 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5568 // requires that the virtual spaces are stable and not deleted.
5569 ClassLoaderDataGraph::set_should_purge(true);
5570 }
5571
5572 _intra_sweep_timer.stop();
5573 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5574
5575 _inter_sweep_timer.reset();
5576 _inter_sweep_timer.start();
5577
5578 // We need to use a monotonically non-decreasing time in ms
5579 // or we will see time-warp warnings and os::javaTimeMillis()
5580 // does not guarantee monotonicity.
5581 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5582 update_time_of_last_gc(now);
5583
5584 // NOTE on abstract state transitions:
5585 // Mutators allocate-live and/or mark the mod-union table dirty
5586 // based on the state of the collection. The former is done in
5587 // the interval [Marking, Sweeping] and the latter in the interval
5588 // [Marking, Sweeping). Thus the transitions into the Marking state
5589 // and out of the Sweeping state must be synchronously visible
5590 // globally to the mutators.
5591 // The transition into the Marking state happens with the world
5592 // stopped so the mutators will globally see it. Sweeping is
5593 // done asynchronously by the background collector so the transition
5594 // from the Sweeping state to the Resizing state must be done
5595 // under the freelistLock (as is the check for whether to
5596 // allocate-live and whether to dirty the mod-union table).
5597 assert(_collectorState == Resizing, "Change of collector state to"
5598 " Resizing must be done under the freelistLocks (plural)");
5599
5600 // Now that sweeping has been completed, we clear
5601 // the incremental_collection_failed flag,
5602 // thus inviting a younger gen collection to promote into
5603 // this generation. If such a promotion may still fail,
5604 // the flag will be set again when a young collection is
5605 // attempted.
5606 GenCollectedHeap* gch = GenCollectedHeap::heap();
5607 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5608 gch->update_full_collections_completed(_collection_count_start);
5609 }
5610
5611 // FIX ME!!! Looks like this belongs in CFLSpace, with
5612 // CMSGen merely delegating to it.
5613 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5614 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5615 HeapWord* minAddr = _cmsSpace->bottom();
5616 HeapWord* largestAddr =
5617 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5618 if (largestAddr == NULL) {
5619 // The dictionary appears to be empty. In this case
5620 // try to coalesce at the end of the heap.
5621 largestAddr = _cmsSpace->end();
5622 }
5623 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5624 size_t nearLargestOffset =
5625 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5626 if (PrintFLSStatistics != 0) {
5627 gclog_or_tty->print_cr(
5628 "CMS: Large Block: " PTR_FORMAT ";"
5629 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5630 p2i(largestAddr),
5631 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5632 }
5633 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5634 }
5635
5636 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5637 return addr >= _cmsSpace->nearLargestChunk();
5638 }
5639
5640 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5641 return _cmsSpace->find_chunk_at_end();
5642 }
5643
5644 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5645 bool full) {
5646 // If the young generation has been collected, gather any statistics
5647 // that are of interest at this point.
5648 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation);
5649 if (!full && current_is_young) {
5650 // Gather statistics on the young generation collection.
5651 collector()->stats().record_gc0_end(used());
5652 }
5653 }
5654
5655 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5656 // We iterate over the space(s) underlying this generation,
5657 // checking the mark bit map to see if the bits corresponding
5658 // to specific blocks are marked or not. Blocks that are
5659 // marked are live and are not swept up. All remaining blocks
5660 // are swept up, with coalescing on-the-fly as we sweep up
5661 // contiguous free and/or garbage blocks:
5662 // We need to ensure that the sweeper synchronizes with allocators
5663 // and stop-the-world collectors. In particular, the following
5664 // locks are used:
5665 // . CMS token: if this is held, a stop the world collection cannot occur
5666 // . freelistLock: if this is held no allocation can occur from this
5667 // generation by another thread
5668 // . bitMapLock: if this is held, no other thread can access or update
5669 //
5670
5671 // Note that we need to hold the freelistLock if we use
5672 // block iterate below; else the iterator might go awry if
5673 // a mutator (or promotion) causes block contents to change
5674 // (for instance if the allocator divvies up a block).
5675 // If we hold the free list lock, for all practical purposes
5676 // young generation GC's can't occur (they'll usually need to
5677 // promote), so we might as well prevent all young generation
5678 // GC's while we do a sweeping step. For the same reason, we might
5679 // as well take the bit map lock for the entire duration
5680
5681 // check that we hold the requisite locks
5682 assert(have_cms_token(), "Should hold cms token");
5683 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5684 assert_lock_strong(old_gen->freelistLock());
5685 assert_lock_strong(bitMapLock());
5686
5687 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5688 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5689 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5690 _inter_sweep_estimate.padded_average(),
5691 _intra_sweep_estimate.padded_average());
5692 old_gen->setNearLargestChunk();
5693
5694 {
5695 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5696 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5697 // We need to free-up/coalesce garbage/blocks from a
5698 // co-terminal free run. This is done in the SweepClosure
5699 // destructor; so, do not remove this scope, else the
5700 // end-of-sweep-census below will be off by a little bit.
5701 }
5702 old_gen->cmsSpace()->sweep_completed();
5703 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5704 if (should_unload_classes()) { // unloaded classes this cycle,
5705 _concurrent_cycles_since_last_unload = 0; // ... reset count
5706 } else { // did not unload classes,
5707 _concurrent_cycles_since_last_unload++; // ... increment count
5708 }
5709 }
5710
5711 // Reset CMS data structures (for now just the marking bit map)
5712 // preparatory for the next cycle.
5713 void CMSCollector::reset(bool concurrent) {
5714 if (concurrent) {
5715 CMSTokenSyncWithLocks ts(true, bitMapLock());
5716
5717 // If the state is not "Resetting", the foreground thread
5718 // has done a collection and the resetting.
5719 if (_collectorState != Resetting) {
5720 assert(_collectorState == Idling, "The state should only change"
5721 " because the foreground collector has finished the collection");
5722 return;
5723 }
5724
5725 // Clear the mark bitmap (no grey objects to start with)
5726 // for the next cycle.
5727 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5728 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails);
5729
5730 HeapWord* curAddr = _markBitMap.startWord();
5731 while (curAddr < _markBitMap.endWord()) {
5732 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5733 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5734 _markBitMap.clear_large_range(chunk);
5735 if (ConcurrentMarkSweepThread::should_yield() &&
5736 !foregroundGCIsActive() &&
5737 CMSYield) {
5738 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5739 "CMS thread should hold CMS token");
5740 assert_lock_strong(bitMapLock());
5741 bitMapLock()->unlock();
5742 ConcurrentMarkSweepThread::desynchronize(true);
5743 stopTimer();
5744 if (PrintCMSStatistics != 0) {
5745 incrementYields();
5746 }
5747
5748 // See the comment in coordinator_yield()
5749 for (unsigned i = 0; i < CMSYieldSleepCount &&
5750 ConcurrentMarkSweepThread::should_yield() &&
5751 !CMSCollector::foregroundGCIsActive(); ++i) {
5752 os::sleep(Thread::current(), 1, false);
5753 }
5754
5755 ConcurrentMarkSweepThread::synchronize(true);
5756 bitMapLock()->lock_without_safepoint_check();
5757 startTimer();
5758 }
5759 curAddr = chunk.end();
5760 }
5761 // A successful mostly concurrent collection has been done.
5762 // Because only the full (i.e., concurrent mode failure) collections
5763 // are being measured for gc overhead limits, clean the "near" flag
5764 // and count.
5765 size_policy()->reset_gc_overhead_limit_count();
5766 _collectorState = Idling;
5767 } else {
5768 // already have the lock
5769 assert(_collectorState == Resetting, "just checking");
5770 assert_lock_strong(bitMapLock());
5771 _markBitMap.clear_all();
5772 _collectorState = Idling;
5773 }
5774
5775 register_gc_end();
5776 }
5777
5778 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5779 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5780 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id());
5781 TraceCollectorStats tcs(counters());
5782
5783 switch (op) {
5784 case CMS_op_checkpointRootsInitial: {
5785 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5786 checkpointRootsInitial();
5787 if (PrintGC) {
5788 _cmsGen->printOccupancy("initial-mark");
5789 }
5790 break;
5791 }
5792 case CMS_op_checkpointRootsFinal: {
5793 SvcGCMarker sgcm(SvcGCMarker::OTHER);
5794 checkpointRootsFinal();
5795 if (PrintGC) {
5796 _cmsGen->printOccupancy("remark");
5797 }
5798 break;
5799 }
5800 default:
5801 fatal("No such CMS_op");
5802 }
5803 }
5804
5805 #ifndef PRODUCT
5806 size_t const CMSCollector::skip_header_HeapWords() {
5807 return FreeChunk::header_size();
5808 }
5809
5810 // Try and collect here conditions that should hold when
5811 // CMS thread is exiting. The idea is that the foreground GC
5812 // thread should not be blocked if it wants to terminate
5813 // the CMS thread and yet continue to run the VM for a while
5814 // after that.
5815 void CMSCollector::verify_ok_to_terminate() const {
5816 assert(Thread::current()->is_ConcurrentGC_thread(),
5817 "should be called by CMS thread");
5818 assert(!_foregroundGCShouldWait, "should be false");
5819 // We could check here that all the various low-level locks
5820 // are not held by the CMS thread, but that is overkill; see
5821 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5822 // is checked.
5823 }
5824 #endif
5825
5826 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5827 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5828 "missing Printezis mark?");
5829 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5830 size_t size = pointer_delta(nextOneAddr + 1, addr);
5831 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5832 "alignment problem");
5833 assert(size >= 3, "Necessary for Printezis marks to work");
5834 return size;
5835 }
5836
5837 // A variant of the above (block_size_using_printezis_bits()) except
5838 // that we return 0 if the P-bits are not yet set.
5839 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5840 if (_markBitMap.isMarked(addr + 1)) {
5841 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5842 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5843 size_t size = pointer_delta(nextOneAddr + 1, addr);
5844 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5845 "alignment problem");
5846 assert(size >= 3, "Necessary for Printezis marks to work");
5847 return size;
5848 }
5849 return 0;
5850 }
5851
5852 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5853 size_t sz = 0;
5854 oop p = (oop)addr;
5855 if (p->klass_or_null() != NULL) {
5856 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5857 } else {
5858 sz = block_size_using_printezis_bits(addr);
5859 }
5860 assert(sz > 0, "size must be nonzero");
5861 HeapWord* next_block = addr + sz;
5862 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
5863 CardTableModRefBS::card_size);
5864 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
5865 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
5866 "must be different cards");
5867 return next_card;
5868 }
5869
5870
5871 // CMS Bit Map Wrapper /////////////////////////////////////////
5872
5873 // Construct a CMS bit map infrastructure, but don't create the
5874 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5875 // further below.
5876 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5877 _bm(),
5878 _shifter(shifter),
5879 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5880 Monitor::_safepoint_check_sometimes) : NULL)
5881 {
5882 _bmStartWord = 0;
5883 _bmWordSize = 0;
5884 }
5885
5886 bool CMSBitMap::allocate(MemRegion mr) {
5887 _bmStartWord = mr.start();
5888 _bmWordSize = mr.word_size();
5889 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5890 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5891 if (!brs.is_reserved()) {
5892 warning("CMS bit map allocation failure");
5893 return false;
5894 }
5895 // For now we'll just commit all of the bit map up front.
5896 // Later on we'll try to be more parsimonious with swap.
5897 if (!_virtual_space.initialize(brs, brs.size())) {
5898 warning("CMS bit map backing store failure");
5899 return false;
5900 }
5901 assert(_virtual_space.committed_size() == brs.size(),
5902 "didn't reserve backing store for all of CMS bit map?");
5903 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
5904 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5905 _bmWordSize, "inconsistency in bit map sizing");
5906 _bm.set_size(_bmWordSize >> _shifter);
5907
5908 // bm.clear(); // can we rely on getting zero'd memory? verify below
5909 assert(isAllClear(),
5910 "Expected zero'd memory from ReservedSpace constructor");
5911 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5912 "consistency check");
5913 return true;
5914 }
5915
5916 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5917 HeapWord *next_addr, *end_addr, *last_addr;
5918 assert_locked();
5919 assert(covers(mr), "out-of-range error");
5920 // XXX assert that start and end are appropriately aligned
5921 for (next_addr = mr.start(), end_addr = mr.end();
5922 next_addr < end_addr; next_addr = last_addr) {
5923 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5924 last_addr = dirty_region.end();
5925 if (!dirty_region.is_empty()) {
5926 cl->do_MemRegion(dirty_region);
5927 } else {
5928 assert(last_addr == end_addr, "program logic");
5929 return;
5930 }
5931 }
5932 }
5933
5934 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5935 _bm.print_on_error(st, prefix);
5936 }
5937
5938 #ifndef PRODUCT
5939 void CMSBitMap::assert_locked() const {
5940 CMSLockVerifier::assert_locked(lock());
5941 }
5942
5943 bool CMSBitMap::covers(MemRegion mr) const {
5944 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5945 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5946 "size inconsistency");
5947 return (mr.start() >= _bmStartWord) &&
5948 (mr.end() <= endWord());
5949 }
5950
5951 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5952 return (start >= _bmStartWord && (start + size) <= endWord());
5953 }
5954
5955 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5956 // verify that there are no 1 bits in the interval [left, right)
5957 FalseBitMapClosure falseBitMapClosure;
5958 iterate(&falseBitMapClosure, left, right);
5959 }
5960
5961 void CMSBitMap::region_invariant(MemRegion mr)
5962 {
5963 assert_locked();
5964 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5965 assert(!mr.is_empty(), "unexpected empty region");
5966 assert(covers(mr), "mr should be covered by bit map");
5967 // convert address range into offset range
5968 size_t start_ofs = heapWordToOffset(mr.start());
5969 // Make sure that end() is appropriately aligned
5970 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
5971 (1 << (_shifter+LogHeapWordSize))),
5972 "Misaligned mr.end()");
5973 size_t end_ofs = heapWordToOffset(mr.end());
5974 assert(end_ofs > start_ofs, "Should mark at least one bit");
5975 }
5976
5977 #endif
5978
5979 bool CMSMarkStack::allocate(size_t size) {
5980 // allocate a stack of the requisite depth
5981 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5982 size * sizeof(oop)));
5983 if (!rs.is_reserved()) {
5984 warning("CMSMarkStack allocation failure");
5985 return false;
5986 }
5987 if (!_virtual_space.initialize(rs, rs.size())) {
5988 warning("CMSMarkStack backing store failure");
5989 return false;
5990 }
5991 assert(_virtual_space.committed_size() == rs.size(),
5992 "didn't reserve backing store for all of CMS stack?");
5993 _base = (oop*)(_virtual_space.low());
5994 _index = 0;
5995 _capacity = size;
5996 NOT_PRODUCT(_max_depth = 0);
5997 return true;
5998 }
5999
6000 // XXX FIX ME !!! In the MT case we come in here holding a
6001 // leaf lock. For printing we need to take a further lock
6002 // which has lower rank. We need to recalibrate the two
6003 // lock-ranks involved in order to be able to print the
6004 // messages below. (Or defer the printing to the caller.
6005 // For now we take the expedient path of just disabling the
6006 // messages for the problematic case.)
6007 void CMSMarkStack::expand() {
6008 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6009 if (_capacity == MarkStackSizeMax) {
6010 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6011 // We print a warning message only once per CMS cycle.
6012 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6013 }
6014 return;
6015 }
6016 // Double capacity if possible
6017 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6018 // Do not give up existing stack until we have managed to
6019 // get the double capacity that we desired.
6020 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6021 new_capacity * sizeof(oop)));
6022 if (rs.is_reserved()) {
6023 // Release the backing store associated with old stack
6024 _virtual_space.release();
6025 // Reinitialize virtual space for new stack
6026 if (!_virtual_space.initialize(rs, rs.size())) {
6027 fatal("Not enough swap for expanded marking stack");
6028 }
6029 _base = (oop*)(_virtual_space.low());
6030 _index = 0;
6031 _capacity = new_capacity;
6032 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6033 // Failed to double capacity, continue;
6034 // we print a detail message only once per CMS cycle.
6035 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to "
6036 SIZE_FORMAT "K",
6037 _capacity / K, new_capacity / K);
6038 }
6039 }
6040
6041
6042 // Closures
6043 // XXX: there seems to be a lot of code duplication here;
6044 // should refactor and consolidate common code.
6045
6046 // This closure is used to mark refs into the CMS generation in
6047 // the CMS bit map. Called at the first checkpoint. This closure
6048 // assumes that we do not need to re-mark dirty cards; if the CMS
6049 // generation on which this is used is not an oldest
6050 // generation then this will lose younger_gen cards!
6051
6052 MarkRefsIntoClosure::MarkRefsIntoClosure(
6053 MemRegion span, CMSBitMap* bitMap):
6054 _span(span),
6055 _bitMap(bitMap)
6056 {
6057 assert(_ref_processor == NULL, "deliberately left NULL");
6058 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6059 }
6060
6061 void MarkRefsIntoClosure::do_oop(oop obj) {
6062 // if p points into _span, then mark corresponding bit in _markBitMap
6063 assert(obj->is_oop(), "expected an oop");
6064 HeapWord* addr = (HeapWord*)obj;
6065 if (_span.contains(addr)) {
6066 // this should be made more efficient
6067 _bitMap->mark(addr);
6068 }
6069 }
6070
6071 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6072 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6073
6074 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6075 MemRegion span, CMSBitMap* bitMap):
6076 _span(span),
6077 _bitMap(bitMap)
6078 {
6079 assert(_ref_processor == NULL, "deliberately left NULL");
6080 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6081 }
6082
6083 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6084 // if p points into _span, then mark corresponding bit in _markBitMap
6085 assert(obj->is_oop(), "expected an oop");
6086 HeapWord* addr = (HeapWord*)obj;
6087 if (_span.contains(addr)) {
6088 // this should be made more efficient
6089 _bitMap->par_mark(addr);
6090 }
6091 }
6092
6093 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6094 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6095
6096 // A variant of the above, used for CMS marking verification.
6097 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6098 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6099 _span(span),
6100 _verification_bm(verification_bm),
6101 _cms_bm(cms_bm)
6102 {
6103 assert(_ref_processor == NULL, "deliberately left NULL");
6104 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6105 }
6106
6107 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6108 // if p points into _span, then mark corresponding bit in _markBitMap
6109 assert(obj->is_oop(), "expected an oop");
6110 HeapWord* addr = (HeapWord*)obj;
6111 if (_span.contains(addr)) {
6112 _verification_bm->mark(addr);
6113 if (!_cms_bm->isMarked(addr)) {
6114 oop(addr)->print();
6115 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6116 fatal("... aborting");
6117 }
6118 }
6119 }
6120
6121 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6122 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6123
6124 //////////////////////////////////////////////////
6125 // MarkRefsIntoAndScanClosure
6126 //////////////////////////////////////////////////
6127
6128 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6129 ReferenceProcessor* rp,
6130 CMSBitMap* bit_map,
6131 CMSBitMap* mod_union_table,
6132 CMSMarkStack* mark_stack,
6133 CMSCollector* collector,
6134 bool should_yield,
6135 bool concurrent_precleaning):
6136 _collector(collector),
6137 _span(span),
6138 _bit_map(bit_map),
6139 _mark_stack(mark_stack),
6140 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6141 mark_stack, concurrent_precleaning),
6142 _yield(should_yield),
6143 _concurrent_precleaning(concurrent_precleaning),
6144 _freelistLock(NULL)
6145 {
6146 _ref_processor = rp;
6147 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6148 }
6149
6150 // This closure is used to mark refs into the CMS generation at the
6151 // second (final) checkpoint, and to scan and transitively follow
6152 // the unmarked oops. It is also used during the concurrent precleaning
6153 // phase while scanning objects on dirty cards in the CMS generation.
6154 // The marks are made in the marking bit map and the marking stack is
6155 // used for keeping the (newly) grey objects during the scan.
6156 // The parallel version (Par_...) appears further below.
6157 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6158 if (obj != NULL) {
6159 assert(obj->is_oop(), "expected an oop");
6160 HeapWord* addr = (HeapWord*)obj;
6161 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6162 assert(_collector->overflow_list_is_empty(),
6163 "overflow list should be empty");
6164 if (_span.contains(addr) &&
6165 !_bit_map->isMarked(addr)) {
6166 // mark bit map (object is now grey)
6167 _bit_map->mark(addr);
6168 // push on marking stack (stack should be empty), and drain the
6169 // stack by applying this closure to the oops in the oops popped
6170 // from the stack (i.e. blacken the grey objects)
6171 bool res = _mark_stack->push(obj);
6172 assert(res, "Should have space to push on empty stack");
6173 do {
6174 oop new_oop = _mark_stack->pop();
6175 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6176 assert(_bit_map->isMarked((HeapWord*)new_oop),
6177 "only grey objects on this stack");
6178 // iterate over the oops in this oop, marking and pushing
6179 // the ones in CMS heap (i.e. in _span).
6180 new_oop->oop_iterate(&_pushAndMarkClosure);
6181 // check if it's time to yield
6182 do_yield_check();
6183 } while (!_mark_stack->isEmpty() ||
6184 (!_concurrent_precleaning && take_from_overflow_list()));
6185 // if marking stack is empty, and we are not doing this
6186 // during precleaning, then check the overflow list
6187 }
6188 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6189 assert(_collector->overflow_list_is_empty(),
6190 "overflow list was drained above");
6191 // We could restore evacuated mark words, if any, used for
6192 // overflow list links here because the overflow list is
6193 // provably empty here. That would reduce the maximum
6194 // size requirements for preserved_{oop,mark}_stack.
6195 // But we'll just postpone it until we are all done
6196 // so we can just stream through.
6197 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6198 _collector->restore_preserved_marks_if_any();
6199 assert(_collector->no_preserved_marks(), "No preserved marks");
6200 }
6201 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6202 "All preserved marks should have been restored above");
6203 }
6204 }
6205
6206 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6207 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6208
6209 void MarkRefsIntoAndScanClosure::do_yield_work() {
6210 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6211 "CMS thread should hold CMS token");
6212 assert_lock_strong(_freelistLock);
6213 assert_lock_strong(_bit_map->lock());
6214 // relinquish the free_list_lock and bitMaplock()
6215 _bit_map->lock()->unlock();
6216 _freelistLock->unlock();
6217 ConcurrentMarkSweepThread::desynchronize(true);
6218 _collector->stopTimer();
6219 if (PrintCMSStatistics != 0) {
6220 _collector->incrementYields();
6221 }
6222
6223 // See the comment in coordinator_yield()
6224 for (unsigned i = 0;
6225 i < CMSYieldSleepCount &&
6226 ConcurrentMarkSweepThread::should_yield() &&
6227 !CMSCollector::foregroundGCIsActive();
6228 ++i) {
6229 os::sleep(Thread::current(), 1, false);
6230 }
6231
6232 ConcurrentMarkSweepThread::synchronize(true);
6233 _freelistLock->lock_without_safepoint_check();
6234 _bit_map->lock()->lock_without_safepoint_check();
6235 _collector->startTimer();
6236 }
6237
6238 ///////////////////////////////////////////////////////////
6239 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6240 // MarkRefsIntoAndScanClosure
6241 ///////////////////////////////////////////////////////////
6242 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6243 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6244 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6245 _span(span),
6246 _bit_map(bit_map),
6247 _work_queue(work_queue),
6248 _low_water_mark(MIN2((work_queue->max_elems()/4),
6249 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6250 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
6251 {
6252 _ref_processor = rp;
6253 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6254 }
6255
6256 // This closure is used to mark refs into the CMS generation at the
6257 // second (final) checkpoint, and to scan and transitively follow
6258 // the unmarked oops. The marks are made in the marking bit map and
6259 // the work_queue is used for keeping the (newly) grey objects during
6260 // the scan phase whence they are also available for stealing by parallel
6261 // threads. Since the marking bit map is shared, updates are
6262 // synchronized (via CAS).
6263 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6264 if (obj != NULL) {
6265 // Ignore mark word because this could be an already marked oop
6266 // that may be chained at the end of the overflow list.
6267 assert(obj->is_oop(true), "expected an oop");
6268 HeapWord* addr = (HeapWord*)obj;
6269 if (_span.contains(addr) &&
6270 !_bit_map->isMarked(addr)) {
6271 // mark bit map (object will become grey):
6272 // It is possible for several threads to be
6273 // trying to "claim" this object concurrently;
6274 // the unique thread that succeeds in marking the
6275 // object first will do the subsequent push on
6276 // to the work queue (or overflow list).
6277 if (_bit_map->par_mark(addr)) {
6278 // push on work_queue (which may not be empty), and trim the
6279 // queue to an appropriate length by applying this closure to
6280 // the oops in the oops popped from the stack (i.e. blacken the
6281 // grey objects)
6282 bool res = _work_queue->push(obj);
6283 assert(res, "Low water mark should be less than capacity?");
6284 trim_queue(_low_water_mark);
6285 } // Else, another thread claimed the object
6286 }
6287 }
6288 }
6289
6290 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6291 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6292
6293 // This closure is used to rescan the marked objects on the dirty cards
6294 // in the mod union table and the card table proper.
6295 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6296 oop p, MemRegion mr) {
6297
6298 size_t size = 0;
6299 HeapWord* addr = (HeapWord*)p;
6300 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6301 assert(_span.contains(addr), "we are scanning the CMS generation");
6302 // check if it's time to yield
6303 if (do_yield_check()) {
6304 // We yielded for some foreground stop-world work,
6305 // and we have been asked to abort this ongoing preclean cycle.
6306 return 0;
6307 }
6308 if (_bitMap->isMarked(addr)) {
6309 // it's marked; is it potentially uninitialized?
6310 if (p->klass_or_null() != NULL) {
6311 // an initialized object; ignore mark word in verification below
6312 // since we are running concurrent with mutators
6313 assert(p->is_oop(true), "should be an oop");
6314 if (p->is_objArray()) {
6315 // objArrays are precisely marked; restrict scanning
6316 // to dirty cards only.
6317 size = CompactibleFreeListSpace::adjustObjectSize(
6318 p->oop_iterate_size(_scanningClosure, mr));
6319 } else {
6320 // A non-array may have been imprecisely marked; we need
6321 // to scan object in its entirety.
6322 size = CompactibleFreeListSpace::adjustObjectSize(
6323 p->oop_iterate_size(_scanningClosure));
6324 }
6325 #ifdef ASSERT
6326 size_t direct_size =
6327 CompactibleFreeListSpace::adjustObjectSize(p->size());
6328 assert(size == direct_size, "Inconsistency in size");
6329 assert(size >= 3, "Necessary for Printezis marks to work");
6330 if (!_bitMap->isMarked(addr+1)) {
6331 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6332 } else {
6333 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6334 assert(_bitMap->isMarked(addr+size-1),
6335 "inconsistent Printezis mark");
6336 }
6337 #endif // ASSERT
6338 } else {
6339 // An uninitialized object.
6340 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6341 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6342 size = pointer_delta(nextOneAddr + 1, addr);
6343 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6344 "alignment problem");
6345 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6346 // will dirty the card when the klass pointer is installed in the
6347 // object (signaling the completion of initialization).
6348 }
6349 } else {
6350 // Either a not yet marked object or an uninitialized object
6351 if (p->klass_or_null() == NULL) {
6352 // An uninitialized object, skip to the next card, since
6353 // we may not be able to read its P-bits yet.
6354 assert(size == 0, "Initial value");
6355 } else {
6356 // An object not (yet) reached by marking: we merely need to
6357 // compute its size so as to go look at the next block.
6358 assert(p->is_oop(true), "should be an oop");
6359 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6360 }
6361 }
6362 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6363 return size;
6364 }
6365
6366 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6367 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6368 "CMS thread should hold CMS token");
6369 assert_lock_strong(_freelistLock);
6370 assert_lock_strong(_bitMap->lock());
6371 // relinquish the free_list_lock and bitMaplock()
6372 _bitMap->lock()->unlock();
6373 _freelistLock->unlock();
6374 ConcurrentMarkSweepThread::desynchronize(true);
6375 _collector->stopTimer();
6376 if (PrintCMSStatistics != 0) {
6377 _collector->incrementYields();
6378 }
6379
6380 // See the comment in coordinator_yield()
6381 for (unsigned i = 0; i < CMSYieldSleepCount &&
6382 ConcurrentMarkSweepThread::should_yield() &&
6383 !CMSCollector::foregroundGCIsActive(); ++i) {
6384 os::sleep(Thread::current(), 1, false);
6385 }
6386
6387 ConcurrentMarkSweepThread::synchronize(true);
6388 _freelistLock->lock_without_safepoint_check();
6389 _bitMap->lock()->lock_without_safepoint_check();
6390 _collector->startTimer();
6391 }
6392
6393
6394 //////////////////////////////////////////////////////////////////
6395 // SurvivorSpacePrecleanClosure
6396 //////////////////////////////////////////////////////////////////
6397 // This (single-threaded) closure is used to preclean the oops in
6398 // the survivor spaces.
6399 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6400
6401 HeapWord* addr = (HeapWord*)p;
6402 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6403 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6404 assert(p->klass_or_null() != NULL, "object should be initialized");
6405 // an initialized object; ignore mark word in verification below
6406 // since we are running concurrent with mutators
6407 assert(p->is_oop(true), "should be an oop");
6408 // Note that we do not yield while we iterate over
6409 // the interior oops of p, pushing the relevant ones
6410 // on our marking stack.
6411 size_t size = p->oop_iterate_size(_scanning_closure);
6412 do_yield_check();
6413 // Observe that below, we do not abandon the preclean
6414 // phase as soon as we should; rather we empty the
6415 // marking stack before returning. This is to satisfy
6416 // some existing assertions. In general, it may be a
6417 // good idea to abort immediately and complete the marking
6418 // from the grey objects at a later time.
6419 while (!_mark_stack->isEmpty()) {
6420 oop new_oop = _mark_stack->pop();
6421 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6422 assert(_bit_map->isMarked((HeapWord*)new_oop),
6423 "only grey objects on this stack");
6424 // iterate over the oops in this oop, marking and pushing
6425 // the ones in CMS heap (i.e. in _span).
6426 new_oop->oop_iterate(_scanning_closure);
6427 // check if it's time to yield
6428 do_yield_check();
6429 }
6430 unsigned int after_count =
6431 GenCollectedHeap::heap()->total_collections();
6432 bool abort = (_before_count != after_count) ||
6433 _collector->should_abort_preclean();
6434 return abort ? 0 : size;
6435 }
6436
6437 void SurvivorSpacePrecleanClosure::do_yield_work() {
6438 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6439 "CMS thread should hold CMS token");
6440 assert_lock_strong(_bit_map->lock());
6441 // Relinquish the bit map lock
6442 _bit_map->lock()->unlock();
6443 ConcurrentMarkSweepThread::desynchronize(true);
6444 _collector->stopTimer();
6445 if (PrintCMSStatistics != 0) {
6446 _collector->incrementYields();
6447 }
6448
6449 // See the comment in coordinator_yield()
6450 for (unsigned i = 0; i < CMSYieldSleepCount &&
6451 ConcurrentMarkSweepThread::should_yield() &&
6452 !CMSCollector::foregroundGCIsActive(); ++i) {
6453 os::sleep(Thread::current(), 1, false);
6454 }
6455
6456 ConcurrentMarkSweepThread::synchronize(true);
6457 _bit_map->lock()->lock_without_safepoint_check();
6458 _collector->startTimer();
6459 }
6460
6461 // This closure is used to rescan the marked objects on the dirty cards
6462 // in the mod union table and the card table proper. In the parallel
6463 // case, although the bitMap is shared, we do a single read so the
6464 // isMarked() query is "safe".
6465 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6466 // Ignore mark word because we are running concurrent with mutators
6467 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p)));
6468 HeapWord* addr = (HeapWord*)p;
6469 assert(_span.contains(addr), "we are scanning the CMS generation");
6470 bool is_obj_array = false;
6471 #ifdef ASSERT
6472 if (!_parallel) {
6473 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6474 assert(_collector->overflow_list_is_empty(),
6475 "overflow list should be empty");
6476
6477 }
6478 #endif // ASSERT
6479 if (_bit_map->isMarked(addr)) {
6480 // Obj arrays are precisely marked, non-arrays are not;
6481 // so we scan objArrays precisely and non-arrays in their
6482 // entirety.
6483 if (p->is_objArray()) {
6484 is_obj_array = true;
6485 if (_parallel) {
6486 p->oop_iterate(_par_scan_closure, mr);
6487 } else {
6488 p->oop_iterate(_scan_closure, mr);
6489 }
6490 } else {
6491 if (_parallel) {
6492 p->oop_iterate(_par_scan_closure);
6493 } else {
6494 p->oop_iterate(_scan_closure);
6495 }
6496 }
6497 }
6498 #ifdef ASSERT
6499 if (!_parallel) {
6500 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6501 assert(_collector->overflow_list_is_empty(),
6502 "overflow list should be empty");
6503
6504 }
6505 #endif // ASSERT
6506 return is_obj_array;
6507 }
6508
6509 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6510 MemRegion span,
6511 CMSBitMap* bitMap, CMSMarkStack* markStack,
6512 bool should_yield, bool verifying):
6513 _collector(collector),
6514 _span(span),
6515 _bitMap(bitMap),
6516 _mut(&collector->_modUnionTable),
6517 _markStack(markStack),
6518 _yield(should_yield),
6519 _skipBits(0)
6520 {
6521 assert(_markStack->isEmpty(), "stack should be empty");
6522 _finger = _bitMap->startWord();
6523 _threshold = _finger;
6524 assert(_collector->_restart_addr == NULL, "Sanity check");
6525 assert(_span.contains(_finger), "Out of bounds _finger?");
6526 DEBUG_ONLY(_verifying = verifying;)
6527 }
6528
6529 void MarkFromRootsClosure::reset(HeapWord* addr) {
6530 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6531 assert(_span.contains(addr), "Out of bounds _finger?");
6532 _finger = addr;
6533 _threshold = (HeapWord*)round_to(
6534 (intptr_t)_finger, CardTableModRefBS::card_size);
6535 }
6536
6537 // Should revisit to see if this should be restructured for
6538 // greater efficiency.
6539 bool MarkFromRootsClosure::do_bit(size_t offset) {
6540 if (_skipBits > 0) {
6541 _skipBits--;
6542 return true;
6543 }
6544 // convert offset into a HeapWord*
6545 HeapWord* addr = _bitMap->startWord() + offset;
6546 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6547 "address out of range");
6548 assert(_bitMap->isMarked(addr), "tautology");
6549 if (_bitMap->isMarked(addr+1)) {
6550 // this is an allocated but not yet initialized object
6551 assert(_skipBits == 0, "tautology");
6552 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6553 oop p = oop(addr);
6554 if (p->klass_or_null() == NULL) {
6555 DEBUG_ONLY(if (!_verifying) {)
6556 // We re-dirty the cards on which this object lies and increase
6557 // the _threshold so that we'll come back to scan this object
6558 // during the preclean or remark phase. (CMSCleanOnEnter)
6559 if (CMSCleanOnEnter) {
6560 size_t sz = _collector->block_size_using_printezis_bits(addr);
6561 HeapWord* end_card_addr = (HeapWord*)round_to(
6562 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6563 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6564 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6565 // Bump _threshold to end_card_addr; note that
6566 // _threshold cannot possibly exceed end_card_addr, anyhow.
6567 // This prevents future clearing of the card as the scan proceeds
6568 // to the right.
6569 assert(_threshold <= end_card_addr,
6570 "Because we are just scanning into this object");
6571 if (_threshold < end_card_addr) {
6572 _threshold = end_card_addr;
6573 }
6574 if (p->klass_or_null() != NULL) {
6575 // Redirty the range of cards...
6576 _mut->mark_range(redirty_range);
6577 } // ...else the setting of klass will dirty the card anyway.
6578 }
6579 DEBUG_ONLY(})
6580 return true;
6581 }
6582 }
6583 scanOopsInOop(addr);
6584 return true;
6585 }
6586
6587 // We take a break if we've been at this for a while,
6588 // so as to avoid monopolizing the locks involved.
6589 void MarkFromRootsClosure::do_yield_work() {
6590 // First give up the locks, then yield, then re-lock
6591 // We should probably use a constructor/destructor idiom to
6592 // do this unlock/lock or modify the MutexUnlocker class to
6593 // serve our purpose. XXX
6594 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6595 "CMS thread should hold CMS token");
6596 assert_lock_strong(_bitMap->lock());
6597 _bitMap->lock()->unlock();
6598 ConcurrentMarkSweepThread::desynchronize(true);
6599 _collector->stopTimer();
6600 if (PrintCMSStatistics != 0) {
6601 _collector->incrementYields();
6602 }
6603
6604 // See the comment in coordinator_yield()
6605 for (unsigned i = 0; i < CMSYieldSleepCount &&
6606 ConcurrentMarkSweepThread::should_yield() &&
6607 !CMSCollector::foregroundGCIsActive(); ++i) {
6608 os::sleep(Thread::current(), 1, false);
6609 }
6610
6611 ConcurrentMarkSweepThread::synchronize(true);
6612 _bitMap->lock()->lock_without_safepoint_check();
6613 _collector->startTimer();
6614 }
6615
6616 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6617 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6618 assert(_markStack->isEmpty(),
6619 "should drain stack to limit stack usage");
6620 // convert ptr to an oop preparatory to scanning
6621 oop obj = oop(ptr);
6622 // Ignore mark word in verification below, since we
6623 // may be running concurrent with mutators.
6624 assert(obj->is_oop(true), "should be an oop");
6625 assert(_finger <= ptr, "_finger runneth ahead");
6626 // advance the finger to right end of this object
6627 _finger = ptr + obj->size();
6628 assert(_finger > ptr, "we just incremented it above");
6629 // On large heaps, it may take us some time to get through
6630 // the marking phase. During
6631 // this time it's possible that a lot of mutations have
6632 // accumulated in the card table and the mod union table --
6633 // these mutation records are redundant until we have
6634 // actually traced into the corresponding card.
6635 // Here, we check whether advancing the finger would make
6636 // us cross into a new card, and if so clear corresponding
6637 // cards in the MUT (preclean them in the card-table in the
6638 // future).
6639
6640 DEBUG_ONLY(if (!_verifying) {)
6641 // The clean-on-enter optimization is disabled by default,
6642 // until we fix 6178663.
6643 if (CMSCleanOnEnter && (_finger > _threshold)) {
6644 // [_threshold, _finger) represents the interval
6645 // of cards to be cleared in MUT (or precleaned in card table).
6646 // The set of cards to be cleared is all those that overlap
6647 // with the interval [_threshold, _finger); note that
6648 // _threshold is always kept card-aligned but _finger isn't
6649 // always card-aligned.
6650 HeapWord* old_threshold = _threshold;
6651 assert(old_threshold == (HeapWord*)round_to(
6652 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6653 "_threshold should always be card-aligned");
6654 _threshold = (HeapWord*)round_to(
6655 (intptr_t)_finger, CardTableModRefBS::card_size);
6656 MemRegion mr(old_threshold, _threshold);
6657 assert(!mr.is_empty(), "Control point invariant");
6658 assert(_span.contains(mr), "Should clear within span");
6659 _mut->clear_range(mr);
6660 }
6661 DEBUG_ONLY(})
6662 // Note: the finger doesn't advance while we drain
6663 // the stack below.
6664 PushOrMarkClosure pushOrMarkClosure(_collector,
6665 _span, _bitMap, _markStack,
6666 _finger, this);
6667 bool res = _markStack->push(obj);
6668 assert(res, "Empty non-zero size stack should have space for single push");
6669 while (!_markStack->isEmpty()) {
6670 oop new_oop = _markStack->pop();
6671 // Skip verifying header mark word below because we are
6672 // running concurrent with mutators.
6673 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6674 // now scan this oop's oops
6675 new_oop->oop_iterate(&pushOrMarkClosure);
6676 do_yield_check();
6677 }
6678 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6679 }
6680
6681 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
6682 CMSCollector* collector, MemRegion span,
6683 CMSBitMap* bit_map,
6684 OopTaskQueue* work_queue,
6685 CMSMarkStack* overflow_stack):
6686 _collector(collector),
6687 _whole_span(collector->_span),
6688 _span(span),
6689 _bit_map(bit_map),
6690 _mut(&collector->_modUnionTable),
6691 _work_queue(work_queue),
6692 _overflow_stack(overflow_stack),
6693 _skip_bits(0),
6694 _task(task)
6695 {
6696 assert(_work_queue->size() == 0, "work_queue should be empty");
6697 _finger = span.start();
6698 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6699 assert(_span.contains(_finger), "Out of bounds _finger?");
6700 }
6701
6702 // Should revisit to see if this should be restructured for
6703 // greater efficiency.
6704 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
6705 if (_skip_bits > 0) {
6706 _skip_bits--;
6707 return true;
6708 }
6709 // convert offset into a HeapWord*
6710 HeapWord* addr = _bit_map->startWord() + offset;
6711 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6712 "address out of range");
6713 assert(_bit_map->isMarked(addr), "tautology");
6714 if (_bit_map->isMarked(addr+1)) {
6715 // this is an allocated object that might not yet be initialized
6716 assert(_skip_bits == 0, "tautology");
6717 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6718 oop p = oop(addr);
6719 if (p->klass_or_null() == NULL) {
6720 // in the case of Clean-on-Enter optimization, redirty card
6721 // and avoid clearing card by increasing the threshold.
6722 return true;
6723 }
6724 }
6725 scan_oops_in_oop(addr);
6726 return true;
6727 }
6728
6729 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6730 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6731 // Should we assert that our work queue is empty or
6732 // below some drain limit?
6733 assert(_work_queue->size() == 0,
6734 "should drain stack to limit stack usage");
6735 // convert ptr to an oop preparatory to scanning
6736 oop obj = oop(ptr);
6737 // Ignore mark word in verification below, since we
6738 // may be running concurrent with mutators.
6739 assert(obj->is_oop(true), "should be an oop");
6740 assert(_finger <= ptr, "_finger runneth ahead");
6741 // advance the finger to right end of this object
6742 _finger = ptr + obj->size();
6743 assert(_finger > ptr, "we just incremented it above");
6744 // On large heaps, it may take us some time to get through
6745 // the marking phase. During
6746 // this time it's possible that a lot of mutations have
6747 // accumulated in the card table and the mod union table --
6748 // these mutation records are redundant until we have
6749 // actually traced into the corresponding card.
6750 // Here, we check whether advancing the finger would make
6751 // us cross into a new card, and if so clear corresponding
6752 // cards in the MUT (preclean them in the card-table in the
6753 // future).
6754
6755 // The clean-on-enter optimization is disabled by default,
6756 // until we fix 6178663.
6757 if (CMSCleanOnEnter && (_finger > _threshold)) {
6758 // [_threshold, _finger) represents the interval
6759 // of cards to be cleared in MUT (or precleaned in card table).
6760 // The set of cards to be cleared is all those that overlap
6761 // with the interval [_threshold, _finger); note that
6762 // _threshold is always kept card-aligned but _finger isn't
6763 // always card-aligned.
6764 HeapWord* old_threshold = _threshold;
6765 assert(old_threshold == (HeapWord*)round_to(
6766 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6767 "_threshold should always be card-aligned");
6768 _threshold = (HeapWord*)round_to(
6769 (intptr_t)_finger, CardTableModRefBS::card_size);
6770 MemRegion mr(old_threshold, _threshold);
6771 assert(!mr.is_empty(), "Control point invariant");
6772 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6773 _mut->clear_range(mr);
6774 }
6775
6776 // Note: the local finger doesn't advance while we drain
6777 // the stack below, but the global finger sure can and will.
6778 HeapWord** gfa = _task->global_finger_addr();
6779 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
6780 _span, _bit_map,
6781 _work_queue,
6782 _overflow_stack,
6783 _finger,
6784 gfa, this);
6785 bool res = _work_queue->push(obj); // overflow could occur here
6786 assert(res, "Will hold once we use workqueues");
6787 while (true) {
6788 oop new_oop;
6789 if (!_work_queue->pop_local(new_oop)) {
6790 // We emptied our work_queue; check if there's stuff that can
6791 // be gotten from the overflow stack.
6792 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6793 _overflow_stack, _work_queue)) {
6794 do_yield_check();
6795 continue;
6796 } else { // done
6797 break;
6798 }
6799 }
6800 // Skip verifying header mark word below because we are
6801 // running concurrent with mutators.
6802 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6803 // now scan this oop's oops
6804 new_oop->oop_iterate(&pushOrMarkClosure);
6805 do_yield_check();
6806 }
6807 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6808 }
6809
6810 // Yield in response to a request from VM Thread or
6811 // from mutators.
6812 void Par_MarkFromRootsClosure::do_yield_work() {
6813 assert(_task != NULL, "sanity");
6814 _task->yield();
6815 }
6816
6817 // A variant of the above used for verifying CMS marking work.
6818 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6819 MemRegion span,
6820 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6821 CMSMarkStack* mark_stack):
6822 _collector(collector),
6823 _span(span),
6824 _verification_bm(verification_bm),
6825 _cms_bm(cms_bm),
6826 _mark_stack(mark_stack),
6827 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6828 mark_stack)
6829 {
6830 assert(_mark_stack->isEmpty(), "stack should be empty");
6831 _finger = _verification_bm->startWord();
6832 assert(_collector->_restart_addr == NULL, "Sanity check");
6833 assert(_span.contains(_finger), "Out of bounds _finger?");
6834 }
6835
6836 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6837 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6838 assert(_span.contains(addr), "Out of bounds _finger?");
6839 _finger = addr;
6840 }
6841
6842 // Should revisit to see if this should be restructured for
6843 // greater efficiency.
6844 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6845 // convert offset into a HeapWord*
6846 HeapWord* addr = _verification_bm->startWord() + offset;
6847 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6848 "address out of range");
6849 assert(_verification_bm->isMarked(addr), "tautology");
6850 assert(_cms_bm->isMarked(addr), "tautology");
6851
6852 assert(_mark_stack->isEmpty(),
6853 "should drain stack to limit stack usage");
6854 // convert addr to an oop preparatory to scanning
6855 oop obj = oop(addr);
6856 assert(obj->is_oop(), "should be an oop");
6857 assert(_finger <= addr, "_finger runneth ahead");
6858 // advance the finger to right end of this object
6859 _finger = addr + obj->size();
6860 assert(_finger > addr, "we just incremented it above");
6861 // Note: the finger doesn't advance while we drain
6862 // the stack below.
6863 bool res = _mark_stack->push(obj);
6864 assert(res, "Empty non-zero size stack should have space for single push");
6865 while (!_mark_stack->isEmpty()) {
6866 oop new_oop = _mark_stack->pop();
6867 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
6868 // now scan this oop's oops
6869 new_oop->oop_iterate(&_pam_verify_closure);
6870 }
6871 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6872 return true;
6873 }
6874
6875 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6876 CMSCollector* collector, MemRegion span,
6877 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6878 CMSMarkStack* mark_stack):
6879 MetadataAwareOopClosure(collector->ref_processor()),
6880 _collector(collector),
6881 _span(span),
6882 _verification_bm(verification_bm),
6883 _cms_bm(cms_bm),
6884 _mark_stack(mark_stack)
6885 { }
6886
6887 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6888 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6889
6890 // Upon stack overflow, we discard (part of) the stack,
6891 // remembering the least address amongst those discarded
6892 // in CMSCollector's _restart_address.
6893 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6894 // Remember the least grey address discarded
6895 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6896 _collector->lower_restart_addr(ra);
6897 _mark_stack->reset(); // discard stack contents
6898 _mark_stack->expand(); // expand the stack if possible
6899 }
6900
6901 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6902 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
6903 HeapWord* addr = (HeapWord*)obj;
6904 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6905 // Oop lies in _span and isn't yet grey or black
6906 _verification_bm->mark(addr); // now grey
6907 if (!_cms_bm->isMarked(addr)) {
6908 oop(addr)->print();
6909 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
6910 p2i(addr));
6911 fatal("... aborting");
6912 }
6913
6914 if (!_mark_stack->push(obj)) { // stack overflow
6915 if (PrintCMSStatistics != 0) {
6916 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
6917 SIZE_FORMAT, _mark_stack->capacity());
6918 }
6919 assert(_mark_stack->isFull(), "Else push should have succeeded");
6920 handle_stack_overflow(addr);
6921 }
6922 // anything including and to the right of _finger
6923 // will be scanned as we iterate over the remainder of the
6924 // bit map
6925 }
6926 }
6927
6928 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6929 MemRegion span,
6930 CMSBitMap* bitMap, CMSMarkStack* markStack,
6931 HeapWord* finger, MarkFromRootsClosure* parent) :
6932 MetadataAwareOopClosure(collector->ref_processor()),
6933 _collector(collector),
6934 _span(span),
6935 _bitMap(bitMap),
6936 _markStack(markStack),
6937 _finger(finger),
6938 _parent(parent)
6939 { }
6940
6941 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
6942 MemRegion span,
6943 CMSBitMap* bit_map,
6944 OopTaskQueue* work_queue,
6945 CMSMarkStack* overflow_stack,
6946 HeapWord* finger,
6947 HeapWord** global_finger_addr,
6948 Par_MarkFromRootsClosure* parent) :
6949 MetadataAwareOopClosure(collector->ref_processor()),
6950 _collector(collector),
6951 _whole_span(collector->_span),
6952 _span(span),
6953 _bit_map(bit_map),
6954 _work_queue(work_queue),
6955 _overflow_stack(overflow_stack),
6956 _finger(finger),
6957 _global_finger_addr(global_finger_addr),
6958 _parent(parent)
6959 { }
6960
6961 // Assumes thread-safe access by callers, who are
6962 // responsible for mutual exclusion.
6963 void CMSCollector::lower_restart_addr(HeapWord* low) {
6964 assert(_span.contains(low), "Out of bounds addr");
6965 if (_restart_addr == NULL) {
6966 _restart_addr = low;
6967 } else {
6968 _restart_addr = MIN2(_restart_addr, low);
6969 }
6970 }
6971
6972 // Upon stack overflow, we discard (part of) the stack,
6973 // remembering the least address amongst those discarded
6974 // in CMSCollector's _restart_address.
6975 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6976 // Remember the least grey address discarded
6977 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6978 _collector->lower_restart_addr(ra);
6979 _markStack->reset(); // discard stack contents
6980 _markStack->expand(); // expand the stack if possible
6981 }
6982
6983 // Upon stack overflow, we discard (part of) the stack,
6984 // remembering the least address amongst those discarded
6985 // in CMSCollector's _restart_address.
6986 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6987 // We need to do this under a mutex to prevent other
6988 // workers from interfering with the work done below.
6989 MutexLockerEx ml(_overflow_stack->par_lock(),
6990 Mutex::_no_safepoint_check_flag);
6991 // Remember the least grey address discarded
6992 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6993 _collector->lower_restart_addr(ra);
6994 _overflow_stack->reset(); // discard stack contents
6995 _overflow_stack->expand(); // expand the stack if possible
6996 }
6997
6998 void PushOrMarkClosure::do_oop(oop obj) {
6999 // Ignore mark word because we are running concurrent with mutators.
7000 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7001 HeapWord* addr = (HeapWord*)obj;
7002 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7003 // Oop lies in _span and isn't yet grey or black
7004 _bitMap->mark(addr); // now grey
7005 if (addr < _finger) {
7006 // the bit map iteration has already either passed, or
7007 // sampled, this bit in the bit map; we'll need to
7008 // use the marking stack to scan this oop's oops.
7009 bool simulate_overflow = false;
7010 NOT_PRODUCT(
7011 if (CMSMarkStackOverflowALot &&
7012 _collector->simulate_overflow()) {
7013 // simulate a stack overflow
7014 simulate_overflow = true;
7015 }
7016 )
7017 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7018 if (PrintCMSStatistics != 0) {
7019 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7020 SIZE_FORMAT, _markStack->capacity());
7021 }
7022 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7023 handle_stack_overflow(addr);
7024 }
7025 }
7026 // anything including and to the right of _finger
7027 // will be scanned as we iterate over the remainder of the
7028 // bit map
7029 do_yield_check();
7030 }
7031 }
7032
7033 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7034 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7035
7036 void Par_PushOrMarkClosure::do_oop(oop obj) {
7037 // Ignore mark word because we are running concurrent with mutators.
7038 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7039 HeapWord* addr = (HeapWord*)obj;
7040 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7041 // Oop lies in _span and isn't yet grey or black
7042 // We read the global_finger (volatile read) strictly after marking oop
7043 bool res = _bit_map->par_mark(addr); // now grey
7044 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7045 // Should we push this marked oop on our stack?
7046 // -- if someone else marked it, nothing to do
7047 // -- if target oop is above global finger nothing to do
7048 // -- if target oop is in chunk and above local finger
7049 // then nothing to do
7050 // -- else push on work queue
7051 if ( !res // someone else marked it, they will deal with it
7052 || (addr >= *gfa) // will be scanned in a later task
7053 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7054 return;
7055 }
7056 // the bit map iteration has already either passed, or
7057 // sampled, this bit in the bit map; we'll need to
7058 // use the marking stack to scan this oop's oops.
7059 bool simulate_overflow = false;
7060 NOT_PRODUCT(
7061 if (CMSMarkStackOverflowALot &&
7062 _collector->simulate_overflow()) {
7063 // simulate a stack overflow
7064 simulate_overflow = true;
7065 }
7066 )
7067 if (simulate_overflow ||
7068 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7069 // stack overflow
7070 if (PrintCMSStatistics != 0) {
7071 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7072 SIZE_FORMAT, _overflow_stack->capacity());
7073 }
7074 // We cannot assert that the overflow stack is full because
7075 // it may have been emptied since.
7076 assert(simulate_overflow ||
7077 _work_queue->size() == _work_queue->max_elems(),
7078 "Else push should have succeeded");
7079 handle_stack_overflow(addr);
7080 }
7081 do_yield_check();
7082 }
7083 }
7084
7085 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7086 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7087
7088 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7089 MemRegion span,
7090 ReferenceProcessor* rp,
7091 CMSBitMap* bit_map,
7092 CMSBitMap* mod_union_table,
7093 CMSMarkStack* mark_stack,
7094 bool concurrent_precleaning):
7095 MetadataAwareOopClosure(rp),
7096 _collector(collector),
7097 _span(span),
7098 _bit_map(bit_map),
7099 _mod_union_table(mod_union_table),
7100 _mark_stack(mark_stack),
7101 _concurrent_precleaning(concurrent_precleaning)
7102 {
7103 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7104 }
7105
7106 // Grey object rescan during pre-cleaning and second checkpoint phases --
7107 // the non-parallel version (the parallel version appears further below.)
7108 void PushAndMarkClosure::do_oop(oop obj) {
7109 // Ignore mark word verification. If during concurrent precleaning,
7110 // the object monitor may be locked. If during the checkpoint
7111 // phases, the object may already have been reached by a different
7112 // path and may be at the end of the global overflow list (so
7113 // the mark word may be NULL).
7114 assert(obj->is_oop_or_null(true /* ignore mark word */),
7115 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7116 HeapWord* addr = (HeapWord*)obj;
7117 // Check if oop points into the CMS generation
7118 // and is not marked
7119 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7120 // a white object ...
7121 _bit_map->mark(addr); // ... now grey
7122 // push on the marking stack (grey set)
7123 bool simulate_overflow = false;
7124 NOT_PRODUCT(
7125 if (CMSMarkStackOverflowALot &&
7126 _collector->simulate_overflow()) {
7127 // simulate a stack overflow
7128 simulate_overflow = true;
7129 }
7130 )
7131 if (simulate_overflow || !_mark_stack->push(obj)) {
7132 if (_concurrent_precleaning) {
7133 // During precleaning we can just dirty the appropriate card(s)
7134 // in the mod union table, thus ensuring that the object remains
7135 // in the grey set and continue. In the case of object arrays
7136 // we need to dirty all of the cards that the object spans,
7137 // since the rescan of object arrays will be limited to the
7138 // dirty cards.
7139 // Note that no one can be interfering with us in this action
7140 // of dirtying the mod union table, so no locking or atomics
7141 // are required.
7142 if (obj->is_objArray()) {
7143 size_t sz = obj->size();
7144 HeapWord* end_card_addr = (HeapWord*)round_to(
7145 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7146 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7147 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7148 _mod_union_table->mark_range(redirty_range);
7149 } else {
7150 _mod_union_table->mark(addr);
7151 }
7152 _collector->_ser_pmc_preclean_ovflw++;
7153 } else {
7154 // During the remark phase, we need to remember this oop
7155 // in the overflow list.
7156 _collector->push_on_overflow_list(obj);
7157 _collector->_ser_pmc_remark_ovflw++;
7158 }
7159 }
7160 }
7161 }
7162
7163 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7164 MemRegion span,
7165 ReferenceProcessor* rp,
7166 CMSBitMap* bit_map,
7167 OopTaskQueue* work_queue):
7168 MetadataAwareOopClosure(rp),
7169 _collector(collector),
7170 _span(span),
7171 _bit_map(bit_map),
7172 _work_queue(work_queue)
7173 {
7174 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7175 }
7176
7177 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7178 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7179
7180 // Grey object rescan during second checkpoint phase --
7181 // the parallel version.
7182 void Par_PushAndMarkClosure::do_oop(oop obj) {
7183 // In the assert below, we ignore the mark word because
7184 // this oop may point to an already visited object that is
7185 // on the overflow stack (in which case the mark word has
7186 // been hijacked for chaining into the overflow stack --
7187 // if this is the last object in the overflow stack then
7188 // its mark word will be NULL). Because this object may
7189 // have been subsequently popped off the global overflow
7190 // stack, and the mark word possibly restored to the prototypical
7191 // value, by the time we get to examined this failing assert in
7192 // the debugger, is_oop_or_null(false) may subsequently start
7193 // to hold.
7194 assert(obj->is_oop_or_null(true),
7195 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj)));
7196 HeapWord* addr = (HeapWord*)obj;
7197 // Check if oop points into the CMS generation
7198 // and is not marked
7199 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7200 // a white object ...
7201 // If we manage to "claim" the object, by being the
7202 // first thread to mark it, then we push it on our
7203 // marking stack
7204 if (_bit_map->par_mark(addr)) { // ... now grey
7205 // push on work queue (grey set)
7206 bool simulate_overflow = false;
7207 NOT_PRODUCT(
7208 if (CMSMarkStackOverflowALot &&
7209 _collector->par_simulate_overflow()) {
7210 // simulate a stack overflow
7211 simulate_overflow = true;
7212 }
7213 )
7214 if (simulate_overflow || !_work_queue->push(obj)) {
7215 _collector->par_push_on_overflow_list(obj);
7216 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7217 }
7218 } // Else, some other thread got there first
7219 }
7220 }
7221
7222 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7223 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7224
7225 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7226 Mutex* bml = _collector->bitMapLock();
7227 assert_lock_strong(bml);
7228 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7229 "CMS thread should hold CMS token");
7230
7231 bml->unlock();
7232 ConcurrentMarkSweepThread::desynchronize(true);
7233
7234 _collector->stopTimer();
7235 if (PrintCMSStatistics != 0) {
7236 _collector->incrementYields();
7237 }
7238
7239 // See the comment in coordinator_yield()
7240 for (unsigned i = 0; i < CMSYieldSleepCount &&
7241 ConcurrentMarkSweepThread::should_yield() &&
7242 !CMSCollector::foregroundGCIsActive(); ++i) {
7243 os::sleep(Thread::current(), 1, false);
7244 }
7245
7246 ConcurrentMarkSweepThread::synchronize(true);
7247 bml->lock();
7248
7249 _collector->startTimer();
7250 }
7251
7252 bool CMSPrecleanRefsYieldClosure::should_return() {
7253 if (ConcurrentMarkSweepThread::should_yield()) {
7254 do_yield_work();
7255 }
7256 return _collector->foregroundGCIsActive();
7257 }
7258
7259 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7260 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7261 "mr should be aligned to start at a card boundary");
7262 // We'd like to assert:
7263 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7264 // "mr should be a range of cards");
7265 // However, that would be too strong in one case -- the last
7266 // partition ends at _unallocated_block which, in general, can be
7267 // an arbitrary boundary, not necessarily card aligned.
7268 if (PrintCMSStatistics != 0) {
7269 _num_dirty_cards +=
7270 mr.word_size()/CardTableModRefBS::card_size_in_words;
7271 }
7272 _space->object_iterate_mem(mr, &_scan_cl);
7273 }
7274
7275 SweepClosure::SweepClosure(CMSCollector* collector,
7276 ConcurrentMarkSweepGeneration* g,
7277 CMSBitMap* bitMap, bool should_yield) :
7278 _collector(collector),
7279 _g(g),
7280 _sp(g->cmsSpace()),
7281 _limit(_sp->sweep_limit()),
7282 _freelistLock(_sp->freelistLock()),
7283 _bitMap(bitMap),
7284 _yield(should_yield),
7285 _inFreeRange(false), // No free range at beginning of sweep
7286 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7287 _lastFreeRangeCoalesced(false),
7288 _freeFinger(g->used_region().start())
7289 {
7290 NOT_PRODUCT(
7291 _numObjectsFreed = 0;
7292 _numWordsFreed = 0;
7293 _numObjectsLive = 0;
7294 _numWordsLive = 0;
7295 _numObjectsAlreadyFree = 0;
7296 _numWordsAlreadyFree = 0;
7297 _last_fc = NULL;
7298
7299 _sp->initializeIndexedFreeListArrayReturnedBytes();
7300 _sp->dictionary()->initialize_dict_returned_bytes();
7301 )
7302 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7303 "sweep _limit out of bounds");
7304 if (CMSTraceSweeper) {
7305 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7306 p2i(_limit));
7307 }
7308 }
7309
7310 void SweepClosure::print_on(outputStream* st) const {
7311 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7312 p2i(_sp->bottom()), p2i(_sp->end()));
7313 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7314 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7315 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7316 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7317 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7318 }
7319
7320 #ifndef PRODUCT
7321 // Assertion checking only: no useful work in product mode --
7322 // however, if any of the flags below become product flags,
7323 // you may need to review this code to see if it needs to be
7324 // enabled in product mode.
7325 SweepClosure::~SweepClosure() {
7326 assert_lock_strong(_freelistLock);
7327 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7328 "sweep _limit out of bounds");
7329 if (inFreeRange()) {
7330 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7331 print();
7332 ShouldNotReachHere();
7333 }
7334 if (Verbose && PrintGC) {
7335 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7336 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7337 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, "
7338 SIZE_FORMAT " bytes "
7339 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7340 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7341 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7342 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7343 * sizeof(HeapWord);
7344 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7345
7346 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7347 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7348 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7349 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7350 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes);
7351 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes",
7352 indexListReturnedBytes);
7353 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes",
7354 dict_returned_bytes);
7355 }
7356 }
7357 if (CMSTraceSweeper) {
7358 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7359 p2i(_limit));
7360 }
7361 }
7362 #endif // PRODUCT
7363
7364 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7365 bool freeRangeInFreeLists) {
7366 if (CMSTraceSweeper) {
7367 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n",
7368 p2i(freeFinger), freeRangeInFreeLists);
7369 }
7370 assert(!inFreeRange(), "Trampling existing free range");
7371 set_inFreeRange(true);
7372 set_lastFreeRangeCoalesced(false);
7373
7374 set_freeFinger(freeFinger);
7375 set_freeRangeInFreeLists(freeRangeInFreeLists);
7376 if (CMSTestInFreeList) {
7377 if (freeRangeInFreeLists) {
7378 FreeChunk* fc = (FreeChunk*) freeFinger;
7379 assert(fc->is_free(), "A chunk on the free list should be free.");
7380 assert(fc->size() > 0, "Free range should have a size");
7381 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7382 }
7383 }
7384 }
7385
7386 // Note that the sweeper runs concurrently with mutators. Thus,
7387 // it is possible for direct allocation in this generation to happen
7388 // in the middle of the sweep. Note that the sweeper also coalesces
7389 // contiguous free blocks. Thus, unless the sweeper and the allocator
7390 // synchronize appropriately freshly allocated blocks may get swept up.
7391 // This is accomplished by the sweeper locking the free lists while
7392 // it is sweeping. Thus blocks that are determined to be free are
7393 // indeed free. There is however one additional complication:
7394 // blocks that have been allocated since the final checkpoint and
7395 // mark, will not have been marked and so would be treated as
7396 // unreachable and swept up. To prevent this, the allocator marks
7397 // the bit map when allocating during the sweep phase. This leads,
7398 // however, to a further complication -- objects may have been allocated
7399 // but not yet initialized -- in the sense that the header isn't yet
7400 // installed. The sweeper can not then determine the size of the block
7401 // in order to skip over it. To deal with this case, we use a technique
7402 // (due to Printezis) to encode such uninitialized block sizes in the
7403 // bit map. Since the bit map uses a bit per every HeapWord, but the
7404 // CMS generation has a minimum object size of 3 HeapWords, it follows
7405 // that "normal marks" won't be adjacent in the bit map (there will
7406 // always be at least two 0 bits between successive 1 bits). We make use
7407 // of these "unused" bits to represent uninitialized blocks -- the bit
7408 // corresponding to the start of the uninitialized object and the next
7409 // bit are both set. Finally, a 1 bit marks the end of the object that
7410 // started with the two consecutive 1 bits to indicate its potentially
7411 // uninitialized state.
7412
7413 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7414 FreeChunk* fc = (FreeChunk*)addr;
7415 size_t res;
7416
7417 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7418 // than "addr == _limit" because although _limit was a block boundary when
7419 // we started the sweep, it may no longer be one because heap expansion
7420 // may have caused us to coalesce the block ending at the address _limit
7421 // with a newly expanded chunk (this happens when _limit was set to the
7422 // previous _end of the space), so we may have stepped past _limit:
7423 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7424 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7425 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7426 "sweep _limit out of bounds");
7427 assert(addr < _sp->end(), "addr out of bounds");
7428 // Flush any free range we might be holding as a single
7429 // coalesced chunk to the appropriate free list.
7430 if (inFreeRange()) {
7431 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7432 err_msg("freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())));
7433 flush_cur_free_chunk(freeFinger(),
7434 pointer_delta(addr, freeFinger()));
7435 if (CMSTraceSweeper) {
7436 gclog_or_tty->print("Sweep: last chunk: ");
7437 gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") "
7438 "[coalesced:%d]\n",
7439 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7440 lastFreeRangeCoalesced() ? 1 : 0);
7441 }
7442 }
7443
7444 // help the iterator loop finish
7445 return pointer_delta(_sp->end(), addr);
7446 }
7447
7448 assert(addr < _limit, "sweep invariant");
7449 // check if we should yield
7450 do_yield_check(addr);
7451 if (fc->is_free()) {
7452 // Chunk that is already free
7453 res = fc->size();
7454 do_already_free_chunk(fc);
7455 debug_only(_sp->verifyFreeLists());
7456 // If we flush the chunk at hand in lookahead_and_flush()
7457 // and it's coalesced with a preceding chunk, then the
7458 // process of "mangling" the payload of the coalesced block
7459 // will cause erasure of the size information from the
7460 // (erstwhile) header of all the coalesced blocks but the
7461 // first, so the first disjunct in the assert will not hold
7462 // in that specific case (in which case the second disjunct
7463 // will hold).
7464 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7465 "Otherwise the size info doesn't change at this step");
7466 NOT_PRODUCT(
7467 _numObjectsAlreadyFree++;
7468 _numWordsAlreadyFree += res;
7469 )
7470 NOT_PRODUCT(_last_fc = fc;)
7471 } else if (!_bitMap->isMarked(addr)) {
7472 // Chunk is fresh garbage
7473 res = do_garbage_chunk(fc);
7474 debug_only(_sp->verifyFreeLists());
7475 NOT_PRODUCT(
7476 _numObjectsFreed++;
7477 _numWordsFreed += res;
7478 )
7479 } else {
7480 // Chunk that is alive.
7481 res = do_live_chunk(fc);
7482 debug_only(_sp->verifyFreeLists());
7483 NOT_PRODUCT(
7484 _numObjectsLive++;
7485 _numWordsLive += res;
7486 )
7487 }
7488 return res;
7489 }
7490
7491 // For the smart allocation, record following
7492 // split deaths - a free chunk is removed from its free list because
7493 // it is being split into two or more chunks.
7494 // split birth - a free chunk is being added to its free list because
7495 // a larger free chunk has been split and resulted in this free chunk.
7496 // coal death - a free chunk is being removed from its free list because
7497 // it is being coalesced into a large free chunk.
7498 // coal birth - a free chunk is being added to its free list because
7499 // it was created when two or more free chunks where coalesced into
7500 // this free chunk.
7501 //
7502 // These statistics are used to determine the desired number of free
7503 // chunks of a given size. The desired number is chosen to be relative
7504 // to the end of a CMS sweep. The desired number at the end of a sweep
7505 // is the
7506 // count-at-end-of-previous-sweep (an amount that was enough)
7507 // - count-at-beginning-of-current-sweep (the excess)
7508 // + split-births (gains in this size during interval)
7509 // - split-deaths (demands on this size during interval)
7510 // where the interval is from the end of one sweep to the end of the
7511 // next.
7512 //
7513 // When sweeping the sweeper maintains an accumulated chunk which is
7514 // the chunk that is made up of chunks that have been coalesced. That
7515 // will be termed the left-hand chunk. A new chunk of garbage that
7516 // is being considered for coalescing will be referred to as the
7517 // right-hand chunk.
7518 //
7519 // When making a decision on whether to coalesce a right-hand chunk with
7520 // the current left-hand chunk, the current count vs. the desired count
7521 // of the left-hand chunk is considered. Also if the right-hand chunk
7522 // is near the large chunk at the end of the heap (see
7523 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7524 // left-hand chunk is coalesced.
7525 //
7526 // When making a decision about whether to split a chunk, the desired count
7527 // vs. the current count of the candidate to be split is also considered.
7528 // If the candidate is underpopulated (currently fewer chunks than desired)
7529 // a chunk of an overpopulated (currently more chunks than desired) size may
7530 // be chosen. The "hint" associated with a free list, if non-null, points
7531 // to a free list which may be overpopulated.
7532 //
7533
7534 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7535 const size_t size = fc->size();
7536 // Chunks that cannot be coalesced are not in the
7537 // free lists.
7538 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7539 assert(_sp->verify_chunk_in_free_list(fc),
7540 "free chunk should be in free lists");
7541 }
7542 // a chunk that is already free, should not have been
7543 // marked in the bit map
7544 HeapWord* const addr = (HeapWord*) fc;
7545 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7546 // Verify that the bit map has no bits marked between
7547 // addr and purported end of this block.
7548 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7549
7550 // Some chunks cannot be coalesced under any circumstances.
7551 // See the definition of cantCoalesce().
7552 if (!fc->cantCoalesce()) {
7553 // This chunk can potentially be coalesced.
7554 if (_sp->adaptive_freelists()) {
7555 // All the work is done in
7556 do_post_free_or_garbage_chunk(fc, size);
7557 } else { // Not adaptive free lists
7558 // this is a free chunk that can potentially be coalesced by the sweeper;
7559 if (!inFreeRange()) {
7560 // if the next chunk is a free block that can't be coalesced
7561 // it doesn't make sense to remove this chunk from the free lists
7562 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7563 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
7564 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
7565 nextChunk->is_free() && // ... which is free...
7566 nextChunk->cantCoalesce()) { // ... but can't be coalesced
7567 // nothing to do
7568 } else {
7569 // Potentially the start of a new free range:
7570 // Don't eagerly remove it from the free lists.
7571 // No need to remove it if it will just be put
7572 // back again. (Also from a pragmatic point of view
7573 // if it is a free block in a region that is beyond
7574 // any allocated blocks, an assertion will fail)
7575 // Remember the start of a free run.
7576 initialize_free_range(addr, true);
7577 // end - can coalesce with next chunk
7578 }
7579 } else {
7580 // the midst of a free range, we are coalescing
7581 print_free_block_coalesced(fc);
7582 if (CMSTraceSweeper) {
7583 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7584 }
7585 // remove it from the free lists
7586 _sp->removeFreeChunkFromFreeLists(fc);
7587 set_lastFreeRangeCoalesced(true);
7588 // If the chunk is being coalesced and the current free range is
7589 // in the free lists, remove the current free range so that it
7590 // will be returned to the free lists in its entirety - all
7591 // the coalesced pieces included.
7592 if (freeRangeInFreeLists()) {
7593 FreeChunk* ffc = (FreeChunk*) freeFinger();
7594 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7595 "Size of free range is inconsistent with chunk size.");
7596 if (CMSTestInFreeList) {
7597 assert(_sp->verify_chunk_in_free_list(ffc),
7598 "free range is not in free lists");
7599 }
7600 _sp->removeFreeChunkFromFreeLists(ffc);
7601 set_freeRangeInFreeLists(false);
7602 }
7603 }
7604 }
7605 // Note that if the chunk is not coalescable (the else arm
7606 // below), we unconditionally flush, without needing to do
7607 // a "lookahead," as we do below.
7608 if (inFreeRange()) lookahead_and_flush(fc, size);
7609 } else {
7610 // Code path common to both original and adaptive free lists.
7611
7612 // cant coalesce with previous block; this should be treated
7613 // as the end of a free run if any
7614 if (inFreeRange()) {
7615 // we kicked some butt; time to pick up the garbage
7616 assert(freeFinger() < addr, "freeFinger points too high");
7617 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7618 }
7619 // else, nothing to do, just continue
7620 }
7621 }
7622
7623 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7624 // This is a chunk of garbage. It is not in any free list.
7625 // Add it to a free list or let it possibly be coalesced into
7626 // a larger chunk.
7627 HeapWord* const addr = (HeapWord*) fc;
7628 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7629
7630 if (_sp->adaptive_freelists()) {
7631 // Verify that the bit map has no bits marked between
7632 // addr and purported end of just dead object.
7633 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7634
7635 do_post_free_or_garbage_chunk(fc, size);
7636 } else {
7637 if (!inFreeRange()) {
7638 // start of a new free range
7639 assert(size > 0, "A free range should have a size");
7640 initialize_free_range(addr, false);
7641 } else {
7642 // this will be swept up when we hit the end of the
7643 // free range
7644 if (CMSTraceSweeper) {
7645 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", p2i(fc), size);
7646 }
7647 // If the chunk is being coalesced and the current free range is
7648 // in the free lists, remove the current free range so that it
7649 // will be returned to the free lists in its entirety - all
7650 // the coalesced pieces included.
7651 if (freeRangeInFreeLists()) {
7652 FreeChunk* ffc = (FreeChunk*)freeFinger();
7653 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7654 "Size of free range is inconsistent with chunk size.");
7655 if (CMSTestInFreeList) {
7656 assert(_sp->verify_chunk_in_free_list(ffc),
7657 "free range is not in free lists");
7658 }
7659 _sp->removeFreeChunkFromFreeLists(ffc);
7660 set_freeRangeInFreeLists(false);
7661 }
7662 set_lastFreeRangeCoalesced(true);
7663 }
7664 // this will be swept up when we hit the end of the free range
7665
7666 // Verify that the bit map has no bits marked between
7667 // addr and purported end of just dead object.
7668 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7669 }
7670 assert(_limit >= addr + size,
7671 "A freshly garbage chunk can't possibly straddle over _limit");
7672 if (inFreeRange()) lookahead_and_flush(fc, size);
7673 return size;
7674 }
7675
7676 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7677 HeapWord* addr = (HeapWord*) fc;
7678 // The sweeper has just found a live object. Return any accumulated
7679 // left hand chunk to the free lists.
7680 if (inFreeRange()) {
7681 assert(freeFinger() < addr, "freeFinger points too high");
7682 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7683 }
7684
7685 // This object is live: we'd normally expect this to be
7686 // an oop, and like to assert the following:
7687 // assert(oop(addr)->is_oop(), "live block should be an oop");
7688 // However, as we commented above, this may be an object whose
7689 // header hasn't yet been initialized.
7690 size_t size;
7691 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7692 if (_bitMap->isMarked(addr + 1)) {
7693 // Determine the size from the bit map, rather than trying to
7694 // compute it from the object header.
7695 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7696 size = pointer_delta(nextOneAddr + 1, addr);
7697 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7698 "alignment problem");
7699
7700 #ifdef ASSERT
7701 if (oop(addr)->klass_or_null() != NULL) {
7702 // Ignore mark word because we are running concurrent with mutators
7703 assert(oop(addr)->is_oop(true), "live block should be an oop");
7704 assert(size ==
7705 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7706 "P-mark and computed size do not agree");
7707 }
7708 #endif
7709
7710 } else {
7711 // This should be an initialized object that's alive.
7712 assert(oop(addr)->klass_or_null() != NULL,
7713 "Should be an initialized object");
7714 // Ignore mark word because we are running concurrent with mutators
7715 assert(oop(addr)->is_oop(true), "live block should be an oop");
7716 // Verify that the bit map has no bits marked between
7717 // addr and purported end of this block.
7718 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7719 assert(size >= 3, "Necessary for Printezis marks to work");
7720 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7721 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7722 }
7723 return size;
7724 }
7725
7726 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7727 size_t chunkSize) {
7728 // do_post_free_or_garbage_chunk() should only be called in the case
7729 // of the adaptive free list allocator.
7730 const bool fcInFreeLists = fc->is_free();
7731 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
7732 assert((HeapWord*)fc <= _limit, "sweep invariant");
7733 if (CMSTestInFreeList && fcInFreeLists) {
7734 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7735 }
7736
7737 if (CMSTraceSweeper) {
7738 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7739 }
7740
7741 HeapWord* const fc_addr = (HeapWord*) fc;
7742
7743 bool coalesce;
7744 const size_t left = pointer_delta(fc_addr, freeFinger());
7745 const size_t right = chunkSize;
7746 switch (FLSCoalescePolicy) {
7747 // numeric value forms a coalition aggressiveness metric
7748 case 0: { // never coalesce
7749 coalesce = false;
7750 break;
7751 }
7752 case 1: { // coalesce if left & right chunks on overpopulated lists
7753 coalesce = _sp->coalOverPopulated(left) &&
7754 _sp->coalOverPopulated(right);
7755 break;
7756 }
7757 case 2: { // coalesce if left chunk on overpopulated list (default)
7758 coalesce = _sp->coalOverPopulated(left);
7759 break;
7760 }
7761 case 3: { // coalesce if left OR right chunk on overpopulated list
7762 coalesce = _sp->coalOverPopulated(left) ||
7763 _sp->coalOverPopulated(right);
7764 break;
7765 }
7766 case 4: { // always coalesce
7767 coalesce = true;
7768 break;
7769 }
7770 default:
7771 ShouldNotReachHere();
7772 }
7773
7774 // Should the current free range be coalesced?
7775 // If the chunk is in a free range and either we decided to coalesce above
7776 // or the chunk is near the large block at the end of the heap
7777 // (isNearLargestChunk() returns true), then coalesce this chunk.
7778 const bool doCoalesce = inFreeRange()
7779 && (coalesce || _g->isNearLargestChunk(fc_addr));
7780 if (doCoalesce) {
7781 // Coalesce the current free range on the left with the new
7782 // chunk on the right. If either is on a free list,
7783 // it must be removed from the list and stashed in the closure.
7784 if (freeRangeInFreeLists()) {
7785 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7786 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7787 "Size of free range is inconsistent with chunk size.");
7788 if (CMSTestInFreeList) {
7789 assert(_sp->verify_chunk_in_free_list(ffc),
7790 "Chunk is not in free lists");
7791 }
7792 _sp->coalDeath(ffc->size());
7793 _sp->removeFreeChunkFromFreeLists(ffc);
7794 set_freeRangeInFreeLists(false);
7795 }
7796 if (fcInFreeLists) {
7797 _sp->coalDeath(chunkSize);
7798 assert(fc->size() == chunkSize,
7799 "The chunk has the wrong size or is not in the free lists");
7800 _sp->removeFreeChunkFromFreeLists(fc);
7801 }
7802 set_lastFreeRangeCoalesced(true);
7803 print_free_block_coalesced(fc);
7804 } else { // not in a free range and/or should not coalesce
7805 // Return the current free range and start a new one.
7806 if (inFreeRange()) {
7807 // In a free range but cannot coalesce with the right hand chunk.
7808 // Put the current free range into the free lists.
7809 flush_cur_free_chunk(freeFinger(),
7810 pointer_delta(fc_addr, freeFinger()));
7811 }
7812 // Set up for new free range. Pass along whether the right hand
7813 // chunk is in the free lists.
7814 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7815 }
7816 }
7817
7818 // Lookahead flush:
7819 // If we are tracking a free range, and this is the last chunk that
7820 // we'll look at because its end crosses past _limit, we'll preemptively
7821 // flush it along with any free range we may be holding on to. Note that
7822 // this can be the case only for an already free or freshly garbage
7823 // chunk. If this block is an object, it can never straddle
7824 // over _limit. The "straddling" occurs when _limit is set at
7825 // the previous end of the space when this cycle started, and
7826 // a subsequent heap expansion caused the previously co-terminal
7827 // free block to be coalesced with the newly expanded portion,
7828 // thus rendering _limit a non-block-boundary making it dangerous
7829 // for the sweeper to step over and examine.
7830 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7831 assert(inFreeRange(), "Should only be called if currently in a free range.");
7832 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7833 assert(_sp->used_region().contains(eob - 1),
7834 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7835 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7836 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7837 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size));
7838 if (eob >= _limit) {
7839 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7840 if (CMSTraceSweeper) {
7841 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
7842 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7843 "[" PTR_FORMAT "," PTR_FORMAT ")",
7844 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7845 }
7846 // Return the storage we are tracking back into the free lists.
7847 if (CMSTraceSweeper) {
7848 gclog_or_tty->print_cr("Flushing ... ");
7849 }
7850 assert(freeFinger() < eob, "Error");
7851 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7852 }
7853 }
7854
7855 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7856 assert(inFreeRange(), "Should only be called if currently in a free range.");
7857 assert(size > 0,
7858 "A zero sized chunk cannot be added to the free lists.");
7859 if (!freeRangeInFreeLists()) {
7860 if (CMSTestInFreeList) {
7861 FreeChunk* fc = (FreeChunk*) chunk;
7862 fc->set_size(size);
7863 assert(!_sp->verify_chunk_in_free_list(fc),
7864 "chunk should not be in free lists yet");
7865 }
7866 if (CMSTraceSweeper) {
7867 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists",
7868 p2i(chunk), size);
7869 }
7870 // A new free range is going to be starting. The current
7871 // free range has not been added to the free lists yet or
7872 // was removed so add it back.
7873 // If the current free range was coalesced, then the death
7874 // of the free range was recorded. Record a birth now.
7875 if (lastFreeRangeCoalesced()) {
7876 _sp->coalBirth(size);
7877 }
7878 _sp->addChunkAndRepairOffsetTable(chunk, size,
7879 lastFreeRangeCoalesced());
7880 } else if (CMSTraceSweeper) {
7881 gclog_or_tty->print_cr("Already in free list: nothing to flush");
7882 }
7883 set_inFreeRange(false);
7884 set_freeRangeInFreeLists(false);
7885 }
7886
7887 // We take a break if we've been at this for a while,
7888 // so as to avoid monopolizing the locks involved.
7889 void SweepClosure::do_yield_work(HeapWord* addr) {
7890 // Return current free chunk being used for coalescing (if any)
7891 // to the appropriate freelist. After yielding, the next
7892 // free block encountered will start a coalescing range of
7893 // free blocks. If the next free block is adjacent to the
7894 // chunk just flushed, they will need to wait for the next
7895 // sweep to be coalesced.
7896 if (inFreeRange()) {
7897 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7898 }
7899
7900 // First give up the locks, then yield, then re-lock.
7901 // We should probably use a constructor/destructor idiom to
7902 // do this unlock/lock or modify the MutexUnlocker class to
7903 // serve our purpose. XXX
7904 assert_lock_strong(_bitMap->lock());
7905 assert_lock_strong(_freelistLock);
7906 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7907 "CMS thread should hold CMS token");
7908 _bitMap->lock()->unlock();
7909 _freelistLock->unlock();
7910 ConcurrentMarkSweepThread::desynchronize(true);
7911 _collector->stopTimer();
7912 if (PrintCMSStatistics != 0) {
7913 _collector->incrementYields();
7914 }
7915
7916 // See the comment in coordinator_yield()
7917 for (unsigned i = 0; i < CMSYieldSleepCount &&
7918 ConcurrentMarkSweepThread::should_yield() &&
7919 !CMSCollector::foregroundGCIsActive(); ++i) {
7920 os::sleep(Thread::current(), 1, false);
7921 }
7922
7923 ConcurrentMarkSweepThread::synchronize(true);
7924 _freelistLock->lock();
7925 _bitMap->lock()->lock_without_safepoint_check();
7926 _collector->startTimer();
7927 }
7928
7929 #ifndef PRODUCT
7930 // This is actually very useful in a product build if it can
7931 // be called from the debugger. Compile it into the product
7932 // as needed.
7933 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7934 return debug_cms_space->verify_chunk_in_free_list(fc);
7935 }
7936 #endif
7937
7938 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7939 if (CMSTraceSweeper) {
7940 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7941 p2i(fc), fc->size());
7942 }
7943 }
7944
7945 // CMSIsAliveClosure
7946 bool CMSIsAliveClosure::do_object_b(oop obj) {
7947 HeapWord* addr = (HeapWord*)obj;
7948 return addr != NULL &&
7949 (!_span.contains(addr) || _bit_map->isMarked(addr));
7950 }
7951
7952
7953 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7954 MemRegion span,
7955 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7956 bool cpc):
7957 _collector(collector),
7958 _span(span),
7959 _bit_map(bit_map),
7960 _mark_stack(mark_stack),
7961 _concurrent_precleaning(cpc) {
7962 assert(!_span.is_empty(), "Empty span could spell trouble");
7963 }
7964
7965
7966 // CMSKeepAliveClosure: the serial version
7967 void CMSKeepAliveClosure::do_oop(oop obj) {
7968 HeapWord* addr = (HeapWord*)obj;
7969 if (_span.contains(addr) &&
7970 !_bit_map->isMarked(addr)) {
7971 _bit_map->mark(addr);
7972 bool simulate_overflow = false;
7973 NOT_PRODUCT(
7974 if (CMSMarkStackOverflowALot &&
7975 _collector->simulate_overflow()) {
7976 // simulate a stack overflow
7977 simulate_overflow = true;
7978 }
7979 )
7980 if (simulate_overflow || !_mark_stack->push(obj)) {
7981 if (_concurrent_precleaning) {
7982 // We dirty the overflown object and let the remark
7983 // phase deal with it.
7984 assert(_collector->overflow_list_is_empty(), "Error");
7985 // In the case of object arrays, we need to dirty all of
7986 // the cards that the object spans. No locking or atomics
7987 // are needed since no one else can be mutating the mod union
7988 // table.
7989 if (obj->is_objArray()) {
7990 size_t sz = obj->size();
7991 HeapWord* end_card_addr =
7992 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
7993 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7994 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7995 _collector->_modUnionTable.mark_range(redirty_range);
7996 } else {
7997 _collector->_modUnionTable.mark(addr);
7998 }
7999 _collector->_ser_kac_preclean_ovflw++;
8000 } else {
8001 _collector->push_on_overflow_list(obj);
8002 _collector->_ser_kac_ovflw++;
8003 }
8004 }
8005 }
8006 }
8007
8008 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8009 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8010
8011 // CMSParKeepAliveClosure: a parallel version of the above.
8012 // The work queues are private to each closure (thread),
8013 // but (may be) available for stealing by other threads.
8014 void CMSParKeepAliveClosure::do_oop(oop obj) {
8015 HeapWord* addr = (HeapWord*)obj;
8016 if (_span.contains(addr) &&
8017 !_bit_map->isMarked(addr)) {
8018 // In general, during recursive tracing, several threads
8019 // may be concurrently getting here; the first one to
8020 // "tag" it, claims it.
8021 if (_bit_map->par_mark(addr)) {
8022 bool res = _work_queue->push(obj);
8023 assert(res, "Low water mark should be much less than capacity");
8024 // Do a recursive trim in the hope that this will keep
8025 // stack usage lower, but leave some oops for potential stealers
8026 trim_queue(_low_water_mark);
8027 } // Else, another thread got there first
8028 }
8029 }
8030
8031 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8032 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8033
8034 void CMSParKeepAliveClosure::trim_queue(uint max) {
8035 while (_work_queue->size() > max) {
8036 oop new_oop;
8037 if (_work_queue->pop_local(new_oop)) {
8038 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8039 assert(_bit_map->isMarked((HeapWord*)new_oop),
8040 "no white objects on this stack!");
8041 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8042 // iterate over the oops in this oop, marking and pushing
8043 // the ones in CMS heap (i.e. in _span).
8044 new_oop->oop_iterate(&_mark_and_push);
8045 }
8046 }
8047 }
8048
8049 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8050 CMSCollector* collector,
8051 MemRegion span, CMSBitMap* bit_map,
8052 OopTaskQueue* work_queue):
8053 _collector(collector),
8054 _span(span),
8055 _bit_map(bit_map),
8056 _work_queue(work_queue) { }
8057
8058 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8059 HeapWord* addr = (HeapWord*)obj;
8060 if (_span.contains(addr) &&
8061 !_bit_map->isMarked(addr)) {
8062 if (_bit_map->par_mark(addr)) {
8063 bool simulate_overflow = false;
8064 NOT_PRODUCT(
8065 if (CMSMarkStackOverflowALot &&
8066 _collector->par_simulate_overflow()) {
8067 // simulate a stack overflow
8068 simulate_overflow = true;
8069 }
8070 )
8071 if (simulate_overflow || !_work_queue->push(obj)) {
8072 _collector->par_push_on_overflow_list(obj);
8073 _collector->_par_kac_ovflw++;
8074 }
8075 } // Else another thread got there already
8076 }
8077 }
8078
8079 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8080 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8081
8082 //////////////////////////////////////////////////////////////////
8083 // CMSExpansionCause /////////////////////////////
8084 //////////////////////////////////////////////////////////////////
8085 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8086 switch (cause) {
8087 case _no_expansion:
8088 return "No expansion";
8089 case _satisfy_free_ratio:
8090 return "Free ratio";
8091 case _satisfy_promotion:
8092 return "Satisfy promotion";
8093 case _satisfy_allocation:
8094 return "allocation";
8095 case _allocate_par_lab:
8096 return "Par LAB";
8097 case _allocate_par_spooling_space:
8098 return "Par Spooling Space";
8099 case _adaptive_size_policy:
8100 return "Ergonomics";
8101 default:
8102 return "unknown";
8103 }
8104 }
8105
8106 void CMSDrainMarkingStackClosure::do_void() {
8107 // the max number to take from overflow list at a time
8108 const size_t num = _mark_stack->capacity()/4;
8109 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8110 "Overflow list should be NULL during concurrent phases");
8111 while (!_mark_stack->isEmpty() ||
8112 // if stack is empty, check the overflow list
8113 _collector->take_from_overflow_list(num, _mark_stack)) {
8114 oop obj = _mark_stack->pop();
8115 HeapWord* addr = (HeapWord*)obj;
8116 assert(_span.contains(addr), "Should be within span");
8117 assert(_bit_map->isMarked(addr), "Should be marked");
8118 assert(obj->is_oop(), "Should be an oop");
8119 obj->oop_iterate(_keep_alive);
8120 }
8121 }
8122
8123 void CMSParDrainMarkingStackClosure::do_void() {
8124 // drain queue
8125 trim_queue(0);
8126 }
8127
8128 // Trim our work_queue so its length is below max at return
8129 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8130 while (_work_queue->size() > max) {
8131 oop new_oop;
8132 if (_work_queue->pop_local(new_oop)) {
8133 assert(new_oop->is_oop(), "Expected an oop");
8134 assert(_bit_map->isMarked((HeapWord*)new_oop),
8135 "no white objects on this stack!");
8136 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8137 // iterate over the oops in this oop, marking and pushing
8138 // the ones in CMS heap (i.e. in _span).
8139 new_oop->oop_iterate(&_mark_and_push);
8140 }
8141 }
8142 }
8143
8144 ////////////////////////////////////////////////////////////////////
8145 // Support for Marking Stack Overflow list handling and related code
8146 ////////////////////////////////////////////////////////////////////
8147 // Much of the following code is similar in shape and spirit to the
8148 // code used in ParNewGC. We should try and share that code
8149 // as much as possible in the future.
8150
8151 #ifndef PRODUCT
8152 // Debugging support for CMSStackOverflowALot
8153
8154 // It's OK to call this multi-threaded; the worst thing
8155 // that can happen is that we'll get a bunch of closely
8156 // spaced simulated overflows, but that's OK, in fact
8157 // probably good as it would exercise the overflow code
8158 // under contention.
8159 bool CMSCollector::simulate_overflow() {
8160 if (_overflow_counter-- <= 0) { // just being defensive
8161 _overflow_counter = CMSMarkStackOverflowInterval;
8162 return true;
8163 } else {
8164 return false;
8165 }
8166 }
8167
8168 bool CMSCollector::par_simulate_overflow() {
8169 return simulate_overflow();
8170 }
8171 #endif
8172
8173 // Single-threaded
8174 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8175 assert(stack->isEmpty(), "Expected precondition");
8176 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8177 size_t i = num;
8178 oop cur = _overflow_list;
8179 const markOop proto = markOopDesc::prototype();
8180 NOT_PRODUCT(ssize_t n = 0;)
8181 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8182 next = oop(cur->mark());
8183 cur->set_mark(proto); // until proven otherwise
8184 assert(cur->is_oop(), "Should be an oop");
8185 bool res = stack->push(cur);
8186 assert(res, "Bit off more than can chew?");
8187 NOT_PRODUCT(n++;)
8188 }
8189 _overflow_list = cur;
8190 #ifndef PRODUCT
8191 assert(_num_par_pushes >= n, "Too many pops?");
8192 _num_par_pushes -=n;
8193 #endif
8194 return !stack->isEmpty();
8195 }
8196
8197 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
8198 // (MT-safe) Get a prefix of at most "num" from the list.
8199 // The overflow list is chained through the mark word of
8200 // each object in the list. We fetch the entire list,
8201 // break off a prefix of the right size and return the
8202 // remainder. If other threads try to take objects from
8203 // the overflow list at that time, they will wait for
8204 // some time to see if data becomes available. If (and
8205 // only if) another thread places one or more object(s)
8206 // on the global list before we have returned the suffix
8207 // to the global list, we will walk down our local list
8208 // to find its end and append the global list to
8209 // our suffix before returning it. This suffix walk can
8210 // prove to be expensive (quadratic in the amount of traffic)
8211 // when there are many objects in the overflow list and
8212 // there is much producer-consumer contention on the list.
8213 // *NOTE*: The overflow list manipulation code here and
8214 // in ParNewGeneration:: are very similar in shape,
8215 // except that in the ParNew case we use the old (from/eden)
8216 // copy of the object to thread the list via its klass word.
8217 // Because of the common code, if you make any changes in
8218 // the code below, please check the ParNew version to see if
8219 // similar changes might be needed.
8220 // CR 6797058 has been filed to consolidate the common code.
8221 bool CMSCollector::par_take_from_overflow_list(size_t num,
8222 OopTaskQueue* work_q,
8223 int no_of_gc_threads) {
8224 assert(work_q->size() == 0, "First empty local work queue");
8225 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8226 if (_overflow_list == NULL) {
8227 return false;
8228 }
8229 // Grab the entire list; we'll put back a suffix
8230 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8231 Thread* tid = Thread::current();
8232 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8233 // set to ParallelGCThreads.
8234 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8235 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8236 // If the list is busy, we spin for a short while,
8237 // sleeping between attempts to get the list.
8238 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8239 os::sleep(tid, sleep_time_millis, false);
8240 if (_overflow_list == NULL) {
8241 // Nothing left to take
8242 return false;
8243 } else if (_overflow_list != BUSY) {
8244 // Try and grab the prefix
8245 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
8246 }
8247 }
8248 // If the list was found to be empty, or we spun long
8249 // enough, we give up and return empty-handed. If we leave
8250 // the list in the BUSY state below, it must be the case that
8251 // some other thread holds the overflow list and will set it
8252 // to a non-BUSY state in the future.
8253 if (prefix == NULL || prefix == BUSY) {
8254 // Nothing to take or waited long enough
8255 if (prefix == NULL) {
8256 // Write back the NULL in case we overwrote it with BUSY above
8257 // and it is still the same value.
8258 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8259 }
8260 return false;
8261 }
8262 assert(prefix != NULL && prefix != BUSY, "Error");
8263 size_t i = num;
8264 oop cur = prefix;
8265 // Walk down the first "num" objects, unless we reach the end.
8266 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8267 if (cur->mark() == NULL) {
8268 // We have "num" or fewer elements in the list, so there
8269 // is nothing to return to the global list.
8270 // Write back the NULL in lieu of the BUSY we wrote
8271 // above, if it is still the same value.
8272 if (_overflow_list == BUSY) {
8273 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8274 }
8275 } else {
8276 // Chop off the suffix and return it to the global list.
8277 assert(cur->mark() != BUSY, "Error");
8278 oop suffix_head = cur->mark(); // suffix will be put back on global list
8279 cur->set_mark(NULL); // break off suffix
8280 // It's possible that the list is still in the empty(busy) state
8281 // we left it in a short while ago; in that case we may be
8282 // able to place back the suffix without incurring the cost
8283 // of a walk down the list.
8284 oop observed_overflow_list = _overflow_list;
8285 oop cur_overflow_list = observed_overflow_list;
8286 bool attached = false;
8287 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8288 observed_overflow_list =
8289 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8290 if (cur_overflow_list == observed_overflow_list) {
8291 attached = true;
8292 break;
8293 } else cur_overflow_list = observed_overflow_list;
8294 }
8295 if (!attached) {
8296 // Too bad, someone else sneaked in (at least) an element; we'll need
8297 // to do a splice. Find tail of suffix so we can prepend suffix to global
8298 // list.
8299 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8300 oop suffix_tail = cur;
8301 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8302 "Tautology");
8303 observed_overflow_list = _overflow_list;
8304 do {
8305 cur_overflow_list = observed_overflow_list;
8306 if (cur_overflow_list != BUSY) {
8307 // Do the splice ...
8308 suffix_tail->set_mark(markOop(cur_overflow_list));
8309 } else { // cur_overflow_list == BUSY
8310 suffix_tail->set_mark(NULL);
8311 }
8312 // ... and try to place spliced list back on overflow_list ...
8313 observed_overflow_list =
8314 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8315 } while (cur_overflow_list != observed_overflow_list);
8316 // ... until we have succeeded in doing so.
8317 }
8318 }
8319
8320 // Push the prefix elements on work_q
8321 assert(prefix != NULL, "control point invariant");
8322 const markOop proto = markOopDesc::prototype();
8323 oop next;
8324 NOT_PRODUCT(ssize_t n = 0;)
8325 for (cur = prefix; cur != NULL; cur = next) {
8326 next = oop(cur->mark());
8327 cur->set_mark(proto); // until proven otherwise
8328 assert(cur->is_oop(), "Should be an oop");
8329 bool res = work_q->push(cur);
8330 assert(res, "Bit off more than we can chew?");
8331 NOT_PRODUCT(n++;)
8332 }
8333 #ifndef PRODUCT
8334 assert(_num_par_pushes >= n, "Too many pops?");
8335 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8336 #endif
8337 return true;
8338 }
8339
8340 // Single-threaded
8341 void CMSCollector::push_on_overflow_list(oop p) {
8342 NOT_PRODUCT(_num_par_pushes++;)
8343 assert(p->is_oop(), "Not an oop");
8344 preserve_mark_if_necessary(p);
8345 p->set_mark((markOop)_overflow_list);
8346 _overflow_list = p;
8347 }
8348
8349 // Multi-threaded; use CAS to prepend to overflow list
8350 void CMSCollector::par_push_on_overflow_list(oop p) {
8351 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8352 assert(p->is_oop(), "Not an oop");
8353 par_preserve_mark_if_necessary(p);
8354 oop observed_overflow_list = _overflow_list;
8355 oop cur_overflow_list;
8356 do {
8357 cur_overflow_list = observed_overflow_list;
8358 if (cur_overflow_list != BUSY) {
8359 p->set_mark(markOop(cur_overflow_list));
8360 } else {
8361 p->set_mark(NULL);
8362 }
8363 observed_overflow_list =
8364 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8365 } while (cur_overflow_list != observed_overflow_list);
8366 }
8367 #undef BUSY
8368
8369 // Single threaded
8370 // General Note on GrowableArray: pushes may silently fail
8371 // because we are (temporarily) out of C-heap for expanding
8372 // the stack. The problem is quite ubiquitous and affects
8373 // a lot of code in the JVM. The prudent thing for GrowableArray
8374 // to do (for now) is to exit with an error. However, that may
8375 // be too draconian in some cases because the caller may be
8376 // able to recover without much harm. For such cases, we
8377 // should probably introduce a "soft_push" method which returns
8378 // an indication of success or failure with the assumption that
8379 // the caller may be able to recover from a failure; code in
8380 // the VM can then be changed, incrementally, to deal with such
8381 // failures where possible, thus, incrementally hardening the VM
8382 // in such low resource situations.
8383 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8384 _preserved_oop_stack.push(p);
8385 _preserved_mark_stack.push(m);
8386 assert(m == p->mark(), "Mark word changed");
8387 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8388 "bijection");
8389 }
8390
8391 // Single threaded
8392 void CMSCollector::preserve_mark_if_necessary(oop p) {
8393 markOop m = p->mark();
8394 if (m->must_be_preserved(p)) {
8395 preserve_mark_work(p, m);
8396 }
8397 }
8398
8399 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8400 markOop m = p->mark();
8401 if (m->must_be_preserved(p)) {
8402 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8403 // Even though we read the mark word without holding
8404 // the lock, we are assured that it will not change
8405 // because we "own" this oop, so no other thread can
8406 // be trying to push it on the overflow list; see
8407 // the assertion in preserve_mark_work() that checks
8408 // that m == p->mark().
8409 preserve_mark_work(p, m);
8410 }
8411 }
8412
8413 // We should be able to do this multi-threaded,
8414 // a chunk of stack being a task (this is
8415 // correct because each oop only ever appears
8416 // once in the overflow list. However, it's
8417 // not very easy to completely overlap this with
8418 // other operations, so will generally not be done
8419 // until all work's been completed. Because we
8420 // expect the preserved oop stack (set) to be small,
8421 // it's probably fine to do this single-threaded.
8422 // We can explore cleverer concurrent/overlapped/parallel
8423 // processing of preserved marks if we feel the
8424 // need for this in the future. Stack overflow should
8425 // be so rare in practice and, when it happens, its
8426 // effect on performance so great that this will
8427 // likely just be in the noise anyway.
8428 void CMSCollector::restore_preserved_marks_if_any() {
8429 assert(SafepointSynchronize::is_at_safepoint(),
8430 "world should be stopped");
8431 assert(Thread::current()->is_ConcurrentGC_thread() ||
8432 Thread::current()->is_VM_thread(),
8433 "should be single-threaded");
8434 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8435 "bijection");
8436
8437 while (!_preserved_oop_stack.is_empty()) {
8438 oop p = _preserved_oop_stack.pop();
8439 assert(p->is_oop(), "Should be an oop");
8440 assert(_span.contains(p), "oop should be in _span");
8441 assert(p->mark() == markOopDesc::prototype(),
8442 "Set when taken from overflow list");
8443 markOop m = _preserved_mark_stack.pop();
8444 p->set_mark(m);
8445 }
8446 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8447 "stacks were cleared above");
8448 }
8449
8450 #ifndef PRODUCT
8451 bool CMSCollector::no_preserved_marks() const {
8452 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8453 }
8454 #endif
8455
8456 // Transfer some number of overflown objects to usual marking
8457 // stack. Return true if some objects were transferred.
8458 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8459 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8460 (size_t)ParGCDesiredObjsFromOverflowList);
8461
8462 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8463 assert(_collector->overflow_list_is_empty() || res,
8464 "If list is not empty, we should have taken something");
8465 assert(!res || !_mark_stack->isEmpty(),
8466 "If we took something, it should now be on our stack");
8467 return res;
8468 }
8469
8470 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8471 size_t res = _sp->block_size_no_stall(addr, _collector);
8472 if (_sp->block_is_obj(addr)) {
8473 if (_live_bit_map->isMarked(addr)) {
8474 // It can't have been dead in a previous cycle
8475 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8476 } else {
8477 _dead_bit_map->mark(addr); // mark the dead object
8478 }
8479 }
8480 // Could be 0, if the block size could not be computed without stalling.
8481 return res;
8482 }
8483
8484 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8485
8486 switch (phase) {
8487 case CMSCollector::InitialMarking:
8488 initialize(true /* fullGC */ ,
8489 cause /* cause of the GC */,
8490 true /* recordGCBeginTime */,
8491 true /* recordPreGCUsage */,
8492 false /* recordPeakUsage */,
8493 false /* recordPostGCusage */,
8494 true /* recordAccumulatedGCTime */,
8495 false /* recordGCEndTime */,
8496 false /* countCollection */ );
8497 break;
8498
8499 case CMSCollector::FinalMarking:
8500 initialize(true /* fullGC */ ,
8501 cause /* cause of the GC */,
8502 false /* recordGCBeginTime */,
8503 false /* recordPreGCUsage */,
8504 false /* recordPeakUsage */,
8505 false /* recordPostGCusage */,
8506 true /* recordAccumulatedGCTime */,
8507 false /* recordGCEndTime */,
8508 false /* countCollection */ );
8509 break;
8510
8511 case CMSCollector::Sweeping:
8512 initialize(true /* fullGC */ ,
8513 cause /* cause of the GC */,
8514 false /* recordGCBeginTime */,
8515 false /* recordPreGCUsage */,
8516 true /* recordPeakUsage */,
8517 true /* recordPostGCusage */,
8518 false /* recordAccumulatedGCTime */,
8519 true /* recordGCEndTime */,
8520 true /* countCollection */ );
8521 break;
8522
8523 default:
8524 ShouldNotReachHere();
8525 }
8526 }