Thu, 21 Aug 2008 23:36:31 -0400
Merge
1 /*
2 * Copyright 2001-2008 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 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3200 return CardGeneration::expand(bytes, expand_bytes);
3201 }
3203 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3204 CMSExpansionCause::Cause cause)
3205 {
3207 bool success = expand(bytes, expand_bytes);
3209 // remember why we expanded; this information is used
3210 // by shouldConcurrentCollect() when making decisions on whether to start
3211 // a new CMS cycle.
3212 if (success) {
3213 set_expansion_cause(cause);
3214 if (PrintGCDetails && Verbose) {
3215 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3216 CMSExpansionCause::to_string(cause));
3217 }
3218 }
3219 }
3221 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3222 HeapWord* res = NULL;
3223 MutexLocker x(ParGCRareEvent_lock);
3224 while (true) {
3225 // Expansion by some other thread might make alloc OK now:
3226 res = ps->lab.alloc(word_sz);
3227 if (res != NULL) return res;
3228 // If there's not enough expansion space available, give up.
3229 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3230 return NULL;
3231 }
3232 // Otherwise, we try expansion.
3233 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3234 CMSExpansionCause::_allocate_par_lab);
3235 // Now go around the loop and try alloc again;
3236 // A competing par_promote might beat us to the expansion space,
3237 // so we may go around the loop again if promotion fails agaion.
3238 if (GCExpandToAllocateDelayMillis > 0) {
3239 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3240 }
3241 }
3242 }
3245 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3246 PromotionInfo* promo) {
3247 MutexLocker x(ParGCRareEvent_lock);
3248 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3249 while (true) {
3250 // Expansion by some other thread might make alloc OK now:
3251 if (promo->ensure_spooling_space()) {
3252 assert(promo->has_spooling_space(),
3253 "Post-condition of successful ensure_spooling_space()");
3254 return true;
3255 }
3256 // If there's not enough expansion space available, give up.
3257 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3258 return false;
3259 }
3260 // Otherwise, we try expansion.
3261 expand(refill_size_bytes, MinHeapDeltaBytes,
3262 CMSExpansionCause::_allocate_par_spooling_space);
3263 // Now go around the loop and try alloc again;
3264 // A competing allocation might beat us to the expansion space,
3265 // so we may go around the loop again if allocation fails again.
3266 if (GCExpandToAllocateDelayMillis > 0) {
3267 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3268 }
3269 }
3270 }
3274 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3275 assert_locked_or_safepoint(Heap_lock);
3276 size_t size = ReservedSpace::page_align_size_down(bytes);
3277 if (size > 0) {
3278 shrink_by(size);
3279 }
3280 }
3282 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3283 assert_locked_or_safepoint(Heap_lock);
3284 bool result = _virtual_space.expand_by(bytes);
3285 if (result) {
3286 HeapWord* old_end = _cmsSpace->end();
3287 size_t new_word_size =
3288 heap_word_size(_virtual_space.committed_size());
3289 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3290 _bts->resize(new_word_size); // resize the block offset shared array
3291 Universe::heap()->barrier_set()->resize_covered_region(mr);
3292 // Hmmmm... why doesn't CFLS::set_end verify locking?
3293 // This is quite ugly; FIX ME XXX
3294 _cmsSpace->assert_locked();
3295 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3297 // update the space and generation capacity counters
3298 if (UsePerfData) {
3299 _space_counters->update_capacity();
3300 _gen_counters->update_all();
3301 }
3303 if (Verbose && PrintGC) {
3304 size_t new_mem_size = _virtual_space.committed_size();
3305 size_t old_mem_size = new_mem_size - bytes;
3306 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3307 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3308 }
3309 }
3310 return result;
3311 }
3313 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3314 assert_locked_or_safepoint(Heap_lock);
3315 bool success = true;
3316 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3317 if (remaining_bytes > 0) {
3318 success = grow_by(remaining_bytes);
3319 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3320 }
3321 return success;
3322 }
3324 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3325 assert_locked_or_safepoint(Heap_lock);
3326 assert_lock_strong(freelistLock());
3327 // XXX Fix when compaction is implemented.
3328 warning("Shrinking of CMS not yet implemented");
3329 return;
3330 }
3333 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3334 // phases.
3335 class CMSPhaseAccounting: public StackObj {
3336 public:
3337 CMSPhaseAccounting(CMSCollector *collector,
3338 const char *phase,
3339 bool print_cr = true);
3340 ~CMSPhaseAccounting();
3342 private:
3343 CMSCollector *_collector;
3344 const char *_phase;
3345 elapsedTimer _wallclock;
3346 bool _print_cr;
3348 public:
3349 // Not MT-safe; so do not pass around these StackObj's
3350 // where they may be accessed by other threads.
3351 jlong wallclock_millis() {
3352 assert(_wallclock.is_active(), "Wall clock should not stop");
3353 _wallclock.stop(); // to record time
3354 jlong ret = _wallclock.milliseconds();
3355 _wallclock.start(); // restart
3356 return ret;
3357 }
3358 };
3360 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3361 const char *phase,
3362 bool print_cr) :
3363 _collector(collector), _phase(phase), _print_cr(print_cr) {
3365 if (PrintCMSStatistics != 0) {
3366 _collector->resetYields();
3367 }
3368 if (PrintGCDetails && PrintGCTimeStamps) {
3369 gclog_or_tty->date_stamp(PrintGCDateStamps);
3370 gclog_or_tty->stamp();
3371 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3372 _collector->cmsGen()->short_name(), _phase);
3373 }
3374 _collector->resetTimer();
3375 _wallclock.start();
3376 _collector->startTimer();
3377 }
3379 CMSPhaseAccounting::~CMSPhaseAccounting() {
3380 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3381 _collector->stopTimer();
3382 _wallclock.stop();
3383 if (PrintGCDetails) {
3384 gclog_or_tty->date_stamp(PrintGCDateStamps);
3385 if (PrintGCTimeStamps) {
3386 gclog_or_tty->stamp();
3387 gclog_or_tty->print(": ");
3388 }
3389 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3390 _collector->cmsGen()->short_name(),
3391 _phase, _collector->timerValue(), _wallclock.seconds());
3392 if (_print_cr) {
3393 gclog_or_tty->print_cr("");
3394 }
3395 if (PrintCMSStatistics != 0) {
3396 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3397 _collector->yields());
3398 }
3399 }
3400 }
3402 // CMS work
3404 // Checkpoint the roots into this generation from outside
3405 // this generation. [Note this initial checkpoint need only
3406 // be approximate -- we'll do a catch up phase subsequently.]
3407 void CMSCollector::checkpointRootsInitial(bool asynch) {
3408 assert(_collectorState == InitialMarking, "Wrong collector state");
3409 check_correct_thread_executing();
3410 ReferenceProcessor* rp = ref_processor();
3411 SpecializationStats::clear();
3412 assert(_restart_addr == NULL, "Control point invariant");
3413 if (asynch) {
3414 // acquire locks for subsequent manipulations
3415 MutexLockerEx x(bitMapLock(),
3416 Mutex::_no_safepoint_check_flag);
3417 checkpointRootsInitialWork(asynch);
3418 rp->verify_no_references_recorded();
3419 rp->enable_discovery(); // enable ("weak") refs discovery
3420 _collectorState = Marking;
3421 } else {
3422 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3423 // which recognizes if we are a CMS generation, and doesn't try to turn on
3424 // discovery; verify that they aren't meddling.
3425 assert(!rp->discovery_is_atomic(),
3426 "incorrect setting of discovery predicate");
3427 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3428 "ref discovery for this generation kind");
3429 // already have locks
3430 checkpointRootsInitialWork(asynch);
3431 rp->enable_discovery(); // now enable ("weak") refs discovery
3432 _collectorState = Marking;
3433 }
3434 SpecializationStats::print();
3435 }
3437 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3438 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3439 assert(_collectorState == InitialMarking, "just checking");
3441 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3442 // precede our marking with a collection of all
3443 // younger generations to keep floating garbage to a minimum.
3444 // XXX: we won't do this for now -- it's an optimization to be done later.
3446 // already have locks
3447 assert_lock_strong(bitMapLock());
3448 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3450 // Setup the verification and class unloading state for this
3451 // CMS collection cycle.
3452 setup_cms_unloading_and_verification_state();
3454 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3455 PrintGCDetails && Verbose, true, gclog_or_tty);)
3456 if (UseAdaptiveSizePolicy) {
3457 size_policy()->checkpoint_roots_initial_begin();
3458 }
3460 // Reset all the PLAB chunk arrays if necessary.
3461 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3462 reset_survivor_plab_arrays();
3463 }
3465 ResourceMark rm;
3466 HandleMark hm;
3468 FalseClosure falseClosure;
3469 // In the case of a synchronous collection, we will elide the
3470 // remark step, so it's important to catch all the nmethod oops
3471 // in this step; hence the last argument to the constrcutor below.
3472 MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */);
3473 GenCollectedHeap* gch = GenCollectedHeap::heap();
3475 verify_work_stacks_empty();
3476 verify_overflow_empty();
3478 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3479 // Update the saved marks which may affect the root scans.
3480 gch->save_marks();
3482 // weak reference processing has not started yet.
3483 ref_processor()->set_enqueuing_is_done(false);
3485 {
3486 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3487 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3488 gch->gen_process_strong_roots(_cmsGen->level(),
3489 true, // younger gens are roots
3490 true, // collecting perm gen
3491 SharedHeap::ScanningOption(roots_scanning_options()),
3492 NULL, ¬Older);
3493 }
3495 // Clear mod-union table; it will be dirtied in the prologue of
3496 // CMS generation per each younger generation collection.
3498 assert(_modUnionTable.isAllClear(),
3499 "Was cleared in most recent final checkpoint phase"
3500 " or no bits are set in the gc_prologue before the start of the next "
3501 "subsequent marking phase.");
3503 // Temporarily disabled, since pre/post-consumption closures don't
3504 // care about precleaned cards
3505 #if 0
3506 {
3507 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3508 (HeapWord*)_virtual_space.high());
3509 _ct->ct_bs()->preclean_dirty_cards(mr);
3510 }
3511 #endif
3513 // Save the end of the used_region of the constituent generations
3514 // to be used to limit the extent of sweep in each generation.
3515 save_sweep_limits();
3516 if (UseAdaptiveSizePolicy) {
3517 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3518 }
3519 verify_overflow_empty();
3520 }
3522 bool CMSCollector::markFromRoots(bool asynch) {
3523 // we might be tempted to assert that:
3524 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3525 // "inconsistent argument?");
3526 // However that wouldn't be right, because it's possible that
3527 // a safepoint is indeed in progress as a younger generation
3528 // stop-the-world GC happens even as we mark in this generation.
3529 assert(_collectorState == Marking, "inconsistent state?");
3530 check_correct_thread_executing();
3531 verify_overflow_empty();
3533 bool res;
3534 if (asynch) {
3536 // Start the timers for adaptive size policy for the concurrent phases
3537 // Do it here so that the foreground MS can use the concurrent
3538 // timer since a foreground MS might has the sweep done concurrently
3539 // or STW.
3540 if (UseAdaptiveSizePolicy) {
3541 size_policy()->concurrent_marking_begin();
3542 }
3544 // Weak ref discovery note: We may be discovering weak
3545 // refs in this generation concurrent (but interleaved) with
3546 // weak ref discovery by a younger generation collector.
3548 CMSTokenSyncWithLocks ts(true, bitMapLock());
3549 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3550 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3551 res = markFromRootsWork(asynch);
3552 if (res) {
3553 _collectorState = Precleaning;
3554 } else { // We failed and a foreground collection wants to take over
3555 assert(_foregroundGCIsActive, "internal state inconsistency");
3556 assert(_restart_addr == NULL, "foreground will restart from scratch");
3557 if (PrintGCDetails) {
3558 gclog_or_tty->print_cr("bailing out to foreground collection");
3559 }
3560 }
3561 if (UseAdaptiveSizePolicy) {
3562 size_policy()->concurrent_marking_end();
3563 }
3564 } else {
3565 assert(SafepointSynchronize::is_at_safepoint(),
3566 "inconsistent with asynch == false");
3567 if (UseAdaptiveSizePolicy) {
3568 size_policy()->ms_collection_marking_begin();
3569 }
3570 // already have locks
3571 res = markFromRootsWork(asynch);
3572 _collectorState = FinalMarking;
3573 if (UseAdaptiveSizePolicy) {
3574 GenCollectedHeap* gch = GenCollectedHeap::heap();
3575 size_policy()->ms_collection_marking_end(gch->gc_cause());
3576 }
3577 }
3578 verify_overflow_empty();
3579 return res;
3580 }
3582 bool CMSCollector::markFromRootsWork(bool asynch) {
3583 // iterate over marked bits in bit map, doing a full scan and mark
3584 // from these roots using the following algorithm:
3585 // . if oop is to the right of the current scan pointer,
3586 // mark corresponding bit (we'll process it later)
3587 // . else (oop is to left of current scan pointer)
3588 // push oop on marking stack
3589 // . drain the marking stack
3591 // Note that when we do a marking step we need to hold the
3592 // bit map lock -- recall that direct allocation (by mutators)
3593 // and promotion (by younger generation collectors) is also
3594 // marking the bit map. [the so-called allocate live policy.]
3595 // Because the implementation of bit map marking is not
3596 // robust wrt simultaneous marking of bits in the same word,
3597 // we need to make sure that there is no such interference
3598 // between concurrent such updates.
3600 // already have locks
3601 assert_lock_strong(bitMapLock());
3603 // Clear the revisit stack, just in case there are any
3604 // obsolete contents from a short-circuited previous CMS cycle.
3605 _revisitStack.reset();
3606 verify_work_stacks_empty();
3607 verify_overflow_empty();
3608 assert(_revisitStack.isEmpty(), "tabula rasa");
3610 bool result = false;
3611 if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) {
3612 result = do_marking_mt(asynch);
3613 } else {
3614 result = do_marking_st(asynch);
3615 }
3616 return result;
3617 }
3619 // Forward decl
3620 class CMSConcMarkingTask;
3622 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3623 CMSCollector* _collector;
3624 CMSConcMarkingTask* _task;
3625 bool _yield;
3626 protected:
3627 virtual void yield();
3628 public:
3629 // "n_threads" is the number of threads to be terminated.
3630 // "queue_set" is a set of work queues of other threads.
3631 // "collector" is the CMS collector associated with this task terminator.
3632 // "yield" indicates whether we need the gang as a whole to yield.
3633 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
3634 CMSCollector* collector, bool yield) :
3635 ParallelTaskTerminator(n_threads, queue_set),
3636 _collector(collector),
3637 _yield(yield) { }
3639 void set_task(CMSConcMarkingTask* task) {
3640 _task = task;
3641 }
3642 };
3644 // MT Concurrent Marking Task
3645 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3646 CMSCollector* _collector;
3647 YieldingFlexibleWorkGang* _workers; // the whole gang
3648 int _n_workers; // requested/desired # workers
3649 bool _asynch;
3650 bool _result;
3651 CompactibleFreeListSpace* _cms_space;
3652 CompactibleFreeListSpace* _perm_space;
3653 HeapWord* _global_finger;
3655 // Exposed here for yielding support
3656 Mutex* const _bit_map_lock;
3658 // The per thread work queues, available here for stealing
3659 OopTaskQueueSet* _task_queues;
3660 CMSConcMarkingTerminator _term;
3662 public:
3663 CMSConcMarkingTask(CMSCollector* collector,
3664 CompactibleFreeListSpace* cms_space,
3665 CompactibleFreeListSpace* perm_space,
3666 bool asynch, int n_workers,
3667 YieldingFlexibleWorkGang* workers,
3668 OopTaskQueueSet* task_queues):
3669 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3670 _collector(collector),
3671 _cms_space(cms_space),
3672 _perm_space(perm_space),
3673 _asynch(asynch), _n_workers(n_workers), _result(true),
3674 _workers(workers), _task_queues(task_queues),
3675 _term(n_workers, task_queues, _collector, asynch),
3676 _bit_map_lock(collector->bitMapLock())
3677 {
3678 assert(n_workers <= workers->total_workers(),
3679 "Else termination won't work correctly today"); // XXX FIX ME!
3680 _requested_size = n_workers;
3681 _term.set_task(this);
3682 assert(_cms_space->bottom() < _perm_space->bottom(),
3683 "Finger incorrectly initialized below");
3684 _global_finger = _cms_space->bottom();
3685 }
3688 OopTaskQueueSet* task_queues() { return _task_queues; }
3690 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3692 HeapWord** global_finger_addr() { return &_global_finger; }
3694 CMSConcMarkingTerminator* terminator() { return &_term; }
3696 void work(int i);
3698 virtual void coordinator_yield(); // stuff done by coordinator
3699 bool result() { return _result; }
3701 void reset(HeapWord* ra) {
3702 _term.reset_for_reuse();
3703 }
3705 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3706 OopTaskQueue* work_q);
3708 private:
3709 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3710 void do_work_steal(int i);
3711 void bump_global_finger(HeapWord* f);
3712 };
3714 void CMSConcMarkingTerminator::yield() {
3715 if (ConcurrentMarkSweepThread::should_yield() &&
3716 !_collector->foregroundGCIsActive() &&
3717 _yield) {
3718 _task->yield();
3719 } else {
3720 ParallelTaskTerminator::yield();
3721 }
3722 }
3724 ////////////////////////////////////////////////////////////////
3725 // Concurrent Marking Algorithm Sketch
3726 ////////////////////////////////////////////////////////////////
3727 // Until all tasks exhausted (both spaces):
3728 // -- claim next available chunk
3729 // -- bump global finger via CAS
3730 // -- find first object that starts in this chunk
3731 // and start scanning bitmap from that position
3732 // -- scan marked objects for oops
3733 // -- CAS-mark target, and if successful:
3734 // . if target oop is above global finger (volatile read)
3735 // nothing to do
3736 // . if target oop is in chunk and above local finger
3737 // then nothing to do
3738 // . else push on work-queue
3739 // -- Deal with possible overflow issues:
3740 // . local work-queue overflow causes stuff to be pushed on
3741 // global (common) overflow queue
3742 // . always first empty local work queue
3743 // . then get a batch of oops from global work queue if any
3744 // . then do work stealing
3745 // -- When all tasks claimed (both spaces)
3746 // and local work queue empty,
3747 // then in a loop do:
3748 // . check global overflow stack; steal a batch of oops and trace
3749 // . try to steal from other threads oif GOS is empty
3750 // . if neither is available, offer termination
3751 // -- Terminate and return result
3752 //
3753 void CMSConcMarkingTask::work(int i) {
3754 elapsedTimer _timer;
3755 ResourceMark rm;
3756 HandleMark hm;
3758 DEBUG_ONLY(_collector->verify_overflow_empty();)
3760 // Before we begin work, our work queue should be empty
3761 assert(work_queue(i)->size() == 0, "Expected to be empty");
3762 // Scan the bitmap covering _cms_space, tracing through grey objects.
3763 _timer.start();
3764 do_scan_and_mark(i, _cms_space);
3765 _timer.stop();
3766 if (PrintCMSStatistics != 0) {
3767 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3768 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3769 }
3771 // ... do the same for the _perm_space
3772 _timer.reset();
3773 _timer.start();
3774 do_scan_and_mark(i, _perm_space);
3775 _timer.stop();
3776 if (PrintCMSStatistics != 0) {
3777 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3778 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3779 }
3781 // ... do work stealing
3782 _timer.reset();
3783 _timer.start();
3784 do_work_steal(i);
3785 _timer.stop();
3786 if (PrintCMSStatistics != 0) {
3787 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3788 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3789 }
3790 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3791 assert(work_queue(i)->size() == 0, "Should have been emptied");
3792 // Note that under the current task protocol, the
3793 // following assertion is true even of the spaces
3794 // expanded since the completion of the concurrent
3795 // marking. XXX This will likely change under a strict
3796 // ABORT semantics.
3797 assert(_global_finger > _cms_space->end() &&
3798 _global_finger >= _perm_space->end(),
3799 "All tasks have been completed");
3800 DEBUG_ONLY(_collector->verify_overflow_empty();)
3801 }
3803 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3804 HeapWord* read = _global_finger;
3805 HeapWord* cur = read;
3806 while (f > read) {
3807 cur = read;
3808 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3809 if (cur == read) {
3810 // our cas succeeded
3811 assert(_global_finger >= f, "protocol consistency");
3812 break;
3813 }
3814 }
3815 }
3817 // This is really inefficient, and should be redone by
3818 // using (not yet available) block-read and -write interfaces to the
3819 // stack and the work_queue. XXX FIX ME !!!
3820 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3821 OopTaskQueue* work_q) {
3822 // Fast lock-free check
3823 if (ovflw_stk->length() == 0) {
3824 return false;
3825 }
3826 assert(work_q->size() == 0, "Shouldn't steal");
3827 MutexLockerEx ml(ovflw_stk->par_lock(),
3828 Mutex::_no_safepoint_check_flag);
3829 // Grab up to 1/4 the size of the work queue
3830 size_t num = MIN2((size_t)work_q->max_elems()/4,
3831 (size_t)ParGCDesiredObjsFromOverflowList);
3832 num = MIN2(num, ovflw_stk->length());
3833 for (int i = (int) num; i > 0; i--) {
3834 oop cur = ovflw_stk->pop();
3835 assert(cur != NULL, "Counted wrong?");
3836 work_q->push(cur);
3837 }
3838 return num > 0;
3839 }
3841 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3842 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3843 int n_tasks = pst->n_tasks();
3844 // We allow that there may be no tasks to do here because
3845 // we are restarting after a stack overflow.
3846 assert(pst->valid() || n_tasks == 0, "Uninitializd use?");
3847 int nth_task = 0;
3849 HeapWord* start = sp->bottom();
3850 size_t chunk_size = sp->marking_task_size();
3851 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3852 // Having claimed the nth task in this space,
3853 // compute the chunk that it corresponds to:
3854 MemRegion span = MemRegion(start + nth_task*chunk_size,
3855 start + (nth_task+1)*chunk_size);
3856 // Try and bump the global finger via a CAS;
3857 // note that we need to do the global finger bump
3858 // _before_ taking the intersection below, because
3859 // the task corresponding to that region will be
3860 // deemed done even if the used_region() expands
3861 // because of allocation -- as it almost certainly will
3862 // during start-up while the threads yield in the
3863 // closure below.
3864 HeapWord* finger = span.end();
3865 bump_global_finger(finger); // atomically
3866 // There are null tasks here corresponding to chunks
3867 // beyond the "top" address of the space.
3868 span = span.intersection(sp->used_region());
3869 if (!span.is_empty()) { // Non-null task
3870 // We want to skip the first object because
3871 // the protocol is to scan any object in its entirety
3872 // that _starts_ in this span; a fortiori, any
3873 // object starting in an earlier span is scanned
3874 // as part of an earlier claimed task.
3875 // Below we use the "careful" version of block_start
3876 // so we do not try to navigate uninitialized objects.
3877 HeapWord* prev_obj = sp->block_start_careful(span.start());
3878 // Below we use a variant of block_size that uses the
3879 // Printezis bits to avoid waiting for allocated
3880 // objects to become initialized/parsable.
3881 while (prev_obj < span.start()) {
3882 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3883 if (sz > 0) {
3884 prev_obj += sz;
3885 } else {
3886 // In this case we may end up doing a bit of redundant
3887 // scanning, but that appears unavoidable, short of
3888 // locking the free list locks; see bug 6324141.
3889 break;
3890 }
3891 }
3892 if (prev_obj < span.end()) {
3893 MemRegion my_span = MemRegion(prev_obj, span.end());
3894 // Do the marking work within a non-empty span --
3895 // the last argument to the constructor indicates whether the
3896 // iteration should be incremental with periodic yields.
3897 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3898 &_collector->_markBitMap,
3899 work_queue(i),
3900 &_collector->_markStack,
3901 &_collector->_revisitStack,
3902 _asynch);
3903 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3904 } // else nothing to do for this task
3905 } // else nothing to do for this task
3906 }
3907 // We'd be tempted to assert here that since there are no
3908 // more tasks left to claim in this space, the global_finger
3909 // must exceed space->top() and a fortiori space->end(). However,
3910 // that would not quite be correct because the bumping of
3911 // global_finger occurs strictly after the claiming of a task,
3912 // so by the time we reach here the global finger may not yet
3913 // have been bumped up by the thread that claimed the last
3914 // task.
3915 pst->all_tasks_completed();
3916 }
3918 class Par_ConcMarkingClosure: public OopClosure {
3919 private:
3920 CMSCollector* _collector;
3921 MemRegion _span;
3922 CMSBitMap* _bit_map;
3923 CMSMarkStack* _overflow_stack;
3924 CMSMarkStack* _revisit_stack; // XXXXXX Check proper use
3925 OopTaskQueue* _work_queue;
3926 protected:
3927 DO_OOP_WORK_DEFN
3928 public:
3929 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
3930 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3931 _collector(collector),
3932 _span(_collector->_span),
3933 _work_queue(work_queue),
3934 _bit_map(bit_map),
3935 _overflow_stack(overflow_stack) { } // need to initialize revisit stack etc.
3936 virtual void do_oop(oop* p);
3937 virtual void do_oop(narrowOop* p);
3938 void trim_queue(size_t max);
3939 void handle_stack_overflow(HeapWord* lost);
3940 };
3942 // Grey object rescan during work stealing phase --
3943 // the salient assumption here is that stolen oops must
3944 // always be initialized, so we do not need to check for
3945 // uninitialized objects before scanning here.
3946 void Par_ConcMarkingClosure::do_oop(oop obj) {
3947 assert(obj->is_oop_or_null(), "expected an oop or NULL");
3948 HeapWord* addr = (HeapWord*)obj;
3949 // Check if oop points into the CMS generation
3950 // and is not marked
3951 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3952 // a white object ...
3953 // If we manage to "claim" the object, by being the
3954 // first thread to mark it, then we push it on our
3955 // marking stack
3956 if (_bit_map->par_mark(addr)) { // ... now grey
3957 // push on work queue (grey set)
3958 bool simulate_overflow = false;
3959 NOT_PRODUCT(
3960 if (CMSMarkStackOverflowALot &&
3961 _collector->simulate_overflow()) {
3962 // simulate a stack overflow
3963 simulate_overflow = true;
3964 }
3965 )
3966 if (simulate_overflow ||
3967 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3968 // stack overflow
3969 if (PrintCMSStatistics != 0) {
3970 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3971 SIZE_FORMAT, _overflow_stack->capacity());
3972 }
3973 // We cannot assert that the overflow stack is full because
3974 // it may have been emptied since.
3975 assert(simulate_overflow ||
3976 _work_queue->size() == _work_queue->max_elems(),
3977 "Else push should have succeeded");
3978 handle_stack_overflow(addr);
3979 }
3980 } // Else, some other thread got there first
3981 }
3982 }
3984 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3985 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3987 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3988 while (_work_queue->size() > max) {
3989 oop new_oop;
3990 if (_work_queue->pop_local(new_oop)) {
3991 assert(new_oop->is_oop(), "Should be an oop");
3992 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3993 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3994 assert(new_oop->is_parsable(), "Should be parsable");
3995 new_oop->oop_iterate(this); // do_oop() above
3996 }
3997 }
3998 }
4000 // Upon stack overflow, we discard (part of) the stack,
4001 // remembering the least address amongst those discarded
4002 // in CMSCollector's _restart_address.
4003 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4004 // We need to do this under a mutex to prevent other
4005 // workers from interfering with the expansion below.
4006 MutexLockerEx ml(_overflow_stack->par_lock(),
4007 Mutex::_no_safepoint_check_flag);
4008 // Remember the least grey address discarded
4009 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4010 _collector->lower_restart_addr(ra);
4011 _overflow_stack->reset(); // discard stack contents
4012 _overflow_stack->expand(); // expand the stack if possible
4013 }
4016 void CMSConcMarkingTask::do_work_steal(int i) {
4017 OopTaskQueue* work_q = work_queue(i);
4018 oop obj_to_scan;
4019 CMSBitMap* bm = &(_collector->_markBitMap);
4020 CMSMarkStack* ovflw = &(_collector->_markStack);
4021 int* seed = _collector->hash_seed(i);
4022 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw);
4023 while (true) {
4024 cl.trim_queue(0);
4025 assert(work_q->size() == 0, "Should have been emptied above");
4026 if (get_work_from_overflow_stack(ovflw, work_q)) {
4027 // Can't assert below because the work obtained from the
4028 // overflow stack may already have been stolen from us.
4029 // assert(work_q->size() > 0, "Work from overflow stack");
4030 continue;
4031 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4032 assert(obj_to_scan->is_oop(), "Should be an oop");
4033 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4034 obj_to_scan->oop_iterate(&cl);
4035 } else if (terminator()->offer_termination()) {
4036 assert(work_q->size() == 0, "Impossible!");
4037 break;
4038 }
4039 }
4040 }
4042 // This is run by the CMS (coordinator) thread.
4043 void CMSConcMarkingTask::coordinator_yield() {
4044 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4045 "CMS thread should hold CMS token");
4047 // First give up the locks, then yield, then re-lock
4048 // We should probably use a constructor/destructor idiom to
4049 // do this unlock/lock or modify the MutexUnlocker class to
4050 // serve our purpose. XXX
4051 assert_lock_strong(_bit_map_lock);
4052 _bit_map_lock->unlock();
4053 ConcurrentMarkSweepThread::desynchronize(true);
4054 ConcurrentMarkSweepThread::acknowledge_yield_request();
4055 _collector->stopTimer();
4056 if (PrintCMSStatistics != 0) {
4057 _collector->incrementYields();
4058 }
4059 _collector->icms_wait();
4061 // It is possible for whichever thread initiated the yield request
4062 // not to get a chance to wake up and take the bitmap lock between
4063 // this thread releasing it and reacquiring it. So, while the
4064 // should_yield() flag is on, let's sleep for a bit to give the
4065 // other thread a chance to wake up. The limit imposed on the number
4066 // of iterations is defensive, to avoid any unforseen circumstances
4067 // putting us into an infinite loop. Since it's always been this
4068 // (coordinator_yield()) method that was observed to cause the
4069 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4070 // which is by default non-zero. For the other seven methods that
4071 // also perform the yield operation, as are using a different
4072 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4073 // can enable the sleeping for those methods too, if necessary.
4074 // See 6442774.
4075 //
4076 // We really need to reconsider the synchronization between the GC
4077 // thread and the yield-requesting threads in the future and we
4078 // should really use wait/notify, which is the recommended
4079 // way of doing this type of interaction. Additionally, we should
4080 // consolidate the eight methods that do the yield operation and they
4081 // are almost identical into one for better maintenability and
4082 // readability. See 6445193.
4083 //
4084 // Tony 2006.06.29
4085 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4086 ConcurrentMarkSweepThread::should_yield() &&
4087 !CMSCollector::foregroundGCIsActive(); ++i) {
4088 os::sleep(Thread::current(), 1, false);
4089 ConcurrentMarkSweepThread::acknowledge_yield_request();
4090 }
4092 ConcurrentMarkSweepThread::synchronize(true);
4093 _bit_map_lock->lock_without_safepoint_check();
4094 _collector->startTimer();
4095 }
4097 bool CMSCollector::do_marking_mt(bool asynch) {
4098 assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition");
4099 // In the future this would be determined ergonomically, based
4100 // on #cpu's, # active mutator threads (and load), and mutation rate.
4101 int num_workers = ParallelCMSThreads;
4103 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4104 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4106 CMSConcMarkingTask tsk(this, cms_space, perm_space,
4107 asynch, num_workers /* number requested XXX */,
4108 conc_workers(), task_queues());
4110 // Since the actual number of workers we get may be different
4111 // from the number we requested above, do we need to do anything different
4112 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4113 // class?? XXX
4114 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4115 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4117 // Refs discovery is already non-atomic.
4118 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4119 // Mutate the Refs discovery so it is MT during the
4120 // multi-threaded marking phase.
4121 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4123 conc_workers()->start_task(&tsk);
4124 while (tsk.yielded()) {
4125 tsk.coordinator_yield();
4126 conc_workers()->continue_task(&tsk);
4127 }
4128 // If the task was aborted, _restart_addr will be non-NULL
4129 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4130 while (_restart_addr != NULL) {
4131 // XXX For now we do not make use of ABORTED state and have not
4132 // yet implemented the right abort semantics (even in the original
4133 // single-threaded CMS case). That needs some more investigation
4134 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4135 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4136 // If _restart_addr is non-NULL, a marking stack overflow
4137 // occured; we need to do a fresh marking iteration from the
4138 // indicated restart address.
4139 if (_foregroundGCIsActive && asynch) {
4140 // We may be running into repeated stack overflows, having
4141 // reached the limit of the stack size, while making very
4142 // slow forward progress. It may be best to bail out and
4143 // let the foreground collector do its job.
4144 // Clear _restart_addr, so that foreground GC
4145 // works from scratch. This avoids the headache of
4146 // a "rescan" which would otherwise be needed because
4147 // of the dirty mod union table & card table.
4148 _restart_addr = NULL;
4149 return false;
4150 }
4151 // Adjust the task to restart from _restart_addr
4152 tsk.reset(_restart_addr);
4153 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4154 _restart_addr);
4155 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4156 _restart_addr);
4157 _restart_addr = NULL;
4158 // Get the workers going again
4159 conc_workers()->start_task(&tsk);
4160 while (tsk.yielded()) {
4161 tsk.coordinator_yield();
4162 conc_workers()->continue_task(&tsk);
4163 }
4164 }
4165 assert(tsk.completed(), "Inconsistency");
4166 assert(tsk.result() == true, "Inconsistency");
4167 return true;
4168 }
4170 bool CMSCollector::do_marking_st(bool asynch) {
4171 ResourceMark rm;
4172 HandleMark hm;
4174 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4175 &_markStack, &_revisitStack, CMSYield && asynch);
4176 // the last argument to iterate indicates whether the iteration
4177 // should be incremental with periodic yields.
4178 _markBitMap.iterate(&markFromRootsClosure);
4179 // If _restart_addr is non-NULL, a marking stack overflow
4180 // occured; we need to do a fresh iteration from the
4181 // indicated restart address.
4182 while (_restart_addr != NULL) {
4183 if (_foregroundGCIsActive && asynch) {
4184 // We may be running into repeated stack overflows, having
4185 // reached the limit of the stack size, while making very
4186 // slow forward progress. It may be best to bail out and
4187 // let the foreground collector do its job.
4188 // Clear _restart_addr, so that foreground GC
4189 // works from scratch. This avoids the headache of
4190 // a "rescan" which would otherwise be needed because
4191 // of the dirty mod union table & card table.
4192 _restart_addr = NULL;
4193 return false; // indicating failure to complete marking
4194 }
4195 // Deal with stack overflow:
4196 // we restart marking from _restart_addr
4197 HeapWord* ra = _restart_addr;
4198 markFromRootsClosure.reset(ra);
4199 _restart_addr = NULL;
4200 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4201 }
4202 return true;
4203 }
4205 void CMSCollector::preclean() {
4206 check_correct_thread_executing();
4207 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4208 verify_work_stacks_empty();
4209 verify_overflow_empty();
4210 _abort_preclean = false;
4211 if (CMSPrecleaningEnabled) {
4212 _eden_chunk_index = 0;
4213 size_t used = get_eden_used();
4214 size_t capacity = get_eden_capacity();
4215 // Don't start sampling unless we will get sufficiently
4216 // many samples.
4217 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4218 * CMSScheduleRemarkEdenPenetration)) {
4219 _start_sampling = true;
4220 } else {
4221 _start_sampling = false;
4222 }
4223 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4224 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4225 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4226 }
4227 CMSTokenSync x(true); // is cms thread
4228 if (CMSPrecleaningEnabled) {
4229 sample_eden();
4230 _collectorState = AbortablePreclean;
4231 } else {
4232 _collectorState = FinalMarking;
4233 }
4234 verify_work_stacks_empty();
4235 verify_overflow_empty();
4236 }
4238 // Try and schedule the remark such that young gen
4239 // occupancy is CMSScheduleRemarkEdenPenetration %.
4240 void CMSCollector::abortable_preclean() {
4241 check_correct_thread_executing();
4242 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4243 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4245 // If Eden's current occupancy is below this threshold,
4246 // immediately schedule the remark; else preclean
4247 // past the next scavenge in an effort to
4248 // schedule the pause as described avove. By choosing
4249 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4250 // we will never do an actual abortable preclean cycle.
4251 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4252 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4253 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4254 // We need more smarts in the abortable preclean
4255 // loop below to deal with cases where allocation
4256 // in young gen is very very slow, and our precleaning
4257 // is running a losing race against a horde of
4258 // mutators intent on flooding us with CMS updates
4259 // (dirty cards).
4260 // One, admittedly dumb, strategy is to give up
4261 // after a certain number of abortable precleaning loops
4262 // or after a certain maximum time. We want to make
4263 // this smarter in the next iteration.
4264 // XXX FIX ME!!! YSR
4265 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4266 while (!(should_abort_preclean() ||
4267 ConcurrentMarkSweepThread::should_terminate())) {
4268 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4269 cumworkdone += workdone;
4270 loops++;
4271 // Voluntarily terminate abortable preclean phase if we have
4272 // been at it for too long.
4273 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4274 loops >= CMSMaxAbortablePrecleanLoops) {
4275 if (PrintGCDetails) {
4276 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4277 }
4278 break;
4279 }
4280 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4281 if (PrintGCDetails) {
4282 gclog_or_tty->print(" CMS: abort preclean due to time ");
4283 }
4284 break;
4285 }
4286 // If we are doing little work each iteration, we should
4287 // take a short break.
4288 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4289 // Sleep for some time, waiting for work to accumulate
4290 stopTimer();
4291 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4292 startTimer();
4293 waited++;
4294 }
4295 }
4296 if (PrintCMSStatistics > 0) {
4297 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4298 loops, waited, cumworkdone);
4299 }
4300 }
4301 CMSTokenSync x(true); // is cms thread
4302 if (_collectorState != Idling) {
4303 assert(_collectorState == AbortablePreclean,
4304 "Spontaneous state transition?");
4305 _collectorState = FinalMarking;
4306 } // Else, a foreground collection completed this CMS cycle.
4307 return;
4308 }
4310 // Respond to an Eden sampling opportunity
4311 void CMSCollector::sample_eden() {
4312 // Make sure a young gc cannot sneak in between our
4313 // reading and recording of a sample.
4314 assert(Thread::current()->is_ConcurrentGC_thread(),
4315 "Only the cms thread may collect Eden samples");
4316 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4317 "Should collect samples while holding CMS token");
4318 if (!_start_sampling) {
4319 return;
4320 }
4321 if (_eden_chunk_array) {
4322 if (_eden_chunk_index < _eden_chunk_capacity) {
4323 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4324 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4325 "Unexpected state of Eden");
4326 // We'd like to check that what we just sampled is an oop-start address;
4327 // however, we cannot do that here since the object may not yet have been
4328 // initialized. So we'll instead do the check when we _use_ this sample
4329 // later.
4330 if (_eden_chunk_index == 0 ||
4331 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4332 _eden_chunk_array[_eden_chunk_index-1])
4333 >= CMSSamplingGrain)) {
4334 _eden_chunk_index++; // commit sample
4335 }
4336 }
4337 }
4338 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4339 size_t used = get_eden_used();
4340 size_t capacity = get_eden_capacity();
4341 assert(used <= capacity, "Unexpected state of Eden");
4342 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4343 _abort_preclean = true;
4344 }
4345 }
4346 }
4349 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4350 assert(_collectorState == Precleaning ||
4351 _collectorState == AbortablePreclean, "incorrect state");
4352 ResourceMark rm;
4353 HandleMark hm;
4354 // Do one pass of scrubbing the discovered reference lists
4355 // to remove any reference objects with strongly-reachable
4356 // referents.
4357 if (clean_refs) {
4358 ReferenceProcessor* rp = ref_processor();
4359 CMSPrecleanRefsYieldClosure yield_cl(this);
4360 assert(rp->span().equals(_span), "Spans should be equal");
4361 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4362 &_markStack);
4363 CMSDrainMarkingStackClosure complete_trace(this,
4364 _span, &_markBitMap, &_markStack,
4365 &keep_alive);
4367 // We don't want this step to interfere with a young
4368 // collection because we don't want to take CPU
4369 // or memory bandwidth away from the young GC threads
4370 // (which may be as many as there are CPUs).
4371 // Note that we don't need to protect ourselves from
4372 // interference with mutators because they can't
4373 // manipulate the discovered reference lists nor affect
4374 // the computed reachability of the referents, the
4375 // only properties manipulated by the precleaning
4376 // of these reference lists.
4377 stopTimer();
4378 CMSTokenSyncWithLocks x(true /* is cms thread */,
4379 bitMapLock());
4380 startTimer();
4381 sample_eden();
4382 // The following will yield to allow foreground
4383 // collection to proceed promptly. XXX YSR:
4384 // The code in this method may need further
4385 // tweaking for better performance and some restructuring
4386 // for cleaner interfaces.
4387 rp->preclean_discovered_references(
4388 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4389 &yield_cl);
4390 }
4392 if (clean_survivor) { // preclean the active survivor space(s)
4393 assert(_young_gen->kind() == Generation::DefNew ||
4394 _young_gen->kind() == Generation::ParNew ||
4395 _young_gen->kind() == Generation::ASParNew,
4396 "incorrect type for cast");
4397 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4398 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4399 &_markBitMap, &_modUnionTable,
4400 &_markStack, &_revisitStack,
4401 true /* precleaning phase */);
4402 stopTimer();
4403 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4404 bitMapLock());
4405 startTimer();
4406 unsigned int before_count =
4407 GenCollectedHeap::heap()->total_collections();
4408 SurvivorSpacePrecleanClosure
4409 sss_cl(this, _span, &_markBitMap, &_markStack,
4410 &pam_cl, before_count, CMSYield);
4411 dng->from()->object_iterate_careful(&sss_cl);
4412 dng->to()->object_iterate_careful(&sss_cl);
4413 }
4414 MarkRefsIntoAndScanClosure
4415 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4416 &_markStack, &_revisitStack, this, CMSYield,
4417 true /* precleaning phase */);
4418 // CAUTION: The following closure has persistent state that may need to
4419 // be reset upon a decrease in the sequence of addresses it
4420 // processes.
4421 ScanMarkedObjectsAgainCarefullyClosure
4422 smoac_cl(this, _span,
4423 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4425 // Preclean dirty cards in ModUnionTable and CardTable using
4426 // appropriate convergence criterion;
4427 // repeat CMSPrecleanIter times unless we find that
4428 // we are losing.
4429 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4430 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4431 "Bad convergence multiplier");
4432 assert(CMSPrecleanThreshold >= 100,
4433 "Unreasonably low CMSPrecleanThreshold");
4435 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4436 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4437 numIter < CMSPrecleanIter;
4438 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4439 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4440 if (CMSPermGenPrecleaningEnabled) {
4441 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4442 }
4443 if (Verbose && PrintGCDetails) {
4444 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4445 }
4446 // Either there are very few dirty cards, so re-mark
4447 // pause will be small anyway, or our pre-cleaning isn't
4448 // that much faster than the rate at which cards are being
4449 // dirtied, so we might as well stop and re-mark since
4450 // precleaning won't improve our re-mark time by much.
4451 if (curNumCards <= CMSPrecleanThreshold ||
4452 (numIter > 0 &&
4453 (curNumCards * CMSPrecleanDenominator >
4454 lastNumCards * CMSPrecleanNumerator))) {
4455 numIter++;
4456 cumNumCards += curNumCards;
4457 break;
4458 }
4459 }
4460 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4461 if (CMSPermGenPrecleaningEnabled) {
4462 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4463 }
4464 cumNumCards += curNumCards;
4465 if (PrintGCDetails && PrintCMSStatistics != 0) {
4466 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4467 curNumCards, cumNumCards, numIter);
4468 }
4469 return cumNumCards; // as a measure of useful work done
4470 }
4472 // PRECLEANING NOTES:
4473 // Precleaning involves:
4474 // . reading the bits of the modUnionTable and clearing the set bits.
4475 // . For the cards corresponding to the set bits, we scan the
4476 // objects on those cards. This means we need the free_list_lock
4477 // so that we can safely iterate over the CMS space when scanning
4478 // for oops.
4479 // . When we scan the objects, we'll be both reading and setting
4480 // marks in the marking bit map, so we'll need the marking bit map.
4481 // . For protecting _collector_state transitions, we take the CGC_lock.
4482 // Note that any races in the reading of of card table entries by the
4483 // CMS thread on the one hand and the clearing of those entries by the
4484 // VM thread or the setting of those entries by the mutator threads on the
4485 // other are quite benign. However, for efficiency it makes sense to keep
4486 // the VM thread from racing with the CMS thread while the latter is
4487 // dirty card info to the modUnionTable. We therefore also use the
4488 // CGC_lock to protect the reading of the card table and the mod union
4489 // table by the CM thread.
4490 // . We run concurrently with mutator updates, so scanning
4491 // needs to be done carefully -- we should not try to scan
4492 // potentially uninitialized objects.
4493 //
4494 // Locking strategy: While holding the CGC_lock, we scan over and
4495 // reset a maximal dirty range of the mod union / card tables, then lock
4496 // the free_list_lock and bitmap lock to do a full marking, then
4497 // release these locks; and repeat the cycle. This allows for a
4498 // certain amount of fairness in the sharing of these locks between
4499 // the CMS collector on the one hand, and the VM thread and the
4500 // mutators on the other.
4502 // NOTE: preclean_mod_union_table() and preclean_card_table()
4503 // further below are largely identical; if you need to modify
4504 // one of these methods, please check the other method too.
4506 size_t CMSCollector::preclean_mod_union_table(
4507 ConcurrentMarkSweepGeneration* gen,
4508 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4509 verify_work_stacks_empty();
4510 verify_overflow_empty();
4512 // strategy: starting with the first card, accumulate contiguous
4513 // ranges of dirty cards; clear these cards, then scan the region
4514 // covered by these cards.
4516 // Since all of the MUT is committed ahead, we can just use
4517 // that, in case the generations expand while we are precleaning.
4518 // It might also be fine to just use the committed part of the
4519 // generation, but we might potentially miss cards when the
4520 // generation is rapidly expanding while we are in the midst
4521 // of precleaning.
4522 HeapWord* startAddr = gen->reserved().start();
4523 HeapWord* endAddr = gen->reserved().end();
4525 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4527 size_t numDirtyCards, cumNumDirtyCards;
4528 HeapWord *nextAddr, *lastAddr;
4529 for (cumNumDirtyCards = numDirtyCards = 0,
4530 nextAddr = lastAddr = startAddr;
4531 nextAddr < endAddr;
4532 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4534 ResourceMark rm;
4535 HandleMark hm;
4537 MemRegion dirtyRegion;
4538 {
4539 stopTimer();
4540 CMSTokenSync ts(true);
4541 startTimer();
4542 sample_eden();
4543 // Get dirty region starting at nextOffset (inclusive),
4544 // simultaneously clearing it.
4545 dirtyRegion =
4546 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4547 assert(dirtyRegion.start() >= nextAddr,
4548 "returned region inconsistent?");
4549 }
4550 // Remember where the next search should begin.
4551 // The returned region (if non-empty) is a right open interval,
4552 // so lastOffset is obtained from the right end of that
4553 // interval.
4554 lastAddr = dirtyRegion.end();
4555 // Should do something more transparent and less hacky XXX
4556 numDirtyCards =
4557 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4559 // We'll scan the cards in the dirty region (with periodic
4560 // yields for foreground GC as needed).
4561 if (!dirtyRegion.is_empty()) {
4562 assert(numDirtyCards > 0, "consistency check");
4563 HeapWord* stop_point = NULL;
4564 {
4565 stopTimer();
4566 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4567 bitMapLock());
4568 startTimer();
4569 verify_work_stacks_empty();
4570 verify_overflow_empty();
4571 sample_eden();
4572 stop_point =
4573 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4574 }
4575 if (stop_point != NULL) {
4576 // The careful iteration stopped early either because it found an
4577 // uninitialized object, or because we were in the midst of an
4578 // "abortable preclean", which should now be aborted. Redirty
4579 // the bits corresponding to the partially-scanned or unscanned
4580 // cards. We'll either restart at the next block boundary or
4581 // abort the preclean.
4582 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4583 (_collectorState == AbortablePreclean && should_abort_preclean()),
4584 "Unparsable objects should only be in perm gen.");
4586 stopTimer();
4587 CMSTokenSyncWithLocks ts(true, bitMapLock());
4588 startTimer();
4589 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4590 if (should_abort_preclean()) {
4591 break; // out of preclean loop
4592 } else {
4593 // Compute the next address at which preclean should pick up;
4594 // might need bitMapLock in order to read P-bits.
4595 lastAddr = next_card_start_after_block(stop_point);
4596 }
4597 }
4598 } else {
4599 assert(lastAddr == endAddr, "consistency check");
4600 assert(numDirtyCards == 0, "consistency check");
4601 break;
4602 }
4603 }
4604 verify_work_stacks_empty();
4605 verify_overflow_empty();
4606 return cumNumDirtyCards;
4607 }
4609 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4610 // below are largely identical; if you need to modify
4611 // one of these methods, please check the other method too.
4613 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4614 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4615 // strategy: it's similar to precleamModUnionTable above, in that
4616 // we accumulate contiguous ranges of dirty cards, mark these cards
4617 // precleaned, then scan the region covered by these cards.
4618 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4619 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4621 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4623 size_t numDirtyCards, cumNumDirtyCards;
4624 HeapWord *lastAddr, *nextAddr;
4626 for (cumNumDirtyCards = numDirtyCards = 0,
4627 nextAddr = lastAddr = startAddr;
4628 nextAddr < endAddr;
4629 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4631 ResourceMark rm;
4632 HandleMark hm;
4634 MemRegion dirtyRegion;
4635 {
4636 // See comments in "Precleaning notes" above on why we
4637 // do this locking. XXX Could the locking overheads be
4638 // too high when dirty cards are sparse? [I don't think so.]
4639 stopTimer();
4640 CMSTokenSync x(true); // is cms thread
4641 startTimer();
4642 sample_eden();
4643 // Get and clear dirty region from card table
4644 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4645 MemRegion(nextAddr, endAddr),
4646 true,
4647 CardTableModRefBS::precleaned_card_val());
4649 assert(dirtyRegion.start() >= nextAddr,
4650 "returned region inconsistent?");
4651 }
4652 lastAddr = dirtyRegion.end();
4653 numDirtyCards =
4654 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4656 if (!dirtyRegion.is_empty()) {
4657 stopTimer();
4658 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4659 startTimer();
4660 sample_eden();
4661 verify_work_stacks_empty();
4662 verify_overflow_empty();
4663 HeapWord* stop_point =
4664 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4665 if (stop_point != NULL) {
4666 // The careful iteration stopped early because it found an
4667 // uninitialized object. Redirty the bits corresponding to the
4668 // partially-scanned or unscanned cards, and start again at the
4669 // next block boundary.
4670 assert(CMSPermGenPrecleaningEnabled ||
4671 (_collectorState == AbortablePreclean && should_abort_preclean()),
4672 "Unparsable objects should only be in perm gen.");
4673 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4674 if (should_abort_preclean()) {
4675 break; // out of preclean loop
4676 } else {
4677 // Compute the next address at which preclean should pick up.
4678 lastAddr = next_card_start_after_block(stop_point);
4679 }
4680 }
4681 } else {
4682 break;
4683 }
4684 }
4685 verify_work_stacks_empty();
4686 verify_overflow_empty();
4687 return cumNumDirtyCards;
4688 }
4690 void CMSCollector::checkpointRootsFinal(bool asynch,
4691 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4692 assert(_collectorState == FinalMarking, "incorrect state transition?");
4693 check_correct_thread_executing();
4694 // world is stopped at this checkpoint
4695 assert(SafepointSynchronize::is_at_safepoint(),
4696 "world should be stopped");
4697 verify_work_stacks_empty();
4698 verify_overflow_empty();
4700 SpecializationStats::clear();
4701 if (PrintGCDetails) {
4702 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4703 _young_gen->used() / K,
4704 _young_gen->capacity() / K);
4705 }
4706 if (asynch) {
4707 if (CMSScavengeBeforeRemark) {
4708 GenCollectedHeap* gch = GenCollectedHeap::heap();
4709 // Temporarily set flag to false, GCH->do_collection will
4710 // expect it to be false and set to true
4711 FlagSetting fl(gch->_is_gc_active, false);
4712 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4713 PrintGCDetails && Verbose, true, gclog_or_tty);)
4714 int level = _cmsGen->level() - 1;
4715 if (level >= 0) {
4716 gch->do_collection(true, // full (i.e. force, see below)
4717 false, // !clear_all_soft_refs
4718 0, // size
4719 false, // is_tlab
4720 level // max_level
4721 );
4722 }
4723 }
4724 FreelistLocker x(this);
4725 MutexLockerEx y(bitMapLock(),
4726 Mutex::_no_safepoint_check_flag);
4727 assert(!init_mark_was_synchronous, "but that's impossible!");
4728 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4729 } else {
4730 // already have all the locks
4731 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4732 init_mark_was_synchronous);
4733 }
4734 verify_work_stacks_empty();
4735 verify_overflow_empty();
4736 SpecializationStats::print();
4737 }
4739 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4740 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4742 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4744 assert(haveFreelistLocks(), "must have free list locks");
4745 assert_lock_strong(bitMapLock());
4747 if (UseAdaptiveSizePolicy) {
4748 size_policy()->checkpoint_roots_final_begin();
4749 }
4751 ResourceMark rm;
4752 HandleMark hm;
4754 GenCollectedHeap* gch = GenCollectedHeap::heap();
4756 if (should_unload_classes()) {
4757 CodeCache::gc_prologue();
4758 }
4759 assert(haveFreelistLocks(), "must have free list locks");
4760 assert_lock_strong(bitMapLock());
4762 if (!init_mark_was_synchronous) {
4763 // We might assume that we need not fill TLAB's when
4764 // CMSScavengeBeforeRemark is set, because we may have just done
4765 // a scavenge which would have filled all TLAB's -- and besides
4766 // Eden would be empty. This however may not always be the case --
4767 // for instance although we asked for a scavenge, it may not have
4768 // happened because of a JNI critical section. We probably need
4769 // a policy for deciding whether we can in that case wait until
4770 // the critical section releases and then do the remark following
4771 // the scavenge, and skip it here. In the absence of that policy,
4772 // or of an indication of whether the scavenge did indeed occur,
4773 // we cannot rely on TLAB's having been filled and must do
4774 // so here just in case a scavenge did not happen.
4775 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4776 // Update the saved marks which may affect the root scans.
4777 gch->save_marks();
4779 {
4780 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4782 // Note on the role of the mod union table:
4783 // Since the marker in "markFromRoots" marks concurrently with
4784 // mutators, it is possible for some reachable objects not to have been
4785 // scanned. For instance, an only reference to an object A was
4786 // placed in object B after the marker scanned B. Unless B is rescanned,
4787 // A would be collected. Such updates to references in marked objects
4788 // are detected via the mod union table which is the set of all cards
4789 // dirtied since the first checkpoint in this GC cycle and prior to
4790 // the most recent young generation GC, minus those cleaned up by the
4791 // concurrent precleaning.
4792 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
4793 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4794 do_remark_parallel();
4795 } else {
4796 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4797 gclog_or_tty);
4798 do_remark_non_parallel();
4799 }
4800 }
4801 } else {
4802 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4803 // The initial mark was stop-world, so there's no rescanning to
4804 // do; go straight on to the next step below.
4805 }
4806 verify_work_stacks_empty();
4807 verify_overflow_empty();
4809 {
4810 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4811 refProcessingWork(asynch, clear_all_soft_refs);
4812 }
4813 verify_work_stacks_empty();
4814 verify_overflow_empty();
4816 if (should_unload_classes()) {
4817 CodeCache::gc_epilogue();
4818 }
4820 // If we encountered any (marking stack / work queue) overflow
4821 // events during the current CMS cycle, take appropriate
4822 // remedial measures, where possible, so as to try and avoid
4823 // recurrence of that condition.
4824 assert(_markStack.isEmpty(), "No grey objects");
4825 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4826 _ser_kac_ovflw;
4827 if (ser_ovflw > 0) {
4828 if (PrintCMSStatistics != 0) {
4829 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4830 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4831 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4832 _ser_kac_ovflw);
4833 }
4834 _markStack.expand();
4835 _ser_pmc_remark_ovflw = 0;
4836 _ser_pmc_preclean_ovflw = 0;
4837 _ser_kac_ovflw = 0;
4838 }
4839 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4840 if (PrintCMSStatistics != 0) {
4841 gclog_or_tty->print_cr("Work queue overflow (benign) "
4842 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4843 _par_pmc_remark_ovflw, _par_kac_ovflw);
4844 }
4845 _par_pmc_remark_ovflw = 0;
4846 _par_kac_ovflw = 0;
4847 }
4848 if (PrintCMSStatistics != 0) {
4849 if (_markStack._hit_limit > 0) {
4850 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4851 _markStack._hit_limit);
4852 }
4853 if (_markStack._failed_double > 0) {
4854 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4855 " current capacity "SIZE_FORMAT,
4856 _markStack._failed_double,
4857 _markStack.capacity());
4858 }
4859 }
4860 _markStack._hit_limit = 0;
4861 _markStack._failed_double = 0;
4863 if ((VerifyAfterGC || VerifyDuringGC) &&
4864 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4865 verify_after_remark();
4866 }
4868 // Change under the freelistLocks.
4869 _collectorState = Sweeping;
4870 // Call isAllClear() under bitMapLock
4871 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
4872 " final marking");
4873 if (UseAdaptiveSizePolicy) {
4874 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
4875 }
4876 }
4878 // Parallel remark task
4879 class CMSParRemarkTask: public AbstractGangTask {
4880 CMSCollector* _collector;
4881 WorkGang* _workers;
4882 int _n_workers;
4883 CompactibleFreeListSpace* _cms_space;
4884 CompactibleFreeListSpace* _perm_space;
4886 // The per-thread work queues, available here for stealing.
4887 OopTaskQueueSet* _task_queues;
4888 ParallelTaskTerminator _term;
4890 public:
4891 CMSParRemarkTask(CMSCollector* collector,
4892 CompactibleFreeListSpace* cms_space,
4893 CompactibleFreeListSpace* perm_space,
4894 int n_workers, WorkGang* workers,
4895 OopTaskQueueSet* task_queues):
4896 AbstractGangTask("Rescan roots and grey objects in parallel"),
4897 _collector(collector),
4898 _cms_space(cms_space), _perm_space(perm_space),
4899 _n_workers(n_workers),
4900 _workers(workers),
4901 _task_queues(task_queues),
4902 _term(workers->total_workers(), task_queues) { }
4904 OopTaskQueueSet* task_queues() { return _task_queues; }
4906 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4908 ParallelTaskTerminator* terminator() { return &_term; }
4910 void work(int i);
4912 private:
4913 // Work method in support of parallel rescan ... of young gen spaces
4914 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
4915 ContiguousSpace* space,
4916 HeapWord** chunk_array, size_t chunk_top);
4918 // ... of dirty cards in old space
4919 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4920 Par_MarkRefsIntoAndScanClosure* cl);
4922 // ... work stealing for the above
4923 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4924 };
4926 void CMSParRemarkTask::work(int i) {
4927 elapsedTimer _timer;
4928 ResourceMark rm;
4929 HandleMark hm;
4931 // ---------- rescan from roots --------------
4932 _timer.start();
4933 GenCollectedHeap* gch = GenCollectedHeap::heap();
4934 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4935 _collector->_span, _collector->ref_processor(),
4936 &(_collector->_markBitMap),
4937 work_queue(i), &(_collector->_revisitStack));
4939 // Rescan young gen roots first since these are likely
4940 // coarsely partitioned and may, on that account, constitute
4941 // the critical path; thus, it's best to start off that
4942 // work first.
4943 // ---------- young gen roots --------------
4944 {
4945 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
4946 EdenSpace* eden_space = dng->eden();
4947 ContiguousSpace* from_space = dng->from();
4948 ContiguousSpace* to_space = dng->to();
4950 HeapWord** eca = _collector->_eden_chunk_array;
4951 size_t ect = _collector->_eden_chunk_index;
4952 HeapWord** sca = _collector->_survivor_chunk_array;
4953 size_t sct = _collector->_survivor_chunk_index;
4955 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4956 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4958 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
4959 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
4960 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
4962 _timer.stop();
4963 if (PrintCMSStatistics != 0) {
4964 gclog_or_tty->print_cr(
4965 "Finished young gen rescan work in %dth thread: %3.3f sec",
4966 i, _timer.seconds());
4967 }
4968 }
4970 // ---------- remaining roots --------------
4971 _timer.reset();
4972 _timer.start();
4973 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
4974 false, // yg was scanned above
4975 true, // collecting perm gen
4976 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4977 NULL, &par_mrias_cl);
4978 _timer.stop();
4979 if (PrintCMSStatistics != 0) {
4980 gclog_or_tty->print_cr(
4981 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4982 i, _timer.seconds());
4983 }
4985 // ---------- rescan dirty cards ------------
4986 _timer.reset();
4987 _timer.start();
4989 // Do the rescan tasks for each of the two spaces
4990 // (cms_space and perm_space) in turn.
4991 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
4992 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
4993 _timer.stop();
4994 if (PrintCMSStatistics != 0) {
4995 gclog_or_tty->print_cr(
4996 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4997 i, _timer.seconds());
4998 }
5000 // ---------- steal work from other threads ...
5001 // ---------- ... and drain overflow list.
5002 _timer.reset();
5003 _timer.start();
5004 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5005 _timer.stop();
5006 if (PrintCMSStatistics != 0) {
5007 gclog_or_tty->print_cr(
5008 "Finished work stealing in %dth thread: %3.3f sec",
5009 i, _timer.seconds());
5010 }
5011 }
5013 void
5014 CMSParRemarkTask::do_young_space_rescan(int i,
5015 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5016 HeapWord** chunk_array, size_t chunk_top) {
5017 // Until all tasks completed:
5018 // . claim an unclaimed task
5019 // . compute region boundaries corresponding to task claimed
5020 // using chunk_array
5021 // . par_oop_iterate(cl) over that region
5023 ResourceMark rm;
5024 HandleMark hm;
5026 SequentialSubTasksDone* pst = space->par_seq_tasks();
5027 assert(pst->valid(), "Uninitialized use?");
5029 int nth_task = 0;
5030 int n_tasks = pst->n_tasks();
5032 HeapWord *start, *end;
5033 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5034 // We claimed task # nth_task; compute its boundaries.
5035 if (chunk_top == 0) { // no samples were taken
5036 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5037 start = space->bottom();
5038 end = space->top();
5039 } else if (nth_task == 0) {
5040 start = space->bottom();
5041 end = chunk_array[nth_task];
5042 } else if (nth_task < (jint)chunk_top) {
5043 assert(nth_task >= 1, "Control point invariant");
5044 start = chunk_array[nth_task - 1];
5045 end = chunk_array[nth_task];
5046 } else {
5047 assert(nth_task == (jint)chunk_top, "Control point invariant");
5048 start = chunk_array[chunk_top - 1];
5049 end = space->top();
5050 }
5051 MemRegion mr(start, end);
5052 // Verify that mr is in space
5053 assert(mr.is_empty() || space->used_region().contains(mr),
5054 "Should be in space");
5055 // Verify that "start" is an object boundary
5056 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5057 "Should be an oop");
5058 space->par_oop_iterate(mr, cl);
5059 }
5060 pst->all_tasks_completed();
5061 }
5063 void
5064 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5065 CompactibleFreeListSpace* sp, int i,
5066 Par_MarkRefsIntoAndScanClosure* cl) {
5067 // Until all tasks completed:
5068 // . claim an unclaimed task
5069 // . compute region boundaries corresponding to task claimed
5070 // . transfer dirty bits ct->mut for that region
5071 // . apply rescanclosure to dirty mut bits for that region
5073 ResourceMark rm;
5074 HandleMark hm;
5076 OopTaskQueue* work_q = work_queue(i);
5077 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5078 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5079 // CAUTION: This closure has state that persists across calls to
5080 // the work method dirty_range_iterate_clear() in that it has
5081 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5082 // use of that state in the imbedded UpwardsObjectClosure instance
5083 // assumes that the cards are always iterated (even if in parallel
5084 // by several threads) in monotonically increasing order per each
5085 // thread. This is true of the implementation below which picks
5086 // card ranges (chunks) in monotonically increasing order globally
5087 // and, a-fortiori, in monotonically increasing order per thread
5088 // (the latter order being a subsequence of the former).
5089 // If the work code below is ever reorganized into a more chaotic
5090 // work-partitioning form than the current "sequential tasks"
5091 // paradigm, the use of that persistent state will have to be
5092 // revisited and modified appropriately. See also related
5093 // bug 4756801 work on which should examine this code to make
5094 // sure that the changes there do not run counter to the
5095 // assumptions made here and necessary for correctness and
5096 // efficiency. Note also that this code might yield inefficient
5097 // behaviour in the case of very large objects that span one or
5098 // more work chunks. Such objects would potentially be scanned
5099 // several times redundantly. Work on 4756801 should try and
5100 // address that performance anomaly if at all possible. XXX
5101 MemRegion full_span = _collector->_span;
5102 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5103 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5104 MarkFromDirtyCardsClosure
5105 greyRescanClosure(_collector, full_span, // entire span of interest
5106 sp, bm, work_q, rs, cl);
5108 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5109 assert(pst->valid(), "Uninitialized use?");
5110 int nth_task = 0;
5111 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5112 MemRegion span = sp->used_region();
5113 HeapWord* start_addr = span.start();
5114 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5115 alignment);
5116 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5117 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5118 start_addr, "Check alignment");
5119 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5120 chunk_size, "Check alignment");
5122 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5123 // Having claimed the nth_task, compute corresponding mem-region,
5124 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5125 // The alignment restriction ensures that we do not need any
5126 // synchronization with other gang-workers while setting or
5127 // clearing bits in thus chunk of the MUT.
5128 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5129 start_addr + (nth_task+1)*chunk_size);
5130 // The last chunk's end might be way beyond end of the
5131 // used region. In that case pull back appropriately.
5132 if (this_span.end() > end_addr) {
5133 this_span.set_end(end_addr);
5134 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5135 }
5136 // Iterate over the dirty cards covering this chunk, marking them
5137 // precleaned, and setting the corresponding bits in the mod union
5138 // table. Since we have been careful to partition at Card and MUT-word
5139 // boundaries no synchronization is needed between parallel threads.
5140 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5141 &modUnionClosure);
5143 // Having transferred these marks into the modUnionTable,
5144 // rescan the marked objects on the dirty cards in the modUnionTable.
5145 // Even if this is at a synchronous collection, the initial marking
5146 // may have been done during an asynchronous collection so there
5147 // may be dirty bits in the mod-union table.
5148 _collector->_modUnionTable.dirty_range_iterate_clear(
5149 this_span, &greyRescanClosure);
5150 _collector->_modUnionTable.verifyNoOneBitsInRange(
5151 this_span.start(),
5152 this_span.end());
5153 }
5154 pst->all_tasks_completed(); // declare that i am done
5155 }
5157 // . see if we can share work_queues with ParNew? XXX
5158 void
5159 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5160 int* seed) {
5161 OopTaskQueue* work_q = work_queue(i);
5162 NOT_PRODUCT(int num_steals = 0;)
5163 oop obj_to_scan;
5164 CMSBitMap* bm = &(_collector->_markBitMap);
5165 size_t num_from_overflow_list =
5166 MIN2((size_t)work_q->max_elems()/4,
5167 (size_t)ParGCDesiredObjsFromOverflowList);
5169 while (true) {
5170 // Completely finish any left over work from (an) earlier round(s)
5171 cl->trim_queue(0);
5172 // Now check if there's any work in the overflow list
5173 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5174 work_q)) {
5175 // found something in global overflow list;
5176 // not yet ready to go stealing work from others.
5177 // We'd like to assert(work_q->size() != 0, ...)
5178 // because we just took work from the overflow list,
5179 // but of course we can't since all of that could have
5180 // been already stolen from us.
5181 // "He giveth and He taketh away."
5182 continue;
5183 }
5184 // Verify that we have no work before we resort to stealing
5185 assert(work_q->size() == 0, "Have work, shouldn't steal");
5186 // Try to steal from other queues that have work
5187 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5188 NOT_PRODUCT(num_steals++;)
5189 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5190 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5191 // Do scanning work
5192 obj_to_scan->oop_iterate(cl);
5193 // Loop around, finish this work, and try to steal some more
5194 } else if (terminator()->offer_termination()) {
5195 break; // nirvana from the infinite cycle
5196 }
5197 }
5198 NOT_PRODUCT(
5199 if (PrintCMSStatistics != 0) {
5200 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5201 }
5202 )
5203 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5204 "Else our work is not yet done");
5205 }
5207 // Return a thread-local PLAB recording array, as appropriate.
5208 void* CMSCollector::get_data_recorder(int thr_num) {
5209 if (_survivor_plab_array != NULL &&
5210 (CMSPLABRecordAlways ||
5211 (_collectorState > Marking && _collectorState < FinalMarking))) {
5212 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5213 ChunkArray* ca = &_survivor_plab_array[thr_num];
5214 ca->reset(); // clear it so that fresh data is recorded
5215 return (void*) ca;
5216 } else {
5217 return NULL;
5218 }
5219 }
5221 // Reset all the thread-local PLAB recording arrays
5222 void CMSCollector::reset_survivor_plab_arrays() {
5223 for (uint i = 0; i < ParallelGCThreads; i++) {
5224 _survivor_plab_array[i].reset();
5225 }
5226 }
5228 // Merge the per-thread plab arrays into the global survivor chunk
5229 // array which will provide the partitioning of the survivor space
5230 // for CMS rescan.
5231 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
5232 assert(_survivor_plab_array != NULL, "Error");
5233 assert(_survivor_chunk_array != NULL, "Error");
5234 assert(_collectorState == FinalMarking, "Error");
5235 for (uint j = 0; j < ParallelGCThreads; j++) {
5236 _cursor[j] = 0;
5237 }
5238 HeapWord* top = surv->top();
5239 size_t i;
5240 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5241 HeapWord* min_val = top; // Higher than any PLAB address
5242 uint min_tid = 0; // position of min_val this round
5243 for (uint j = 0; j < ParallelGCThreads; j++) {
5244 ChunkArray* cur_sca = &_survivor_plab_array[j];
5245 if (_cursor[j] == cur_sca->end()) {
5246 continue;
5247 }
5248 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5249 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5250 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5251 if (cur_val < min_val) {
5252 min_tid = j;
5253 min_val = cur_val;
5254 } else {
5255 assert(cur_val < top, "All recorded addresses should be less");
5256 }
5257 }
5258 // At this point min_val and min_tid are respectively
5259 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5260 // and the thread (j) that witnesses that address.
5261 // We record this address in the _survivor_chunk_array[i]
5262 // and increment _cursor[min_tid] prior to the next round i.
5263 if (min_val == top) {
5264 break;
5265 }
5266 _survivor_chunk_array[i] = min_val;
5267 _cursor[min_tid]++;
5268 }
5269 // We are all done; record the size of the _survivor_chunk_array
5270 _survivor_chunk_index = i; // exclusive: [0, i)
5271 if (PrintCMSStatistics > 0) {
5272 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5273 }
5274 // Verify that we used up all the recorded entries
5275 #ifdef ASSERT
5276 size_t total = 0;
5277 for (uint j = 0; j < ParallelGCThreads; j++) {
5278 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5279 total += _cursor[j];
5280 }
5281 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5282 // Check that the merged array is in sorted order
5283 if (total > 0) {
5284 for (size_t i = 0; i < total - 1; i++) {
5285 if (PrintCMSStatistics > 0) {
5286 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5287 i, _survivor_chunk_array[i]);
5288 }
5289 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5290 "Not sorted");
5291 }
5292 }
5293 #endif // ASSERT
5294 }
5296 // Set up the space's par_seq_tasks structure for work claiming
5297 // for parallel rescan of young gen.
5298 // See ParRescanTask where this is currently used.
5299 void
5300 CMSCollector::
5301 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5302 assert(n_threads > 0, "Unexpected n_threads argument");
5303 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5305 // Eden space
5306 {
5307 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5308 assert(!pst->valid(), "Clobbering existing data?");
5309 // Each valid entry in [0, _eden_chunk_index) represents a task.
5310 size_t n_tasks = _eden_chunk_index + 1;
5311 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5312 pst->set_par_threads(n_threads);
5313 pst->set_n_tasks((int)n_tasks);
5314 }
5316 // Merge the survivor plab arrays into _survivor_chunk_array
5317 if (_survivor_plab_array != NULL) {
5318 merge_survivor_plab_arrays(dng->from());
5319 } else {
5320 assert(_survivor_chunk_index == 0, "Error");
5321 }
5323 // To space
5324 {
5325 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5326 assert(!pst->valid(), "Clobbering existing data?");
5327 pst->set_par_threads(n_threads);
5328 pst->set_n_tasks(1);
5329 assert(pst->valid(), "Error");
5330 }
5332 // From space
5333 {
5334 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5335 assert(!pst->valid(), "Clobbering existing data?");
5336 size_t n_tasks = _survivor_chunk_index + 1;
5337 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5338 pst->set_par_threads(n_threads);
5339 pst->set_n_tasks((int)n_tasks);
5340 assert(pst->valid(), "Error");
5341 }
5342 }
5344 // Parallel version of remark
5345 void CMSCollector::do_remark_parallel() {
5346 GenCollectedHeap* gch = GenCollectedHeap::heap();
5347 WorkGang* workers = gch->workers();
5348 assert(workers != NULL, "Need parallel worker threads.");
5349 int n_workers = workers->total_workers();
5350 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5351 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5353 CMSParRemarkTask tsk(this,
5354 cms_space, perm_space,
5355 n_workers, workers, task_queues());
5357 // Set up for parallel process_strong_roots work.
5358 gch->set_par_threads(n_workers);
5359 gch->change_strong_roots_parity();
5360 // We won't be iterating over the cards in the card table updating
5361 // the younger_gen cards, so we shouldn't call the following else
5362 // the verification code as well as subsequent younger_refs_iterate
5363 // code would get confused. XXX
5364 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5366 // The young gen rescan work will not be done as part of
5367 // process_strong_roots (which currently doesn't knw how to
5368 // parallelize such a scan), but rather will be broken up into
5369 // a set of parallel tasks (via the sampling that the [abortable]
5370 // preclean phase did of EdenSpace, plus the [two] tasks of
5371 // scanning the [two] survivor spaces. Further fine-grain
5372 // parallelization of the scanning of the survivor spaces
5373 // themselves, and of precleaning of the younger gen itself
5374 // is deferred to the future.
5375 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5377 // The dirty card rescan work is broken up into a "sequence"
5378 // of parallel tasks (per constituent space) that are dynamically
5379 // claimed by the parallel threads.
5380 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5381 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5383 // It turns out that even when we're using 1 thread, doing the work in a
5384 // separate thread causes wide variance in run times. We can't help this
5385 // in the multi-threaded case, but we special-case n=1 here to get
5386 // repeatable measurements of the 1-thread overhead of the parallel code.
5387 if (n_workers > 1) {
5388 // Make refs discovery MT-safe
5389 ReferenceProcessorMTMutator mt(ref_processor(), true);
5390 workers->run_task(&tsk);
5391 } else {
5392 tsk.work(0);
5393 }
5394 gch->set_par_threads(0); // 0 ==> non-parallel.
5395 // restore, single-threaded for now, any preserved marks
5396 // as a result of work_q overflow
5397 restore_preserved_marks_if_any();
5398 }
5400 // Non-parallel version of remark
5401 void CMSCollector::do_remark_non_parallel() {
5402 ResourceMark rm;
5403 HandleMark hm;
5404 GenCollectedHeap* gch = GenCollectedHeap::heap();
5405 MarkRefsIntoAndScanClosure
5406 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5407 &_markStack, &_revisitStack, this,
5408 false /* should_yield */, false /* not precleaning */);
5409 MarkFromDirtyCardsClosure
5410 markFromDirtyCardsClosure(this, _span,
5411 NULL, // space is set further below
5412 &_markBitMap, &_markStack, &_revisitStack,
5413 &mrias_cl);
5414 {
5415 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5416 // Iterate over the dirty cards, setting the corresponding bits in the
5417 // mod union table.
5418 {
5419 ModUnionClosure modUnionClosure(&_modUnionTable);
5420 _ct->ct_bs()->dirty_card_iterate(
5421 _cmsGen->used_region(),
5422 &modUnionClosure);
5423 _ct->ct_bs()->dirty_card_iterate(
5424 _permGen->used_region(),
5425 &modUnionClosure);
5426 }
5427 // Having transferred these marks into the modUnionTable, we just need
5428 // to rescan the marked objects on the dirty cards in the modUnionTable.
5429 // The initial marking may have been done during an asynchronous
5430 // collection so there may be dirty bits in the mod-union table.
5431 const int alignment =
5432 CardTableModRefBS::card_size * BitsPerWord;
5433 {
5434 // ... First handle dirty cards in CMS gen
5435 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5436 MemRegion ur = _cmsGen->used_region();
5437 HeapWord* lb = ur.start();
5438 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5439 MemRegion cms_span(lb, ub);
5440 _modUnionTable.dirty_range_iterate_clear(cms_span,
5441 &markFromDirtyCardsClosure);
5442 verify_work_stacks_empty();
5443 if (PrintCMSStatistics != 0) {
5444 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5445 markFromDirtyCardsClosure.num_dirty_cards());
5446 }
5447 }
5448 {
5449 // .. and then repeat for dirty cards in perm gen
5450 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5451 MemRegion ur = _permGen->used_region();
5452 HeapWord* lb = ur.start();
5453 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5454 MemRegion perm_span(lb, ub);
5455 _modUnionTable.dirty_range_iterate_clear(perm_span,
5456 &markFromDirtyCardsClosure);
5457 verify_work_stacks_empty();
5458 if (PrintCMSStatistics != 0) {
5459 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5460 markFromDirtyCardsClosure.num_dirty_cards());
5461 }
5462 }
5463 }
5464 if (VerifyDuringGC &&
5465 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5466 HandleMark hm; // Discard invalid handles created during verification
5467 Universe::verify(true);
5468 }
5469 {
5470 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5472 verify_work_stacks_empty();
5474 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5475 gch->gen_process_strong_roots(_cmsGen->level(),
5476 true, // younger gens as roots
5477 true, // collecting perm gen
5478 SharedHeap::ScanningOption(roots_scanning_options()),
5479 NULL, &mrias_cl);
5480 }
5481 verify_work_stacks_empty();
5482 // Restore evacuated mark words, if any, used for overflow list links
5483 if (!CMSOverflowEarlyRestoration) {
5484 restore_preserved_marks_if_any();
5485 }
5486 verify_overflow_empty();
5487 }
5489 ////////////////////////////////////////////////////////
5490 // Parallel Reference Processing Task Proxy Class
5491 ////////////////////////////////////////////////////////
5492 class CMSRefProcTaskProxy: public AbstractGangTask {
5493 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5494 CMSCollector* _collector;
5495 CMSBitMap* _mark_bit_map;
5496 const MemRegion _span;
5497 OopTaskQueueSet* _task_queues;
5498 ParallelTaskTerminator _term;
5499 ProcessTask& _task;
5501 public:
5502 CMSRefProcTaskProxy(ProcessTask& task,
5503 CMSCollector* collector,
5504 const MemRegion& span,
5505 CMSBitMap* mark_bit_map,
5506 int total_workers,
5507 OopTaskQueueSet* task_queues):
5508 AbstractGangTask("Process referents by policy in parallel"),
5509 _task(task),
5510 _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
5511 _task_queues(task_queues),
5512 _term(total_workers, task_queues)
5513 {
5514 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5515 "Inconsistency in _span");
5516 }
5518 OopTaskQueueSet* task_queues() { return _task_queues; }
5520 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5522 ParallelTaskTerminator* terminator() { return &_term; }
5524 void do_work_steal(int i,
5525 CMSParDrainMarkingStackClosure* drain,
5526 CMSParKeepAliveClosure* keep_alive,
5527 int* seed);
5529 virtual void work(int i);
5530 };
5532 void CMSRefProcTaskProxy::work(int i) {
5533 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5534 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5535 _mark_bit_map, work_queue(i));
5536 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5537 _mark_bit_map, work_queue(i));
5538 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5539 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5540 if (_task.marks_oops_alive()) {
5541 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5542 _collector->hash_seed(i));
5543 }
5544 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5545 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5546 }
5548 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5549 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5550 EnqueueTask& _task;
5552 public:
5553 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5554 : AbstractGangTask("Enqueue reference objects in parallel"),
5555 _task(task)
5556 { }
5558 virtual void work(int i)
5559 {
5560 _task.work(i);
5561 }
5562 };
5564 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5565 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5566 _collector(collector),
5567 _span(span),
5568 _bit_map(bit_map),
5569 _work_queue(work_queue),
5570 _mark_and_push(collector, span, bit_map, work_queue),
5571 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5572 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5573 { }
5575 // . see if we can share work_queues with ParNew? XXX
5576 void CMSRefProcTaskProxy::do_work_steal(int i,
5577 CMSParDrainMarkingStackClosure* drain,
5578 CMSParKeepAliveClosure* keep_alive,
5579 int* seed) {
5580 OopTaskQueue* work_q = work_queue(i);
5581 NOT_PRODUCT(int num_steals = 0;)
5582 oop obj_to_scan;
5583 size_t num_from_overflow_list =
5584 MIN2((size_t)work_q->max_elems()/4,
5585 (size_t)ParGCDesiredObjsFromOverflowList);
5587 while (true) {
5588 // Completely finish any left over work from (an) earlier round(s)
5589 drain->trim_queue(0);
5590 // Now check if there's any work in the overflow list
5591 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5592 work_q)) {
5593 // Found something in global overflow list;
5594 // not yet ready to go stealing work from others.
5595 // We'd like to assert(work_q->size() != 0, ...)
5596 // because we just took work from the overflow list,
5597 // but of course we can't, since all of that might have
5598 // been already stolen from us.
5599 continue;
5600 }
5601 // Verify that we have no work before we resort to stealing
5602 assert(work_q->size() == 0, "Have work, shouldn't steal");
5603 // Try to steal from other queues that have work
5604 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5605 NOT_PRODUCT(num_steals++;)
5606 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5607 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5608 // Do scanning work
5609 obj_to_scan->oop_iterate(keep_alive);
5610 // Loop around, finish this work, and try to steal some more
5611 } else if (terminator()->offer_termination()) {
5612 break; // nirvana from the infinite cycle
5613 }
5614 }
5615 NOT_PRODUCT(
5616 if (PrintCMSStatistics != 0) {
5617 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5618 }
5619 )
5620 }
5622 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5623 {
5624 GenCollectedHeap* gch = GenCollectedHeap::heap();
5625 WorkGang* workers = gch->workers();
5626 assert(workers != NULL, "Need parallel worker threads.");
5627 int n_workers = workers->total_workers();
5628 CMSRefProcTaskProxy rp_task(task, &_collector,
5629 _collector.ref_processor()->span(),
5630 _collector.markBitMap(),
5631 n_workers, _collector.task_queues());
5632 workers->run_task(&rp_task);
5633 }
5635 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5636 {
5638 GenCollectedHeap* gch = GenCollectedHeap::heap();
5639 WorkGang* workers = gch->workers();
5640 assert(workers != NULL, "Need parallel worker threads.");
5641 CMSRefEnqueueTaskProxy enq_task(task);
5642 workers->run_task(&enq_task);
5643 }
5645 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5647 ResourceMark rm;
5648 HandleMark hm;
5649 ReferencePolicy* soft_ref_policy;
5651 assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete");
5652 // Process weak references.
5653 if (clear_all_soft_refs) {
5654 soft_ref_policy = new AlwaysClearPolicy();
5655 } else {
5656 #ifdef COMPILER2
5657 soft_ref_policy = new LRUMaxHeapPolicy();
5658 #else
5659 soft_ref_policy = new LRUCurrentHeapPolicy();
5660 #endif // COMPILER2
5661 }
5662 verify_work_stacks_empty();
5664 ReferenceProcessor* rp = ref_processor();
5665 assert(rp->span().equals(_span), "Spans should be equal");
5666 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5667 &_markStack);
5668 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5669 _span, &_markBitMap, &_markStack,
5670 &cmsKeepAliveClosure);
5671 {
5672 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5673 if (rp->processing_is_mt()) {
5674 CMSRefProcTaskExecutor task_executor(*this);
5675 rp->process_discovered_references(soft_ref_policy,
5676 &_is_alive_closure,
5677 &cmsKeepAliveClosure,
5678 &cmsDrainMarkingStackClosure,
5679 &task_executor);
5680 } else {
5681 rp->process_discovered_references(soft_ref_policy,
5682 &_is_alive_closure,
5683 &cmsKeepAliveClosure,
5684 &cmsDrainMarkingStackClosure,
5685 NULL);
5686 }
5687 verify_work_stacks_empty();
5688 }
5690 if (should_unload_classes()) {
5691 {
5692 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5694 // Follow SystemDictionary roots and unload classes
5695 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5697 // Follow CodeCache roots and unload any methods marked for unloading
5698 CodeCache::do_unloading(&_is_alive_closure,
5699 &cmsKeepAliveClosure,
5700 purged_class);
5702 cmsDrainMarkingStackClosure.do_void();
5703 verify_work_stacks_empty();
5705 // Update subklass/sibling/implementor links in KlassKlass descendants
5706 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5707 oop k;
5708 while ((k = _revisitStack.pop()) != NULL) {
5709 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5710 &_is_alive_closure,
5711 &cmsKeepAliveClosure);
5712 }
5713 assert(!ClassUnloading ||
5714 (_markStack.isEmpty() && overflow_list_is_empty()),
5715 "Should not have found new reachable objects");
5716 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5717 cmsDrainMarkingStackClosure.do_void();
5718 verify_work_stacks_empty();
5719 }
5721 {
5722 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5723 // Now clean up stale oops in SymbolTable and StringTable
5724 SymbolTable::unlink(&_is_alive_closure);
5725 StringTable::unlink(&_is_alive_closure);
5726 }
5727 }
5729 verify_work_stacks_empty();
5730 // Restore any preserved marks as a result of mark stack or
5731 // work queue overflow
5732 restore_preserved_marks_if_any(); // done single-threaded for now
5734 rp->set_enqueuing_is_done(true);
5735 if (rp->processing_is_mt()) {
5736 CMSRefProcTaskExecutor task_executor(*this);
5737 rp->enqueue_discovered_references(&task_executor);
5738 } else {
5739 rp->enqueue_discovered_references(NULL);
5740 }
5741 rp->verify_no_references_recorded();
5742 assert(!rp->discovery_enabled(), "should have been disabled");
5744 // JVMTI object tagging is based on JNI weak refs. If any of these
5745 // refs were cleared then JVMTI needs to update its maps and
5746 // maybe post ObjectFrees to agents.
5747 JvmtiExport::cms_ref_processing_epilogue();
5748 }
5750 #ifndef PRODUCT
5751 void CMSCollector::check_correct_thread_executing() {
5752 Thread* t = Thread::current();
5753 // Only the VM thread or the CMS thread should be here.
5754 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5755 "Unexpected thread type");
5756 // If this is the vm thread, the foreground process
5757 // should not be waiting. Note that _foregroundGCIsActive is
5758 // true while the foreground collector is waiting.
5759 if (_foregroundGCShouldWait) {
5760 // We cannot be the VM thread
5761 assert(t->is_ConcurrentGC_thread(),
5762 "Should be CMS thread");
5763 } else {
5764 // We can be the CMS thread only if we are in a stop-world
5765 // phase of CMS collection.
5766 if (t->is_ConcurrentGC_thread()) {
5767 assert(_collectorState == InitialMarking ||
5768 _collectorState == FinalMarking,
5769 "Should be a stop-world phase");
5770 // The CMS thread should be holding the CMS_token.
5771 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5772 "Potential interference with concurrently "
5773 "executing VM thread");
5774 }
5775 }
5776 }
5777 #endif
5779 void CMSCollector::sweep(bool asynch) {
5780 assert(_collectorState == Sweeping, "just checking");
5781 check_correct_thread_executing();
5782 verify_work_stacks_empty();
5783 verify_overflow_empty();
5784 incrementSweepCount();
5785 _sweep_timer.stop();
5786 _sweep_estimate.sample(_sweep_timer.seconds());
5787 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5789 // PermGen verification support: If perm gen sweeping is disabled in
5790 // this cycle, we preserve the perm gen object "deadness" information
5791 // in the perm_gen_verify_bit_map. In order to do that we traverse
5792 // all blocks in perm gen and mark all dead objects.
5793 if (verifying() && !should_unload_classes()) {
5794 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
5795 "Should have already been allocated");
5796 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
5797 markBitMap(), perm_gen_verify_bit_map());
5798 if (asynch) {
5799 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5800 bitMapLock());
5801 _permGen->cmsSpace()->blk_iterate(&mdo);
5802 } else {
5803 // In the case of synchronous sweep, we already have
5804 // the requisite locks/tokens.
5805 _permGen->cmsSpace()->blk_iterate(&mdo);
5806 }
5807 }
5809 if (asynch) {
5810 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5811 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5812 // First sweep the old gen then the perm gen
5813 {
5814 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5815 bitMapLock());
5816 sweepWork(_cmsGen, asynch);
5817 }
5819 // Now repeat for perm gen
5820 if (should_unload_classes()) {
5821 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5822 bitMapLock());
5823 sweepWork(_permGen, asynch);
5824 }
5826 // Update Universe::_heap_*_at_gc figures.
5827 // We need all the free list locks to make the abstract state
5828 // transition from Sweeping to Resetting. See detailed note
5829 // further below.
5830 {
5831 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5832 _permGen->freelistLock());
5833 // Update heap occupancy information which is used as
5834 // input to soft ref clearing policy at the next gc.
5835 Universe::update_heap_info_at_gc();
5836 _collectorState = Resizing;
5837 }
5838 } else {
5839 // already have needed locks
5840 sweepWork(_cmsGen, asynch);
5842 if (should_unload_classes()) {
5843 sweepWork(_permGen, asynch);
5844 }
5845 // Update heap occupancy information which is used as
5846 // input to soft ref clearing policy at the next gc.
5847 Universe::update_heap_info_at_gc();
5848 _collectorState = Resizing;
5849 }
5850 verify_work_stacks_empty();
5851 verify_overflow_empty();
5853 _sweep_timer.reset();
5854 _sweep_timer.start();
5856 update_time_of_last_gc(os::javaTimeMillis());
5858 // NOTE on abstract state transitions:
5859 // Mutators allocate-live and/or mark the mod-union table dirty
5860 // based on the state of the collection. The former is done in
5861 // the interval [Marking, Sweeping] and the latter in the interval
5862 // [Marking, Sweeping). Thus the transitions into the Marking state
5863 // and out of the Sweeping state must be synchronously visible
5864 // globally to the mutators.
5865 // The transition into the Marking state happens with the world
5866 // stopped so the mutators will globally see it. Sweeping is
5867 // done asynchronously by the background collector so the transition
5868 // from the Sweeping state to the Resizing state must be done
5869 // under the freelistLock (as is the check for whether to
5870 // allocate-live and whether to dirty the mod-union table).
5871 assert(_collectorState == Resizing, "Change of collector state to"
5872 " Resizing must be done under the freelistLocks (plural)");
5874 // Now that sweeping has been completed, if the GCH's
5875 // incremental_collection_will_fail flag is set, clear it,
5876 // thus inviting a younger gen collection to promote into
5877 // this generation. If such a promotion may still fail,
5878 // the flag will be set again when a young collection is
5879 // attempted.
5880 // I think the incremental_collection_will_fail flag's use
5881 // is specific to a 2 generation collection policy, so i'll
5882 // assert that that's the configuration we are operating within.
5883 // The use of the flag can and should be generalized appropriately
5884 // in the future to deal with a general n-generation system.
5886 GenCollectedHeap* gch = GenCollectedHeap::heap();
5887 assert(gch->collector_policy()->is_two_generation_policy(),
5888 "Resetting of incremental_collection_will_fail flag"
5889 " may be incorrect otherwise");
5890 gch->clear_incremental_collection_will_fail();
5891 gch->update_full_collections_completed(_collection_count_start);
5892 }
5894 // FIX ME!!! Looks like this belongs in CFLSpace, with
5895 // CMSGen merely delegating to it.
5896 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5897 double nearLargestPercent = 0.999;
5898 HeapWord* minAddr = _cmsSpace->bottom();
5899 HeapWord* largestAddr =
5900 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
5901 if (largestAddr == 0) {
5902 // The dictionary appears to be empty. In this case
5903 // try to coalesce at the end of the heap.
5904 largestAddr = _cmsSpace->end();
5905 }
5906 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5907 size_t nearLargestOffset =
5908 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5909 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5910 }
5912 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5913 return addr >= _cmsSpace->nearLargestChunk();
5914 }
5916 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5917 return _cmsSpace->find_chunk_at_end();
5918 }
5920 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5921 bool full) {
5922 // The next lower level has been collected. Gather any statistics
5923 // that are of interest at this point.
5924 if (!full && (current_level + 1) == level()) {
5925 // Gather statistics on the young generation collection.
5926 collector()->stats().record_gc0_end(used());
5927 }
5928 }
5930 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
5931 GenCollectedHeap* gch = GenCollectedHeap::heap();
5932 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
5933 "Wrong type of heap");
5934 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
5935 gch->gen_policy()->size_policy();
5936 assert(sp->is_gc_cms_adaptive_size_policy(),
5937 "Wrong type of size policy");
5938 return sp;
5939 }
5941 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
5942 if (PrintGCDetails && Verbose) {
5943 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
5944 }
5945 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
5946 _debug_collection_type =
5947 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
5948 if (PrintGCDetails && Verbose) {
5949 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
5950 }
5951 }
5953 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
5954 bool asynch) {
5955 // We iterate over the space(s) underlying this generation,
5956 // checking the mark bit map to see if the bits corresponding
5957 // to specific blocks are marked or not. Blocks that are
5958 // marked are live and are not swept up. All remaining blocks
5959 // are swept up, with coalescing on-the-fly as we sweep up
5960 // contiguous free and/or garbage blocks:
5961 // We need to ensure that the sweeper synchronizes with allocators
5962 // and stop-the-world collectors. In particular, the following
5963 // locks are used:
5964 // . CMS token: if this is held, a stop the world collection cannot occur
5965 // . freelistLock: if this is held no allocation can occur from this
5966 // generation by another thread
5967 // . bitMapLock: if this is held, no other thread can access or update
5968 //
5970 // Note that we need to hold the freelistLock if we use
5971 // block iterate below; else the iterator might go awry if
5972 // a mutator (or promotion) causes block contents to change
5973 // (for instance if the allocator divvies up a block).
5974 // If we hold the free list lock, for all practical purposes
5975 // young generation GC's can't occur (they'll usually need to
5976 // promote), so we might as well prevent all young generation
5977 // GC's while we do a sweeping step. For the same reason, we might
5978 // as well take the bit map lock for the entire duration
5980 // check that we hold the requisite locks
5981 assert(have_cms_token(), "Should hold cms token");
5982 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
5983 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
5984 "Should possess CMS token to sweep");
5985 assert_lock_strong(gen->freelistLock());
5986 assert_lock_strong(bitMapLock());
5988 assert(!_sweep_timer.is_active(), "Was switched off in an outer context");
5989 gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
5990 _sweep_estimate.padded_average());
5991 gen->setNearLargestChunk();
5993 {
5994 SweepClosure sweepClosure(this, gen, &_markBitMap,
5995 CMSYield && asynch);
5996 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5997 // We need to free-up/coalesce garbage/blocks from a
5998 // co-terminal free run. This is done in the SweepClosure
5999 // destructor; so, do not remove this scope, else the
6000 // end-of-sweep-census below will be off by a little bit.
6001 }
6002 gen->cmsSpace()->sweep_completed();
6003 gen->cmsSpace()->endSweepFLCensus(sweepCount());
6004 if (should_unload_classes()) { // unloaded classes this cycle,
6005 _concurrent_cycles_since_last_unload = 0; // ... reset count
6006 } else { // did not unload classes,
6007 _concurrent_cycles_since_last_unload++; // ... increment count
6008 }
6009 }
6011 // Reset CMS data structures (for now just the marking bit map)
6012 // preparatory for the next cycle.
6013 void CMSCollector::reset(bool asynch) {
6014 GenCollectedHeap* gch = GenCollectedHeap::heap();
6015 CMSAdaptiveSizePolicy* sp = size_policy();
6016 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6017 if (asynch) {
6018 CMSTokenSyncWithLocks ts(true, bitMapLock());
6020 // If the state is not "Resetting", the foreground thread
6021 // has done a collection and the resetting.
6022 if (_collectorState != Resetting) {
6023 assert(_collectorState == Idling, "The state should only change"
6024 " because the foreground collector has finished the collection");
6025 return;
6026 }
6028 // Clear the mark bitmap (no grey objects to start with)
6029 // for the next cycle.
6030 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6031 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6033 HeapWord* curAddr = _markBitMap.startWord();
6034 while (curAddr < _markBitMap.endWord()) {
6035 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6036 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6037 _markBitMap.clear_large_range(chunk);
6038 if (ConcurrentMarkSweepThread::should_yield() &&
6039 !foregroundGCIsActive() &&
6040 CMSYield) {
6041 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6042 "CMS thread should hold CMS token");
6043 assert_lock_strong(bitMapLock());
6044 bitMapLock()->unlock();
6045 ConcurrentMarkSweepThread::desynchronize(true);
6046 ConcurrentMarkSweepThread::acknowledge_yield_request();
6047 stopTimer();
6048 if (PrintCMSStatistics != 0) {
6049 incrementYields();
6050 }
6051 icms_wait();
6053 // See the comment in coordinator_yield()
6054 for (unsigned i = 0; i < CMSYieldSleepCount &&
6055 ConcurrentMarkSweepThread::should_yield() &&
6056 !CMSCollector::foregroundGCIsActive(); ++i) {
6057 os::sleep(Thread::current(), 1, false);
6058 ConcurrentMarkSweepThread::acknowledge_yield_request();
6059 }
6061 ConcurrentMarkSweepThread::synchronize(true);
6062 bitMapLock()->lock_without_safepoint_check();
6063 startTimer();
6064 }
6065 curAddr = chunk.end();
6066 }
6067 _collectorState = Idling;
6068 } else {
6069 // already have the lock
6070 assert(_collectorState == Resetting, "just checking");
6071 assert_lock_strong(bitMapLock());
6072 _markBitMap.clear_all();
6073 _collectorState = Idling;
6074 }
6076 // Stop incremental mode after a cycle completes, so that any future cycles
6077 // are triggered by allocation.
6078 stop_icms();
6080 NOT_PRODUCT(
6081 if (RotateCMSCollectionTypes) {
6082 _cmsGen->rotate_debug_collection_type();
6083 }
6084 )
6085 }
6087 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6088 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6089 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6090 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6091 TraceCollectorStats tcs(counters());
6093 switch (op) {
6094 case CMS_op_checkpointRootsInitial: {
6095 checkpointRootsInitial(true); // asynch
6096 if (PrintGC) {
6097 _cmsGen->printOccupancy("initial-mark");
6098 }
6099 break;
6100 }
6101 case CMS_op_checkpointRootsFinal: {
6102 checkpointRootsFinal(true, // asynch
6103 false, // !clear_all_soft_refs
6104 false); // !init_mark_was_synchronous
6105 if (PrintGC) {
6106 _cmsGen->printOccupancy("remark");
6107 }
6108 break;
6109 }
6110 default:
6111 fatal("No such CMS_op");
6112 }
6113 }
6115 #ifndef PRODUCT
6116 size_t const CMSCollector::skip_header_HeapWords() {
6117 return FreeChunk::header_size();
6118 }
6120 // Try and collect here conditions that should hold when
6121 // CMS thread is exiting. The idea is that the foreground GC
6122 // thread should not be blocked if it wants to terminate
6123 // the CMS thread and yet continue to run the VM for a while
6124 // after that.
6125 void CMSCollector::verify_ok_to_terminate() const {
6126 assert(Thread::current()->is_ConcurrentGC_thread(),
6127 "should be called by CMS thread");
6128 assert(!_foregroundGCShouldWait, "should be false");
6129 // We could check here that all the various low-level locks
6130 // are not held by the CMS thread, but that is overkill; see
6131 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6132 // is checked.
6133 }
6134 #endif
6136 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6137 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6138 "missing Printezis mark?");
6139 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6140 size_t size = pointer_delta(nextOneAddr + 1, addr);
6141 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6142 "alignment problem");
6143 assert(size >= 3, "Necessary for Printezis marks to work");
6144 return size;
6145 }
6147 // A variant of the above (block_size_using_printezis_bits()) except
6148 // that we return 0 if the P-bits are not yet set.
6149 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6150 if (_markBitMap.isMarked(addr)) {
6151 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
6152 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6153 size_t size = pointer_delta(nextOneAddr + 1, addr);
6154 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6155 "alignment problem");
6156 assert(size >= 3, "Necessary for Printezis marks to work");
6157 return size;
6158 } else {
6159 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
6160 return 0;
6161 }
6162 }
6164 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6165 size_t sz = 0;
6166 oop p = (oop)addr;
6167 if (p->klass_or_null() != NULL && p->is_parsable()) {
6168 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6169 } else {
6170 sz = block_size_using_printezis_bits(addr);
6171 }
6172 assert(sz > 0, "size must be nonzero");
6173 HeapWord* next_block = addr + sz;
6174 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6175 CardTableModRefBS::card_size);
6176 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6177 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6178 "must be different cards");
6179 return next_card;
6180 }
6183 // CMS Bit Map Wrapper /////////////////////////////////////////
6185 // Construct a CMS bit map infrastructure, but don't create the
6186 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6187 // further below.
6188 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6189 _bm(),
6190 _shifter(shifter),
6191 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6192 {
6193 _bmStartWord = 0;
6194 _bmWordSize = 0;
6195 }
6197 bool CMSBitMap::allocate(MemRegion mr) {
6198 _bmStartWord = mr.start();
6199 _bmWordSize = mr.word_size();
6200 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6201 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6202 if (!brs.is_reserved()) {
6203 warning("CMS bit map allocation failure");
6204 return false;
6205 }
6206 // For now we'll just commit all of the bit map up fromt.
6207 // Later on we'll try to be more parsimonious with swap.
6208 if (!_virtual_space.initialize(brs, brs.size())) {
6209 warning("CMS bit map backing store failure");
6210 return false;
6211 }
6212 assert(_virtual_space.committed_size() == brs.size(),
6213 "didn't reserve backing store for all of CMS bit map?");
6214 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6215 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6216 _bmWordSize, "inconsistency in bit map sizing");
6217 _bm.set_size(_bmWordSize >> _shifter);
6219 // bm.clear(); // can we rely on getting zero'd memory? verify below
6220 assert(isAllClear(),
6221 "Expected zero'd memory from ReservedSpace constructor");
6222 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6223 "consistency check");
6224 return true;
6225 }
6227 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6228 HeapWord *next_addr, *end_addr, *last_addr;
6229 assert_locked();
6230 assert(covers(mr), "out-of-range error");
6231 // XXX assert that start and end are appropriately aligned
6232 for (next_addr = mr.start(), end_addr = mr.end();
6233 next_addr < end_addr; next_addr = last_addr) {
6234 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6235 last_addr = dirty_region.end();
6236 if (!dirty_region.is_empty()) {
6237 cl->do_MemRegion(dirty_region);
6238 } else {
6239 assert(last_addr == end_addr, "program logic");
6240 return;
6241 }
6242 }
6243 }
6245 #ifndef PRODUCT
6246 void CMSBitMap::assert_locked() const {
6247 CMSLockVerifier::assert_locked(lock());
6248 }
6250 bool CMSBitMap::covers(MemRegion mr) const {
6251 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6252 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6253 "size inconsistency");
6254 return (mr.start() >= _bmStartWord) &&
6255 (mr.end() <= endWord());
6256 }
6258 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6259 return (start >= _bmStartWord && (start + size) <= endWord());
6260 }
6262 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6263 // verify that there are no 1 bits in the interval [left, right)
6264 FalseBitMapClosure falseBitMapClosure;
6265 iterate(&falseBitMapClosure, left, right);
6266 }
6268 void CMSBitMap::region_invariant(MemRegion mr)
6269 {
6270 assert_locked();
6271 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6272 assert(!mr.is_empty(), "unexpected empty region");
6273 assert(covers(mr), "mr should be covered by bit map");
6274 // convert address range into offset range
6275 size_t start_ofs = heapWordToOffset(mr.start());
6276 // Make sure that end() is appropriately aligned
6277 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6278 (1 << (_shifter+LogHeapWordSize))),
6279 "Misaligned mr.end()");
6280 size_t end_ofs = heapWordToOffset(mr.end());
6281 assert(end_ofs > start_ofs, "Should mark at least one bit");
6282 }
6284 #endif
6286 bool CMSMarkStack::allocate(size_t size) {
6287 // allocate a stack of the requisite depth
6288 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6289 size * sizeof(oop)));
6290 if (!rs.is_reserved()) {
6291 warning("CMSMarkStack allocation failure");
6292 return false;
6293 }
6294 if (!_virtual_space.initialize(rs, rs.size())) {
6295 warning("CMSMarkStack backing store failure");
6296 return false;
6297 }
6298 assert(_virtual_space.committed_size() == rs.size(),
6299 "didn't reserve backing store for all of CMS stack?");
6300 _base = (oop*)(_virtual_space.low());
6301 _index = 0;
6302 _capacity = size;
6303 NOT_PRODUCT(_max_depth = 0);
6304 return true;
6305 }
6307 // XXX FIX ME !!! In the MT case we come in here holding a
6308 // leaf lock. For printing we need to take a further lock
6309 // which has lower rank. We need to recallibrate the two
6310 // lock-ranks involved in order to be able to rpint the
6311 // messages below. (Or defer the printing to the caller.
6312 // For now we take the expedient path of just disabling the
6313 // messages for the problematic case.)
6314 void CMSMarkStack::expand() {
6315 assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted");
6316 if (_capacity == CMSMarkStackSizeMax) {
6317 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6318 // We print a warning message only once per CMS cycle.
6319 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6320 }
6321 return;
6322 }
6323 // Double capacity if possible
6324 size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax);
6325 // Do not give up existing stack until we have managed to
6326 // get the double capacity that we desired.
6327 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6328 new_capacity * sizeof(oop)));
6329 if (rs.is_reserved()) {
6330 // Release the backing store associated with old stack
6331 _virtual_space.release();
6332 // Reinitialize virtual space for new stack
6333 if (!_virtual_space.initialize(rs, rs.size())) {
6334 fatal("Not enough swap for expanded marking stack");
6335 }
6336 _base = (oop*)(_virtual_space.low());
6337 _index = 0;
6338 _capacity = new_capacity;
6339 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6340 // Failed to double capacity, continue;
6341 // we print a detail message only once per CMS cycle.
6342 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6343 SIZE_FORMAT"K",
6344 _capacity / K, new_capacity / K);
6345 }
6346 }
6349 // Closures
6350 // XXX: there seems to be a lot of code duplication here;
6351 // should refactor and consolidate common code.
6353 // This closure is used to mark refs into the CMS generation in
6354 // the CMS bit map. Called at the first checkpoint. This closure
6355 // assumes that we do not need to re-mark dirty cards; if the CMS
6356 // generation on which this is used is not an oldest (modulo perm gen)
6357 // generation then this will lose younger_gen cards!
6359 MarkRefsIntoClosure::MarkRefsIntoClosure(
6360 MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods):
6361 _span(span),
6362 _bitMap(bitMap),
6363 _should_do_nmethods(should_do_nmethods)
6364 {
6365 assert(_ref_processor == NULL, "deliberately left NULL");
6366 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6367 }
6369 void MarkRefsIntoClosure::do_oop(oop obj) {
6370 // if p points into _span, then mark corresponding bit in _markBitMap
6371 assert(obj->is_oop(), "expected an oop");
6372 HeapWord* addr = (HeapWord*)obj;
6373 if (_span.contains(addr)) {
6374 // this should be made more efficient
6375 _bitMap->mark(addr);
6376 }
6377 }
6379 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6380 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6382 // A variant of the above, used for CMS marking verification.
6383 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6384 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6385 bool should_do_nmethods):
6386 _span(span),
6387 _verification_bm(verification_bm),
6388 _cms_bm(cms_bm),
6389 _should_do_nmethods(should_do_nmethods) {
6390 assert(_ref_processor == NULL, "deliberately left NULL");
6391 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6392 }
6394 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6395 // if p points into _span, then mark corresponding bit in _markBitMap
6396 assert(obj->is_oop(), "expected an oop");
6397 HeapWord* addr = (HeapWord*)obj;
6398 if (_span.contains(addr)) {
6399 _verification_bm->mark(addr);
6400 if (!_cms_bm->isMarked(addr)) {
6401 oop(addr)->print();
6402 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6403 fatal("... aborting");
6404 }
6405 }
6406 }
6408 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6409 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6411 //////////////////////////////////////////////////
6412 // MarkRefsIntoAndScanClosure
6413 //////////////////////////////////////////////////
6415 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6416 ReferenceProcessor* rp,
6417 CMSBitMap* bit_map,
6418 CMSBitMap* mod_union_table,
6419 CMSMarkStack* mark_stack,
6420 CMSMarkStack* revisit_stack,
6421 CMSCollector* collector,
6422 bool should_yield,
6423 bool concurrent_precleaning):
6424 _collector(collector),
6425 _span(span),
6426 _bit_map(bit_map),
6427 _mark_stack(mark_stack),
6428 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6429 mark_stack, revisit_stack, concurrent_precleaning),
6430 _yield(should_yield),
6431 _concurrent_precleaning(concurrent_precleaning),
6432 _freelistLock(NULL)
6433 {
6434 _ref_processor = rp;
6435 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6436 }
6438 // This closure is used to mark refs into the CMS generation at the
6439 // second (final) checkpoint, and to scan and transitively follow
6440 // the unmarked oops. It is also used during the concurrent precleaning
6441 // phase while scanning objects on dirty cards in the CMS generation.
6442 // The marks are made in the marking bit map and the marking stack is
6443 // used for keeping the (newly) grey objects during the scan.
6444 // The parallel version (Par_...) appears further below.
6445 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6446 if (obj != NULL) {
6447 assert(obj->is_oop(), "expected an oop");
6448 HeapWord* addr = (HeapWord*)obj;
6449 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6450 assert(_collector->overflow_list_is_empty(),
6451 "overflow list should be empty");
6452 if (_span.contains(addr) &&
6453 !_bit_map->isMarked(addr)) {
6454 // mark bit map (object is now grey)
6455 _bit_map->mark(addr);
6456 // push on marking stack (stack should be empty), and drain the
6457 // stack by applying this closure to the oops in the oops popped
6458 // from the stack (i.e. blacken the grey objects)
6459 bool res = _mark_stack->push(obj);
6460 assert(res, "Should have space to push on empty stack");
6461 do {
6462 oop new_oop = _mark_stack->pop();
6463 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6464 assert(new_oop->is_parsable(), "Found unparsable oop");
6465 assert(_bit_map->isMarked((HeapWord*)new_oop),
6466 "only grey objects on this stack");
6467 // iterate over the oops in this oop, marking and pushing
6468 // the ones in CMS heap (i.e. in _span).
6469 new_oop->oop_iterate(&_pushAndMarkClosure);
6470 // check if it's time to yield
6471 do_yield_check();
6472 } while (!_mark_stack->isEmpty() ||
6473 (!_concurrent_precleaning && take_from_overflow_list()));
6474 // if marking stack is empty, and we are not doing this
6475 // during precleaning, then check the overflow list
6476 }
6477 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6478 assert(_collector->overflow_list_is_empty(),
6479 "overflow list was drained above");
6480 // We could restore evacuated mark words, if any, used for
6481 // overflow list links here because the overflow list is
6482 // provably empty here. That would reduce the maximum
6483 // size requirements for preserved_{oop,mark}_stack.
6484 // But we'll just postpone it until we are all done
6485 // so we can just stream through.
6486 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6487 _collector->restore_preserved_marks_if_any();
6488 assert(_collector->no_preserved_marks(), "No preserved marks");
6489 }
6490 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6491 "All preserved marks should have been restored above");
6492 }
6493 }
6495 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6496 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6498 void MarkRefsIntoAndScanClosure::do_yield_work() {
6499 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6500 "CMS thread should hold CMS token");
6501 assert_lock_strong(_freelistLock);
6502 assert_lock_strong(_bit_map->lock());
6503 // relinquish the free_list_lock and bitMaplock()
6504 _bit_map->lock()->unlock();
6505 _freelistLock->unlock();
6506 ConcurrentMarkSweepThread::desynchronize(true);
6507 ConcurrentMarkSweepThread::acknowledge_yield_request();
6508 _collector->stopTimer();
6509 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6510 if (PrintCMSStatistics != 0) {
6511 _collector->incrementYields();
6512 }
6513 _collector->icms_wait();
6515 // See the comment in coordinator_yield()
6516 for (unsigned i = 0;
6517 i < CMSYieldSleepCount &&
6518 ConcurrentMarkSweepThread::should_yield() &&
6519 !CMSCollector::foregroundGCIsActive();
6520 ++i) {
6521 os::sleep(Thread::current(), 1, false);
6522 ConcurrentMarkSweepThread::acknowledge_yield_request();
6523 }
6525 ConcurrentMarkSweepThread::synchronize(true);
6526 _freelistLock->lock_without_safepoint_check();
6527 _bit_map->lock()->lock_without_safepoint_check();
6528 _collector->startTimer();
6529 }
6531 ///////////////////////////////////////////////////////////
6532 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6533 // MarkRefsIntoAndScanClosure
6534 ///////////////////////////////////////////////////////////
6535 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6536 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6537 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6538 _span(span),
6539 _bit_map(bit_map),
6540 _work_queue(work_queue),
6541 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6542 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6543 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6544 revisit_stack)
6545 {
6546 _ref_processor = rp;
6547 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6548 }
6550 // This closure is used to mark refs into the CMS generation at the
6551 // second (final) checkpoint, and to scan and transitively follow
6552 // the unmarked oops. The marks are made in the marking bit map and
6553 // the work_queue is used for keeping the (newly) grey objects during
6554 // the scan phase whence they are also available for stealing by parallel
6555 // threads. Since the marking bit map is shared, updates are
6556 // synchronized (via CAS).
6557 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6558 if (obj != NULL) {
6559 // Ignore mark word because this could be an already marked oop
6560 // that may be chained at the end of the overflow list.
6561 assert(obj->is_oop(), "expected an oop");
6562 HeapWord* addr = (HeapWord*)obj;
6563 if (_span.contains(addr) &&
6564 !_bit_map->isMarked(addr)) {
6565 // mark bit map (object will become grey):
6566 // It is possible for several threads to be
6567 // trying to "claim" this object concurrently;
6568 // the unique thread that succeeds in marking the
6569 // object first will do the subsequent push on
6570 // to the work queue (or overflow list).
6571 if (_bit_map->par_mark(addr)) {
6572 // push on work_queue (which may not be empty), and trim the
6573 // queue to an appropriate length by applying this closure to
6574 // the oops in the oops popped from the stack (i.e. blacken the
6575 // grey objects)
6576 bool res = _work_queue->push(obj);
6577 assert(res, "Low water mark should be less than capacity?");
6578 trim_queue(_low_water_mark);
6579 } // Else, another thread claimed the object
6580 }
6581 }
6582 }
6584 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6585 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6587 // This closure is used to rescan the marked objects on the dirty cards
6588 // in the mod union table and the card table proper.
6589 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6590 oop p, MemRegion mr) {
6592 size_t size = 0;
6593 HeapWord* addr = (HeapWord*)p;
6594 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6595 assert(_span.contains(addr), "we are scanning the CMS generation");
6596 // check if it's time to yield
6597 if (do_yield_check()) {
6598 // We yielded for some foreground stop-world work,
6599 // and we have been asked to abort this ongoing preclean cycle.
6600 return 0;
6601 }
6602 if (_bitMap->isMarked(addr)) {
6603 // it's marked; is it potentially uninitialized?
6604 if (p->klass_or_null() != NULL) {
6605 if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) {
6606 // Signal precleaning to redirty the card since
6607 // the klass pointer is already installed.
6608 assert(size == 0, "Initial value");
6609 } else {
6610 assert(p->is_parsable(), "must be parsable.");
6611 // an initialized object; ignore mark word in verification below
6612 // since we are running concurrent with mutators
6613 assert(p->is_oop(true), "should be an oop");
6614 if (p->is_objArray()) {
6615 // objArrays are precisely marked; restrict scanning
6616 // to dirty cards only.
6617 size = CompactibleFreeListSpace::adjustObjectSize(
6618 p->oop_iterate(_scanningClosure, mr));
6619 } else {
6620 // A non-array may have been imprecisely marked; we need
6621 // to scan object in its entirety.
6622 size = CompactibleFreeListSpace::adjustObjectSize(
6623 p->oop_iterate(_scanningClosure));
6624 }
6625 #ifdef DEBUG
6626 size_t direct_size =
6627 CompactibleFreeListSpace::adjustObjectSize(p->size());
6628 assert(size == direct_size, "Inconsistency in size");
6629 assert(size >= 3, "Necessary for Printezis marks to work");
6630 if (!_bitMap->isMarked(addr+1)) {
6631 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6632 } else {
6633 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6634 assert(_bitMap->isMarked(addr+size-1),
6635 "inconsistent Printezis mark");
6636 }
6637 #endif // DEBUG
6638 }
6639 } else {
6640 // an unitialized object
6641 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6642 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6643 size = pointer_delta(nextOneAddr + 1, addr);
6644 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6645 "alignment problem");
6646 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6647 // will dirty the card when the klass pointer is installed in the
6648 // object (signalling the completion of initialization).
6649 }
6650 } else {
6651 // Either a not yet marked object or an uninitialized object
6652 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6653 // An uninitialized object, skip to the next card, since
6654 // we may not be able to read its P-bits yet.
6655 assert(size == 0, "Initial value");
6656 } else {
6657 // An object not (yet) reached by marking: we merely need to
6658 // compute its size so as to go look at the next block.
6659 assert(p->is_oop(true), "should be an oop");
6660 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6661 }
6662 }
6663 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6664 return size;
6665 }
6667 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6668 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6669 "CMS thread should hold CMS token");
6670 assert_lock_strong(_freelistLock);
6671 assert_lock_strong(_bitMap->lock());
6672 // relinquish the free_list_lock and bitMaplock()
6673 _bitMap->lock()->unlock();
6674 _freelistLock->unlock();
6675 ConcurrentMarkSweepThread::desynchronize(true);
6676 ConcurrentMarkSweepThread::acknowledge_yield_request();
6677 _collector->stopTimer();
6678 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6679 if (PrintCMSStatistics != 0) {
6680 _collector->incrementYields();
6681 }
6682 _collector->icms_wait();
6684 // See the comment in coordinator_yield()
6685 for (unsigned i = 0; i < CMSYieldSleepCount &&
6686 ConcurrentMarkSweepThread::should_yield() &&
6687 !CMSCollector::foregroundGCIsActive(); ++i) {
6688 os::sleep(Thread::current(), 1, false);
6689 ConcurrentMarkSweepThread::acknowledge_yield_request();
6690 }
6692 ConcurrentMarkSweepThread::synchronize(true);
6693 _freelistLock->lock_without_safepoint_check();
6694 _bitMap->lock()->lock_without_safepoint_check();
6695 _collector->startTimer();
6696 }
6699 //////////////////////////////////////////////////////////////////
6700 // SurvivorSpacePrecleanClosure
6701 //////////////////////////////////////////////////////////////////
6702 // This (single-threaded) closure is used to preclean the oops in
6703 // the survivor spaces.
6704 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6706 HeapWord* addr = (HeapWord*)p;
6707 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6708 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6709 assert(p->klass_or_null() != NULL, "object should be initializd");
6710 assert(p->is_parsable(), "must be parsable.");
6711 // an initialized object; ignore mark word in verification below
6712 // since we are running concurrent with mutators
6713 assert(p->is_oop(true), "should be an oop");
6714 // Note that we do not yield while we iterate over
6715 // the interior oops of p, pushing the relevant ones
6716 // on our marking stack.
6717 size_t size = p->oop_iterate(_scanning_closure);
6718 do_yield_check();
6719 // Observe that below, we do not abandon the preclean
6720 // phase as soon as we should; rather we empty the
6721 // marking stack before returning. This is to satisfy
6722 // some existing assertions. In general, it may be a
6723 // good idea to abort immediately and complete the marking
6724 // from the grey objects at a later time.
6725 while (!_mark_stack->isEmpty()) {
6726 oop new_oop = _mark_stack->pop();
6727 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6728 assert(new_oop->is_parsable(), "Found unparsable oop");
6729 assert(_bit_map->isMarked((HeapWord*)new_oop),
6730 "only grey objects on this stack");
6731 // iterate over the oops in this oop, marking and pushing
6732 // the ones in CMS heap (i.e. in _span).
6733 new_oop->oop_iterate(_scanning_closure);
6734 // check if it's time to yield
6735 do_yield_check();
6736 }
6737 unsigned int after_count =
6738 GenCollectedHeap::heap()->total_collections();
6739 bool abort = (_before_count != after_count) ||
6740 _collector->should_abort_preclean();
6741 return abort ? 0 : size;
6742 }
6744 void SurvivorSpacePrecleanClosure::do_yield_work() {
6745 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6746 "CMS thread should hold CMS token");
6747 assert_lock_strong(_bit_map->lock());
6748 // Relinquish the bit map lock
6749 _bit_map->lock()->unlock();
6750 ConcurrentMarkSweepThread::desynchronize(true);
6751 ConcurrentMarkSweepThread::acknowledge_yield_request();
6752 _collector->stopTimer();
6753 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6754 if (PrintCMSStatistics != 0) {
6755 _collector->incrementYields();
6756 }
6757 _collector->icms_wait();
6759 // See the comment in coordinator_yield()
6760 for (unsigned i = 0; i < CMSYieldSleepCount &&
6761 ConcurrentMarkSweepThread::should_yield() &&
6762 !CMSCollector::foregroundGCIsActive(); ++i) {
6763 os::sleep(Thread::current(), 1, false);
6764 ConcurrentMarkSweepThread::acknowledge_yield_request();
6765 }
6767 ConcurrentMarkSweepThread::synchronize(true);
6768 _bit_map->lock()->lock_without_safepoint_check();
6769 _collector->startTimer();
6770 }
6772 // This closure is used to rescan the marked objects on the dirty cards
6773 // in the mod union table and the card table proper. In the parallel
6774 // case, although the bitMap is shared, we do a single read so the
6775 // isMarked() query is "safe".
6776 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6777 // Ignore mark word because we are running concurrent with mutators
6778 assert(p->is_oop_or_null(true), "expected an oop or null");
6779 HeapWord* addr = (HeapWord*)p;
6780 assert(_span.contains(addr), "we are scanning the CMS generation");
6781 bool is_obj_array = false;
6782 #ifdef DEBUG
6783 if (!_parallel) {
6784 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6785 assert(_collector->overflow_list_is_empty(),
6786 "overflow list should be empty");
6788 }
6789 #endif // DEBUG
6790 if (_bit_map->isMarked(addr)) {
6791 // Obj arrays are precisely marked, non-arrays are not;
6792 // so we scan objArrays precisely and non-arrays in their
6793 // entirety.
6794 if (p->is_objArray()) {
6795 is_obj_array = true;
6796 if (_parallel) {
6797 p->oop_iterate(_par_scan_closure, mr);
6798 } else {
6799 p->oop_iterate(_scan_closure, mr);
6800 }
6801 } else {
6802 if (_parallel) {
6803 p->oop_iterate(_par_scan_closure);
6804 } else {
6805 p->oop_iterate(_scan_closure);
6806 }
6807 }
6808 }
6809 #ifdef DEBUG
6810 if (!_parallel) {
6811 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6812 assert(_collector->overflow_list_is_empty(),
6813 "overflow list should be empty");
6815 }
6816 #endif // DEBUG
6817 return is_obj_array;
6818 }
6820 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6821 MemRegion span,
6822 CMSBitMap* bitMap, CMSMarkStack* markStack,
6823 CMSMarkStack* revisitStack,
6824 bool should_yield, bool verifying):
6825 _collector(collector),
6826 _span(span),
6827 _bitMap(bitMap),
6828 _mut(&collector->_modUnionTable),
6829 _markStack(markStack),
6830 _revisitStack(revisitStack),
6831 _yield(should_yield),
6832 _skipBits(0)
6833 {
6834 assert(_markStack->isEmpty(), "stack should be empty");
6835 _finger = _bitMap->startWord();
6836 _threshold = _finger;
6837 assert(_collector->_restart_addr == NULL, "Sanity check");
6838 assert(_span.contains(_finger), "Out of bounds _finger?");
6839 DEBUG_ONLY(_verifying = verifying;)
6840 }
6842 void MarkFromRootsClosure::reset(HeapWord* addr) {
6843 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6844 assert(_span.contains(addr), "Out of bounds _finger?");
6845 _finger = addr;
6846 _threshold = (HeapWord*)round_to(
6847 (intptr_t)_finger, CardTableModRefBS::card_size);
6848 }
6850 // Should revisit to see if this should be restructured for
6851 // greater efficiency.
6852 bool MarkFromRootsClosure::do_bit(size_t offset) {
6853 if (_skipBits > 0) {
6854 _skipBits--;
6855 return true;
6856 }
6857 // convert offset into a HeapWord*
6858 HeapWord* addr = _bitMap->startWord() + offset;
6859 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6860 "address out of range");
6861 assert(_bitMap->isMarked(addr), "tautology");
6862 if (_bitMap->isMarked(addr+1)) {
6863 // this is an allocated but not yet initialized object
6864 assert(_skipBits == 0, "tautology");
6865 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6866 oop p = oop(addr);
6867 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6868 DEBUG_ONLY(if (!_verifying) {)
6869 // We re-dirty the cards on which this object lies and increase
6870 // the _threshold so that we'll come back to scan this object
6871 // during the preclean or remark phase. (CMSCleanOnEnter)
6872 if (CMSCleanOnEnter) {
6873 size_t sz = _collector->block_size_using_printezis_bits(addr);
6874 HeapWord* end_card_addr = (HeapWord*)round_to(
6875 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6876 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6877 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6878 // Bump _threshold to end_card_addr; note that
6879 // _threshold cannot possibly exceed end_card_addr, anyhow.
6880 // This prevents future clearing of the card as the scan proceeds
6881 // to the right.
6882 assert(_threshold <= end_card_addr,
6883 "Because we are just scanning into this object");
6884 if (_threshold < end_card_addr) {
6885 _threshold = end_card_addr;
6886 }
6887 if (p->klass_or_null() != NULL) {
6888 // Redirty the range of cards...
6889 _mut->mark_range(redirty_range);
6890 } // ...else the setting of klass will dirty the card anyway.
6891 }
6892 DEBUG_ONLY(})
6893 return true;
6894 }
6895 }
6896 scanOopsInOop(addr);
6897 return true;
6898 }
6900 // We take a break if we've been at this for a while,
6901 // so as to avoid monopolizing the locks involved.
6902 void MarkFromRootsClosure::do_yield_work() {
6903 // First give up the locks, then yield, then re-lock
6904 // We should probably use a constructor/destructor idiom to
6905 // do this unlock/lock or modify the MutexUnlocker class to
6906 // serve our purpose. XXX
6907 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6908 "CMS thread should hold CMS token");
6909 assert_lock_strong(_bitMap->lock());
6910 _bitMap->lock()->unlock();
6911 ConcurrentMarkSweepThread::desynchronize(true);
6912 ConcurrentMarkSweepThread::acknowledge_yield_request();
6913 _collector->stopTimer();
6914 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6915 if (PrintCMSStatistics != 0) {
6916 _collector->incrementYields();
6917 }
6918 _collector->icms_wait();
6920 // See the comment in coordinator_yield()
6921 for (unsigned i = 0; i < CMSYieldSleepCount &&
6922 ConcurrentMarkSweepThread::should_yield() &&
6923 !CMSCollector::foregroundGCIsActive(); ++i) {
6924 os::sleep(Thread::current(), 1, false);
6925 ConcurrentMarkSweepThread::acknowledge_yield_request();
6926 }
6928 ConcurrentMarkSweepThread::synchronize(true);
6929 _bitMap->lock()->lock_without_safepoint_check();
6930 _collector->startTimer();
6931 }
6933 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6934 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6935 assert(_markStack->isEmpty(),
6936 "should drain stack to limit stack usage");
6937 // convert ptr to an oop preparatory to scanning
6938 oop obj = oop(ptr);
6939 // Ignore mark word in verification below, since we
6940 // may be running concurrent with mutators.
6941 assert(obj->is_oop(true), "should be an oop");
6942 assert(_finger <= ptr, "_finger runneth ahead");
6943 // advance the finger to right end of this object
6944 _finger = ptr + obj->size();
6945 assert(_finger > ptr, "we just incremented it above");
6946 // On large heaps, it may take us some time to get through
6947 // the marking phase (especially if running iCMS). During
6948 // this time it's possible that a lot of mutations have
6949 // accumulated in the card table and the mod union table --
6950 // these mutation records are redundant until we have
6951 // actually traced into the corresponding card.
6952 // Here, we check whether advancing the finger would make
6953 // us cross into a new card, and if so clear corresponding
6954 // cards in the MUT (preclean them in the card-table in the
6955 // future).
6957 DEBUG_ONLY(if (!_verifying) {)
6958 // The clean-on-enter optimization is disabled by default,
6959 // until we fix 6178663.
6960 if (CMSCleanOnEnter && (_finger > _threshold)) {
6961 // [_threshold, _finger) represents the interval
6962 // of cards to be cleared in MUT (or precleaned in card table).
6963 // The set of cards to be cleared is all those that overlap
6964 // with the interval [_threshold, _finger); note that
6965 // _threshold is always kept card-aligned but _finger isn't
6966 // always card-aligned.
6967 HeapWord* old_threshold = _threshold;
6968 assert(old_threshold == (HeapWord*)round_to(
6969 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6970 "_threshold should always be card-aligned");
6971 _threshold = (HeapWord*)round_to(
6972 (intptr_t)_finger, CardTableModRefBS::card_size);
6973 MemRegion mr(old_threshold, _threshold);
6974 assert(!mr.is_empty(), "Control point invariant");
6975 assert(_span.contains(mr), "Should clear within span");
6976 // XXX When _finger crosses from old gen into perm gen
6977 // we may be doing unnecessary cleaning; do better in the
6978 // future by detecting that condition and clearing fewer
6979 // MUT/CT entries.
6980 _mut->clear_range(mr);
6981 }
6982 DEBUG_ONLY(})
6984 // Note: the finger doesn't advance while we drain
6985 // the stack below.
6986 PushOrMarkClosure pushOrMarkClosure(_collector,
6987 _span, _bitMap, _markStack,
6988 _revisitStack,
6989 _finger, this);
6990 bool res = _markStack->push(obj);
6991 assert(res, "Empty non-zero size stack should have space for single push");
6992 while (!_markStack->isEmpty()) {
6993 oop new_oop = _markStack->pop();
6994 // Skip verifying header mark word below because we are
6995 // running concurrent with mutators.
6996 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
6997 // now scan this oop's oops
6998 new_oop->oop_iterate(&pushOrMarkClosure);
6999 do_yield_check();
7000 }
7001 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7002 }
7004 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7005 CMSCollector* collector, MemRegion span,
7006 CMSBitMap* bit_map,
7007 OopTaskQueue* work_queue,
7008 CMSMarkStack* overflow_stack,
7009 CMSMarkStack* revisit_stack,
7010 bool should_yield):
7011 _collector(collector),
7012 _whole_span(collector->_span),
7013 _span(span),
7014 _bit_map(bit_map),
7015 _mut(&collector->_modUnionTable),
7016 _work_queue(work_queue),
7017 _overflow_stack(overflow_stack),
7018 _revisit_stack(revisit_stack),
7019 _yield(should_yield),
7020 _skip_bits(0),
7021 _task(task)
7022 {
7023 assert(_work_queue->size() == 0, "work_queue should be empty");
7024 _finger = span.start();
7025 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7026 assert(_span.contains(_finger), "Out of bounds _finger?");
7027 }
7029 // Should revisit to see if this should be restructured for
7030 // greater efficiency.
7031 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7032 if (_skip_bits > 0) {
7033 _skip_bits--;
7034 return true;
7035 }
7036 // convert offset into a HeapWord*
7037 HeapWord* addr = _bit_map->startWord() + offset;
7038 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7039 "address out of range");
7040 assert(_bit_map->isMarked(addr), "tautology");
7041 if (_bit_map->isMarked(addr+1)) {
7042 // this is an allocated object that might not yet be initialized
7043 assert(_skip_bits == 0, "tautology");
7044 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7045 oop p = oop(addr);
7046 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7047 // in the case of Clean-on-Enter optimization, redirty card
7048 // and avoid clearing card by increasing the threshold.
7049 return true;
7050 }
7051 }
7052 scan_oops_in_oop(addr);
7053 return true;
7054 }
7056 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7057 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7058 // Should we assert that our work queue is empty or
7059 // below some drain limit?
7060 assert(_work_queue->size() == 0,
7061 "should drain stack to limit stack usage");
7062 // convert ptr to an oop preparatory to scanning
7063 oop obj = oop(ptr);
7064 // Ignore mark word in verification below, since we
7065 // may be running concurrent with mutators.
7066 assert(obj->is_oop(true), "should be an oop");
7067 assert(_finger <= ptr, "_finger runneth ahead");
7068 // advance the finger to right end of this object
7069 _finger = ptr + obj->size();
7070 assert(_finger > ptr, "we just incremented it above");
7071 // On large heaps, it may take us some time to get through
7072 // the marking phase (especially if running iCMS). During
7073 // this time it's possible that a lot of mutations have
7074 // accumulated in the card table and the mod union table --
7075 // these mutation records are redundant until we have
7076 // actually traced into the corresponding card.
7077 // Here, we check whether advancing the finger would make
7078 // us cross into a new card, and if so clear corresponding
7079 // cards in the MUT (preclean them in the card-table in the
7080 // future).
7082 // The clean-on-enter optimization is disabled by default,
7083 // until we fix 6178663.
7084 if (CMSCleanOnEnter && (_finger > _threshold)) {
7085 // [_threshold, _finger) represents the interval
7086 // of cards to be cleared in MUT (or precleaned in card table).
7087 // The set of cards to be cleared is all those that overlap
7088 // with the interval [_threshold, _finger); note that
7089 // _threshold is always kept card-aligned but _finger isn't
7090 // always card-aligned.
7091 HeapWord* old_threshold = _threshold;
7092 assert(old_threshold == (HeapWord*)round_to(
7093 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7094 "_threshold should always be card-aligned");
7095 _threshold = (HeapWord*)round_to(
7096 (intptr_t)_finger, CardTableModRefBS::card_size);
7097 MemRegion mr(old_threshold, _threshold);
7098 assert(!mr.is_empty(), "Control point invariant");
7099 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7100 // XXX When _finger crosses from old gen into perm gen
7101 // we may be doing unnecessary cleaning; do better in the
7102 // future by detecting that condition and clearing fewer
7103 // MUT/CT entries.
7104 _mut->clear_range(mr);
7105 }
7107 // Note: the local finger doesn't advance while we drain
7108 // the stack below, but the global finger sure can and will.
7109 HeapWord** gfa = _task->global_finger_addr();
7110 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7111 _span, _bit_map,
7112 _work_queue,
7113 _overflow_stack,
7114 _revisit_stack,
7115 _finger,
7116 gfa, this);
7117 bool res = _work_queue->push(obj); // overflow could occur here
7118 assert(res, "Will hold once we use workqueues");
7119 while (true) {
7120 oop new_oop;
7121 if (!_work_queue->pop_local(new_oop)) {
7122 // We emptied our work_queue; check if there's stuff that can
7123 // be gotten from the overflow stack.
7124 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7125 _overflow_stack, _work_queue)) {
7126 do_yield_check();
7127 continue;
7128 } else { // done
7129 break;
7130 }
7131 }
7132 // Skip verifying header mark word below because we are
7133 // running concurrent with mutators.
7134 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7135 // now scan this oop's oops
7136 new_oop->oop_iterate(&pushOrMarkClosure);
7137 do_yield_check();
7138 }
7139 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7140 }
7142 // Yield in response to a request from VM Thread or
7143 // from mutators.
7144 void Par_MarkFromRootsClosure::do_yield_work() {
7145 assert(_task != NULL, "sanity");
7146 _task->yield();
7147 }
7149 // A variant of the above used for verifying CMS marking work.
7150 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7151 MemRegion span,
7152 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7153 CMSMarkStack* mark_stack):
7154 _collector(collector),
7155 _span(span),
7156 _verification_bm(verification_bm),
7157 _cms_bm(cms_bm),
7158 _mark_stack(mark_stack),
7159 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7160 mark_stack)
7161 {
7162 assert(_mark_stack->isEmpty(), "stack should be empty");
7163 _finger = _verification_bm->startWord();
7164 assert(_collector->_restart_addr == NULL, "Sanity check");
7165 assert(_span.contains(_finger), "Out of bounds _finger?");
7166 }
7168 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7169 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7170 assert(_span.contains(addr), "Out of bounds _finger?");
7171 _finger = addr;
7172 }
7174 // Should revisit to see if this should be restructured for
7175 // greater efficiency.
7176 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7177 // convert offset into a HeapWord*
7178 HeapWord* addr = _verification_bm->startWord() + offset;
7179 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7180 "address out of range");
7181 assert(_verification_bm->isMarked(addr), "tautology");
7182 assert(_cms_bm->isMarked(addr), "tautology");
7184 assert(_mark_stack->isEmpty(),
7185 "should drain stack to limit stack usage");
7186 // convert addr to an oop preparatory to scanning
7187 oop obj = oop(addr);
7188 assert(obj->is_oop(), "should be an oop");
7189 assert(_finger <= addr, "_finger runneth ahead");
7190 // advance the finger to right end of this object
7191 _finger = addr + obj->size();
7192 assert(_finger > addr, "we just incremented it above");
7193 // Note: the finger doesn't advance while we drain
7194 // the stack below.
7195 bool res = _mark_stack->push(obj);
7196 assert(res, "Empty non-zero size stack should have space for single push");
7197 while (!_mark_stack->isEmpty()) {
7198 oop new_oop = _mark_stack->pop();
7199 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7200 // now scan this oop's oops
7201 new_oop->oop_iterate(&_pam_verify_closure);
7202 }
7203 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7204 return true;
7205 }
7207 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7208 CMSCollector* collector, MemRegion span,
7209 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7210 CMSMarkStack* mark_stack):
7211 OopClosure(collector->ref_processor()),
7212 _collector(collector),
7213 _span(span),
7214 _verification_bm(verification_bm),
7215 _cms_bm(cms_bm),
7216 _mark_stack(mark_stack)
7217 { }
7219 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7220 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7222 // Upon stack overflow, we discard (part of) the stack,
7223 // remembering the least address amongst those discarded
7224 // in CMSCollector's _restart_address.
7225 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7226 // Remember the least grey address discarded
7227 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7228 _collector->lower_restart_addr(ra);
7229 _mark_stack->reset(); // discard stack contents
7230 _mark_stack->expand(); // expand the stack if possible
7231 }
7233 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7234 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7235 HeapWord* addr = (HeapWord*)obj;
7236 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7237 // Oop lies in _span and isn't yet grey or black
7238 _verification_bm->mark(addr); // now grey
7239 if (!_cms_bm->isMarked(addr)) {
7240 oop(addr)->print();
7241 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7242 addr);
7243 fatal("... aborting");
7244 }
7246 if (!_mark_stack->push(obj)) { // stack overflow
7247 if (PrintCMSStatistics != 0) {
7248 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7249 SIZE_FORMAT, _mark_stack->capacity());
7250 }
7251 assert(_mark_stack->isFull(), "Else push should have succeeded");
7252 handle_stack_overflow(addr);
7253 }
7254 // anything including and to the right of _finger
7255 // will be scanned as we iterate over the remainder of the
7256 // bit map
7257 }
7258 }
7260 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7261 MemRegion span,
7262 CMSBitMap* bitMap, CMSMarkStack* markStack,
7263 CMSMarkStack* revisitStack,
7264 HeapWord* finger, MarkFromRootsClosure* parent) :
7265 OopClosure(collector->ref_processor()),
7266 _collector(collector),
7267 _span(span),
7268 _bitMap(bitMap),
7269 _markStack(markStack),
7270 _revisitStack(revisitStack),
7271 _finger(finger),
7272 _parent(parent),
7273 _should_remember_klasses(collector->should_unload_classes())
7274 { }
7276 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7277 MemRegion span,
7278 CMSBitMap* bit_map,
7279 OopTaskQueue* work_queue,
7280 CMSMarkStack* overflow_stack,
7281 CMSMarkStack* revisit_stack,
7282 HeapWord* finger,
7283 HeapWord** global_finger_addr,
7284 Par_MarkFromRootsClosure* parent) :
7285 OopClosure(collector->ref_processor()),
7286 _collector(collector),
7287 _whole_span(collector->_span),
7288 _span(span),
7289 _bit_map(bit_map),
7290 _work_queue(work_queue),
7291 _overflow_stack(overflow_stack),
7292 _revisit_stack(revisit_stack),
7293 _finger(finger),
7294 _global_finger_addr(global_finger_addr),
7295 _parent(parent),
7296 _should_remember_klasses(collector->should_unload_classes())
7297 { }
7299 void CMSCollector::lower_restart_addr(HeapWord* low) {
7300 assert(_span.contains(low), "Out of bounds addr");
7301 if (_restart_addr == NULL) {
7302 _restart_addr = low;
7303 } else {
7304 _restart_addr = MIN2(_restart_addr, low);
7305 }
7306 }
7308 // Upon stack overflow, we discard (part of) the stack,
7309 // remembering the least address amongst those discarded
7310 // in CMSCollector's _restart_address.
7311 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7312 // Remember the least grey address discarded
7313 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7314 _collector->lower_restart_addr(ra);
7315 _markStack->reset(); // discard stack contents
7316 _markStack->expand(); // expand the stack if possible
7317 }
7319 // Upon stack overflow, we discard (part of) the stack,
7320 // remembering the least address amongst those discarded
7321 // in CMSCollector's _restart_address.
7322 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7323 // We need to do this under a mutex to prevent other
7324 // workers from interfering with the expansion below.
7325 MutexLockerEx ml(_overflow_stack->par_lock(),
7326 Mutex::_no_safepoint_check_flag);
7327 // Remember the least grey address discarded
7328 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7329 _collector->lower_restart_addr(ra);
7330 _overflow_stack->reset(); // discard stack contents
7331 _overflow_stack->expand(); // expand the stack if possible
7332 }
7334 void PushOrMarkClosure::do_oop(oop obj) {
7335 // Ignore mark word because we are running concurrent with mutators.
7336 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7337 HeapWord* addr = (HeapWord*)obj;
7338 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7339 // Oop lies in _span and isn't yet grey or black
7340 _bitMap->mark(addr); // now grey
7341 if (addr < _finger) {
7342 // the bit map iteration has already either passed, or
7343 // sampled, this bit in the bit map; we'll need to
7344 // use the marking stack to scan this oop's oops.
7345 bool simulate_overflow = false;
7346 NOT_PRODUCT(
7347 if (CMSMarkStackOverflowALot &&
7348 _collector->simulate_overflow()) {
7349 // simulate a stack overflow
7350 simulate_overflow = true;
7351 }
7352 )
7353 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7354 if (PrintCMSStatistics != 0) {
7355 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7356 SIZE_FORMAT, _markStack->capacity());
7357 }
7358 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7359 handle_stack_overflow(addr);
7360 }
7361 }
7362 // anything including and to the right of _finger
7363 // will be scanned as we iterate over the remainder of the
7364 // bit map
7365 do_yield_check();
7366 }
7367 }
7369 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7370 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7372 void Par_PushOrMarkClosure::do_oop(oop obj) {
7373 // Ignore mark word because we are running concurrent with mutators.
7374 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7375 HeapWord* addr = (HeapWord*)obj;
7376 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7377 // Oop lies in _span and isn't yet grey or black
7378 // We read the global_finger (volatile read) strictly after marking oop
7379 bool res = _bit_map->par_mark(addr); // now grey
7380 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7381 // Should we push this marked oop on our stack?
7382 // -- if someone else marked it, nothing to do
7383 // -- if target oop is above global finger nothing to do
7384 // -- if target oop is in chunk and above local finger
7385 // then nothing to do
7386 // -- else push on work queue
7387 if ( !res // someone else marked it, they will deal with it
7388 || (addr >= *gfa) // will be scanned in a later task
7389 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7390 return;
7391 }
7392 // the bit map iteration has already either passed, or
7393 // sampled, this bit in the bit map; we'll need to
7394 // use the marking stack to scan this oop's oops.
7395 bool simulate_overflow = false;
7396 NOT_PRODUCT(
7397 if (CMSMarkStackOverflowALot &&
7398 _collector->simulate_overflow()) {
7399 // simulate a stack overflow
7400 simulate_overflow = true;
7401 }
7402 )
7403 if (simulate_overflow ||
7404 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7405 // stack overflow
7406 if (PrintCMSStatistics != 0) {
7407 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7408 SIZE_FORMAT, _overflow_stack->capacity());
7409 }
7410 // We cannot assert that the overflow stack is full because
7411 // it may have been emptied since.
7412 assert(simulate_overflow ||
7413 _work_queue->size() == _work_queue->max_elems(),
7414 "Else push should have succeeded");
7415 handle_stack_overflow(addr);
7416 }
7417 do_yield_check();
7418 }
7419 }
7421 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7422 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7424 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7425 MemRegion span,
7426 ReferenceProcessor* rp,
7427 CMSBitMap* bit_map,
7428 CMSBitMap* mod_union_table,
7429 CMSMarkStack* mark_stack,
7430 CMSMarkStack* revisit_stack,
7431 bool concurrent_precleaning):
7432 OopClosure(rp),
7433 _collector(collector),
7434 _span(span),
7435 _bit_map(bit_map),
7436 _mod_union_table(mod_union_table),
7437 _mark_stack(mark_stack),
7438 _revisit_stack(revisit_stack),
7439 _concurrent_precleaning(concurrent_precleaning),
7440 _should_remember_klasses(collector->should_unload_classes())
7441 {
7442 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7443 }
7445 // Grey object rescan during pre-cleaning and second checkpoint phases --
7446 // the non-parallel version (the parallel version appears further below.)
7447 void PushAndMarkClosure::do_oop(oop obj) {
7448 // Ignore mark word verification. If during concurrent precleaning,
7449 // the object monitor may be locked. If during the checkpoint
7450 // phases, the object may already have been reached by a different
7451 // path and may be at the end of the global overflow list (so
7452 // the mark word may be NULL).
7453 assert(obj->is_oop_or_null(true /* ignore mark word */),
7454 "expected an oop or NULL");
7455 HeapWord* addr = (HeapWord*)obj;
7456 // Check if oop points into the CMS generation
7457 // and is not marked
7458 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7459 // a white object ...
7460 _bit_map->mark(addr); // ... now grey
7461 // push on the marking stack (grey set)
7462 bool simulate_overflow = false;
7463 NOT_PRODUCT(
7464 if (CMSMarkStackOverflowALot &&
7465 _collector->simulate_overflow()) {
7466 // simulate a stack overflow
7467 simulate_overflow = true;
7468 }
7469 )
7470 if (simulate_overflow || !_mark_stack->push(obj)) {
7471 if (_concurrent_precleaning) {
7472 // During precleaning we can just dirty the appropriate card(s)
7473 // in the mod union table, thus ensuring that the object remains
7474 // in the grey set and continue. In the case of object arrays
7475 // we need to dirty all of the cards that the object spans,
7476 // since the rescan of object arrays will be limited to the
7477 // dirty cards.
7478 // Note that no one can be intefering with us in this action
7479 // of dirtying the mod union table, so no locking or atomics
7480 // are required.
7481 if (obj->is_objArray()) {
7482 size_t sz = obj->size();
7483 HeapWord* end_card_addr = (HeapWord*)round_to(
7484 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7485 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7486 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7487 _mod_union_table->mark_range(redirty_range);
7488 } else {
7489 _mod_union_table->mark(addr);
7490 }
7491 _collector->_ser_pmc_preclean_ovflw++;
7492 } else {
7493 // During the remark phase, we need to remember this oop
7494 // in the overflow list.
7495 _collector->push_on_overflow_list(obj);
7496 _collector->_ser_pmc_remark_ovflw++;
7497 }
7498 }
7499 }
7500 }
7502 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7503 MemRegion span,
7504 ReferenceProcessor* rp,
7505 CMSBitMap* bit_map,
7506 OopTaskQueue* work_queue,
7507 CMSMarkStack* revisit_stack):
7508 OopClosure(rp),
7509 _collector(collector),
7510 _span(span),
7511 _bit_map(bit_map),
7512 _work_queue(work_queue),
7513 _revisit_stack(revisit_stack),
7514 _should_remember_klasses(collector->should_unload_classes())
7515 {
7516 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7517 }
7519 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7520 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7522 // Grey object rescan during second checkpoint phase --
7523 // the parallel version.
7524 void Par_PushAndMarkClosure::do_oop(oop obj) {
7525 // In the assert below, we ignore the mark word because
7526 // this oop may point to an already visited object that is
7527 // on the overflow stack (in which case the mark word has
7528 // been hijacked for chaining into the overflow stack --
7529 // if this is the last object in the overflow stack then
7530 // its mark word will be NULL). Because this object may
7531 // have been subsequently popped off the global overflow
7532 // stack, and the mark word possibly restored to the prototypical
7533 // value, by the time we get to examined this failing assert in
7534 // the debugger, is_oop_or_null(false) may subsequently start
7535 // to hold.
7536 assert(obj->is_oop_or_null(true),
7537 "expected an oop or NULL");
7538 HeapWord* addr = (HeapWord*)obj;
7539 // Check if oop points into the CMS generation
7540 // and is not marked
7541 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7542 // a white object ...
7543 // If we manage to "claim" the object, by being the
7544 // first thread to mark it, then we push it on our
7545 // marking stack
7546 if (_bit_map->par_mark(addr)) { // ... now grey
7547 // push on work queue (grey set)
7548 bool simulate_overflow = false;
7549 NOT_PRODUCT(
7550 if (CMSMarkStackOverflowALot &&
7551 _collector->par_simulate_overflow()) {
7552 // simulate a stack overflow
7553 simulate_overflow = true;
7554 }
7555 )
7556 if (simulate_overflow || !_work_queue->push(obj)) {
7557 _collector->par_push_on_overflow_list(obj);
7558 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7559 }
7560 } // Else, some other thread got there first
7561 }
7562 }
7564 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7565 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7567 void PushAndMarkClosure::remember_klass(Klass* k) {
7568 if (!_revisit_stack->push(oop(k))) {
7569 fatal("Revisit stack overflowed in PushAndMarkClosure");
7570 }
7571 }
7573 void Par_PushAndMarkClosure::remember_klass(Klass* k) {
7574 if (!_revisit_stack->par_push(oop(k))) {
7575 fatal("Revist stack overflowed in Par_PushAndMarkClosure");
7576 }
7577 }
7579 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7580 Mutex* bml = _collector->bitMapLock();
7581 assert_lock_strong(bml);
7582 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7583 "CMS thread should hold CMS token");
7585 bml->unlock();
7586 ConcurrentMarkSweepThread::desynchronize(true);
7588 ConcurrentMarkSweepThread::acknowledge_yield_request();
7590 _collector->stopTimer();
7591 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7592 if (PrintCMSStatistics != 0) {
7593 _collector->incrementYields();
7594 }
7595 _collector->icms_wait();
7597 // See the comment in coordinator_yield()
7598 for (unsigned i = 0; i < CMSYieldSleepCount &&
7599 ConcurrentMarkSweepThread::should_yield() &&
7600 !CMSCollector::foregroundGCIsActive(); ++i) {
7601 os::sleep(Thread::current(), 1, false);
7602 ConcurrentMarkSweepThread::acknowledge_yield_request();
7603 }
7605 ConcurrentMarkSweepThread::synchronize(true);
7606 bml->lock();
7608 _collector->startTimer();
7609 }
7611 bool CMSPrecleanRefsYieldClosure::should_return() {
7612 if (ConcurrentMarkSweepThread::should_yield()) {
7613 do_yield_work();
7614 }
7615 return _collector->foregroundGCIsActive();
7616 }
7618 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7619 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7620 "mr should be aligned to start at a card boundary");
7621 // We'd like to assert:
7622 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7623 // "mr should be a range of cards");
7624 // However, that would be too strong in one case -- the last
7625 // partition ends at _unallocated_block which, in general, can be
7626 // an arbitrary boundary, not necessarily card aligned.
7627 if (PrintCMSStatistics != 0) {
7628 _num_dirty_cards +=
7629 mr.word_size()/CardTableModRefBS::card_size_in_words;
7630 }
7631 _space->object_iterate_mem(mr, &_scan_cl);
7632 }
7634 SweepClosure::SweepClosure(CMSCollector* collector,
7635 ConcurrentMarkSweepGeneration* g,
7636 CMSBitMap* bitMap, bool should_yield) :
7637 _collector(collector),
7638 _g(g),
7639 _sp(g->cmsSpace()),
7640 _limit(_sp->sweep_limit()),
7641 _freelistLock(_sp->freelistLock()),
7642 _bitMap(bitMap),
7643 _yield(should_yield),
7644 _inFreeRange(false), // No free range at beginning of sweep
7645 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7646 _lastFreeRangeCoalesced(false),
7647 _freeFinger(g->used_region().start())
7648 {
7649 NOT_PRODUCT(
7650 _numObjectsFreed = 0;
7651 _numWordsFreed = 0;
7652 _numObjectsLive = 0;
7653 _numWordsLive = 0;
7654 _numObjectsAlreadyFree = 0;
7655 _numWordsAlreadyFree = 0;
7656 _last_fc = NULL;
7658 _sp->initializeIndexedFreeListArrayReturnedBytes();
7659 _sp->dictionary()->initializeDictReturnedBytes();
7660 )
7661 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7662 "sweep _limit out of bounds");
7663 if (CMSTraceSweeper) {
7664 gclog_or_tty->print("\n====================\nStarting new sweep\n");
7665 }
7666 }
7668 // We need this destructor to reclaim any space at the end
7669 // of the space, which do_blk below may not have added back to
7670 // the free lists. [basically dealing with the "fringe effect"]
7671 SweepClosure::~SweepClosure() {
7672 assert_lock_strong(_freelistLock);
7673 // this should be treated as the end of a free run if any
7674 // The current free range should be returned to the free lists
7675 // as one coalesced chunk.
7676 if (inFreeRange()) {
7677 flushCurFreeChunk(freeFinger(),
7678 pointer_delta(_limit, freeFinger()));
7679 assert(freeFinger() < _limit, "the finger pointeth off base");
7680 if (CMSTraceSweeper) {
7681 gclog_or_tty->print("destructor:");
7682 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
7683 "[coalesced:"SIZE_FORMAT"]\n",
7684 freeFinger(), pointer_delta(_limit, freeFinger()),
7685 lastFreeRangeCoalesced());
7686 }
7687 }
7688 NOT_PRODUCT(
7689 if (Verbose && PrintGC) {
7690 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7691 SIZE_FORMAT " bytes",
7692 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7693 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7694 SIZE_FORMAT" bytes "
7695 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7696 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7697 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7698 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7699 sizeof(HeapWord);
7700 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7702 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7703 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7704 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7705 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7706 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7707 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7708 indexListReturnedBytes);
7709 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7710 dictReturnedBytes);
7711 }
7712 }
7713 )
7714 // Now, in debug mode, just null out the sweep_limit
7715 NOT_PRODUCT(_sp->clear_sweep_limit();)
7716 if (CMSTraceSweeper) {
7717 gclog_or_tty->print("end of sweep\n================\n");
7718 }
7719 }
7721 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7722 bool freeRangeInFreeLists) {
7723 if (CMSTraceSweeper) {
7724 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
7725 freeFinger, _sp->block_size(freeFinger),
7726 freeRangeInFreeLists);
7727 }
7728 assert(!inFreeRange(), "Trampling existing free range");
7729 set_inFreeRange(true);
7730 set_lastFreeRangeCoalesced(false);
7732 set_freeFinger(freeFinger);
7733 set_freeRangeInFreeLists(freeRangeInFreeLists);
7734 if (CMSTestInFreeList) {
7735 if (freeRangeInFreeLists) {
7736 FreeChunk* fc = (FreeChunk*) freeFinger;
7737 assert(fc->isFree(), "A chunk on the free list should be free.");
7738 assert(fc->size() > 0, "Free range should have a size");
7739 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7740 }
7741 }
7742 }
7744 // Note that the sweeper runs concurrently with mutators. Thus,
7745 // it is possible for direct allocation in this generation to happen
7746 // in the middle of the sweep. Note that the sweeper also coalesces
7747 // contiguous free blocks. Thus, unless the sweeper and the allocator
7748 // synchronize appropriately freshly allocated blocks may get swept up.
7749 // This is accomplished by the sweeper locking the free lists while
7750 // it is sweeping. Thus blocks that are determined to be free are
7751 // indeed free. There is however one additional complication:
7752 // blocks that have been allocated since the final checkpoint and
7753 // mark, will not have been marked and so would be treated as
7754 // unreachable and swept up. To prevent this, the allocator marks
7755 // the bit map when allocating during the sweep phase. This leads,
7756 // however, to a further complication -- objects may have been allocated
7757 // but not yet initialized -- in the sense that the header isn't yet
7758 // installed. The sweeper can not then determine the size of the block
7759 // in order to skip over it. To deal with this case, we use a technique
7760 // (due to Printezis) to encode such uninitialized block sizes in the
7761 // bit map. Since the bit map uses a bit per every HeapWord, but the
7762 // CMS generation has a minimum object size of 3 HeapWords, it follows
7763 // that "normal marks" won't be adjacent in the bit map (there will
7764 // always be at least two 0 bits between successive 1 bits). We make use
7765 // of these "unused" bits to represent uninitialized blocks -- the bit
7766 // corresponding to the start of the uninitialized object and the next
7767 // bit are both set. Finally, a 1 bit marks the end of the object that
7768 // started with the two consecutive 1 bits to indicate its potentially
7769 // uninitialized state.
7771 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7772 FreeChunk* fc = (FreeChunk*)addr;
7773 size_t res;
7775 // check if we are done sweepinrg
7776 if (addr == _limit) { // we have swept up to the limit, do nothing more
7777 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7778 "sweep _limit out of bounds");
7779 // help the closure application finish
7780 return pointer_delta(_sp->end(), _limit);
7781 }
7782 assert(addr <= _limit, "sweep invariant");
7784 // check if we should yield
7785 do_yield_check(addr);
7786 if (fc->isFree()) {
7787 // Chunk that is already free
7788 res = fc->size();
7789 doAlreadyFreeChunk(fc);
7790 debug_only(_sp->verifyFreeLists());
7791 assert(res == fc->size(), "Don't expect the size to change");
7792 NOT_PRODUCT(
7793 _numObjectsAlreadyFree++;
7794 _numWordsAlreadyFree += res;
7795 )
7796 NOT_PRODUCT(_last_fc = fc;)
7797 } else if (!_bitMap->isMarked(addr)) {
7798 // Chunk is fresh garbage
7799 res = doGarbageChunk(fc);
7800 debug_only(_sp->verifyFreeLists());
7801 NOT_PRODUCT(
7802 _numObjectsFreed++;
7803 _numWordsFreed += res;
7804 )
7805 } else {
7806 // Chunk that is alive.
7807 res = doLiveChunk(fc);
7808 debug_only(_sp->verifyFreeLists());
7809 NOT_PRODUCT(
7810 _numObjectsLive++;
7811 _numWordsLive += res;
7812 )
7813 }
7814 return res;
7815 }
7817 // For the smart allocation, record following
7818 // split deaths - a free chunk is removed from its free list because
7819 // it is being split into two or more chunks.
7820 // split birth - a free chunk is being added to its free list because
7821 // a larger free chunk has been split and resulted in this free chunk.
7822 // coal death - a free chunk is being removed from its free list because
7823 // it is being coalesced into a large free chunk.
7824 // coal birth - a free chunk is being added to its free list because
7825 // it was created when two or more free chunks where coalesced into
7826 // this free chunk.
7827 //
7828 // These statistics are used to determine the desired number of free
7829 // chunks of a given size. The desired number is chosen to be relative
7830 // to the end of a CMS sweep. The desired number at the end of a sweep
7831 // is the
7832 // count-at-end-of-previous-sweep (an amount that was enough)
7833 // - count-at-beginning-of-current-sweep (the excess)
7834 // + split-births (gains in this size during interval)
7835 // - split-deaths (demands on this size during interval)
7836 // where the interval is from the end of one sweep to the end of the
7837 // next.
7838 //
7839 // When sweeping the sweeper maintains an accumulated chunk which is
7840 // the chunk that is made up of chunks that have been coalesced. That
7841 // will be termed the left-hand chunk. A new chunk of garbage that
7842 // is being considered for coalescing will be referred to as the
7843 // right-hand chunk.
7844 //
7845 // When making a decision on whether to coalesce a right-hand chunk with
7846 // the current left-hand chunk, the current count vs. the desired count
7847 // of the left-hand chunk is considered. Also if the right-hand chunk
7848 // is near the large chunk at the end of the heap (see
7849 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7850 // left-hand chunk is coalesced.
7851 //
7852 // When making a decision about whether to split a chunk, the desired count
7853 // vs. the current count of the candidate to be split is also considered.
7854 // If the candidate is underpopulated (currently fewer chunks than desired)
7855 // a chunk of an overpopulated (currently more chunks than desired) size may
7856 // be chosen. The "hint" associated with a free list, if non-null, points
7857 // to a free list which may be overpopulated.
7858 //
7860 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
7861 size_t size = fc->size();
7862 // Chunks that cannot be coalesced are not in the
7863 // free lists.
7864 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7865 assert(_sp->verifyChunkInFreeLists(fc),
7866 "free chunk should be in free lists");
7867 }
7868 // a chunk that is already free, should not have been
7869 // marked in the bit map
7870 HeapWord* addr = (HeapWord*) fc;
7871 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7872 // Verify that the bit map has no bits marked between
7873 // addr and purported end of this block.
7874 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7876 // Some chunks cannot be coalesced in under any circumstances.
7877 // See the definition of cantCoalesce().
7878 if (!fc->cantCoalesce()) {
7879 // This chunk can potentially be coalesced.
7880 if (_sp->adaptive_freelists()) {
7881 // All the work is done in
7882 doPostIsFreeOrGarbageChunk(fc, size);
7883 } else { // Not adaptive free lists
7884 // this is a free chunk that can potentially be coalesced by the sweeper;
7885 if (!inFreeRange()) {
7886 // if the next chunk is a free block that can't be coalesced
7887 // it doesn't make sense to remove this chunk from the free lists
7888 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7889 assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
7890 if ((HeapWord*)nextChunk < _limit && // there's a next chunk...
7891 nextChunk->isFree() && // which is free...
7892 nextChunk->cantCoalesce()) { // ... but cant be coalesced
7893 // nothing to do
7894 } else {
7895 // Potentially the start of a new free range:
7896 // Don't eagerly remove it from the free lists.
7897 // No need to remove it if it will just be put
7898 // back again. (Also from a pragmatic point of view
7899 // if it is a free block in a region that is beyond
7900 // any allocated blocks, an assertion will fail)
7901 // Remember the start of a free run.
7902 initialize_free_range(addr, true);
7903 // end - can coalesce with next chunk
7904 }
7905 } else {
7906 // the midst of a free range, we are coalescing
7907 debug_only(record_free_block_coalesced(fc);)
7908 if (CMSTraceSweeper) {
7909 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
7910 }
7911 // remove it from the free lists
7912 _sp->removeFreeChunkFromFreeLists(fc);
7913 set_lastFreeRangeCoalesced(true);
7914 // If the chunk is being coalesced and the current free range is
7915 // in the free lists, remove the current free range so that it
7916 // will be returned to the free lists in its entirety - all
7917 // the coalesced pieces included.
7918 if (freeRangeInFreeLists()) {
7919 FreeChunk* ffc = (FreeChunk*) freeFinger();
7920 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7921 "Size of free range is inconsistent with chunk size.");
7922 if (CMSTestInFreeList) {
7923 assert(_sp->verifyChunkInFreeLists(ffc),
7924 "free range is not in free lists");
7925 }
7926 _sp->removeFreeChunkFromFreeLists(ffc);
7927 set_freeRangeInFreeLists(false);
7928 }
7929 }
7930 }
7931 } else {
7932 // Code path common to both original and adaptive free lists.
7934 // cant coalesce with previous block; this should be treated
7935 // as the end of a free run if any
7936 if (inFreeRange()) {
7937 // we kicked some butt; time to pick up the garbage
7938 assert(freeFinger() < addr, "the finger pointeth off base");
7939 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
7940 }
7941 // else, nothing to do, just continue
7942 }
7943 }
7945 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
7946 // This is a chunk of garbage. It is not in any free list.
7947 // Add it to a free list or let it possibly be coalesced into
7948 // a larger chunk.
7949 HeapWord* addr = (HeapWord*) fc;
7950 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7952 if (_sp->adaptive_freelists()) {
7953 // Verify that the bit map has no bits marked between
7954 // addr and purported end of just dead object.
7955 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7957 doPostIsFreeOrGarbageChunk(fc, size);
7958 } else {
7959 if (!inFreeRange()) {
7960 // start of a new free range
7961 assert(size > 0, "A free range should have a size");
7962 initialize_free_range(addr, false);
7964 } else {
7965 // this will be swept up when we hit the end of the
7966 // free range
7967 if (CMSTraceSweeper) {
7968 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
7969 }
7970 // If the chunk is being coalesced and the current free range is
7971 // in the free lists, remove the current free range so that it
7972 // will be returned to the free lists in its entirety - all
7973 // the coalesced pieces included.
7974 if (freeRangeInFreeLists()) {
7975 FreeChunk* ffc = (FreeChunk*)freeFinger();
7976 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7977 "Size of free range is inconsistent with chunk size.");
7978 if (CMSTestInFreeList) {
7979 assert(_sp->verifyChunkInFreeLists(ffc),
7980 "free range is not in free lists");
7981 }
7982 _sp->removeFreeChunkFromFreeLists(ffc);
7983 set_freeRangeInFreeLists(false);
7984 }
7985 set_lastFreeRangeCoalesced(true);
7986 }
7987 // this will be swept up when we hit the end of the free range
7989 // Verify that the bit map has no bits marked between
7990 // addr and purported end of just dead object.
7991 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7992 }
7993 return size;
7994 }
7996 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
7997 HeapWord* addr = (HeapWord*) fc;
7998 // The sweeper has just found a live object. Return any accumulated
7999 // left hand chunk to the free lists.
8000 if (inFreeRange()) {
8001 if (_sp->adaptive_freelists()) {
8002 flushCurFreeChunk(freeFinger(),
8003 pointer_delta(addr, freeFinger()));
8004 } else { // not adaptive freelists
8005 set_inFreeRange(false);
8006 // Add the free range back to the free list if it is not already
8007 // there.
8008 if (!freeRangeInFreeLists()) {
8009 assert(freeFinger() < addr, "the finger pointeth off base");
8010 if (CMSTraceSweeper) {
8011 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
8012 "[coalesced:%d]\n",
8013 freeFinger(), pointer_delta(addr, freeFinger()),
8014 lastFreeRangeCoalesced());
8015 }
8016 _sp->addChunkAndRepairOffsetTable(freeFinger(),
8017 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
8018 }
8019 }
8020 }
8022 // Common code path for original and adaptive free lists.
8024 // this object is live: we'd normally expect this to be
8025 // an oop, and like to assert the following:
8026 // assert(oop(addr)->is_oop(), "live block should be an oop");
8027 // However, as we commented above, this may be an object whose
8028 // header hasn't yet been initialized.
8029 size_t size;
8030 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8031 if (_bitMap->isMarked(addr + 1)) {
8032 // Determine the size from the bit map, rather than trying to
8033 // compute it from the object header.
8034 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8035 size = pointer_delta(nextOneAddr + 1, addr);
8036 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8037 "alignment problem");
8039 #ifdef DEBUG
8040 if (oop(addr)->klass_or_null() != NULL &&
8041 ( !_collector->should_unload_classes()
8042 || oop(addr)->is_parsable())) {
8043 // Ignore mark word because we are running concurrent with mutators
8044 assert(oop(addr)->is_oop(true), "live block should be an oop");
8045 assert(size ==
8046 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8047 "P-mark and computed size do not agree");
8048 }
8049 #endif
8051 } else {
8052 // This should be an initialized object that's alive.
8053 assert(oop(addr)->klass_or_null() != NULL &&
8054 (!_collector->should_unload_classes()
8055 || oop(addr)->is_parsable()),
8056 "Should be an initialized object");
8057 // Ignore mark word because we are running concurrent with mutators
8058 assert(oop(addr)->is_oop(true), "live block should be an oop");
8059 // Verify that the bit map has no bits marked between
8060 // addr and purported end of this block.
8061 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8062 assert(size >= 3, "Necessary for Printezis marks to work");
8063 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8064 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8065 }
8066 return size;
8067 }
8069 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
8070 size_t chunkSize) {
8071 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
8072 // scheme.
8073 bool fcInFreeLists = fc->isFree();
8074 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8075 assert((HeapWord*)fc <= _limit, "sweep invariant");
8076 if (CMSTestInFreeList && fcInFreeLists) {
8077 assert(_sp->verifyChunkInFreeLists(fc),
8078 "free chunk is not in free lists");
8079 }
8082 if (CMSTraceSweeper) {
8083 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8084 }
8086 HeapWord* addr = (HeapWord*) fc;
8088 bool coalesce;
8089 size_t left = pointer_delta(addr, freeFinger());
8090 size_t right = chunkSize;
8091 switch (FLSCoalescePolicy) {
8092 // numeric value forms a coalition aggressiveness metric
8093 case 0: { // never coalesce
8094 coalesce = false;
8095 break;
8096 }
8097 case 1: { // coalesce if left & right chunks on overpopulated lists
8098 coalesce = _sp->coalOverPopulated(left) &&
8099 _sp->coalOverPopulated(right);
8100 break;
8101 }
8102 case 2: { // coalesce if left chunk on overpopulated list (default)
8103 coalesce = _sp->coalOverPopulated(left);
8104 break;
8105 }
8106 case 3: { // coalesce if left OR right chunk on overpopulated list
8107 coalesce = _sp->coalOverPopulated(left) ||
8108 _sp->coalOverPopulated(right);
8109 break;
8110 }
8111 case 4: { // always coalesce
8112 coalesce = true;
8113 break;
8114 }
8115 default:
8116 ShouldNotReachHere();
8117 }
8119 // Should the current free range be coalesced?
8120 // If the chunk is in a free range and either we decided to coalesce above
8121 // or the chunk is near the large block at the end of the heap
8122 // (isNearLargestChunk() returns true), then coalesce this chunk.
8123 bool doCoalesce = inFreeRange() &&
8124 (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8125 if (doCoalesce) {
8126 // Coalesce the current free range on the left with the new
8127 // chunk on the right. If either is on a free list,
8128 // it must be removed from the list and stashed in the closure.
8129 if (freeRangeInFreeLists()) {
8130 FreeChunk* ffc = (FreeChunk*)freeFinger();
8131 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8132 "Size of free range is inconsistent with chunk size.");
8133 if (CMSTestInFreeList) {
8134 assert(_sp->verifyChunkInFreeLists(ffc),
8135 "Chunk is not in free lists");
8136 }
8137 _sp->coalDeath(ffc->size());
8138 _sp->removeFreeChunkFromFreeLists(ffc);
8139 set_freeRangeInFreeLists(false);
8140 }
8141 if (fcInFreeLists) {
8142 _sp->coalDeath(chunkSize);
8143 assert(fc->size() == chunkSize,
8144 "The chunk has the wrong size or is not in the free lists");
8145 _sp->removeFreeChunkFromFreeLists(fc);
8146 }
8147 set_lastFreeRangeCoalesced(true);
8148 } else { // not in a free range and/or should not coalesce
8149 // Return the current free range and start a new one.
8150 if (inFreeRange()) {
8151 // In a free range but cannot coalesce with the right hand chunk.
8152 // Put the current free range into the free lists.
8153 flushCurFreeChunk(freeFinger(),
8154 pointer_delta(addr, freeFinger()));
8155 }
8156 // Set up for new free range. Pass along whether the right hand
8157 // chunk is in the free lists.
8158 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8159 }
8160 }
8161 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
8162 assert(inFreeRange(), "Should only be called if currently in a free range.");
8163 assert(size > 0,
8164 "A zero sized chunk cannot be added to the free lists.");
8165 if (!freeRangeInFreeLists()) {
8166 if(CMSTestInFreeList) {
8167 FreeChunk* fc = (FreeChunk*) chunk;
8168 fc->setSize(size);
8169 assert(!_sp->verifyChunkInFreeLists(fc),
8170 "chunk should not be in free lists yet");
8171 }
8172 if (CMSTraceSweeper) {
8173 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8174 chunk, size);
8175 }
8176 // A new free range is going to be starting. The current
8177 // free range has not been added to the free lists yet or
8178 // was removed so add it back.
8179 // If the current free range was coalesced, then the death
8180 // of the free range was recorded. Record a birth now.
8181 if (lastFreeRangeCoalesced()) {
8182 _sp->coalBirth(size);
8183 }
8184 _sp->addChunkAndRepairOffsetTable(chunk, size,
8185 lastFreeRangeCoalesced());
8186 }
8187 set_inFreeRange(false);
8188 set_freeRangeInFreeLists(false);
8189 }
8191 // We take a break if we've been at this for a while,
8192 // so as to avoid monopolizing the locks involved.
8193 void SweepClosure::do_yield_work(HeapWord* addr) {
8194 // Return current free chunk being used for coalescing (if any)
8195 // to the appropriate freelist. After yielding, the next
8196 // free block encountered will start a coalescing range of
8197 // free blocks. If the next free block is adjacent to the
8198 // chunk just flushed, they will need to wait for the next
8199 // sweep to be coalesced.
8200 if (inFreeRange()) {
8201 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
8202 }
8204 // First give up the locks, then yield, then re-lock.
8205 // We should probably use a constructor/destructor idiom to
8206 // do this unlock/lock or modify the MutexUnlocker class to
8207 // serve our purpose. XXX
8208 assert_lock_strong(_bitMap->lock());
8209 assert_lock_strong(_freelistLock);
8210 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8211 "CMS thread should hold CMS token");
8212 _bitMap->lock()->unlock();
8213 _freelistLock->unlock();
8214 ConcurrentMarkSweepThread::desynchronize(true);
8215 ConcurrentMarkSweepThread::acknowledge_yield_request();
8216 _collector->stopTimer();
8217 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8218 if (PrintCMSStatistics != 0) {
8219 _collector->incrementYields();
8220 }
8221 _collector->icms_wait();
8223 // See the comment in coordinator_yield()
8224 for (unsigned i = 0; i < CMSYieldSleepCount &&
8225 ConcurrentMarkSweepThread::should_yield() &&
8226 !CMSCollector::foregroundGCIsActive(); ++i) {
8227 os::sleep(Thread::current(), 1, false);
8228 ConcurrentMarkSweepThread::acknowledge_yield_request();
8229 }
8231 ConcurrentMarkSweepThread::synchronize(true);
8232 _freelistLock->lock();
8233 _bitMap->lock()->lock_without_safepoint_check();
8234 _collector->startTimer();
8235 }
8237 #ifndef PRODUCT
8238 // This is actually very useful in a product build if it can
8239 // be called from the debugger. Compile it into the product
8240 // as needed.
8241 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8242 return debug_cms_space->verifyChunkInFreeLists(fc);
8243 }
8245 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8246 if (CMSTraceSweeper) {
8247 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8248 }
8249 }
8250 #endif
8252 // CMSIsAliveClosure
8253 bool CMSIsAliveClosure::do_object_b(oop obj) {
8254 HeapWord* addr = (HeapWord*)obj;
8255 return addr != NULL &&
8256 (!_span.contains(addr) || _bit_map->isMarked(addr));
8257 }
8259 // CMSKeepAliveClosure: the serial version
8260 void CMSKeepAliveClosure::do_oop(oop obj) {
8261 HeapWord* addr = (HeapWord*)obj;
8262 if (_span.contains(addr) &&
8263 !_bit_map->isMarked(addr)) {
8264 _bit_map->mark(addr);
8265 bool simulate_overflow = false;
8266 NOT_PRODUCT(
8267 if (CMSMarkStackOverflowALot &&
8268 _collector->simulate_overflow()) {
8269 // simulate a stack overflow
8270 simulate_overflow = true;
8271 }
8272 )
8273 if (simulate_overflow || !_mark_stack->push(obj)) {
8274 _collector->push_on_overflow_list(obj);
8275 _collector->_ser_kac_ovflw++;
8276 }
8277 }
8278 }
8280 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8281 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8283 // CMSParKeepAliveClosure: a parallel version of the above.
8284 // The work queues are private to each closure (thread),
8285 // but (may be) available for stealing by other threads.
8286 void CMSParKeepAliveClosure::do_oop(oop obj) {
8287 HeapWord* addr = (HeapWord*)obj;
8288 if (_span.contains(addr) &&
8289 !_bit_map->isMarked(addr)) {
8290 // In general, during recursive tracing, several threads
8291 // may be concurrently getting here; the first one to
8292 // "tag" it, claims it.
8293 if (_bit_map->par_mark(addr)) {
8294 bool res = _work_queue->push(obj);
8295 assert(res, "Low water mark should be much less than capacity");
8296 // Do a recursive trim in the hope that this will keep
8297 // stack usage lower, but leave some oops for potential stealers
8298 trim_queue(_low_water_mark);
8299 } // Else, another thread got there first
8300 }
8301 }
8303 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8304 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8306 void CMSParKeepAliveClosure::trim_queue(uint max) {
8307 while (_work_queue->size() > max) {
8308 oop new_oop;
8309 if (_work_queue->pop_local(new_oop)) {
8310 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8311 assert(_bit_map->isMarked((HeapWord*)new_oop),
8312 "no white objects on this stack!");
8313 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8314 // iterate over the oops in this oop, marking and pushing
8315 // the ones in CMS heap (i.e. in _span).
8316 new_oop->oop_iterate(&_mark_and_push);
8317 }
8318 }
8319 }
8321 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8322 HeapWord* addr = (HeapWord*)obj;
8323 if (_span.contains(addr) &&
8324 !_bit_map->isMarked(addr)) {
8325 if (_bit_map->par_mark(addr)) {
8326 bool simulate_overflow = false;
8327 NOT_PRODUCT(
8328 if (CMSMarkStackOverflowALot &&
8329 _collector->par_simulate_overflow()) {
8330 // simulate a stack overflow
8331 simulate_overflow = true;
8332 }
8333 )
8334 if (simulate_overflow || !_work_queue->push(obj)) {
8335 _collector->par_push_on_overflow_list(obj);
8336 _collector->_par_kac_ovflw++;
8337 }
8338 } // Else another thread got there already
8339 }
8340 }
8342 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8343 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8345 //////////////////////////////////////////////////////////////////
8346 // CMSExpansionCause /////////////////////////////
8347 //////////////////////////////////////////////////////////////////
8348 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8349 switch (cause) {
8350 case _no_expansion:
8351 return "No expansion";
8352 case _satisfy_free_ratio:
8353 return "Free ratio";
8354 case _satisfy_promotion:
8355 return "Satisfy promotion";
8356 case _satisfy_allocation:
8357 return "allocation";
8358 case _allocate_par_lab:
8359 return "Par LAB";
8360 case _allocate_par_spooling_space:
8361 return "Par Spooling Space";
8362 case _adaptive_size_policy:
8363 return "Ergonomics";
8364 default:
8365 return "unknown";
8366 }
8367 }
8369 void CMSDrainMarkingStackClosure::do_void() {
8370 // the max number to take from overflow list at a time
8371 const size_t num = _mark_stack->capacity()/4;
8372 while (!_mark_stack->isEmpty() ||
8373 // if stack is empty, check the overflow list
8374 _collector->take_from_overflow_list(num, _mark_stack)) {
8375 oop obj = _mark_stack->pop();
8376 HeapWord* addr = (HeapWord*)obj;
8377 assert(_span.contains(addr), "Should be within span");
8378 assert(_bit_map->isMarked(addr), "Should be marked");
8379 assert(obj->is_oop(), "Should be an oop");
8380 obj->oop_iterate(_keep_alive);
8381 }
8382 }
8384 void CMSParDrainMarkingStackClosure::do_void() {
8385 // drain queue
8386 trim_queue(0);
8387 }
8389 // Trim our work_queue so its length is below max at return
8390 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8391 while (_work_queue->size() > max) {
8392 oop new_oop;
8393 if (_work_queue->pop_local(new_oop)) {
8394 assert(new_oop->is_oop(), "Expected an oop");
8395 assert(_bit_map->isMarked((HeapWord*)new_oop),
8396 "no white objects on this stack!");
8397 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8398 // iterate over the oops in this oop, marking and pushing
8399 // the ones in CMS heap (i.e. in _span).
8400 new_oop->oop_iterate(&_mark_and_push);
8401 }
8402 }
8403 }
8405 ////////////////////////////////////////////////////////////////////
8406 // Support for Marking Stack Overflow list handling and related code
8407 ////////////////////////////////////////////////////////////////////
8408 // Much of the following code is similar in shape and spirit to the
8409 // code used in ParNewGC. We should try and share that code
8410 // as much as possible in the future.
8412 #ifndef PRODUCT
8413 // Debugging support for CMSStackOverflowALot
8415 // It's OK to call this multi-threaded; the worst thing
8416 // that can happen is that we'll get a bunch of closely
8417 // spaced simulated oveflows, but that's OK, in fact
8418 // probably good as it would exercise the overflow code
8419 // under contention.
8420 bool CMSCollector::simulate_overflow() {
8421 if (_overflow_counter-- <= 0) { // just being defensive
8422 _overflow_counter = CMSMarkStackOverflowInterval;
8423 return true;
8424 } else {
8425 return false;
8426 }
8427 }
8429 bool CMSCollector::par_simulate_overflow() {
8430 return simulate_overflow();
8431 }
8432 #endif
8434 // Single-threaded
8435 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8436 assert(stack->isEmpty(), "Expected precondition");
8437 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8438 size_t i = num;
8439 oop cur = _overflow_list;
8440 const markOop proto = markOopDesc::prototype();
8441 NOT_PRODUCT(size_t n = 0;)
8442 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8443 next = oop(cur->mark());
8444 cur->set_mark(proto); // until proven otherwise
8445 assert(cur->is_oop(), "Should be an oop");
8446 bool res = stack->push(cur);
8447 assert(res, "Bit off more than can chew?");
8448 NOT_PRODUCT(n++;)
8449 }
8450 _overflow_list = cur;
8451 #ifndef PRODUCT
8452 assert(_num_par_pushes >= n, "Too many pops?");
8453 _num_par_pushes -=n;
8454 #endif
8455 return !stack->isEmpty();
8456 }
8458 // Multi-threaded; use CAS to break off a prefix
8459 bool CMSCollector::par_take_from_overflow_list(size_t num,
8460 OopTaskQueue* work_q) {
8461 assert(work_q->size() == 0, "That's the current policy");
8462 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8463 if (_overflow_list == NULL) {
8464 return false;
8465 }
8466 // Grab the entire list; we'll put back a suffix
8467 oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
8468 if (prefix == NULL) { // someone grabbed it before we did ...
8469 // ... we could spin for a short while, but for now we don't
8470 return false;
8471 }
8472 size_t i = num;
8473 oop cur = prefix;
8474 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8475 if (cur->mark() != NULL) {
8476 oop suffix_head = cur->mark(); // suffix will be put back on global list
8477 cur->set_mark(NULL); // break off suffix
8478 // Find tail of suffix so we can prepend suffix to global list
8479 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8480 oop suffix_tail = cur;
8481 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8482 "Tautology");
8483 oop observed_overflow_list = _overflow_list;
8484 do {
8485 cur = observed_overflow_list;
8486 suffix_tail->set_mark(markOop(cur));
8487 observed_overflow_list =
8488 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
8489 } while (cur != observed_overflow_list);
8490 }
8492 // Push the prefix elements on work_q
8493 assert(prefix != NULL, "control point invariant");
8494 const markOop proto = markOopDesc::prototype();
8495 oop next;
8496 NOT_PRODUCT(size_t n = 0;)
8497 for (cur = prefix; cur != NULL; cur = next) {
8498 next = oop(cur->mark());
8499 cur->set_mark(proto); // until proven otherwise
8500 assert(cur->is_oop(), "Should be an oop");
8501 bool res = work_q->push(cur);
8502 assert(res, "Bit off more than we can chew?");
8503 NOT_PRODUCT(n++;)
8504 }
8505 #ifndef PRODUCT
8506 assert(_num_par_pushes >= n, "Too many pops?");
8507 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8508 #endif
8509 return true;
8510 }
8512 // Single-threaded
8513 void CMSCollector::push_on_overflow_list(oop p) {
8514 NOT_PRODUCT(_num_par_pushes++;)
8515 assert(p->is_oop(), "Not an oop");
8516 preserve_mark_if_necessary(p);
8517 p->set_mark((markOop)_overflow_list);
8518 _overflow_list = p;
8519 }
8521 // Multi-threaded; use CAS to prepend to overflow list
8522 void CMSCollector::par_push_on_overflow_list(oop p) {
8523 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8524 assert(p->is_oop(), "Not an oop");
8525 par_preserve_mark_if_necessary(p);
8526 oop observed_overflow_list = _overflow_list;
8527 oop cur_overflow_list;
8528 do {
8529 cur_overflow_list = observed_overflow_list;
8530 p->set_mark(markOop(cur_overflow_list));
8531 observed_overflow_list =
8532 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8533 } while (cur_overflow_list != observed_overflow_list);
8534 }
8536 // Single threaded
8537 // General Note on GrowableArray: pushes may silently fail
8538 // because we are (temporarily) out of C-heap for expanding
8539 // the stack. The problem is quite ubiquitous and affects
8540 // a lot of code in the JVM. The prudent thing for GrowableArray
8541 // to do (for now) is to exit with an error. However, that may
8542 // be too draconian in some cases because the caller may be
8543 // able to recover without much harm. For suych cases, we
8544 // should probably introduce a "soft_push" method which returns
8545 // an indication of success or failure with the assumption that
8546 // the caller may be able to recover from a failure; code in
8547 // the VM can then be changed, incrementally, to deal with such
8548 // failures where possible, thus, incrementally hardening the VM
8549 // in such low resource situations.
8550 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8551 int PreserveMarkStackSize = 128;
8553 if (_preserved_oop_stack == NULL) {
8554 assert(_preserved_mark_stack == NULL,
8555 "bijection with preserved_oop_stack");
8556 // Allocate the stacks
8557 _preserved_oop_stack = new (ResourceObj::C_HEAP)
8558 GrowableArray<oop>(PreserveMarkStackSize, true);
8559 _preserved_mark_stack = new (ResourceObj::C_HEAP)
8560 GrowableArray<markOop>(PreserveMarkStackSize, true);
8561 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
8562 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
8563 "Preserved Mark/Oop Stack for CMS (C-heap)");
8564 }
8565 }
8566 _preserved_oop_stack->push(p);
8567 _preserved_mark_stack->push(m);
8568 assert(m == p->mark(), "Mark word changed");
8569 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
8570 "bijection");
8571 }
8573 // Single threaded
8574 void CMSCollector::preserve_mark_if_necessary(oop p) {
8575 markOop m = p->mark();
8576 if (m->must_be_preserved(p)) {
8577 preserve_mark_work(p, m);
8578 }
8579 }
8581 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8582 markOop m = p->mark();
8583 if (m->must_be_preserved(p)) {
8584 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8585 // Even though we read the mark word without holding
8586 // the lock, we are assured that it will not change
8587 // because we "own" this oop, so no other thread can
8588 // be trying to push it on the overflow list; see
8589 // the assertion in preserve_mark_work() that checks
8590 // that m == p->mark().
8591 preserve_mark_work(p, m);
8592 }
8593 }
8595 // We should be able to do this multi-threaded,
8596 // a chunk of stack being a task (this is
8597 // correct because each oop only ever appears
8598 // once in the overflow list. However, it's
8599 // not very easy to completely overlap this with
8600 // other operations, so will generally not be done
8601 // until all work's been completed. Because we
8602 // expect the preserved oop stack (set) to be small,
8603 // it's probably fine to do this single-threaded.
8604 // We can explore cleverer concurrent/overlapped/parallel
8605 // processing of preserved marks if we feel the
8606 // need for this in the future. Stack overflow should
8607 // be so rare in practice and, when it happens, its
8608 // effect on performance so great that this will
8609 // likely just be in the noise anyway.
8610 void CMSCollector::restore_preserved_marks_if_any() {
8611 if (_preserved_oop_stack == NULL) {
8612 assert(_preserved_mark_stack == NULL,
8613 "bijection with preserved_oop_stack");
8614 return;
8615 }
8617 assert(SafepointSynchronize::is_at_safepoint(),
8618 "world should be stopped");
8619 assert(Thread::current()->is_ConcurrentGC_thread() ||
8620 Thread::current()->is_VM_thread(),
8621 "should be single-threaded");
8623 int length = _preserved_oop_stack->length();
8624 assert(_preserved_mark_stack->length() == length, "bijection");
8625 for (int i = 0; i < length; i++) {
8626 oop p = _preserved_oop_stack->at(i);
8627 assert(p->is_oop(), "Should be an oop");
8628 assert(_span.contains(p), "oop should be in _span");
8629 assert(p->mark() == markOopDesc::prototype(),
8630 "Set when taken from overflow list");
8631 markOop m = _preserved_mark_stack->at(i);
8632 p->set_mark(m);
8633 }
8634 _preserved_mark_stack->clear();
8635 _preserved_oop_stack->clear();
8636 assert(_preserved_mark_stack->is_empty() &&
8637 _preserved_oop_stack->is_empty(),
8638 "stacks were cleared above");
8639 }
8641 #ifndef PRODUCT
8642 bool CMSCollector::no_preserved_marks() const {
8643 return ( ( _preserved_mark_stack == NULL
8644 && _preserved_oop_stack == NULL)
8645 || ( _preserved_mark_stack->is_empty()
8646 && _preserved_oop_stack->is_empty()));
8647 }
8648 #endif
8650 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8651 {
8652 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8653 CMSAdaptiveSizePolicy* size_policy =
8654 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8655 assert(size_policy->is_gc_cms_adaptive_size_policy(),
8656 "Wrong type for size policy");
8657 return size_policy;
8658 }
8660 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
8661 size_t desired_promo_size) {
8662 if (cur_promo_size < desired_promo_size) {
8663 size_t expand_bytes = desired_promo_size - cur_promo_size;
8664 if (PrintAdaptiveSizePolicy && Verbose) {
8665 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8666 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
8667 expand_bytes);
8668 }
8669 expand(expand_bytes,
8670 MinHeapDeltaBytes,
8671 CMSExpansionCause::_adaptive_size_policy);
8672 } else if (desired_promo_size < cur_promo_size) {
8673 size_t shrink_bytes = cur_promo_size - desired_promo_size;
8674 if (PrintAdaptiveSizePolicy && Verbose) {
8675 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8676 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
8677 shrink_bytes);
8678 }
8679 shrink(shrink_bytes);
8680 }
8681 }
8683 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
8684 GenCollectedHeap* gch = GenCollectedHeap::heap();
8685 CMSGCAdaptivePolicyCounters* counters =
8686 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
8687 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
8688 "Wrong kind of counters");
8689 return counters;
8690 }
8693 void ASConcurrentMarkSweepGeneration::update_counters() {
8694 if (UsePerfData) {
8695 _space_counters->update_all();
8696 _gen_counters->update_all();
8697 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8698 GenCollectedHeap* gch = GenCollectedHeap::heap();
8699 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8700 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8701 "Wrong gc statistics type");
8702 counters->update_counters(gc_stats_l);
8703 }
8704 }
8706 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
8707 if (UsePerfData) {
8708 _space_counters->update_used(used);
8709 _space_counters->update_capacity();
8710 _gen_counters->update_all();
8712 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8713 GenCollectedHeap* gch = GenCollectedHeap::heap();
8714 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8715 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8716 "Wrong gc statistics type");
8717 counters->update_counters(gc_stats_l);
8718 }
8719 }
8721 // The desired expansion delta is computed so that:
8722 // . desired free percentage or greater is used
8723 void ASConcurrentMarkSweepGeneration::compute_new_size() {
8724 assert_locked_or_safepoint(Heap_lock);
8726 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8728 // If incremental collection failed, we just want to expand
8729 // to the limit.
8730 if (incremental_collection_failed()) {
8731 clear_incremental_collection_failed();
8732 grow_to_reserved();
8733 return;
8734 }
8736 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
8738 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
8739 "Wrong type of heap");
8740 int prev_level = level() - 1;
8741 assert(prev_level >= 0, "The cms generation is the lowest generation");
8742 Generation* prev_gen = gch->get_gen(prev_level);
8743 assert(prev_gen->kind() == Generation::ASParNew,
8744 "Wrong type of young generation");
8745 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
8746 size_t cur_eden = younger_gen->eden()->capacity();
8747 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
8748 size_t cur_promo = free();
8749 size_policy->compute_tenured_generation_free_space(cur_promo,
8750 max_available(),
8751 cur_eden);
8752 resize(cur_promo, size_policy->promo_size());
8754 // Record the new size of the space in the cms generation
8755 // that is available for promotions. This is temporary.
8756 // It should be the desired promo size.
8757 size_policy->avg_cms_promo()->sample(free());
8758 size_policy->avg_old_live()->sample(used());
8760 if (UsePerfData) {
8761 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8762 counters->update_cms_capacity_counter(capacity());
8763 }
8764 }
8766 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
8767 assert_locked_or_safepoint(Heap_lock);
8768 assert_lock_strong(freelistLock());
8769 HeapWord* old_end = _cmsSpace->end();
8770 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
8771 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
8772 FreeChunk* chunk_at_end = find_chunk_at_end();
8773 if (chunk_at_end == NULL) {
8774 // No room to shrink
8775 if (PrintGCDetails && Verbose) {
8776 gclog_or_tty->print_cr("No room to shrink: old_end "
8777 PTR_FORMAT " unallocated_start " PTR_FORMAT
8778 " chunk_at_end " PTR_FORMAT,
8779 old_end, unallocated_start, chunk_at_end);
8780 }
8781 return;
8782 } else {
8784 // Find the chunk at the end of the space and determine
8785 // how much it can be shrunk.
8786 size_t shrinkable_size_in_bytes = chunk_at_end->size();
8787 size_t aligned_shrinkable_size_in_bytes =
8788 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
8789 assert(unallocated_start <= chunk_at_end->end(),
8790 "Inconsistent chunk at end of space");
8791 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
8792 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
8794 // Shrink the underlying space
8795 _virtual_space.shrink_by(bytes);
8796 if (PrintGCDetails && Verbose) {
8797 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
8798 " desired_bytes " SIZE_FORMAT
8799 " shrinkable_size_in_bytes " SIZE_FORMAT
8800 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
8801 " bytes " SIZE_FORMAT,
8802 desired_bytes, shrinkable_size_in_bytes,
8803 aligned_shrinkable_size_in_bytes, bytes);
8804 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
8805 " unallocated_start " SIZE_FORMAT,
8806 old_end, unallocated_start);
8807 }
8809 // If the space did shrink (shrinking is not guaranteed),
8810 // shrink the chunk at the end by the appropriate amount.
8811 if (((HeapWord*)_virtual_space.high()) < old_end) {
8812 size_t new_word_size =
8813 heap_word_size(_virtual_space.committed_size());
8815 // Have to remove the chunk from the dictionary because it is changing
8816 // size and might be someplace elsewhere in the dictionary.
8818 // Get the chunk at end, shrink it, and put it
8819 // back.
8820 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
8821 size_t word_size_change = word_size_before - new_word_size;
8822 size_t chunk_at_end_old_size = chunk_at_end->size();
8823 assert(chunk_at_end_old_size >= word_size_change,
8824 "Shrink is too large");
8825 chunk_at_end->setSize(chunk_at_end_old_size -
8826 word_size_change);
8827 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
8828 word_size_change);
8830 _cmsSpace->returnChunkToDictionary(chunk_at_end);
8832 MemRegion mr(_cmsSpace->bottom(), new_word_size);
8833 _bts->resize(new_word_size); // resize the block offset shared array
8834 Universe::heap()->barrier_set()->resize_covered_region(mr);
8835 _cmsSpace->assert_locked();
8836 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
8838 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
8840 // update the space and generation capacity counters
8841 if (UsePerfData) {
8842 _space_counters->update_capacity();
8843 _gen_counters->update_all();
8844 }
8846 if (Verbose && PrintGCDetails) {
8847 size_t new_mem_size = _virtual_space.committed_size();
8848 size_t old_mem_size = new_mem_size + bytes;
8849 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
8850 name(), old_mem_size/K, bytes/K, new_mem_size/K);
8851 }
8852 }
8854 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
8855 "Inconsistency at end of space");
8856 assert(chunk_at_end->end() == _cmsSpace->end(),
8857 "Shrinking is inconsistent");
8858 return;
8859 }
8860 }
8862 // Transfer some number of overflown objects to usual marking
8863 // stack. Return true if some objects were transferred.
8864 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8865 size_t num = MIN2((size_t)_mark_stack->capacity()/4,
8866 (size_t)ParGCDesiredObjsFromOverflowList);
8868 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8869 assert(_collector->overflow_list_is_empty() || res,
8870 "If list is not empty, we should have taken something");
8871 assert(!res || !_mark_stack->isEmpty(),
8872 "If we took something, it should now be on our stack");
8873 return res;
8874 }
8876 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8877 size_t res = _sp->block_size_no_stall(addr, _collector);
8878 assert(res != 0, "Should always be able to compute a size");
8879 if (_sp->block_is_obj(addr)) {
8880 if (_live_bit_map->isMarked(addr)) {
8881 // It can't have been dead in a previous cycle
8882 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8883 } else {
8884 _dead_bit_map->mark(addr); // mark the dead object
8885 }
8886 }
8887 return res;
8888 }