Thu, 12 Jun 2008 13:50:55 -0700
Merge
1 /*
2 * Copyright 2001-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 # include "incls/_precompiled.incl"
26 # include "incls/_concurrentMarkSweepGeneration.cpp.incl"
28 // statics
29 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
30 bool CMSCollector::_full_gc_requested = false;
32 //////////////////////////////////////////////////////////////////
33 // In support of CMS/VM thread synchronization
34 //////////////////////////////////////////////////////////////////
35 // We split use of the CGC_lock into 2 "levels".
36 // The low-level locking is of the usual CGC_lock monitor. We introduce
37 // a higher level "token" (hereafter "CMS token") built on top of the
38 // low level monitor (hereafter "CGC lock").
39 // The token-passing protocol gives priority to the VM thread. The
40 // CMS-lock doesn't provide any fairness guarantees, but clients
41 // should ensure that it is only held for very short, bounded
42 // durations.
43 //
44 // When either of the CMS thread or the VM thread is involved in
45 // collection operations during which it does not want the other
46 // thread to interfere, it obtains the CMS token.
47 //
48 // If either thread tries to get the token while the other has
49 // it, that thread waits. However, if the VM thread and CMS thread
50 // both want the token, then the VM thread gets priority while the
51 // CMS thread waits. This ensures, for instance, that the "concurrent"
52 // phases of the CMS thread's work do not block out the VM thread
53 // for long periods of time as the CMS thread continues to hog
54 // the token. (See bug 4616232).
55 //
56 // The baton-passing functions are, however, controlled by the
57 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
58 // and here the low-level CMS lock, not the high level token,
59 // ensures mutual exclusion.
60 //
61 // Two important conditions that we have to satisfy:
62 // 1. if a thread does a low-level wait on the CMS lock, then it
63 // relinquishes the CMS token if it were holding that token
64 // when it acquired the low-level CMS lock.
65 // 2. any low-level notifications on the low-level lock
66 // should only be sent when a thread has relinquished the token.
67 //
68 // In the absence of either property, we'd have potential deadlock.
69 //
70 // We protect each of the CMS (concurrent and sequential) phases
71 // with the CMS _token_, not the CMS _lock_.
72 //
73 // The only code protected by CMS lock is the token acquisition code
74 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
75 // baton-passing code.
76 //
77 // Unfortunately, i couldn't come up with a good abstraction to factor and
78 // hide the naked CGC_lock manipulation in the baton-passing code
79 // further below. That's something we should try to do. Also, the proof
80 // of correctness of this 2-level locking scheme is far from obvious,
81 // and potentially quite slippery. We have an uneasy supsicion, for instance,
82 // that there may be a theoretical possibility of delay/starvation in the
83 // low-level lock/wait/notify scheme used for the baton-passing because of
84 // potential intereference with the priority scheme embodied in the
85 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
86 // invocation further below and marked with "XXX 20011219YSR".
87 // Indeed, as we note elsewhere, this may become yet more slippery
88 // in the presence of multiple CMS and/or multiple VM threads. XXX
90 class CMSTokenSync: public StackObj {
91 private:
92 bool _is_cms_thread;
93 public:
94 CMSTokenSync(bool is_cms_thread):
95 _is_cms_thread(is_cms_thread) {
96 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
97 "Incorrect argument to constructor");
98 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
99 }
101 ~CMSTokenSync() {
102 assert(_is_cms_thread ?
103 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
104 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
105 "Incorrect state");
106 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
107 }
108 };
110 // Convenience class that does a CMSTokenSync, and then acquires
111 // upto three locks.
112 class CMSTokenSyncWithLocks: public CMSTokenSync {
113 private:
114 // Note: locks are acquired in textual declaration order
115 // and released in the opposite order
116 MutexLockerEx _locker1, _locker2, _locker3;
117 public:
118 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
119 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
120 CMSTokenSync(is_cms_thread),
121 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
122 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
123 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
124 { }
125 };
128 // Wrapper class to temporarily disable icms during a foreground cms collection.
129 class ICMSDisabler: public StackObj {
130 public:
131 // The ctor disables icms and wakes up the thread so it notices the change;
132 // the dtor re-enables icms. Note that the CMSCollector methods will check
133 // CMSIncrementalMode.
134 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
135 ~ICMSDisabler() { CMSCollector::enable_icms(); }
136 };
138 //////////////////////////////////////////////////////////////////
139 // Concurrent Mark-Sweep Generation /////////////////////////////
140 //////////////////////////////////////////////////////////////////
142 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
144 // This struct contains per-thread things necessary to support parallel
145 // young-gen collection.
146 class CMSParGCThreadState: public CHeapObj {
147 public:
148 CFLS_LAB lab;
149 PromotionInfo promo;
151 // Constructor.
152 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
153 promo.setSpace(cfls);
154 }
155 };
157 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
158 ReservedSpace rs, size_t initial_byte_size, int level,
159 CardTableRS* ct, bool use_adaptive_freelists,
160 FreeBlockDictionary::DictionaryChoice dictionaryChoice) :
161 CardGeneration(rs, initial_byte_size, level, ct),
162 _dilatation_factor(((double)MinChunkSize)/((double)(oopDesc::header_size()))),
163 _debug_collection_type(Concurrent_collection_type)
164 {
165 HeapWord* bottom = (HeapWord*) _virtual_space.low();
166 HeapWord* end = (HeapWord*) _virtual_space.high();
168 _direct_allocated_words = 0;
169 NOT_PRODUCT(
170 _numObjectsPromoted = 0;
171 _numWordsPromoted = 0;
172 _numObjectsAllocated = 0;
173 _numWordsAllocated = 0;
174 )
176 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
177 use_adaptive_freelists,
178 dictionaryChoice);
179 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
180 if (_cmsSpace == NULL) {
181 vm_exit_during_initialization(
182 "CompactibleFreeListSpace allocation failure");
183 }
184 _cmsSpace->_gen = this;
186 _gc_stats = new CMSGCStats();
188 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
189 // offsets match. The ability to tell free chunks from objects
190 // depends on this property.
191 debug_only(
192 FreeChunk* junk = NULL;
193 assert(UseCompressedOops ||
194 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
195 "Offset of FreeChunk::_prev within FreeChunk must match"
196 " that of OopDesc::_klass within OopDesc");
197 )
198 if (ParallelGCThreads > 0) {
199 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
200 _par_gc_thread_states =
201 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
202 if (_par_gc_thread_states == NULL) {
203 vm_exit_during_initialization("Could not allocate par gc structs");
204 }
205 for (uint i = 0; i < ParallelGCThreads; i++) {
206 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
207 if (_par_gc_thread_states[i] == NULL) {
208 vm_exit_during_initialization("Could not allocate par gc structs");
209 }
210 }
211 } else {
212 _par_gc_thread_states = NULL;
213 }
214 _incremental_collection_failed = false;
215 // The "dilatation_factor" is the expansion that can occur on
216 // account of the fact that the minimum object size in the CMS
217 // generation may be larger than that in, say, a contiguous young
218 // generation.
219 // Ideally, in the calculation below, we'd compute the dilatation
220 // factor as: MinChunkSize/(promoting_gen's min object size)
221 // Since we do not have such a general query interface for the
222 // promoting generation, we'll instead just use the mimimum
223 // object size (which today is a header's worth of space);
224 // note that all arithmetic is in units of HeapWords.
225 assert(MinChunkSize >= oopDesc::header_size(), "just checking");
226 assert(_dilatation_factor >= 1.0, "from previous assert");
227 }
230 // The field "_initiating_occupancy" represents the occupancy percentage
231 // at which we trigger a new collection cycle. Unless explicitly specified
232 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
233 // is calculated by:
234 //
235 // Let "f" be MinHeapFreeRatio in
236 //
237 // _intiating_occupancy = 100-f +
238 // f * (CMSTrigger[Perm]Ratio/100)
239 // where CMSTrigger[Perm]Ratio is the argument "tr" below.
240 //
241 // That is, if we assume the heap is at its desired maximum occupancy at the
242 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
243 // space be allocated before initiating a new collection cycle.
244 //
245 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
246 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
247 if (io >= 0) {
248 _initiating_occupancy = (double)io / 100.0;
249 } else {
250 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
251 (double)(tr * MinHeapFreeRatio) / 100.0)
252 / 100.0;
253 }
254 }
257 void ConcurrentMarkSweepGeneration::ref_processor_init() {
258 assert(collector() != NULL, "no collector");
259 collector()->ref_processor_init();
260 }
262 void CMSCollector::ref_processor_init() {
263 if (_ref_processor == NULL) {
264 // Allocate and initialize a reference processor
265 _ref_processor = ReferenceProcessor::create_ref_processor(
266 _span, // span
267 _cmsGen->refs_discovery_is_atomic(), // atomic_discovery
268 _cmsGen->refs_discovery_is_mt(), // mt_discovery
269 &_is_alive_closure,
270 ParallelGCThreads,
271 ParallelRefProcEnabled);
272 // Initialize the _ref_processor field of CMSGen
273 _cmsGen->set_ref_processor(_ref_processor);
275 // Allocate a dummy ref processor for perm gen.
276 ReferenceProcessor* rp2 = new ReferenceProcessor();
277 if (rp2 == NULL) {
278 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
279 }
280 _permGen->set_ref_processor(rp2);
281 }
282 }
284 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
285 GenCollectedHeap* gch = GenCollectedHeap::heap();
286 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
287 "Wrong type of heap");
288 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
289 gch->gen_policy()->size_policy();
290 assert(sp->is_gc_cms_adaptive_size_policy(),
291 "Wrong type of size policy");
292 return sp;
293 }
295 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
296 CMSGCAdaptivePolicyCounters* results =
297 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
298 assert(
299 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
300 "Wrong gc policy counter kind");
301 return results;
302 }
305 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
307 const char* gen_name = "old";
309 // Generation Counters - generation 1, 1 subspace
310 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
312 _space_counters = new GSpaceCounters(gen_name, 0,
313 _virtual_space.reserved_size(),
314 this, _gen_counters);
315 }
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;
323 // Initialize the alphas to the bootstrap value of 100.
324 _gc0_alpha = _cms_alpha = 100;
326 _cms_begin_time.update();
327 _cms_end_time.update();
329 _gc0_duration = 0.0;
330 _gc0_period = 0.0;
331 _gc0_promoted = 0;
333 _cms_duration = 0.0;
334 _cms_period = 0.0;
335 _cms_allocated = 0;
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 _icms_duty_cycle = CMSIncrementalDutyCycle;
342 }
344 // If promotion failure handling is on use
345 // the padded average size of the promotion for each
346 // young generation collection.
347 double CMSStats::time_until_cms_gen_full() const {
348 size_t cms_free = _cms_gen->cmsSpace()->free();
349 GenCollectedHeap* gch = GenCollectedHeap::heap();
350 size_t expected_promotion = gch->get_gen(0)->capacity();
351 if (HandlePromotionFailure) {
352 expected_promotion = MIN2(
353 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(),
354 expected_promotion);
355 }
356 if (cms_free > expected_promotion) {
357 // Start a cms collection if there isn't enough space to promote
358 // for the next minor collection. Use the padded average as
359 // a safety factor.
360 cms_free -= expected_promotion;
362 // Adjust by the safety factor.
363 double cms_free_dbl = (double)cms_free;
364 cms_free_dbl = cms_free_dbl * (100.0 - CMSIncrementalSafetyFactor) / 100.0;
366 if (PrintGCDetails && Verbose) {
367 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
368 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
369 cms_free, expected_promotion);
370 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
371 cms_free_dbl, cms_consumption_rate() + 1.0);
372 }
373 // Add 1 in case the consumption rate goes to zero.
374 return cms_free_dbl / (cms_consumption_rate() + 1.0);
375 }
376 return 0.0;
377 }
379 // Compare the duration of the cms collection to the
380 // time remaining before the cms generation is empty.
381 // Note that the time from the start of the cms collection
382 // to the start of the cms sweep (less than the total
383 // duration of the cms collection) can be used. This
384 // has been tried and some applications experienced
385 // promotion failures early in execution. This was
386 // possibly because the averages were not accurate
387 // enough at the beginning.
388 double CMSStats::time_until_cms_start() const {
389 // We add "gc0_period" to the "work" calculation
390 // below because this query is done (mostly) at the
391 // end of a scavenge, so we need to conservatively
392 // account for that much possible delay
393 // in the query so as to avoid concurrent mode failures
394 // due to starting the collection just a wee bit too
395 // late.
396 double work = cms_duration() + gc0_period();
397 double deadline = time_until_cms_gen_full();
398 if (work > deadline) {
399 if (Verbose && PrintGCDetails) {
400 gclog_or_tty->print(
401 " CMSCollector: collect because of anticipated promotion "
402 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
403 gc0_period(), time_until_cms_gen_full());
404 }
405 return 0.0;
406 }
407 return work - deadline;
408 }
410 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
411 // amount of change to prevent wild oscillation.
412 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
413 unsigned int new_duty_cycle) {
414 assert(old_duty_cycle <= 100, "bad input value");
415 assert(new_duty_cycle <= 100, "bad input value");
417 // Note: use subtraction with caution since it may underflow (values are
418 // unsigned). Addition is safe since we're in the range 0-100.
419 unsigned int damped_duty_cycle = new_duty_cycle;
420 if (new_duty_cycle < old_duty_cycle) {
421 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
422 if (new_duty_cycle + largest_delta < old_duty_cycle) {
423 damped_duty_cycle = old_duty_cycle - largest_delta;
424 }
425 } else if (new_duty_cycle > old_duty_cycle) {
426 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
427 if (new_duty_cycle > old_duty_cycle + largest_delta) {
428 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
429 }
430 }
431 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
433 if (CMSTraceIncrementalPacing) {
434 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
435 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
436 }
437 return damped_duty_cycle;
438 }
440 unsigned int CMSStats::icms_update_duty_cycle_impl() {
441 assert(CMSIncrementalPacing && valid(),
442 "should be handled in icms_update_duty_cycle()");
444 double cms_time_so_far = cms_timer().seconds();
445 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
446 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
448 // Avoid division by 0.
449 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
450 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
452 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
453 if (new_duty_cycle > _icms_duty_cycle) {
454 // Avoid very small duty cycles (1 or 2); 0 is allowed.
455 if (new_duty_cycle > 2) {
456 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
457 new_duty_cycle);
458 }
459 } else if (_allow_duty_cycle_reduction) {
460 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
461 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
462 // Respect the minimum duty cycle.
463 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
464 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
465 }
467 if (PrintGCDetails || CMSTraceIncrementalPacing) {
468 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
469 }
471 _allow_duty_cycle_reduction = false;
472 return _icms_duty_cycle;
473 }
475 #ifndef PRODUCT
476 void CMSStats::print_on(outputStream *st) const {
477 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
478 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
479 gc0_duration(), gc0_period(), gc0_promoted());
480 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
481 cms_duration(), cms_duration_per_mb(),
482 cms_period(), cms_allocated());
483 st->print(",cms_since_beg=%g,cms_since_end=%g",
484 cms_time_since_begin(), cms_time_since_end());
485 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
486 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
487 if (CMSIncrementalMode) {
488 st->print(",dc=%d", icms_duty_cycle());
489 }
491 if (valid()) {
492 st->print(",promo_rate=%g,cms_alloc_rate=%g",
493 promotion_rate(), cms_allocation_rate());
494 st->print(",cms_consumption_rate=%g,time_until_full=%g",
495 cms_consumption_rate(), time_until_cms_gen_full());
496 }
497 st->print(" ");
498 }
499 #endif // #ifndef PRODUCT
501 CMSCollector::CollectorState CMSCollector::_collectorState =
502 CMSCollector::Idling;
503 bool CMSCollector::_foregroundGCIsActive = false;
504 bool CMSCollector::_foregroundGCShouldWait = false;
506 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
507 ConcurrentMarkSweepGeneration* permGen,
508 CardTableRS* ct,
509 ConcurrentMarkSweepPolicy* cp):
510 _cmsGen(cmsGen),
511 _permGen(permGen),
512 _ct(ct),
513 _ref_processor(NULL), // will be set later
514 _conc_workers(NULL), // may be set later
515 _abort_preclean(false),
516 _start_sampling(false),
517 _between_prologue_and_epilogue(false),
518 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
519 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
520 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
521 -1 /* lock-free */, "No_lock" /* dummy */),
522 _modUnionClosure(&_modUnionTable),
523 _modUnionClosurePar(&_modUnionTable),
524 // Adjust my span to cover old (cms) gen and perm gen
525 _span(cmsGen->reserved()._union(permGen->reserved())),
526 // Construct the is_alive_closure with _span & markBitMap
527 _is_alive_closure(_span, &_markBitMap),
528 _restart_addr(NULL),
529 _overflow_list(NULL),
530 _preserved_oop_stack(NULL),
531 _preserved_mark_stack(NULL),
532 _stats(cmsGen),
533 _eden_chunk_array(NULL), // may be set in ctor body
534 _eden_chunk_capacity(0), // -- ditto --
535 _eden_chunk_index(0), // -- ditto --
536 _survivor_plab_array(NULL), // -- ditto --
537 _survivor_chunk_array(NULL), // -- ditto --
538 _survivor_chunk_capacity(0), // -- ditto --
539 _survivor_chunk_index(0), // -- ditto --
540 _ser_pmc_preclean_ovflw(0),
541 _ser_pmc_remark_ovflw(0),
542 _par_pmc_remark_ovflw(0),
543 _ser_kac_ovflw(0),
544 _par_kac_ovflw(0),
545 #ifndef PRODUCT
546 _num_par_pushes(0),
547 #endif
548 _collection_count_start(0),
549 _verifying(false),
550 _icms_start_limit(NULL),
551 _icms_stop_limit(NULL),
552 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
553 _completed_initialization(false),
554 _collector_policy(cp),
555 _should_unload_classes(false),
556 _concurrent_cycles_since_last_unload(0),
557 _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
558 {
559 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
560 ExplicitGCInvokesConcurrent = true;
561 }
562 // Now expand the span and allocate the collection support structures
563 // (MUT, marking bit map etc.) to cover both generations subject to
564 // collection.
566 // First check that _permGen is adjacent to _cmsGen and above it.
567 assert( _cmsGen->reserved().word_size() > 0
568 && _permGen->reserved().word_size() > 0,
569 "generations should not be of zero size");
570 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
571 "_cmsGen and _permGen should not overlap");
572 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
573 "_cmsGen->end() different from _permGen->start()");
575 // For use by dirty card to oop closures.
576 _cmsGen->cmsSpace()->set_collector(this);
577 _permGen->cmsSpace()->set_collector(this);
579 // Allocate MUT and marking bit map
580 {
581 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
582 if (!_markBitMap.allocate(_span)) {
583 warning("Failed to allocate CMS Bit Map");
584 return;
585 }
586 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
587 }
588 {
589 _modUnionTable.allocate(_span);
590 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
591 }
593 if (!_markStack.allocate(CMSMarkStackSize)) {
594 warning("Failed to allocate CMS Marking Stack");
595 return;
596 }
597 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
598 warning("Failed to allocate CMS Revisit Stack");
599 return;
600 }
602 // Support for multi-threaded concurrent phases
603 if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) {
604 if (FLAG_IS_DEFAULT(ParallelCMSThreads)) {
605 // just for now
606 FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4);
607 }
608 if (ParallelCMSThreads > 1) {
609 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
610 ParallelCMSThreads, true);
611 if (_conc_workers == NULL) {
612 warning("GC/CMS: _conc_workers allocation failure: "
613 "forcing -CMSConcurrentMTEnabled");
614 CMSConcurrentMTEnabled = false;
615 }
616 } else {
617 CMSConcurrentMTEnabled = false;
618 }
619 }
620 if (!CMSConcurrentMTEnabled) {
621 ParallelCMSThreads = 0;
622 } else {
623 // Turn off CMSCleanOnEnter optimization temporarily for
624 // the MT case where it's not fixed yet; see 6178663.
625 CMSCleanOnEnter = false;
626 }
627 assert((_conc_workers != NULL) == (ParallelCMSThreads > 1),
628 "Inconsistency");
630 // Parallel task queues; these are shared for the
631 // concurrent and stop-world phases of CMS, but
632 // are not shared with parallel scavenge (ParNew).
633 {
634 uint i;
635 uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads);
637 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
638 || ParallelRefProcEnabled)
639 && num_queues > 0) {
640 _task_queues = new OopTaskQueueSet(num_queues);
641 if (_task_queues == NULL) {
642 warning("task_queues allocation failure.");
643 return;
644 }
645 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
646 if (_hash_seed == NULL) {
647 warning("_hash_seed array allocation failure");
648 return;
649 }
651 // XXX use a global constant instead of 64!
652 typedef struct OopTaskQueuePadded {
653 OopTaskQueue work_queue;
654 char pad[64 - sizeof(OopTaskQueue)]; // prevent false sharing
655 } OopTaskQueuePadded;
657 for (i = 0; i < num_queues; i++) {
658 OopTaskQueuePadded *q_padded = new OopTaskQueuePadded();
659 if (q_padded == NULL) {
660 warning("work_queue allocation failure.");
661 return;
662 }
663 _task_queues->register_queue(i, &q_padded->work_queue);
664 }
665 for (i = 0; i < num_queues; i++) {
666 _task_queues->queue(i)->initialize();
667 _hash_seed[i] = 17; // copied from ParNew
668 }
669 }
670 }
672 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
673 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
675 // Clip CMSBootstrapOccupancy between 0 and 100.
676 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
677 /(double)100;
679 _full_gcs_since_conc_gc = 0;
681 // Now tell CMS generations the identity of their collector
682 ConcurrentMarkSweepGeneration::set_collector(this);
684 // Create & start a CMS thread for this CMS collector
685 _cmsThread = ConcurrentMarkSweepThread::start(this);
686 assert(cmsThread() != NULL, "CMS Thread should have been created");
687 assert(cmsThread()->collector() == this,
688 "CMS Thread should refer to this gen");
689 assert(CGC_lock != NULL, "Where's the CGC_lock?");
691 // Support for parallelizing young gen rescan
692 GenCollectedHeap* gch = GenCollectedHeap::heap();
693 _young_gen = gch->prev_gen(_cmsGen);
694 if (gch->supports_inline_contig_alloc()) {
695 _top_addr = gch->top_addr();
696 _end_addr = gch->end_addr();
697 assert(_young_gen != NULL, "no _young_gen");
698 _eden_chunk_index = 0;
699 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
700 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
701 if (_eden_chunk_array == NULL) {
702 _eden_chunk_capacity = 0;
703 warning("GC/CMS: _eden_chunk_array allocation failure");
704 }
705 }
706 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
708 // Support for parallelizing survivor space rescan
709 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
710 size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize);
711 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
712 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
713 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
714 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
715 || _cursor == NULL) {
716 warning("Failed to allocate survivor plab/chunk array");
717 if (_survivor_plab_array != NULL) {
718 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
719 _survivor_plab_array = NULL;
720 }
721 if (_survivor_chunk_array != NULL) {
722 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
723 _survivor_chunk_array = NULL;
724 }
725 if (_cursor != NULL) {
726 FREE_C_HEAP_ARRAY(size_t, _cursor);
727 _cursor = NULL;
728 }
729 } else {
730 _survivor_chunk_capacity = 2*max_plab_samples;
731 for (uint i = 0; i < ParallelGCThreads; i++) {
732 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
733 if (vec == NULL) {
734 warning("Failed to allocate survivor plab array");
735 for (int j = i; j > 0; j--) {
736 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
737 }
738 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
739 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
740 _survivor_plab_array = NULL;
741 _survivor_chunk_array = NULL;
742 _survivor_chunk_capacity = 0;
743 break;
744 } else {
745 ChunkArray* cur =
746 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
747 max_plab_samples);
748 assert(cur->end() == 0, "Should be 0");
749 assert(cur->array() == vec, "Should be vec");
750 assert(cur->capacity() == max_plab_samples, "Error");
751 }
752 }
753 }
754 }
755 assert( ( _survivor_plab_array != NULL
756 && _survivor_chunk_array != NULL)
757 || ( _survivor_chunk_capacity == 0
758 && _survivor_chunk_index == 0),
759 "Error");
761 // Choose what strong roots should be scanned depending on verification options
762 // and perm gen collection mode.
763 if (!CMSClassUnloadingEnabled) {
764 // If class unloading is disabled we want to include all classes into the root set.
765 add_root_scanning_option(SharedHeap::SO_AllClasses);
766 } else {
767 add_root_scanning_option(SharedHeap::SO_SystemClasses);
768 }
770 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
771 _gc_counters = new CollectorCounters("CMS", 1);
772 _completed_initialization = true;
773 _sweep_timer.start(); // start of time
774 }
776 const char* ConcurrentMarkSweepGeneration::name() const {
777 return "concurrent mark-sweep generation";
778 }
779 void ConcurrentMarkSweepGeneration::update_counters() {
780 if (UsePerfData) {
781 _space_counters->update_all();
782 _gen_counters->update_all();
783 }
784 }
786 // this is an optimized version of update_counters(). it takes the
787 // used value as a parameter rather than computing it.
788 //
789 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
790 if (UsePerfData) {
791 _space_counters->update_used(used);
792 _space_counters->update_capacity();
793 _gen_counters->update_all();
794 }
795 }
797 void ConcurrentMarkSweepGeneration::print() const {
798 Generation::print();
799 cmsSpace()->print();
800 }
802 #ifndef PRODUCT
803 void ConcurrentMarkSweepGeneration::print_statistics() {
804 cmsSpace()->printFLCensus(0);
805 }
806 #endif
808 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
809 GenCollectedHeap* gch = GenCollectedHeap::heap();
810 if (PrintGCDetails) {
811 if (Verbose) {
812 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
813 level(), short_name(), s, used(), capacity());
814 } else {
815 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
816 level(), short_name(), s, used() / K, capacity() / K);
817 }
818 }
819 if (Verbose) {
820 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
821 gch->used(), gch->capacity());
822 } else {
823 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
824 gch->used() / K, gch->capacity() / K);
825 }
826 }
828 size_t
829 ConcurrentMarkSweepGeneration::contiguous_available() const {
830 // dld proposes an improvement in precision here. If the committed
831 // part of the space ends in a free block we should add that to
832 // uncommitted size in the calculation below. Will make this
833 // change later, staying with the approximation below for the
834 // time being. -- ysr.
835 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
836 }
838 size_t
839 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
840 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
841 }
843 size_t ConcurrentMarkSweepGeneration::max_available() const {
844 return free() + _virtual_space.uncommitted_size();
845 }
847 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(
848 size_t max_promotion_in_bytes,
849 bool younger_handles_promotion_failure) const {
851 // This is the most conservative test. Full promotion is
852 // guaranteed if this is used. The multiplicative factor is to
853 // account for the worst case "dilatation".
854 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes;
855 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t
856 adjusted_max_promo_bytes = (double)max_uintx;
857 }
858 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes);
860 if (younger_handles_promotion_failure && !result) {
861 // Full promotion is not guaranteed because fragmentation
862 // of the cms generation can prevent the full promotion.
863 result = (max_available() >= (size_t)adjusted_max_promo_bytes);
865 if (!result) {
866 // With promotion failure handling the test for the ability
867 // to support the promotion does not have to be guaranteed.
868 // Use an average of the amount promoted.
869 result = max_available() >= (size_t)
870 gc_stats()->avg_promoted()->padded_average();
871 if (PrintGC && Verbose && result) {
872 gclog_or_tty->print_cr(
873 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
874 " max_available: " SIZE_FORMAT
875 " avg_promoted: " SIZE_FORMAT,
876 max_available(), (size_t)
877 gc_stats()->avg_promoted()->padded_average());
878 }
879 } else {
880 if (PrintGC && Verbose) {
881 gclog_or_tty->print_cr(
882 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
883 " max_available: " SIZE_FORMAT
884 " adj_max_promo_bytes: " SIZE_FORMAT,
885 max_available(), (size_t)adjusted_max_promo_bytes);
886 }
887 }
888 } else {
889 if (PrintGC && Verbose) {
890 gclog_or_tty->print_cr(
891 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
892 " contiguous_available: " SIZE_FORMAT
893 " adj_max_promo_bytes: " SIZE_FORMAT,
894 max_contiguous_available(), (size_t)adjusted_max_promo_bytes);
895 }
896 }
897 return result;
898 }
900 CompactibleSpace*
901 ConcurrentMarkSweepGeneration::first_compaction_space() const {
902 return _cmsSpace;
903 }
905 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
906 // Clear the promotion information. These pointers can be adjusted
907 // along with all the other pointers into the heap but
908 // compaction is expected to be a rare event with
909 // a heap using cms so don't do it without seeing the need.
910 if (ParallelGCThreads > 0) {
911 for (uint i = 0; i < ParallelGCThreads; i++) {
912 _par_gc_thread_states[i]->promo.reset();
913 }
914 }
915 }
917 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
918 blk->do_space(_cmsSpace);
919 }
921 void ConcurrentMarkSweepGeneration::compute_new_size() {
922 assert_locked_or_safepoint(Heap_lock);
924 // If incremental collection failed, we just want to expand
925 // to the limit.
926 if (incremental_collection_failed()) {
927 clear_incremental_collection_failed();
928 grow_to_reserved();
929 return;
930 }
932 size_t expand_bytes = 0;
933 double free_percentage = ((double) free()) / capacity();
934 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
935 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
937 // compute expansion delta needed for reaching desired free percentage
938 if (free_percentage < desired_free_percentage) {
939 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
940 assert(desired_capacity >= capacity(), "invalid expansion size");
941 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
942 }
943 if (expand_bytes > 0) {
944 if (PrintGCDetails && Verbose) {
945 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
946 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
947 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
948 gclog_or_tty->print_cr(" Desired free fraction %f",
949 desired_free_percentage);
950 gclog_or_tty->print_cr(" Maximum free fraction %f",
951 maximum_free_percentage);
952 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
953 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
954 desired_capacity/1000);
955 int prev_level = level() - 1;
956 if (prev_level >= 0) {
957 size_t prev_size = 0;
958 GenCollectedHeap* gch = GenCollectedHeap::heap();
959 Generation* prev_gen = gch->_gens[prev_level];
960 prev_size = prev_gen->capacity();
961 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
962 prev_size/1000);
963 }
964 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
965 unsafe_max_alloc_nogc()/1000);
966 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
967 contiguous_available()/1000);
968 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
969 expand_bytes);
970 }
971 // safe if expansion fails
972 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
973 if (PrintGCDetails && Verbose) {
974 gclog_or_tty->print_cr(" Expanded free fraction %f",
975 ((double) free()) / capacity());
976 }
977 }
978 }
980 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
981 return cmsSpace()->freelistLock();
982 }
984 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
985 bool tlab) {
986 CMSSynchronousYieldRequest yr;
987 MutexLockerEx x(freelistLock(),
988 Mutex::_no_safepoint_check_flag);
989 return have_lock_and_allocate(size, tlab);
990 }
992 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
993 bool tlab) {
994 assert_lock_strong(freelistLock());
995 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
996 HeapWord* res = cmsSpace()->allocate(adjustedSize);
997 // Allocate the object live (grey) if the background collector has
998 // started marking. This is necessary because the marker may
999 // have passed this address and consequently this object will
1000 // not otherwise be greyed and would be incorrectly swept up.
1001 // Note that if this object contains references, the writing
1002 // of those references will dirty the card containing this object
1003 // allowing the object to be blackened (and its references scanned)
1004 // either during a preclean phase or at the final checkpoint.
1005 if (res != NULL) {
1006 collector()->direct_allocated(res, adjustedSize);
1007 _direct_allocated_words += adjustedSize;
1008 // allocation counters
1009 NOT_PRODUCT(
1010 _numObjectsAllocated++;
1011 _numWordsAllocated += (int)adjustedSize;
1012 )
1013 }
1014 return res;
1015 }
1017 // In the case of direct allocation by mutators in a generation that
1018 // is being concurrently collected, the object must be allocated
1019 // live (grey) if the background collector has started marking.
1020 // This is necessary because the marker may
1021 // have passed this address and consequently this object will
1022 // not otherwise be greyed and would be incorrectly swept up.
1023 // Note that if this object contains references, the writing
1024 // of those references will dirty the card containing this object
1025 // allowing the object to be blackened (and its references scanned)
1026 // either during a preclean phase or at the final checkpoint.
1027 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1028 assert(_markBitMap.covers(start, size), "Out of bounds");
1029 if (_collectorState >= Marking) {
1030 MutexLockerEx y(_markBitMap.lock(),
1031 Mutex::_no_safepoint_check_flag);
1032 // [see comments preceding SweepClosure::do_blk() below for details]
1033 // 1. need to mark the object as live so it isn't collected
1034 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1035 // 3. need to mark the end of the object so sweeper can skip over it
1036 // if it's uninitialized when the sweeper reaches it.
1037 _markBitMap.mark(start); // object is live
1038 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1039 _markBitMap.mark(start + size - 1);
1040 // mark end of object
1041 }
1042 // check that oop looks uninitialized
1043 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1044 }
1046 void CMSCollector::promoted(bool par, HeapWord* start,
1047 bool is_obj_array, size_t obj_size) {
1048 assert(_markBitMap.covers(start), "Out of bounds");
1049 // See comment in direct_allocated() about when objects should
1050 // be allocated live.
1051 if (_collectorState >= Marking) {
1052 // we already hold the marking bit map lock, taken in
1053 // the prologue
1054 if (par) {
1055 _markBitMap.par_mark(start);
1056 } else {
1057 _markBitMap.mark(start);
1058 }
1059 // We don't need to mark the object as uninitialized (as
1060 // in direct_allocated above) because this is being done with the
1061 // world stopped and the object will be initialized by the
1062 // time the sweeper gets to look at it.
1063 assert(SafepointSynchronize::is_at_safepoint(),
1064 "expect promotion only at safepoints");
1066 if (_collectorState < Sweeping) {
1067 // Mark the appropriate cards in the modUnionTable, so that
1068 // this object gets scanned before the sweep. If this is
1069 // not done, CMS generation references in the object might
1070 // not get marked.
1071 // For the case of arrays, which are otherwise precisely
1072 // marked, we need to dirty the entire array, not just its head.
1073 if (is_obj_array) {
1074 // The [par_]mark_range() method expects mr.end() below to
1075 // be aligned to the granularity of a bit's representation
1076 // in the heap. In the case of the MUT below, that's a
1077 // card size.
1078 MemRegion mr(start,
1079 (HeapWord*)round_to((intptr_t)(start + obj_size),
1080 CardTableModRefBS::card_size /* bytes */));
1081 if (par) {
1082 _modUnionTable.par_mark_range(mr);
1083 } else {
1084 _modUnionTable.mark_range(mr);
1085 }
1086 } else { // not an obj array; we can just mark the head
1087 if (par) {
1088 _modUnionTable.par_mark(start);
1089 } else {
1090 _modUnionTable.mark(start);
1091 }
1092 }
1093 }
1094 }
1095 }
1097 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1098 {
1099 size_t delta = pointer_delta(addr, space->bottom());
1100 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1101 }
1103 void CMSCollector::icms_update_allocation_limits()
1104 {
1105 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1106 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1108 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1109 if (CMSTraceIncrementalPacing) {
1110 stats().print();
1111 }
1113 assert(duty_cycle <= 100, "invalid duty cycle");
1114 if (duty_cycle != 0) {
1115 // The duty_cycle is a percentage between 0 and 100; convert to words and
1116 // then compute the offset from the endpoints of the space.
1117 size_t free_words = eden->free() / HeapWordSize;
1118 double free_words_dbl = (double)free_words;
1119 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1120 size_t offset_words = (free_words - duty_cycle_words) / 2;
1122 _icms_start_limit = eden->top() + offset_words;
1123 _icms_stop_limit = eden->end() - offset_words;
1125 // The limits may be adjusted (shifted to the right) by
1126 // CMSIncrementalOffset, to allow the application more mutator time after a
1127 // young gen gc (when all mutators were stopped) and before CMS starts and
1128 // takes away one or more cpus.
1129 if (CMSIncrementalOffset != 0) {
1130 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1131 size_t adjustment = (size_t)adjustment_dbl;
1132 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1133 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1134 _icms_start_limit += adjustment;
1135 _icms_stop_limit = tmp_stop;
1136 }
1137 }
1138 }
1139 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1140 _icms_start_limit = _icms_stop_limit = eden->end();
1141 }
1143 // Install the new start limit.
1144 eden->set_soft_end(_icms_start_limit);
1146 if (CMSTraceIncrementalMode) {
1147 gclog_or_tty->print(" icms alloc limits: "
1148 PTR_FORMAT "," PTR_FORMAT
1149 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1150 _icms_start_limit, _icms_stop_limit,
1151 percent_of_space(eden, _icms_start_limit),
1152 percent_of_space(eden, _icms_stop_limit));
1153 if (Verbose) {
1154 gclog_or_tty->print("eden: ");
1155 eden->print_on(gclog_or_tty);
1156 }
1157 }
1158 }
1160 // Any changes here should try to maintain the invariant
1161 // that if this method is called with _icms_start_limit
1162 // and _icms_stop_limit both NULL, then it should return NULL
1163 // and not notify the icms thread.
1164 HeapWord*
1165 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1166 size_t word_size)
1167 {
1168 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1169 // nop.
1170 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1171 if (top <= _icms_start_limit) {
1172 if (CMSTraceIncrementalMode) {
1173 space->print_on(gclog_or_tty);
1174 gclog_or_tty->stamp();
1175 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1176 ", new limit=" PTR_FORMAT
1177 " (" SIZE_FORMAT "%%)",
1178 top, _icms_stop_limit,
1179 percent_of_space(space, _icms_stop_limit));
1180 }
1181 ConcurrentMarkSweepThread::start_icms();
1182 assert(top < _icms_stop_limit, "Tautology");
1183 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1184 return _icms_stop_limit;
1185 }
1187 // The allocation will cross both the _start and _stop limits, so do the
1188 // stop notification also and return end().
1189 if (CMSTraceIncrementalMode) {
1190 space->print_on(gclog_or_tty);
1191 gclog_or_tty->stamp();
1192 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1193 ", new limit=" PTR_FORMAT
1194 " (" SIZE_FORMAT "%%)",
1195 top, space->end(),
1196 percent_of_space(space, space->end()));
1197 }
1198 ConcurrentMarkSweepThread::stop_icms();
1199 return space->end();
1200 }
1202 if (top <= _icms_stop_limit) {
1203 if (CMSTraceIncrementalMode) {
1204 space->print_on(gclog_or_tty);
1205 gclog_or_tty->stamp();
1206 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1207 ", new limit=" PTR_FORMAT
1208 " (" SIZE_FORMAT "%%)",
1209 top, space->end(),
1210 percent_of_space(space, space->end()));
1211 }
1212 ConcurrentMarkSweepThread::stop_icms();
1213 return space->end();
1214 }
1216 if (CMSTraceIncrementalMode) {
1217 space->print_on(gclog_or_tty);
1218 gclog_or_tty->stamp();
1219 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1220 ", new limit=" PTR_FORMAT,
1221 top, NULL);
1222 }
1223 }
1225 return NULL;
1226 }
1228 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1229 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1230 // allocate, copy and if necessary update promoinfo --
1231 // delegate to underlying space.
1232 assert_lock_strong(freelistLock());
1234 #ifndef PRODUCT
1235 if (Universe::heap()->promotion_should_fail()) {
1236 return NULL;
1237 }
1238 #endif // #ifndef PRODUCT
1240 oop res = _cmsSpace->promote(obj, obj_size);
1241 if (res == NULL) {
1242 // expand and retry
1243 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1244 expand(s*HeapWordSize, MinHeapDeltaBytes,
1245 CMSExpansionCause::_satisfy_promotion);
1246 // Since there's currently no next generation, we don't try to promote
1247 // into a more senior generation.
1248 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1249 "is made to pass on a possibly failing "
1250 "promotion to next generation");
1251 res = _cmsSpace->promote(obj, obj_size);
1252 }
1253 if (res != NULL) {
1254 // See comment in allocate() about when objects should
1255 // be allocated live.
1256 assert(obj->is_oop(), "Will dereference klass pointer below");
1257 collector()->promoted(false, // Not parallel
1258 (HeapWord*)res, obj->is_objArray(), obj_size);
1259 // promotion counters
1260 NOT_PRODUCT(
1261 _numObjectsPromoted++;
1262 _numWordsPromoted +=
1263 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1264 )
1265 }
1266 return res;
1267 }
1270 HeapWord*
1271 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1272 HeapWord* top,
1273 size_t word_sz)
1274 {
1275 return collector()->allocation_limit_reached(space, top, word_sz);
1276 }
1278 // Things to support parallel young-gen collection.
1279 oop
1280 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1281 oop old, markOop m,
1282 size_t word_sz) {
1283 #ifndef PRODUCT
1284 if (Universe::heap()->promotion_should_fail()) {
1285 return NULL;
1286 }
1287 #endif // #ifndef PRODUCT
1289 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1290 PromotionInfo* promoInfo = &ps->promo;
1291 // if we are tracking promotions, then first ensure space for
1292 // promotion (including spooling space for saving header if necessary).
1293 // then allocate and copy, then track promoted info if needed.
1294 // When tracking (see PromotionInfo::track()), the mark word may
1295 // be displaced and in this case restoration of the mark word
1296 // occurs in the (oop_since_save_marks_)iterate phase.
1297 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1298 // Out of space for allocating spooling buffers;
1299 // try expanding and allocating spooling buffers.
1300 if (!expand_and_ensure_spooling_space(promoInfo)) {
1301 return NULL;
1302 }
1303 }
1304 assert(promoInfo->has_spooling_space(), "Control point invariant");
1305 HeapWord* obj_ptr = ps->lab.alloc(word_sz);
1306 if (obj_ptr == NULL) {
1307 obj_ptr = expand_and_par_lab_allocate(ps, word_sz);
1308 if (obj_ptr == NULL) {
1309 return NULL;
1310 }
1311 }
1312 oop obj = oop(obj_ptr);
1313 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1314 // Otherwise, copy the object. Here we must be careful to insert the
1315 // klass pointer last, since this marks the block as an allocated object.
1316 // Except with compressed oops it's the mark word.
1317 HeapWord* old_ptr = (HeapWord*)old;
1318 if (word_sz > (size_t)oopDesc::header_size()) {
1319 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1320 obj_ptr + oopDesc::header_size(),
1321 word_sz - oopDesc::header_size());
1322 }
1324 if (UseCompressedOops) {
1325 // Copy gap missed by (aligned) header size calculation above
1326 obj->set_klass_gap(old->klass_gap());
1327 }
1329 // Restore the mark word copied above.
1330 obj->set_mark(m);
1332 // Now we can track the promoted object, if necessary. We take care
1333 // To delay the transition from uninitialized to full object
1334 // (i.e., insertion of klass pointer) until after, so that it
1335 // atomically becomes a promoted object.
1336 if (promoInfo->tracking()) {
1337 promoInfo->track((PromotedObject*)obj, old->klass());
1338 }
1340 // Finally, install the klass pointer (this should be volatile).
1341 obj->set_klass(old->klass());
1343 assert(old->is_oop(), "Will dereference klass ptr below");
1344 collector()->promoted(true, // parallel
1345 obj_ptr, old->is_objArray(), word_sz);
1347 NOT_PRODUCT(
1348 Atomic::inc(&_numObjectsPromoted);
1349 Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()),
1350 &_numWordsPromoted);
1351 )
1353 return obj;
1354 }
1356 void
1357 ConcurrentMarkSweepGeneration::
1358 par_promote_alloc_undo(int thread_num,
1359 HeapWord* obj, size_t word_sz) {
1360 // CMS does not support promotion undo.
1361 ShouldNotReachHere();
1362 }
1364 void
1365 ConcurrentMarkSweepGeneration::
1366 par_promote_alloc_done(int thread_num) {
1367 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1368 ps->lab.retire();
1369 #if CFLS_LAB_REFILL_STATS
1370 if (thread_num == 0) {
1371 _cmsSpace->print_par_alloc_stats();
1372 }
1373 #endif
1374 }
1376 void
1377 ConcurrentMarkSweepGeneration::
1378 par_oop_since_save_marks_iterate_done(int thread_num) {
1379 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1380 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1381 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1382 }
1384 // XXXPERM
1385 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1386 size_t size,
1387 bool tlab)
1388 {
1389 // We allow a STW collection only if a full
1390 // collection was requested.
1391 return full || should_allocate(size, tlab); // FIX ME !!!
1392 // This and promotion failure handling are connected at the
1393 // hip and should be fixed by untying them.
1394 }
1396 bool CMSCollector::shouldConcurrentCollect() {
1397 if (_full_gc_requested) {
1398 assert(ExplicitGCInvokesConcurrent, "Unexpected state");
1399 if (Verbose && PrintGCDetails) {
1400 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1401 " gc request");
1402 }
1403 return true;
1404 }
1406 // For debugging purposes, change the type of collection.
1407 // If the rotation is not on the concurrent collection
1408 // type, don't start a concurrent collection.
1409 NOT_PRODUCT(
1410 if (RotateCMSCollectionTypes &&
1411 (_cmsGen->debug_collection_type() !=
1412 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1413 assert(_cmsGen->debug_collection_type() !=
1414 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1415 "Bad cms collection type");
1416 return false;
1417 }
1418 )
1420 FreelistLocker x(this);
1421 // ------------------------------------------------------------------
1422 // Print out lots of information which affects the initiation of
1423 // a collection.
1424 if (PrintCMSInitiationStatistics && stats().valid()) {
1425 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1426 gclog_or_tty->stamp();
1427 gclog_or_tty->print_cr("");
1428 stats().print_on(gclog_or_tty);
1429 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1430 stats().time_until_cms_gen_full());
1431 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1432 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1433 _cmsGen->contiguous_available());
1434 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1435 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1436 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1437 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1438 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1439 }
1440 // ------------------------------------------------------------------
1442 // If the estimated time to complete a cms collection (cms_duration())
1443 // is less than the estimated time remaining until the cms generation
1444 // is full, start a collection.
1445 if (!UseCMSInitiatingOccupancyOnly) {
1446 if (stats().valid()) {
1447 if (stats().time_until_cms_start() == 0.0) {
1448 return true;
1449 }
1450 } else {
1451 // We want to conservatively collect somewhat early in order
1452 // to try and "bootstrap" our CMS/promotion statistics;
1453 // this branch will not fire after the first successful CMS
1454 // collection because the stats should then be valid.
1455 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1456 if (Verbose && PrintGCDetails) {
1457 gclog_or_tty->print_cr(
1458 " CMSCollector: collect for bootstrapping statistics:"
1459 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1460 _bootstrap_occupancy);
1461 }
1462 return true;
1463 }
1464 }
1465 }
1467 // Otherwise, we start a collection cycle if either the perm gen or
1468 // old gen want a collection cycle started. Each may use
1469 // an appropriate criterion for making this decision.
1470 // XXX We need to make sure that the gen expansion
1471 // criterion dovetails well with this. XXX NEED TO FIX THIS
1472 if (_cmsGen->should_concurrent_collect()) {
1473 if (Verbose && PrintGCDetails) {
1474 gclog_or_tty->print_cr("CMS old gen initiated");
1475 }
1476 return true;
1477 }
1479 // We start a collection if we believe an incremental collection may fail;
1480 // this is not likely to be productive in practice because it's probably too
1481 // late anyway.
1482 GenCollectedHeap* gch = GenCollectedHeap::heap();
1483 assert(gch->collector_policy()->is_two_generation_policy(),
1484 "You may want to check the correctness of the following");
1485 if (gch->incremental_collection_will_fail()) {
1486 if (PrintGCDetails && Verbose) {
1487 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1488 }
1489 return true;
1490 }
1492 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1493 bool res = update_should_unload_classes();
1494 if (res) {
1495 if (Verbose && PrintGCDetails) {
1496 gclog_or_tty->print_cr("CMS perm gen initiated");
1497 }
1498 return true;
1499 }
1500 }
1501 return false;
1502 }
1504 // Clear _expansion_cause fields of constituent generations
1505 void CMSCollector::clear_expansion_cause() {
1506 _cmsGen->clear_expansion_cause();
1507 _permGen->clear_expansion_cause();
1508 }
1510 // We should be conservative in starting a collection cycle. To
1511 // start too eagerly runs the risk of collecting too often in the
1512 // extreme. To collect too rarely falls back on full collections,
1513 // which works, even if not optimum in terms of concurrent work.
1514 // As a work around for too eagerly collecting, use the flag
1515 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1516 // giving the user an easily understandable way of controlling the
1517 // collections.
1518 // We want to start a new collection cycle if any of the following
1519 // conditions hold:
1520 // . our current occupancy exceeds the configured initiating occupancy
1521 // for this generation, or
1522 // . we recently needed to expand this space and have not, since that
1523 // expansion, done a collection of this generation, or
1524 // . the underlying space believes that it may be a good idea to initiate
1525 // a concurrent collection (this may be based on criteria such as the
1526 // following: the space uses linear allocation and linear allocation is
1527 // going to fail, or there is believed to be excessive fragmentation in
1528 // the generation, etc... or ...
1529 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1530 // the case of the old generation, not the perm generation; see CR 6543076):
1531 // we may be approaching a point at which allocation requests may fail because
1532 // we will be out of sufficient free space given allocation rate estimates.]
1533 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1535 assert_lock_strong(freelistLock());
1536 if (occupancy() > initiating_occupancy()) {
1537 if (PrintGCDetails && Verbose) {
1538 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1539 short_name(), occupancy(), initiating_occupancy());
1540 }
1541 return true;
1542 }
1543 if (UseCMSInitiatingOccupancyOnly) {
1544 return false;
1545 }
1546 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1547 if (PrintGCDetails && Verbose) {
1548 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1549 short_name());
1550 }
1551 return true;
1552 }
1553 if (_cmsSpace->should_concurrent_collect()) {
1554 if (PrintGCDetails && Verbose) {
1555 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1556 short_name());
1557 }
1558 return true;
1559 }
1560 return false;
1561 }
1563 void ConcurrentMarkSweepGeneration::collect(bool full,
1564 bool clear_all_soft_refs,
1565 size_t size,
1566 bool tlab)
1567 {
1568 collector()->collect(full, clear_all_soft_refs, size, tlab);
1569 }
1571 void CMSCollector::collect(bool full,
1572 bool clear_all_soft_refs,
1573 size_t size,
1574 bool tlab)
1575 {
1576 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1577 // For debugging purposes skip the collection if the state
1578 // is not currently idle
1579 if (TraceCMSState) {
1580 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1581 Thread::current(), full, _collectorState);
1582 }
1583 return;
1584 }
1586 // The following "if" branch is present for defensive reasons.
1587 // In the current uses of this interface, it can be replaced with:
1588 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1589 // But I am not placing that assert here to allow future
1590 // generality in invoking this interface.
1591 if (GC_locker::is_active()) {
1592 // A consistency test for GC_locker
1593 assert(GC_locker::needs_gc(), "Should have been set already");
1594 // Skip this foreground collection, instead
1595 // expanding the heap if necessary.
1596 // Need the free list locks for the call to free() in compute_new_size()
1597 compute_new_size();
1598 return;
1599 }
1600 acquire_control_and_collect(full, clear_all_soft_refs);
1601 _full_gcs_since_conc_gc++;
1603 }
1605 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1606 GenCollectedHeap* gch = GenCollectedHeap::heap();
1607 unsigned int gc_count = gch->total_full_collections();
1608 if (gc_count == full_gc_count) {
1609 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1610 _full_gc_requested = true;
1611 CGC_lock->notify(); // nudge CMS thread
1612 }
1613 }
1616 // The foreground and background collectors need to coordinate in order
1617 // to make sure that they do not mutually interfere with CMS collections.
1618 // When a background collection is active,
1619 // the foreground collector may need to take over (preempt) and
1620 // synchronously complete an ongoing collection. Depending on the
1621 // frequency of the background collections and the heap usage
1622 // of the application, this preemption can be seldom or frequent.
1623 // There are only certain
1624 // points in the background collection that the "collection-baton"
1625 // can be passed to the foreground collector.
1626 //
1627 // The foreground collector will wait for the baton before
1628 // starting any part of the collection. The foreground collector
1629 // will only wait at one location.
1630 //
1631 // The background collector will yield the baton before starting a new
1632 // phase of the collection (e.g., before initial marking, marking from roots,
1633 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1634 // of the loop which switches the phases. The background collector does some
1635 // of the phases (initial mark, final re-mark) with the world stopped.
1636 // Because of locking involved in stopping the world,
1637 // the foreground collector should not block waiting for the background
1638 // collector when it is doing a stop-the-world phase. The background
1639 // collector will yield the baton at an additional point just before
1640 // it enters a stop-the-world phase. Once the world is stopped, the
1641 // background collector checks the phase of the collection. If the
1642 // phase has not changed, it proceeds with the collection. If the
1643 // phase has changed, it skips that phase of the collection. See
1644 // the comments on the use of the Heap_lock in collect_in_background().
1645 //
1646 // Variable used in baton passing.
1647 // _foregroundGCIsActive - Set to true by the foreground collector when
1648 // it wants the baton. The foreground clears it when it has finished
1649 // the collection.
1650 // _foregroundGCShouldWait - Set to true by the background collector
1651 // when it is running. The foreground collector waits while
1652 // _foregroundGCShouldWait is true.
1653 // CGC_lock - monitor used to protect access to the above variables
1654 // and to notify the foreground and background collectors.
1655 // _collectorState - current state of the CMS collection.
1656 //
1657 // The foreground collector
1658 // acquires the CGC_lock
1659 // sets _foregroundGCIsActive
1660 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1661 // various locks acquired in preparation for the collection
1662 // are released so as not to block the background collector
1663 // that is in the midst of a collection
1664 // proceeds with the collection
1665 // clears _foregroundGCIsActive
1666 // returns
1667 //
1668 // The background collector in a loop iterating on the phases of the
1669 // collection
1670 // acquires the CGC_lock
1671 // sets _foregroundGCShouldWait
1672 // if _foregroundGCIsActive is set
1673 // clears _foregroundGCShouldWait, notifies _CGC_lock
1674 // waits on _CGC_lock for _foregroundGCIsActive to become false
1675 // and exits the loop.
1676 // otherwise
1677 // proceed with that phase of the collection
1678 // if the phase is a stop-the-world phase,
1679 // yield the baton once more just before enqueueing
1680 // the stop-world CMS operation (executed by the VM thread).
1681 // returns after all phases of the collection are done
1682 //
1684 void CMSCollector::acquire_control_and_collect(bool full,
1685 bool clear_all_soft_refs) {
1686 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1687 assert(!Thread::current()->is_ConcurrentGC_thread(),
1688 "shouldn't try to acquire control from self!");
1690 // Start the protocol for acquiring control of the
1691 // collection from the background collector (aka CMS thread).
1692 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1693 "VM thread should have CMS token");
1694 // Remember the possibly interrupted state of an ongoing
1695 // concurrent collection
1696 CollectorState first_state = _collectorState;
1698 // Signal to a possibly ongoing concurrent collection that
1699 // we want to do a foreground collection.
1700 _foregroundGCIsActive = true;
1702 // Disable incremental mode during a foreground collection.
1703 ICMSDisabler icms_disabler;
1705 // release locks and wait for a notify from the background collector
1706 // releasing the locks in only necessary for phases which
1707 // do yields to improve the granularity of the collection.
1708 assert_lock_strong(bitMapLock());
1709 // We need to lock the Free list lock for the space that we are
1710 // currently collecting.
1711 assert(haveFreelistLocks(), "Must be holding free list locks");
1712 bitMapLock()->unlock();
1713 releaseFreelistLocks();
1714 {
1715 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1716 if (_foregroundGCShouldWait) {
1717 // We are going to be waiting for action for the CMS thread;
1718 // it had better not be gone (for instance at shutdown)!
1719 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1720 "CMS thread must be running");
1721 // Wait here until the background collector gives us the go-ahead
1722 ConcurrentMarkSweepThread::clear_CMS_flag(
1723 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1724 // Get a possibly blocked CMS thread going:
1725 // Note that we set _foregroundGCIsActive true above,
1726 // without protection of the CGC_lock.
1727 CGC_lock->notify();
1728 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1729 "Possible deadlock");
1730 while (_foregroundGCShouldWait) {
1731 // wait for notification
1732 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1733 // Possibility of delay/starvation here, since CMS token does
1734 // not know to give priority to VM thread? Actually, i think
1735 // there wouldn't be any delay/starvation, but the proof of
1736 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1737 }
1738 ConcurrentMarkSweepThread::set_CMS_flag(
1739 ConcurrentMarkSweepThread::CMS_vm_has_token);
1740 }
1741 }
1742 // The CMS_token is already held. Get back the other locks.
1743 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1744 "VM thread should have CMS token");
1745 getFreelistLocks();
1746 bitMapLock()->lock_without_safepoint_check();
1747 if (TraceCMSState) {
1748 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1749 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1750 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1751 }
1753 // Check if we need to do a compaction, or if not, whether
1754 // we need to start the mark-sweep from scratch.
1755 bool should_compact = false;
1756 bool should_start_over = false;
1757 decide_foreground_collection_type(clear_all_soft_refs,
1758 &should_compact, &should_start_over);
1760 NOT_PRODUCT(
1761 if (RotateCMSCollectionTypes) {
1762 if (_cmsGen->debug_collection_type() ==
1763 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1764 should_compact = true;
1765 } else if (_cmsGen->debug_collection_type() ==
1766 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1767 should_compact = false;
1768 }
1769 }
1770 )
1772 if (PrintGCDetails && first_state > Idling) {
1773 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1774 if (GCCause::is_user_requested_gc(cause) ||
1775 GCCause::is_serviceability_requested_gc(cause)) {
1776 gclog_or_tty->print(" (concurrent mode interrupted)");
1777 } else {
1778 gclog_or_tty->print(" (concurrent mode failure)");
1779 }
1780 }
1782 if (should_compact) {
1783 // If the collection is being acquired from the background
1784 // collector, there may be references on the discovered
1785 // references lists that have NULL referents (being those
1786 // that were concurrently cleared by a mutator) or
1787 // that are no longer active (having been enqueued concurrently
1788 // by the mutator).
1789 // Scrub the list of those references because Mark-Sweep-Compact
1790 // code assumes referents are not NULL and that all discovered
1791 // Reference objects are active.
1792 ref_processor()->clean_up_discovered_references();
1794 do_compaction_work(clear_all_soft_refs);
1796 // Has the GC time limit been exceeded?
1797 check_gc_time_limit();
1799 } else {
1800 do_mark_sweep_work(clear_all_soft_refs, first_state,
1801 should_start_over);
1802 }
1803 // Reset the expansion cause, now that we just completed
1804 // a collection cycle.
1805 clear_expansion_cause();
1806 _foregroundGCIsActive = false;
1807 return;
1808 }
1810 void CMSCollector::check_gc_time_limit() {
1812 // Ignore explicit GC's. Exiting here does not set the flag and
1813 // does not reset the count. Updating of the averages for system
1814 // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC.
1815 GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause();
1816 if (GCCause::is_user_requested_gc(gc_cause) ||
1817 GCCause::is_serviceability_requested_gc(gc_cause)) {
1818 return;
1819 }
1821 // Calculate the fraction of the CMS generation was freed during
1822 // the last collection.
1823 // Only consider the STW compacting cost for now.
1824 //
1825 // Note that the gc time limit test only works for the collections
1826 // of the young gen + tenured gen and not for collections of the
1827 // permanent gen. That is because the calculation of the space
1828 // freed by the collection is the free space in the young gen +
1829 // tenured gen.
1831 double fraction_free =
1832 ((double)_cmsGen->free())/((double)_cmsGen->max_capacity());
1833 if ((100.0 * size_policy()->compacting_gc_cost()) >
1834 ((double) GCTimeLimit) &&
1835 ((fraction_free * 100) < GCHeapFreeLimit)) {
1836 size_policy()->inc_gc_time_limit_count();
1837 if (UseGCOverheadLimit &&
1838 (size_policy()->gc_time_limit_count() >
1839 AdaptiveSizePolicyGCTimeLimitThreshold)) {
1840 size_policy()->set_gc_time_limit_exceeded(true);
1841 // Avoid consecutive OOM due to the gc time limit by resetting
1842 // the counter.
1843 size_policy()->reset_gc_time_limit_count();
1844 if (PrintGCDetails) {
1845 gclog_or_tty->print_cr(" GC is exceeding overhead limit "
1846 "of %d%%", GCTimeLimit);
1847 }
1848 } else {
1849 if (PrintGCDetails) {
1850 gclog_or_tty->print_cr(" GC would exceed overhead limit "
1851 "of %d%%", GCTimeLimit);
1852 }
1853 }
1854 } else {
1855 size_policy()->reset_gc_time_limit_count();
1856 }
1857 }
1859 // Resize the perm generation and the tenured generation
1860 // after obtaining the free list locks for the
1861 // two generations.
1862 void CMSCollector::compute_new_size() {
1863 assert_locked_or_safepoint(Heap_lock);
1864 FreelistLocker z(this);
1865 _permGen->compute_new_size();
1866 _cmsGen->compute_new_size();
1867 }
1869 // A work method used by foreground collection to determine
1870 // what type of collection (compacting or not, continuing or fresh)
1871 // it should do.
1872 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1873 // and CMSCompactWhenClearAllSoftRefs the default in the future
1874 // and do away with the flags after a suitable period.
1875 void CMSCollector::decide_foreground_collection_type(
1876 bool clear_all_soft_refs, bool* should_compact,
1877 bool* should_start_over) {
1878 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1879 // flag is set, and we have either requested a System.gc() or
1880 // the number of full gc's since the last concurrent cycle
1881 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1882 // or if an incremental collection has failed
1883 GenCollectedHeap* gch = GenCollectedHeap::heap();
1884 assert(gch->collector_policy()->is_two_generation_policy(),
1885 "You may want to check the correctness of the following");
1886 // Inform cms gen if this was due to partial collection failing.
1887 // The CMS gen may use this fact to determine its expansion policy.
1888 if (gch->incremental_collection_will_fail()) {
1889 assert(!_cmsGen->incremental_collection_failed(),
1890 "Should have been noticed, reacted to and cleared");
1891 _cmsGen->set_incremental_collection_failed();
1892 }
1893 *should_compact =
1894 UseCMSCompactAtFullCollection &&
1895 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1896 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1897 gch->incremental_collection_will_fail());
1898 *should_start_over = false;
1899 if (clear_all_soft_refs && !*should_compact) {
1900 // We are about to do a last ditch collection attempt
1901 // so it would normally make sense to do a compaction
1902 // to reclaim as much space as possible.
1903 if (CMSCompactWhenClearAllSoftRefs) {
1904 // Default: The rationale is that in this case either
1905 // we are past the final marking phase, in which case
1906 // we'd have to start over, or so little has been done
1907 // that there's little point in saving that work. Compaction
1908 // appears to be the sensible choice in either case.
1909 *should_compact = true;
1910 } else {
1911 // We have been asked to clear all soft refs, but not to
1912 // compact. Make sure that we aren't past the final checkpoint
1913 // phase, for that is where we process soft refs. If we are already
1914 // past that phase, we'll need to redo the refs discovery phase and
1915 // if necessary clear soft refs that weren't previously
1916 // cleared. We do so by remembering the phase in which
1917 // we came in, and if we are past the refs processing
1918 // phase, we'll choose to just redo the mark-sweep
1919 // collection from scratch.
1920 if (_collectorState > FinalMarking) {
1921 // We are past the refs processing phase;
1922 // start over and do a fresh synchronous CMS cycle
1923 _collectorState = Resetting; // skip to reset to start new cycle
1924 reset(false /* == !asynch */);
1925 *should_start_over = true;
1926 } // else we can continue a possibly ongoing current cycle
1927 }
1928 }
1929 }
1931 // A work method used by the foreground collector to do
1932 // a mark-sweep-compact.
1933 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1934 GenCollectedHeap* gch = GenCollectedHeap::heap();
1935 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1936 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1937 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1938 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1939 }
1941 // Sample collection interval time and reset for collection pause.
1942 if (UseAdaptiveSizePolicy) {
1943 size_policy()->msc_collection_begin();
1944 }
1946 // Temporarily widen the span of the weak reference processing to
1947 // the entire heap.
1948 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1949 ReferenceProcessorSpanMutator x(ref_processor(), new_span);
1951 // Temporarily, clear the "is_alive_non_header" field of the
1952 // reference processor.
1953 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
1955 // Temporarily make reference _processing_ single threaded (non-MT).
1956 ReferenceProcessorMTProcMutator z(ref_processor(), false);
1958 // Temporarily make refs discovery atomic
1959 ReferenceProcessorAtomicMutator w(ref_processor(), true);
1961 ref_processor()->set_enqueuing_is_done(false);
1962 ref_processor()->enable_discovery();
1963 // If an asynchronous collection finishes, the _modUnionTable is
1964 // all clear. If we are assuming the collection from an asynchronous
1965 // collection, clear the _modUnionTable.
1966 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1967 "_modUnionTable should be clear if the baton was not passed");
1968 _modUnionTable.clear_all();
1970 // We must adjust the allocation statistics being maintained
1971 // in the free list space. We do so by reading and clearing
1972 // the sweep timer and updating the block flux rate estimates below.
1973 assert(_sweep_timer.is_active(), "We should never see the timer inactive");
1974 _sweep_timer.stop();
1975 // Note that we do not use this sample to update the _sweep_estimate.
1976 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
1977 _sweep_estimate.padded_average());
1979 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1980 ref_processor(), clear_all_soft_refs);
1981 #ifdef ASSERT
1982 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1983 size_t free_size = cms_space->free();
1984 assert(free_size ==
1985 pointer_delta(cms_space->end(), cms_space->compaction_top())
1986 * HeapWordSize,
1987 "All the free space should be compacted into one chunk at top");
1988 assert(cms_space->dictionary()->totalChunkSize(
1989 debug_only(cms_space->freelistLock())) == 0 ||
1990 cms_space->totalSizeInIndexedFreeLists() == 0,
1991 "All the free space should be in a single chunk");
1992 size_t num = cms_space->totalCount();
1993 assert((free_size == 0 && num == 0) ||
1994 (free_size > 0 && (num == 1 || num == 2)),
1995 "There should be at most 2 free chunks after compaction");
1996 #endif // ASSERT
1997 _collectorState = Resetting;
1998 assert(_restart_addr == NULL,
1999 "Should have been NULL'd before baton was passed");
2000 reset(false /* == !asynch */);
2001 _cmsGen->reset_after_compaction();
2002 _concurrent_cycles_since_last_unload = 0;
2004 if (verifying() && !should_unload_classes()) {
2005 perm_gen_verify_bit_map()->clear_all();
2006 }
2008 // Clear any data recorded in the PLAB chunk arrays.
2009 if (_survivor_plab_array != NULL) {
2010 reset_survivor_plab_arrays();
2011 }
2013 // Adjust the per-size allocation stats for the next epoch.
2014 _cmsGen->cmsSpace()->endSweepFLCensus(sweepCount() /* fake */);
2015 // Restart the "sweep timer" for next epoch.
2016 _sweep_timer.reset();
2017 _sweep_timer.start();
2019 // Sample collection pause time and reset for collection interval.
2020 if (UseAdaptiveSizePolicy) {
2021 size_policy()->msc_collection_end(gch->gc_cause());
2022 }
2024 // For a mark-sweep-compact, compute_new_size() will be called
2025 // in the heap's do_collection() method.
2026 }
2028 // A work method used by the foreground collector to do
2029 // a mark-sweep, after taking over from a possibly on-going
2030 // concurrent mark-sweep collection.
2031 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2032 CollectorState first_state, bool should_start_over) {
2033 if (PrintGC && Verbose) {
2034 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2035 "collector with count %d",
2036 _full_gcs_since_conc_gc);
2037 }
2038 switch (_collectorState) {
2039 case Idling:
2040 if (first_state == Idling || should_start_over) {
2041 // The background GC was not active, or should
2042 // restarted from scratch; start the cycle.
2043 _collectorState = InitialMarking;
2044 }
2045 // If first_state was not Idling, then a background GC
2046 // was in progress and has now finished. No need to do it
2047 // again. Leave the state as Idling.
2048 break;
2049 case Precleaning:
2050 // In the foreground case don't do the precleaning since
2051 // it is not done concurrently and there is extra work
2052 // required.
2053 _collectorState = FinalMarking;
2054 }
2055 if (PrintGCDetails &&
2056 (_collectorState > Idling ||
2057 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2058 gclog_or_tty->print(" (concurrent mode failure)");
2059 }
2060 collect_in_foreground(clear_all_soft_refs);
2062 // For a mark-sweep, compute_new_size() will be called
2063 // in the heap's do_collection() method.
2064 }
2067 void CMSCollector::getFreelistLocks() const {
2068 // Get locks for all free lists in all generations that this
2069 // collector is responsible for
2070 _cmsGen->freelistLock()->lock_without_safepoint_check();
2071 _permGen->freelistLock()->lock_without_safepoint_check();
2072 }
2074 void CMSCollector::releaseFreelistLocks() const {
2075 // Release locks for all free lists in all generations that this
2076 // collector is responsible for
2077 _cmsGen->freelistLock()->unlock();
2078 _permGen->freelistLock()->unlock();
2079 }
2081 bool CMSCollector::haveFreelistLocks() const {
2082 // Check locks for all free lists in all generations that this
2083 // collector is responsible for
2084 assert_lock_strong(_cmsGen->freelistLock());
2085 assert_lock_strong(_permGen->freelistLock());
2086 PRODUCT_ONLY(ShouldNotReachHere());
2087 return true;
2088 }
2090 // A utility class that is used by the CMS collector to
2091 // temporarily "release" the foreground collector from its
2092 // usual obligation to wait for the background collector to
2093 // complete an ongoing phase before proceeding.
2094 class ReleaseForegroundGC: public StackObj {
2095 private:
2096 CMSCollector* _c;
2097 public:
2098 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2099 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2100 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2101 // allow a potentially blocked foreground collector to proceed
2102 _c->_foregroundGCShouldWait = false;
2103 if (_c->_foregroundGCIsActive) {
2104 CGC_lock->notify();
2105 }
2106 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2107 "Possible deadlock");
2108 }
2110 ~ReleaseForegroundGC() {
2111 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2112 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2113 _c->_foregroundGCShouldWait = true;
2114 }
2115 };
2117 // There are separate collect_in_background and collect_in_foreground because of
2118 // the different locking requirements of the background collector and the
2119 // foreground collector. There was originally an attempt to share
2120 // one "collect" method between the background collector and the foreground
2121 // collector but the if-then-else required made it cleaner to have
2122 // separate methods.
2123 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2124 assert(Thread::current()->is_ConcurrentGC_thread(),
2125 "A CMS asynchronous collection is only allowed on a CMS thread.");
2127 GenCollectedHeap* gch = GenCollectedHeap::heap();
2128 {
2129 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2130 MutexLockerEx hl(Heap_lock, safepoint_check);
2131 FreelistLocker fll(this);
2132 MutexLockerEx x(CGC_lock, safepoint_check);
2133 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2134 // The foreground collector is active or we're
2135 // not using asynchronous collections. Skip this
2136 // background collection.
2137 assert(!_foregroundGCShouldWait, "Should be clear");
2138 return;
2139 } else {
2140 assert(_collectorState == Idling, "Should be idling before start.");
2141 _collectorState = InitialMarking;
2142 // Reset the expansion cause, now that we are about to begin
2143 // a new cycle.
2144 clear_expansion_cause();
2145 }
2146 // Decide if we want to enable class unloading as part of the
2147 // ensuing concurrent GC cycle.
2148 update_should_unload_classes();
2149 _full_gc_requested = false; // acks all outstanding full gc requests
2150 // Signal that we are about to start a collection
2151 gch->increment_total_full_collections(); // ... starting a collection cycle
2152 _collection_count_start = gch->total_full_collections();
2153 }
2155 // Used for PrintGC
2156 size_t prev_used;
2157 if (PrintGC && Verbose) {
2158 prev_used = _cmsGen->used(); // XXXPERM
2159 }
2161 // The change of the collection state is normally done at this level;
2162 // the exceptions are phases that are executed while the world is
2163 // stopped. For those phases the change of state is done while the
2164 // world is stopped. For baton passing purposes this allows the
2165 // background collector to finish the phase and change state atomically.
2166 // The foreground collector cannot wait on a phase that is done
2167 // while the world is stopped because the foreground collector already
2168 // has the world stopped and would deadlock.
2169 while (_collectorState != Idling) {
2170 if (TraceCMSState) {
2171 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2172 Thread::current(), _collectorState);
2173 }
2174 // The foreground collector
2175 // holds the Heap_lock throughout its collection.
2176 // holds the CMS token (but not the lock)
2177 // except while it is waiting for the background collector to yield.
2178 //
2179 // The foreground collector should be blocked (not for long)
2180 // if the background collector is about to start a phase
2181 // executed with world stopped. If the background
2182 // collector has already started such a phase, the
2183 // foreground collector is blocked waiting for the
2184 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2185 // are executed in the VM thread.
2186 //
2187 // The locking order is
2188 // PendingListLock (PLL) -- if applicable (FinalMarking)
2189 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2190 // CMS token (claimed in
2191 // stop_world_and_do() -->
2192 // safepoint_synchronize() -->
2193 // CMSThread::synchronize())
2195 {
2196 // Check if the FG collector wants us to yield.
2197 CMSTokenSync x(true); // is cms thread
2198 if (waitForForegroundGC()) {
2199 // We yielded to a foreground GC, nothing more to be
2200 // done this round.
2201 assert(_foregroundGCShouldWait == false, "We set it to false in "
2202 "waitForForegroundGC()");
2203 if (TraceCMSState) {
2204 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2205 " exiting collection CMS state %d",
2206 Thread::current(), _collectorState);
2207 }
2208 return;
2209 } else {
2210 // The background collector can run but check to see if the
2211 // foreground collector has done a collection while the
2212 // background collector was waiting to get the CGC_lock
2213 // above. If yes, break so that _foregroundGCShouldWait
2214 // is cleared before returning.
2215 if (_collectorState == Idling) {
2216 break;
2217 }
2218 }
2219 }
2221 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2222 "should be waiting");
2224 switch (_collectorState) {
2225 case InitialMarking:
2226 {
2227 ReleaseForegroundGC x(this);
2228 stats().record_cms_begin();
2230 VM_CMS_Initial_Mark initial_mark_op(this);
2231 VMThread::execute(&initial_mark_op);
2232 }
2233 // The collector state may be any legal state at this point
2234 // since the background collector may have yielded to the
2235 // foreground collector.
2236 break;
2237 case Marking:
2238 // initial marking in checkpointRootsInitialWork has been completed
2239 if (markFromRoots(true)) { // we were successful
2240 assert(_collectorState == Precleaning, "Collector state should "
2241 "have changed");
2242 } else {
2243 assert(_foregroundGCIsActive, "Internal state inconsistency");
2244 }
2245 break;
2246 case Precleaning:
2247 if (UseAdaptiveSizePolicy) {
2248 size_policy()->concurrent_precleaning_begin();
2249 }
2250 // marking from roots in markFromRoots has been completed
2251 preclean();
2252 if (UseAdaptiveSizePolicy) {
2253 size_policy()->concurrent_precleaning_end();
2254 }
2255 assert(_collectorState == AbortablePreclean ||
2256 _collectorState == FinalMarking,
2257 "Collector state should have changed");
2258 break;
2259 case AbortablePreclean:
2260 if (UseAdaptiveSizePolicy) {
2261 size_policy()->concurrent_phases_resume();
2262 }
2263 abortable_preclean();
2264 if (UseAdaptiveSizePolicy) {
2265 size_policy()->concurrent_precleaning_end();
2266 }
2267 assert(_collectorState == FinalMarking, "Collector state should "
2268 "have changed");
2269 break;
2270 case FinalMarking:
2271 {
2272 ReleaseForegroundGC x(this);
2274 VM_CMS_Final_Remark final_remark_op(this);
2275 VMThread::execute(&final_remark_op);
2276 }
2277 assert(_foregroundGCShouldWait, "block post-condition");
2278 break;
2279 case Sweeping:
2280 if (UseAdaptiveSizePolicy) {
2281 size_policy()->concurrent_sweeping_begin();
2282 }
2283 // final marking in checkpointRootsFinal has been completed
2284 sweep(true);
2285 assert(_collectorState == Resizing, "Collector state change "
2286 "to Resizing must be done under the free_list_lock");
2287 _full_gcs_since_conc_gc = 0;
2289 // Stop the timers for adaptive size policy for the concurrent phases
2290 if (UseAdaptiveSizePolicy) {
2291 size_policy()->concurrent_sweeping_end();
2292 size_policy()->concurrent_phases_end(gch->gc_cause(),
2293 gch->prev_gen(_cmsGen)->capacity(),
2294 _cmsGen->free());
2295 }
2297 case Resizing: {
2298 // Sweeping has been completed...
2299 // At this point the background collection has completed.
2300 // Don't move the call to compute_new_size() down
2301 // into code that might be executed if the background
2302 // collection was preempted.
2303 {
2304 ReleaseForegroundGC x(this); // unblock FG collection
2305 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2306 CMSTokenSync z(true); // not strictly needed.
2307 if (_collectorState == Resizing) {
2308 compute_new_size();
2309 _collectorState = Resetting;
2310 } else {
2311 assert(_collectorState == Idling, "The state should only change"
2312 " because the foreground collector has finished the collection");
2313 }
2314 }
2315 break;
2316 }
2317 case Resetting:
2318 // CMS heap resizing has been completed
2319 reset(true);
2320 assert(_collectorState == Idling, "Collector state should "
2321 "have changed");
2322 stats().record_cms_end();
2323 // Don't move the concurrent_phases_end() and compute_new_size()
2324 // calls to here because a preempted background collection
2325 // has it's state set to "Resetting".
2326 break;
2327 case Idling:
2328 default:
2329 ShouldNotReachHere();
2330 break;
2331 }
2332 if (TraceCMSState) {
2333 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2334 Thread::current(), _collectorState);
2335 }
2336 assert(_foregroundGCShouldWait, "block post-condition");
2337 }
2339 // Should this be in gc_epilogue?
2340 collector_policy()->counters()->update_counters();
2342 {
2343 // Clear _foregroundGCShouldWait and, in the event that the
2344 // foreground collector is waiting, notify it, before
2345 // returning.
2346 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2347 _foregroundGCShouldWait = false;
2348 if (_foregroundGCIsActive) {
2349 CGC_lock->notify();
2350 }
2351 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2352 "Possible deadlock");
2353 }
2354 if (TraceCMSState) {
2355 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2356 " exiting collection CMS state %d",
2357 Thread::current(), _collectorState);
2358 }
2359 if (PrintGC && Verbose) {
2360 _cmsGen->print_heap_change(prev_used);
2361 }
2362 }
2364 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2365 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2366 "Foreground collector should be waiting, not executing");
2367 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2368 "may only be done by the VM Thread with the world stopped");
2369 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2370 "VM thread should have CMS token");
2372 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2373 true, gclog_or_tty);)
2374 if (UseAdaptiveSizePolicy) {
2375 size_policy()->ms_collection_begin();
2376 }
2377 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2379 HandleMark hm; // Discard invalid handles created during verification
2381 if (VerifyBeforeGC &&
2382 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2383 Universe::verify(true);
2384 }
2386 bool init_mark_was_synchronous = false; // until proven otherwise
2387 while (_collectorState != Idling) {
2388 if (TraceCMSState) {
2389 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2390 Thread::current(), _collectorState);
2391 }
2392 switch (_collectorState) {
2393 case InitialMarking:
2394 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2395 checkpointRootsInitial(false);
2396 assert(_collectorState == Marking, "Collector state should have changed"
2397 " within checkpointRootsInitial()");
2398 break;
2399 case Marking:
2400 // initial marking in checkpointRootsInitialWork has been completed
2401 if (VerifyDuringGC &&
2402 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2403 gclog_or_tty->print("Verify before initial mark: ");
2404 Universe::verify(true);
2405 }
2406 {
2407 bool res = markFromRoots(false);
2408 assert(res && _collectorState == FinalMarking, "Collector state should "
2409 "have changed");
2410 break;
2411 }
2412 case FinalMarking:
2413 if (VerifyDuringGC &&
2414 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2415 gclog_or_tty->print("Verify before re-mark: ");
2416 Universe::verify(true);
2417 }
2418 checkpointRootsFinal(false, clear_all_soft_refs,
2419 init_mark_was_synchronous);
2420 assert(_collectorState == Sweeping, "Collector state should not "
2421 "have changed within checkpointRootsFinal()");
2422 break;
2423 case Sweeping:
2424 // final marking in checkpointRootsFinal has been completed
2425 if (VerifyDuringGC &&
2426 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2427 gclog_or_tty->print("Verify before sweep: ");
2428 Universe::verify(true);
2429 }
2430 sweep(false);
2431 assert(_collectorState == Resizing, "Incorrect state");
2432 break;
2433 case Resizing: {
2434 // Sweeping has been completed; the actual resize in this case
2435 // is done separately; nothing to be done in this state.
2436 _collectorState = Resetting;
2437 break;
2438 }
2439 case Resetting:
2440 // The heap has been resized.
2441 if (VerifyDuringGC &&
2442 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2443 gclog_or_tty->print("Verify before reset: ");
2444 Universe::verify(true);
2445 }
2446 reset(false);
2447 assert(_collectorState == Idling, "Collector state should "
2448 "have changed");
2449 break;
2450 case Precleaning:
2451 case AbortablePreclean:
2452 // Elide the preclean phase
2453 _collectorState = FinalMarking;
2454 break;
2455 default:
2456 ShouldNotReachHere();
2457 }
2458 if (TraceCMSState) {
2459 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2460 Thread::current(), _collectorState);
2461 }
2462 }
2464 if (UseAdaptiveSizePolicy) {
2465 GenCollectedHeap* gch = GenCollectedHeap::heap();
2466 size_policy()->ms_collection_end(gch->gc_cause());
2467 }
2469 if (VerifyAfterGC &&
2470 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2471 Universe::verify(true);
2472 }
2473 if (TraceCMSState) {
2474 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2475 " exiting collection CMS state %d",
2476 Thread::current(), _collectorState);
2477 }
2478 }
2480 bool CMSCollector::waitForForegroundGC() {
2481 bool res = false;
2482 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2483 "CMS thread should have CMS token");
2484 // Block the foreground collector until the
2485 // background collectors decides whether to
2486 // yield.
2487 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2488 _foregroundGCShouldWait = true;
2489 if (_foregroundGCIsActive) {
2490 // The background collector yields to the
2491 // foreground collector and returns a value
2492 // indicating that it has yielded. The foreground
2493 // collector can proceed.
2494 res = true;
2495 _foregroundGCShouldWait = false;
2496 ConcurrentMarkSweepThread::clear_CMS_flag(
2497 ConcurrentMarkSweepThread::CMS_cms_has_token);
2498 ConcurrentMarkSweepThread::set_CMS_flag(
2499 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2500 // Get a possibly blocked foreground thread going
2501 CGC_lock->notify();
2502 if (TraceCMSState) {
2503 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2504 Thread::current(), _collectorState);
2505 }
2506 while (_foregroundGCIsActive) {
2507 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2508 }
2509 ConcurrentMarkSweepThread::set_CMS_flag(
2510 ConcurrentMarkSweepThread::CMS_cms_has_token);
2511 ConcurrentMarkSweepThread::clear_CMS_flag(
2512 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2513 }
2514 if (TraceCMSState) {
2515 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2516 Thread::current(), _collectorState);
2517 }
2518 return res;
2519 }
2521 // Because of the need to lock the free lists and other structures in
2522 // the collector, common to all the generations that the collector is
2523 // collecting, we need the gc_prologues of individual CMS generations
2524 // delegate to their collector. It may have been simpler had the
2525 // current infrastructure allowed one to call a prologue on a
2526 // collector. In the absence of that we have the generation's
2527 // prologue delegate to the collector, which delegates back
2528 // some "local" work to a worker method in the individual generations
2529 // that it's responsible for collecting, while itself doing any
2530 // work common to all generations it's responsible for. A similar
2531 // comment applies to the gc_epilogue()'s.
2532 // The role of the varaible _between_prologue_and_epilogue is to
2533 // enforce the invocation protocol.
2534 void CMSCollector::gc_prologue(bool full) {
2535 // Call gc_prologue_work() for each CMSGen and PermGen that
2536 // we are responsible for.
2538 // The following locking discipline assumes that we are only called
2539 // when the world is stopped.
2540 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2542 // The CMSCollector prologue must call the gc_prologues for the
2543 // "generations" (including PermGen if any) that it's responsible
2544 // for.
2546 assert( Thread::current()->is_VM_thread()
2547 || ( CMSScavengeBeforeRemark
2548 && Thread::current()->is_ConcurrentGC_thread()),
2549 "Incorrect thread type for prologue execution");
2551 if (_between_prologue_and_epilogue) {
2552 // We have already been invoked; this is a gc_prologue delegation
2553 // from yet another CMS generation that we are responsible for, just
2554 // ignore it since all relevant work has already been done.
2555 return;
2556 }
2558 // set a bit saying prologue has been called; cleared in epilogue
2559 _between_prologue_and_epilogue = true;
2560 // Claim locks for common data structures, then call gc_prologue_work()
2561 // for each CMSGen and PermGen that we are responsible for.
2563 getFreelistLocks(); // gets free list locks on constituent spaces
2564 bitMapLock()->lock_without_safepoint_check();
2566 // Should call gc_prologue_work() for all cms gens we are responsible for
2567 bool registerClosure = _collectorState >= Marking
2568 && _collectorState < Sweeping;
2569 ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar
2570 : &_modUnionClosure;
2571 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2572 _permGen->gc_prologue_work(full, registerClosure, muc);
2574 if (!full) {
2575 stats().record_gc0_begin();
2576 }
2577 }
2579 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2580 // Delegate to CMScollector which knows how to coordinate between
2581 // this and any other CMS generations that it is responsible for
2582 // collecting.
2583 collector()->gc_prologue(full);
2584 }
2586 // This is a "private" interface for use by this generation's CMSCollector.
2587 // Not to be called directly by any other entity (for instance,
2588 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2589 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2590 bool registerClosure, ModUnionClosure* modUnionClosure) {
2591 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2592 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2593 "Should be NULL");
2594 if (registerClosure) {
2595 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2596 }
2597 cmsSpace()->gc_prologue();
2598 // Clear stat counters
2599 NOT_PRODUCT(
2600 assert(_numObjectsPromoted == 0, "check");
2601 assert(_numWordsPromoted == 0, "check");
2602 if (Verbose && PrintGC) {
2603 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2604 SIZE_FORMAT" bytes concurrently",
2605 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2606 }
2607 _numObjectsAllocated = 0;
2608 _numWordsAllocated = 0;
2609 )
2610 }
2612 void CMSCollector::gc_epilogue(bool full) {
2613 // The following locking discipline assumes that we are only called
2614 // when the world is stopped.
2615 assert(SafepointSynchronize::is_at_safepoint(),
2616 "world is stopped assumption");
2618 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2619 // if linear allocation blocks need to be appropriately marked to allow the
2620 // the blocks to be parsable. We also check here whether we need to nudge the
2621 // CMS collector thread to start a new cycle (if it's not already active).
2622 assert( Thread::current()->is_VM_thread()
2623 || ( CMSScavengeBeforeRemark
2624 && Thread::current()->is_ConcurrentGC_thread()),
2625 "Incorrect thread type for epilogue execution");
2627 if (!_between_prologue_and_epilogue) {
2628 // We have already been invoked; this is a gc_epilogue delegation
2629 // from yet another CMS generation that we are responsible for, just
2630 // ignore it since all relevant work has already been done.
2631 return;
2632 }
2633 assert(haveFreelistLocks(), "must have freelist locks");
2634 assert_lock_strong(bitMapLock());
2636 _cmsGen->gc_epilogue_work(full);
2637 _permGen->gc_epilogue_work(full);
2639 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2640 // in case sampling was not already enabled, enable it
2641 _start_sampling = true;
2642 }
2643 // reset _eden_chunk_array so sampling starts afresh
2644 _eden_chunk_index = 0;
2646 size_t cms_used = _cmsGen->cmsSpace()->used();
2647 size_t perm_used = _permGen->cmsSpace()->used();
2649 // update performance counters - this uses a special version of
2650 // update_counters() that allows the utilization to be passed as a
2651 // parameter, avoiding multiple calls to used().
2652 //
2653 _cmsGen->update_counters(cms_used);
2654 _permGen->update_counters(perm_used);
2656 if (CMSIncrementalMode) {
2657 icms_update_allocation_limits();
2658 }
2660 bitMapLock()->unlock();
2661 releaseFreelistLocks();
2663 _between_prologue_and_epilogue = false; // ready for next cycle
2664 }
2666 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2667 collector()->gc_epilogue(full);
2669 // Also reset promotion tracking in par gc thread states.
2670 if (ParallelGCThreads > 0) {
2671 for (uint i = 0; i < ParallelGCThreads; i++) {
2672 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2673 }
2674 }
2675 }
2677 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2678 assert(!incremental_collection_failed(), "Should have been cleared");
2679 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2680 cmsSpace()->gc_epilogue();
2681 // Print stat counters
2682 NOT_PRODUCT(
2683 assert(_numObjectsAllocated == 0, "check");
2684 assert(_numWordsAllocated == 0, "check");
2685 if (Verbose && PrintGC) {
2686 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2687 SIZE_FORMAT" bytes",
2688 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2689 }
2690 _numObjectsPromoted = 0;
2691 _numWordsPromoted = 0;
2692 )
2694 if (PrintGC && Verbose) {
2695 // Call down the chain in contiguous_available needs the freelistLock
2696 // so print this out before releasing the freeListLock.
2697 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2698 contiguous_available());
2699 }
2700 }
2702 #ifndef PRODUCT
2703 bool CMSCollector::have_cms_token() {
2704 Thread* thr = Thread::current();
2705 if (thr->is_VM_thread()) {
2706 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2707 } else if (thr->is_ConcurrentGC_thread()) {
2708 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2709 } else if (thr->is_GC_task_thread()) {
2710 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2711 ParGCRareEvent_lock->owned_by_self();
2712 }
2713 return false;
2714 }
2715 #endif
2717 // Check reachability of the given heap address in CMS generation,
2718 // treating all other generations as roots.
2719 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2720 // We could "guarantee" below, rather than assert, but i'll
2721 // leave these as "asserts" so that an adventurous debugger
2722 // could try this in the product build provided some subset of
2723 // the conditions were met, provided they were intersted in the
2724 // results and knew that the computation below wouldn't interfere
2725 // with other concurrent computations mutating the structures
2726 // being read or written.
2727 assert(SafepointSynchronize::is_at_safepoint(),
2728 "Else mutations in object graph will make answer suspect");
2729 assert(have_cms_token(), "Should hold cms token");
2730 assert(haveFreelistLocks(), "must hold free list locks");
2731 assert_lock_strong(bitMapLock());
2733 // Clear the marking bit map array before starting, but, just
2734 // for kicks, first report if the given address is already marked
2735 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2736 _markBitMap.isMarked(addr) ? "" : " not");
2738 if (verify_after_remark()) {
2739 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2740 bool result = verification_mark_bm()->isMarked(addr);
2741 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2742 result ? "IS" : "is NOT");
2743 return result;
2744 } else {
2745 gclog_or_tty->print_cr("Could not compute result");
2746 return false;
2747 }
2748 }
2750 ////////////////////////////////////////////////////////
2751 // CMS Verification Support
2752 ////////////////////////////////////////////////////////
2753 // Following the remark phase, the following invariant
2754 // should hold -- each object in the CMS heap which is
2755 // marked in markBitMap() should be marked in the verification_mark_bm().
2757 class VerifyMarkedClosure: public BitMapClosure {
2758 CMSBitMap* _marks;
2759 bool _failed;
2761 public:
2762 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2764 bool do_bit(size_t offset) {
2765 HeapWord* addr = _marks->offsetToHeapWord(offset);
2766 if (!_marks->isMarked(addr)) {
2767 oop(addr)->print();
2768 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2769 _failed = true;
2770 }
2771 return true;
2772 }
2774 bool failed() { return _failed; }
2775 };
2777 bool CMSCollector::verify_after_remark() {
2778 gclog_or_tty->print(" [Verifying CMS Marking... ");
2779 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2780 static bool init = false;
2782 assert(SafepointSynchronize::is_at_safepoint(),
2783 "Else mutations in object graph will make answer suspect");
2784 assert(have_cms_token(),
2785 "Else there may be mutual interference in use of "
2786 " verification data structures");
2787 assert(_collectorState > Marking && _collectorState <= Sweeping,
2788 "Else marking info checked here may be obsolete");
2789 assert(haveFreelistLocks(), "must hold free list locks");
2790 assert_lock_strong(bitMapLock());
2793 // Allocate marking bit map if not already allocated
2794 if (!init) { // first time
2795 if (!verification_mark_bm()->allocate(_span)) {
2796 return false;
2797 }
2798 init = true;
2799 }
2801 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2803 // Turn off refs discovery -- so we will be tracing through refs.
2804 // This is as intended, because by this time
2805 // GC must already have cleared any refs that need to be cleared,
2806 // and traced those that need to be marked; moreover,
2807 // the marking done here is not going to intefere in any
2808 // way with the marking information used by GC.
2809 NoRefDiscovery no_discovery(ref_processor());
2811 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2813 // Clear any marks from a previous round
2814 verification_mark_bm()->clear_all();
2815 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2816 assert(overflow_list_is_empty(), "overflow list should be empty");
2818 GenCollectedHeap* gch = GenCollectedHeap::heap();
2819 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2820 // Update the saved marks which may affect the root scans.
2821 gch->save_marks();
2823 if (CMSRemarkVerifyVariant == 1) {
2824 // In this first variant of verification, we complete
2825 // all marking, then check if the new marks-verctor is
2826 // a subset of the CMS marks-vector.
2827 verify_after_remark_work_1();
2828 } else if (CMSRemarkVerifyVariant == 2) {
2829 // In this second variant of verification, we flag an error
2830 // (i.e. an object reachable in the new marks-vector not reachable
2831 // in the CMS marks-vector) immediately, also indicating the
2832 // identify of an object (A) that references the unmarked object (B) --
2833 // presumably, a mutation to A failed to be picked up by preclean/remark?
2834 verify_after_remark_work_2();
2835 } else {
2836 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2837 CMSRemarkVerifyVariant);
2838 }
2839 gclog_or_tty->print(" done] ");
2840 return true;
2841 }
2843 void CMSCollector::verify_after_remark_work_1() {
2844 ResourceMark rm;
2845 HandleMark hm;
2846 GenCollectedHeap* gch = GenCollectedHeap::heap();
2848 // Mark from roots one level into CMS
2849 MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */);
2850 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2852 gch->gen_process_strong_roots(_cmsGen->level(),
2853 true, // younger gens are roots
2854 true, // collecting perm gen
2855 SharedHeap::ScanningOption(roots_scanning_options()),
2856 NULL, ¬Older);
2858 // Now mark from the roots
2859 assert(_revisitStack.isEmpty(), "Should be empty");
2860 MarkFromRootsClosure markFromRootsClosure(this, _span,
2861 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2862 false /* don't yield */, true /* verifying */);
2863 assert(_restart_addr == NULL, "Expected pre-condition");
2864 verification_mark_bm()->iterate(&markFromRootsClosure);
2865 while (_restart_addr != NULL) {
2866 // Deal with stack overflow: by restarting at the indicated
2867 // address.
2868 HeapWord* ra = _restart_addr;
2869 markFromRootsClosure.reset(ra);
2870 _restart_addr = NULL;
2871 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2872 }
2873 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2874 verify_work_stacks_empty();
2875 // Should reset the revisit stack above, since no class tree
2876 // surgery is forthcoming.
2877 _revisitStack.reset(); // throwing away all contents
2879 // Marking completed -- now verify that each bit marked in
2880 // verification_mark_bm() is also marked in markBitMap(); flag all
2881 // errors by printing corresponding objects.
2882 VerifyMarkedClosure vcl(markBitMap());
2883 verification_mark_bm()->iterate(&vcl);
2884 if (vcl.failed()) {
2885 gclog_or_tty->print("Verification failed");
2886 Universe::heap()->print();
2887 fatal(" ... aborting");
2888 }
2889 }
2891 void CMSCollector::verify_after_remark_work_2() {
2892 ResourceMark rm;
2893 HandleMark hm;
2894 GenCollectedHeap* gch = GenCollectedHeap::heap();
2896 // Mark from roots one level into CMS
2897 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2898 markBitMap(), true /* nmethods */);
2899 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2900 gch->gen_process_strong_roots(_cmsGen->level(),
2901 true, // younger gens are roots
2902 true, // collecting perm gen
2903 SharedHeap::ScanningOption(roots_scanning_options()),
2904 NULL, ¬Older);
2906 // Now mark from the roots
2907 assert(_revisitStack.isEmpty(), "Should be empty");
2908 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2909 verification_mark_bm(), markBitMap(), verification_mark_stack());
2910 assert(_restart_addr == NULL, "Expected pre-condition");
2911 verification_mark_bm()->iterate(&markFromRootsClosure);
2912 while (_restart_addr != NULL) {
2913 // Deal with stack overflow: by restarting at the indicated
2914 // address.
2915 HeapWord* ra = _restart_addr;
2916 markFromRootsClosure.reset(ra);
2917 _restart_addr = NULL;
2918 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2919 }
2920 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2921 verify_work_stacks_empty();
2922 // Should reset the revisit stack above, since no class tree
2923 // surgery is forthcoming.
2924 _revisitStack.reset(); // throwing away all contents
2926 // Marking completed -- now verify that each bit marked in
2927 // verification_mark_bm() is also marked in markBitMap(); flag all
2928 // errors by printing corresponding objects.
2929 VerifyMarkedClosure vcl(markBitMap());
2930 verification_mark_bm()->iterate(&vcl);
2931 assert(!vcl.failed(), "Else verification above should not have succeeded");
2932 }
2934 void ConcurrentMarkSweepGeneration::save_marks() {
2935 // delegate to CMS space
2936 cmsSpace()->save_marks();
2937 for (uint i = 0; i < ParallelGCThreads; i++) {
2938 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2939 }
2940 }
2942 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2943 return cmsSpace()->no_allocs_since_save_marks();
2944 }
2946 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2947 \
2948 void ConcurrentMarkSweepGeneration:: \
2949 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2950 cl->set_generation(this); \
2951 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2952 cl->reset_generation(); \
2953 save_marks(); \
2954 }
2956 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2958 void
2959 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
2960 {
2961 // Not currently implemented; need to do the following. -- ysr.
2962 // dld -- I think that is used for some sort of allocation profiler. So it
2963 // really means the objects allocated by the mutator since the last
2964 // GC. We could potentially implement this cheaply by recording only
2965 // the direct allocations in a side data structure.
2966 //
2967 // I think we probably ought not to be required to support these
2968 // iterations at any arbitrary point; I think there ought to be some
2969 // call to enable/disable allocation profiling in a generation/space,
2970 // and the iterator ought to return the objects allocated in the
2971 // gen/space since the enable call, or the last iterator call (which
2972 // will probably be at a GC.) That way, for gens like CM&S that would
2973 // require some extra data structure to support this, we only pay the
2974 // cost when it's in use...
2975 cmsSpace()->object_iterate_since_last_GC(blk);
2976 }
2978 void
2979 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
2980 cl->set_generation(this);
2981 younger_refs_in_space_iterate(_cmsSpace, cl);
2982 cl->reset_generation();
2983 }
2985 void
2986 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
2987 if (freelistLock()->owned_by_self()) {
2988 Generation::oop_iterate(mr, cl);
2989 } else {
2990 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2991 Generation::oop_iterate(mr, cl);
2992 }
2993 }
2995 void
2996 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
2997 if (freelistLock()->owned_by_self()) {
2998 Generation::oop_iterate(cl);
2999 } else {
3000 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3001 Generation::oop_iterate(cl);
3002 }
3003 }
3005 void
3006 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3007 if (freelistLock()->owned_by_self()) {
3008 Generation::object_iterate(cl);
3009 } else {
3010 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3011 Generation::object_iterate(cl);
3012 }
3013 }
3015 void
3016 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3017 }
3019 void
3020 ConcurrentMarkSweepGeneration::post_compact() {
3021 }
3023 void
3024 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3025 // Fix the linear allocation blocks to look like free blocks.
3027 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3028 // are not called when the heap is verified during universe initialization and
3029 // at vm shutdown.
3030 if (freelistLock()->owned_by_self()) {
3031 cmsSpace()->prepare_for_verify();
3032 } else {
3033 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3034 cmsSpace()->prepare_for_verify();
3035 }
3036 }
3038 void
3039 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3040 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3041 // are not called when the heap is verified during universe initialization and
3042 // at vm shutdown.
3043 if (freelistLock()->owned_by_self()) {
3044 cmsSpace()->verify(false /* ignored */);
3045 } else {
3046 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3047 cmsSpace()->verify(false /* ignored */);
3048 }
3049 }
3051 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3052 _cmsGen->verify(allow_dirty);
3053 _permGen->verify(allow_dirty);
3054 }
3056 #ifndef PRODUCT
3057 bool CMSCollector::overflow_list_is_empty() const {
3058 assert(_num_par_pushes >= 0, "Inconsistency");
3059 if (_overflow_list == NULL) {
3060 assert(_num_par_pushes == 0, "Inconsistency");
3061 }
3062 return _overflow_list == NULL;
3063 }
3065 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3066 // merely consolidate assertion checks that appear to occur together frequently.
3067 void CMSCollector::verify_work_stacks_empty() const {
3068 assert(_markStack.isEmpty(), "Marking stack should be empty");
3069 assert(overflow_list_is_empty(), "Overflow list should be empty");
3070 }
3072 void CMSCollector::verify_overflow_empty() const {
3073 assert(overflow_list_is_empty(), "Overflow list should be empty");
3074 assert(no_preserved_marks(), "No preserved marks");
3075 }
3076 #endif // PRODUCT
3078 // Decide if we want to enable class unloading as part of the
3079 // ensuing concurrent GC cycle. We will collect the perm gen and
3080 // unload classes if it's the case that:
3081 // (1) an explicit gc request has been made and the flag
3082 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3083 // (2) (a) class unloading is enabled at the command line, and
3084 // (b) (i) perm gen threshold has been crossed, or
3085 // (ii) old gen is getting really full, or
3086 // (iii) the previous N CMS collections did not collect the
3087 // perm gen
3088 // NOTE: Provided there is no change in the state of the heap between
3089 // calls to this method, it should have idempotent results. Moreover,
3090 // its results should be monotonically increasing (i.e. going from 0 to 1,
3091 // but not 1 to 0) between successive calls between which the heap was
3092 // not collected. For the implementation below, it must thus rely on
3093 // the property that concurrent_cycles_since_last_unload()
3094 // will not decrease unless a collection cycle happened and that
3095 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3096 // themselves also monotonic in that sense. See check_monotonicity()
3097 // below.
3098 bool CMSCollector::update_should_unload_classes() {
3099 _should_unload_classes = false;
3100 // Condition 1 above
3101 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3102 _should_unload_classes = true;
3103 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3104 // Disjuncts 2.b.(i,ii,iii) above
3105 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3106 CMSClassUnloadingMaxInterval)
3107 || _permGen->should_concurrent_collect()
3108 || _cmsGen->is_too_full();
3109 }
3110 return _should_unload_classes;
3111 }
3113 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3114 bool res = should_concurrent_collect();
3115 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3116 return res;
3117 }
3119 void CMSCollector::setup_cms_unloading_and_verification_state() {
3120 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3121 || VerifyBeforeExit;
3122 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings
3123 | SharedHeap::SO_CodeCache;
3125 if (should_unload_classes()) { // Should unload classes this cycle
3126 remove_root_scanning_option(rso); // Shrink the root set appropriately
3127 set_verifying(should_verify); // Set verification state for this cycle
3128 return; // Nothing else needs to be done at this time
3129 }
3131 // Not unloading classes this cycle
3132 assert(!should_unload_classes(), "Inconsitency!");
3133 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3134 // We were not verifying, or we _were_ unloading classes in the last cycle,
3135 // AND some verification options are enabled this cycle; in this case,
3136 // we must make sure that the deadness map is allocated if not already so,
3137 // and cleared (if already allocated previously --
3138 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3139 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3140 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3141 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3142 "permanent generation verification disabled");
3143 return; // Note that we leave verification disabled, so we'll retry this
3144 // allocation next cycle. We _could_ remember this failure
3145 // and skip further attempts and permanently disable verification
3146 // attempts if that is considered more desirable.
3147 }
3148 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3149 "_perm_gen_ver_bit_map inconsistency?");
3150 } else {
3151 perm_gen_verify_bit_map()->clear_all();
3152 }
3153 // Include symbols, strings and code cache elements to prevent their resurrection.
3154 add_root_scanning_option(rso);
3155 set_verifying(true);
3156 } else if (verifying() && !should_verify) {
3157 // We were verifying, but some verification flags got disabled.
3158 set_verifying(false);
3159 // Exclude symbols, strings and code cache elements from root scanning to
3160 // reduce IM and RM pauses.
3161 remove_root_scanning_option(rso);
3162 }
3163 }
3166 #ifndef PRODUCT
3167 HeapWord* CMSCollector::block_start(const void* p) const {
3168 const HeapWord* addr = (HeapWord*)p;
3169 if (_span.contains(p)) {
3170 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3171 return _cmsGen->cmsSpace()->block_start(p);
3172 } else {
3173 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3174 "Inconsistent _span?");
3175 return _permGen->cmsSpace()->block_start(p);
3176 }
3177 }
3178 return NULL;
3179 }
3180 #endif
3182 HeapWord*
3183 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3184 bool tlab,
3185 bool parallel) {
3186 assert(!tlab, "Can't deal with TLAB allocation");
3187 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3188 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3189 CMSExpansionCause::_satisfy_allocation);
3190 if (GCExpandToAllocateDelayMillis > 0) {
3191 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3192 }
3193 return have_lock_and_allocate(word_size, tlab);
3194 }
3196 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3197 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3198 // to CardGeneration and share it...
3199 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3200 CMSExpansionCause::Cause cause)
3201 {
3202 assert_locked_or_safepoint(Heap_lock);
3204 size_t aligned_bytes = ReservedSpace::page_align_size_up(bytes);
3205 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
3206 bool success = false;
3207 if (aligned_expand_bytes > aligned_bytes) {
3208 success = grow_by(aligned_expand_bytes);
3209 }
3210 if (!success) {
3211 success = grow_by(aligned_bytes);
3212 }
3213 if (!success) {
3214 size_t remaining_bytes = _virtual_space.uncommitted_size();
3215 if (remaining_bytes > 0) {
3216 success = grow_by(remaining_bytes);
3217 }
3218 }
3219 if (GC_locker::is_active()) {
3220 if (PrintGC && Verbose) {
3221 gclog_or_tty->print_cr("Garbage collection disabled, expanded heap instead");
3222 }
3223 }
3224 // remember why we expanded; this information is used
3225 // by shouldConcurrentCollect() when making decisions on whether to start
3226 // a new CMS cycle.
3227 if (success) {
3228 set_expansion_cause(cause);
3229 if (PrintGCDetails && Verbose) {
3230 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3231 CMSExpansionCause::to_string(cause));
3232 }
3233 }
3234 }
3236 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3237 HeapWord* res = NULL;
3238 MutexLocker x(ParGCRareEvent_lock);
3239 while (true) {
3240 // Expansion by some other thread might make alloc OK now:
3241 res = ps->lab.alloc(word_sz);
3242 if (res != NULL) return res;
3243 // If there's not enough expansion space available, give up.
3244 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3245 return NULL;
3246 }
3247 // Otherwise, we try expansion.
3248 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3249 CMSExpansionCause::_allocate_par_lab);
3250 // Now go around the loop and try alloc again;
3251 // A competing par_promote might beat us to the expansion space,
3252 // so we may go around the loop again if promotion fails agaion.
3253 if (GCExpandToAllocateDelayMillis > 0) {
3254 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3255 }
3256 }
3257 }
3260 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3261 PromotionInfo* promo) {
3262 MutexLocker x(ParGCRareEvent_lock);
3263 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3264 while (true) {
3265 // Expansion by some other thread might make alloc OK now:
3266 if (promo->ensure_spooling_space()) {
3267 assert(promo->has_spooling_space(),
3268 "Post-condition of successful ensure_spooling_space()");
3269 return true;
3270 }
3271 // If there's not enough expansion space available, give up.
3272 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3273 return false;
3274 }
3275 // Otherwise, we try expansion.
3276 expand(refill_size_bytes, MinHeapDeltaBytes,
3277 CMSExpansionCause::_allocate_par_spooling_space);
3278 // Now go around the loop and try alloc again;
3279 // A competing allocation might beat us to the expansion space,
3280 // so we may go around the loop again if allocation fails again.
3281 if (GCExpandToAllocateDelayMillis > 0) {
3282 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3283 }
3284 }
3285 }
3289 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3290 assert_locked_or_safepoint(Heap_lock);
3291 size_t size = ReservedSpace::page_align_size_down(bytes);
3292 if (size > 0) {
3293 shrink_by(size);
3294 }
3295 }
3297 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3298 assert_locked_or_safepoint(Heap_lock);
3299 bool result = _virtual_space.expand_by(bytes);
3300 if (result) {
3301 HeapWord* old_end = _cmsSpace->end();
3302 size_t new_word_size =
3303 heap_word_size(_virtual_space.committed_size());
3304 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3305 _bts->resize(new_word_size); // resize the block offset shared array
3306 Universe::heap()->barrier_set()->resize_covered_region(mr);
3307 // Hmmmm... why doesn't CFLS::set_end verify locking?
3308 // This is quite ugly; FIX ME XXX
3309 _cmsSpace->assert_locked();
3310 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3312 // update the space and generation capacity counters
3313 if (UsePerfData) {
3314 _space_counters->update_capacity();
3315 _gen_counters->update_all();
3316 }
3318 if (Verbose && PrintGC) {
3319 size_t new_mem_size = _virtual_space.committed_size();
3320 size_t old_mem_size = new_mem_size - bytes;
3321 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3322 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3323 }
3324 }
3325 return result;
3326 }
3328 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3329 assert_locked_or_safepoint(Heap_lock);
3330 bool success = true;
3331 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3332 if (remaining_bytes > 0) {
3333 success = grow_by(remaining_bytes);
3334 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3335 }
3336 return success;
3337 }
3339 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3340 assert_locked_or_safepoint(Heap_lock);
3341 assert_lock_strong(freelistLock());
3342 // XXX Fix when compaction is implemented.
3343 warning("Shrinking of CMS not yet implemented");
3344 return;
3345 }
3348 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3349 // phases.
3350 class CMSPhaseAccounting: public StackObj {
3351 public:
3352 CMSPhaseAccounting(CMSCollector *collector,
3353 const char *phase,
3354 bool print_cr = true);
3355 ~CMSPhaseAccounting();
3357 private:
3358 CMSCollector *_collector;
3359 const char *_phase;
3360 elapsedTimer _wallclock;
3361 bool _print_cr;
3363 public:
3364 // Not MT-safe; so do not pass around these StackObj's
3365 // where they may be accessed by other threads.
3366 jlong wallclock_millis() {
3367 assert(_wallclock.is_active(), "Wall clock should not stop");
3368 _wallclock.stop(); // to record time
3369 jlong ret = _wallclock.milliseconds();
3370 _wallclock.start(); // restart
3371 return ret;
3372 }
3373 };
3375 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3376 const char *phase,
3377 bool print_cr) :
3378 _collector(collector), _phase(phase), _print_cr(print_cr) {
3380 if (PrintCMSStatistics != 0) {
3381 _collector->resetYields();
3382 }
3383 if (PrintGCDetails && PrintGCTimeStamps) {
3384 gclog_or_tty->date_stamp(PrintGCDateStamps);
3385 gclog_or_tty->stamp();
3386 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3387 _collector->cmsGen()->short_name(), _phase);
3388 }
3389 _collector->resetTimer();
3390 _wallclock.start();
3391 _collector->startTimer();
3392 }
3394 CMSPhaseAccounting::~CMSPhaseAccounting() {
3395 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3396 _collector->stopTimer();
3397 _wallclock.stop();
3398 if (PrintGCDetails) {
3399 gclog_or_tty->date_stamp(PrintGCDateStamps);
3400 if (PrintGCTimeStamps) {
3401 gclog_or_tty->stamp();
3402 gclog_or_tty->print(": ");
3403 }
3404 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3405 _collector->cmsGen()->short_name(),
3406 _phase, _collector->timerValue(), _wallclock.seconds());
3407 if (_print_cr) {
3408 gclog_or_tty->print_cr("");
3409 }
3410 if (PrintCMSStatistics != 0) {
3411 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3412 _collector->yields());
3413 }
3414 }
3415 }
3417 // CMS work
3419 // Checkpoint the roots into this generation from outside
3420 // this generation. [Note this initial checkpoint need only
3421 // be approximate -- we'll do a catch up phase subsequently.]
3422 void CMSCollector::checkpointRootsInitial(bool asynch) {
3423 assert(_collectorState == InitialMarking, "Wrong collector state");
3424 check_correct_thread_executing();
3425 ReferenceProcessor* rp = ref_processor();
3426 SpecializationStats::clear();
3427 assert(_restart_addr == NULL, "Control point invariant");
3428 if (asynch) {
3429 // acquire locks for subsequent manipulations
3430 MutexLockerEx x(bitMapLock(),
3431 Mutex::_no_safepoint_check_flag);
3432 checkpointRootsInitialWork(asynch);
3433 rp->verify_no_references_recorded();
3434 rp->enable_discovery(); // enable ("weak") refs discovery
3435 _collectorState = Marking;
3436 } else {
3437 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3438 // which recognizes if we are a CMS generation, and doesn't try to turn on
3439 // discovery; verify that they aren't meddling.
3440 assert(!rp->discovery_is_atomic(),
3441 "incorrect setting of discovery predicate");
3442 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3443 "ref discovery for this generation kind");
3444 // already have locks
3445 checkpointRootsInitialWork(asynch);
3446 rp->enable_discovery(); // now enable ("weak") refs discovery
3447 _collectorState = Marking;
3448 }
3449 SpecializationStats::print();
3450 }
3452 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3453 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3454 assert(_collectorState == InitialMarking, "just checking");
3456 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3457 // precede our marking with a collection of all
3458 // younger generations to keep floating garbage to a minimum.
3459 // XXX: we won't do this for now -- it's an optimization to be done later.
3461 // already have locks
3462 assert_lock_strong(bitMapLock());
3463 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3465 // Setup the verification and class unloading state for this
3466 // CMS collection cycle.
3467 setup_cms_unloading_and_verification_state();
3469 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3470 PrintGCDetails && Verbose, true, gclog_or_tty);)
3471 if (UseAdaptiveSizePolicy) {
3472 size_policy()->checkpoint_roots_initial_begin();
3473 }
3475 // Reset all the PLAB chunk arrays if necessary.
3476 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3477 reset_survivor_plab_arrays();
3478 }
3480 ResourceMark rm;
3481 HandleMark hm;
3483 FalseClosure falseClosure;
3484 // In the case of a synchronous collection, we will elide the
3485 // remark step, so it's important to catch all the nmethod oops
3486 // in this step; hence the last argument to the constrcutor below.
3487 MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */);
3488 GenCollectedHeap* gch = GenCollectedHeap::heap();
3490 verify_work_stacks_empty();
3491 verify_overflow_empty();
3493 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3494 // Update the saved marks which may affect the root scans.
3495 gch->save_marks();
3497 // weak reference processing has not started yet.
3498 ref_processor()->set_enqueuing_is_done(false);
3500 {
3501 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3502 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3503 gch->gen_process_strong_roots(_cmsGen->level(),
3504 true, // younger gens are roots
3505 true, // collecting perm gen
3506 SharedHeap::ScanningOption(roots_scanning_options()),
3507 NULL, ¬Older);
3508 }
3510 // Clear mod-union table; it will be dirtied in the prologue of
3511 // CMS generation per each younger generation collection.
3513 assert(_modUnionTable.isAllClear(),
3514 "Was cleared in most recent final checkpoint phase"
3515 " or no bits are set in the gc_prologue before the start of the next "
3516 "subsequent marking phase.");
3518 // Temporarily disabled, since pre/post-consumption closures don't
3519 // care about precleaned cards
3520 #if 0
3521 {
3522 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3523 (HeapWord*)_virtual_space.high());
3524 _ct->ct_bs()->preclean_dirty_cards(mr);
3525 }
3526 #endif
3528 // Save the end of the used_region of the constituent generations
3529 // to be used to limit the extent of sweep in each generation.
3530 save_sweep_limits();
3531 if (UseAdaptiveSizePolicy) {
3532 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3533 }
3534 verify_overflow_empty();
3535 }
3537 bool CMSCollector::markFromRoots(bool asynch) {
3538 // we might be tempted to assert that:
3539 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3540 // "inconsistent argument?");
3541 // However that wouldn't be right, because it's possible that
3542 // a safepoint is indeed in progress as a younger generation
3543 // stop-the-world GC happens even as we mark in this generation.
3544 assert(_collectorState == Marking, "inconsistent state?");
3545 check_correct_thread_executing();
3546 verify_overflow_empty();
3548 bool res;
3549 if (asynch) {
3551 // Start the timers for adaptive size policy for the concurrent phases
3552 // Do it here so that the foreground MS can use the concurrent
3553 // timer since a foreground MS might has the sweep done concurrently
3554 // or STW.
3555 if (UseAdaptiveSizePolicy) {
3556 size_policy()->concurrent_marking_begin();
3557 }
3559 // Weak ref discovery note: We may be discovering weak
3560 // refs in this generation concurrent (but interleaved) with
3561 // weak ref discovery by a younger generation collector.
3563 CMSTokenSyncWithLocks ts(true, bitMapLock());
3564 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3565 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3566 res = markFromRootsWork(asynch);
3567 if (res) {
3568 _collectorState = Precleaning;
3569 } else { // We failed and a foreground collection wants to take over
3570 assert(_foregroundGCIsActive, "internal state inconsistency");
3571 assert(_restart_addr == NULL, "foreground will restart from scratch");
3572 if (PrintGCDetails) {
3573 gclog_or_tty->print_cr("bailing out to foreground collection");
3574 }
3575 }
3576 if (UseAdaptiveSizePolicy) {
3577 size_policy()->concurrent_marking_end();
3578 }
3579 } else {
3580 assert(SafepointSynchronize::is_at_safepoint(),
3581 "inconsistent with asynch == false");
3582 if (UseAdaptiveSizePolicy) {
3583 size_policy()->ms_collection_marking_begin();
3584 }
3585 // already have locks
3586 res = markFromRootsWork(asynch);
3587 _collectorState = FinalMarking;
3588 if (UseAdaptiveSizePolicy) {
3589 GenCollectedHeap* gch = GenCollectedHeap::heap();
3590 size_policy()->ms_collection_marking_end(gch->gc_cause());
3591 }
3592 }
3593 verify_overflow_empty();
3594 return res;
3595 }
3597 bool CMSCollector::markFromRootsWork(bool asynch) {
3598 // iterate over marked bits in bit map, doing a full scan and mark
3599 // from these roots using the following algorithm:
3600 // . if oop is to the right of the current scan pointer,
3601 // mark corresponding bit (we'll process it later)
3602 // . else (oop is to left of current scan pointer)
3603 // push oop on marking stack
3604 // . drain the marking stack
3606 // Note that when we do a marking step we need to hold the
3607 // bit map lock -- recall that direct allocation (by mutators)
3608 // and promotion (by younger generation collectors) is also
3609 // marking the bit map. [the so-called allocate live policy.]
3610 // Because the implementation of bit map marking is not
3611 // robust wrt simultaneous marking of bits in the same word,
3612 // we need to make sure that there is no such interference
3613 // between concurrent such updates.
3615 // already have locks
3616 assert_lock_strong(bitMapLock());
3618 // Clear the revisit stack, just in case there are any
3619 // obsolete contents from a short-circuited previous CMS cycle.
3620 _revisitStack.reset();
3621 verify_work_stacks_empty();
3622 verify_overflow_empty();
3623 assert(_revisitStack.isEmpty(), "tabula rasa");
3625 bool result = false;
3626 if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) {
3627 result = do_marking_mt(asynch);
3628 } else {
3629 result = do_marking_st(asynch);
3630 }
3631 return result;
3632 }
3634 // Forward decl
3635 class CMSConcMarkingTask;
3637 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3638 CMSCollector* _collector;
3639 CMSConcMarkingTask* _task;
3640 bool _yield;
3641 protected:
3642 virtual void yield();
3643 public:
3644 // "n_threads" is the number of threads to be terminated.
3645 // "queue_set" is a set of work queues of other threads.
3646 // "collector" is the CMS collector associated with this task terminator.
3647 // "yield" indicates whether we need the gang as a whole to yield.
3648 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
3649 CMSCollector* collector, bool yield) :
3650 ParallelTaskTerminator(n_threads, queue_set),
3651 _collector(collector),
3652 _yield(yield) { }
3654 void set_task(CMSConcMarkingTask* task) {
3655 _task = task;
3656 }
3657 };
3659 // MT Concurrent Marking Task
3660 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3661 CMSCollector* _collector;
3662 YieldingFlexibleWorkGang* _workers; // the whole gang
3663 int _n_workers; // requested/desired # workers
3664 bool _asynch;
3665 bool _result;
3666 CompactibleFreeListSpace* _cms_space;
3667 CompactibleFreeListSpace* _perm_space;
3668 HeapWord* _global_finger;
3670 // Exposed here for yielding support
3671 Mutex* const _bit_map_lock;
3673 // The per thread work queues, available here for stealing
3674 OopTaskQueueSet* _task_queues;
3675 CMSConcMarkingTerminator _term;
3677 public:
3678 CMSConcMarkingTask(CMSCollector* collector,
3679 CompactibleFreeListSpace* cms_space,
3680 CompactibleFreeListSpace* perm_space,
3681 bool asynch, int n_workers,
3682 YieldingFlexibleWorkGang* workers,
3683 OopTaskQueueSet* task_queues):
3684 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3685 _collector(collector),
3686 _cms_space(cms_space),
3687 _perm_space(perm_space),
3688 _asynch(asynch), _n_workers(n_workers), _result(true),
3689 _workers(workers), _task_queues(task_queues),
3690 _term(n_workers, task_queues, _collector, asynch),
3691 _bit_map_lock(collector->bitMapLock())
3692 {
3693 assert(n_workers <= workers->total_workers(),
3694 "Else termination won't work correctly today"); // XXX FIX ME!
3695 _requested_size = n_workers;
3696 _term.set_task(this);
3697 assert(_cms_space->bottom() < _perm_space->bottom(),
3698 "Finger incorrectly initialized below");
3699 _global_finger = _cms_space->bottom();
3700 }
3703 OopTaskQueueSet* task_queues() { return _task_queues; }
3705 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3707 HeapWord** global_finger_addr() { return &_global_finger; }
3709 CMSConcMarkingTerminator* terminator() { return &_term; }
3711 void work(int i);
3713 virtual void coordinator_yield(); // stuff done by coordinator
3714 bool result() { return _result; }
3716 void reset(HeapWord* ra) {
3717 _term.reset_for_reuse();
3718 }
3720 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3721 OopTaskQueue* work_q);
3723 private:
3724 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3725 void do_work_steal(int i);
3726 void bump_global_finger(HeapWord* f);
3727 };
3729 void CMSConcMarkingTerminator::yield() {
3730 if (ConcurrentMarkSweepThread::should_yield() &&
3731 !_collector->foregroundGCIsActive() &&
3732 _yield) {
3733 _task->yield();
3734 } else {
3735 ParallelTaskTerminator::yield();
3736 }
3737 }
3739 ////////////////////////////////////////////////////////////////
3740 // Concurrent Marking Algorithm Sketch
3741 ////////////////////////////////////////////////////////////////
3742 // Until all tasks exhausted (both spaces):
3743 // -- claim next available chunk
3744 // -- bump global finger via CAS
3745 // -- find first object that starts in this chunk
3746 // and start scanning bitmap from that position
3747 // -- scan marked objects for oops
3748 // -- CAS-mark target, and if successful:
3749 // . if target oop is above global finger (volatile read)
3750 // nothing to do
3751 // . if target oop is in chunk and above local finger
3752 // then nothing to do
3753 // . else push on work-queue
3754 // -- Deal with possible overflow issues:
3755 // . local work-queue overflow causes stuff to be pushed on
3756 // global (common) overflow queue
3757 // . always first empty local work queue
3758 // . then get a batch of oops from global work queue if any
3759 // . then do work stealing
3760 // -- When all tasks claimed (both spaces)
3761 // and local work queue empty,
3762 // then in a loop do:
3763 // . check global overflow stack; steal a batch of oops and trace
3764 // . try to steal from other threads oif GOS is empty
3765 // . if neither is available, offer termination
3766 // -- Terminate and return result
3767 //
3768 void CMSConcMarkingTask::work(int i) {
3769 elapsedTimer _timer;
3770 ResourceMark rm;
3771 HandleMark hm;
3773 DEBUG_ONLY(_collector->verify_overflow_empty();)
3775 // Before we begin work, our work queue should be empty
3776 assert(work_queue(i)->size() == 0, "Expected to be empty");
3777 // Scan the bitmap covering _cms_space, tracing through grey objects.
3778 _timer.start();
3779 do_scan_and_mark(i, _cms_space);
3780 _timer.stop();
3781 if (PrintCMSStatistics != 0) {
3782 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3783 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3784 }
3786 // ... do the same for the _perm_space
3787 _timer.reset();
3788 _timer.start();
3789 do_scan_and_mark(i, _perm_space);
3790 _timer.stop();
3791 if (PrintCMSStatistics != 0) {
3792 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3793 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3794 }
3796 // ... do work stealing
3797 _timer.reset();
3798 _timer.start();
3799 do_work_steal(i);
3800 _timer.stop();
3801 if (PrintCMSStatistics != 0) {
3802 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3803 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3804 }
3805 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3806 assert(work_queue(i)->size() == 0, "Should have been emptied");
3807 // Note that under the current task protocol, the
3808 // following assertion is true even of the spaces
3809 // expanded since the completion of the concurrent
3810 // marking. XXX This will likely change under a strict
3811 // ABORT semantics.
3812 assert(_global_finger > _cms_space->end() &&
3813 _global_finger >= _perm_space->end(),
3814 "All tasks have been completed");
3815 DEBUG_ONLY(_collector->verify_overflow_empty();)
3816 }
3818 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3819 HeapWord* read = _global_finger;
3820 HeapWord* cur = read;
3821 while (f > read) {
3822 cur = read;
3823 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3824 if (cur == read) {
3825 // our cas succeeded
3826 assert(_global_finger >= f, "protocol consistency");
3827 break;
3828 }
3829 }
3830 }
3832 // This is really inefficient, and should be redone by
3833 // using (not yet available) block-read and -write interfaces to the
3834 // stack and the work_queue. XXX FIX ME !!!
3835 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3836 OopTaskQueue* work_q) {
3837 // Fast lock-free check
3838 if (ovflw_stk->length() == 0) {
3839 return false;
3840 }
3841 assert(work_q->size() == 0, "Shouldn't steal");
3842 MutexLockerEx ml(ovflw_stk->par_lock(),
3843 Mutex::_no_safepoint_check_flag);
3844 // Grab up to 1/4 the size of the work queue
3845 size_t num = MIN2((size_t)work_q->max_elems()/4,
3846 (size_t)ParGCDesiredObjsFromOverflowList);
3847 num = MIN2(num, ovflw_stk->length());
3848 for (int i = (int) num; i > 0; i--) {
3849 oop cur = ovflw_stk->pop();
3850 assert(cur != NULL, "Counted wrong?");
3851 work_q->push(cur);
3852 }
3853 return num > 0;
3854 }
3856 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3857 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3858 int n_tasks = pst->n_tasks();
3859 // We allow that there may be no tasks to do here because
3860 // we are restarting after a stack overflow.
3861 assert(pst->valid() || n_tasks == 0, "Uninitializd use?");
3862 int nth_task = 0;
3864 HeapWord* start = sp->bottom();
3865 size_t chunk_size = sp->marking_task_size();
3866 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3867 // Having claimed the nth task in this space,
3868 // compute the chunk that it corresponds to:
3869 MemRegion span = MemRegion(start + nth_task*chunk_size,
3870 start + (nth_task+1)*chunk_size);
3871 // Try and bump the global finger via a CAS;
3872 // note that we need to do the global finger bump
3873 // _before_ taking the intersection below, because
3874 // the task corresponding to that region will be
3875 // deemed done even if the used_region() expands
3876 // because of allocation -- as it almost certainly will
3877 // during start-up while the threads yield in the
3878 // closure below.
3879 HeapWord* finger = span.end();
3880 bump_global_finger(finger); // atomically
3881 // There are null tasks here corresponding to chunks
3882 // beyond the "top" address of the space.
3883 span = span.intersection(sp->used_region());
3884 if (!span.is_empty()) { // Non-null task
3885 // We want to skip the first object because
3886 // the protocol is to scan any object in its entirety
3887 // that _starts_ in this span; a fortiori, any
3888 // object starting in an earlier span is scanned
3889 // as part of an earlier claimed task.
3890 // Below we use the "careful" version of block_start
3891 // so we do not try to navigate uninitialized objects.
3892 HeapWord* prev_obj = sp->block_start_careful(span.start());
3893 // Below we use a variant of block_size that uses the
3894 // Printezis bits to avoid waiting for allocated
3895 // objects to become initialized/parsable.
3896 while (prev_obj < span.start()) {
3897 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3898 if (sz > 0) {
3899 prev_obj += sz;
3900 } else {
3901 // In this case we may end up doing a bit of redundant
3902 // scanning, but that appears unavoidable, short of
3903 // locking the free list locks; see bug 6324141.
3904 break;
3905 }
3906 }
3907 if (prev_obj < span.end()) {
3908 MemRegion my_span = MemRegion(prev_obj, span.end());
3909 // Do the marking work within a non-empty span --
3910 // the last argument to the constructor indicates whether the
3911 // iteration should be incremental with periodic yields.
3912 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3913 &_collector->_markBitMap,
3914 work_queue(i),
3915 &_collector->_markStack,
3916 &_collector->_revisitStack,
3917 _asynch);
3918 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3919 } // else nothing to do for this task
3920 } // else nothing to do for this task
3921 }
3922 // We'd be tempted to assert here that since there are no
3923 // more tasks left to claim in this space, the global_finger
3924 // must exceed space->top() and a fortiori space->end(). However,
3925 // that would not quite be correct because the bumping of
3926 // global_finger occurs strictly after the claiming of a task,
3927 // so by the time we reach here the global finger may not yet
3928 // have been bumped up by the thread that claimed the last
3929 // task.
3930 pst->all_tasks_completed();
3931 }
3933 class Par_ConcMarkingClosure: public OopClosure {
3934 private:
3935 CMSCollector* _collector;
3936 MemRegion _span;
3937 CMSBitMap* _bit_map;
3938 CMSMarkStack* _overflow_stack;
3939 CMSMarkStack* _revisit_stack; // XXXXXX Check proper use
3940 OopTaskQueue* _work_queue;
3941 protected:
3942 DO_OOP_WORK_DEFN
3943 public:
3944 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
3945 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3946 _collector(collector),
3947 _span(_collector->_span),
3948 _work_queue(work_queue),
3949 _bit_map(bit_map),
3950 _overflow_stack(overflow_stack) { } // need to initialize revisit stack etc.
3951 virtual void do_oop(oop* p);
3952 virtual void do_oop(narrowOop* p);
3953 void trim_queue(size_t max);
3954 void handle_stack_overflow(HeapWord* lost);
3955 };
3957 // Grey object rescan during work stealing phase --
3958 // the salient assumption here is that stolen oops must
3959 // always be initialized, so we do not need to check for
3960 // uninitialized objects before scanning here.
3961 void Par_ConcMarkingClosure::do_oop(oop obj) {
3962 assert(obj->is_oop_or_null(), "expected an oop or NULL");
3963 HeapWord* addr = (HeapWord*)obj;
3964 // Check if oop points into the CMS generation
3965 // and is not marked
3966 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3967 // a white object ...
3968 // If we manage to "claim" the object, by being the
3969 // first thread to mark it, then we push it on our
3970 // marking stack
3971 if (_bit_map->par_mark(addr)) { // ... now grey
3972 // push on work queue (grey set)
3973 bool simulate_overflow = false;
3974 NOT_PRODUCT(
3975 if (CMSMarkStackOverflowALot &&
3976 _collector->simulate_overflow()) {
3977 // simulate a stack overflow
3978 simulate_overflow = true;
3979 }
3980 )
3981 if (simulate_overflow ||
3982 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3983 // stack overflow
3984 if (PrintCMSStatistics != 0) {
3985 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3986 SIZE_FORMAT, _overflow_stack->capacity());
3987 }
3988 // We cannot assert that the overflow stack is full because
3989 // it may have been emptied since.
3990 assert(simulate_overflow ||
3991 _work_queue->size() == _work_queue->max_elems(),
3992 "Else push should have succeeded");
3993 handle_stack_overflow(addr);
3994 }
3995 } // Else, some other thread got there first
3996 }
3997 }
3999 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4000 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4002 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4003 while (_work_queue->size() > max) {
4004 oop new_oop;
4005 if (_work_queue->pop_local(new_oop)) {
4006 assert(new_oop->is_oop(), "Should be an oop");
4007 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4008 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4009 assert(new_oop->is_parsable(), "Should be parsable");
4010 new_oop->oop_iterate(this); // do_oop() above
4011 }
4012 }
4013 }
4015 // Upon stack overflow, we discard (part of) the stack,
4016 // remembering the least address amongst those discarded
4017 // in CMSCollector's _restart_address.
4018 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4019 // We need to do this under a mutex to prevent other
4020 // workers from interfering with the expansion below.
4021 MutexLockerEx ml(_overflow_stack->par_lock(),
4022 Mutex::_no_safepoint_check_flag);
4023 // Remember the least grey address discarded
4024 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4025 _collector->lower_restart_addr(ra);
4026 _overflow_stack->reset(); // discard stack contents
4027 _overflow_stack->expand(); // expand the stack if possible
4028 }
4031 void CMSConcMarkingTask::do_work_steal(int i) {
4032 OopTaskQueue* work_q = work_queue(i);
4033 oop obj_to_scan;
4034 CMSBitMap* bm = &(_collector->_markBitMap);
4035 CMSMarkStack* ovflw = &(_collector->_markStack);
4036 int* seed = _collector->hash_seed(i);
4037 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw);
4038 while (true) {
4039 cl.trim_queue(0);
4040 assert(work_q->size() == 0, "Should have been emptied above");
4041 if (get_work_from_overflow_stack(ovflw, work_q)) {
4042 // Can't assert below because the work obtained from the
4043 // overflow stack may already have been stolen from us.
4044 // assert(work_q->size() > 0, "Work from overflow stack");
4045 continue;
4046 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4047 assert(obj_to_scan->is_oop(), "Should be an oop");
4048 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4049 obj_to_scan->oop_iterate(&cl);
4050 } else if (terminator()->offer_termination()) {
4051 assert(work_q->size() == 0, "Impossible!");
4052 break;
4053 }
4054 }
4055 }
4057 // This is run by the CMS (coordinator) thread.
4058 void CMSConcMarkingTask::coordinator_yield() {
4059 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4060 "CMS thread should hold CMS token");
4062 // First give up the locks, then yield, then re-lock
4063 // We should probably use a constructor/destructor idiom to
4064 // do this unlock/lock or modify the MutexUnlocker class to
4065 // serve our purpose. XXX
4066 assert_lock_strong(_bit_map_lock);
4067 _bit_map_lock->unlock();
4068 ConcurrentMarkSweepThread::desynchronize(true);
4069 ConcurrentMarkSweepThread::acknowledge_yield_request();
4070 _collector->stopTimer();
4071 if (PrintCMSStatistics != 0) {
4072 _collector->incrementYields();
4073 }
4074 _collector->icms_wait();
4076 // It is possible for whichever thread initiated the yield request
4077 // not to get a chance to wake up and take the bitmap lock between
4078 // this thread releasing it and reacquiring it. So, while the
4079 // should_yield() flag is on, let's sleep for a bit to give the
4080 // other thread a chance to wake up. The limit imposed on the number
4081 // of iterations is defensive, to avoid any unforseen circumstances
4082 // putting us into an infinite loop. Since it's always been this
4083 // (coordinator_yield()) method that was observed to cause the
4084 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4085 // which is by default non-zero. For the other seven methods that
4086 // also perform the yield operation, as are using a different
4087 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4088 // can enable the sleeping for those methods too, if necessary.
4089 // See 6442774.
4090 //
4091 // We really need to reconsider the synchronization between the GC
4092 // thread and the yield-requesting threads in the future and we
4093 // should really use wait/notify, which is the recommended
4094 // way of doing this type of interaction. Additionally, we should
4095 // consolidate the eight methods that do the yield operation and they
4096 // are almost identical into one for better maintenability and
4097 // readability. See 6445193.
4098 //
4099 // Tony 2006.06.29
4100 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4101 ConcurrentMarkSweepThread::should_yield() &&
4102 !CMSCollector::foregroundGCIsActive(); ++i) {
4103 os::sleep(Thread::current(), 1, false);
4104 ConcurrentMarkSweepThread::acknowledge_yield_request();
4105 }
4107 ConcurrentMarkSweepThread::synchronize(true);
4108 _bit_map_lock->lock_without_safepoint_check();
4109 _collector->startTimer();
4110 }
4112 bool CMSCollector::do_marking_mt(bool asynch) {
4113 assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition");
4114 // In the future this would be determined ergonomically, based
4115 // on #cpu's, # active mutator threads (and load), and mutation rate.
4116 int num_workers = ParallelCMSThreads;
4118 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4119 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4121 CMSConcMarkingTask tsk(this, cms_space, perm_space,
4122 asynch, num_workers /* number requested XXX */,
4123 conc_workers(), task_queues());
4125 // Since the actual number of workers we get may be different
4126 // from the number we requested above, do we need to do anything different
4127 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4128 // class?? XXX
4129 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4130 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4132 // Refs discovery is already non-atomic.
4133 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4134 // Mutate the Refs discovery so it is MT during the
4135 // multi-threaded marking phase.
4136 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4138 conc_workers()->start_task(&tsk);
4139 while (tsk.yielded()) {
4140 tsk.coordinator_yield();
4141 conc_workers()->continue_task(&tsk);
4142 }
4143 // If the task was aborted, _restart_addr will be non-NULL
4144 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4145 while (_restart_addr != NULL) {
4146 // XXX For now we do not make use of ABORTED state and have not
4147 // yet implemented the right abort semantics (even in the original
4148 // single-threaded CMS case). That needs some more investigation
4149 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4150 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4151 // If _restart_addr is non-NULL, a marking stack overflow
4152 // occured; we need to do a fresh marking iteration from the
4153 // indicated restart address.
4154 if (_foregroundGCIsActive && asynch) {
4155 // We may be running into repeated stack overflows, having
4156 // reached the limit of the stack size, while making very
4157 // slow forward progress. It may be best to bail out and
4158 // let the foreground collector do its job.
4159 // Clear _restart_addr, so that foreground GC
4160 // works from scratch. This avoids the headache of
4161 // a "rescan" which would otherwise be needed because
4162 // of the dirty mod union table & card table.
4163 _restart_addr = NULL;
4164 return false;
4165 }
4166 // Adjust the task to restart from _restart_addr
4167 tsk.reset(_restart_addr);
4168 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4169 _restart_addr);
4170 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4171 _restart_addr);
4172 _restart_addr = NULL;
4173 // Get the workers going again
4174 conc_workers()->start_task(&tsk);
4175 while (tsk.yielded()) {
4176 tsk.coordinator_yield();
4177 conc_workers()->continue_task(&tsk);
4178 }
4179 }
4180 assert(tsk.completed(), "Inconsistency");
4181 assert(tsk.result() == true, "Inconsistency");
4182 return true;
4183 }
4185 bool CMSCollector::do_marking_st(bool asynch) {
4186 ResourceMark rm;
4187 HandleMark hm;
4189 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4190 &_markStack, &_revisitStack, CMSYield && asynch);
4191 // the last argument to iterate indicates whether the iteration
4192 // should be incremental with periodic yields.
4193 _markBitMap.iterate(&markFromRootsClosure);
4194 // If _restart_addr is non-NULL, a marking stack overflow
4195 // occured; we need to do a fresh iteration from the
4196 // indicated restart address.
4197 while (_restart_addr != NULL) {
4198 if (_foregroundGCIsActive && asynch) {
4199 // We may be running into repeated stack overflows, having
4200 // reached the limit of the stack size, while making very
4201 // slow forward progress. It may be best to bail out and
4202 // let the foreground collector do its job.
4203 // Clear _restart_addr, so that foreground GC
4204 // works from scratch. This avoids the headache of
4205 // a "rescan" which would otherwise be needed because
4206 // of the dirty mod union table & card table.
4207 _restart_addr = NULL;
4208 return false; // indicating failure to complete marking
4209 }
4210 // Deal with stack overflow:
4211 // we restart marking from _restart_addr
4212 HeapWord* ra = _restart_addr;
4213 markFromRootsClosure.reset(ra);
4214 _restart_addr = NULL;
4215 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4216 }
4217 return true;
4218 }
4220 void CMSCollector::preclean() {
4221 check_correct_thread_executing();
4222 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4223 verify_work_stacks_empty();
4224 verify_overflow_empty();
4225 _abort_preclean = false;
4226 if (CMSPrecleaningEnabled) {
4227 _eden_chunk_index = 0;
4228 size_t used = get_eden_used();
4229 size_t capacity = get_eden_capacity();
4230 // Don't start sampling unless we will get sufficiently
4231 // many samples.
4232 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4233 * CMSScheduleRemarkEdenPenetration)) {
4234 _start_sampling = true;
4235 } else {
4236 _start_sampling = false;
4237 }
4238 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4239 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4240 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4241 }
4242 CMSTokenSync x(true); // is cms thread
4243 if (CMSPrecleaningEnabled) {
4244 sample_eden();
4245 _collectorState = AbortablePreclean;
4246 } else {
4247 _collectorState = FinalMarking;
4248 }
4249 verify_work_stacks_empty();
4250 verify_overflow_empty();
4251 }
4253 // Try and schedule the remark such that young gen
4254 // occupancy is CMSScheduleRemarkEdenPenetration %.
4255 void CMSCollector::abortable_preclean() {
4256 check_correct_thread_executing();
4257 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4258 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4260 // If Eden's current occupancy is below this threshold,
4261 // immediately schedule the remark; else preclean
4262 // past the next scavenge in an effort to
4263 // schedule the pause as described avove. By choosing
4264 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4265 // we will never do an actual abortable preclean cycle.
4266 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4267 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4268 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4269 // We need more smarts in the abortable preclean
4270 // loop below to deal with cases where allocation
4271 // in young gen is very very slow, and our precleaning
4272 // is running a losing race against a horde of
4273 // mutators intent on flooding us with CMS updates
4274 // (dirty cards).
4275 // One, admittedly dumb, strategy is to give up
4276 // after a certain number of abortable precleaning loops
4277 // or after a certain maximum time. We want to make
4278 // this smarter in the next iteration.
4279 // XXX FIX ME!!! YSR
4280 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4281 while (!(should_abort_preclean() ||
4282 ConcurrentMarkSweepThread::should_terminate())) {
4283 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4284 cumworkdone += workdone;
4285 loops++;
4286 // Voluntarily terminate abortable preclean phase if we have
4287 // been at it for too long.
4288 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4289 loops >= CMSMaxAbortablePrecleanLoops) {
4290 if (PrintGCDetails) {
4291 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4292 }
4293 break;
4294 }
4295 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4296 if (PrintGCDetails) {
4297 gclog_or_tty->print(" CMS: abort preclean due to time ");
4298 }
4299 break;
4300 }
4301 // If we are doing little work each iteration, we should
4302 // take a short break.
4303 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4304 // Sleep for some time, waiting for work to accumulate
4305 stopTimer();
4306 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4307 startTimer();
4308 waited++;
4309 }
4310 }
4311 if (PrintCMSStatistics > 0) {
4312 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4313 loops, waited, cumworkdone);
4314 }
4315 }
4316 CMSTokenSync x(true); // is cms thread
4317 if (_collectorState != Idling) {
4318 assert(_collectorState == AbortablePreclean,
4319 "Spontaneous state transition?");
4320 _collectorState = FinalMarking;
4321 } // Else, a foreground collection completed this CMS cycle.
4322 return;
4323 }
4325 // Respond to an Eden sampling opportunity
4326 void CMSCollector::sample_eden() {
4327 // Make sure a young gc cannot sneak in between our
4328 // reading and recording of a sample.
4329 assert(Thread::current()->is_ConcurrentGC_thread(),
4330 "Only the cms thread may collect Eden samples");
4331 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4332 "Should collect samples while holding CMS token");
4333 if (!_start_sampling) {
4334 return;
4335 }
4336 if (_eden_chunk_array) {
4337 if (_eden_chunk_index < _eden_chunk_capacity) {
4338 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4339 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4340 "Unexpected state of Eden");
4341 // We'd like to check that what we just sampled is an oop-start address;
4342 // however, we cannot do that here since the object may not yet have been
4343 // initialized. So we'll instead do the check when we _use_ this sample
4344 // later.
4345 if (_eden_chunk_index == 0 ||
4346 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4347 _eden_chunk_array[_eden_chunk_index-1])
4348 >= CMSSamplingGrain)) {
4349 _eden_chunk_index++; // commit sample
4350 }
4351 }
4352 }
4353 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4354 size_t used = get_eden_used();
4355 size_t capacity = get_eden_capacity();
4356 assert(used <= capacity, "Unexpected state of Eden");
4357 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4358 _abort_preclean = true;
4359 }
4360 }
4361 }
4364 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4365 assert(_collectorState == Precleaning ||
4366 _collectorState == AbortablePreclean, "incorrect state");
4367 ResourceMark rm;
4368 HandleMark hm;
4369 // Do one pass of scrubbing the discovered reference lists
4370 // to remove any reference objects with strongly-reachable
4371 // referents.
4372 if (clean_refs) {
4373 ReferenceProcessor* rp = ref_processor();
4374 CMSPrecleanRefsYieldClosure yield_cl(this);
4375 assert(rp->span().equals(_span), "Spans should be equal");
4376 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4377 &_markStack);
4378 CMSDrainMarkingStackClosure complete_trace(this,
4379 _span, &_markBitMap, &_markStack,
4380 &keep_alive);
4382 // We don't want this step to interfere with a young
4383 // collection because we don't want to take CPU
4384 // or memory bandwidth away from the young GC threads
4385 // (which may be as many as there are CPUs).
4386 // Note that we don't need to protect ourselves from
4387 // interference with mutators because they can't
4388 // manipulate the discovered reference lists nor affect
4389 // the computed reachability of the referents, the
4390 // only properties manipulated by the precleaning
4391 // of these reference lists.
4392 stopTimer();
4393 CMSTokenSyncWithLocks x(true /* is cms thread */,
4394 bitMapLock());
4395 startTimer();
4396 sample_eden();
4397 // The following will yield to allow foreground
4398 // collection to proceed promptly. XXX YSR:
4399 // The code in this method may need further
4400 // tweaking for better performance and some restructuring
4401 // for cleaner interfaces.
4402 rp->preclean_discovered_references(
4403 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4404 &yield_cl);
4405 }
4407 if (clean_survivor) { // preclean the active survivor space(s)
4408 assert(_young_gen->kind() == Generation::DefNew ||
4409 _young_gen->kind() == Generation::ParNew ||
4410 _young_gen->kind() == Generation::ASParNew,
4411 "incorrect type for cast");
4412 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4413 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4414 &_markBitMap, &_modUnionTable,
4415 &_markStack, &_revisitStack,
4416 true /* precleaning phase */);
4417 stopTimer();
4418 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4419 bitMapLock());
4420 startTimer();
4421 unsigned int before_count =
4422 GenCollectedHeap::heap()->total_collections();
4423 SurvivorSpacePrecleanClosure
4424 sss_cl(this, _span, &_markBitMap, &_markStack,
4425 &pam_cl, before_count, CMSYield);
4426 dng->from()->object_iterate_careful(&sss_cl);
4427 dng->to()->object_iterate_careful(&sss_cl);
4428 }
4429 MarkRefsIntoAndScanClosure
4430 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4431 &_markStack, &_revisitStack, this, CMSYield,
4432 true /* precleaning phase */);
4433 // CAUTION: The following closure has persistent state that may need to
4434 // be reset upon a decrease in the sequence of addresses it
4435 // processes.
4436 ScanMarkedObjectsAgainCarefullyClosure
4437 smoac_cl(this, _span,
4438 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4440 // Preclean dirty cards in ModUnionTable and CardTable using
4441 // appropriate convergence criterion;
4442 // repeat CMSPrecleanIter times unless we find that
4443 // we are losing.
4444 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4445 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4446 "Bad convergence multiplier");
4447 assert(CMSPrecleanThreshold >= 100,
4448 "Unreasonably low CMSPrecleanThreshold");
4450 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4451 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4452 numIter < CMSPrecleanIter;
4453 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4454 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4455 if (CMSPermGenPrecleaningEnabled) {
4456 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4457 }
4458 if (Verbose && PrintGCDetails) {
4459 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4460 }
4461 // Either there are very few dirty cards, so re-mark
4462 // pause will be small anyway, or our pre-cleaning isn't
4463 // that much faster than the rate at which cards are being
4464 // dirtied, so we might as well stop and re-mark since
4465 // precleaning won't improve our re-mark time by much.
4466 if (curNumCards <= CMSPrecleanThreshold ||
4467 (numIter > 0 &&
4468 (curNumCards * CMSPrecleanDenominator >
4469 lastNumCards * CMSPrecleanNumerator))) {
4470 numIter++;
4471 cumNumCards += curNumCards;
4472 break;
4473 }
4474 }
4475 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4476 if (CMSPermGenPrecleaningEnabled) {
4477 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4478 }
4479 cumNumCards += curNumCards;
4480 if (PrintGCDetails && PrintCMSStatistics != 0) {
4481 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4482 curNumCards, cumNumCards, numIter);
4483 }
4484 return cumNumCards; // as a measure of useful work done
4485 }
4487 // PRECLEANING NOTES:
4488 // Precleaning involves:
4489 // . reading the bits of the modUnionTable and clearing the set bits.
4490 // . For the cards corresponding to the set bits, we scan the
4491 // objects on those cards. This means we need the free_list_lock
4492 // so that we can safely iterate over the CMS space when scanning
4493 // for oops.
4494 // . When we scan the objects, we'll be both reading and setting
4495 // marks in the marking bit map, so we'll need the marking bit map.
4496 // . For protecting _collector_state transitions, we take the CGC_lock.
4497 // Note that any races in the reading of of card table entries by the
4498 // CMS thread on the one hand and the clearing of those entries by the
4499 // VM thread or the setting of those entries by the mutator threads on the
4500 // other are quite benign. However, for efficiency it makes sense to keep
4501 // the VM thread from racing with the CMS thread while the latter is
4502 // dirty card info to the modUnionTable. We therefore also use the
4503 // CGC_lock to protect the reading of the card table and the mod union
4504 // table by the CM thread.
4505 // . We run concurrently with mutator updates, so scanning
4506 // needs to be done carefully -- we should not try to scan
4507 // potentially uninitialized objects.
4508 //
4509 // Locking strategy: While holding the CGC_lock, we scan over and
4510 // reset a maximal dirty range of the mod union / card tables, then lock
4511 // the free_list_lock and bitmap lock to do a full marking, then
4512 // release these locks; and repeat the cycle. This allows for a
4513 // certain amount of fairness in the sharing of these locks between
4514 // the CMS collector on the one hand, and the VM thread and the
4515 // mutators on the other.
4517 // NOTE: preclean_mod_union_table() and preclean_card_table()
4518 // further below are largely identical; if you need to modify
4519 // one of these methods, please check the other method too.
4521 size_t CMSCollector::preclean_mod_union_table(
4522 ConcurrentMarkSweepGeneration* gen,
4523 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4524 verify_work_stacks_empty();
4525 verify_overflow_empty();
4527 // strategy: starting with the first card, accumulate contiguous
4528 // ranges of dirty cards; clear these cards, then scan the region
4529 // covered by these cards.
4531 // Since all of the MUT is committed ahead, we can just use
4532 // that, in case the generations expand while we are precleaning.
4533 // It might also be fine to just use the committed part of the
4534 // generation, but we might potentially miss cards when the
4535 // generation is rapidly expanding while we are in the midst
4536 // of precleaning.
4537 HeapWord* startAddr = gen->reserved().start();
4538 HeapWord* endAddr = gen->reserved().end();
4540 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4542 size_t numDirtyCards, cumNumDirtyCards;
4543 HeapWord *nextAddr, *lastAddr;
4544 for (cumNumDirtyCards = numDirtyCards = 0,
4545 nextAddr = lastAddr = startAddr;
4546 nextAddr < endAddr;
4547 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4549 ResourceMark rm;
4550 HandleMark hm;
4552 MemRegion dirtyRegion;
4553 {
4554 stopTimer();
4555 CMSTokenSync ts(true);
4556 startTimer();
4557 sample_eden();
4558 // Get dirty region starting at nextOffset (inclusive),
4559 // simultaneously clearing it.
4560 dirtyRegion =
4561 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4562 assert(dirtyRegion.start() >= nextAddr,
4563 "returned region inconsistent?");
4564 }
4565 // Remember where the next search should begin.
4566 // The returned region (if non-empty) is a right open interval,
4567 // so lastOffset is obtained from the right end of that
4568 // interval.
4569 lastAddr = dirtyRegion.end();
4570 // Should do something more transparent and less hacky XXX
4571 numDirtyCards =
4572 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4574 // We'll scan the cards in the dirty region (with periodic
4575 // yields for foreground GC as needed).
4576 if (!dirtyRegion.is_empty()) {
4577 assert(numDirtyCards > 0, "consistency check");
4578 HeapWord* stop_point = NULL;
4579 {
4580 stopTimer();
4581 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4582 bitMapLock());
4583 startTimer();
4584 verify_work_stacks_empty();
4585 verify_overflow_empty();
4586 sample_eden();
4587 stop_point =
4588 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4589 }
4590 if (stop_point != NULL) {
4591 // The careful iteration stopped early either because it found an
4592 // uninitialized object, or because we were in the midst of an
4593 // "abortable preclean", which should now be aborted. Redirty
4594 // the bits corresponding to the partially-scanned or unscanned
4595 // cards. We'll either restart at the next block boundary or
4596 // abort the preclean.
4597 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4598 (_collectorState == AbortablePreclean && should_abort_preclean()),
4599 "Unparsable objects should only be in perm gen.");
4601 stopTimer();
4602 CMSTokenSyncWithLocks ts(true, bitMapLock());
4603 startTimer();
4604 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4605 if (should_abort_preclean()) {
4606 break; // out of preclean loop
4607 } else {
4608 // Compute the next address at which preclean should pick up;
4609 // might need bitMapLock in order to read P-bits.
4610 lastAddr = next_card_start_after_block(stop_point);
4611 }
4612 }
4613 } else {
4614 assert(lastAddr == endAddr, "consistency check");
4615 assert(numDirtyCards == 0, "consistency check");
4616 break;
4617 }
4618 }
4619 verify_work_stacks_empty();
4620 verify_overflow_empty();
4621 return cumNumDirtyCards;
4622 }
4624 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4625 // below are largely identical; if you need to modify
4626 // one of these methods, please check the other method too.
4628 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4629 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4630 // strategy: it's similar to precleamModUnionTable above, in that
4631 // we accumulate contiguous ranges of dirty cards, mark these cards
4632 // precleaned, then scan the region covered by these cards.
4633 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4634 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4636 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4638 size_t numDirtyCards, cumNumDirtyCards;
4639 HeapWord *lastAddr, *nextAddr;
4641 for (cumNumDirtyCards = numDirtyCards = 0,
4642 nextAddr = lastAddr = startAddr;
4643 nextAddr < endAddr;
4644 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4646 ResourceMark rm;
4647 HandleMark hm;
4649 MemRegion dirtyRegion;
4650 {
4651 // See comments in "Precleaning notes" above on why we
4652 // do this locking. XXX Could the locking overheads be
4653 // too high when dirty cards are sparse? [I don't think so.]
4654 stopTimer();
4655 CMSTokenSync x(true); // is cms thread
4656 startTimer();
4657 sample_eden();
4658 // Get and clear dirty region from card table
4659 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4660 MemRegion(nextAddr, endAddr),
4661 true,
4662 CardTableModRefBS::precleaned_card_val());
4664 assert(dirtyRegion.start() >= nextAddr,
4665 "returned region inconsistent?");
4666 }
4667 lastAddr = dirtyRegion.end();
4668 numDirtyCards =
4669 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4671 if (!dirtyRegion.is_empty()) {
4672 stopTimer();
4673 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4674 startTimer();
4675 sample_eden();
4676 verify_work_stacks_empty();
4677 verify_overflow_empty();
4678 HeapWord* stop_point =
4679 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4680 if (stop_point != NULL) {
4681 // The careful iteration stopped early because it found an
4682 // uninitialized object. Redirty the bits corresponding to the
4683 // partially-scanned or unscanned cards, and start again at the
4684 // next block boundary.
4685 assert(CMSPermGenPrecleaningEnabled ||
4686 (_collectorState == AbortablePreclean && should_abort_preclean()),
4687 "Unparsable objects should only be in perm gen.");
4688 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4689 if (should_abort_preclean()) {
4690 break; // out of preclean loop
4691 } else {
4692 // Compute the next address at which preclean should pick up.
4693 lastAddr = next_card_start_after_block(stop_point);
4694 }
4695 }
4696 } else {
4697 break;
4698 }
4699 }
4700 verify_work_stacks_empty();
4701 verify_overflow_empty();
4702 return cumNumDirtyCards;
4703 }
4705 void CMSCollector::checkpointRootsFinal(bool asynch,
4706 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4707 assert(_collectorState == FinalMarking, "incorrect state transition?");
4708 check_correct_thread_executing();
4709 // world is stopped at this checkpoint
4710 assert(SafepointSynchronize::is_at_safepoint(),
4711 "world should be stopped");
4712 verify_work_stacks_empty();
4713 verify_overflow_empty();
4715 SpecializationStats::clear();
4716 if (PrintGCDetails) {
4717 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4718 _young_gen->used() / K,
4719 _young_gen->capacity() / K);
4720 }
4721 if (asynch) {
4722 if (CMSScavengeBeforeRemark) {
4723 GenCollectedHeap* gch = GenCollectedHeap::heap();
4724 // Temporarily set flag to false, GCH->do_collection will
4725 // expect it to be false and set to true
4726 FlagSetting fl(gch->_is_gc_active, false);
4727 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4728 PrintGCDetails && Verbose, true, gclog_or_tty);)
4729 int level = _cmsGen->level() - 1;
4730 if (level >= 0) {
4731 gch->do_collection(true, // full (i.e. force, see below)
4732 false, // !clear_all_soft_refs
4733 0, // size
4734 false, // is_tlab
4735 level // max_level
4736 );
4737 }
4738 }
4739 FreelistLocker x(this);
4740 MutexLockerEx y(bitMapLock(),
4741 Mutex::_no_safepoint_check_flag);
4742 assert(!init_mark_was_synchronous, "but that's impossible!");
4743 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4744 } else {
4745 // already have all the locks
4746 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4747 init_mark_was_synchronous);
4748 }
4749 verify_work_stacks_empty();
4750 verify_overflow_empty();
4751 SpecializationStats::print();
4752 }
4754 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4755 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4757 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4759 assert(haveFreelistLocks(), "must have free list locks");
4760 assert_lock_strong(bitMapLock());
4762 if (UseAdaptiveSizePolicy) {
4763 size_policy()->checkpoint_roots_final_begin();
4764 }
4766 ResourceMark rm;
4767 HandleMark hm;
4769 GenCollectedHeap* gch = GenCollectedHeap::heap();
4771 if (should_unload_classes()) {
4772 CodeCache::gc_prologue();
4773 }
4774 assert(haveFreelistLocks(), "must have free list locks");
4775 assert_lock_strong(bitMapLock());
4777 if (!init_mark_was_synchronous) {
4778 // We might assume that we need not fill TLAB's when
4779 // CMSScavengeBeforeRemark is set, because we may have just done
4780 // a scavenge which would have filled all TLAB's -- and besides
4781 // Eden would be empty. This however may not always be the case --
4782 // for instance although we asked for a scavenge, it may not have
4783 // happened because of a JNI critical section. We probably need
4784 // a policy for deciding whether we can in that case wait until
4785 // the critical section releases and then do the remark following
4786 // the scavenge, and skip it here. In the absence of that policy,
4787 // or of an indication of whether the scavenge did indeed occur,
4788 // we cannot rely on TLAB's having been filled and must do
4789 // so here just in case a scavenge did not happen.
4790 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4791 // Update the saved marks which may affect the root scans.
4792 gch->save_marks();
4794 {
4795 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4797 // Note on the role of the mod union table:
4798 // Since the marker in "markFromRoots" marks concurrently with
4799 // mutators, it is possible for some reachable objects not to have been
4800 // scanned. For instance, an only reference to an object A was
4801 // placed in object B after the marker scanned B. Unless B is rescanned,
4802 // A would be collected. Such updates to references in marked objects
4803 // are detected via the mod union table which is the set of all cards
4804 // dirtied since the first checkpoint in this GC cycle and prior to
4805 // the most recent young generation GC, minus those cleaned up by the
4806 // concurrent precleaning.
4807 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
4808 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4809 do_remark_parallel();
4810 } else {
4811 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4812 gclog_or_tty);
4813 do_remark_non_parallel();
4814 }
4815 }
4816 } else {
4817 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4818 // The initial mark was stop-world, so there's no rescanning to
4819 // do; go straight on to the next step below.
4820 }
4821 verify_work_stacks_empty();
4822 verify_overflow_empty();
4824 {
4825 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4826 refProcessingWork(asynch, clear_all_soft_refs);
4827 }
4828 verify_work_stacks_empty();
4829 verify_overflow_empty();
4831 if (should_unload_classes()) {
4832 CodeCache::gc_epilogue();
4833 }
4835 // If we encountered any (marking stack / work queue) overflow
4836 // events during the current CMS cycle, take appropriate
4837 // remedial measures, where possible, so as to try and avoid
4838 // recurrence of that condition.
4839 assert(_markStack.isEmpty(), "No grey objects");
4840 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4841 _ser_kac_ovflw;
4842 if (ser_ovflw > 0) {
4843 if (PrintCMSStatistics != 0) {
4844 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4845 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4846 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4847 _ser_kac_ovflw);
4848 }
4849 _markStack.expand();
4850 _ser_pmc_remark_ovflw = 0;
4851 _ser_pmc_preclean_ovflw = 0;
4852 _ser_kac_ovflw = 0;
4853 }
4854 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4855 if (PrintCMSStatistics != 0) {
4856 gclog_or_tty->print_cr("Work queue overflow (benign) "
4857 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4858 _par_pmc_remark_ovflw, _par_kac_ovflw);
4859 }
4860 _par_pmc_remark_ovflw = 0;
4861 _par_kac_ovflw = 0;
4862 }
4863 if (PrintCMSStatistics != 0) {
4864 if (_markStack._hit_limit > 0) {
4865 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4866 _markStack._hit_limit);
4867 }
4868 if (_markStack._failed_double > 0) {
4869 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4870 " current capacity "SIZE_FORMAT,
4871 _markStack._failed_double,
4872 _markStack.capacity());
4873 }
4874 }
4875 _markStack._hit_limit = 0;
4876 _markStack._failed_double = 0;
4878 if ((VerifyAfterGC || VerifyDuringGC) &&
4879 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4880 verify_after_remark();
4881 }
4883 // Change under the freelistLocks.
4884 _collectorState = Sweeping;
4885 // Call isAllClear() under bitMapLock
4886 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
4887 " final marking");
4888 if (UseAdaptiveSizePolicy) {
4889 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
4890 }
4891 }
4893 // Parallel remark task
4894 class CMSParRemarkTask: public AbstractGangTask {
4895 CMSCollector* _collector;
4896 WorkGang* _workers;
4897 int _n_workers;
4898 CompactibleFreeListSpace* _cms_space;
4899 CompactibleFreeListSpace* _perm_space;
4901 // The per-thread work queues, available here for stealing.
4902 OopTaskQueueSet* _task_queues;
4903 ParallelTaskTerminator _term;
4905 public:
4906 CMSParRemarkTask(CMSCollector* collector,
4907 CompactibleFreeListSpace* cms_space,
4908 CompactibleFreeListSpace* perm_space,
4909 int n_workers, WorkGang* workers,
4910 OopTaskQueueSet* task_queues):
4911 AbstractGangTask("Rescan roots and grey objects in parallel"),
4912 _collector(collector),
4913 _cms_space(cms_space), _perm_space(perm_space),
4914 _n_workers(n_workers),
4915 _workers(workers),
4916 _task_queues(task_queues),
4917 _term(workers->total_workers(), task_queues) { }
4919 OopTaskQueueSet* task_queues() { return _task_queues; }
4921 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4923 ParallelTaskTerminator* terminator() { return &_term; }
4925 void work(int i);
4927 private:
4928 // Work method in support of parallel rescan ... of young gen spaces
4929 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
4930 ContiguousSpace* space,
4931 HeapWord** chunk_array, size_t chunk_top);
4933 // ... of dirty cards in old space
4934 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4935 Par_MarkRefsIntoAndScanClosure* cl);
4937 // ... work stealing for the above
4938 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4939 };
4941 void CMSParRemarkTask::work(int i) {
4942 elapsedTimer _timer;
4943 ResourceMark rm;
4944 HandleMark hm;
4946 // ---------- rescan from roots --------------
4947 _timer.start();
4948 GenCollectedHeap* gch = GenCollectedHeap::heap();
4949 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4950 _collector->_span, _collector->ref_processor(),
4951 &(_collector->_markBitMap),
4952 work_queue(i), &(_collector->_revisitStack));
4954 // Rescan young gen roots first since these are likely
4955 // coarsely partitioned and may, on that account, constitute
4956 // the critical path; thus, it's best to start off that
4957 // work first.
4958 // ---------- young gen roots --------------
4959 {
4960 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
4961 EdenSpace* eden_space = dng->eden();
4962 ContiguousSpace* from_space = dng->from();
4963 ContiguousSpace* to_space = dng->to();
4965 HeapWord** eca = _collector->_eden_chunk_array;
4966 size_t ect = _collector->_eden_chunk_index;
4967 HeapWord** sca = _collector->_survivor_chunk_array;
4968 size_t sct = _collector->_survivor_chunk_index;
4970 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4971 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4973 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
4974 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
4975 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
4977 _timer.stop();
4978 if (PrintCMSStatistics != 0) {
4979 gclog_or_tty->print_cr(
4980 "Finished young gen rescan work in %dth thread: %3.3f sec",
4981 i, _timer.seconds());
4982 }
4983 }
4985 // ---------- remaining roots --------------
4986 _timer.reset();
4987 _timer.start();
4988 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
4989 false, // yg was scanned above
4990 true, // collecting perm gen
4991 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4992 NULL, &par_mrias_cl);
4993 _timer.stop();
4994 if (PrintCMSStatistics != 0) {
4995 gclog_or_tty->print_cr(
4996 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4997 i, _timer.seconds());
4998 }
5000 // ---------- rescan dirty cards ------------
5001 _timer.reset();
5002 _timer.start();
5004 // Do the rescan tasks for each of the two spaces
5005 // (cms_space and perm_space) in turn.
5006 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
5007 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
5008 _timer.stop();
5009 if (PrintCMSStatistics != 0) {
5010 gclog_or_tty->print_cr(
5011 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5012 i, _timer.seconds());
5013 }
5015 // ---------- steal work from other threads ...
5016 // ---------- ... and drain overflow list.
5017 _timer.reset();
5018 _timer.start();
5019 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5020 _timer.stop();
5021 if (PrintCMSStatistics != 0) {
5022 gclog_or_tty->print_cr(
5023 "Finished work stealing in %dth thread: %3.3f sec",
5024 i, _timer.seconds());
5025 }
5026 }
5028 void
5029 CMSParRemarkTask::do_young_space_rescan(int i,
5030 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5031 HeapWord** chunk_array, size_t chunk_top) {
5032 // Until all tasks completed:
5033 // . claim an unclaimed task
5034 // . compute region boundaries corresponding to task claimed
5035 // using chunk_array
5036 // . par_oop_iterate(cl) over that region
5038 ResourceMark rm;
5039 HandleMark hm;
5041 SequentialSubTasksDone* pst = space->par_seq_tasks();
5042 assert(pst->valid(), "Uninitialized use?");
5044 int nth_task = 0;
5045 int n_tasks = pst->n_tasks();
5047 HeapWord *start, *end;
5048 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5049 // We claimed task # nth_task; compute its boundaries.
5050 if (chunk_top == 0) { // no samples were taken
5051 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5052 start = space->bottom();
5053 end = space->top();
5054 } else if (nth_task == 0) {
5055 start = space->bottom();
5056 end = chunk_array[nth_task];
5057 } else if (nth_task < (jint)chunk_top) {
5058 assert(nth_task >= 1, "Control point invariant");
5059 start = chunk_array[nth_task - 1];
5060 end = chunk_array[nth_task];
5061 } else {
5062 assert(nth_task == (jint)chunk_top, "Control point invariant");
5063 start = chunk_array[chunk_top - 1];
5064 end = space->top();
5065 }
5066 MemRegion mr(start, end);
5067 // Verify that mr is in space
5068 assert(mr.is_empty() || space->used_region().contains(mr),
5069 "Should be in space");
5070 // Verify that "start" is an object boundary
5071 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5072 "Should be an oop");
5073 space->par_oop_iterate(mr, cl);
5074 }
5075 pst->all_tasks_completed();
5076 }
5078 void
5079 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5080 CompactibleFreeListSpace* sp, int i,
5081 Par_MarkRefsIntoAndScanClosure* cl) {
5082 // Until all tasks completed:
5083 // . claim an unclaimed task
5084 // . compute region boundaries corresponding to task claimed
5085 // . transfer dirty bits ct->mut for that region
5086 // . apply rescanclosure to dirty mut bits for that region
5088 ResourceMark rm;
5089 HandleMark hm;
5091 OopTaskQueue* work_q = work_queue(i);
5092 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5093 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5094 // CAUTION: This closure has state that persists across calls to
5095 // the work method dirty_range_iterate_clear() in that it has
5096 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5097 // use of that state in the imbedded UpwardsObjectClosure instance
5098 // assumes that the cards are always iterated (even if in parallel
5099 // by several threads) in monotonically increasing order per each
5100 // thread. This is true of the implementation below which picks
5101 // card ranges (chunks) in monotonically increasing order globally
5102 // and, a-fortiori, in monotonically increasing order per thread
5103 // (the latter order being a subsequence of the former).
5104 // If the work code below is ever reorganized into a more chaotic
5105 // work-partitioning form than the current "sequential tasks"
5106 // paradigm, the use of that persistent state will have to be
5107 // revisited and modified appropriately. See also related
5108 // bug 4756801 work on which should examine this code to make
5109 // sure that the changes there do not run counter to the
5110 // assumptions made here and necessary for correctness and
5111 // efficiency. Note also that this code might yield inefficient
5112 // behaviour in the case of very large objects that span one or
5113 // more work chunks. Such objects would potentially be scanned
5114 // several times redundantly. Work on 4756801 should try and
5115 // address that performance anomaly if at all possible. XXX
5116 MemRegion full_span = _collector->_span;
5117 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5118 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5119 MarkFromDirtyCardsClosure
5120 greyRescanClosure(_collector, full_span, // entire span of interest
5121 sp, bm, work_q, rs, cl);
5123 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5124 assert(pst->valid(), "Uninitialized use?");
5125 int nth_task = 0;
5126 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5127 MemRegion span = sp->used_region();
5128 HeapWord* start_addr = span.start();
5129 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5130 alignment);
5131 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5132 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5133 start_addr, "Check alignment");
5134 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5135 chunk_size, "Check alignment");
5137 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5138 // Having claimed the nth_task, compute corresponding mem-region,
5139 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5140 // The alignment restriction ensures that we do not need any
5141 // synchronization with other gang-workers while setting or
5142 // clearing bits in thus chunk of the MUT.
5143 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5144 start_addr + (nth_task+1)*chunk_size);
5145 // The last chunk's end might be way beyond end of the
5146 // used region. In that case pull back appropriately.
5147 if (this_span.end() > end_addr) {
5148 this_span.set_end(end_addr);
5149 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5150 }
5151 // Iterate over the dirty cards covering this chunk, marking them
5152 // precleaned, and setting the corresponding bits in the mod union
5153 // table. Since we have been careful to partition at Card and MUT-word
5154 // boundaries no synchronization is needed between parallel threads.
5155 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5156 &modUnionClosure);
5158 // Having transferred these marks into the modUnionTable,
5159 // rescan the marked objects on the dirty cards in the modUnionTable.
5160 // Even if this is at a synchronous collection, the initial marking
5161 // may have been done during an asynchronous collection so there
5162 // may be dirty bits in the mod-union table.
5163 _collector->_modUnionTable.dirty_range_iterate_clear(
5164 this_span, &greyRescanClosure);
5165 _collector->_modUnionTable.verifyNoOneBitsInRange(
5166 this_span.start(),
5167 this_span.end());
5168 }
5169 pst->all_tasks_completed(); // declare that i am done
5170 }
5172 // . see if we can share work_queues with ParNew? XXX
5173 void
5174 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5175 int* seed) {
5176 OopTaskQueue* work_q = work_queue(i);
5177 NOT_PRODUCT(int num_steals = 0;)
5178 oop obj_to_scan;
5179 CMSBitMap* bm = &(_collector->_markBitMap);
5180 size_t num_from_overflow_list =
5181 MIN2((size_t)work_q->max_elems()/4,
5182 (size_t)ParGCDesiredObjsFromOverflowList);
5184 while (true) {
5185 // Completely finish any left over work from (an) earlier round(s)
5186 cl->trim_queue(0);
5187 // Now check if there's any work in the overflow list
5188 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5189 work_q)) {
5190 // found something in global overflow list;
5191 // not yet ready to go stealing work from others.
5192 // We'd like to assert(work_q->size() != 0, ...)
5193 // because we just took work from the overflow list,
5194 // but of course we can't since all of that could have
5195 // been already stolen from us.
5196 // "He giveth and He taketh away."
5197 continue;
5198 }
5199 // Verify that we have no work before we resort to stealing
5200 assert(work_q->size() == 0, "Have work, shouldn't steal");
5201 // Try to steal from other queues that have work
5202 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5203 NOT_PRODUCT(num_steals++;)
5204 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5205 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5206 // Do scanning work
5207 obj_to_scan->oop_iterate(cl);
5208 // Loop around, finish this work, and try to steal some more
5209 } else if (terminator()->offer_termination()) {
5210 break; // nirvana from the infinite cycle
5211 }
5212 }
5213 NOT_PRODUCT(
5214 if (PrintCMSStatistics != 0) {
5215 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5216 }
5217 )
5218 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5219 "Else our work is not yet done");
5220 }
5222 // Return a thread-local PLAB recording array, as appropriate.
5223 void* CMSCollector::get_data_recorder(int thr_num) {
5224 if (_survivor_plab_array != NULL &&
5225 (CMSPLABRecordAlways ||
5226 (_collectorState > Marking && _collectorState < FinalMarking))) {
5227 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5228 ChunkArray* ca = &_survivor_plab_array[thr_num];
5229 ca->reset(); // clear it so that fresh data is recorded
5230 return (void*) ca;
5231 } else {
5232 return NULL;
5233 }
5234 }
5236 // Reset all the thread-local PLAB recording arrays
5237 void CMSCollector::reset_survivor_plab_arrays() {
5238 for (uint i = 0; i < ParallelGCThreads; i++) {
5239 _survivor_plab_array[i].reset();
5240 }
5241 }
5243 // Merge the per-thread plab arrays into the global survivor chunk
5244 // array which will provide the partitioning of the survivor space
5245 // for CMS rescan.
5246 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
5247 assert(_survivor_plab_array != NULL, "Error");
5248 assert(_survivor_chunk_array != NULL, "Error");
5249 assert(_collectorState == FinalMarking, "Error");
5250 for (uint j = 0; j < ParallelGCThreads; j++) {
5251 _cursor[j] = 0;
5252 }
5253 HeapWord* top = surv->top();
5254 size_t i;
5255 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5256 HeapWord* min_val = top; // Higher than any PLAB address
5257 uint min_tid = 0; // position of min_val this round
5258 for (uint j = 0; j < ParallelGCThreads; j++) {
5259 ChunkArray* cur_sca = &_survivor_plab_array[j];
5260 if (_cursor[j] == cur_sca->end()) {
5261 continue;
5262 }
5263 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5264 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5265 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5266 if (cur_val < min_val) {
5267 min_tid = j;
5268 min_val = cur_val;
5269 } else {
5270 assert(cur_val < top, "All recorded addresses should be less");
5271 }
5272 }
5273 // At this point min_val and min_tid are respectively
5274 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5275 // and the thread (j) that witnesses that address.
5276 // We record this address in the _survivor_chunk_array[i]
5277 // and increment _cursor[min_tid] prior to the next round i.
5278 if (min_val == top) {
5279 break;
5280 }
5281 _survivor_chunk_array[i] = min_val;
5282 _cursor[min_tid]++;
5283 }
5284 // We are all done; record the size of the _survivor_chunk_array
5285 _survivor_chunk_index = i; // exclusive: [0, i)
5286 if (PrintCMSStatistics > 0) {
5287 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5288 }
5289 // Verify that we used up all the recorded entries
5290 #ifdef ASSERT
5291 size_t total = 0;
5292 for (uint j = 0; j < ParallelGCThreads; j++) {
5293 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5294 total += _cursor[j];
5295 }
5296 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5297 // Check that the merged array is in sorted order
5298 if (total > 0) {
5299 for (size_t i = 0; i < total - 1; i++) {
5300 if (PrintCMSStatistics > 0) {
5301 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5302 i, _survivor_chunk_array[i]);
5303 }
5304 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5305 "Not sorted");
5306 }
5307 }
5308 #endif // ASSERT
5309 }
5311 // Set up the space's par_seq_tasks structure for work claiming
5312 // for parallel rescan of young gen.
5313 // See ParRescanTask where this is currently used.
5314 void
5315 CMSCollector::
5316 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5317 assert(n_threads > 0, "Unexpected n_threads argument");
5318 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5320 // Eden space
5321 {
5322 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5323 assert(!pst->valid(), "Clobbering existing data?");
5324 // Each valid entry in [0, _eden_chunk_index) represents a task.
5325 size_t n_tasks = _eden_chunk_index + 1;
5326 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5327 pst->set_par_threads(n_threads);
5328 pst->set_n_tasks((int)n_tasks);
5329 }
5331 // Merge the survivor plab arrays into _survivor_chunk_array
5332 if (_survivor_plab_array != NULL) {
5333 merge_survivor_plab_arrays(dng->from());
5334 } else {
5335 assert(_survivor_chunk_index == 0, "Error");
5336 }
5338 // To space
5339 {
5340 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5341 assert(!pst->valid(), "Clobbering existing data?");
5342 pst->set_par_threads(n_threads);
5343 pst->set_n_tasks(1);
5344 assert(pst->valid(), "Error");
5345 }
5347 // From space
5348 {
5349 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5350 assert(!pst->valid(), "Clobbering existing data?");
5351 size_t n_tasks = _survivor_chunk_index + 1;
5352 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5353 pst->set_par_threads(n_threads);
5354 pst->set_n_tasks((int)n_tasks);
5355 assert(pst->valid(), "Error");
5356 }
5357 }
5359 // Parallel version of remark
5360 void CMSCollector::do_remark_parallel() {
5361 GenCollectedHeap* gch = GenCollectedHeap::heap();
5362 WorkGang* workers = gch->workers();
5363 assert(workers != NULL, "Need parallel worker threads.");
5364 int n_workers = workers->total_workers();
5365 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5366 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5368 CMSParRemarkTask tsk(this,
5369 cms_space, perm_space,
5370 n_workers, workers, task_queues());
5372 // Set up for parallel process_strong_roots work.
5373 gch->set_par_threads(n_workers);
5374 gch->change_strong_roots_parity();
5375 // We won't be iterating over the cards in the card table updating
5376 // the younger_gen cards, so we shouldn't call the following else
5377 // the verification code as well as subsequent younger_refs_iterate
5378 // code would get confused. XXX
5379 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5381 // The young gen rescan work will not be done as part of
5382 // process_strong_roots (which currently doesn't knw how to
5383 // parallelize such a scan), but rather will be broken up into
5384 // a set of parallel tasks (via the sampling that the [abortable]
5385 // preclean phase did of EdenSpace, plus the [two] tasks of
5386 // scanning the [two] survivor spaces. Further fine-grain
5387 // parallelization of the scanning of the survivor spaces
5388 // themselves, and of precleaning of the younger gen itself
5389 // is deferred to the future.
5390 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5392 // The dirty card rescan work is broken up into a "sequence"
5393 // of parallel tasks (per constituent space) that are dynamically
5394 // claimed by the parallel threads.
5395 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5396 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5398 // It turns out that even when we're using 1 thread, doing the work in a
5399 // separate thread causes wide variance in run times. We can't help this
5400 // in the multi-threaded case, but we special-case n=1 here to get
5401 // repeatable measurements of the 1-thread overhead of the parallel code.
5402 if (n_workers > 1) {
5403 // Make refs discovery MT-safe
5404 ReferenceProcessorMTMutator mt(ref_processor(), true);
5405 workers->run_task(&tsk);
5406 } else {
5407 tsk.work(0);
5408 }
5409 gch->set_par_threads(0); // 0 ==> non-parallel.
5410 // restore, single-threaded for now, any preserved marks
5411 // as a result of work_q overflow
5412 restore_preserved_marks_if_any();
5413 }
5415 // Non-parallel version of remark
5416 void CMSCollector::do_remark_non_parallel() {
5417 ResourceMark rm;
5418 HandleMark hm;
5419 GenCollectedHeap* gch = GenCollectedHeap::heap();
5420 MarkRefsIntoAndScanClosure
5421 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5422 &_markStack, &_revisitStack, this,
5423 false /* should_yield */, false /* not precleaning */);
5424 MarkFromDirtyCardsClosure
5425 markFromDirtyCardsClosure(this, _span,
5426 NULL, // space is set further below
5427 &_markBitMap, &_markStack, &_revisitStack,
5428 &mrias_cl);
5429 {
5430 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5431 // Iterate over the dirty cards, setting the corresponding bits in the
5432 // mod union table.
5433 {
5434 ModUnionClosure modUnionClosure(&_modUnionTable);
5435 _ct->ct_bs()->dirty_card_iterate(
5436 _cmsGen->used_region(),
5437 &modUnionClosure);
5438 _ct->ct_bs()->dirty_card_iterate(
5439 _permGen->used_region(),
5440 &modUnionClosure);
5441 }
5442 // Having transferred these marks into the modUnionTable, we just need
5443 // to rescan the marked objects on the dirty cards in the modUnionTable.
5444 // The initial marking may have been done during an asynchronous
5445 // collection so there may be dirty bits in the mod-union table.
5446 const int alignment =
5447 CardTableModRefBS::card_size * BitsPerWord;
5448 {
5449 // ... First handle dirty cards in CMS gen
5450 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5451 MemRegion ur = _cmsGen->used_region();
5452 HeapWord* lb = ur.start();
5453 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5454 MemRegion cms_span(lb, ub);
5455 _modUnionTable.dirty_range_iterate_clear(cms_span,
5456 &markFromDirtyCardsClosure);
5457 verify_work_stacks_empty();
5458 if (PrintCMSStatistics != 0) {
5459 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5460 markFromDirtyCardsClosure.num_dirty_cards());
5461 }
5462 }
5463 {
5464 // .. and then repeat for dirty cards in perm gen
5465 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5466 MemRegion ur = _permGen->used_region();
5467 HeapWord* lb = ur.start();
5468 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5469 MemRegion perm_span(lb, ub);
5470 _modUnionTable.dirty_range_iterate_clear(perm_span,
5471 &markFromDirtyCardsClosure);
5472 verify_work_stacks_empty();
5473 if (PrintCMSStatistics != 0) {
5474 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5475 markFromDirtyCardsClosure.num_dirty_cards());
5476 }
5477 }
5478 }
5479 if (VerifyDuringGC &&
5480 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5481 HandleMark hm; // Discard invalid handles created during verification
5482 Universe::verify(true);
5483 }
5484 {
5485 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5487 verify_work_stacks_empty();
5489 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5490 gch->gen_process_strong_roots(_cmsGen->level(),
5491 true, // younger gens as roots
5492 true, // collecting perm gen
5493 SharedHeap::ScanningOption(roots_scanning_options()),
5494 NULL, &mrias_cl);
5495 }
5496 verify_work_stacks_empty();
5497 // Restore evacuated mark words, if any, used for overflow list links
5498 if (!CMSOverflowEarlyRestoration) {
5499 restore_preserved_marks_if_any();
5500 }
5501 verify_overflow_empty();
5502 }
5504 ////////////////////////////////////////////////////////
5505 // Parallel Reference Processing Task Proxy Class
5506 ////////////////////////////////////////////////////////
5507 class CMSRefProcTaskProxy: public AbstractGangTask {
5508 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5509 CMSCollector* _collector;
5510 CMSBitMap* _mark_bit_map;
5511 const MemRegion _span;
5512 OopTaskQueueSet* _task_queues;
5513 ParallelTaskTerminator _term;
5514 ProcessTask& _task;
5516 public:
5517 CMSRefProcTaskProxy(ProcessTask& task,
5518 CMSCollector* collector,
5519 const MemRegion& span,
5520 CMSBitMap* mark_bit_map,
5521 int total_workers,
5522 OopTaskQueueSet* task_queues):
5523 AbstractGangTask("Process referents by policy in parallel"),
5524 _task(task),
5525 _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
5526 _task_queues(task_queues),
5527 _term(total_workers, task_queues)
5528 {
5529 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5530 "Inconsistency in _span");
5531 }
5533 OopTaskQueueSet* task_queues() { return _task_queues; }
5535 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5537 ParallelTaskTerminator* terminator() { return &_term; }
5539 void do_work_steal(int i,
5540 CMSParDrainMarkingStackClosure* drain,
5541 CMSParKeepAliveClosure* keep_alive,
5542 int* seed);
5544 virtual void work(int i);
5545 };
5547 void CMSRefProcTaskProxy::work(int i) {
5548 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5549 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5550 _mark_bit_map, work_queue(i));
5551 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5552 _mark_bit_map, work_queue(i));
5553 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5554 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5555 if (_task.marks_oops_alive()) {
5556 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5557 _collector->hash_seed(i));
5558 }
5559 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5560 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5561 }
5563 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5564 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5565 EnqueueTask& _task;
5567 public:
5568 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5569 : AbstractGangTask("Enqueue reference objects in parallel"),
5570 _task(task)
5571 { }
5573 virtual void work(int i)
5574 {
5575 _task.work(i);
5576 }
5577 };
5579 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5580 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5581 _collector(collector),
5582 _span(span),
5583 _bit_map(bit_map),
5584 _work_queue(work_queue),
5585 _mark_and_push(collector, span, bit_map, work_queue),
5586 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5587 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5588 { }
5590 // . see if we can share work_queues with ParNew? XXX
5591 void CMSRefProcTaskProxy::do_work_steal(int i,
5592 CMSParDrainMarkingStackClosure* drain,
5593 CMSParKeepAliveClosure* keep_alive,
5594 int* seed) {
5595 OopTaskQueue* work_q = work_queue(i);
5596 NOT_PRODUCT(int num_steals = 0;)
5597 oop obj_to_scan;
5598 size_t num_from_overflow_list =
5599 MIN2((size_t)work_q->max_elems()/4,
5600 (size_t)ParGCDesiredObjsFromOverflowList);
5602 while (true) {
5603 // Completely finish any left over work from (an) earlier round(s)
5604 drain->trim_queue(0);
5605 // Now check if there's any work in the overflow list
5606 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5607 work_q)) {
5608 // Found something in global overflow list;
5609 // not yet ready to go stealing work from others.
5610 // We'd like to assert(work_q->size() != 0, ...)
5611 // because we just took work from the overflow list,
5612 // but of course we can't, since all of that might have
5613 // been already stolen from us.
5614 continue;
5615 }
5616 // Verify that we have no work before we resort to stealing
5617 assert(work_q->size() == 0, "Have work, shouldn't steal");
5618 // Try to steal from other queues that have work
5619 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5620 NOT_PRODUCT(num_steals++;)
5621 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5622 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5623 // Do scanning work
5624 obj_to_scan->oop_iterate(keep_alive);
5625 // Loop around, finish this work, and try to steal some more
5626 } else if (terminator()->offer_termination()) {
5627 break; // nirvana from the infinite cycle
5628 }
5629 }
5630 NOT_PRODUCT(
5631 if (PrintCMSStatistics != 0) {
5632 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5633 }
5634 )
5635 }
5637 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5638 {
5639 GenCollectedHeap* gch = GenCollectedHeap::heap();
5640 WorkGang* workers = gch->workers();
5641 assert(workers != NULL, "Need parallel worker threads.");
5642 int n_workers = workers->total_workers();
5643 CMSRefProcTaskProxy rp_task(task, &_collector,
5644 _collector.ref_processor()->span(),
5645 _collector.markBitMap(),
5646 n_workers, _collector.task_queues());
5647 workers->run_task(&rp_task);
5648 }
5650 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5651 {
5653 GenCollectedHeap* gch = GenCollectedHeap::heap();
5654 WorkGang* workers = gch->workers();
5655 assert(workers != NULL, "Need parallel worker threads.");
5656 CMSRefEnqueueTaskProxy enq_task(task);
5657 workers->run_task(&enq_task);
5658 }
5660 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5662 ResourceMark rm;
5663 HandleMark hm;
5664 ReferencePolicy* soft_ref_policy;
5666 assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete");
5667 // Process weak references.
5668 if (clear_all_soft_refs) {
5669 soft_ref_policy = new AlwaysClearPolicy();
5670 } else {
5671 #ifdef COMPILER2
5672 soft_ref_policy = new LRUMaxHeapPolicy();
5673 #else
5674 soft_ref_policy = new LRUCurrentHeapPolicy();
5675 #endif // COMPILER2
5676 }
5677 verify_work_stacks_empty();
5679 ReferenceProcessor* rp = ref_processor();
5680 assert(rp->span().equals(_span), "Spans should be equal");
5681 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5682 &_markStack);
5683 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5684 _span, &_markBitMap, &_markStack,
5685 &cmsKeepAliveClosure);
5686 {
5687 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5688 if (rp->processing_is_mt()) {
5689 CMSRefProcTaskExecutor task_executor(*this);
5690 rp->process_discovered_references(soft_ref_policy,
5691 &_is_alive_closure,
5692 &cmsKeepAliveClosure,
5693 &cmsDrainMarkingStackClosure,
5694 &task_executor);
5695 } else {
5696 rp->process_discovered_references(soft_ref_policy,
5697 &_is_alive_closure,
5698 &cmsKeepAliveClosure,
5699 &cmsDrainMarkingStackClosure,
5700 NULL);
5701 }
5702 verify_work_stacks_empty();
5703 }
5705 if (should_unload_classes()) {
5706 {
5707 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5709 // Follow SystemDictionary roots and unload classes
5710 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5712 // Follow CodeCache roots and unload any methods marked for unloading
5713 CodeCache::do_unloading(&_is_alive_closure,
5714 &cmsKeepAliveClosure,
5715 purged_class);
5717 cmsDrainMarkingStackClosure.do_void();
5718 verify_work_stacks_empty();
5720 // Update subklass/sibling/implementor links in KlassKlass descendants
5721 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5722 oop k;
5723 while ((k = _revisitStack.pop()) != NULL) {
5724 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5725 &_is_alive_closure,
5726 &cmsKeepAliveClosure);
5727 }
5728 assert(!ClassUnloading ||
5729 (_markStack.isEmpty() && overflow_list_is_empty()),
5730 "Should not have found new reachable objects");
5731 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5732 cmsDrainMarkingStackClosure.do_void();
5733 verify_work_stacks_empty();
5734 }
5736 {
5737 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5738 // Now clean up stale oops in SymbolTable and StringTable
5739 SymbolTable::unlink(&_is_alive_closure);
5740 StringTable::unlink(&_is_alive_closure);
5741 }
5742 }
5744 verify_work_stacks_empty();
5745 // Restore any preserved marks as a result of mark stack or
5746 // work queue overflow
5747 restore_preserved_marks_if_any(); // done single-threaded for now
5749 rp->set_enqueuing_is_done(true);
5750 if (rp->processing_is_mt()) {
5751 CMSRefProcTaskExecutor task_executor(*this);
5752 rp->enqueue_discovered_references(&task_executor);
5753 } else {
5754 rp->enqueue_discovered_references(NULL);
5755 }
5756 rp->verify_no_references_recorded();
5757 assert(!rp->discovery_enabled(), "should have been disabled");
5759 // JVMTI object tagging is based on JNI weak refs. If any of these
5760 // refs were cleared then JVMTI needs to update its maps and
5761 // maybe post ObjectFrees to agents.
5762 JvmtiExport::cms_ref_processing_epilogue();
5763 }
5765 #ifndef PRODUCT
5766 void CMSCollector::check_correct_thread_executing() {
5767 Thread* t = Thread::current();
5768 // Only the VM thread or the CMS thread should be here.
5769 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5770 "Unexpected thread type");
5771 // If this is the vm thread, the foreground process
5772 // should not be waiting. Note that _foregroundGCIsActive is
5773 // true while the foreground collector is waiting.
5774 if (_foregroundGCShouldWait) {
5775 // We cannot be the VM thread
5776 assert(t->is_ConcurrentGC_thread(),
5777 "Should be CMS thread");
5778 } else {
5779 // We can be the CMS thread only if we are in a stop-world
5780 // phase of CMS collection.
5781 if (t->is_ConcurrentGC_thread()) {
5782 assert(_collectorState == InitialMarking ||
5783 _collectorState == FinalMarking,
5784 "Should be a stop-world phase");
5785 // The CMS thread should be holding the CMS_token.
5786 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5787 "Potential interference with concurrently "
5788 "executing VM thread");
5789 }
5790 }
5791 }
5792 #endif
5794 void CMSCollector::sweep(bool asynch) {
5795 assert(_collectorState == Sweeping, "just checking");
5796 check_correct_thread_executing();
5797 verify_work_stacks_empty();
5798 verify_overflow_empty();
5799 incrementSweepCount();
5800 _sweep_timer.stop();
5801 _sweep_estimate.sample(_sweep_timer.seconds());
5802 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5804 // PermGen verification support: If perm gen sweeping is disabled in
5805 // this cycle, we preserve the perm gen object "deadness" information
5806 // in the perm_gen_verify_bit_map. In order to do that we traverse
5807 // all blocks in perm gen and mark all dead objects.
5808 if (verifying() && !should_unload_classes()) {
5809 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
5810 "Should have already been allocated");
5811 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
5812 markBitMap(), perm_gen_verify_bit_map());
5813 if (asynch) {
5814 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5815 bitMapLock());
5816 _permGen->cmsSpace()->blk_iterate(&mdo);
5817 } else {
5818 // In the case of synchronous sweep, we already have
5819 // the requisite locks/tokens.
5820 _permGen->cmsSpace()->blk_iterate(&mdo);
5821 }
5822 }
5824 if (asynch) {
5825 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5826 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5827 // First sweep the old gen then the perm gen
5828 {
5829 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5830 bitMapLock());
5831 sweepWork(_cmsGen, asynch);
5832 }
5834 // Now repeat for perm gen
5835 if (should_unload_classes()) {
5836 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5837 bitMapLock());
5838 sweepWork(_permGen, asynch);
5839 }
5841 // Update Universe::_heap_*_at_gc figures.
5842 // We need all the free list locks to make the abstract state
5843 // transition from Sweeping to Resetting. See detailed note
5844 // further below.
5845 {
5846 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5847 _permGen->freelistLock());
5848 // Update heap occupancy information which is used as
5849 // input to soft ref clearing policy at the next gc.
5850 Universe::update_heap_info_at_gc();
5851 _collectorState = Resizing;
5852 }
5853 } else {
5854 // already have needed locks
5855 sweepWork(_cmsGen, asynch);
5857 if (should_unload_classes()) {
5858 sweepWork(_permGen, asynch);
5859 }
5860 // Update heap occupancy information which is used as
5861 // input to soft ref clearing policy at the next gc.
5862 Universe::update_heap_info_at_gc();
5863 _collectorState = Resizing;
5864 }
5865 verify_work_stacks_empty();
5866 verify_overflow_empty();
5868 _sweep_timer.reset();
5869 _sweep_timer.start();
5871 update_time_of_last_gc(os::javaTimeMillis());
5873 // NOTE on abstract state transitions:
5874 // Mutators allocate-live and/or mark the mod-union table dirty
5875 // based on the state of the collection. The former is done in
5876 // the interval [Marking, Sweeping] and the latter in the interval
5877 // [Marking, Sweeping). Thus the transitions into the Marking state
5878 // and out of the Sweeping state must be synchronously visible
5879 // globally to the mutators.
5880 // The transition into the Marking state happens with the world
5881 // stopped so the mutators will globally see it. Sweeping is
5882 // done asynchronously by the background collector so the transition
5883 // from the Sweeping state to the Resizing state must be done
5884 // under the freelistLock (as is the check for whether to
5885 // allocate-live and whether to dirty the mod-union table).
5886 assert(_collectorState == Resizing, "Change of collector state to"
5887 " Resizing must be done under the freelistLocks (plural)");
5889 // Now that sweeping has been completed, if the GCH's
5890 // incremental_collection_will_fail flag is set, clear it,
5891 // thus inviting a younger gen collection to promote into
5892 // this generation. If such a promotion may still fail,
5893 // the flag will be set again when a young collection is
5894 // attempted.
5895 // I think the incremental_collection_will_fail flag's use
5896 // is specific to a 2 generation collection policy, so i'll
5897 // assert that that's the configuration we are operating within.
5898 // The use of the flag can and should be generalized appropriately
5899 // in the future to deal with a general n-generation system.
5901 GenCollectedHeap* gch = GenCollectedHeap::heap();
5902 assert(gch->collector_policy()->is_two_generation_policy(),
5903 "Resetting of incremental_collection_will_fail flag"
5904 " may be incorrect otherwise");
5905 gch->clear_incremental_collection_will_fail();
5906 gch->update_full_collections_completed(_collection_count_start);
5907 }
5909 // FIX ME!!! Looks like this belongs in CFLSpace, with
5910 // CMSGen merely delegating to it.
5911 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5912 double nearLargestPercent = 0.999;
5913 HeapWord* minAddr = _cmsSpace->bottom();
5914 HeapWord* largestAddr =
5915 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
5916 if (largestAddr == 0) {
5917 // The dictionary appears to be empty. In this case
5918 // try to coalesce at the end of the heap.
5919 largestAddr = _cmsSpace->end();
5920 }
5921 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5922 size_t nearLargestOffset =
5923 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5924 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5925 }
5927 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5928 return addr >= _cmsSpace->nearLargestChunk();
5929 }
5931 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5932 return _cmsSpace->find_chunk_at_end();
5933 }
5935 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5936 bool full) {
5937 // The next lower level has been collected. Gather any statistics
5938 // that are of interest at this point.
5939 if (!full && (current_level + 1) == level()) {
5940 // Gather statistics on the young generation collection.
5941 collector()->stats().record_gc0_end(used());
5942 }
5943 }
5945 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
5946 GenCollectedHeap* gch = GenCollectedHeap::heap();
5947 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
5948 "Wrong type of heap");
5949 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
5950 gch->gen_policy()->size_policy();
5951 assert(sp->is_gc_cms_adaptive_size_policy(),
5952 "Wrong type of size policy");
5953 return sp;
5954 }
5956 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
5957 if (PrintGCDetails && Verbose) {
5958 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
5959 }
5960 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
5961 _debug_collection_type =
5962 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
5963 if (PrintGCDetails && Verbose) {
5964 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
5965 }
5966 }
5968 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
5969 bool asynch) {
5970 // We iterate over the space(s) underlying this generation,
5971 // checking the mark bit map to see if the bits corresponding
5972 // to specific blocks are marked or not. Blocks that are
5973 // marked are live and are not swept up. All remaining blocks
5974 // are swept up, with coalescing on-the-fly as we sweep up
5975 // contiguous free and/or garbage blocks:
5976 // We need to ensure that the sweeper synchronizes with allocators
5977 // and stop-the-world collectors. In particular, the following
5978 // locks are used:
5979 // . CMS token: if this is held, a stop the world collection cannot occur
5980 // . freelistLock: if this is held no allocation can occur from this
5981 // generation by another thread
5982 // . bitMapLock: if this is held, no other thread can access or update
5983 //
5985 // Note that we need to hold the freelistLock if we use
5986 // block iterate below; else the iterator might go awry if
5987 // a mutator (or promotion) causes block contents to change
5988 // (for instance if the allocator divvies up a block).
5989 // If we hold the free list lock, for all practical purposes
5990 // young generation GC's can't occur (they'll usually need to
5991 // promote), so we might as well prevent all young generation
5992 // GC's while we do a sweeping step. For the same reason, we might
5993 // as well take the bit map lock for the entire duration
5995 // check that we hold the requisite locks
5996 assert(have_cms_token(), "Should hold cms token");
5997 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
5998 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
5999 "Should possess CMS token to sweep");
6000 assert_lock_strong(gen->freelistLock());
6001 assert_lock_strong(bitMapLock());
6003 assert(!_sweep_timer.is_active(), "Was switched off in an outer context");
6004 gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
6005 _sweep_estimate.padded_average());
6006 gen->setNearLargestChunk();
6008 {
6009 SweepClosure sweepClosure(this, gen, &_markBitMap,
6010 CMSYield && asynch);
6011 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6012 // We need to free-up/coalesce garbage/blocks from a
6013 // co-terminal free run. This is done in the SweepClosure
6014 // destructor; so, do not remove this scope, else the
6015 // end-of-sweep-census below will be off by a little bit.
6016 }
6017 gen->cmsSpace()->sweep_completed();
6018 gen->cmsSpace()->endSweepFLCensus(sweepCount());
6019 if (should_unload_classes()) { // unloaded classes this cycle,
6020 _concurrent_cycles_since_last_unload = 0; // ... reset count
6021 } else { // did not unload classes,
6022 _concurrent_cycles_since_last_unload++; // ... increment count
6023 }
6024 }
6026 // Reset CMS data structures (for now just the marking bit map)
6027 // preparatory for the next cycle.
6028 void CMSCollector::reset(bool asynch) {
6029 GenCollectedHeap* gch = GenCollectedHeap::heap();
6030 CMSAdaptiveSizePolicy* sp = size_policy();
6031 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6032 if (asynch) {
6033 CMSTokenSyncWithLocks ts(true, bitMapLock());
6035 // If the state is not "Resetting", the foreground thread
6036 // has done a collection and the resetting.
6037 if (_collectorState != Resetting) {
6038 assert(_collectorState == Idling, "The state should only change"
6039 " because the foreground collector has finished the collection");
6040 return;
6041 }
6043 // Clear the mark bitmap (no grey objects to start with)
6044 // for the next cycle.
6045 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6046 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6048 HeapWord* curAddr = _markBitMap.startWord();
6049 while (curAddr < _markBitMap.endWord()) {
6050 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6051 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6052 _markBitMap.clear_large_range(chunk);
6053 if (ConcurrentMarkSweepThread::should_yield() &&
6054 !foregroundGCIsActive() &&
6055 CMSYield) {
6056 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6057 "CMS thread should hold CMS token");
6058 assert_lock_strong(bitMapLock());
6059 bitMapLock()->unlock();
6060 ConcurrentMarkSweepThread::desynchronize(true);
6061 ConcurrentMarkSweepThread::acknowledge_yield_request();
6062 stopTimer();
6063 if (PrintCMSStatistics != 0) {
6064 incrementYields();
6065 }
6066 icms_wait();
6068 // See the comment in coordinator_yield()
6069 for (unsigned i = 0; i < CMSYieldSleepCount &&
6070 ConcurrentMarkSweepThread::should_yield() &&
6071 !CMSCollector::foregroundGCIsActive(); ++i) {
6072 os::sleep(Thread::current(), 1, false);
6073 ConcurrentMarkSweepThread::acknowledge_yield_request();
6074 }
6076 ConcurrentMarkSweepThread::synchronize(true);
6077 bitMapLock()->lock_without_safepoint_check();
6078 startTimer();
6079 }
6080 curAddr = chunk.end();
6081 }
6082 _collectorState = Idling;
6083 } else {
6084 // already have the lock
6085 assert(_collectorState == Resetting, "just checking");
6086 assert_lock_strong(bitMapLock());
6087 _markBitMap.clear_all();
6088 _collectorState = Idling;
6089 }
6091 // Stop incremental mode after a cycle completes, so that any future cycles
6092 // are triggered by allocation.
6093 stop_icms();
6095 NOT_PRODUCT(
6096 if (RotateCMSCollectionTypes) {
6097 _cmsGen->rotate_debug_collection_type();
6098 }
6099 )
6100 }
6102 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6103 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6104 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6105 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6106 TraceCollectorStats tcs(counters());
6108 switch (op) {
6109 case CMS_op_checkpointRootsInitial: {
6110 checkpointRootsInitial(true); // asynch
6111 if (PrintGC) {
6112 _cmsGen->printOccupancy("initial-mark");
6113 }
6114 break;
6115 }
6116 case CMS_op_checkpointRootsFinal: {
6117 checkpointRootsFinal(true, // asynch
6118 false, // !clear_all_soft_refs
6119 false); // !init_mark_was_synchronous
6120 if (PrintGC) {
6121 _cmsGen->printOccupancy("remark");
6122 }
6123 break;
6124 }
6125 default:
6126 fatal("No such CMS_op");
6127 }
6128 }
6130 #ifndef PRODUCT
6131 size_t const CMSCollector::skip_header_HeapWords() {
6132 return FreeChunk::header_size();
6133 }
6135 // Try and collect here conditions that should hold when
6136 // CMS thread is exiting. The idea is that the foreground GC
6137 // thread should not be blocked if it wants to terminate
6138 // the CMS thread and yet continue to run the VM for a while
6139 // after that.
6140 void CMSCollector::verify_ok_to_terminate() const {
6141 assert(Thread::current()->is_ConcurrentGC_thread(),
6142 "should be called by CMS thread");
6143 assert(!_foregroundGCShouldWait, "should be false");
6144 // We could check here that all the various low-level locks
6145 // are not held by the CMS thread, but that is overkill; see
6146 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6147 // is checked.
6148 }
6149 #endif
6151 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6152 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6153 "missing Printezis mark?");
6154 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6155 size_t size = pointer_delta(nextOneAddr + 1, addr);
6156 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6157 "alignment problem");
6158 assert(size >= 3, "Necessary for Printezis marks to work");
6159 return size;
6160 }
6162 // A variant of the above (block_size_using_printezis_bits()) except
6163 // that we return 0 if the P-bits are not yet set.
6164 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6165 if (_markBitMap.isMarked(addr)) {
6166 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
6167 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6168 size_t size = pointer_delta(nextOneAddr + 1, addr);
6169 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6170 "alignment problem");
6171 assert(size >= 3, "Necessary for Printezis marks to work");
6172 return size;
6173 } else {
6174 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
6175 return 0;
6176 }
6177 }
6179 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6180 size_t sz = 0;
6181 oop p = (oop)addr;
6182 if (p->klass_or_null() != NULL && p->is_parsable()) {
6183 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6184 } else {
6185 sz = block_size_using_printezis_bits(addr);
6186 }
6187 assert(sz > 0, "size must be nonzero");
6188 HeapWord* next_block = addr + sz;
6189 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6190 CardTableModRefBS::card_size);
6191 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6192 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6193 "must be different cards");
6194 return next_card;
6195 }
6198 // CMS Bit Map Wrapper /////////////////////////////////////////
6200 // Construct a CMS bit map infrastructure, but don't create the
6201 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6202 // further below.
6203 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6204 _bm(),
6205 _shifter(shifter),
6206 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6207 {
6208 _bmStartWord = 0;
6209 _bmWordSize = 0;
6210 }
6212 bool CMSBitMap::allocate(MemRegion mr) {
6213 _bmStartWord = mr.start();
6214 _bmWordSize = mr.word_size();
6215 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6216 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6217 if (!brs.is_reserved()) {
6218 warning("CMS bit map allocation failure");
6219 return false;
6220 }
6221 // For now we'll just commit all of the bit map up fromt.
6222 // Later on we'll try to be more parsimonious with swap.
6223 if (!_virtual_space.initialize(brs, brs.size())) {
6224 warning("CMS bit map backing store failure");
6225 return false;
6226 }
6227 assert(_virtual_space.committed_size() == brs.size(),
6228 "didn't reserve backing store for all of CMS bit map?");
6229 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6230 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6231 _bmWordSize, "inconsistency in bit map sizing");
6232 _bm.set_size(_bmWordSize >> _shifter);
6234 // bm.clear(); // can we rely on getting zero'd memory? verify below
6235 assert(isAllClear(),
6236 "Expected zero'd memory from ReservedSpace constructor");
6237 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6238 "consistency check");
6239 return true;
6240 }
6242 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6243 HeapWord *next_addr, *end_addr, *last_addr;
6244 assert_locked();
6245 assert(covers(mr), "out-of-range error");
6246 // XXX assert that start and end are appropriately aligned
6247 for (next_addr = mr.start(), end_addr = mr.end();
6248 next_addr < end_addr; next_addr = last_addr) {
6249 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6250 last_addr = dirty_region.end();
6251 if (!dirty_region.is_empty()) {
6252 cl->do_MemRegion(dirty_region);
6253 } else {
6254 assert(last_addr == end_addr, "program logic");
6255 return;
6256 }
6257 }
6258 }
6260 #ifndef PRODUCT
6261 void CMSBitMap::assert_locked() const {
6262 CMSLockVerifier::assert_locked(lock());
6263 }
6265 bool CMSBitMap::covers(MemRegion mr) const {
6266 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6267 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6268 "size inconsistency");
6269 return (mr.start() >= _bmStartWord) &&
6270 (mr.end() <= endWord());
6271 }
6273 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6274 return (start >= _bmStartWord && (start + size) <= endWord());
6275 }
6277 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6278 // verify that there are no 1 bits in the interval [left, right)
6279 FalseBitMapClosure falseBitMapClosure;
6280 iterate(&falseBitMapClosure, left, right);
6281 }
6283 void CMSBitMap::region_invariant(MemRegion mr)
6284 {
6285 assert_locked();
6286 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6287 assert(!mr.is_empty(), "unexpected empty region");
6288 assert(covers(mr), "mr should be covered by bit map");
6289 // convert address range into offset range
6290 size_t start_ofs = heapWordToOffset(mr.start());
6291 // Make sure that end() is appropriately aligned
6292 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6293 (1 << (_shifter+LogHeapWordSize))),
6294 "Misaligned mr.end()");
6295 size_t end_ofs = heapWordToOffset(mr.end());
6296 assert(end_ofs > start_ofs, "Should mark at least one bit");
6297 }
6299 #endif
6301 bool CMSMarkStack::allocate(size_t size) {
6302 // allocate a stack of the requisite depth
6303 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6304 size * sizeof(oop)));
6305 if (!rs.is_reserved()) {
6306 warning("CMSMarkStack allocation failure");
6307 return false;
6308 }
6309 if (!_virtual_space.initialize(rs, rs.size())) {
6310 warning("CMSMarkStack backing store failure");
6311 return false;
6312 }
6313 assert(_virtual_space.committed_size() == rs.size(),
6314 "didn't reserve backing store for all of CMS stack?");
6315 _base = (oop*)(_virtual_space.low());
6316 _index = 0;
6317 _capacity = size;
6318 NOT_PRODUCT(_max_depth = 0);
6319 return true;
6320 }
6322 // XXX FIX ME !!! In the MT case we come in here holding a
6323 // leaf lock. For printing we need to take a further lock
6324 // which has lower rank. We need to recallibrate the two
6325 // lock-ranks involved in order to be able to rpint the
6326 // messages below. (Or defer the printing to the caller.
6327 // For now we take the expedient path of just disabling the
6328 // messages for the problematic case.)
6329 void CMSMarkStack::expand() {
6330 assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted");
6331 if (_capacity == CMSMarkStackSizeMax) {
6332 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6333 // We print a warning message only once per CMS cycle.
6334 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6335 }
6336 return;
6337 }
6338 // Double capacity if possible
6339 size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax);
6340 // Do not give up existing stack until we have managed to
6341 // get the double capacity that we desired.
6342 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6343 new_capacity * sizeof(oop)));
6344 if (rs.is_reserved()) {
6345 // Release the backing store associated with old stack
6346 _virtual_space.release();
6347 // Reinitialize virtual space for new stack
6348 if (!_virtual_space.initialize(rs, rs.size())) {
6349 fatal("Not enough swap for expanded marking stack");
6350 }
6351 _base = (oop*)(_virtual_space.low());
6352 _index = 0;
6353 _capacity = new_capacity;
6354 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6355 // Failed to double capacity, continue;
6356 // we print a detail message only once per CMS cycle.
6357 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6358 SIZE_FORMAT"K",
6359 _capacity / K, new_capacity / K);
6360 }
6361 }
6364 // Closures
6365 // XXX: there seems to be a lot of code duplication here;
6366 // should refactor and consolidate common code.
6368 // This closure is used to mark refs into the CMS generation in
6369 // the CMS bit map. Called at the first checkpoint. This closure
6370 // assumes that we do not need to re-mark dirty cards; if the CMS
6371 // generation on which this is used is not an oldest (modulo perm gen)
6372 // generation then this will lose younger_gen cards!
6374 MarkRefsIntoClosure::MarkRefsIntoClosure(
6375 MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods):
6376 _span(span),
6377 _bitMap(bitMap),
6378 _should_do_nmethods(should_do_nmethods)
6379 {
6380 assert(_ref_processor == NULL, "deliberately left NULL");
6381 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6382 }
6384 void MarkRefsIntoClosure::do_oop(oop obj) {
6385 // if p points into _span, then mark corresponding bit in _markBitMap
6386 assert(obj->is_oop(), "expected an oop");
6387 HeapWord* addr = (HeapWord*)obj;
6388 if (_span.contains(addr)) {
6389 // this should be made more efficient
6390 _bitMap->mark(addr);
6391 }
6392 }
6394 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6395 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6397 // A variant of the above, used for CMS marking verification.
6398 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6399 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6400 bool should_do_nmethods):
6401 _span(span),
6402 _verification_bm(verification_bm),
6403 _cms_bm(cms_bm),
6404 _should_do_nmethods(should_do_nmethods) {
6405 assert(_ref_processor == NULL, "deliberately left NULL");
6406 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6407 }
6409 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6410 // if p points into _span, then mark corresponding bit in _markBitMap
6411 assert(obj->is_oop(), "expected an oop");
6412 HeapWord* addr = (HeapWord*)obj;
6413 if (_span.contains(addr)) {
6414 _verification_bm->mark(addr);
6415 if (!_cms_bm->isMarked(addr)) {
6416 oop(addr)->print();
6417 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6418 fatal("... aborting");
6419 }
6420 }
6421 }
6423 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6424 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6426 //////////////////////////////////////////////////
6427 // MarkRefsIntoAndScanClosure
6428 //////////////////////////////////////////////////
6430 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6431 ReferenceProcessor* rp,
6432 CMSBitMap* bit_map,
6433 CMSBitMap* mod_union_table,
6434 CMSMarkStack* mark_stack,
6435 CMSMarkStack* revisit_stack,
6436 CMSCollector* collector,
6437 bool should_yield,
6438 bool concurrent_precleaning):
6439 _collector(collector),
6440 _span(span),
6441 _bit_map(bit_map),
6442 _mark_stack(mark_stack),
6443 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6444 mark_stack, revisit_stack, concurrent_precleaning),
6445 _yield(should_yield),
6446 _concurrent_precleaning(concurrent_precleaning),
6447 _freelistLock(NULL)
6448 {
6449 _ref_processor = rp;
6450 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6451 }
6453 // This closure is used to mark refs into the CMS generation at the
6454 // second (final) checkpoint, and to scan and transitively follow
6455 // the unmarked oops. It is also used during the concurrent precleaning
6456 // phase while scanning objects on dirty cards in the CMS generation.
6457 // The marks are made in the marking bit map and the marking stack is
6458 // used for keeping the (newly) grey objects during the scan.
6459 // The parallel version (Par_...) appears further below.
6460 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6461 if (obj != NULL) {
6462 assert(obj->is_oop(), "expected an oop");
6463 HeapWord* addr = (HeapWord*)obj;
6464 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6465 assert(_collector->overflow_list_is_empty(),
6466 "overflow list should be empty");
6467 if (_span.contains(addr) &&
6468 !_bit_map->isMarked(addr)) {
6469 // mark bit map (object is now grey)
6470 _bit_map->mark(addr);
6471 // push on marking stack (stack should be empty), and drain the
6472 // stack by applying this closure to the oops in the oops popped
6473 // from the stack (i.e. blacken the grey objects)
6474 bool res = _mark_stack->push(obj);
6475 assert(res, "Should have space to push on empty stack");
6476 do {
6477 oop new_oop = _mark_stack->pop();
6478 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6479 assert(new_oop->is_parsable(), "Found unparsable oop");
6480 assert(_bit_map->isMarked((HeapWord*)new_oop),
6481 "only grey objects on this stack");
6482 // iterate over the oops in this oop, marking and pushing
6483 // the ones in CMS heap (i.e. in _span).
6484 new_oop->oop_iterate(&_pushAndMarkClosure);
6485 // check if it's time to yield
6486 do_yield_check();
6487 } while (!_mark_stack->isEmpty() ||
6488 (!_concurrent_precleaning && take_from_overflow_list()));
6489 // if marking stack is empty, and we are not doing this
6490 // during precleaning, then check the overflow list
6491 }
6492 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6493 assert(_collector->overflow_list_is_empty(),
6494 "overflow list was drained above");
6495 // We could restore evacuated mark words, if any, used for
6496 // overflow list links here because the overflow list is
6497 // provably empty here. That would reduce the maximum
6498 // size requirements for preserved_{oop,mark}_stack.
6499 // But we'll just postpone it until we are all done
6500 // so we can just stream through.
6501 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6502 _collector->restore_preserved_marks_if_any();
6503 assert(_collector->no_preserved_marks(), "No preserved marks");
6504 }
6505 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6506 "All preserved marks should have been restored above");
6507 }
6508 }
6510 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6511 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6513 void MarkRefsIntoAndScanClosure::do_yield_work() {
6514 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6515 "CMS thread should hold CMS token");
6516 assert_lock_strong(_freelistLock);
6517 assert_lock_strong(_bit_map->lock());
6518 // relinquish the free_list_lock and bitMaplock()
6519 _bit_map->lock()->unlock();
6520 _freelistLock->unlock();
6521 ConcurrentMarkSweepThread::desynchronize(true);
6522 ConcurrentMarkSweepThread::acknowledge_yield_request();
6523 _collector->stopTimer();
6524 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6525 if (PrintCMSStatistics != 0) {
6526 _collector->incrementYields();
6527 }
6528 _collector->icms_wait();
6530 // See the comment in coordinator_yield()
6531 for (unsigned i = 0;
6532 i < CMSYieldSleepCount &&
6533 ConcurrentMarkSweepThread::should_yield() &&
6534 !CMSCollector::foregroundGCIsActive();
6535 ++i) {
6536 os::sleep(Thread::current(), 1, false);
6537 ConcurrentMarkSweepThread::acknowledge_yield_request();
6538 }
6540 ConcurrentMarkSweepThread::synchronize(true);
6541 _freelistLock->lock_without_safepoint_check();
6542 _bit_map->lock()->lock_without_safepoint_check();
6543 _collector->startTimer();
6544 }
6546 ///////////////////////////////////////////////////////////
6547 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6548 // MarkRefsIntoAndScanClosure
6549 ///////////////////////////////////////////////////////////
6550 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6551 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6552 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6553 _span(span),
6554 _bit_map(bit_map),
6555 _work_queue(work_queue),
6556 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6557 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6558 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6559 revisit_stack)
6560 {
6561 _ref_processor = rp;
6562 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6563 }
6565 // This closure is used to mark refs into the CMS generation at the
6566 // second (final) checkpoint, and to scan and transitively follow
6567 // the unmarked oops. The marks are made in the marking bit map and
6568 // the work_queue is used for keeping the (newly) grey objects during
6569 // the scan phase whence they are also available for stealing by parallel
6570 // threads. Since the marking bit map is shared, updates are
6571 // synchronized (via CAS).
6572 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6573 if (obj != NULL) {
6574 // Ignore mark word because this could be an already marked oop
6575 // that may be chained at the end of the overflow list.
6576 assert(obj->is_oop(), "expected an oop");
6577 HeapWord* addr = (HeapWord*)obj;
6578 if (_span.contains(addr) &&
6579 !_bit_map->isMarked(addr)) {
6580 // mark bit map (object will become grey):
6581 // It is possible for several threads to be
6582 // trying to "claim" this object concurrently;
6583 // the unique thread that succeeds in marking the
6584 // object first will do the subsequent push on
6585 // to the work queue (or overflow list).
6586 if (_bit_map->par_mark(addr)) {
6587 // push on work_queue (which may not be empty), and trim the
6588 // queue to an appropriate length by applying this closure to
6589 // the oops in the oops popped from the stack (i.e. blacken the
6590 // grey objects)
6591 bool res = _work_queue->push(obj);
6592 assert(res, "Low water mark should be less than capacity?");
6593 trim_queue(_low_water_mark);
6594 } // Else, another thread claimed the object
6595 }
6596 }
6597 }
6599 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6600 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6602 // This closure is used to rescan the marked objects on the dirty cards
6603 // in the mod union table and the card table proper.
6604 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6605 oop p, MemRegion mr) {
6607 size_t size = 0;
6608 HeapWord* addr = (HeapWord*)p;
6609 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6610 assert(_span.contains(addr), "we are scanning the CMS generation");
6611 // check if it's time to yield
6612 if (do_yield_check()) {
6613 // We yielded for some foreground stop-world work,
6614 // and we have been asked to abort this ongoing preclean cycle.
6615 return 0;
6616 }
6617 if (_bitMap->isMarked(addr)) {
6618 // it's marked; is it potentially uninitialized?
6619 if (p->klass_or_null() != NULL) {
6620 if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) {
6621 // Signal precleaning to redirty the card since
6622 // the klass pointer is already installed.
6623 assert(size == 0, "Initial value");
6624 } else {
6625 assert(p->is_parsable(), "must be parsable.");
6626 // an initialized object; ignore mark word in verification below
6627 // since we are running concurrent with mutators
6628 assert(p->is_oop(true), "should be an oop");
6629 if (p->is_objArray()) {
6630 // objArrays are precisely marked; restrict scanning
6631 // to dirty cards only.
6632 size = CompactibleFreeListSpace::adjustObjectSize(
6633 p->oop_iterate(_scanningClosure, mr));
6634 } else {
6635 // A non-array may have been imprecisely marked; we need
6636 // to scan object in its entirety.
6637 size = CompactibleFreeListSpace::adjustObjectSize(
6638 p->oop_iterate(_scanningClosure));
6639 }
6640 #ifdef DEBUG
6641 size_t direct_size =
6642 CompactibleFreeListSpace::adjustObjectSize(p->size());
6643 assert(size == direct_size, "Inconsistency in size");
6644 assert(size >= 3, "Necessary for Printezis marks to work");
6645 if (!_bitMap->isMarked(addr+1)) {
6646 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6647 } else {
6648 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6649 assert(_bitMap->isMarked(addr+size-1),
6650 "inconsistent Printezis mark");
6651 }
6652 #endif // DEBUG
6653 }
6654 } else {
6655 // an unitialized object
6656 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6657 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6658 size = pointer_delta(nextOneAddr + 1, addr);
6659 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6660 "alignment problem");
6661 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6662 // will dirty the card when the klass pointer is installed in the
6663 // object (signalling the completion of initialization).
6664 }
6665 } else {
6666 // Either a not yet marked object or an uninitialized object
6667 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6668 // An uninitialized object, skip to the next card, since
6669 // we may not be able to read its P-bits yet.
6670 assert(size == 0, "Initial value");
6671 } else {
6672 // An object not (yet) reached by marking: we merely need to
6673 // compute its size so as to go look at the next block.
6674 assert(p->is_oop(true), "should be an oop");
6675 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6676 }
6677 }
6678 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6679 return size;
6680 }
6682 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6683 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6684 "CMS thread should hold CMS token");
6685 assert_lock_strong(_freelistLock);
6686 assert_lock_strong(_bitMap->lock());
6687 // relinquish the free_list_lock and bitMaplock()
6688 _bitMap->lock()->unlock();
6689 _freelistLock->unlock();
6690 ConcurrentMarkSweepThread::desynchronize(true);
6691 ConcurrentMarkSweepThread::acknowledge_yield_request();
6692 _collector->stopTimer();
6693 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6694 if (PrintCMSStatistics != 0) {
6695 _collector->incrementYields();
6696 }
6697 _collector->icms_wait();
6699 // See the comment in coordinator_yield()
6700 for (unsigned i = 0; i < CMSYieldSleepCount &&
6701 ConcurrentMarkSweepThread::should_yield() &&
6702 !CMSCollector::foregroundGCIsActive(); ++i) {
6703 os::sleep(Thread::current(), 1, false);
6704 ConcurrentMarkSweepThread::acknowledge_yield_request();
6705 }
6707 ConcurrentMarkSweepThread::synchronize(true);
6708 _freelistLock->lock_without_safepoint_check();
6709 _bitMap->lock()->lock_without_safepoint_check();
6710 _collector->startTimer();
6711 }
6714 //////////////////////////////////////////////////////////////////
6715 // SurvivorSpacePrecleanClosure
6716 //////////////////////////////////////////////////////////////////
6717 // This (single-threaded) closure is used to preclean the oops in
6718 // the survivor spaces.
6719 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6721 HeapWord* addr = (HeapWord*)p;
6722 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6723 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6724 assert(p->klass_or_null() != NULL, "object should be initializd");
6725 assert(p->is_parsable(), "must be parsable.");
6726 // an initialized object; ignore mark word in verification below
6727 // since we are running concurrent with mutators
6728 assert(p->is_oop(true), "should be an oop");
6729 // Note that we do not yield while we iterate over
6730 // the interior oops of p, pushing the relevant ones
6731 // on our marking stack.
6732 size_t size = p->oop_iterate(_scanning_closure);
6733 do_yield_check();
6734 // Observe that below, we do not abandon the preclean
6735 // phase as soon as we should; rather we empty the
6736 // marking stack before returning. This is to satisfy
6737 // some existing assertions. In general, it may be a
6738 // good idea to abort immediately and complete the marking
6739 // from the grey objects at a later time.
6740 while (!_mark_stack->isEmpty()) {
6741 oop new_oop = _mark_stack->pop();
6742 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6743 assert(new_oop->is_parsable(), "Found unparsable oop");
6744 assert(_bit_map->isMarked((HeapWord*)new_oop),
6745 "only grey objects on this stack");
6746 // iterate over the oops in this oop, marking and pushing
6747 // the ones in CMS heap (i.e. in _span).
6748 new_oop->oop_iterate(_scanning_closure);
6749 // check if it's time to yield
6750 do_yield_check();
6751 }
6752 unsigned int after_count =
6753 GenCollectedHeap::heap()->total_collections();
6754 bool abort = (_before_count != after_count) ||
6755 _collector->should_abort_preclean();
6756 return abort ? 0 : size;
6757 }
6759 void SurvivorSpacePrecleanClosure::do_yield_work() {
6760 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6761 "CMS thread should hold CMS token");
6762 assert_lock_strong(_bit_map->lock());
6763 // Relinquish the bit map lock
6764 _bit_map->lock()->unlock();
6765 ConcurrentMarkSweepThread::desynchronize(true);
6766 ConcurrentMarkSweepThread::acknowledge_yield_request();
6767 _collector->stopTimer();
6768 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6769 if (PrintCMSStatistics != 0) {
6770 _collector->incrementYields();
6771 }
6772 _collector->icms_wait();
6774 // See the comment in coordinator_yield()
6775 for (unsigned i = 0; i < CMSYieldSleepCount &&
6776 ConcurrentMarkSweepThread::should_yield() &&
6777 !CMSCollector::foregroundGCIsActive(); ++i) {
6778 os::sleep(Thread::current(), 1, false);
6779 ConcurrentMarkSweepThread::acknowledge_yield_request();
6780 }
6782 ConcurrentMarkSweepThread::synchronize(true);
6783 _bit_map->lock()->lock_without_safepoint_check();
6784 _collector->startTimer();
6785 }
6787 // This closure is used to rescan the marked objects on the dirty cards
6788 // in the mod union table and the card table proper. In the parallel
6789 // case, although the bitMap is shared, we do a single read so the
6790 // isMarked() query is "safe".
6791 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6792 // Ignore mark word because we are running concurrent with mutators
6793 assert(p->is_oop_or_null(true), "expected an oop or null");
6794 HeapWord* addr = (HeapWord*)p;
6795 assert(_span.contains(addr), "we are scanning the CMS generation");
6796 bool is_obj_array = false;
6797 #ifdef DEBUG
6798 if (!_parallel) {
6799 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6800 assert(_collector->overflow_list_is_empty(),
6801 "overflow list should be empty");
6803 }
6804 #endif // DEBUG
6805 if (_bit_map->isMarked(addr)) {
6806 // Obj arrays are precisely marked, non-arrays are not;
6807 // so we scan objArrays precisely and non-arrays in their
6808 // entirety.
6809 if (p->is_objArray()) {
6810 is_obj_array = true;
6811 if (_parallel) {
6812 p->oop_iterate(_par_scan_closure, mr);
6813 } else {
6814 p->oop_iterate(_scan_closure, mr);
6815 }
6816 } else {
6817 if (_parallel) {
6818 p->oop_iterate(_par_scan_closure);
6819 } else {
6820 p->oop_iterate(_scan_closure);
6821 }
6822 }
6823 }
6824 #ifdef DEBUG
6825 if (!_parallel) {
6826 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6827 assert(_collector->overflow_list_is_empty(),
6828 "overflow list should be empty");
6830 }
6831 #endif // DEBUG
6832 return is_obj_array;
6833 }
6835 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6836 MemRegion span,
6837 CMSBitMap* bitMap, CMSMarkStack* markStack,
6838 CMSMarkStack* revisitStack,
6839 bool should_yield, bool verifying):
6840 _collector(collector),
6841 _span(span),
6842 _bitMap(bitMap),
6843 _mut(&collector->_modUnionTable),
6844 _markStack(markStack),
6845 _revisitStack(revisitStack),
6846 _yield(should_yield),
6847 _skipBits(0)
6848 {
6849 assert(_markStack->isEmpty(), "stack should be empty");
6850 _finger = _bitMap->startWord();
6851 _threshold = _finger;
6852 assert(_collector->_restart_addr == NULL, "Sanity check");
6853 assert(_span.contains(_finger), "Out of bounds _finger?");
6854 DEBUG_ONLY(_verifying = verifying;)
6855 }
6857 void MarkFromRootsClosure::reset(HeapWord* addr) {
6858 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6859 assert(_span.contains(addr), "Out of bounds _finger?");
6860 _finger = addr;
6861 _threshold = (HeapWord*)round_to(
6862 (intptr_t)_finger, CardTableModRefBS::card_size);
6863 }
6865 // Should revisit to see if this should be restructured for
6866 // greater efficiency.
6867 bool MarkFromRootsClosure::do_bit(size_t offset) {
6868 if (_skipBits > 0) {
6869 _skipBits--;
6870 return true;
6871 }
6872 // convert offset into a HeapWord*
6873 HeapWord* addr = _bitMap->startWord() + offset;
6874 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6875 "address out of range");
6876 assert(_bitMap->isMarked(addr), "tautology");
6877 if (_bitMap->isMarked(addr+1)) {
6878 // this is an allocated but not yet initialized object
6879 assert(_skipBits == 0, "tautology");
6880 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6881 oop p = oop(addr);
6882 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6883 DEBUG_ONLY(if (!_verifying) {)
6884 // We re-dirty the cards on which this object lies and increase
6885 // the _threshold so that we'll come back to scan this object
6886 // during the preclean or remark phase. (CMSCleanOnEnter)
6887 if (CMSCleanOnEnter) {
6888 size_t sz = _collector->block_size_using_printezis_bits(addr);
6889 HeapWord* start_card_addr = (HeapWord*)round_down(
6890 (intptr_t)addr, CardTableModRefBS::card_size);
6891 HeapWord* end_card_addr = (HeapWord*)round_to(
6892 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6893 MemRegion redirty_range = MemRegion(start_card_addr, end_card_addr);
6894 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6895 // Bump _threshold to end_card_addr; note that
6896 // _threshold cannot possibly exceed end_card_addr, anyhow.
6897 // This prevents future clearing of the card as the scan proceeds
6898 // to the right.
6899 assert(_threshold <= end_card_addr,
6900 "Because we are just scanning into this object");
6901 if (_threshold < end_card_addr) {
6902 _threshold = end_card_addr;
6903 }
6904 if (p->klass_or_null() != NULL) {
6905 // Redirty the range of cards...
6906 _mut->mark_range(redirty_range);
6907 } // ...else the setting of klass will dirty the card anyway.
6908 }
6909 DEBUG_ONLY(})
6910 return true;
6911 }
6912 }
6913 scanOopsInOop(addr);
6914 return true;
6915 }
6917 // We take a break if we've been at this for a while,
6918 // so as to avoid monopolizing the locks involved.
6919 void MarkFromRootsClosure::do_yield_work() {
6920 // First give up the locks, then yield, then re-lock
6921 // We should probably use a constructor/destructor idiom to
6922 // do this unlock/lock or modify the MutexUnlocker class to
6923 // serve our purpose. XXX
6924 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6925 "CMS thread should hold CMS token");
6926 assert_lock_strong(_bitMap->lock());
6927 _bitMap->lock()->unlock();
6928 ConcurrentMarkSweepThread::desynchronize(true);
6929 ConcurrentMarkSweepThread::acknowledge_yield_request();
6930 _collector->stopTimer();
6931 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6932 if (PrintCMSStatistics != 0) {
6933 _collector->incrementYields();
6934 }
6935 _collector->icms_wait();
6937 // See the comment in coordinator_yield()
6938 for (unsigned i = 0; i < CMSYieldSleepCount &&
6939 ConcurrentMarkSweepThread::should_yield() &&
6940 !CMSCollector::foregroundGCIsActive(); ++i) {
6941 os::sleep(Thread::current(), 1, false);
6942 ConcurrentMarkSweepThread::acknowledge_yield_request();
6943 }
6945 ConcurrentMarkSweepThread::synchronize(true);
6946 _bitMap->lock()->lock_without_safepoint_check();
6947 _collector->startTimer();
6948 }
6950 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6951 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6952 assert(_markStack->isEmpty(),
6953 "should drain stack to limit stack usage");
6954 // convert ptr to an oop preparatory to scanning
6955 oop obj = oop(ptr);
6956 // Ignore mark word in verification below, since we
6957 // may be running concurrent with mutators.
6958 assert(obj->is_oop(true), "should be an oop");
6959 assert(_finger <= ptr, "_finger runneth ahead");
6960 // advance the finger to right end of this object
6961 _finger = ptr + obj->size();
6962 assert(_finger > ptr, "we just incremented it above");
6963 // On large heaps, it may take us some time to get through
6964 // the marking phase (especially if running iCMS). During
6965 // this time it's possible that a lot of mutations have
6966 // accumulated in the card table and the mod union table --
6967 // these mutation records are redundant until we have
6968 // actually traced into the corresponding card.
6969 // Here, we check whether advancing the finger would make
6970 // us cross into a new card, and if so clear corresponding
6971 // cards in the MUT (preclean them in the card-table in the
6972 // future).
6974 DEBUG_ONLY(if (!_verifying) {)
6975 // The clean-on-enter optimization is disabled by default,
6976 // until we fix 6178663.
6977 if (CMSCleanOnEnter && (_finger > _threshold)) {
6978 // [_threshold, _finger) represents the interval
6979 // of cards to be cleared in MUT (or precleaned in card table).
6980 // The set of cards to be cleared is all those that overlap
6981 // with the interval [_threshold, _finger); note that
6982 // _threshold is always kept card-aligned but _finger isn't
6983 // always card-aligned.
6984 HeapWord* old_threshold = _threshold;
6985 assert(old_threshold == (HeapWord*)round_to(
6986 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6987 "_threshold should always be card-aligned");
6988 _threshold = (HeapWord*)round_to(
6989 (intptr_t)_finger, CardTableModRefBS::card_size);
6990 MemRegion mr(old_threshold, _threshold);
6991 assert(!mr.is_empty(), "Control point invariant");
6992 assert(_span.contains(mr), "Should clear within span");
6993 // XXX When _finger crosses from old gen into perm gen
6994 // we may be doing unnecessary cleaning; do better in the
6995 // future by detecting that condition and clearing fewer
6996 // MUT/CT entries.
6997 _mut->clear_range(mr);
6998 }
6999 DEBUG_ONLY(})
7001 // Note: the finger doesn't advance while we drain
7002 // the stack below.
7003 PushOrMarkClosure pushOrMarkClosure(_collector,
7004 _span, _bitMap, _markStack,
7005 _revisitStack,
7006 _finger, this);
7007 bool res = _markStack->push(obj);
7008 assert(res, "Empty non-zero size stack should have space for single push");
7009 while (!_markStack->isEmpty()) {
7010 oop new_oop = _markStack->pop();
7011 // Skip verifying header mark word below because we are
7012 // running concurrent with mutators.
7013 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7014 // now scan this oop's oops
7015 new_oop->oop_iterate(&pushOrMarkClosure);
7016 do_yield_check();
7017 }
7018 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7019 }
7021 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7022 CMSCollector* collector, MemRegion span,
7023 CMSBitMap* bit_map,
7024 OopTaskQueue* work_queue,
7025 CMSMarkStack* overflow_stack,
7026 CMSMarkStack* revisit_stack,
7027 bool should_yield):
7028 _collector(collector),
7029 _whole_span(collector->_span),
7030 _span(span),
7031 _bit_map(bit_map),
7032 _mut(&collector->_modUnionTable),
7033 _work_queue(work_queue),
7034 _overflow_stack(overflow_stack),
7035 _revisit_stack(revisit_stack),
7036 _yield(should_yield),
7037 _skip_bits(0),
7038 _task(task)
7039 {
7040 assert(_work_queue->size() == 0, "work_queue should be empty");
7041 _finger = span.start();
7042 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7043 assert(_span.contains(_finger), "Out of bounds _finger?");
7044 }
7046 // Should revisit to see if this should be restructured for
7047 // greater efficiency.
7048 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7049 if (_skip_bits > 0) {
7050 _skip_bits--;
7051 return true;
7052 }
7053 // convert offset into a HeapWord*
7054 HeapWord* addr = _bit_map->startWord() + offset;
7055 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7056 "address out of range");
7057 assert(_bit_map->isMarked(addr), "tautology");
7058 if (_bit_map->isMarked(addr+1)) {
7059 // this is an allocated object that might not yet be initialized
7060 assert(_skip_bits == 0, "tautology");
7061 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7062 oop p = oop(addr);
7063 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7064 // in the case of Clean-on-Enter optimization, redirty card
7065 // and avoid clearing card by increasing the threshold.
7066 return true;
7067 }
7068 }
7069 scan_oops_in_oop(addr);
7070 return true;
7071 }
7073 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7074 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7075 // Should we assert that our work queue is empty or
7076 // below some drain limit?
7077 assert(_work_queue->size() == 0,
7078 "should drain stack to limit stack usage");
7079 // convert ptr to an oop preparatory to scanning
7080 oop obj = oop(ptr);
7081 // Ignore mark word in verification below, since we
7082 // may be running concurrent with mutators.
7083 assert(obj->is_oop(true), "should be an oop");
7084 assert(_finger <= ptr, "_finger runneth ahead");
7085 // advance the finger to right end of this object
7086 _finger = ptr + obj->size();
7087 assert(_finger > ptr, "we just incremented it above");
7088 // On large heaps, it may take us some time to get through
7089 // the marking phase (especially if running iCMS). During
7090 // this time it's possible that a lot of mutations have
7091 // accumulated in the card table and the mod union table --
7092 // these mutation records are redundant until we have
7093 // actually traced into the corresponding card.
7094 // Here, we check whether advancing the finger would make
7095 // us cross into a new card, and if so clear corresponding
7096 // cards in the MUT (preclean them in the card-table in the
7097 // future).
7099 // The clean-on-enter optimization is disabled by default,
7100 // until we fix 6178663.
7101 if (CMSCleanOnEnter && (_finger > _threshold)) {
7102 // [_threshold, _finger) represents the interval
7103 // of cards to be cleared in MUT (or precleaned in card table).
7104 // The set of cards to be cleared is all those that overlap
7105 // with the interval [_threshold, _finger); note that
7106 // _threshold is always kept card-aligned but _finger isn't
7107 // always card-aligned.
7108 HeapWord* old_threshold = _threshold;
7109 assert(old_threshold == (HeapWord*)round_to(
7110 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7111 "_threshold should always be card-aligned");
7112 _threshold = (HeapWord*)round_to(
7113 (intptr_t)_finger, CardTableModRefBS::card_size);
7114 MemRegion mr(old_threshold, _threshold);
7115 assert(!mr.is_empty(), "Control point invariant");
7116 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7117 // XXX When _finger crosses from old gen into perm gen
7118 // we may be doing unnecessary cleaning; do better in the
7119 // future by detecting that condition and clearing fewer
7120 // MUT/CT entries.
7121 _mut->clear_range(mr);
7122 }
7124 // Note: the local finger doesn't advance while we drain
7125 // the stack below, but the global finger sure can and will.
7126 HeapWord** gfa = _task->global_finger_addr();
7127 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7128 _span, _bit_map,
7129 _work_queue,
7130 _overflow_stack,
7131 _revisit_stack,
7132 _finger,
7133 gfa, this);
7134 bool res = _work_queue->push(obj); // overflow could occur here
7135 assert(res, "Will hold once we use workqueues");
7136 while (true) {
7137 oop new_oop;
7138 if (!_work_queue->pop_local(new_oop)) {
7139 // We emptied our work_queue; check if there's stuff that can
7140 // be gotten from the overflow stack.
7141 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7142 _overflow_stack, _work_queue)) {
7143 do_yield_check();
7144 continue;
7145 } else { // done
7146 break;
7147 }
7148 }
7149 // Skip verifying header mark word below because we are
7150 // running concurrent with mutators.
7151 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7152 // now scan this oop's oops
7153 new_oop->oop_iterate(&pushOrMarkClosure);
7154 do_yield_check();
7155 }
7156 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7157 }
7159 // Yield in response to a request from VM Thread or
7160 // from mutators.
7161 void Par_MarkFromRootsClosure::do_yield_work() {
7162 assert(_task != NULL, "sanity");
7163 _task->yield();
7164 }
7166 // A variant of the above used for verifying CMS marking work.
7167 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7168 MemRegion span,
7169 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7170 CMSMarkStack* mark_stack):
7171 _collector(collector),
7172 _span(span),
7173 _verification_bm(verification_bm),
7174 _cms_bm(cms_bm),
7175 _mark_stack(mark_stack),
7176 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7177 mark_stack)
7178 {
7179 assert(_mark_stack->isEmpty(), "stack should be empty");
7180 _finger = _verification_bm->startWord();
7181 assert(_collector->_restart_addr == NULL, "Sanity check");
7182 assert(_span.contains(_finger), "Out of bounds _finger?");
7183 }
7185 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7186 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7187 assert(_span.contains(addr), "Out of bounds _finger?");
7188 _finger = addr;
7189 }
7191 // Should revisit to see if this should be restructured for
7192 // greater efficiency.
7193 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7194 // convert offset into a HeapWord*
7195 HeapWord* addr = _verification_bm->startWord() + offset;
7196 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7197 "address out of range");
7198 assert(_verification_bm->isMarked(addr), "tautology");
7199 assert(_cms_bm->isMarked(addr), "tautology");
7201 assert(_mark_stack->isEmpty(),
7202 "should drain stack to limit stack usage");
7203 // convert addr to an oop preparatory to scanning
7204 oop obj = oop(addr);
7205 assert(obj->is_oop(), "should be an oop");
7206 assert(_finger <= addr, "_finger runneth ahead");
7207 // advance the finger to right end of this object
7208 _finger = addr + obj->size();
7209 assert(_finger > addr, "we just incremented it above");
7210 // Note: the finger doesn't advance while we drain
7211 // the stack below.
7212 bool res = _mark_stack->push(obj);
7213 assert(res, "Empty non-zero size stack should have space for single push");
7214 while (!_mark_stack->isEmpty()) {
7215 oop new_oop = _mark_stack->pop();
7216 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7217 // now scan this oop's oops
7218 new_oop->oop_iterate(&_pam_verify_closure);
7219 }
7220 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7221 return true;
7222 }
7224 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7225 CMSCollector* collector, MemRegion span,
7226 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7227 CMSMarkStack* mark_stack):
7228 OopClosure(collector->ref_processor()),
7229 _collector(collector),
7230 _span(span),
7231 _verification_bm(verification_bm),
7232 _cms_bm(cms_bm),
7233 _mark_stack(mark_stack)
7234 { }
7236 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7237 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7239 // Upon stack overflow, we discard (part of) the stack,
7240 // remembering the least address amongst those discarded
7241 // in CMSCollector's _restart_address.
7242 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7243 // Remember the least grey address discarded
7244 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7245 _collector->lower_restart_addr(ra);
7246 _mark_stack->reset(); // discard stack contents
7247 _mark_stack->expand(); // expand the stack if possible
7248 }
7250 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7251 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7252 HeapWord* addr = (HeapWord*)obj;
7253 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7254 // Oop lies in _span and isn't yet grey or black
7255 _verification_bm->mark(addr); // now grey
7256 if (!_cms_bm->isMarked(addr)) {
7257 oop(addr)->print();
7258 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7259 addr);
7260 fatal("... aborting");
7261 }
7263 if (!_mark_stack->push(obj)) { // stack overflow
7264 if (PrintCMSStatistics != 0) {
7265 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7266 SIZE_FORMAT, _mark_stack->capacity());
7267 }
7268 assert(_mark_stack->isFull(), "Else push should have succeeded");
7269 handle_stack_overflow(addr);
7270 }
7271 // anything including and to the right of _finger
7272 // will be scanned as we iterate over the remainder of the
7273 // bit map
7274 }
7275 }
7277 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7278 MemRegion span,
7279 CMSBitMap* bitMap, CMSMarkStack* markStack,
7280 CMSMarkStack* revisitStack,
7281 HeapWord* finger, MarkFromRootsClosure* parent) :
7282 OopClosure(collector->ref_processor()),
7283 _collector(collector),
7284 _span(span),
7285 _bitMap(bitMap),
7286 _markStack(markStack),
7287 _revisitStack(revisitStack),
7288 _finger(finger),
7289 _parent(parent),
7290 _should_remember_klasses(collector->should_unload_classes())
7291 { }
7293 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7294 MemRegion span,
7295 CMSBitMap* bit_map,
7296 OopTaskQueue* work_queue,
7297 CMSMarkStack* overflow_stack,
7298 CMSMarkStack* revisit_stack,
7299 HeapWord* finger,
7300 HeapWord** global_finger_addr,
7301 Par_MarkFromRootsClosure* parent) :
7302 OopClosure(collector->ref_processor()),
7303 _collector(collector),
7304 _whole_span(collector->_span),
7305 _span(span),
7306 _bit_map(bit_map),
7307 _work_queue(work_queue),
7308 _overflow_stack(overflow_stack),
7309 _revisit_stack(revisit_stack),
7310 _finger(finger),
7311 _global_finger_addr(global_finger_addr),
7312 _parent(parent),
7313 _should_remember_klasses(collector->should_unload_classes())
7314 { }
7316 void CMSCollector::lower_restart_addr(HeapWord* low) {
7317 assert(_span.contains(low), "Out of bounds addr");
7318 if (_restart_addr == NULL) {
7319 _restart_addr = low;
7320 } else {
7321 _restart_addr = MIN2(_restart_addr, low);
7322 }
7323 }
7325 // Upon stack overflow, we discard (part of) the stack,
7326 // remembering the least address amongst those discarded
7327 // in CMSCollector's _restart_address.
7328 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7329 // Remember the least grey address discarded
7330 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7331 _collector->lower_restart_addr(ra);
7332 _markStack->reset(); // discard stack contents
7333 _markStack->expand(); // expand the stack if possible
7334 }
7336 // Upon stack overflow, we discard (part of) the stack,
7337 // remembering the least address amongst those discarded
7338 // in CMSCollector's _restart_address.
7339 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7340 // We need to do this under a mutex to prevent other
7341 // workers from interfering with the expansion below.
7342 MutexLockerEx ml(_overflow_stack->par_lock(),
7343 Mutex::_no_safepoint_check_flag);
7344 // Remember the least grey address discarded
7345 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7346 _collector->lower_restart_addr(ra);
7347 _overflow_stack->reset(); // discard stack contents
7348 _overflow_stack->expand(); // expand the stack if possible
7349 }
7351 void PushOrMarkClosure::do_oop(oop obj) {
7352 // Ignore mark word because we are running concurrent with mutators.
7353 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7354 HeapWord* addr = (HeapWord*)obj;
7355 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7356 // Oop lies in _span and isn't yet grey or black
7357 _bitMap->mark(addr); // now grey
7358 if (addr < _finger) {
7359 // the bit map iteration has already either passed, or
7360 // sampled, this bit in the bit map; we'll need to
7361 // use the marking stack to scan this oop's oops.
7362 bool simulate_overflow = false;
7363 NOT_PRODUCT(
7364 if (CMSMarkStackOverflowALot &&
7365 _collector->simulate_overflow()) {
7366 // simulate a stack overflow
7367 simulate_overflow = true;
7368 }
7369 )
7370 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7371 if (PrintCMSStatistics != 0) {
7372 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7373 SIZE_FORMAT, _markStack->capacity());
7374 }
7375 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7376 handle_stack_overflow(addr);
7377 }
7378 }
7379 // anything including and to the right of _finger
7380 // will be scanned as we iterate over the remainder of the
7381 // bit map
7382 do_yield_check();
7383 }
7384 }
7386 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7387 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7389 void Par_PushOrMarkClosure::do_oop(oop obj) {
7390 // Ignore mark word because we are running concurrent with mutators.
7391 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7392 HeapWord* addr = (HeapWord*)obj;
7393 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7394 // Oop lies in _span and isn't yet grey or black
7395 // We read the global_finger (volatile read) strictly after marking oop
7396 bool res = _bit_map->par_mark(addr); // now grey
7397 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7398 // Should we push this marked oop on our stack?
7399 // -- if someone else marked it, nothing to do
7400 // -- if target oop is above global finger nothing to do
7401 // -- if target oop is in chunk and above local finger
7402 // then nothing to do
7403 // -- else push on work queue
7404 if ( !res // someone else marked it, they will deal with it
7405 || (addr >= *gfa) // will be scanned in a later task
7406 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7407 return;
7408 }
7409 // the bit map iteration has already either passed, or
7410 // sampled, this bit in the bit map; we'll need to
7411 // use the marking stack to scan this oop's oops.
7412 bool simulate_overflow = false;
7413 NOT_PRODUCT(
7414 if (CMSMarkStackOverflowALot &&
7415 _collector->simulate_overflow()) {
7416 // simulate a stack overflow
7417 simulate_overflow = true;
7418 }
7419 )
7420 if (simulate_overflow ||
7421 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7422 // stack overflow
7423 if (PrintCMSStatistics != 0) {
7424 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7425 SIZE_FORMAT, _overflow_stack->capacity());
7426 }
7427 // We cannot assert that the overflow stack is full because
7428 // it may have been emptied since.
7429 assert(simulate_overflow ||
7430 _work_queue->size() == _work_queue->max_elems(),
7431 "Else push should have succeeded");
7432 handle_stack_overflow(addr);
7433 }
7434 do_yield_check();
7435 }
7436 }
7438 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7439 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7441 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7442 MemRegion span,
7443 ReferenceProcessor* rp,
7444 CMSBitMap* bit_map,
7445 CMSBitMap* mod_union_table,
7446 CMSMarkStack* mark_stack,
7447 CMSMarkStack* revisit_stack,
7448 bool concurrent_precleaning):
7449 OopClosure(rp),
7450 _collector(collector),
7451 _span(span),
7452 _bit_map(bit_map),
7453 _mod_union_table(mod_union_table),
7454 _mark_stack(mark_stack),
7455 _revisit_stack(revisit_stack),
7456 _concurrent_precleaning(concurrent_precleaning),
7457 _should_remember_klasses(collector->should_unload_classes())
7458 {
7459 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7460 }
7462 // Grey object rescan during pre-cleaning and second checkpoint phases --
7463 // the non-parallel version (the parallel version appears further below.)
7464 void PushAndMarkClosure::do_oop(oop obj) {
7465 // Ignore mark word verification. If during concurrent precleaning,
7466 // the object monitor may be locked. If during the checkpoint
7467 // phases, the object may already have been reached by a different
7468 // path and may be at the end of the global overflow list (so
7469 // the mark word may be NULL).
7470 assert(obj->is_oop_or_null(true /* ignore mark word */),
7471 "expected an oop or NULL");
7472 HeapWord* addr = (HeapWord*)obj;
7473 // Check if oop points into the CMS generation
7474 // and is not marked
7475 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7476 // a white object ...
7477 _bit_map->mark(addr); // ... now grey
7478 // push on the marking stack (grey set)
7479 bool simulate_overflow = false;
7480 NOT_PRODUCT(
7481 if (CMSMarkStackOverflowALot &&
7482 _collector->simulate_overflow()) {
7483 // simulate a stack overflow
7484 simulate_overflow = true;
7485 }
7486 )
7487 if (simulate_overflow || !_mark_stack->push(obj)) {
7488 if (_concurrent_precleaning) {
7489 // During precleaning we can just dirty the appropriate card
7490 // in the mod union table, thus ensuring that the object remains
7491 // in the grey set and continue. Note that no one can be intefering
7492 // with us in this action of dirtying the mod union table, so
7493 // no locking is required.
7494 _mod_union_table->mark(addr);
7495 _collector->_ser_pmc_preclean_ovflw++;
7496 } else {
7497 // During the remark phase, we need to remember this oop
7498 // in the overflow list.
7499 _collector->push_on_overflow_list(obj);
7500 _collector->_ser_pmc_remark_ovflw++;
7501 }
7502 }
7503 }
7504 }
7506 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7507 MemRegion span,
7508 ReferenceProcessor* rp,
7509 CMSBitMap* bit_map,
7510 OopTaskQueue* work_queue,
7511 CMSMarkStack* revisit_stack):
7512 OopClosure(rp),
7513 _collector(collector),
7514 _span(span),
7515 _bit_map(bit_map),
7516 _work_queue(work_queue),
7517 _revisit_stack(revisit_stack),
7518 _should_remember_klasses(collector->should_unload_classes())
7519 {
7520 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7521 }
7523 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7524 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7526 // Grey object rescan during second checkpoint phase --
7527 // the parallel version.
7528 void Par_PushAndMarkClosure::do_oop(oop obj) {
7529 // In the assert below, we ignore the mark word because
7530 // this oop may point to an already visited object that is
7531 // on the overflow stack (in which case the mark word has
7532 // been hijacked for chaining into the overflow stack --
7533 // if this is the last object in the overflow stack then
7534 // its mark word will be NULL). Because this object may
7535 // have been subsequently popped off the global overflow
7536 // stack, and the mark word possibly restored to the prototypical
7537 // value, by the time we get to examined this failing assert in
7538 // the debugger, is_oop_or_null(false) may subsequently start
7539 // to hold.
7540 assert(obj->is_oop_or_null(true),
7541 "expected an oop or NULL");
7542 HeapWord* addr = (HeapWord*)obj;
7543 // Check if oop points into the CMS generation
7544 // and is not marked
7545 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7546 // a white object ...
7547 // If we manage to "claim" the object, by being the
7548 // first thread to mark it, then we push it on our
7549 // marking stack
7550 if (_bit_map->par_mark(addr)) { // ... now grey
7551 // push on work queue (grey set)
7552 bool simulate_overflow = false;
7553 NOT_PRODUCT(
7554 if (CMSMarkStackOverflowALot &&
7555 _collector->par_simulate_overflow()) {
7556 // simulate a stack overflow
7557 simulate_overflow = true;
7558 }
7559 )
7560 if (simulate_overflow || !_work_queue->push(obj)) {
7561 _collector->par_push_on_overflow_list(obj);
7562 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7563 }
7564 } // Else, some other thread got there first
7565 }
7566 }
7568 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7569 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7571 void PushAndMarkClosure::remember_klass(Klass* k) {
7572 if (!_revisit_stack->push(oop(k))) {
7573 fatal("Revisit stack overflowed in PushAndMarkClosure");
7574 }
7575 }
7577 void Par_PushAndMarkClosure::remember_klass(Klass* k) {
7578 if (!_revisit_stack->par_push(oop(k))) {
7579 fatal("Revist stack overflowed in Par_PushAndMarkClosure");
7580 }
7581 }
7583 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7584 Mutex* bml = _collector->bitMapLock();
7585 assert_lock_strong(bml);
7586 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7587 "CMS thread should hold CMS token");
7589 bml->unlock();
7590 ConcurrentMarkSweepThread::desynchronize(true);
7592 ConcurrentMarkSweepThread::acknowledge_yield_request();
7594 _collector->stopTimer();
7595 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7596 if (PrintCMSStatistics != 0) {
7597 _collector->incrementYields();
7598 }
7599 _collector->icms_wait();
7601 // See the comment in coordinator_yield()
7602 for (unsigned i = 0; i < CMSYieldSleepCount &&
7603 ConcurrentMarkSweepThread::should_yield() &&
7604 !CMSCollector::foregroundGCIsActive(); ++i) {
7605 os::sleep(Thread::current(), 1, false);
7606 ConcurrentMarkSweepThread::acknowledge_yield_request();
7607 }
7609 ConcurrentMarkSweepThread::synchronize(true);
7610 bml->lock();
7612 _collector->startTimer();
7613 }
7615 bool CMSPrecleanRefsYieldClosure::should_return() {
7616 if (ConcurrentMarkSweepThread::should_yield()) {
7617 do_yield_work();
7618 }
7619 return _collector->foregroundGCIsActive();
7620 }
7622 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7623 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7624 "mr should be aligned to start at a card boundary");
7625 // We'd like to assert:
7626 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7627 // "mr should be a range of cards");
7628 // However, that would be too strong in one case -- the last
7629 // partition ends at _unallocated_block which, in general, can be
7630 // an arbitrary boundary, not necessarily card aligned.
7631 if (PrintCMSStatistics != 0) {
7632 _num_dirty_cards +=
7633 mr.word_size()/CardTableModRefBS::card_size_in_words;
7634 }
7635 _space->object_iterate_mem(mr, &_scan_cl);
7636 }
7638 SweepClosure::SweepClosure(CMSCollector* collector,
7639 ConcurrentMarkSweepGeneration* g,
7640 CMSBitMap* bitMap, bool should_yield) :
7641 _collector(collector),
7642 _g(g),
7643 _sp(g->cmsSpace()),
7644 _limit(_sp->sweep_limit()),
7645 _freelistLock(_sp->freelistLock()),
7646 _bitMap(bitMap),
7647 _yield(should_yield),
7648 _inFreeRange(false), // No free range at beginning of sweep
7649 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7650 _lastFreeRangeCoalesced(false),
7651 _freeFinger(g->used_region().start())
7652 {
7653 NOT_PRODUCT(
7654 _numObjectsFreed = 0;
7655 _numWordsFreed = 0;
7656 _numObjectsLive = 0;
7657 _numWordsLive = 0;
7658 _numObjectsAlreadyFree = 0;
7659 _numWordsAlreadyFree = 0;
7660 _last_fc = NULL;
7662 _sp->initializeIndexedFreeListArrayReturnedBytes();
7663 _sp->dictionary()->initializeDictReturnedBytes();
7664 )
7665 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7666 "sweep _limit out of bounds");
7667 if (CMSTraceSweeper) {
7668 gclog_or_tty->print("\n====================\nStarting new sweep\n");
7669 }
7670 }
7672 // We need this destructor to reclaim any space at the end
7673 // of the space, which do_blk below may not have added back to
7674 // the free lists. [basically dealing with the "fringe effect"]
7675 SweepClosure::~SweepClosure() {
7676 assert_lock_strong(_freelistLock);
7677 // this should be treated as the end of a free run if any
7678 // The current free range should be returned to the free lists
7679 // as one coalesced chunk.
7680 if (inFreeRange()) {
7681 flushCurFreeChunk(freeFinger(),
7682 pointer_delta(_limit, freeFinger()));
7683 assert(freeFinger() < _limit, "the finger pointeth off base");
7684 if (CMSTraceSweeper) {
7685 gclog_or_tty->print("destructor:");
7686 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
7687 "[coalesced:"SIZE_FORMAT"]\n",
7688 freeFinger(), pointer_delta(_limit, freeFinger()),
7689 lastFreeRangeCoalesced());
7690 }
7691 }
7692 NOT_PRODUCT(
7693 if (Verbose && PrintGC) {
7694 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7695 SIZE_FORMAT " bytes",
7696 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7697 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7698 SIZE_FORMAT" bytes "
7699 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7700 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7701 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7702 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7703 sizeof(HeapWord);
7704 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7706 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7707 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7708 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7709 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7710 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7711 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7712 indexListReturnedBytes);
7713 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7714 dictReturnedBytes);
7715 }
7716 }
7717 )
7718 // Now, in debug mode, just null out the sweep_limit
7719 NOT_PRODUCT(_sp->clear_sweep_limit();)
7720 if (CMSTraceSweeper) {
7721 gclog_or_tty->print("end of sweep\n================\n");
7722 }
7723 }
7725 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7726 bool freeRangeInFreeLists) {
7727 if (CMSTraceSweeper) {
7728 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
7729 freeFinger, _sp->block_size(freeFinger),
7730 freeRangeInFreeLists);
7731 }
7732 assert(!inFreeRange(), "Trampling existing free range");
7733 set_inFreeRange(true);
7734 set_lastFreeRangeCoalesced(false);
7736 set_freeFinger(freeFinger);
7737 set_freeRangeInFreeLists(freeRangeInFreeLists);
7738 if (CMSTestInFreeList) {
7739 if (freeRangeInFreeLists) {
7740 FreeChunk* fc = (FreeChunk*) freeFinger;
7741 assert(fc->isFree(), "A chunk on the free list should be free.");
7742 assert(fc->size() > 0, "Free range should have a size");
7743 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7744 }
7745 }
7746 }
7748 // Note that the sweeper runs concurrently with mutators. Thus,
7749 // it is possible for direct allocation in this generation to happen
7750 // in the middle of the sweep. Note that the sweeper also coalesces
7751 // contiguous free blocks. Thus, unless the sweeper and the allocator
7752 // synchronize appropriately freshly allocated blocks may get swept up.
7753 // This is accomplished by the sweeper locking the free lists while
7754 // it is sweeping. Thus blocks that are determined to be free are
7755 // indeed free. There is however one additional complication:
7756 // blocks that have been allocated since the final checkpoint and
7757 // mark, will not have been marked and so would be treated as
7758 // unreachable and swept up. To prevent this, the allocator marks
7759 // the bit map when allocating during the sweep phase. This leads,
7760 // however, to a further complication -- objects may have been allocated
7761 // but not yet initialized -- in the sense that the header isn't yet
7762 // installed. The sweeper can not then determine the size of the block
7763 // in order to skip over it. To deal with this case, we use a technique
7764 // (due to Printezis) to encode such uninitialized block sizes in the
7765 // bit map. Since the bit map uses a bit per every HeapWord, but the
7766 // CMS generation has a minimum object size of 3 HeapWords, it follows
7767 // that "normal marks" won't be adjacent in the bit map (there will
7768 // always be at least two 0 bits between successive 1 bits). We make use
7769 // of these "unused" bits to represent uninitialized blocks -- the bit
7770 // corresponding to the start of the uninitialized object and the next
7771 // bit are both set. Finally, a 1 bit marks the end of the object that
7772 // started with the two consecutive 1 bits to indicate its potentially
7773 // uninitialized state.
7775 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7776 FreeChunk* fc = (FreeChunk*)addr;
7777 size_t res;
7779 // check if we are done sweepinrg
7780 if (addr == _limit) { // we have swept up to the limit, do nothing more
7781 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7782 "sweep _limit out of bounds");
7783 // help the closure application finish
7784 return pointer_delta(_sp->end(), _limit);
7785 }
7786 assert(addr <= _limit, "sweep invariant");
7788 // check if we should yield
7789 do_yield_check(addr);
7790 if (fc->isFree()) {
7791 // Chunk that is already free
7792 res = fc->size();
7793 doAlreadyFreeChunk(fc);
7794 debug_only(_sp->verifyFreeLists());
7795 assert(res == fc->size(), "Don't expect the size to change");
7796 NOT_PRODUCT(
7797 _numObjectsAlreadyFree++;
7798 _numWordsAlreadyFree += res;
7799 )
7800 NOT_PRODUCT(_last_fc = fc;)
7801 } else if (!_bitMap->isMarked(addr)) {
7802 // Chunk is fresh garbage
7803 res = doGarbageChunk(fc);
7804 debug_only(_sp->verifyFreeLists());
7805 NOT_PRODUCT(
7806 _numObjectsFreed++;
7807 _numWordsFreed += res;
7808 )
7809 } else {
7810 // Chunk that is alive.
7811 res = doLiveChunk(fc);
7812 debug_only(_sp->verifyFreeLists());
7813 NOT_PRODUCT(
7814 _numObjectsLive++;
7815 _numWordsLive += res;
7816 )
7817 }
7818 return res;
7819 }
7821 // For the smart allocation, record following
7822 // split deaths - a free chunk is removed from its free list because
7823 // it is being split into two or more chunks.
7824 // split birth - a free chunk is being added to its free list because
7825 // a larger free chunk has been split and resulted in this free chunk.
7826 // coal death - a free chunk is being removed from its free list because
7827 // it is being coalesced into a large free chunk.
7828 // coal birth - a free chunk is being added to its free list because
7829 // it was created when two or more free chunks where coalesced into
7830 // this free chunk.
7831 //
7832 // These statistics are used to determine the desired number of free
7833 // chunks of a given size. The desired number is chosen to be relative
7834 // to the end of a CMS sweep. The desired number at the end of a sweep
7835 // is the
7836 // count-at-end-of-previous-sweep (an amount that was enough)
7837 // - count-at-beginning-of-current-sweep (the excess)
7838 // + split-births (gains in this size during interval)
7839 // - split-deaths (demands on this size during interval)
7840 // where the interval is from the end of one sweep to the end of the
7841 // next.
7842 //
7843 // When sweeping the sweeper maintains an accumulated chunk which is
7844 // the chunk that is made up of chunks that have been coalesced. That
7845 // will be termed the left-hand chunk. A new chunk of garbage that
7846 // is being considered for coalescing will be referred to as the
7847 // right-hand chunk.
7848 //
7849 // When making a decision on whether to coalesce a right-hand chunk with
7850 // the current left-hand chunk, the current count vs. the desired count
7851 // of the left-hand chunk is considered. Also if the right-hand chunk
7852 // is near the large chunk at the end of the heap (see
7853 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7854 // left-hand chunk is coalesced.
7855 //
7856 // When making a decision about whether to split a chunk, the desired count
7857 // vs. the current count of the candidate to be split is also considered.
7858 // If the candidate is underpopulated (currently fewer chunks than desired)
7859 // a chunk of an overpopulated (currently more chunks than desired) size may
7860 // be chosen. The "hint" associated with a free list, if non-null, points
7861 // to a free list which may be overpopulated.
7862 //
7864 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
7865 size_t size = fc->size();
7866 // Chunks that cannot be coalesced are not in the
7867 // free lists.
7868 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7869 assert(_sp->verifyChunkInFreeLists(fc),
7870 "free chunk should be in free lists");
7871 }
7872 // a chunk that is already free, should not have been
7873 // marked in the bit map
7874 HeapWord* addr = (HeapWord*) fc;
7875 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7876 // Verify that the bit map has no bits marked between
7877 // addr and purported end of this block.
7878 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7880 // Some chunks cannot be coalesced in under any circumstances.
7881 // See the definition of cantCoalesce().
7882 if (!fc->cantCoalesce()) {
7883 // This chunk can potentially be coalesced.
7884 if (_sp->adaptive_freelists()) {
7885 // All the work is done in
7886 doPostIsFreeOrGarbageChunk(fc, size);
7887 } else { // Not adaptive free lists
7888 // this is a free chunk that can potentially be coalesced by the sweeper;
7889 if (!inFreeRange()) {
7890 // if the next chunk is a free block that can't be coalesced
7891 // it doesn't make sense to remove this chunk from the free lists
7892 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7893 assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
7894 if ((HeapWord*)nextChunk < _limit && // there's a next chunk...
7895 nextChunk->isFree() && // which is free...
7896 nextChunk->cantCoalesce()) { // ... but cant be coalesced
7897 // nothing to do
7898 } else {
7899 // Potentially the start of a new free range:
7900 // Don't eagerly remove it from the free lists.
7901 // No need to remove it if it will just be put
7902 // back again. (Also from a pragmatic point of view
7903 // if it is a free block in a region that is beyond
7904 // any allocated blocks, an assertion will fail)
7905 // Remember the start of a free run.
7906 initialize_free_range(addr, true);
7907 // end - can coalesce with next chunk
7908 }
7909 } else {
7910 // the midst of a free range, we are coalescing
7911 debug_only(record_free_block_coalesced(fc);)
7912 if (CMSTraceSweeper) {
7913 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
7914 }
7915 // remove it from the free lists
7916 _sp->removeFreeChunkFromFreeLists(fc);
7917 set_lastFreeRangeCoalesced(true);
7918 // If the chunk is being coalesced and the current free range is
7919 // in the free lists, remove the current free range so that it
7920 // will be returned to the free lists in its entirety - all
7921 // the coalesced pieces included.
7922 if (freeRangeInFreeLists()) {
7923 FreeChunk* ffc = (FreeChunk*) freeFinger();
7924 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7925 "Size of free range is inconsistent with chunk size.");
7926 if (CMSTestInFreeList) {
7927 assert(_sp->verifyChunkInFreeLists(ffc),
7928 "free range is not in free lists");
7929 }
7930 _sp->removeFreeChunkFromFreeLists(ffc);
7931 set_freeRangeInFreeLists(false);
7932 }
7933 }
7934 }
7935 } else {
7936 // Code path common to both original and adaptive free lists.
7938 // cant coalesce with previous block; this should be treated
7939 // as the end of a free run if any
7940 if (inFreeRange()) {
7941 // we kicked some butt; time to pick up the garbage
7942 assert(freeFinger() < addr, "the finger pointeth off base");
7943 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
7944 }
7945 // else, nothing to do, just continue
7946 }
7947 }
7949 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
7950 // This is a chunk of garbage. It is not in any free list.
7951 // Add it to a free list or let it possibly be coalesced into
7952 // a larger chunk.
7953 HeapWord* addr = (HeapWord*) fc;
7954 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7956 if (_sp->adaptive_freelists()) {
7957 // Verify that the bit map has no bits marked between
7958 // addr and purported end of just dead object.
7959 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7961 doPostIsFreeOrGarbageChunk(fc, size);
7962 } else {
7963 if (!inFreeRange()) {
7964 // start of a new free range
7965 assert(size > 0, "A free range should have a size");
7966 initialize_free_range(addr, false);
7968 } else {
7969 // this will be swept up when we hit the end of the
7970 // free range
7971 if (CMSTraceSweeper) {
7972 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
7973 }
7974 // If the chunk is being coalesced and the current free range is
7975 // in the free lists, remove the current free range so that it
7976 // will be returned to the free lists in its entirety - all
7977 // the coalesced pieces included.
7978 if (freeRangeInFreeLists()) {
7979 FreeChunk* ffc = (FreeChunk*)freeFinger();
7980 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7981 "Size of free range is inconsistent with chunk size.");
7982 if (CMSTestInFreeList) {
7983 assert(_sp->verifyChunkInFreeLists(ffc),
7984 "free range is not in free lists");
7985 }
7986 _sp->removeFreeChunkFromFreeLists(ffc);
7987 set_freeRangeInFreeLists(false);
7988 }
7989 set_lastFreeRangeCoalesced(true);
7990 }
7991 // this will be swept up when we hit the end of the free range
7993 // Verify that the bit map has no bits marked between
7994 // addr and purported end of just dead object.
7995 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7996 }
7997 return size;
7998 }
8000 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
8001 HeapWord* addr = (HeapWord*) fc;
8002 // The sweeper has just found a live object. Return any accumulated
8003 // left hand chunk to the free lists.
8004 if (inFreeRange()) {
8005 if (_sp->adaptive_freelists()) {
8006 flushCurFreeChunk(freeFinger(),
8007 pointer_delta(addr, freeFinger()));
8008 } else { // not adaptive freelists
8009 set_inFreeRange(false);
8010 // Add the free range back to the free list if it is not already
8011 // there.
8012 if (!freeRangeInFreeLists()) {
8013 assert(freeFinger() < addr, "the finger pointeth off base");
8014 if (CMSTraceSweeper) {
8015 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
8016 "[coalesced:%d]\n",
8017 freeFinger(), pointer_delta(addr, freeFinger()),
8018 lastFreeRangeCoalesced());
8019 }
8020 _sp->addChunkAndRepairOffsetTable(freeFinger(),
8021 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
8022 }
8023 }
8024 }
8026 // Common code path for original and adaptive free lists.
8028 // this object is live: we'd normally expect this to be
8029 // an oop, and like to assert the following:
8030 // assert(oop(addr)->is_oop(), "live block should be an oop");
8031 // However, as we commented above, this may be an object whose
8032 // header hasn't yet been initialized.
8033 size_t size;
8034 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8035 if (_bitMap->isMarked(addr + 1)) {
8036 // Determine the size from the bit map, rather than trying to
8037 // compute it from the object header.
8038 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8039 size = pointer_delta(nextOneAddr + 1, addr);
8040 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8041 "alignment problem");
8043 #ifdef DEBUG
8044 if (oop(addr)->klass_or_null() != NULL &&
8045 ( !_collector->should_unload_classes()
8046 || oop(addr)->is_parsable())) {
8047 // Ignore mark word because we are running concurrent with mutators
8048 assert(oop(addr)->is_oop(true), "live block should be an oop");
8049 assert(size ==
8050 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8051 "P-mark and computed size do not agree");
8052 }
8053 #endif
8055 } else {
8056 // This should be an initialized object that's alive.
8057 assert(oop(addr)->klass_or_null() != NULL &&
8058 (!_collector->should_unload_classes()
8059 || oop(addr)->is_parsable()),
8060 "Should be an initialized object");
8061 // Ignore mark word because we are running concurrent with mutators
8062 assert(oop(addr)->is_oop(true), "live block should be an oop");
8063 // Verify that the bit map has no bits marked between
8064 // addr and purported end of this block.
8065 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8066 assert(size >= 3, "Necessary for Printezis marks to work");
8067 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8068 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8069 }
8070 return size;
8071 }
8073 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
8074 size_t chunkSize) {
8075 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
8076 // scheme.
8077 bool fcInFreeLists = fc->isFree();
8078 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8079 assert((HeapWord*)fc <= _limit, "sweep invariant");
8080 if (CMSTestInFreeList && fcInFreeLists) {
8081 assert(_sp->verifyChunkInFreeLists(fc),
8082 "free chunk is not in free lists");
8083 }
8086 if (CMSTraceSweeper) {
8087 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8088 }
8090 HeapWord* addr = (HeapWord*) fc;
8092 bool coalesce;
8093 size_t left = pointer_delta(addr, freeFinger());
8094 size_t right = chunkSize;
8095 switch (FLSCoalescePolicy) {
8096 // numeric value forms a coalition aggressiveness metric
8097 case 0: { // never coalesce
8098 coalesce = false;
8099 break;
8100 }
8101 case 1: { // coalesce if left & right chunks on overpopulated lists
8102 coalesce = _sp->coalOverPopulated(left) &&
8103 _sp->coalOverPopulated(right);
8104 break;
8105 }
8106 case 2: { // coalesce if left chunk on overpopulated list (default)
8107 coalesce = _sp->coalOverPopulated(left);
8108 break;
8109 }
8110 case 3: { // coalesce if left OR right chunk on overpopulated list
8111 coalesce = _sp->coalOverPopulated(left) ||
8112 _sp->coalOverPopulated(right);
8113 break;
8114 }
8115 case 4: { // always coalesce
8116 coalesce = true;
8117 break;
8118 }
8119 default:
8120 ShouldNotReachHere();
8121 }
8123 // Should the current free range be coalesced?
8124 // If the chunk is in a free range and either we decided to coalesce above
8125 // or the chunk is near the large block at the end of the heap
8126 // (isNearLargestChunk() returns true), then coalesce this chunk.
8127 bool doCoalesce = inFreeRange() &&
8128 (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8129 if (doCoalesce) {
8130 // Coalesce the current free range on the left with the new
8131 // chunk on the right. If either is on a free list,
8132 // it must be removed from the list and stashed in the closure.
8133 if (freeRangeInFreeLists()) {
8134 FreeChunk* ffc = (FreeChunk*)freeFinger();
8135 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8136 "Size of free range is inconsistent with chunk size.");
8137 if (CMSTestInFreeList) {
8138 assert(_sp->verifyChunkInFreeLists(ffc),
8139 "Chunk is not in free lists");
8140 }
8141 _sp->coalDeath(ffc->size());
8142 _sp->removeFreeChunkFromFreeLists(ffc);
8143 set_freeRangeInFreeLists(false);
8144 }
8145 if (fcInFreeLists) {
8146 _sp->coalDeath(chunkSize);
8147 assert(fc->size() == chunkSize,
8148 "The chunk has the wrong size or is not in the free lists");
8149 _sp->removeFreeChunkFromFreeLists(fc);
8150 }
8151 set_lastFreeRangeCoalesced(true);
8152 } else { // not in a free range and/or should not coalesce
8153 // Return the current free range and start a new one.
8154 if (inFreeRange()) {
8155 // In a free range but cannot coalesce with the right hand chunk.
8156 // Put the current free range into the free lists.
8157 flushCurFreeChunk(freeFinger(),
8158 pointer_delta(addr, freeFinger()));
8159 }
8160 // Set up for new free range. Pass along whether the right hand
8161 // chunk is in the free lists.
8162 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8163 }
8164 }
8165 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
8166 assert(inFreeRange(), "Should only be called if currently in a free range.");
8167 assert(size > 0,
8168 "A zero sized chunk cannot be added to the free lists.");
8169 if (!freeRangeInFreeLists()) {
8170 if(CMSTestInFreeList) {
8171 FreeChunk* fc = (FreeChunk*) chunk;
8172 fc->setSize(size);
8173 assert(!_sp->verifyChunkInFreeLists(fc),
8174 "chunk should not be in free lists yet");
8175 }
8176 if (CMSTraceSweeper) {
8177 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8178 chunk, size);
8179 }
8180 // A new free range is going to be starting. The current
8181 // free range has not been added to the free lists yet or
8182 // was removed so add it back.
8183 // If the current free range was coalesced, then the death
8184 // of the free range was recorded. Record a birth now.
8185 if (lastFreeRangeCoalesced()) {
8186 _sp->coalBirth(size);
8187 }
8188 _sp->addChunkAndRepairOffsetTable(chunk, size,
8189 lastFreeRangeCoalesced());
8190 }
8191 set_inFreeRange(false);
8192 set_freeRangeInFreeLists(false);
8193 }
8195 // We take a break if we've been at this for a while,
8196 // so as to avoid monopolizing the locks involved.
8197 void SweepClosure::do_yield_work(HeapWord* addr) {
8198 // Return current free chunk being used for coalescing (if any)
8199 // to the appropriate freelist. After yielding, the next
8200 // free block encountered will start a coalescing range of
8201 // free blocks. If the next free block is adjacent to the
8202 // chunk just flushed, they will need to wait for the next
8203 // sweep to be coalesced.
8204 if (inFreeRange()) {
8205 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
8206 }
8208 // First give up the locks, then yield, then re-lock.
8209 // We should probably use a constructor/destructor idiom to
8210 // do this unlock/lock or modify the MutexUnlocker class to
8211 // serve our purpose. XXX
8212 assert_lock_strong(_bitMap->lock());
8213 assert_lock_strong(_freelistLock);
8214 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8215 "CMS thread should hold CMS token");
8216 _bitMap->lock()->unlock();
8217 _freelistLock->unlock();
8218 ConcurrentMarkSweepThread::desynchronize(true);
8219 ConcurrentMarkSweepThread::acknowledge_yield_request();
8220 _collector->stopTimer();
8221 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8222 if (PrintCMSStatistics != 0) {
8223 _collector->incrementYields();
8224 }
8225 _collector->icms_wait();
8227 // See the comment in coordinator_yield()
8228 for (unsigned i = 0; i < CMSYieldSleepCount &&
8229 ConcurrentMarkSweepThread::should_yield() &&
8230 !CMSCollector::foregroundGCIsActive(); ++i) {
8231 os::sleep(Thread::current(), 1, false);
8232 ConcurrentMarkSweepThread::acknowledge_yield_request();
8233 }
8235 ConcurrentMarkSweepThread::synchronize(true);
8236 _freelistLock->lock();
8237 _bitMap->lock()->lock_without_safepoint_check();
8238 _collector->startTimer();
8239 }
8241 #ifndef PRODUCT
8242 // This is actually very useful in a product build if it can
8243 // be called from the debugger. Compile it into the product
8244 // as needed.
8245 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8246 return debug_cms_space->verifyChunkInFreeLists(fc);
8247 }
8249 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8250 if (CMSTraceSweeper) {
8251 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8252 }
8253 }
8254 #endif
8256 // CMSIsAliveClosure
8257 bool CMSIsAliveClosure::do_object_b(oop obj) {
8258 HeapWord* addr = (HeapWord*)obj;
8259 return addr != NULL &&
8260 (!_span.contains(addr) || _bit_map->isMarked(addr));
8261 }
8263 // CMSKeepAliveClosure: the serial version
8264 void CMSKeepAliveClosure::do_oop(oop obj) {
8265 HeapWord* addr = (HeapWord*)obj;
8266 if (_span.contains(addr) &&
8267 !_bit_map->isMarked(addr)) {
8268 _bit_map->mark(addr);
8269 bool simulate_overflow = false;
8270 NOT_PRODUCT(
8271 if (CMSMarkStackOverflowALot &&
8272 _collector->simulate_overflow()) {
8273 // simulate a stack overflow
8274 simulate_overflow = true;
8275 }
8276 )
8277 if (simulate_overflow || !_mark_stack->push(obj)) {
8278 _collector->push_on_overflow_list(obj);
8279 _collector->_ser_kac_ovflw++;
8280 }
8281 }
8282 }
8284 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8285 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8287 // CMSParKeepAliveClosure: a parallel version of the above.
8288 // The work queues are private to each closure (thread),
8289 // but (may be) available for stealing by other threads.
8290 void CMSParKeepAliveClosure::do_oop(oop obj) {
8291 HeapWord* addr = (HeapWord*)obj;
8292 if (_span.contains(addr) &&
8293 !_bit_map->isMarked(addr)) {
8294 // In general, during recursive tracing, several threads
8295 // may be concurrently getting here; the first one to
8296 // "tag" it, claims it.
8297 if (_bit_map->par_mark(addr)) {
8298 bool res = _work_queue->push(obj);
8299 assert(res, "Low water mark should be much less than capacity");
8300 // Do a recursive trim in the hope that this will keep
8301 // stack usage lower, but leave some oops for potential stealers
8302 trim_queue(_low_water_mark);
8303 } // Else, another thread got there first
8304 }
8305 }
8307 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8308 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8310 void CMSParKeepAliveClosure::trim_queue(uint max) {
8311 while (_work_queue->size() > max) {
8312 oop new_oop;
8313 if (_work_queue->pop_local(new_oop)) {
8314 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8315 assert(_bit_map->isMarked((HeapWord*)new_oop),
8316 "no white objects on this stack!");
8317 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8318 // iterate over the oops in this oop, marking and pushing
8319 // the ones in CMS heap (i.e. in _span).
8320 new_oop->oop_iterate(&_mark_and_push);
8321 }
8322 }
8323 }
8325 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8326 HeapWord* addr = (HeapWord*)obj;
8327 if (_span.contains(addr) &&
8328 !_bit_map->isMarked(addr)) {
8329 if (_bit_map->par_mark(addr)) {
8330 bool simulate_overflow = false;
8331 NOT_PRODUCT(
8332 if (CMSMarkStackOverflowALot &&
8333 _collector->par_simulate_overflow()) {
8334 // simulate a stack overflow
8335 simulate_overflow = true;
8336 }
8337 )
8338 if (simulate_overflow || !_work_queue->push(obj)) {
8339 _collector->par_push_on_overflow_list(obj);
8340 _collector->_par_kac_ovflw++;
8341 }
8342 } // Else another thread got there already
8343 }
8344 }
8346 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8347 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8349 //////////////////////////////////////////////////////////////////
8350 // CMSExpansionCause /////////////////////////////
8351 //////////////////////////////////////////////////////////////////
8352 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8353 switch (cause) {
8354 case _no_expansion:
8355 return "No expansion";
8356 case _satisfy_free_ratio:
8357 return "Free ratio";
8358 case _satisfy_promotion:
8359 return "Satisfy promotion";
8360 case _satisfy_allocation:
8361 return "allocation";
8362 case _allocate_par_lab:
8363 return "Par LAB";
8364 case _allocate_par_spooling_space:
8365 return "Par Spooling Space";
8366 case _adaptive_size_policy:
8367 return "Ergonomics";
8368 default:
8369 return "unknown";
8370 }
8371 }
8373 void CMSDrainMarkingStackClosure::do_void() {
8374 // the max number to take from overflow list at a time
8375 const size_t num = _mark_stack->capacity()/4;
8376 while (!_mark_stack->isEmpty() ||
8377 // if stack is empty, check the overflow list
8378 _collector->take_from_overflow_list(num, _mark_stack)) {
8379 oop obj = _mark_stack->pop();
8380 HeapWord* addr = (HeapWord*)obj;
8381 assert(_span.contains(addr), "Should be within span");
8382 assert(_bit_map->isMarked(addr), "Should be marked");
8383 assert(obj->is_oop(), "Should be an oop");
8384 obj->oop_iterate(_keep_alive);
8385 }
8386 }
8388 void CMSParDrainMarkingStackClosure::do_void() {
8389 // drain queue
8390 trim_queue(0);
8391 }
8393 // Trim our work_queue so its length is below max at return
8394 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8395 while (_work_queue->size() > max) {
8396 oop new_oop;
8397 if (_work_queue->pop_local(new_oop)) {
8398 assert(new_oop->is_oop(), "Expected an oop");
8399 assert(_bit_map->isMarked((HeapWord*)new_oop),
8400 "no white objects on this stack!");
8401 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8402 // iterate over the oops in this oop, marking and pushing
8403 // the ones in CMS heap (i.e. in _span).
8404 new_oop->oop_iterate(&_mark_and_push);
8405 }
8406 }
8407 }
8409 ////////////////////////////////////////////////////////////////////
8410 // Support for Marking Stack Overflow list handling and related code
8411 ////////////////////////////////////////////////////////////////////
8412 // Much of the following code is similar in shape and spirit to the
8413 // code used in ParNewGC. We should try and share that code
8414 // as much as possible in the future.
8416 #ifndef PRODUCT
8417 // Debugging support for CMSStackOverflowALot
8419 // It's OK to call this multi-threaded; the worst thing
8420 // that can happen is that we'll get a bunch of closely
8421 // spaced simulated oveflows, but that's OK, in fact
8422 // probably good as it would exercise the overflow code
8423 // under contention.
8424 bool CMSCollector::simulate_overflow() {
8425 if (_overflow_counter-- <= 0) { // just being defensive
8426 _overflow_counter = CMSMarkStackOverflowInterval;
8427 return true;
8428 } else {
8429 return false;
8430 }
8431 }
8433 bool CMSCollector::par_simulate_overflow() {
8434 return simulate_overflow();
8435 }
8436 #endif
8438 // Single-threaded
8439 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8440 assert(stack->isEmpty(), "Expected precondition");
8441 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8442 size_t i = num;
8443 oop cur = _overflow_list;
8444 const markOop proto = markOopDesc::prototype();
8445 NOT_PRODUCT(size_t n = 0;)
8446 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8447 next = oop(cur->mark());
8448 cur->set_mark(proto); // until proven otherwise
8449 assert(cur->is_oop(), "Should be an oop");
8450 bool res = stack->push(cur);
8451 assert(res, "Bit off more than can chew?");
8452 NOT_PRODUCT(n++;)
8453 }
8454 _overflow_list = cur;
8455 #ifndef PRODUCT
8456 assert(_num_par_pushes >= n, "Too many pops?");
8457 _num_par_pushes -=n;
8458 #endif
8459 return !stack->isEmpty();
8460 }
8462 // Multi-threaded; use CAS to break off a prefix
8463 bool CMSCollector::par_take_from_overflow_list(size_t num,
8464 OopTaskQueue* work_q) {
8465 assert(work_q->size() == 0, "That's the current policy");
8466 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8467 if (_overflow_list == NULL) {
8468 return false;
8469 }
8470 // Grab the entire list; we'll put back a suffix
8471 oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
8472 if (prefix == NULL) { // someone grabbed it before we did ...
8473 // ... we could spin for a short while, but for now we don't
8474 return false;
8475 }
8476 size_t i = num;
8477 oop cur = prefix;
8478 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8479 if (cur->mark() != NULL) {
8480 oop suffix_head = cur->mark(); // suffix will be put back on global list
8481 cur->set_mark(NULL); // break off suffix
8482 // Find tail of suffix so we can prepend suffix to global list
8483 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8484 oop suffix_tail = cur;
8485 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8486 "Tautology");
8487 oop observed_overflow_list = _overflow_list;
8488 do {
8489 cur = observed_overflow_list;
8490 suffix_tail->set_mark(markOop(cur));
8491 observed_overflow_list =
8492 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
8493 } while (cur != observed_overflow_list);
8494 }
8496 // Push the prefix elements on work_q
8497 assert(prefix != NULL, "control point invariant");
8498 const markOop proto = markOopDesc::prototype();
8499 oop next;
8500 NOT_PRODUCT(size_t n = 0;)
8501 for (cur = prefix; cur != NULL; cur = next) {
8502 next = oop(cur->mark());
8503 cur->set_mark(proto); // until proven otherwise
8504 assert(cur->is_oop(), "Should be an oop");
8505 bool res = work_q->push(cur);
8506 assert(res, "Bit off more than we can chew?");
8507 NOT_PRODUCT(n++;)
8508 }
8509 #ifndef PRODUCT
8510 assert(_num_par_pushes >= n, "Too many pops?");
8511 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8512 #endif
8513 return true;
8514 }
8516 // Single-threaded
8517 void CMSCollector::push_on_overflow_list(oop p) {
8518 NOT_PRODUCT(_num_par_pushes++;)
8519 assert(p->is_oop(), "Not an oop");
8520 preserve_mark_if_necessary(p);
8521 p->set_mark((markOop)_overflow_list);
8522 _overflow_list = p;
8523 }
8525 // Multi-threaded; use CAS to prepend to overflow list
8526 void CMSCollector::par_push_on_overflow_list(oop p) {
8527 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8528 assert(p->is_oop(), "Not an oop");
8529 par_preserve_mark_if_necessary(p);
8530 oop observed_overflow_list = _overflow_list;
8531 oop cur_overflow_list;
8532 do {
8533 cur_overflow_list = observed_overflow_list;
8534 p->set_mark(markOop(cur_overflow_list));
8535 observed_overflow_list =
8536 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8537 } while (cur_overflow_list != observed_overflow_list);
8538 }
8540 // Single threaded
8541 // General Note on GrowableArray: pushes may silently fail
8542 // because we are (temporarily) out of C-heap for expanding
8543 // the stack. The problem is quite ubiquitous and affects
8544 // a lot of code in the JVM. The prudent thing for GrowableArray
8545 // to do (for now) is to exit with an error. However, that may
8546 // be too draconian in some cases because the caller may be
8547 // able to recover without much harm. For suych cases, we
8548 // should probably introduce a "soft_push" method which returns
8549 // an indication of success or failure with the assumption that
8550 // the caller may be able to recover from a failure; code in
8551 // the VM can then be changed, incrementally, to deal with such
8552 // failures where possible, thus, incrementally hardening the VM
8553 // in such low resource situations.
8554 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8555 int PreserveMarkStackSize = 128;
8557 if (_preserved_oop_stack == NULL) {
8558 assert(_preserved_mark_stack == NULL,
8559 "bijection with preserved_oop_stack");
8560 // Allocate the stacks
8561 _preserved_oop_stack = new (ResourceObj::C_HEAP)
8562 GrowableArray<oop>(PreserveMarkStackSize, true);
8563 _preserved_mark_stack = new (ResourceObj::C_HEAP)
8564 GrowableArray<markOop>(PreserveMarkStackSize, true);
8565 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
8566 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
8567 "Preserved Mark/Oop Stack for CMS (C-heap)");
8568 }
8569 }
8570 _preserved_oop_stack->push(p);
8571 _preserved_mark_stack->push(m);
8572 assert(m == p->mark(), "Mark word changed");
8573 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
8574 "bijection");
8575 }
8577 // Single threaded
8578 void CMSCollector::preserve_mark_if_necessary(oop p) {
8579 markOop m = p->mark();
8580 if (m->must_be_preserved(p)) {
8581 preserve_mark_work(p, m);
8582 }
8583 }
8585 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8586 markOop m = p->mark();
8587 if (m->must_be_preserved(p)) {
8588 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8589 // Even though we read the mark word without holding
8590 // the lock, we are assured that it will not change
8591 // because we "own" this oop, so no other thread can
8592 // be trying to push it on the overflow list; see
8593 // the assertion in preserve_mark_work() that checks
8594 // that m == p->mark().
8595 preserve_mark_work(p, m);
8596 }
8597 }
8599 // We should be able to do this multi-threaded,
8600 // a chunk of stack being a task (this is
8601 // correct because each oop only ever appears
8602 // once in the overflow list. However, it's
8603 // not very easy to completely overlap this with
8604 // other operations, so will generally not be done
8605 // until all work's been completed. Because we
8606 // expect the preserved oop stack (set) to be small,
8607 // it's probably fine to do this single-threaded.
8608 // We can explore cleverer concurrent/overlapped/parallel
8609 // processing of preserved marks if we feel the
8610 // need for this in the future. Stack overflow should
8611 // be so rare in practice and, when it happens, its
8612 // effect on performance so great that this will
8613 // likely just be in the noise anyway.
8614 void CMSCollector::restore_preserved_marks_if_any() {
8615 if (_preserved_oop_stack == NULL) {
8616 assert(_preserved_mark_stack == NULL,
8617 "bijection with preserved_oop_stack");
8618 return;
8619 }
8621 assert(SafepointSynchronize::is_at_safepoint(),
8622 "world should be stopped");
8623 assert(Thread::current()->is_ConcurrentGC_thread() ||
8624 Thread::current()->is_VM_thread(),
8625 "should be single-threaded");
8627 int length = _preserved_oop_stack->length();
8628 assert(_preserved_mark_stack->length() == length, "bijection");
8629 for (int i = 0; i < length; i++) {
8630 oop p = _preserved_oop_stack->at(i);
8631 assert(p->is_oop(), "Should be an oop");
8632 assert(_span.contains(p), "oop should be in _span");
8633 assert(p->mark() == markOopDesc::prototype(),
8634 "Set when taken from overflow list");
8635 markOop m = _preserved_mark_stack->at(i);
8636 p->set_mark(m);
8637 }
8638 _preserved_mark_stack->clear();
8639 _preserved_oop_stack->clear();
8640 assert(_preserved_mark_stack->is_empty() &&
8641 _preserved_oop_stack->is_empty(),
8642 "stacks were cleared above");
8643 }
8645 #ifndef PRODUCT
8646 bool CMSCollector::no_preserved_marks() const {
8647 return ( ( _preserved_mark_stack == NULL
8648 && _preserved_oop_stack == NULL)
8649 || ( _preserved_mark_stack->is_empty()
8650 && _preserved_oop_stack->is_empty()));
8651 }
8652 #endif
8654 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8655 {
8656 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8657 CMSAdaptiveSizePolicy* size_policy =
8658 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8659 assert(size_policy->is_gc_cms_adaptive_size_policy(),
8660 "Wrong type for size policy");
8661 return size_policy;
8662 }
8664 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
8665 size_t desired_promo_size) {
8666 if (cur_promo_size < desired_promo_size) {
8667 size_t expand_bytes = desired_promo_size - cur_promo_size;
8668 if (PrintAdaptiveSizePolicy && Verbose) {
8669 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8670 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
8671 expand_bytes);
8672 }
8673 expand(expand_bytes,
8674 MinHeapDeltaBytes,
8675 CMSExpansionCause::_adaptive_size_policy);
8676 } else if (desired_promo_size < cur_promo_size) {
8677 size_t shrink_bytes = cur_promo_size - desired_promo_size;
8678 if (PrintAdaptiveSizePolicy && Verbose) {
8679 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8680 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
8681 shrink_bytes);
8682 }
8683 shrink(shrink_bytes);
8684 }
8685 }
8687 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
8688 GenCollectedHeap* gch = GenCollectedHeap::heap();
8689 CMSGCAdaptivePolicyCounters* counters =
8690 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
8691 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
8692 "Wrong kind of counters");
8693 return counters;
8694 }
8697 void ASConcurrentMarkSweepGeneration::update_counters() {
8698 if (UsePerfData) {
8699 _space_counters->update_all();
8700 _gen_counters->update_all();
8701 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8702 GenCollectedHeap* gch = GenCollectedHeap::heap();
8703 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8704 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8705 "Wrong gc statistics type");
8706 counters->update_counters(gc_stats_l);
8707 }
8708 }
8710 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
8711 if (UsePerfData) {
8712 _space_counters->update_used(used);
8713 _space_counters->update_capacity();
8714 _gen_counters->update_all();
8716 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8717 GenCollectedHeap* gch = GenCollectedHeap::heap();
8718 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8719 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8720 "Wrong gc statistics type");
8721 counters->update_counters(gc_stats_l);
8722 }
8723 }
8725 // The desired expansion delta is computed so that:
8726 // . desired free percentage or greater is used
8727 void ASConcurrentMarkSweepGeneration::compute_new_size() {
8728 assert_locked_or_safepoint(Heap_lock);
8730 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8732 // If incremental collection failed, we just want to expand
8733 // to the limit.
8734 if (incremental_collection_failed()) {
8735 clear_incremental_collection_failed();
8736 grow_to_reserved();
8737 return;
8738 }
8740 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
8742 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
8743 "Wrong type of heap");
8744 int prev_level = level() - 1;
8745 assert(prev_level >= 0, "The cms generation is the lowest generation");
8746 Generation* prev_gen = gch->get_gen(prev_level);
8747 assert(prev_gen->kind() == Generation::ASParNew,
8748 "Wrong type of young generation");
8749 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
8750 size_t cur_eden = younger_gen->eden()->capacity();
8751 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
8752 size_t cur_promo = free();
8753 size_policy->compute_tenured_generation_free_space(cur_promo,
8754 max_available(),
8755 cur_eden);
8756 resize(cur_promo, size_policy->promo_size());
8758 // Record the new size of the space in the cms generation
8759 // that is available for promotions. This is temporary.
8760 // It should be the desired promo size.
8761 size_policy->avg_cms_promo()->sample(free());
8762 size_policy->avg_old_live()->sample(used());
8764 if (UsePerfData) {
8765 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8766 counters->update_cms_capacity_counter(capacity());
8767 }
8768 }
8770 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
8771 assert_locked_or_safepoint(Heap_lock);
8772 assert_lock_strong(freelistLock());
8773 HeapWord* old_end = _cmsSpace->end();
8774 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
8775 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
8776 FreeChunk* chunk_at_end = find_chunk_at_end();
8777 if (chunk_at_end == NULL) {
8778 // No room to shrink
8779 if (PrintGCDetails && Verbose) {
8780 gclog_or_tty->print_cr("No room to shrink: old_end "
8781 PTR_FORMAT " unallocated_start " PTR_FORMAT
8782 " chunk_at_end " PTR_FORMAT,
8783 old_end, unallocated_start, chunk_at_end);
8784 }
8785 return;
8786 } else {
8788 // Find the chunk at the end of the space and determine
8789 // how much it can be shrunk.
8790 size_t shrinkable_size_in_bytes = chunk_at_end->size();
8791 size_t aligned_shrinkable_size_in_bytes =
8792 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
8793 assert(unallocated_start <= chunk_at_end->end(),
8794 "Inconsistent chunk at end of space");
8795 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
8796 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
8798 // Shrink the underlying space
8799 _virtual_space.shrink_by(bytes);
8800 if (PrintGCDetails && Verbose) {
8801 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
8802 " desired_bytes " SIZE_FORMAT
8803 " shrinkable_size_in_bytes " SIZE_FORMAT
8804 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
8805 " bytes " SIZE_FORMAT,
8806 desired_bytes, shrinkable_size_in_bytes,
8807 aligned_shrinkable_size_in_bytes, bytes);
8808 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
8809 " unallocated_start " SIZE_FORMAT,
8810 old_end, unallocated_start);
8811 }
8813 // If the space did shrink (shrinking is not guaranteed),
8814 // shrink the chunk at the end by the appropriate amount.
8815 if (((HeapWord*)_virtual_space.high()) < old_end) {
8816 size_t new_word_size =
8817 heap_word_size(_virtual_space.committed_size());
8819 // Have to remove the chunk from the dictionary because it is changing
8820 // size and might be someplace elsewhere in the dictionary.
8822 // Get the chunk at end, shrink it, and put it
8823 // back.
8824 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
8825 size_t word_size_change = word_size_before - new_word_size;
8826 size_t chunk_at_end_old_size = chunk_at_end->size();
8827 assert(chunk_at_end_old_size >= word_size_change,
8828 "Shrink is too large");
8829 chunk_at_end->setSize(chunk_at_end_old_size -
8830 word_size_change);
8831 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
8832 word_size_change);
8834 _cmsSpace->returnChunkToDictionary(chunk_at_end);
8836 MemRegion mr(_cmsSpace->bottom(), new_word_size);
8837 _bts->resize(new_word_size); // resize the block offset shared array
8838 Universe::heap()->barrier_set()->resize_covered_region(mr);
8839 _cmsSpace->assert_locked();
8840 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
8842 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
8844 // update the space and generation capacity counters
8845 if (UsePerfData) {
8846 _space_counters->update_capacity();
8847 _gen_counters->update_all();
8848 }
8850 if (Verbose && PrintGCDetails) {
8851 size_t new_mem_size = _virtual_space.committed_size();
8852 size_t old_mem_size = new_mem_size + bytes;
8853 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
8854 name(), old_mem_size/K, bytes/K, new_mem_size/K);
8855 }
8856 }
8858 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
8859 "Inconsistency at end of space");
8860 assert(chunk_at_end->end() == _cmsSpace->end(),
8861 "Shrinking is inconsistent");
8862 return;
8863 }
8864 }
8866 // Transfer some number of overflown objects to usual marking
8867 // stack. Return true if some objects were transferred.
8868 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8869 size_t num = MIN2((size_t)_mark_stack->capacity()/4,
8870 (size_t)ParGCDesiredObjsFromOverflowList);
8872 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8873 assert(_collector->overflow_list_is_empty() || res,
8874 "If list is not empty, we should have taken something");
8875 assert(!res || !_mark_stack->isEmpty(),
8876 "If we took something, it should now be on our stack");
8877 return res;
8878 }
8880 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8881 size_t res = _sp->block_size_no_stall(addr, _collector);
8882 assert(res != 0, "Should always be able to compute a size");
8883 if (_sp->block_is_obj(addr)) {
8884 if (_live_bit_map->isMarked(addr)) {
8885 // It can't have been dead in a previous cycle
8886 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8887 } else {
8888 _dead_bit_map->mark(addr); // mark the dead object
8889 }
8890 }
8891 return res;
8892 }