1 /*
2 * Copyright 2001-2007 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 # include "incls/_precompiled.incl"
26 # include "incls/_concurrentMarkSweepGeneration.cpp.incl"
27
28 // statics
29 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
30 bool CMSCollector::_full_gc_requested = false;
31
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
89
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 }
100
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 };
109
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 };
126
127
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 };
137
138 //////////////////////////////////////////////////////////////////
139 // Concurrent Mark-Sweep Generation /////////////////////////////
140 //////////////////////////////////////////////////////////////////
141
142 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
143
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;
150
151 // Constructor.
152 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
153 promo.setSpace(cfls);
154 }
155 };
156
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();
167
168 _direct_allocated_words = 0;
169 NOT_PRODUCT(
170 _numObjectsPromoted = 0;
171 _numWordsPromoted = 0;
172 _numObjectsAllocated = 0;
173 _numWordsAllocated = 0;
174 )
175
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;
185
186 _gc_stats = new CMSGCStats();
187
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(junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
194 "Offset of FreeChunk::_prev within FreeChunk must match"
195 " that of OopDesc::_klass within OopDesc");
196 )
197 if (ParallelGCThreads > 0) {
198 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
199 _par_gc_thread_states =
200 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
201 if (_par_gc_thread_states == NULL) {
202 vm_exit_during_initialization("Could not allocate par gc structs");
203 }
204 for (uint i = 0; i < ParallelGCThreads; i++) {
205 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
206 if (_par_gc_thread_states[i] == NULL) {
207 vm_exit_during_initialization("Could not allocate par gc structs");
208 }
209 }
210 } else {
211 _par_gc_thread_states = NULL;
212 }
213 _incremental_collection_failed = false;
214 // The "dilatation_factor" is the expansion that can occur on
215 // account of the fact that the minimum object size in the CMS
216 // generation may be larger than that in, say, a contiguous young
217 // generation.
218 // Ideally, in the calculation below, we'd compute the dilatation
219 // factor as: MinChunkSize/(promoting_gen's min object size)
220 // Since we do not have such a general query interface for the
221 // promoting generation, we'll instead just use the mimimum
222 // object size (which today is a header's worth of space);
223 // note that all arithmetic is in units of HeapWords.
224 assert(MinChunkSize >= oopDesc::header_size(), "just checking");
225 assert(_dilatation_factor >= 1.0, "from previous assert");
226 }
227
228
229 // The field "_initiating_occupancy" represents the occupancy percentage
230 // at which we trigger a new collection cycle. Unless explicitly specified
231 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
232 // is calculated by:
233 //
234 // Let "f" be MinHeapFreeRatio in
235 //
236 // _intiating_occupancy = 100-f +
237 // f * (CMSTrigger[Perm]Ratio/100)
238 // where CMSTrigger[Perm]Ratio is the argument "tr" below.
239 //
240 // That is, if we assume the heap is at its desired maximum occupancy at the
241 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
242 // space be allocated before initiating a new collection cycle.
243 //
244 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
245 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
246 if (io >= 0) {
247 _initiating_occupancy = (double)io / 100.0;
248 } else {
249 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
250 (double)(tr * MinHeapFreeRatio) / 100.0)
251 / 100.0;
252 }
253 }
254
255
256 void ConcurrentMarkSweepGeneration::ref_processor_init() {
257 assert(collector() != NULL, "no collector");
258 collector()->ref_processor_init();
259 }
260
261 void CMSCollector::ref_processor_init() {
262 if (_ref_processor == NULL) {
263 // Allocate and initialize a reference processor
264 _ref_processor = ReferenceProcessor::create_ref_processor(
265 _span, // span
266 _cmsGen->refs_discovery_is_atomic(), // atomic_discovery
267 _cmsGen->refs_discovery_is_mt(), // mt_discovery
268 &_is_alive_closure,
269 ParallelGCThreads,
270 ParallelRefProcEnabled);
271 // Initialize the _ref_processor field of CMSGen
272 _cmsGen->set_ref_processor(_ref_processor);
273
274 // Allocate a dummy ref processor for perm gen.
275 ReferenceProcessor* rp2 = new ReferenceProcessor();
276 if (rp2 == NULL) {
277 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
278 }
279 _permGen->set_ref_processor(rp2);
280 }
281 }
282
283 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
284 GenCollectedHeap* gch = GenCollectedHeap::heap();
285 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
286 "Wrong type of heap");
287 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
288 gch->gen_policy()->size_policy();
289 assert(sp->is_gc_cms_adaptive_size_policy(),
290 "Wrong type of size policy");
291 return sp;
292 }
293
294 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
295 CMSGCAdaptivePolicyCounters* results =
296 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
297 assert(
298 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
299 "Wrong gc policy counter kind");
300 return results;
301 }
302
303
304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
305
306 const char* gen_name = "old";
307
308 // Generation Counters - generation 1, 1 subspace
309 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
310
311 _space_counters = new GSpaceCounters(gen_name, 0,
312 _virtual_space.reserved_size(),
313 this, _gen_counters);
314 }
315
316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
317 _cms_gen(cms_gen)
318 {
319 assert(alpha <= 100, "bad value");
320 _saved_alpha = alpha;
321
322 // Initialize the alphas to the bootstrap value of 100.
323 _gc0_alpha = _cms_alpha = 100;
324
325 _cms_begin_time.update();
326 _cms_end_time.update();
327
328 _gc0_duration = 0.0;
329 _gc0_period = 0.0;
330 _gc0_promoted = 0;
331
332 _cms_duration = 0.0;
333 _cms_period = 0.0;
334 _cms_allocated = 0;
335
336 _cms_used_at_gc0_begin = 0;
337 _cms_used_at_gc0_end = 0;
338 _allow_duty_cycle_reduction = false;
339 _valid_bits = 0;
340 _icms_duty_cycle = CMSIncrementalDutyCycle;
341 }
342
343 // If promotion failure handling is on use
344 // the padded average size of the promotion for each
345 // young generation collection.
346 double CMSStats::time_until_cms_gen_full() const {
347 size_t cms_free = _cms_gen->cmsSpace()->free();
348 GenCollectedHeap* gch = GenCollectedHeap::heap();
349 size_t expected_promotion = gch->get_gen(0)->capacity();
350 if (HandlePromotionFailure) {
351 expected_promotion = MIN2(
352 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(),
353 expected_promotion);
354 }
355 if (cms_free > expected_promotion) {
356 // Start a cms collection if there isn't enough space to promote
357 // for the next minor collection. Use the padded average as
358 // a safety factor.
359 cms_free -= expected_promotion;
360
361 // Adjust by the safety factor.
362 double cms_free_dbl = (double)cms_free;
363 cms_free_dbl = cms_free_dbl * (100.0 - CMSIncrementalSafetyFactor) / 100.0;
364
365 if (PrintGCDetails && Verbose) {
366 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
367 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
368 cms_free, expected_promotion);
369 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
370 cms_free_dbl, cms_consumption_rate() + 1.0);
371 }
372 // Add 1 in case the consumption rate goes to zero.
373 return cms_free_dbl / (cms_consumption_rate() + 1.0);
374 }
375 return 0.0;
376 }
377
378 // Compare the duration of the cms collection to the
379 // time remaining before the cms generation is empty.
380 // Note that the time from the start of the cms collection
381 // to the start of the cms sweep (less than the total
382 // duration of the cms collection) can be used. This
383 // has been tried and some applications experienced
384 // promotion failures early in execution. This was
385 // possibly because the averages were not accurate
386 // enough at the beginning.
387 double CMSStats::time_until_cms_start() const {
388 // We add "gc0_period" to the "work" calculation
389 // below because this query is done (mostly) at the
390 // end of a scavenge, so we need to conservatively
391 // account for that much possible delay
392 // in the query so as to avoid concurrent mode failures
393 // due to starting the collection just a wee bit too
394 // late.
395 double work = cms_duration() + gc0_period();
396 double deadline = time_until_cms_gen_full();
397 if (work > deadline) {
398 if (Verbose && PrintGCDetails) {
399 gclog_or_tty->print(
400 " CMSCollector: collect because of anticipated promotion "
401 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
402 gc0_period(), time_until_cms_gen_full());
403 }
404 return 0.0;
405 }
406 return work - deadline;
407 }
408
409 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
410 // amount of change to prevent wild oscillation.
411 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
412 unsigned int new_duty_cycle) {
413 assert(old_duty_cycle <= 100, "bad input value");
414 assert(new_duty_cycle <= 100, "bad input value");
415
416 // Note: use subtraction with caution since it may underflow (values are
417 // unsigned). Addition is safe since we're in the range 0-100.
418 unsigned int damped_duty_cycle = new_duty_cycle;
419 if (new_duty_cycle < old_duty_cycle) {
420 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
421 if (new_duty_cycle + largest_delta < old_duty_cycle) {
422 damped_duty_cycle = old_duty_cycle - largest_delta;
423 }
424 } else if (new_duty_cycle > old_duty_cycle) {
425 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
426 if (new_duty_cycle > old_duty_cycle + largest_delta) {
427 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
428 }
429 }
430 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
431
432 if (CMSTraceIncrementalPacing) {
433 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
434 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
435 }
436 return damped_duty_cycle;
437 }
438
439 unsigned int CMSStats::icms_update_duty_cycle_impl() {
440 assert(CMSIncrementalPacing && valid(),
441 "should be handled in icms_update_duty_cycle()");
442
443 double cms_time_so_far = cms_timer().seconds();
444 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
445 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
446
447 // Avoid division by 0.
448 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
449 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
450
451 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
452 if (new_duty_cycle > _icms_duty_cycle) {
453 // Avoid very small duty cycles (1 or 2); 0 is allowed.
454 if (new_duty_cycle > 2) {
455 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
456 new_duty_cycle);
457 }
458 } else if (_allow_duty_cycle_reduction) {
459 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
460 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
461 // Respect the minimum duty cycle.
462 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
463 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
464 }
465
466 if (PrintGCDetails || CMSTraceIncrementalPacing) {
467 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
468 }
469
470 _allow_duty_cycle_reduction = false;
471 return _icms_duty_cycle;
472 }
473
474 #ifndef PRODUCT
475 void CMSStats::print_on(outputStream *st) const {
476 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
477 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
478 gc0_duration(), gc0_period(), gc0_promoted());
479 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
480 cms_duration(), cms_duration_per_mb(),
481 cms_period(), cms_allocated());
482 st->print(",cms_since_beg=%g,cms_since_end=%g",
483 cms_time_since_begin(), cms_time_since_end());
484 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
485 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
486 if (CMSIncrementalMode) {
487 st->print(",dc=%d", icms_duty_cycle());
488 }
489
490 if (valid()) {
491 st->print(",promo_rate=%g,cms_alloc_rate=%g",
492 promotion_rate(), cms_allocation_rate());
493 st->print(",cms_consumption_rate=%g,time_until_full=%g",
494 cms_consumption_rate(), time_until_cms_gen_full());
495 }
496 st->print(" ");
497 }
498 #endif // #ifndef PRODUCT
499
500 CMSCollector::CollectorState CMSCollector::_collectorState =
501 CMSCollector::Idling;
502 bool CMSCollector::_foregroundGCIsActive = false;
503 bool CMSCollector::_foregroundGCShouldWait = false;
504
505 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
506 ConcurrentMarkSweepGeneration* permGen,
507 CardTableRS* ct,
508 ConcurrentMarkSweepPolicy* cp):
509 _cmsGen(cmsGen),
510 _permGen(permGen),
511 _ct(ct),
512 _ref_processor(NULL), // will be set later
513 _conc_workers(NULL), // may be set later
514 _abort_preclean(false),
515 _start_sampling(false),
516 _between_prologue_and_epilogue(false),
517 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
518 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
519 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
520 -1 /* lock-free */, "No_lock" /* dummy */),
521 _modUnionClosure(&_modUnionTable),
522 _modUnionClosurePar(&_modUnionTable),
523 // Adjust my span to cover old (cms) gen and perm gen
524 _span(cmsGen->reserved()._union(permGen->reserved())),
525 // Construct the is_alive_closure with _span & markBitMap
526 _is_alive_closure(_span, &_markBitMap),
527 _restart_addr(NULL),
528 _overflow_list(NULL),
529 _preserved_oop_stack(NULL),
530 _preserved_mark_stack(NULL),
531 _stats(cmsGen),
532 _eden_chunk_array(NULL), // may be set in ctor body
533 _eden_chunk_capacity(0), // -- ditto --
534 _eden_chunk_index(0), // -- ditto --
535 _survivor_plab_array(NULL), // -- ditto --
536 _survivor_chunk_array(NULL), // -- ditto --
537 _survivor_chunk_capacity(0), // -- ditto --
538 _survivor_chunk_index(0), // -- ditto --
539 _ser_pmc_preclean_ovflw(0),
540 _ser_pmc_remark_ovflw(0),
541 _par_pmc_remark_ovflw(0),
542 _ser_kac_ovflw(0),
543 _par_kac_ovflw(0),
544 #ifndef PRODUCT
545 _num_par_pushes(0),
546 #endif
547 _collection_count_start(0),
548 _verifying(false),
549 _icms_start_limit(NULL),
550 _icms_stop_limit(NULL),
551 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
552 _completed_initialization(false),
553 _collector_policy(cp),
554 _should_unload_classes(false),
555 _concurrent_cycles_since_last_unload(0),
556 _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
557 {
558 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
559 ExplicitGCInvokesConcurrent = true;
560 }
561 // Now expand the span and allocate the collection support structures
562 // (MUT, marking bit map etc.) to cover both generations subject to
563 // collection.
564
565 // First check that _permGen is adjacent to _cmsGen and above it.
566 assert( _cmsGen->reserved().word_size() > 0
567 && _permGen->reserved().word_size() > 0,
568 "generations should not be of zero size");
569 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
570 "_cmsGen and _permGen should not overlap");
571 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
572 "_cmsGen->end() different from _permGen->start()");
573
574 // For use by dirty card to oop closures.
575 _cmsGen->cmsSpace()->set_collector(this);
576 _permGen->cmsSpace()->set_collector(this);
577
578 // Allocate MUT and marking bit map
579 {
580 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
581 if (!_markBitMap.allocate(_span)) {
582 warning("Failed to allocate CMS Bit Map");
583 return;
584 }
585 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
586 }
587 {
588 _modUnionTable.allocate(_span);
589 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
590 }
591
592 if (!_markStack.allocate(CMSMarkStackSize)) {
593 warning("Failed to allocate CMS Marking Stack");
594 return;
595 }
596 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
597 warning("Failed to allocate CMS Revisit Stack");
598 return;
599 }
600
601 // Support for multi-threaded concurrent phases
602 if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) {
603 if (FLAG_IS_DEFAULT(ParallelCMSThreads)) {
604 // just for now
605 FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4);
606 }
607 if (ParallelCMSThreads > 1) {
608 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
609 ParallelCMSThreads, true);
610 if (_conc_workers == NULL) {
611 warning("GC/CMS: _conc_workers allocation failure: "
612 "forcing -CMSConcurrentMTEnabled");
613 CMSConcurrentMTEnabled = false;
614 }
615 } else {
616 CMSConcurrentMTEnabled = false;
617 }
618 }
619 if (!CMSConcurrentMTEnabled) {
620 ParallelCMSThreads = 0;
621 } else {
622 // Turn off CMSCleanOnEnter optimization temporarily for
623 // the MT case where it's not fixed yet; see 6178663.
624 CMSCleanOnEnter = false;
625 }
626 assert((_conc_workers != NULL) == (ParallelCMSThreads > 1),
627 "Inconsistency");
628
629 // Parallel task queues; these are shared for the
630 // concurrent and stop-world phases of CMS, but
631 // are not shared with parallel scavenge (ParNew).
632 {
633 uint i;
634 uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads);
635
636 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
637 || ParallelRefProcEnabled)
638 && num_queues > 0) {
639 _task_queues = new OopTaskQueueSet(num_queues);
640 if (_task_queues == NULL) {
641 warning("task_queues allocation failure.");
642 return;
643 }
644 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
645 if (_hash_seed == NULL) {
646 warning("_hash_seed array allocation failure");
647 return;
648 }
649
650 // XXX use a global constant instead of 64!
651 typedef struct OopTaskQueuePadded {
652 OopTaskQueue work_queue;
653 char pad[64 - sizeof(OopTaskQueue)]; // prevent false sharing
654 } OopTaskQueuePadded;
655
656 for (i = 0; i < num_queues; i++) {
657 OopTaskQueuePadded *q_padded = new OopTaskQueuePadded();
658 if (q_padded == NULL) {
659 warning("work_queue allocation failure.");
660 return;
661 }
662 _task_queues->register_queue(i, &q_padded->work_queue);
663 }
664 for (i = 0; i < num_queues; i++) {
665 _task_queues->queue(i)->initialize();
666 _hash_seed[i] = 17; // copied from ParNew
667 }
668 }
669 }
670
671 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
672 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
673
674 // Clip CMSBootstrapOccupancy between 0 and 100.
675 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
676 /(double)100;
677
678 _full_gcs_since_conc_gc = 0;
679
680 // Now tell CMS generations the identity of their collector
681 ConcurrentMarkSweepGeneration::set_collector(this);
682
683 // Create & start a CMS thread for this CMS collector
684 _cmsThread = ConcurrentMarkSweepThread::start(this);
685 assert(cmsThread() != NULL, "CMS Thread should have been created");
686 assert(cmsThread()->collector() == this,
687 "CMS Thread should refer to this gen");
688 assert(CGC_lock != NULL, "Where's the CGC_lock?");
689
690 // Support for parallelizing young gen rescan
691 GenCollectedHeap* gch = GenCollectedHeap::heap();
692 _young_gen = gch->prev_gen(_cmsGen);
693 if (gch->supports_inline_contig_alloc()) {
694 _top_addr = gch->top_addr();
695 _end_addr = gch->end_addr();
696 assert(_young_gen != NULL, "no _young_gen");
697 _eden_chunk_index = 0;
698 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
699 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
700 if (_eden_chunk_array == NULL) {
701 _eden_chunk_capacity = 0;
702 warning("GC/CMS: _eden_chunk_array allocation failure");
703 }
704 }
705 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
706
707 // Support for parallelizing survivor space rescan
708 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
709 size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize);
710 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
711 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
712 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
713 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
714 || _cursor == NULL) {
715 warning("Failed to allocate survivor plab/chunk array");
716 if (_survivor_plab_array != NULL) {
717 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
718 _survivor_plab_array = NULL;
719 }
720 if (_survivor_chunk_array != NULL) {
721 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
722 _survivor_chunk_array = NULL;
723 }
724 if (_cursor != NULL) {
725 FREE_C_HEAP_ARRAY(size_t, _cursor);
726 _cursor = NULL;
727 }
728 } else {
729 _survivor_chunk_capacity = 2*max_plab_samples;
730 for (uint i = 0; i < ParallelGCThreads; i++) {
731 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
732 if (vec == NULL) {
733 warning("Failed to allocate survivor plab array");
734 for (int j = i; j > 0; j--) {
735 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
736 }
737 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
738 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
739 _survivor_plab_array = NULL;
740 _survivor_chunk_array = NULL;
741 _survivor_chunk_capacity = 0;
742 break;
743 } else {
744 ChunkArray* cur =
745 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
746 max_plab_samples);
747 assert(cur->end() == 0, "Should be 0");
748 assert(cur->array() == vec, "Should be vec");
749 assert(cur->capacity() == max_plab_samples, "Error");
750 }
751 }
752 }
753 }
754 assert( ( _survivor_plab_array != NULL
755 && _survivor_chunk_array != NULL)
756 || ( _survivor_chunk_capacity == 0
757 && _survivor_chunk_index == 0),
758 "Error");
759
760 // Choose what strong roots should be scanned depending on verification options
761 // and perm gen collection mode.
762 if (!CMSClassUnloadingEnabled) {
763 // If class unloading is disabled we want to include all classes into the root set.
764 add_root_scanning_option(SharedHeap::SO_AllClasses);
765 } else {
766 add_root_scanning_option(SharedHeap::SO_SystemClasses);
767 }
768
769 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
770 _gc_counters = new CollectorCounters("CMS", 1);
771 _completed_initialization = true;
772 _sweep_timer.start(); // start of time
773 }
774
775 const char* ConcurrentMarkSweepGeneration::name() const {
776 return "concurrent mark-sweep generation";
777 }
778 void ConcurrentMarkSweepGeneration::update_counters() {
779 if (UsePerfData) {
780 _space_counters->update_all();
781 _gen_counters->update_all();
782 }
783 }
784
785 // this is an optimized version of update_counters(). it takes the
786 // used value as a parameter rather than computing it.
787 //
788 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
789 if (UsePerfData) {
790 _space_counters->update_used(used);
791 _space_counters->update_capacity();
792 _gen_counters->update_all();
793 }
794 }
795
796 void ConcurrentMarkSweepGeneration::print() const {
797 Generation::print();
798 cmsSpace()->print();
799 }
800
801 #ifndef PRODUCT
802 void ConcurrentMarkSweepGeneration::print_statistics() {
803 cmsSpace()->printFLCensus(0);
804 }
805 #endif
806
807 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
808 GenCollectedHeap* gch = GenCollectedHeap::heap();
809 if (PrintGCDetails) {
810 if (Verbose) {
811 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
812 level(), short_name(), s, used(), capacity());
813 } else {
814 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
815 level(), short_name(), s, used() / K, capacity() / K);
816 }
817 }
818 if (Verbose) {
819 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
820 gch->used(), gch->capacity());
821 } else {
822 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
823 gch->used() / K, gch->capacity() / K);
824 }
825 }
826
827 size_t
828 ConcurrentMarkSweepGeneration::contiguous_available() const {
829 // dld proposes an improvement in precision here. If the committed
830 // part of the space ends in a free block we should add that to
831 // uncommitted size in the calculation below. Will make this
832 // change later, staying with the approximation below for the
833 // time being. -- ysr.
834 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
835 }
836
837 size_t
838 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
839 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
840 }
841
842 size_t ConcurrentMarkSweepGeneration::max_available() const {
843 return free() + _virtual_space.uncommitted_size();
844 }
845
846 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(
847 size_t max_promotion_in_bytes,
848 bool younger_handles_promotion_failure) const {
849
850 // This is the most conservative test. Full promotion is
851 // guaranteed if this is used. The multiplicative factor is to
852 // account for the worst case "dilatation".
853 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes;
854 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t
855 adjusted_max_promo_bytes = (double)max_uintx;
856 }
857 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes);
858
859 if (younger_handles_promotion_failure && !result) {
860 // Full promotion is not guaranteed because fragmentation
861 // of the cms generation can prevent the full promotion.
862 result = (max_available() >= (size_t)adjusted_max_promo_bytes);
863
864 if (!result) {
865 // With promotion failure handling the test for the ability
866 // to support the promotion does not have to be guaranteed.
867 // Use an average of the amount promoted.
868 result = max_available() >= (size_t)
869 gc_stats()->avg_promoted()->padded_average();
870 if (PrintGC && Verbose && result) {
871 gclog_or_tty->print_cr(
872 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
873 " max_available: " SIZE_FORMAT
874 " avg_promoted: " SIZE_FORMAT,
875 max_available(), (size_t)
876 gc_stats()->avg_promoted()->padded_average());
877 }
878 } else {
879 if (PrintGC && Verbose) {
880 gclog_or_tty->print_cr(
881 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
882 " max_available: " SIZE_FORMAT
883 " adj_max_promo_bytes: " SIZE_FORMAT,
884 max_available(), (size_t)adjusted_max_promo_bytes);
885 }
886 }
887 } else {
888 if (PrintGC && Verbose) {
889 gclog_or_tty->print_cr(
890 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
891 " contiguous_available: " SIZE_FORMAT
892 " adj_max_promo_bytes: " SIZE_FORMAT,
893 max_contiguous_available(), (size_t)adjusted_max_promo_bytes);
894 }
895 }
896 return result;
897 }
898
899 CompactibleSpace*
900 ConcurrentMarkSweepGeneration::first_compaction_space() const {
901 return _cmsSpace;
902 }
903
904 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
905 // Clear the promotion information. These pointers can be adjusted
906 // along with all the other pointers into the heap but
907 // compaction is expected to be a rare event with
908 // a heap using cms so don't do it without seeing the need.
909 if (ParallelGCThreads > 0) {
910 for (uint i = 0; i < ParallelGCThreads; i++) {
911 _par_gc_thread_states[i]->promo.reset();
912 }
913 }
914 }
915
916 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
917 blk->do_space(_cmsSpace);
918 }
919
920 void ConcurrentMarkSweepGeneration::compute_new_size() {
921 assert_locked_or_safepoint(Heap_lock);
922
923 // If incremental collection failed, we just want to expand
924 // to the limit.
925 if (incremental_collection_failed()) {
926 clear_incremental_collection_failed();
927 grow_to_reserved();
928 return;
929 }
930
931 size_t expand_bytes = 0;
932 double free_percentage = ((double) free()) / capacity();
933 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
934 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
935
936 // compute expansion delta needed for reaching desired free percentage
937 if (free_percentage < desired_free_percentage) {
938 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
939 assert(desired_capacity >= capacity(), "invalid expansion size");
940 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
941 }
942 if (expand_bytes > 0) {
943 if (PrintGCDetails && Verbose) {
944 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
945 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
946 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
947 gclog_or_tty->print_cr(" Desired free fraction %f",
948 desired_free_percentage);
949 gclog_or_tty->print_cr(" Maximum free fraction %f",
950 maximum_free_percentage);
951 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
952 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
953 desired_capacity/1000);
954 int prev_level = level() - 1;
955 if (prev_level >= 0) {
956 size_t prev_size = 0;
957 GenCollectedHeap* gch = GenCollectedHeap::heap();
958 Generation* prev_gen = gch->_gens[prev_level];
959 prev_size = prev_gen->capacity();
960 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
961 prev_size/1000);
962 }
963 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
964 unsafe_max_alloc_nogc()/1000);
965 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
966 contiguous_available()/1000);
967 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
968 expand_bytes);
969 }
970 // safe if expansion fails
971 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
972 if (PrintGCDetails && Verbose) {
973 gclog_or_tty->print_cr(" Expanded free fraction %f",
974 ((double) free()) / capacity());
975 }
976 }
977 }
978
979 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
980 return cmsSpace()->freelistLock();
981 }
982
983 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
984 bool tlab) {
985 CMSSynchronousYieldRequest yr;
986 MutexLockerEx x(freelistLock(),
987 Mutex::_no_safepoint_check_flag);
988 return have_lock_and_allocate(size, tlab);
989 }
990
991 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
992 bool tlab) {
993 assert_lock_strong(freelistLock());
994 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
995 HeapWord* res = cmsSpace()->allocate(adjustedSize);
996 // Allocate the object live (grey) if the background collector has
997 // started marking. This is necessary because the marker may
998 // have passed this address and consequently this object will
999 // not otherwise be greyed and would be incorrectly swept up.
1000 // Note that if this object contains references, the writing
1001 // of those references will dirty the card containing this object
1002 // allowing the object to be blackened (and its references scanned)
1003 // either during a preclean phase or at the final checkpoint.
1004 if (res != NULL) {
1005 collector()->direct_allocated(res, adjustedSize);
1006 _direct_allocated_words += adjustedSize;
1007 // allocation counters
1008 NOT_PRODUCT(
1009 _numObjectsAllocated++;
1010 _numWordsAllocated += (int)adjustedSize;
1011 )
1012 }
1013 return res;
1014 }
1015
1016 // In the case of direct allocation by mutators in a generation that
1017 // is being concurrently collected, the object must be allocated
1018 // live (grey) if the background collector has started marking.
1019 // This is necessary because the marker may
1020 // have passed this address and consequently this object will
1021 // not otherwise be greyed and would be incorrectly swept up.
1022 // Note that if this object contains references, the writing
1023 // of those references will dirty the card containing this object
1024 // allowing the object to be blackened (and its references scanned)
1025 // either during a preclean phase or at the final checkpoint.
1026 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1027 assert(_markBitMap.covers(start, size), "Out of bounds");
1028 if (_collectorState >= Marking) {
1029 MutexLockerEx y(_markBitMap.lock(),
1030 Mutex::_no_safepoint_check_flag);
1031 // [see comments preceding SweepClosure::do_blk() below for details]
1032 // 1. need to mark the object as live so it isn't collected
1033 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1034 // 3. need to mark the end of the object so sweeper can skip over it
1035 // if it's uninitialized when the sweeper reaches it.
1036 _markBitMap.mark(start); // object is live
1037 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1038 _markBitMap.mark(start + size - 1);
1039 // mark end of object
1040 }
1041 // check that oop looks uninitialized
1042 assert(oop(start)->klass() == NULL, "_klass should be NULL");
1043 }
1044
1045 void CMSCollector::promoted(bool par, HeapWord* start,
1046 bool is_obj_array, size_t obj_size) {
1047 assert(_markBitMap.covers(start), "Out of bounds");
1048 // See comment in direct_allocated() about when objects should
1049 // be allocated live.
1050 if (_collectorState >= Marking) {
1051 // we already hold the marking bit map lock, taken in
1052 // the prologue
1053 if (par) {
1054 _markBitMap.par_mark(start);
1055 } else {
1056 _markBitMap.mark(start);
1057 }
1058 // We don't need to mark the object as uninitialized (as
1059 // in direct_allocated above) because this is being done with the
1060 // world stopped and the object will be initialized by the
1061 // time the sweeper gets to look at it.
1062 assert(SafepointSynchronize::is_at_safepoint(),
1063 "expect promotion only at safepoints");
1064
1065 if (_collectorState < Sweeping) {
1066 // Mark the appropriate cards in the modUnionTable, so that
1067 // this object gets scanned before the sweep. If this is
1068 // not done, CMS generation references in the object might
1069 // not get marked.
1070 // For the case of arrays, which are otherwise precisely
1071 // marked, we need to dirty the entire array, not just its head.
1072 if (is_obj_array) {
1073 // The [par_]mark_range() method expects mr.end() below to
1074 // be aligned to the granularity of a bit's representation
1075 // in the heap. In the case of the MUT below, that's a
1076 // card size.
1077 MemRegion mr(start,
1078 (HeapWord*)round_to((intptr_t)(start + obj_size),
1079 CardTableModRefBS::card_size /* bytes */));
1080 if (par) {
1081 _modUnionTable.par_mark_range(mr);
1082 } else {
1083 _modUnionTable.mark_range(mr);
1084 }
1085 } else { // not an obj array; we can just mark the head
1086 if (par) {
1087 _modUnionTable.par_mark(start);
1088 } else {
1089 _modUnionTable.mark(start);
1090 }
1091 }
1092 }
1093 }
1094 }
1095
1096 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1097 {
1098 size_t delta = pointer_delta(addr, space->bottom());
1099 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1100 }
1101
1102 void CMSCollector::icms_update_allocation_limits()
1103 {
1104 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1105 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1106
1107 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1108 if (CMSTraceIncrementalPacing) {
1109 stats().print();
1110 }
1111
1112 assert(duty_cycle <= 100, "invalid duty cycle");
1113 if (duty_cycle != 0) {
1114 // The duty_cycle is a percentage between 0 and 100; convert to words and
1115 // then compute the offset from the endpoints of the space.
1116 size_t free_words = eden->free() / HeapWordSize;
1117 double free_words_dbl = (double)free_words;
1118 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1119 size_t offset_words = (free_words - duty_cycle_words) / 2;
1120
1121 _icms_start_limit = eden->top() + offset_words;
1122 _icms_stop_limit = eden->end() - offset_words;
1123
1124 // The limits may be adjusted (shifted to the right) by
1125 // CMSIncrementalOffset, to allow the application more mutator time after a
1126 // young gen gc (when all mutators were stopped) and before CMS starts and
1127 // takes away one or more cpus.
1128 if (CMSIncrementalOffset != 0) {
1129 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1130 size_t adjustment = (size_t)adjustment_dbl;
1131 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1132 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1133 _icms_start_limit += adjustment;
1134 _icms_stop_limit = tmp_stop;
1135 }
1136 }
1137 }
1138 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1139 _icms_start_limit = _icms_stop_limit = eden->end();
1140 }
1141
1142 // Install the new start limit.
1143 eden->set_soft_end(_icms_start_limit);
1144
1145 if (CMSTraceIncrementalMode) {
1146 gclog_or_tty->print(" icms alloc limits: "
1147 PTR_FORMAT "," PTR_FORMAT
1148 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1149 _icms_start_limit, _icms_stop_limit,
1150 percent_of_space(eden, _icms_start_limit),
1151 percent_of_space(eden, _icms_stop_limit));
1152 if (Verbose) {
1153 gclog_or_tty->print("eden: ");
1154 eden->print_on(gclog_or_tty);
1155 }
1156 }
1157 }
1158
1159 // Any changes here should try to maintain the invariant
1160 // that if this method is called with _icms_start_limit
1161 // and _icms_stop_limit both NULL, then it should return NULL
1162 // and not notify the icms thread.
1163 HeapWord*
1164 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1165 size_t word_size)
1166 {
1167 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1168 // nop.
1169 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1170 if (top <= _icms_start_limit) {
1171 if (CMSTraceIncrementalMode) {
1172 space->print_on(gclog_or_tty);
1173 gclog_or_tty->stamp();
1174 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1175 ", new limit=" PTR_FORMAT
1176 " (" SIZE_FORMAT "%%)",
1177 top, _icms_stop_limit,
1178 percent_of_space(space, _icms_stop_limit));
1179 }
1180 ConcurrentMarkSweepThread::start_icms();
1181 assert(top < _icms_stop_limit, "Tautology");
1182 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1183 return _icms_stop_limit;
1184 }
1185
1186 // The allocation will cross both the _start and _stop limits, so do the
1187 // stop notification also and return end().
1188 if (CMSTraceIncrementalMode) {
1189 space->print_on(gclog_or_tty);
1190 gclog_or_tty->stamp();
1191 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1192 ", new limit=" PTR_FORMAT
1193 " (" SIZE_FORMAT "%%)",
1194 top, space->end(),
1195 percent_of_space(space, space->end()));
1196 }
1197 ConcurrentMarkSweepThread::stop_icms();
1198 return space->end();
1199 }
1200
1201 if (top <= _icms_stop_limit) {
1202 if (CMSTraceIncrementalMode) {
1203 space->print_on(gclog_or_tty);
1204 gclog_or_tty->stamp();
1205 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1206 ", new limit=" PTR_FORMAT
1207 " (" SIZE_FORMAT "%%)",
1208 top, space->end(),
1209 percent_of_space(space, space->end()));
1210 }
1211 ConcurrentMarkSweepThread::stop_icms();
1212 return space->end();
1213 }
1214
1215 if (CMSTraceIncrementalMode) {
1216 space->print_on(gclog_or_tty);
1217 gclog_or_tty->stamp();
1218 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1219 ", new limit=" PTR_FORMAT,
1220 top, NULL);
1221 }
1222 }
1223
1224 return NULL;
1225 }
1226
1227 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1228 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1229 // allocate, copy and if necessary update promoinfo --
1230 // delegate to underlying space.
1231 assert_lock_strong(freelistLock());
1232
1233 #ifndef PRODUCT
1234 if (Universe::heap()->promotion_should_fail()) {
1235 return NULL;
1236 }
1237 #endif // #ifndef PRODUCT
1238
1239 oop res = _cmsSpace->promote(obj, obj_size);
1240 if (res == NULL) {
1241 // expand and retry
1242 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1243 expand(s*HeapWordSize, MinHeapDeltaBytes,
1244 CMSExpansionCause::_satisfy_promotion);
1245 // Since there's currently no next generation, we don't try to promote
1246 // into a more senior generation.
1247 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1248 "is made to pass on a possibly failing "
1249 "promotion to next generation");
1250 res = _cmsSpace->promote(obj, obj_size);
1251 }
1252 if (res != NULL) {
1253 // See comment in allocate() about when objects should
1254 // be allocated live.
1255 assert(obj->is_oop(), "Will dereference klass pointer below");
1256 collector()->promoted(false, // Not parallel
1257 (HeapWord*)res, obj->is_objArray(), obj_size);
1258 // promotion counters
1259 NOT_PRODUCT(
1260 _numObjectsPromoted++;
1261 _numWordsPromoted +=
1262 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1263 )
1264 }
1265 return res;
1266 }
1267
1268
1269 HeapWord*
1270 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1271 HeapWord* top,
1272 size_t word_sz)
1273 {
1274 return collector()->allocation_limit_reached(space, top, word_sz);
1275 }
1276
1277 // Things to support parallel young-gen collection.
1278 oop
1279 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1280 oop old, markOop m,
1281 size_t word_sz) {
1282 #ifndef PRODUCT
1283 if (Universe::heap()->promotion_should_fail()) {
1284 return NULL;
1285 }
1286 #endif // #ifndef PRODUCT
1287
1288 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1289 PromotionInfo* promoInfo = &ps->promo;
1290 // if we are tracking promotions, then first ensure space for
1291 // promotion (including spooling space for saving header if necessary).
1292 // then allocate and copy, then track promoted info if needed.
1293 // When tracking (see PromotionInfo::track()), the mark word may
1294 // be displaced and in this case restoration of the mark word
1295 // occurs in the (oop_since_save_marks_)iterate phase.
1296 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1297 // Out of space for allocating spooling buffers;
1298 // try expanding and allocating spooling buffers.
1299 if (!expand_and_ensure_spooling_space(promoInfo)) {
1300 return NULL;
1301 }
1302 }
1303 assert(promoInfo->has_spooling_space(), "Control point invariant");
1304 HeapWord* obj_ptr = ps->lab.alloc(word_sz);
1305 if (obj_ptr == NULL) {
1306 obj_ptr = expand_and_par_lab_allocate(ps, word_sz);
1307 if (obj_ptr == NULL) {
1308 return NULL;
1309 }
1310 }
1311 oop obj = oop(obj_ptr);
1312 assert(obj->klass() == NULL, "Object should be uninitialized here.");
1313 // Otherwise, copy the object. Here we must be careful to insert the
1314 // klass pointer last, since this marks the block as an allocated object.
1315 HeapWord* old_ptr = (HeapWord*)old;
1316 if (word_sz > (size_t)oopDesc::header_size()) {
1317 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1318 obj_ptr + oopDesc::header_size(),
1319 word_sz - oopDesc::header_size());
1320 }
1321 // Restore the mark word copied above.
1322 obj->set_mark(m);
1323 // Now we can track the promoted object, if necessary. We take care
1324 // To delay the transition from uninitialized to full object
1325 // (i.e., insertion of klass pointer) until after, so that it
1326 // atomically becomes a promoted object.
1327 if (promoInfo->tracking()) {
1328 promoInfo->track((PromotedObject*)obj, old->klass());
1329 }
1330 // Finally, install the klass pointer.
1331 obj->set_klass(old->klass());
1332
1333 assert(old->is_oop(), "Will dereference klass ptr below");
1334 collector()->promoted(true, // parallel
1335 obj_ptr, old->is_objArray(), word_sz);
1336
1337 NOT_PRODUCT(
1338 Atomic::inc(&_numObjectsPromoted);
1339 Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()),
1340 &_numWordsPromoted);
1341 )
1342
1343 return obj;
1344 }
1345
1346 void
1347 ConcurrentMarkSweepGeneration::
1348 par_promote_alloc_undo(int thread_num,
1349 HeapWord* obj, size_t word_sz) {
1350 // CMS does not support promotion undo.
1351 ShouldNotReachHere();
1352 }
1353
1354 void
1355 ConcurrentMarkSweepGeneration::
1356 par_promote_alloc_done(int thread_num) {
1357 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1358 ps->lab.retire();
1359 #if CFLS_LAB_REFILL_STATS
1360 if (thread_num == 0) {
1361 _cmsSpace->print_par_alloc_stats();
1362 }
1363 #endif
1364 }
1365
1366 void
1367 ConcurrentMarkSweepGeneration::
1368 par_oop_since_save_marks_iterate_done(int thread_num) {
1369 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1370 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1371 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1372 }
1373
1374 // XXXPERM
1375 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1376 size_t size,
1377 bool tlab)
1378 {
1379 // We allow a STW collection only if a full
1380 // collection was requested.
1381 return full || should_allocate(size, tlab); // FIX ME !!!
1382 // This and promotion failure handling are connected at the
1383 // hip and should be fixed by untying them.
1384 }
1385
1386 bool CMSCollector::shouldConcurrentCollect() {
1387 if (_full_gc_requested) {
1388 assert(ExplicitGCInvokesConcurrent, "Unexpected state");
1389 if (Verbose && PrintGCDetails) {
1390 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1391 " gc request");
1392 }
1393 return true;
1394 }
1395
1396 // For debugging purposes, change the type of collection.
1397 // If the rotation is not on the concurrent collection
1398 // type, don't start a concurrent collection.
1399 NOT_PRODUCT(
1400 if (RotateCMSCollectionTypes &&
1401 (_cmsGen->debug_collection_type() !=
1402 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1403 assert(_cmsGen->debug_collection_type() !=
1404 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1405 "Bad cms collection type");
1406 return false;
1407 }
1408 )
1409
1410 FreelistLocker x(this);
1411 // ------------------------------------------------------------------
1412 // Print out lots of information which affects the initiation of
1413 // a collection.
1414 if (PrintCMSInitiationStatistics && stats().valid()) {
1415 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1416 gclog_or_tty->stamp();
1417 gclog_or_tty->print_cr("");
1418 stats().print_on(gclog_or_tty);
1419 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1420 stats().time_until_cms_gen_full());
1421 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1422 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1423 _cmsGen->contiguous_available());
1424 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1425 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1426 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1427 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1428 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1429 }
1430 // ------------------------------------------------------------------
1431
1432 // If the estimated time to complete a cms collection (cms_duration())
1433 // is less than the estimated time remaining until the cms generation
1434 // is full, start a collection.
1435 if (!UseCMSInitiatingOccupancyOnly) {
1436 if (stats().valid()) {
1437 if (stats().time_until_cms_start() == 0.0) {
1438 return true;
1439 }
1440 } else {
1441 // We want to conservatively collect somewhat early in order
1442 // to try and "bootstrap" our CMS/promotion statistics;
1443 // this branch will not fire after the first successful CMS
1444 // collection because the stats should then be valid.
1445 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1446 if (Verbose && PrintGCDetails) {
1447 gclog_or_tty->print_cr(
1448 " CMSCollector: collect for bootstrapping statistics:"
1449 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1450 _bootstrap_occupancy);
1451 }
1452 return true;
1453 }
1454 }
1455 }
1456
1457 // Otherwise, we start a collection cycle if either the perm gen or
1458 // old gen want a collection cycle started. Each may use
1459 // an appropriate criterion for making this decision.
1460 // XXX We need to make sure that the gen expansion
1461 // criterion dovetails well with this. XXX NEED TO FIX THIS
1462 if (_cmsGen->should_concurrent_collect()) {
1463 if (Verbose && PrintGCDetails) {
1464 gclog_or_tty->print_cr("CMS old gen initiated");
1465 }
1466 return true;
1467 }
1468
1469 // We start a collection if we believe an incremental collection may fail;
1470 // this is not likely to be productive in practice because it's probably too
1471 // late anyway.
1472 GenCollectedHeap* gch = GenCollectedHeap::heap();
1473 assert(gch->collector_policy()->is_two_generation_policy(),
1474 "You may want to check the correctness of the following");
1475 if (gch->incremental_collection_will_fail()) {
1476 if (PrintGCDetails && Verbose) {
1477 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1478 }
1479 return true;
1480 }
1481
1482 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1483 bool res = update_should_unload_classes();
1484 if (res) {
1485 if (Verbose && PrintGCDetails) {
1486 gclog_or_tty->print_cr("CMS perm gen initiated");
1487 }
1488 return true;
1489 }
1490 }
1491 return false;
1492 }
1493
1494 // Clear _expansion_cause fields of constituent generations
1495 void CMSCollector::clear_expansion_cause() {
1496 _cmsGen->clear_expansion_cause();
1497 _permGen->clear_expansion_cause();
1498 }
1499
1500 // We should be conservative in starting a collection cycle. To
1501 // start too eagerly runs the risk of collecting too often in the
1502 // extreme. To collect too rarely falls back on full collections,
1503 // which works, even if not optimum in terms of concurrent work.
1504 // As a work around for too eagerly collecting, use the flag
1505 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1506 // giving the user an easily understandable way of controlling the
1507 // collections.
1508 // We want to start a new collection cycle if any of the following
1509 // conditions hold:
1510 // . our current occupancy exceeds the configured initiating occupancy
1511 // for this generation, or
1512 // . we recently needed to expand this space and have not, since that
1513 // expansion, done a collection of this generation, or
1514 // . the underlying space believes that it may be a good idea to initiate
1515 // a concurrent collection (this may be based on criteria such as the
1516 // following: the space uses linear allocation and linear allocation is
1517 // going to fail, or there is believed to be excessive fragmentation in
1518 // the generation, etc... or ...
1519 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1520 // the case of the old generation, not the perm generation; see CR 6543076):
1521 // we may be approaching a point at which allocation requests may fail because
1522 // we will be out of sufficient free space given allocation rate estimates.]
1523 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1524
1525 assert_lock_strong(freelistLock());
1526 if (occupancy() > initiating_occupancy()) {
1527 if (PrintGCDetails && Verbose) {
1528 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1529 short_name(), occupancy(), initiating_occupancy());
1530 }
1531 return true;
1532 }
1533 if (UseCMSInitiatingOccupancyOnly) {
1534 return false;
1535 }
1536 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1537 if (PrintGCDetails && Verbose) {
1538 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1539 short_name());
1540 }
1541 return true;
1542 }
1543 if (_cmsSpace->should_concurrent_collect()) {
1544 if (PrintGCDetails && Verbose) {
1545 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1546 short_name());
1547 }
1548 return true;
1549 }
1550 return false;
1551 }
1552
1553 void ConcurrentMarkSweepGeneration::collect(bool full,
1554 bool clear_all_soft_refs,
1555 size_t size,
1556 bool tlab)
1557 {
1558 collector()->collect(full, clear_all_soft_refs, size, tlab);
1559 }
1560
1561 void CMSCollector::collect(bool full,
1562 bool clear_all_soft_refs,
1563 size_t size,
1564 bool tlab)
1565 {
1566 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1567 // For debugging purposes skip the collection if the state
1568 // is not currently idle
1569 if (TraceCMSState) {
1570 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1571 Thread::current(), full, _collectorState);
1572 }
1573 return;
1574 }
1575
1576 // The following "if" branch is present for defensive reasons.
1577 // In the current uses of this interface, it can be replaced with:
1578 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1579 // But I am not placing that assert here to allow future
1580 // generality in invoking this interface.
1581 if (GC_locker::is_active()) {
1582 // A consistency test for GC_locker
1583 assert(GC_locker::needs_gc(), "Should have been set already");
1584 // Skip this foreground collection, instead
1585 // expanding the heap if necessary.
1586 // Need the free list locks for the call to free() in compute_new_size()
1587 compute_new_size();
1588 return;
1589 }
1590 acquire_control_and_collect(full, clear_all_soft_refs);
1591 _full_gcs_since_conc_gc++;
1592
1593 }
1594
1595 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1596 GenCollectedHeap* gch = GenCollectedHeap::heap();
1597 unsigned int gc_count = gch->total_full_collections();
1598 if (gc_count == full_gc_count) {
1599 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1600 _full_gc_requested = true;
1601 CGC_lock->notify(); // nudge CMS thread
1602 }
1603 }
1604
1605
1606 // The foreground and background collectors need to coordinate in order
1607 // to make sure that they do not mutually interfere with CMS collections.
1608 // When a background collection is active,
1609 // the foreground collector may need to take over (preempt) and
1610 // synchronously complete an ongoing collection. Depending on the
1611 // frequency of the background collections and the heap usage
1612 // of the application, this preemption can be seldom or frequent.
1613 // There are only certain
1614 // points in the background collection that the "collection-baton"
1615 // can be passed to the foreground collector.
1616 //
1617 // The foreground collector will wait for the baton before
1618 // starting any part of the collection. The foreground collector
1619 // will only wait at one location.
1620 //
1621 // The background collector will yield the baton before starting a new
1622 // phase of the collection (e.g., before initial marking, marking from roots,
1623 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1624 // of the loop which switches the phases. The background collector does some
1625 // of the phases (initial mark, final re-mark) with the world stopped.
1626 // Because of locking involved in stopping the world,
1627 // the foreground collector should not block waiting for the background
1628 // collector when it is doing a stop-the-world phase. The background
1629 // collector will yield the baton at an additional point just before
1630 // it enters a stop-the-world phase. Once the world is stopped, the
1631 // background collector checks the phase of the collection. If the
1632 // phase has not changed, it proceeds with the collection. If the
1633 // phase has changed, it skips that phase of the collection. See
1634 // the comments on the use of the Heap_lock in collect_in_background().
1635 //
1636 // Variable used in baton passing.
1637 // _foregroundGCIsActive - Set to true by the foreground collector when
1638 // it wants the baton. The foreground clears it when it has finished
1639 // the collection.
1640 // _foregroundGCShouldWait - Set to true by the background collector
1641 // when it is running. The foreground collector waits while
1642 // _foregroundGCShouldWait is true.
1643 // CGC_lock - monitor used to protect access to the above variables
1644 // and to notify the foreground and background collectors.
1645 // _collectorState - current state of the CMS collection.
1646 //
1647 // The foreground collector
1648 // acquires the CGC_lock
1649 // sets _foregroundGCIsActive
1650 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1651 // various locks acquired in preparation for the collection
1652 // are released so as not to block the background collector
1653 // that is in the midst of a collection
1654 // proceeds with the collection
1655 // clears _foregroundGCIsActive
1656 // returns
1657 //
1658 // The background collector in a loop iterating on the phases of the
1659 // collection
1660 // acquires the CGC_lock
1661 // sets _foregroundGCShouldWait
1662 // if _foregroundGCIsActive is set
1663 // clears _foregroundGCShouldWait, notifies _CGC_lock
1664 // waits on _CGC_lock for _foregroundGCIsActive to become false
1665 // and exits the loop.
1666 // otherwise
1667 // proceed with that phase of the collection
1668 // if the phase is a stop-the-world phase,
1669 // yield the baton once more just before enqueueing
1670 // the stop-world CMS operation (executed by the VM thread).
1671 // returns after all phases of the collection are done
1672 //
1673
1674 void CMSCollector::acquire_control_and_collect(bool full,
1675 bool clear_all_soft_refs) {
1676 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1677 assert(!Thread::current()->is_ConcurrentGC_thread(),
1678 "shouldn't try to acquire control from self!");
1679
1680 // Start the protocol for acquiring control of the
1681 // collection from the background collector (aka CMS thread).
1682 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1683 "VM thread should have CMS token");
1684 // Remember the possibly interrupted state of an ongoing
1685 // concurrent collection
1686 CollectorState first_state = _collectorState;
1687
1688 // Signal to a possibly ongoing concurrent collection that
1689 // we want to do a foreground collection.
1690 _foregroundGCIsActive = true;
1691
1692 // Disable incremental mode during a foreground collection.
1693 ICMSDisabler icms_disabler;
1694
1695 // release locks and wait for a notify from the background collector
1696 // releasing the locks in only necessary for phases which
1697 // do yields to improve the granularity of the collection.
1698 assert_lock_strong(bitMapLock());
1699 // We need to lock the Free list lock for the space that we are
1700 // currently collecting.
1701 assert(haveFreelistLocks(), "Must be holding free list locks");
1702 bitMapLock()->unlock();
1703 releaseFreelistLocks();
1704 {
1705 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1706 if (_foregroundGCShouldWait) {
1707 // We are going to be waiting for action for the CMS thread;
1708 // it had better not be gone (for instance at shutdown)!
1709 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1710 "CMS thread must be running");
1711 // Wait here until the background collector gives us the go-ahead
1712 ConcurrentMarkSweepThread::clear_CMS_flag(
1713 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1714 // Get a possibly blocked CMS thread going:
1715 // Note that we set _foregroundGCIsActive true above,
1716 // without protection of the CGC_lock.
1717 CGC_lock->notify();
1718 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1719 "Possible deadlock");
1720 while (_foregroundGCShouldWait) {
1721 // wait for notification
1722 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1723 // Possibility of delay/starvation here, since CMS token does
1724 // not know to give priority to VM thread? Actually, i think
1725 // there wouldn't be any delay/starvation, but the proof of
1726 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1727 }
1728 ConcurrentMarkSweepThread::set_CMS_flag(
1729 ConcurrentMarkSweepThread::CMS_vm_has_token);
1730 }
1731 }
1732 // The CMS_token is already held. Get back the other locks.
1733 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1734 "VM thread should have CMS token");
1735 getFreelistLocks();
1736 bitMapLock()->lock_without_safepoint_check();
1737 if (TraceCMSState) {
1738 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1739 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1740 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1741 }
1742
1743 // Check if we need to do a compaction, or if not, whether
1744 // we need to start the mark-sweep from scratch.
1745 bool should_compact = false;
1746 bool should_start_over = false;
1747 decide_foreground_collection_type(clear_all_soft_refs,
1748 &should_compact, &should_start_over);
1749
1750 NOT_PRODUCT(
1751 if (RotateCMSCollectionTypes) {
1752 if (_cmsGen->debug_collection_type() ==
1753 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1754 should_compact = true;
1755 } else if (_cmsGen->debug_collection_type() ==
1756 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1757 should_compact = false;
1758 }
1759 }
1760 )
1761
1762 if (PrintGCDetails && first_state > Idling) {
1763 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1764 if (GCCause::is_user_requested_gc(cause) ||
1765 GCCause::is_serviceability_requested_gc(cause)) {
1766 gclog_or_tty->print(" (concurrent mode interrupted)");
1767 } else {
1768 gclog_or_tty->print(" (concurrent mode failure)");
1769 }
1770 }
1771
1772 if (should_compact) {
1773 // If the collection is being acquired from the background
1774 // collector, there may be references on the discovered
1775 // references lists that have NULL referents (being those
1776 // that were concurrently cleared by a mutator) or
1777 // that are no longer active (having been enqueued concurrently
1778 // by the mutator).
1779 // Scrub the list of those references because Mark-Sweep-Compact
1780 // code assumes referents are not NULL and that all discovered
1781 // Reference objects are active.
1782 ref_processor()->clean_up_discovered_references();
1783
1784 do_compaction_work(clear_all_soft_refs);
1785
1786 // Has the GC time limit been exceeded?
1787 check_gc_time_limit();
1788
1789 } else {
1790 do_mark_sweep_work(clear_all_soft_refs, first_state,
1791 should_start_over);
1792 }
1793 // Reset the expansion cause, now that we just completed
1794 // a collection cycle.
1795 clear_expansion_cause();
1796 _foregroundGCIsActive = false;
1797 return;
1798 }
1799
1800 void CMSCollector::check_gc_time_limit() {
1801
1802 // Ignore explicit GC's. Exiting here does not set the flag and
1803 // does not reset the count. Updating of the averages for system
1804 // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC.
1805 GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause();
1806 if (GCCause::is_user_requested_gc(gc_cause) ||
1807 GCCause::is_serviceability_requested_gc(gc_cause)) {
1808 return;
1809 }
1810
1811 // Calculate the fraction of the CMS generation was freed during
1812 // the last collection.
1813 // Only consider the STW compacting cost for now.
1814 //
1815 // Note that the gc time limit test only works for the collections
1816 // of the young gen + tenured gen and not for collections of the
1817 // permanent gen. That is because the calculation of the space
1818 // freed by the collection is the free space in the young gen +
1819 // tenured gen.
1820
1821 double fraction_free =
1822 ((double)_cmsGen->free())/((double)_cmsGen->max_capacity());
1823 if ((100.0 * size_policy()->compacting_gc_cost()) >
1824 ((double) GCTimeLimit) &&
1825 ((fraction_free * 100) < GCHeapFreeLimit)) {
1826 size_policy()->inc_gc_time_limit_count();
1827 if (UseGCOverheadLimit &&
1828 (size_policy()->gc_time_limit_count() >
1829 AdaptiveSizePolicyGCTimeLimitThreshold)) {
1830 size_policy()->set_gc_time_limit_exceeded(true);
1831 // Avoid consecutive OOM due to the gc time limit by resetting
1832 // the counter.
1833 size_policy()->reset_gc_time_limit_count();
1834 if (PrintGCDetails) {
1835 gclog_or_tty->print_cr(" GC is exceeding overhead limit "
1836 "of %d%%", GCTimeLimit);
1837 }
1838 } else {
1839 if (PrintGCDetails) {
1840 gclog_or_tty->print_cr(" GC would exceed overhead limit "
1841 "of %d%%", GCTimeLimit);
1842 }
1843 }
1844 } else {
1845 size_policy()->reset_gc_time_limit_count();
1846 }
1847 }
1848
1849 // Resize the perm generation and the tenured generation
1850 // after obtaining the free list locks for the
1851 // two generations.
1852 void CMSCollector::compute_new_size() {
1853 assert_locked_or_safepoint(Heap_lock);
1854 FreelistLocker z(this);
1855 _permGen->compute_new_size();
1856 _cmsGen->compute_new_size();
1857 }
1858
1859 // A work method used by foreground collection to determine
1860 // what type of collection (compacting or not, continuing or fresh)
1861 // it should do.
1862 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1863 // and CMSCompactWhenClearAllSoftRefs the default in the future
1864 // and do away with the flags after a suitable period.
1865 void CMSCollector::decide_foreground_collection_type(
1866 bool clear_all_soft_refs, bool* should_compact,
1867 bool* should_start_over) {
1868 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1869 // flag is set, and we have either requested a System.gc() or
1870 // the number of full gc's since the last concurrent cycle
1871 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1872 // or if an incremental collection has failed
1873 GenCollectedHeap* gch = GenCollectedHeap::heap();
1874 assert(gch->collector_policy()->is_two_generation_policy(),
1875 "You may want to check the correctness of the following");
1876 // Inform cms gen if this was due to partial collection failing.
1877 // The CMS gen may use this fact to determine its expansion policy.
1878 if (gch->incremental_collection_will_fail()) {
1879 assert(!_cmsGen->incremental_collection_failed(),
1880 "Should have been noticed, reacted to and cleared");
1881 _cmsGen->set_incremental_collection_failed();
1882 }
1883 *should_compact =
1884 UseCMSCompactAtFullCollection &&
1885 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1886 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1887 gch->incremental_collection_will_fail());
1888 *should_start_over = false;
1889 if (clear_all_soft_refs && !*should_compact) {
1890 // We are about to do a last ditch collection attempt
1891 // so it would normally make sense to do a compaction
1892 // to reclaim as much space as possible.
1893 if (CMSCompactWhenClearAllSoftRefs) {
1894 // Default: The rationale is that in this case either
1895 // we are past the final marking phase, in which case
1896 // we'd have to start over, or so little has been done
1897 // that there's little point in saving that work. Compaction
1898 // appears to be the sensible choice in either case.
1899 *should_compact = true;
1900 } else {
1901 // We have been asked to clear all soft refs, but not to
1902 // compact. Make sure that we aren't past the final checkpoint
1903 // phase, for that is where we process soft refs. If we are already
1904 // past that phase, we'll need to redo the refs discovery phase and
1905 // if necessary clear soft refs that weren't previously
1906 // cleared. We do so by remembering the phase in which
1907 // we came in, and if we are past the refs processing
1908 // phase, we'll choose to just redo the mark-sweep
1909 // collection from scratch.
1910 if (_collectorState > FinalMarking) {
1911 // We are past the refs processing phase;
1912 // start over and do a fresh synchronous CMS cycle
1913 _collectorState = Resetting; // skip to reset to start new cycle
1914 reset(false /* == !asynch */);
1915 *should_start_over = true;
1916 } // else we can continue a possibly ongoing current cycle
1917 }
1918 }
1919 }
1920
1921 // A work method used by the foreground collector to do
1922 // a mark-sweep-compact.
1923 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1924 GenCollectedHeap* gch = GenCollectedHeap::heap();
1925 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1926 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1927 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1928 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1929 }
1930
1931 // Sample collection interval time and reset for collection pause.
1932 if (UseAdaptiveSizePolicy) {
1933 size_policy()->msc_collection_begin();
1934 }
1935
1936 // Temporarily widen the span of the weak reference processing to
1937 // the entire heap.
1938 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1939 ReferenceProcessorSpanMutator x(ref_processor(), new_span);
1940
1941 // Temporarily, clear the "is_alive_non_header" field of the
1942 // reference processor.
1943 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
1944
1945 // Temporarily make reference _processing_ single threaded (non-MT).
1946 ReferenceProcessorMTProcMutator z(ref_processor(), false);
1947
1948 // Temporarily make refs discovery atomic
1949 ReferenceProcessorAtomicMutator w(ref_processor(), true);
1950
1951 ref_processor()->set_enqueuing_is_done(false);
1952 ref_processor()->enable_discovery();
1953 // If an asynchronous collection finishes, the _modUnionTable is
1954 // all clear. If we are assuming the collection from an asynchronous
1955 // collection, clear the _modUnionTable.
1956 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1957 "_modUnionTable should be clear if the baton was not passed");
1958 _modUnionTable.clear_all();
1959
1960 // We must adjust the allocation statistics being maintained
1961 // in the free list space. We do so by reading and clearing
1962 // the sweep timer and updating the block flux rate estimates below.
1963 assert(_sweep_timer.is_active(), "We should never see the timer inactive");
1964 _sweep_timer.stop();
1965 // Note that we do not use this sample to update the _sweep_estimate.
1966 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
1967 _sweep_estimate.padded_average());
1968
1969 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1970 ref_processor(), clear_all_soft_refs);
1971 #ifdef ASSERT
1972 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1973 size_t free_size = cms_space->free();
1974 assert(free_size ==
1975 pointer_delta(cms_space->end(), cms_space->compaction_top())
1976 * HeapWordSize,
1977 "All the free space should be compacted into one chunk at top");
1978 assert(cms_space->dictionary()->totalChunkSize(
1979 debug_only(cms_space->freelistLock())) == 0 ||
1980 cms_space->totalSizeInIndexedFreeLists() == 0,
1981 "All the free space should be in a single chunk");
1982 size_t num = cms_space->totalCount();
1983 assert((free_size == 0 && num == 0) ||
1984 (free_size > 0 && (num == 1 || num == 2)),
1985 "There should be at most 2 free chunks after compaction");
1986 #endif // ASSERT
1987 _collectorState = Resetting;
1988 assert(_restart_addr == NULL,
1989 "Should have been NULL'd before baton was passed");
1990 reset(false /* == !asynch */);
1991 _cmsGen->reset_after_compaction();
1992 _concurrent_cycles_since_last_unload = 0;
1993
1994 if (verifying() && !should_unload_classes()) {
1995 perm_gen_verify_bit_map()->clear_all();
1996 }
1997
1998 // Clear any data recorded in the PLAB chunk arrays.
1999 if (_survivor_plab_array != NULL) {
2000 reset_survivor_plab_arrays();
2001 }
2002
2003 // Adjust the per-size allocation stats for the next epoch.
2004 _cmsGen->cmsSpace()->endSweepFLCensus(sweepCount() /* fake */);
2005 // Restart the "sweep timer" for next epoch.
2006 _sweep_timer.reset();
2007 _sweep_timer.start();
2008
2009 // Sample collection pause time and reset for collection interval.
2010 if (UseAdaptiveSizePolicy) {
2011 size_policy()->msc_collection_end(gch->gc_cause());
2012 }
2013
2014 // For a mark-sweep-compact, compute_new_size() will be called
2015 // in the heap's do_collection() method.
2016 }
2017
2018 // A work method used by the foreground collector to do
2019 // a mark-sweep, after taking over from a possibly on-going
2020 // concurrent mark-sweep collection.
2021 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2022 CollectorState first_state, bool should_start_over) {
2023 if (PrintGC && Verbose) {
2024 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2025 "collector with count %d",
2026 _full_gcs_since_conc_gc);
2027 }
2028 switch (_collectorState) {
2029 case Idling:
2030 if (first_state == Idling || should_start_over) {
2031 // The background GC was not active, or should
2032 // restarted from scratch; start the cycle.
2033 _collectorState = InitialMarking;
2034 }
2035 // If first_state was not Idling, then a background GC
2036 // was in progress and has now finished. No need to do it
2037 // again. Leave the state as Idling.
2038 break;
2039 case Precleaning:
2040 // In the foreground case don't do the precleaning since
2041 // it is not done concurrently and there is extra work
2042 // required.
2043 _collectorState = FinalMarking;
2044 }
2045 if (PrintGCDetails &&
2046 (_collectorState > Idling ||
2047 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2048 gclog_or_tty->print(" (concurrent mode failure)");
2049 }
2050 collect_in_foreground(clear_all_soft_refs);
2051
2052 // For a mark-sweep, compute_new_size() will be called
2053 // in the heap's do_collection() method.
2054 }
2055
2056
2057 void CMSCollector::getFreelistLocks() const {
2058 // Get locks for all free lists in all generations that this
2059 // collector is responsible for
2060 _cmsGen->freelistLock()->lock_without_safepoint_check();
2061 _permGen->freelistLock()->lock_without_safepoint_check();
2062 }
2063
2064 void CMSCollector::releaseFreelistLocks() const {
2065 // Release locks for all free lists in all generations that this
2066 // collector is responsible for
2067 _cmsGen->freelistLock()->unlock();
2068 _permGen->freelistLock()->unlock();
2069 }
2070
2071 bool CMSCollector::haveFreelistLocks() const {
2072 // Check locks for all free lists in all generations that this
2073 // collector is responsible for
2074 assert_lock_strong(_cmsGen->freelistLock());
2075 assert_lock_strong(_permGen->freelistLock());
2076 PRODUCT_ONLY(ShouldNotReachHere());
2077 return true;
2078 }
2079
2080 // A utility class that is used by the CMS collector to
2081 // temporarily "release" the foreground collector from its
2082 // usual obligation to wait for the background collector to
2083 // complete an ongoing phase before proceeding.
2084 class ReleaseForegroundGC: public StackObj {
2085 private:
2086 CMSCollector* _c;
2087 public:
2088 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2089 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2090 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2091 // allow a potentially blocked foreground collector to proceed
2092 _c->_foregroundGCShouldWait = false;
2093 if (_c->_foregroundGCIsActive) {
2094 CGC_lock->notify();
2095 }
2096 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2097 "Possible deadlock");
2098 }
2099
2100 ~ReleaseForegroundGC() {
2101 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2102 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2103 _c->_foregroundGCShouldWait = true;
2104 }
2105 };
2106
2107 // There are separate collect_in_background and collect_in_foreground because of
2108 // the different locking requirements of the background collector and the
2109 // foreground collector. There was originally an attempt to share
2110 // one "collect" method between the background collector and the foreground
2111 // collector but the if-then-else required made it cleaner to have
2112 // separate methods.
2113 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2114 assert(Thread::current()->is_ConcurrentGC_thread(),
2115 "A CMS asynchronous collection is only allowed on a CMS thread.");
2116
2117 GenCollectedHeap* gch = GenCollectedHeap::heap();
2118 {
2119 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2120 MutexLockerEx hl(Heap_lock, safepoint_check);
2121 FreelistLocker fll(this);
2122 MutexLockerEx x(CGC_lock, safepoint_check);
2123 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2124 // The foreground collector is active or we're
2125 // not using asynchronous collections. Skip this
2126 // background collection.
2127 assert(!_foregroundGCShouldWait, "Should be clear");
2128 return;
2129 } else {
2130 assert(_collectorState == Idling, "Should be idling before start.");
2131 _collectorState = InitialMarking;
2132 // Reset the expansion cause, now that we are about to begin
2133 // a new cycle.
2134 clear_expansion_cause();
2135 }
2136 // Decide if we want to enable class unloading as part of the
2137 // ensuing concurrent GC cycle.
2138 update_should_unload_classes();
2139 _full_gc_requested = false; // acks all outstanding full gc requests
2140 // Signal that we are about to start a collection
2141 gch->increment_total_full_collections(); // ... starting a collection cycle
2142 _collection_count_start = gch->total_full_collections();
2143 }
2144
2145 // Used for PrintGC
2146 size_t prev_used;
2147 if (PrintGC && Verbose) {
2148 prev_used = _cmsGen->used(); // XXXPERM
2149 }
2150
2151 // The change of the collection state is normally done at this level;
2152 // the exceptions are phases that are executed while the world is
2153 // stopped. For those phases the change of state is done while the
2154 // world is stopped. For baton passing purposes this allows the
2155 // background collector to finish the phase and change state atomically.
2156 // The foreground collector cannot wait on a phase that is done
2157 // while the world is stopped because the foreground collector already
2158 // has the world stopped and would deadlock.
2159 while (_collectorState != Idling) {
2160 if (TraceCMSState) {
2161 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2162 Thread::current(), _collectorState);
2163 }
2164 // The foreground collector
2165 // holds the Heap_lock throughout its collection.
2166 // holds the CMS token (but not the lock)
2167 // except while it is waiting for the background collector to yield.
2168 //
2169 // The foreground collector should be blocked (not for long)
2170 // if the background collector is about to start a phase
2171 // executed with world stopped. If the background
2172 // collector has already started such a phase, the
2173 // foreground collector is blocked waiting for the
2174 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2175 // are executed in the VM thread.
2176 //
2177 // The locking order is
2178 // PendingListLock (PLL) -- if applicable (FinalMarking)
2179 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2180 // CMS token (claimed in
2181 // stop_world_and_do() -->
2182 // safepoint_synchronize() -->
2183 // CMSThread::synchronize())
2184
2185 {
2186 // Check if the FG collector wants us to yield.
2187 CMSTokenSync x(true); // is cms thread
2188 if (waitForForegroundGC()) {
2189 // We yielded to a foreground GC, nothing more to be
2190 // done this round.
2191 assert(_foregroundGCShouldWait == false, "We set it to false in "
2192 "waitForForegroundGC()");
2193 if (TraceCMSState) {
2194 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2195 " exiting collection CMS state %d",
2196 Thread::current(), _collectorState);
2197 }
2198 return;
2199 } else {
2200 // The background collector can run but check to see if the
2201 // foreground collector has done a collection while the
2202 // background collector was waiting to get the CGC_lock
2203 // above. If yes, break so that _foregroundGCShouldWait
2204 // is cleared before returning.
2205 if (_collectorState == Idling) {
2206 break;
2207 }
2208 }
2209 }
2210
2211 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2212 "should be waiting");
2213
2214 switch (_collectorState) {
2215 case InitialMarking:
2216 {
2217 ReleaseForegroundGC x(this);
2218 stats().record_cms_begin();
2219
2220 VM_CMS_Initial_Mark initial_mark_op(this);
2221 VMThread::execute(&initial_mark_op);
2222 }
2223 // The collector state may be any legal state at this point
2224 // since the background collector may have yielded to the
2225 // foreground collector.
2226 break;
2227 case Marking:
2228 // initial marking in checkpointRootsInitialWork has been completed
2229 if (markFromRoots(true)) { // we were successful
2230 assert(_collectorState == Precleaning, "Collector state should "
2231 "have changed");
2232 } else {
2233 assert(_foregroundGCIsActive, "Internal state inconsistency");
2234 }
2235 break;
2236 case Precleaning:
2237 if (UseAdaptiveSizePolicy) {
2238 size_policy()->concurrent_precleaning_begin();
2239 }
2240 // marking from roots in markFromRoots has been completed
2241 preclean();
2242 if (UseAdaptiveSizePolicy) {
2243 size_policy()->concurrent_precleaning_end();
2244 }
2245 assert(_collectorState == AbortablePreclean ||
2246 _collectorState == FinalMarking,
2247 "Collector state should have changed");
2248 break;
2249 case AbortablePreclean:
2250 if (UseAdaptiveSizePolicy) {
2251 size_policy()->concurrent_phases_resume();
2252 }
2253 abortable_preclean();
2254 if (UseAdaptiveSizePolicy) {
2255 size_policy()->concurrent_precleaning_end();
2256 }
2257 assert(_collectorState == FinalMarking, "Collector state should "
2258 "have changed");
2259 break;
2260 case FinalMarking:
2261 {
2262 ReleaseForegroundGC x(this);
2263
2264 VM_CMS_Final_Remark final_remark_op(this);
2265 VMThread::execute(&final_remark_op);
2266 }
2267 assert(_foregroundGCShouldWait, "block post-condition");
2268 break;
2269 case Sweeping:
2270 if (UseAdaptiveSizePolicy) {
2271 size_policy()->concurrent_sweeping_begin();
2272 }
2273 // final marking in checkpointRootsFinal has been completed
2274 sweep(true);
2275 assert(_collectorState == Resizing, "Collector state change "
2276 "to Resizing must be done under the free_list_lock");
2277 _full_gcs_since_conc_gc = 0;
2278
2279 // Stop the timers for adaptive size policy for the concurrent phases
2280 if (UseAdaptiveSizePolicy) {
2281 size_policy()->concurrent_sweeping_end();
2282 size_policy()->concurrent_phases_end(gch->gc_cause(),
2283 gch->prev_gen(_cmsGen)->capacity(),
2284 _cmsGen->free());
2285 }
2286
2287 case Resizing: {
2288 // Sweeping has been completed...
2289 // At this point the background collection has completed.
2290 // Don't move the call to compute_new_size() down
2291 // into code that might be executed if the background
2292 // collection was preempted.
2293 {
2294 ReleaseForegroundGC x(this); // unblock FG collection
2295 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2296 CMSTokenSync z(true); // not strictly needed.
2297 if (_collectorState == Resizing) {
2298 compute_new_size();
2299 _collectorState = Resetting;
2300 } else {
2301 assert(_collectorState == Idling, "The state should only change"
2302 " because the foreground collector has finished the collection");
2303 }
2304 }
2305 break;
2306 }
2307 case Resetting:
2308 // CMS heap resizing has been completed
2309 reset(true);
2310 assert(_collectorState == Idling, "Collector state should "
2311 "have changed");
2312 stats().record_cms_end();
2313 // Don't move the concurrent_phases_end() and compute_new_size()
2314 // calls to here because a preempted background collection
2315 // has it's state set to "Resetting".
2316 break;
2317 case Idling:
2318 default:
2319 ShouldNotReachHere();
2320 break;
2321 }
2322 if (TraceCMSState) {
2323 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2324 Thread::current(), _collectorState);
2325 }
2326 assert(_foregroundGCShouldWait, "block post-condition");
2327 }
2328
2329 // Should this be in gc_epilogue?
2330 collector_policy()->counters()->update_counters();
2331
2332 {
2333 // Clear _foregroundGCShouldWait and, in the event that the
2334 // foreground collector is waiting, notify it, before
2335 // returning.
2336 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2337 _foregroundGCShouldWait = false;
2338 if (_foregroundGCIsActive) {
2339 CGC_lock->notify();
2340 }
2341 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2342 "Possible deadlock");
2343 }
2344 if (TraceCMSState) {
2345 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2346 " exiting collection CMS state %d",
2347 Thread::current(), _collectorState);
2348 }
2349 if (PrintGC && Verbose) {
2350 _cmsGen->print_heap_change(prev_used);
2351 }
2352 }
2353
2354 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2355 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2356 "Foreground collector should be waiting, not executing");
2357 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2358 "may only be done by the VM Thread with the world stopped");
2359 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2360 "VM thread should have CMS token");
2361
2362 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2363 true, gclog_or_tty);)
2364 if (UseAdaptiveSizePolicy) {
2365 size_policy()->ms_collection_begin();
2366 }
2367 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2368
2369 HandleMark hm; // Discard invalid handles created during verification
2370
2371 if (VerifyBeforeGC &&
2372 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2373 Universe::verify(true);
2374 }
2375
2376 bool init_mark_was_synchronous = false; // until proven otherwise
2377 while (_collectorState != Idling) {
2378 if (TraceCMSState) {
2379 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2380 Thread::current(), _collectorState);
2381 }
2382 switch (_collectorState) {
2383 case InitialMarking:
2384 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2385 checkpointRootsInitial(false);
2386 assert(_collectorState == Marking, "Collector state should have changed"
2387 " within checkpointRootsInitial()");
2388 break;
2389 case Marking:
2390 // initial marking in checkpointRootsInitialWork has been completed
2391 if (VerifyDuringGC &&
2392 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2393 gclog_or_tty->print("Verify before initial mark: ");
2394 Universe::verify(true);
2395 }
2396 {
2397 bool res = markFromRoots(false);
2398 assert(res && _collectorState == FinalMarking, "Collector state should "
2399 "have changed");
2400 break;
2401 }
2402 case FinalMarking:
2403 if (VerifyDuringGC &&
2404 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2405 gclog_or_tty->print("Verify before re-mark: ");
2406 Universe::verify(true);
2407 }
2408 checkpointRootsFinal(false, clear_all_soft_refs,
2409 init_mark_was_synchronous);
2410 assert(_collectorState == Sweeping, "Collector state should not "
2411 "have changed within checkpointRootsFinal()");
2412 break;
2413 case Sweeping:
2414 // final marking in checkpointRootsFinal has been completed
2415 if (VerifyDuringGC &&
2416 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2417 gclog_or_tty->print("Verify before sweep: ");
2418 Universe::verify(true);
2419 }
2420 sweep(false);
2421 assert(_collectorState == Resizing, "Incorrect state");
2422 break;
2423 case Resizing: {
2424 // Sweeping has been completed; the actual resize in this case
2425 // is done separately; nothing to be done in this state.
2426 _collectorState = Resetting;
2427 break;
2428 }
2429 case Resetting:
2430 // The heap has been resized.
2431 if (VerifyDuringGC &&
2432 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2433 gclog_or_tty->print("Verify before reset: ");
2434 Universe::verify(true);
2435 }
2436 reset(false);
2437 assert(_collectorState == Idling, "Collector state should "
2438 "have changed");
2439 break;
2440 case Precleaning:
2441 case AbortablePreclean:
2442 // Elide the preclean phase
2443 _collectorState = FinalMarking;
2444 break;
2445 default:
2446 ShouldNotReachHere();
2447 }
2448 if (TraceCMSState) {
2449 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2450 Thread::current(), _collectorState);
2451 }
2452 }
2453
2454 if (UseAdaptiveSizePolicy) {
2455 GenCollectedHeap* gch = GenCollectedHeap::heap();
2456 size_policy()->ms_collection_end(gch->gc_cause());
2457 }
2458
2459 if (VerifyAfterGC &&
2460 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2461 Universe::verify(true);
2462 }
2463 if (TraceCMSState) {
2464 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2465 " exiting collection CMS state %d",
2466 Thread::current(), _collectorState);
2467 }
2468 }
2469
2470 bool CMSCollector::waitForForegroundGC() {
2471 bool res = false;
2472 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2473 "CMS thread should have CMS token");
2474 // Block the foreground collector until the
2475 // background collectors decides whether to
2476 // yield.
2477 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2478 _foregroundGCShouldWait = true;
2479 if (_foregroundGCIsActive) {
2480 // The background collector yields to the
2481 // foreground collector and returns a value
2482 // indicating that it has yielded. The foreground
2483 // collector can proceed.
2484 res = true;
2485 _foregroundGCShouldWait = false;
2486 ConcurrentMarkSweepThread::clear_CMS_flag(
2487 ConcurrentMarkSweepThread::CMS_cms_has_token);
2488 ConcurrentMarkSweepThread::set_CMS_flag(
2489 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2490 // Get a possibly blocked foreground thread going
2491 CGC_lock->notify();
2492 if (TraceCMSState) {
2493 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2494 Thread::current(), _collectorState);
2495 }
2496 while (_foregroundGCIsActive) {
2497 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2498 }
2499 ConcurrentMarkSweepThread::set_CMS_flag(
2500 ConcurrentMarkSweepThread::CMS_cms_has_token);
2501 ConcurrentMarkSweepThread::clear_CMS_flag(
2502 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2503 }
2504 if (TraceCMSState) {
2505 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2506 Thread::current(), _collectorState);
2507 }
2508 return res;
2509 }
2510
2511 // Because of the need to lock the free lists and other structures in
2512 // the collector, common to all the generations that the collector is
2513 // collecting, we need the gc_prologues of individual CMS generations
2514 // delegate to their collector. It may have been simpler had the
2515 // current infrastructure allowed one to call a prologue on a
2516 // collector. In the absence of that we have the generation's
2517 // prologue delegate to the collector, which delegates back
2518 // some "local" work to a worker method in the individual generations
2519 // that it's responsible for collecting, while itself doing any
2520 // work common to all generations it's responsible for. A similar
2521 // comment applies to the gc_epilogue()'s.
2522 // The role of the varaible _between_prologue_and_epilogue is to
2523 // enforce the invocation protocol.
2524 void CMSCollector::gc_prologue(bool full) {
2525 // Call gc_prologue_work() for each CMSGen and PermGen that
2526 // we are responsible for.
2527
2528 // The following locking discipline assumes that we are only called
2529 // when the world is stopped.
2530 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2531
2532 // The CMSCollector prologue must call the gc_prologues for the
2533 // "generations" (including PermGen if any) that it's responsible
2534 // for.
2535
2536 assert( Thread::current()->is_VM_thread()
2537 || ( CMSScavengeBeforeRemark
2538 && Thread::current()->is_ConcurrentGC_thread()),
2539 "Incorrect thread type for prologue execution");
2540
2541 if (_between_prologue_and_epilogue) {
2542 // We have already been invoked; this is a gc_prologue delegation
2543 // from yet another CMS generation that we are responsible for, just
2544 // ignore it since all relevant work has already been done.
2545 return;
2546 }
2547
2548 // set a bit saying prologue has been called; cleared in epilogue
2549 _between_prologue_and_epilogue = true;
2550 // Claim locks for common data structures, then call gc_prologue_work()
2551 // for each CMSGen and PermGen that we are responsible for.
2552
2553 getFreelistLocks(); // gets free list locks on constituent spaces
2554 bitMapLock()->lock_without_safepoint_check();
2555
2556 // Should call gc_prologue_work() for all cms gens we are responsible for
2557 bool registerClosure = _collectorState >= Marking
2558 && _collectorState < Sweeping;
2559 ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar
2560 : &_modUnionClosure;
2561 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2562 _permGen->gc_prologue_work(full, registerClosure, muc);
2563
2564 if (!full) {
2565 stats().record_gc0_begin();
2566 }
2567 }
2568
2569 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2570 // Delegate to CMScollector which knows how to coordinate between
2571 // this and any other CMS generations that it is responsible for
2572 // collecting.
2573 collector()->gc_prologue(full);
2574 }
2575
2576 // This is a "private" interface for use by this generation's CMSCollector.
2577 // Not to be called directly by any other entity (for instance,
2578 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2579 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2580 bool registerClosure, ModUnionClosure* modUnionClosure) {
2581 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2582 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2583 "Should be NULL");
2584 if (registerClosure) {
2585 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2586 }
2587 cmsSpace()->gc_prologue();
2588 // Clear stat counters
2589 NOT_PRODUCT(
2590 assert(_numObjectsPromoted == 0, "check");
2591 assert(_numWordsPromoted == 0, "check");
2592 if (Verbose && PrintGC) {
2593 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2594 SIZE_FORMAT" bytes concurrently",
2595 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2596 }
2597 _numObjectsAllocated = 0;
2598 _numWordsAllocated = 0;
2599 )
2600 }
2601
2602 void CMSCollector::gc_epilogue(bool full) {
2603 // The following locking discipline assumes that we are only called
2604 // when the world is stopped.
2605 assert(SafepointSynchronize::is_at_safepoint(),
2606 "world is stopped assumption");
2607
2608 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2609 // if linear allocation blocks need to be appropriately marked to allow the
2610 // the blocks to be parsable. We also check here whether we need to nudge the
2611 // CMS collector thread to start a new cycle (if it's not already active).
2612 assert( Thread::current()->is_VM_thread()
2613 || ( CMSScavengeBeforeRemark
2614 && Thread::current()->is_ConcurrentGC_thread()),
2615 "Incorrect thread type for epilogue execution");
2616
2617 if (!_between_prologue_and_epilogue) {
2618 // We have already been invoked; this is a gc_epilogue delegation
2619 // from yet another CMS generation that we are responsible for, just
2620 // ignore it since all relevant work has already been done.
2621 return;
2622 }
2623 assert(haveFreelistLocks(), "must have freelist locks");
2624 assert_lock_strong(bitMapLock());
2625
2626 _cmsGen->gc_epilogue_work(full);
2627 _permGen->gc_epilogue_work(full);
2628
2629 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2630 // in case sampling was not already enabled, enable it
2631 _start_sampling = true;
2632 }
2633 // reset _eden_chunk_array so sampling starts afresh
2634 _eden_chunk_index = 0;
2635
2636 size_t cms_used = _cmsGen->cmsSpace()->used();
2637 size_t perm_used = _permGen->cmsSpace()->used();
2638
2639 // update performance counters - this uses a special version of
2640 // update_counters() that allows the utilization to be passed as a
2641 // parameter, avoiding multiple calls to used().
2642 //
2643 _cmsGen->update_counters(cms_used);
2644 _permGen->update_counters(perm_used);
2645
2646 if (CMSIncrementalMode) {
2647 icms_update_allocation_limits();
2648 }
2649
2650 bitMapLock()->unlock();
2651 releaseFreelistLocks();
2652
2653 _between_prologue_and_epilogue = false; // ready for next cycle
2654 }
2655
2656 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2657 collector()->gc_epilogue(full);
2658
2659 // Also reset promotion tracking in par gc thread states.
2660 if (ParallelGCThreads > 0) {
2661 for (uint i = 0; i < ParallelGCThreads; i++) {
2662 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2663 }
2664 }
2665 }
2666
2667 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2668 assert(!incremental_collection_failed(), "Should have been cleared");
2669 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2670 cmsSpace()->gc_epilogue();
2671 // Print stat counters
2672 NOT_PRODUCT(
2673 assert(_numObjectsAllocated == 0, "check");
2674 assert(_numWordsAllocated == 0, "check");
2675 if (Verbose && PrintGC) {
2676 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2677 SIZE_FORMAT" bytes",
2678 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2679 }
2680 _numObjectsPromoted = 0;
2681 _numWordsPromoted = 0;
2682 )
2683
2684 if (PrintGC && Verbose) {
2685 // Call down the chain in contiguous_available needs the freelistLock
2686 // so print this out before releasing the freeListLock.
2687 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2688 contiguous_available());
2689 }
2690 }
2691
2692 #ifndef PRODUCT
2693 bool CMSCollector::have_cms_token() {
2694 Thread* thr = Thread::current();
2695 if (thr->is_VM_thread()) {
2696 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2697 } else if (thr->is_ConcurrentGC_thread()) {
2698 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2699 } else if (thr->is_GC_task_thread()) {
2700 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2701 ParGCRareEvent_lock->owned_by_self();
2702 }
2703 return false;
2704 }
2705 #endif
2706
2707 // Check reachability of the given heap address in CMS generation,
2708 // treating all other generations as roots.
2709 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2710 // We could "guarantee" below, rather than assert, but i'll
2711 // leave these as "asserts" so that an adventurous debugger
2712 // could try this in the product build provided some subset of
2713 // the conditions were met, provided they were intersted in the
2714 // results and knew that the computation below wouldn't interfere
2715 // with other concurrent computations mutating the structures
2716 // being read or written.
2717 assert(SafepointSynchronize::is_at_safepoint(),
2718 "Else mutations in object graph will make answer suspect");
2719 assert(have_cms_token(), "Should hold cms token");
2720 assert(haveFreelistLocks(), "must hold free list locks");
2721 assert_lock_strong(bitMapLock());
2722
2723 // Clear the marking bit map array before starting, but, just
2724 // for kicks, first report if the given address is already marked
2725 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2726 _markBitMap.isMarked(addr) ? "" : " not");
2727
2728 if (verify_after_remark()) {
2729 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2730 bool result = verification_mark_bm()->isMarked(addr);
2731 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2732 result ? "IS" : "is NOT");
2733 return result;
2734 } else {
2735 gclog_or_tty->print_cr("Could not compute result");
2736 return false;
2737 }
2738 }
2739
2740 ////////////////////////////////////////////////////////
2741 // CMS Verification Support
2742 ////////////////////////////////////////////////////////
2743 // Following the remark phase, the following invariant
2744 // should hold -- each object in the CMS heap which is
2745 // marked in markBitMap() should be marked in the verification_mark_bm().
2746
2747 class VerifyMarkedClosure: public BitMapClosure {
2748 CMSBitMap* _marks;
2749 bool _failed;
2750
2751 public:
2752 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2753
2754 void do_bit(size_t offset) {
2755 HeapWord* addr = _marks->offsetToHeapWord(offset);
2756 if (!_marks->isMarked(addr)) {
2757 oop(addr)->print();
2758 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2759 _failed = true;
2760 }
2761 }
2762
2763 bool failed() { return _failed; }
2764 };
2765
2766 bool CMSCollector::verify_after_remark() {
2767 gclog_or_tty->print(" [Verifying CMS Marking... ");
2768 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2769 static bool init = false;
2770
2771 assert(SafepointSynchronize::is_at_safepoint(),
2772 "Else mutations in object graph will make answer suspect");
2773 assert(have_cms_token(),
2774 "Else there may be mutual interference in use of "
2775 " verification data structures");
2776 assert(_collectorState > Marking && _collectorState <= Sweeping,
2777 "Else marking info checked here may be obsolete");
2778 assert(haveFreelistLocks(), "must hold free list locks");
2779 assert_lock_strong(bitMapLock());
2780
2781
2782 // Allocate marking bit map if not already allocated
2783 if (!init) { // first time
2784 if (!verification_mark_bm()->allocate(_span)) {
2785 return false;
2786 }
2787 init = true;
2788 }
2789
2790 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2791
2792 // Turn off refs discovery -- so we will be tracing through refs.
2793 // This is as intended, because by this time
2794 // GC must already have cleared any refs that need to be cleared,
2795 // and traced those that need to be marked; moreover,
2796 // the marking done here is not going to intefere in any
2797 // way with the marking information used by GC.
2798 NoRefDiscovery no_discovery(ref_processor());
2799
2800 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2801
2802 // Clear any marks from a previous round
2803 verification_mark_bm()->clear_all();
2804 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2805 assert(overflow_list_is_empty(), "overflow list should be empty");
2806
2807 GenCollectedHeap* gch = GenCollectedHeap::heap();
2808 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2809 // Update the saved marks which may affect the root scans.
2810 gch->save_marks();
2811
2812 if (CMSRemarkVerifyVariant == 1) {
2813 // In this first variant of verification, we complete
2814 // all marking, then check if the new marks-verctor is
2815 // a subset of the CMS marks-vector.
2816 verify_after_remark_work_1();
2817 } else if (CMSRemarkVerifyVariant == 2) {
2818 // In this second variant of verification, we flag an error
2819 // (i.e. an object reachable in the new marks-vector not reachable
2820 // in the CMS marks-vector) immediately, also indicating the
2821 // identify of an object (A) that references the unmarked object (B) --
2822 // presumably, a mutation to A failed to be picked up by preclean/remark?
2823 verify_after_remark_work_2();
2824 } else {
2825 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2826 CMSRemarkVerifyVariant);
2827 }
2828 gclog_or_tty->print(" done] ");
2829 return true;
2830 }
2831
2832 void CMSCollector::verify_after_remark_work_1() {
2833 ResourceMark rm;
2834 HandleMark hm;
2835 GenCollectedHeap* gch = GenCollectedHeap::heap();
2836
2837 // Mark from roots one level into CMS
2838 MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */);
2839 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2840
2841 gch->gen_process_strong_roots(_cmsGen->level(),
2842 true, // younger gens are roots
2843 true, // collecting perm gen
2844 SharedHeap::ScanningOption(roots_scanning_options()),
2845 NULL, ¬Older);
2846
2847 // Now mark from the roots
2848 assert(_revisitStack.isEmpty(), "Should be empty");
2849 MarkFromRootsClosure markFromRootsClosure(this, _span,
2850 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2851 false /* don't yield */, true /* verifying */);
2852 assert(_restart_addr == NULL, "Expected pre-condition");
2853 verification_mark_bm()->iterate(&markFromRootsClosure);
2854 while (_restart_addr != NULL) {
2855 // Deal with stack overflow: by restarting at the indicated
2856 // address.
2857 HeapWord* ra = _restart_addr;
2858 markFromRootsClosure.reset(ra);
2859 _restart_addr = NULL;
2860 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2861 }
2862 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2863 verify_work_stacks_empty();
2864 // Should reset the revisit stack above, since no class tree
2865 // surgery is forthcoming.
2866 _revisitStack.reset(); // throwing away all contents
2867
2868 // Marking completed -- now verify that each bit marked in
2869 // verification_mark_bm() is also marked in markBitMap(); flag all
2870 // errors by printing corresponding objects.
2871 VerifyMarkedClosure vcl(markBitMap());
2872 verification_mark_bm()->iterate(&vcl);
2873 if (vcl.failed()) {
2874 gclog_or_tty->print("Verification failed");
2875 Universe::heap()->print();
2876 fatal(" ... aborting");
2877 }
2878 }
2879
2880 void CMSCollector::verify_after_remark_work_2() {
2881 ResourceMark rm;
2882 HandleMark hm;
2883 GenCollectedHeap* gch = GenCollectedHeap::heap();
2884
2885 // Mark from roots one level into CMS
2886 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2887 markBitMap(), true /* nmethods */);
2888 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2889 gch->gen_process_strong_roots(_cmsGen->level(),
2890 true, // younger gens are roots
2891 true, // collecting perm gen
2892 SharedHeap::ScanningOption(roots_scanning_options()),
2893 NULL, ¬Older);
2894
2895 // Now mark from the roots
2896 assert(_revisitStack.isEmpty(), "Should be empty");
2897 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2898 verification_mark_bm(), markBitMap(), verification_mark_stack());
2899 assert(_restart_addr == NULL, "Expected pre-condition");
2900 verification_mark_bm()->iterate(&markFromRootsClosure);
2901 while (_restart_addr != NULL) {
2902 // Deal with stack overflow: by restarting at the indicated
2903 // address.
2904 HeapWord* ra = _restart_addr;
2905 markFromRootsClosure.reset(ra);
2906 _restart_addr = NULL;
2907 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2908 }
2909 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2910 verify_work_stacks_empty();
2911 // Should reset the revisit stack above, since no class tree
2912 // surgery is forthcoming.
2913 _revisitStack.reset(); // throwing away all contents
2914
2915 // Marking completed -- now verify that each bit marked in
2916 // verification_mark_bm() is also marked in markBitMap(); flag all
2917 // errors by printing corresponding objects.
2918 VerifyMarkedClosure vcl(markBitMap());
2919 verification_mark_bm()->iterate(&vcl);
2920 assert(!vcl.failed(), "Else verification above should not have succeeded");
2921 }
2922
2923 void ConcurrentMarkSweepGeneration::save_marks() {
2924 // delegate to CMS space
2925 cmsSpace()->save_marks();
2926 for (uint i = 0; i < ParallelGCThreads; i++) {
2927 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2928 }
2929 }
2930
2931 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2932 return cmsSpace()->no_allocs_since_save_marks();
2933 }
2934
2935 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2936 \
2937 void ConcurrentMarkSweepGeneration:: \
2938 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2939 cl->set_generation(this); \
2940 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2941 cl->reset_generation(); \
2942 save_marks(); \
2943 }
2944
2945 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2946
2947 void
2948 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
2949 {
2950 // Not currently implemented; need to do the following. -- ysr.
2951 // dld -- I think that is used for some sort of allocation profiler. So it
2952 // really means the objects allocated by the mutator since the last
2953 // GC. We could potentially implement this cheaply by recording only
2954 // the direct allocations in a side data structure.
2955 //
2956 // I think we probably ought not to be required to support these
2957 // iterations at any arbitrary point; I think there ought to be some
2958 // call to enable/disable allocation profiling in a generation/space,
2959 // and the iterator ought to return the objects allocated in the
2960 // gen/space since the enable call, or the last iterator call (which
2961 // will probably be at a GC.) That way, for gens like CM&S that would
2962 // require some extra data structure to support this, we only pay the
2963 // cost when it's in use...
2964 cmsSpace()->object_iterate_since_last_GC(blk);
2965 }
2966
2967 void
2968 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
2969 cl->set_generation(this);
2970 younger_refs_in_space_iterate(_cmsSpace, cl);
2971 cl->reset_generation();
2972 }
2973
2974 void
2975 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
2976 if (freelistLock()->owned_by_self()) {
2977 Generation::oop_iterate(mr, cl);
2978 } else {
2979 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2980 Generation::oop_iterate(mr, cl);
2981 }
2982 }
2983
2984 void
2985 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
2986 if (freelistLock()->owned_by_self()) {
2987 Generation::oop_iterate(cl);
2988 } else {
2989 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2990 Generation::oop_iterate(cl);
2991 }
2992 }
2993
2994 void
2995 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2996 if (freelistLock()->owned_by_self()) {
2997 Generation::object_iterate(cl);
2998 } else {
2999 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3000 Generation::object_iterate(cl);
3001 }
3002 }
3003
3004 void
3005 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3006 }
3007
3008 void
3009 ConcurrentMarkSweepGeneration::post_compact() {
3010 }
3011
3012 void
3013 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3014 // Fix the linear allocation blocks to look like free blocks.
3015
3016 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3017 // are not called when the heap is verified during universe initialization and
3018 // at vm shutdown.
3019 if (freelistLock()->owned_by_self()) {
3020 cmsSpace()->prepare_for_verify();
3021 } else {
3022 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3023 cmsSpace()->prepare_for_verify();
3024 }
3025 }
3026
3027 void
3028 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3029 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3030 // are not called when the heap is verified during universe initialization and
3031 // at vm shutdown.
3032 if (freelistLock()->owned_by_self()) {
3033 cmsSpace()->verify(false /* ignored */);
3034 } else {
3035 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3036 cmsSpace()->verify(false /* ignored */);
3037 }
3038 }
3039
3040 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3041 _cmsGen->verify(allow_dirty);
3042 _permGen->verify(allow_dirty);
3043 }
3044
3045 #ifndef PRODUCT
3046 bool CMSCollector::overflow_list_is_empty() const {
3047 assert(_num_par_pushes >= 0, "Inconsistency");
3048 if (_overflow_list == NULL) {
3049 assert(_num_par_pushes == 0, "Inconsistency");
3050 }
3051 return _overflow_list == NULL;
3052 }
3053
3054 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3055 // merely consolidate assertion checks that appear to occur together frequently.
3056 void CMSCollector::verify_work_stacks_empty() const {
3057 assert(_markStack.isEmpty(), "Marking stack should be empty");
3058 assert(overflow_list_is_empty(), "Overflow list should be empty");
3059 }
3060
3061 void CMSCollector::verify_overflow_empty() const {
3062 assert(overflow_list_is_empty(), "Overflow list should be empty");
3063 assert(no_preserved_marks(), "No preserved marks");
3064 }
3065 #endif // PRODUCT
3066
3067 // Decide if we want to enable class unloading as part of the
3068 // ensuing concurrent GC cycle. We will collect the perm gen and
3069 // unload classes if it's the case that:
3070 // (1) an explicit gc request has been made and the flag
3071 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3072 // (2) (a) class unloading is enabled at the command line, and
3073 // (b) (i) perm gen threshold has been crossed, or
3074 // (ii) old gen is getting really full, or
3075 // (iii) the previous N CMS collections did not collect the
3076 // perm gen
3077 // NOTE: Provided there is no change in the state of the heap between
3078 // calls to this method, it should have idempotent results. Moreover,
3079 // its results should be monotonically increasing (i.e. going from 0 to 1,
3080 // but not 1 to 0) between successive calls between which the heap was
3081 // not collected. For the implementation below, it must thus rely on
3082 // the property that concurrent_cycles_since_last_unload()
3083 // will not decrease unless a collection cycle happened and that
3084 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3085 // themselves also monotonic in that sense. See check_monotonicity()
3086 // below.
3087 bool CMSCollector::update_should_unload_classes() {
3088 _should_unload_classes = false;
3089 // Condition 1 above
3090 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3091 _should_unload_classes = true;
3092 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3093 // Disjuncts 2.b.(i,ii,iii) above
3094 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3095 CMSClassUnloadingMaxInterval)
3096 || _permGen->should_concurrent_collect()
3097 || _cmsGen->is_too_full();
3098 }
3099 return _should_unload_classes;
3100 }
3101
3102 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3103 bool res = should_concurrent_collect();
3104 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3105 return res;
3106 }
3107
3108 void CMSCollector::setup_cms_unloading_and_verification_state() {
3109 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3110 || VerifyBeforeExit;
3111 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings
3112 | SharedHeap::SO_CodeCache;
3113
3114 if (should_unload_classes()) { // Should unload classes this cycle
3115 remove_root_scanning_option(rso); // Shrink the root set appropriately
3116 set_verifying(should_verify); // Set verification state for this cycle
3117 return; // Nothing else needs to be done at this time
3118 }
3119
3120 // Not unloading classes this cycle
3121 assert(!should_unload_classes(), "Inconsitency!");
3122 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3123 // We were not verifying, or we _were_ unloading classes in the last cycle,
3124 // AND some verification options are enabled this cycle; in this case,
3125 // we must make sure that the deadness map is allocated if not already so,
3126 // and cleared (if already allocated previously --
3127 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3128 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3129 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3130 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3131 "permanent generation verification disabled");
3132 return; // Note that we leave verification disabled, so we'll retry this
3133 // allocation next cycle. We _could_ remember this failure
3134 // and skip further attempts and permanently disable verification
3135 // attempts if that is considered more desirable.
3136 }
3137 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3138 "_perm_gen_ver_bit_map inconsistency?");
3139 } else {
3140 perm_gen_verify_bit_map()->clear_all();
3141 }
3142 // Include symbols, strings and code cache elements to prevent their resurrection.
3143 add_root_scanning_option(rso);
3144 set_verifying(true);
3145 } else if (verifying() && !should_verify) {
3146 // We were verifying, but some verification flags got disabled.
3147 set_verifying(false);
3148 // Exclude symbols, strings and code cache elements from root scanning to
3149 // reduce IM and RM pauses.
3150 remove_root_scanning_option(rso);
3151 }
3152 }
3153
3154
3155 #ifndef PRODUCT
3156 HeapWord* CMSCollector::block_start(const void* p) const {
3157 const HeapWord* addr = (HeapWord*)p;
3158 if (_span.contains(p)) {
3159 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3160 return _cmsGen->cmsSpace()->block_start(p);
3161 } else {
3162 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3163 "Inconsistent _span?");
3164 return _permGen->cmsSpace()->block_start(p);
3165 }
3166 }
3167 return NULL;
3168 }
3169 #endif
3170
3171 HeapWord*
3172 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3173 bool tlab,
3174 bool parallel) {
3175 assert(!tlab, "Can't deal with TLAB allocation");
3176 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3177 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3178 CMSExpansionCause::_satisfy_allocation);
3179 if (GCExpandToAllocateDelayMillis > 0) {
3180 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3181 }
3182 return have_lock_and_allocate(word_size, tlab);
3183 }
3184
3185 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3186 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3187 // to CardGeneration and share it...
3188 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3189 CMSExpansionCause::Cause cause)
3190 {
3191 assert_locked_or_safepoint(Heap_lock);
3192
3193 size_t aligned_bytes = ReservedSpace::page_align_size_up(bytes);
3194 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
3195 bool success = false;
3196 if (aligned_expand_bytes > aligned_bytes) {
3197 success = grow_by(aligned_expand_bytes);
3198 }
3199 if (!success) {
3200 success = grow_by(aligned_bytes);
3201 }
3202 if (!success) {
3203 size_t remaining_bytes = _virtual_space.uncommitted_size();
3204 if (remaining_bytes > 0) {
3205 success = grow_by(remaining_bytes);
3206 }
3207 }
3208 if (GC_locker::is_active()) {
3209 if (PrintGC && Verbose) {
3210 gclog_or_tty->print_cr("Garbage collection disabled, expanded heap instead");
3211 }
3212 }
3213 // remember why we expanded; this information is used
3214 // by shouldConcurrentCollect() when making decisions on whether to start
3215 // a new CMS cycle.
3216 if (success) {
3217 set_expansion_cause(cause);
3218 if (PrintGCDetails && Verbose) {
3219 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3220 CMSExpansionCause::to_string(cause));
3221 }
3222 }
3223 }
3224
3225 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3226 HeapWord* res = NULL;
3227 MutexLocker x(ParGCRareEvent_lock);
3228 while (true) {
3229 // Expansion by some other thread might make alloc OK now:
3230 res = ps->lab.alloc(word_sz);
3231 if (res != NULL) return res;
3232 // If there's not enough expansion space available, give up.
3233 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3234 return NULL;
3235 }
3236 // Otherwise, we try expansion.
3237 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3238 CMSExpansionCause::_allocate_par_lab);
3239 // Now go around the loop and try alloc again;
3240 // A competing par_promote might beat us to the expansion space,
3241 // so we may go around the loop again if promotion fails agaion.
3242 if (GCExpandToAllocateDelayMillis > 0) {
3243 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3244 }
3245 }
3246 }
3247
3248
3249 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3250 PromotionInfo* promo) {
3251 MutexLocker x(ParGCRareEvent_lock);
3252 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3253 while (true) {
3254 // Expansion by some other thread might make alloc OK now:
3255 if (promo->ensure_spooling_space()) {
3256 assert(promo->has_spooling_space(),
3257 "Post-condition of successful ensure_spooling_space()");
3258 return true;
3259 }
3260 // If there's not enough expansion space available, give up.
3261 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3262 return false;
3263 }
3264 // Otherwise, we try expansion.
3265 expand(refill_size_bytes, MinHeapDeltaBytes,
3266 CMSExpansionCause::_allocate_par_spooling_space);
3267 // Now go around the loop and try alloc again;
3268 // A competing allocation might beat us to the expansion space,
3269 // so we may go around the loop again if allocation fails again.
3270 if (GCExpandToAllocateDelayMillis > 0) {
3271 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3272 }
3273 }
3274 }
3275
3276
3277
3278 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3279 assert_locked_or_safepoint(Heap_lock);
3280 size_t size = ReservedSpace::page_align_size_down(bytes);
3281 if (size > 0) {
3282 shrink_by(size);
3283 }
3284 }
3285
3286 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3287 assert_locked_or_safepoint(Heap_lock);
3288 bool result = _virtual_space.expand_by(bytes);
3289 if (result) {
3290 HeapWord* old_end = _cmsSpace->end();
3291 size_t new_word_size =
3292 heap_word_size(_virtual_space.committed_size());
3293 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3294 _bts->resize(new_word_size); // resize the block offset shared array
3295 Universe::heap()->barrier_set()->resize_covered_region(mr);
3296 // Hmmmm... why doesn't CFLS::set_end verify locking?
3297 // This is quite ugly; FIX ME XXX
3298 _cmsSpace->assert_locked();
3299 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3300
3301 // update the space and generation capacity counters
3302 if (UsePerfData) {
3303 _space_counters->update_capacity();
3304 _gen_counters->update_all();
3305 }
3306
3307 if (Verbose && PrintGC) {
3308 size_t new_mem_size = _virtual_space.committed_size();
3309 size_t old_mem_size = new_mem_size - bytes;
3310 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3311 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3312 }
3313 }
3314 return result;
3315 }
3316
3317 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3318 assert_locked_or_safepoint(Heap_lock);
3319 bool success = true;
3320 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3321 if (remaining_bytes > 0) {
3322 success = grow_by(remaining_bytes);
3323 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3324 }
3325 return success;
3326 }
3327
3328 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3329 assert_locked_or_safepoint(Heap_lock);
3330 assert_lock_strong(freelistLock());
3331 // XXX Fix when compaction is implemented.
3332 warning("Shrinking of CMS not yet implemented");
3333 return;
3334 }
3335
3336
3337 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3338 // phases.
3339 class CMSPhaseAccounting: public StackObj {
3340 public:
3341 CMSPhaseAccounting(CMSCollector *collector,
3342 const char *phase,
3343 bool print_cr = true);
3344 ~CMSPhaseAccounting();
3345
3346 private:
3347 CMSCollector *_collector;
3348 const char *_phase;
3349 elapsedTimer _wallclock;
3350 bool _print_cr;
3351
3352 public:
3353 // Not MT-safe; so do not pass around these StackObj's
3354 // where they may be accessed by other threads.
3355 jlong wallclock_millis() {
3356 assert(_wallclock.is_active(), "Wall clock should not stop");
3357 _wallclock.stop(); // to record time
3358 jlong ret = _wallclock.milliseconds();
3359 _wallclock.start(); // restart
3360 return ret;
3361 }
3362 };
3363
3364 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3365 const char *phase,
3366 bool print_cr) :
3367 _collector(collector), _phase(phase), _print_cr(print_cr) {
3368
3369 if (PrintCMSStatistics != 0) {
3370 _collector->resetYields();
3371 }
3372 if (PrintGCDetails && PrintGCTimeStamps) {
3373 gclog_or_tty->date_stamp(PrintGCDateStamps);
3374 gclog_or_tty->stamp();
3375 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3376 _collector->cmsGen()->short_name(), _phase);
3377 }
3378 _collector->resetTimer();
3379 _wallclock.start();
3380 _collector->startTimer();
3381 }
3382
3383 CMSPhaseAccounting::~CMSPhaseAccounting() {
3384 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3385 _collector->stopTimer();
3386 _wallclock.stop();
3387 if (PrintGCDetails) {
3388 gclog_or_tty->date_stamp(PrintGCDateStamps);
3389 if (PrintGCTimeStamps) {
3390 gclog_or_tty->stamp();
3391 gclog_or_tty->print(": ");
3392 }
3393 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3394 _collector->cmsGen()->short_name(),
3395 _phase, _collector->timerValue(), _wallclock.seconds());
3396 if (_print_cr) {
3397 gclog_or_tty->print_cr("");
3398 }
3399 if (PrintCMSStatistics != 0) {
3400 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3401 _collector->yields());
3402 }
3403 }
3404 }
3405
3406 // CMS work
3407
3408 // Checkpoint the roots into this generation from outside
3409 // this generation. [Note this initial checkpoint need only
3410 // be approximate -- we'll do a catch up phase subsequently.]
3411 void CMSCollector::checkpointRootsInitial(bool asynch) {
3412 assert(_collectorState == InitialMarking, "Wrong collector state");
3413 check_correct_thread_executing();
3414 ReferenceProcessor* rp = ref_processor();
3415 SpecializationStats::clear();
3416 assert(_restart_addr == NULL, "Control point invariant");
3417 if (asynch) {
3418 // acquire locks for subsequent manipulations
3419 MutexLockerEx x(bitMapLock(),
3420 Mutex::_no_safepoint_check_flag);
3421 checkpointRootsInitialWork(asynch);
3422 rp->verify_no_references_recorded();
3423 rp->enable_discovery(); // enable ("weak") refs discovery
3424 _collectorState = Marking;
3425 } else {
3426 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3427 // which recognizes if we are a CMS generation, and doesn't try to turn on
3428 // discovery; verify that they aren't meddling.
3429 assert(!rp->discovery_is_atomic(),
3430 "incorrect setting of discovery predicate");
3431 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3432 "ref discovery for this generation kind");
3433 // already have locks
3434 checkpointRootsInitialWork(asynch);
3435 rp->enable_discovery(); // now enable ("weak") refs discovery
3436 _collectorState = Marking;
3437 }
3438 SpecializationStats::print();
3439 }
3440
3441 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3442 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3443 assert(_collectorState == InitialMarking, "just checking");
3444
3445 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3446 // precede our marking with a collection of all
3447 // younger generations to keep floating garbage to a minimum.
3448 // XXX: we won't do this for now -- it's an optimization to be done later.
3449
3450 // already have locks
3451 assert_lock_strong(bitMapLock());
3452 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3453
3454 // Setup the verification and class unloading state for this
3455 // CMS collection cycle.
3456 setup_cms_unloading_and_verification_state();
3457
3458 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3459 PrintGCDetails && Verbose, true, gclog_or_tty);)
3460 if (UseAdaptiveSizePolicy) {
3461 size_policy()->checkpoint_roots_initial_begin();
3462 }
3463
3464 // Reset all the PLAB chunk arrays if necessary.
3465 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3466 reset_survivor_plab_arrays();
3467 }
3468
3469 ResourceMark rm;
3470 HandleMark hm;
3471
3472 FalseClosure falseClosure;
3473 // In the case of a synchronous collection, we will elide the
3474 // remark step, so it's important to catch all the nmethod oops
3475 // in this step; hence the last argument to the constrcutor below.
3476 MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */);
3477 GenCollectedHeap* gch = GenCollectedHeap::heap();
3478
3479 verify_work_stacks_empty();
3480 verify_overflow_empty();
3481
3482 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3483 // Update the saved marks which may affect the root scans.
3484 gch->save_marks();
3485
3486 // weak reference processing has not started yet.
3487 ref_processor()->set_enqueuing_is_done(false);
3488
3489 {
3490 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3491 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3492 gch->gen_process_strong_roots(_cmsGen->level(),
3493 true, // younger gens are roots
3494 true, // collecting perm gen
3495 SharedHeap::ScanningOption(roots_scanning_options()),
3496 NULL, ¬Older);
3497 }
3498
3499 // Clear mod-union table; it will be dirtied in the prologue of
3500 // CMS generation per each younger generation collection.
3501
3502 assert(_modUnionTable.isAllClear(),
3503 "Was cleared in most recent final checkpoint phase"
3504 " or no bits are set in the gc_prologue before the start of the next "
3505 "subsequent marking phase.");
3506
3507 // Temporarily disabled, since pre/post-consumption closures don't
3508 // care about precleaned cards
3509 #if 0
3510 {
3511 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3512 (HeapWord*)_virtual_space.high());
3513 _ct->ct_bs()->preclean_dirty_cards(mr);
3514 }
3515 #endif
3516
3517 // Save the end of the used_region of the constituent generations
3518 // to be used to limit the extent of sweep in each generation.
3519 save_sweep_limits();
3520 if (UseAdaptiveSizePolicy) {
3521 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3522 }
3523 verify_overflow_empty();
3524 }
3525
3526 bool CMSCollector::markFromRoots(bool asynch) {
3527 // we might be tempted to assert that:
3528 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3529 // "inconsistent argument?");
3530 // However that wouldn't be right, because it's possible that
3531 // a safepoint is indeed in progress as a younger generation
3532 // stop-the-world GC happens even as we mark in this generation.
3533 assert(_collectorState == Marking, "inconsistent state?");
3534 check_correct_thread_executing();
3535 verify_overflow_empty();
3536
3537 bool res;
3538 if (asynch) {
3539
3540 // Start the timers for adaptive size policy for the concurrent phases
3541 // Do it here so that the foreground MS can use the concurrent
3542 // timer since a foreground MS might has the sweep done concurrently
3543 // or STW.
3544 if (UseAdaptiveSizePolicy) {
3545 size_policy()->concurrent_marking_begin();
3546 }
3547
3548 // Weak ref discovery note: We may be discovering weak
3549 // refs in this generation concurrent (but interleaved) with
3550 // weak ref discovery by a younger generation collector.
3551
3552 CMSTokenSyncWithLocks ts(true, bitMapLock());
3553 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3554 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3555 res = markFromRootsWork(asynch);
3556 if (res) {
3557 _collectorState = Precleaning;
3558 } else { // We failed and a foreground collection wants to take over
3559 assert(_foregroundGCIsActive, "internal state inconsistency");
3560 assert(_restart_addr == NULL, "foreground will restart from scratch");
3561 if (PrintGCDetails) {
3562 gclog_or_tty->print_cr("bailing out to foreground collection");
3563 }
3564 }
3565 if (UseAdaptiveSizePolicy) {
3566 size_policy()->concurrent_marking_end();
3567 }
3568 } else {
3569 assert(SafepointSynchronize::is_at_safepoint(),
3570 "inconsistent with asynch == false");
3571 if (UseAdaptiveSizePolicy) {
3572 size_policy()->ms_collection_marking_begin();
3573 }
3574 // already have locks
3575 res = markFromRootsWork(asynch);
3576 _collectorState = FinalMarking;
3577 if (UseAdaptiveSizePolicy) {
3578 GenCollectedHeap* gch = GenCollectedHeap::heap();
3579 size_policy()->ms_collection_marking_end(gch->gc_cause());
3580 }
3581 }
3582 verify_overflow_empty();
3583 return res;
3584 }
3585
3586 bool CMSCollector::markFromRootsWork(bool asynch) {
3587 // iterate over marked bits in bit map, doing a full scan and mark
3588 // from these roots using the following algorithm:
3589 // . if oop is to the right of the current scan pointer,
3590 // mark corresponding bit (we'll process it later)
3591 // . else (oop is to left of current scan pointer)
3592 // push oop on marking stack
3593 // . drain the marking stack
3594
3595 // Note that when we do a marking step we need to hold the
3596 // bit map lock -- recall that direct allocation (by mutators)
3597 // and promotion (by younger generation collectors) is also
3598 // marking the bit map. [the so-called allocate live policy.]
3599 // Because the implementation of bit map marking is not
3600 // robust wrt simultaneous marking of bits in the same word,
3601 // we need to make sure that there is no such interference
3602 // between concurrent such updates.
3603
3604 // already have locks
3605 assert_lock_strong(bitMapLock());
3606
3607 // Clear the revisit stack, just in case there are any
3608 // obsolete contents from a short-circuited previous CMS cycle.
3609 _revisitStack.reset();
3610 verify_work_stacks_empty();
3611 verify_overflow_empty();
3612 assert(_revisitStack.isEmpty(), "tabula rasa");
3613
3614 bool result = false;
3615 if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) {
3616 result = do_marking_mt(asynch);
3617 } else {
3618 result = do_marking_st(asynch);
3619 }
3620 return result;
3621 }
3622
3623 // Forward decl
3624 class CMSConcMarkingTask;
3625
3626 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3627 CMSCollector* _collector;
3628 CMSConcMarkingTask* _task;
3629 bool _yield;
3630 protected:
3631 virtual void yield();
3632 public:
3633 // "n_threads" is the number of threads to be terminated.
3634 // "queue_set" is a set of work queues of other threads.
3635 // "collector" is the CMS collector associated with this task terminator.
3636 // "yield" indicates whether we need the gang as a whole to yield.
3637 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
3638 CMSCollector* collector, bool yield) :
3639 ParallelTaskTerminator(n_threads, queue_set),
3640 _collector(collector),
3641 _yield(yield) { }
3642
3643 void set_task(CMSConcMarkingTask* task) {
3644 _task = task;
3645 }
3646 };
3647
3648 // MT Concurrent Marking Task
3649 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3650 CMSCollector* _collector;
3651 YieldingFlexibleWorkGang* _workers; // the whole gang
3652 int _n_workers; // requested/desired # workers
3653 bool _asynch;
3654 bool _result;
3655 CompactibleFreeListSpace* _cms_space;
3656 CompactibleFreeListSpace* _perm_space;
3657 HeapWord* _global_finger;
3658
3659 // Exposed here for yielding support
3660 Mutex* const _bit_map_lock;
3661
3662 // The per thread work queues, available here for stealing
3663 OopTaskQueueSet* _task_queues;
3664 CMSConcMarkingTerminator _term;
3665
3666 public:
3667 CMSConcMarkingTask(CMSCollector* collector,
3668 CompactibleFreeListSpace* cms_space,
3669 CompactibleFreeListSpace* perm_space,
3670 bool asynch, int n_workers,
3671 YieldingFlexibleWorkGang* workers,
3672 OopTaskQueueSet* task_queues):
3673 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3674 _collector(collector),
3675 _cms_space(cms_space),
3676 _perm_space(perm_space),
3677 _asynch(asynch), _n_workers(n_workers), _result(true),
3678 _workers(workers), _task_queues(task_queues),
3679 _term(n_workers, task_queues, _collector, asynch),
3680 _bit_map_lock(collector->bitMapLock())
3681 {
3682 assert(n_workers <= workers->total_workers(),
3683 "Else termination won't work correctly today"); // XXX FIX ME!
3684 _requested_size = n_workers;
3685 _term.set_task(this);
3686 assert(_cms_space->bottom() < _perm_space->bottom(),
3687 "Finger incorrectly initialized below");
3688 _global_finger = _cms_space->bottom();
3689 }
3690
3691
3692 OopTaskQueueSet* task_queues() { return _task_queues; }
3693
3694 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3695
3696 HeapWord** global_finger_addr() { return &_global_finger; }
3697
3698 CMSConcMarkingTerminator* terminator() { return &_term; }
3699
3700 void work(int i);
3701
3702 virtual void coordinator_yield(); // stuff done by coordinator
3703 bool result() { return _result; }
3704
3705 void reset(HeapWord* ra) {
3706 _term.reset_for_reuse();
3707 }
3708
3709 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3710 OopTaskQueue* work_q);
3711
3712 private:
3713 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3714 void do_work_steal(int i);
3715 void bump_global_finger(HeapWord* f);
3716 };
3717
3718 void CMSConcMarkingTerminator::yield() {
3719 if (ConcurrentMarkSweepThread::should_yield() &&
3720 !_collector->foregroundGCIsActive() &&
3721 _yield) {
3722 _task->yield();
3723 } else {
3724 ParallelTaskTerminator::yield();
3725 }
3726 }
3727
3728 ////////////////////////////////////////////////////////////////
3729 // Concurrent Marking Algorithm Sketch
3730 ////////////////////////////////////////////////////////////////
3731 // Until all tasks exhausted (both spaces):
3732 // -- claim next available chunk
3733 // -- bump global finger via CAS
3734 // -- find first object that starts in this chunk
3735 // and start scanning bitmap from that position
3736 // -- scan marked objects for oops
3737 // -- CAS-mark target, and if successful:
3738 // . if target oop is above global finger (volatile read)
3739 // nothing to do
3740 // . if target oop is in chunk and above local finger
3741 // then nothing to do
3742 // . else push on work-queue
3743 // -- Deal with possible overflow issues:
3744 // . local work-queue overflow causes stuff to be pushed on
3745 // global (common) overflow queue
3746 // . always first empty local work queue
3747 // . then get a batch of oops from global work queue if any
3748 // . then do work stealing
3749 // -- When all tasks claimed (both spaces)
3750 // and local work queue empty,
3751 // then in a loop do:
3752 // . check global overflow stack; steal a batch of oops and trace
3753 // . try to steal from other threads oif GOS is empty
3754 // . if neither is available, offer termination
3755 // -- Terminate and return result
3756 //
3757 void CMSConcMarkingTask::work(int i) {
3758 elapsedTimer _timer;
3759 ResourceMark rm;
3760 HandleMark hm;
3761
3762 DEBUG_ONLY(_collector->verify_overflow_empty();)
3763
3764 // Before we begin work, our work queue should be empty
3765 assert(work_queue(i)->size() == 0, "Expected to be empty");
3766 // Scan the bitmap covering _cms_space, tracing through grey objects.
3767 _timer.start();
3768 do_scan_and_mark(i, _cms_space);
3769 _timer.stop();
3770 if (PrintCMSStatistics != 0) {
3771 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3772 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3773 }
3774
3775 // ... do the same for the _perm_space
3776 _timer.reset();
3777 _timer.start();
3778 do_scan_and_mark(i, _perm_space);
3779 _timer.stop();
3780 if (PrintCMSStatistics != 0) {
3781 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3782 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3783 }
3784
3785 // ... do work stealing
3786 _timer.reset();
3787 _timer.start();
3788 do_work_steal(i);
3789 _timer.stop();
3790 if (PrintCMSStatistics != 0) {
3791 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3792 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3793 }
3794 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3795 assert(work_queue(i)->size() == 0, "Should have been emptied");
3796 // Note that under the current task protocol, the
3797 // following assertion is true even of the spaces
3798 // expanded since the completion of the concurrent
3799 // marking. XXX This will likely change under a strict
3800 // ABORT semantics.
3801 assert(_global_finger > _cms_space->end() &&
3802 _global_finger >= _perm_space->end(),
3803 "All tasks have been completed");
3804 DEBUG_ONLY(_collector->verify_overflow_empty();)
3805 }
3806
3807 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3808 HeapWord* read = _global_finger;
3809 HeapWord* cur = read;
3810 while (f > read) {
3811 cur = read;
3812 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3813 if (cur == read) {
3814 // our cas succeeded
3815 assert(_global_finger >= f, "protocol consistency");
3816 break;
3817 }
3818 }
3819 }
3820
3821 // This is really inefficient, and should be redone by
3822 // using (not yet available) block-read and -write interfaces to the
3823 // stack and the work_queue. XXX FIX ME !!!
3824 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3825 OopTaskQueue* work_q) {
3826 // Fast lock-free check
3827 if (ovflw_stk->length() == 0) {
3828 return false;
3829 }
3830 assert(work_q->size() == 0, "Shouldn't steal");
3831 MutexLockerEx ml(ovflw_stk->par_lock(),
3832 Mutex::_no_safepoint_check_flag);
3833 // Grab up to 1/4 the size of the work queue
3834 size_t num = MIN2((size_t)work_q->max_elems()/4,
3835 (size_t)ParGCDesiredObjsFromOverflowList);
3836 num = MIN2(num, ovflw_stk->length());
3837 for (int i = (int) num; i > 0; i--) {
3838 oop cur = ovflw_stk->pop();
3839 assert(cur != NULL, "Counted wrong?");
3840 work_q->push(cur);
3841 }
3842 return num > 0;
3843 }
3844
3845 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3846 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3847 int n_tasks = pst->n_tasks();
3848 // We allow that there may be no tasks to do here because
3849 // we are restarting after a stack overflow.
3850 assert(pst->valid() || n_tasks == 0, "Uninitializd use?");
3851 int nth_task = 0;
3852
3853 HeapWord* start = sp->bottom();
3854 size_t chunk_size = sp->marking_task_size();
3855 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3856 // Having claimed the nth task in this space,
3857 // compute the chunk that it corresponds to:
3858 MemRegion span = MemRegion(start + nth_task*chunk_size,
3859 start + (nth_task+1)*chunk_size);
3860 // Try and bump the global finger via a CAS;
3861 // note that we need to do the global finger bump
3862 // _before_ taking the intersection below, because
3863 // the task corresponding to that region will be
3864 // deemed done even if the used_region() expands
3865 // because of allocation -- as it almost certainly will
3866 // during start-up while the threads yield in the
3867 // closure below.
3868 HeapWord* finger = span.end();
3869 bump_global_finger(finger); // atomically
3870 // There are null tasks here corresponding to chunks
3871 // beyond the "top" address of the space.
3872 span = span.intersection(sp->used_region());
3873 if (!span.is_empty()) { // Non-null task
3874 // We want to skip the first object because
3875 // the protocol is to scan any object in its entirety
3876 // that _starts_ in this span; a fortiori, any
3877 // object starting in an earlier span is scanned
3878 // as part of an earlier claimed task.
3879 // Below we use the "careful" version of block_start
3880 // so we do not try to navigate uninitialized objects.
3881 HeapWord* prev_obj = sp->block_start_careful(span.start());
3882 // Below we use a variant of block_size that uses the
3883 // Printezis bits to avoid waiting for allocated
3884 // objects to become initialized/parsable.
3885 while (prev_obj < span.start()) {
3886 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3887 if (sz > 0) {
3888 prev_obj += sz;
3889 } else {
3890 // In this case we may end up doing a bit of redundant
3891 // scanning, but that appears unavoidable, short of
3892 // locking the free list locks; see bug 6324141.
3893 break;
3894 }
3895 }
3896 if (prev_obj < span.end()) {
3897 MemRegion my_span = MemRegion(prev_obj, span.end());
3898 // Do the marking work within a non-empty span --
3899 // the last argument to the constructor indicates whether the
3900 // iteration should be incremental with periodic yields.
3901 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3902 &_collector->_markBitMap,
3903 work_queue(i),
3904 &_collector->_markStack,
3905 &_collector->_revisitStack,
3906 _asynch);
3907 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3908 } // else nothing to do for this task
3909 } // else nothing to do for this task
3910 }
3911 // We'd be tempted to assert here that since there are no
3912 // more tasks left to claim in this space, the global_finger
3913 // must exceed space->top() and a fortiori space->end(). However,
3914 // that would not quite be correct because the bumping of
3915 // global_finger occurs strictly after the claiming of a task,
3916 // so by the time we reach here the global finger may not yet
3917 // have been bumped up by the thread that claimed the last
3918 // task.
3919 pst->all_tasks_completed();
3920 }
3921
3922 class Par_ConcMarkingClosure: public OopClosure {
3923 private:
3924 CMSCollector* _collector;
3925 MemRegion _span;
3926 CMSBitMap* _bit_map;
3927 CMSMarkStack* _overflow_stack;
3928 CMSMarkStack* _revisit_stack; // XXXXXX Check proper use
3929 OopTaskQueue* _work_queue;
3930 protected:
3931 DO_OOP_WORK_DEFN
3932 public:
3933 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
3934 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3935 _collector(collector),
3936 _span(_collector->_span),
3937 _work_queue(work_queue),
3938 _bit_map(bit_map),
3939 _overflow_stack(overflow_stack) { } // need to initialize revisit stack etc.
3940 virtual void do_oop(oop* p);
3941 virtual void do_oop(narrowOop* p);
3942 void trim_queue(size_t max);
3943 void handle_stack_overflow(HeapWord* lost);
3944 };
3945
3946 // Grey object rescan during work stealing phase --
3947 // the salient assumption here is that stolen oops must
3948 // always be initialized, so we do not need to check for
3949 // uninitialized objects before scanning here.
3950 void Par_ConcMarkingClosure::do_oop(oop obj) {
3951 assert(obj->is_oop_or_null(), "expected an oop or NULL");
3952 HeapWord* addr = (HeapWord*)obj;
3953 // Check if oop points into the CMS generation
3954 // and is not marked
3955 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3956 // a white object ...
3957 // If we manage to "claim" the object, by being the
3958 // first thread to mark it, then we push it on our
3959 // marking stack
3960 if (_bit_map->par_mark(addr)) { // ... now grey
3961 // push on work queue (grey set)
3962 bool simulate_overflow = false;
3963 NOT_PRODUCT(
3964 if (CMSMarkStackOverflowALot &&
3965 _collector->simulate_overflow()) {
3966 // simulate a stack overflow
3967 simulate_overflow = true;
3968 }
3969 )
3970 if (simulate_overflow ||
3971 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3972 // stack overflow
3973 if (PrintCMSStatistics != 0) {
3974 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3975 SIZE_FORMAT, _overflow_stack->capacity());
3976 }
3977 // We cannot assert that the overflow stack is full because
3978 // it may have been emptied since.
3979 assert(simulate_overflow ||
3980 _work_queue->size() == _work_queue->max_elems(),
3981 "Else push should have succeeded");
3982 handle_stack_overflow(addr);
3983 }
3984 } // Else, some other thread got there first
3985 }
3986 }
3987
3988 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3989 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3990
3991 void Par_ConcMarkingClosure::trim_queue(size_t max) {
3992 while (_work_queue->size() > max) {
3993 oop new_oop;
3994 if (_work_queue->pop_local(new_oop)) {
3995 assert(new_oop->is_oop(), "Should be an oop");
3996 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3997 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3998 assert(new_oop->is_parsable(), "Should be parsable");
3999 new_oop->oop_iterate(this); // do_oop() above
4000 }
4001 }
4002 }
4003
4004 // Upon stack overflow, we discard (part of) the stack,
4005 // remembering the least address amongst those discarded
4006 // in CMSCollector's _restart_address.
4007 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4008 // We need to do this under a mutex to prevent other
4009 // workers from interfering with the expansion below.
4010 MutexLockerEx ml(_overflow_stack->par_lock(),
4011 Mutex::_no_safepoint_check_flag);
4012 // Remember the least grey address discarded
4013 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4014 _collector->lower_restart_addr(ra);
4015 _overflow_stack->reset(); // discard stack contents
4016 _overflow_stack->expand(); // expand the stack if possible
4017 }
4018
4019
4020 void CMSConcMarkingTask::do_work_steal(int i) {
4021 OopTaskQueue* work_q = work_queue(i);
4022 oop obj_to_scan;
4023 CMSBitMap* bm = &(_collector->_markBitMap);
4024 CMSMarkStack* ovflw = &(_collector->_markStack);
4025 int* seed = _collector->hash_seed(i);
4026 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw);
4027 while (true) {
4028 cl.trim_queue(0);
4029 assert(work_q->size() == 0, "Should have been emptied above");
4030 if (get_work_from_overflow_stack(ovflw, work_q)) {
4031 // Can't assert below because the work obtained from the
4032 // overflow stack may already have been stolen from us.
4033 // assert(work_q->size() > 0, "Work from overflow stack");
4034 continue;
4035 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4036 assert(obj_to_scan->is_oop(), "Should be an oop");
4037 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4038 obj_to_scan->oop_iterate(&cl);
4039 } else if (terminator()->offer_termination()) {
4040 assert(work_q->size() == 0, "Impossible!");
4041 break;
4042 }
4043 }
4044 }
4045
4046 // This is run by the CMS (coordinator) thread.
4047 void CMSConcMarkingTask::coordinator_yield() {
4048 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4049 "CMS thread should hold CMS token");
4050
4051 // First give up the locks, then yield, then re-lock
4052 // We should probably use a constructor/destructor idiom to
4053 // do this unlock/lock or modify the MutexUnlocker class to
4054 // serve our purpose. XXX
4055 assert_lock_strong(_bit_map_lock);
4056 _bit_map_lock->unlock();
4057 ConcurrentMarkSweepThread::desynchronize(true);
4058 ConcurrentMarkSweepThread::acknowledge_yield_request();
4059 _collector->stopTimer();
4060 if (PrintCMSStatistics != 0) {
4061 _collector->incrementYields();
4062 }
4063 _collector->icms_wait();
4064
4065 // It is possible for whichever thread initiated the yield request
4066 // not to get a chance to wake up and take the bitmap lock between
4067 // this thread releasing it and reacquiring it. So, while the
4068 // should_yield() flag is on, let's sleep for a bit to give the
4069 // other thread a chance to wake up. The limit imposed on the number
4070 // of iterations is defensive, to avoid any unforseen circumstances
4071 // putting us into an infinite loop. Since it's always been this
4072 // (coordinator_yield()) method that was observed to cause the
4073 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4074 // which is by default non-zero. For the other seven methods that
4075 // also perform the yield operation, as are using a different
4076 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4077 // can enable the sleeping for those methods too, if necessary.
4078 // See 6442774.
4079 //
4080 // We really need to reconsider the synchronization between the GC
4081 // thread and the yield-requesting threads in the future and we
4082 // should really use wait/notify, which is the recommended
4083 // way of doing this type of interaction. Additionally, we should
4084 // consolidate the eight methods that do the yield operation and they
4085 // are almost identical into one for better maintenability and
4086 // readability. See 6445193.
4087 //
4088 // Tony 2006.06.29
4089 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4090 ConcurrentMarkSweepThread::should_yield() &&
4091 !CMSCollector::foregroundGCIsActive(); ++i) {
4092 os::sleep(Thread::current(), 1, false);
4093 ConcurrentMarkSweepThread::acknowledge_yield_request();
4094 }
4095
4096 ConcurrentMarkSweepThread::synchronize(true);
4097 _bit_map_lock->lock_without_safepoint_check();
4098 _collector->startTimer();
4099 }
4100
4101 bool CMSCollector::do_marking_mt(bool asynch) {
4102 assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition");
4103 // In the future this would be determined ergonomically, based
4104 // on #cpu's, # active mutator threads (and load), and mutation rate.
4105 int num_workers = ParallelCMSThreads;
4106
4107 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4108 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4109
4110 CMSConcMarkingTask tsk(this, cms_space, perm_space,
4111 asynch, num_workers /* number requested XXX */,
4112 conc_workers(), task_queues());
4113
4114 // Since the actual number of workers we get may be different
4115 // from the number we requested above, do we need to do anything different
4116 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4117 // class?? XXX
4118 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4119 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4120
4121 // Refs discovery is already non-atomic.
4122 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4123 // Mutate the Refs discovery so it is MT during the
4124 // multi-threaded marking phase.
4125 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4126
4127 conc_workers()->start_task(&tsk);
4128 while (tsk.yielded()) {
4129 tsk.coordinator_yield();
4130 conc_workers()->continue_task(&tsk);
4131 }
4132 // If the task was aborted, _restart_addr will be non-NULL
4133 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4134 while (_restart_addr != NULL) {
4135 // XXX For now we do not make use of ABORTED state and have not
4136 // yet implemented the right abort semantics (even in the original
4137 // single-threaded CMS case). That needs some more investigation
4138 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4139 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4140 // If _restart_addr is non-NULL, a marking stack overflow
4141 // occured; we need to do a fresh marking iteration from the
4142 // indicated restart address.
4143 if (_foregroundGCIsActive && asynch) {
4144 // We may be running into repeated stack overflows, having
4145 // reached the limit of the stack size, while making very
4146 // slow forward progress. It may be best to bail out and
4147 // let the foreground collector do its job.
4148 // Clear _restart_addr, so that foreground GC
4149 // works from scratch. This avoids the headache of
4150 // a "rescan" which would otherwise be needed because
4151 // of the dirty mod union table & card table.
4152 _restart_addr = NULL;
4153 return false;
4154 }
4155 // Adjust the task to restart from _restart_addr
4156 tsk.reset(_restart_addr);
4157 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4158 _restart_addr);
4159 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4160 _restart_addr);
4161 _restart_addr = NULL;
4162 // Get the workers going again
4163 conc_workers()->start_task(&tsk);
4164 while (tsk.yielded()) {
4165 tsk.coordinator_yield();
4166 conc_workers()->continue_task(&tsk);
4167 }
4168 }
4169 assert(tsk.completed(), "Inconsistency");
4170 assert(tsk.result() == true, "Inconsistency");
4171 return true;
4172 }
4173
4174 bool CMSCollector::do_marking_st(bool asynch) {
4175 ResourceMark rm;
4176 HandleMark hm;
4177
4178 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4179 &_markStack, &_revisitStack, CMSYield && asynch);
4180 // the last argument to iterate indicates whether the iteration
4181 // should be incremental with periodic yields.
4182 _markBitMap.iterate(&markFromRootsClosure);
4183 // If _restart_addr is non-NULL, a marking stack overflow
4184 // occured; we need to do a fresh iteration from the
4185 // indicated restart address.
4186 while (_restart_addr != NULL) {
4187 if (_foregroundGCIsActive && asynch) {
4188 // We may be running into repeated stack overflows, having
4189 // reached the limit of the stack size, while making very
4190 // slow forward progress. It may be best to bail out and
4191 // let the foreground collector do its job.
4192 // Clear _restart_addr, so that foreground GC
4193 // works from scratch. This avoids the headache of
4194 // a "rescan" which would otherwise be needed because
4195 // of the dirty mod union table & card table.
4196 _restart_addr = NULL;
4197 return false; // indicating failure to complete marking
4198 }
4199 // Deal with stack overflow:
4200 // we restart marking from _restart_addr
4201 HeapWord* ra = _restart_addr;
4202 markFromRootsClosure.reset(ra);
4203 _restart_addr = NULL;
4204 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4205 }
4206 return true;
4207 }
4208
4209 void CMSCollector::preclean() {
4210 check_correct_thread_executing();
4211 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4212 verify_work_stacks_empty();
4213 verify_overflow_empty();
4214 _abort_preclean = false;
4215 if (CMSPrecleaningEnabled) {
4216 _eden_chunk_index = 0;
4217 size_t used = get_eden_used();
4218 size_t capacity = get_eden_capacity();
4219 // Don't start sampling unless we will get sufficiently
4220 // many samples.
4221 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4222 * CMSScheduleRemarkEdenPenetration)) {
4223 _start_sampling = true;
4224 } else {
4225 _start_sampling = false;
4226 }
4227 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4228 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4229 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4230 }
4231 CMSTokenSync x(true); // is cms thread
4232 if (CMSPrecleaningEnabled) {
4233 sample_eden();
4234 _collectorState = AbortablePreclean;
4235 } else {
4236 _collectorState = FinalMarking;
4237 }
4238 verify_work_stacks_empty();
4239 verify_overflow_empty();
4240 }
4241
4242 // Try and schedule the remark such that young gen
4243 // occupancy is CMSScheduleRemarkEdenPenetration %.
4244 void CMSCollector::abortable_preclean() {
4245 check_correct_thread_executing();
4246 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4247 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4248
4249 // If Eden's current occupancy is below this threshold,
4250 // immediately schedule the remark; else preclean
4251 // past the next scavenge in an effort to
4252 // schedule the pause as described avove. By choosing
4253 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4254 // we will never do an actual abortable preclean cycle.
4255 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4256 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4257 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4258 // We need more smarts in the abortable preclean
4259 // loop below to deal with cases where allocation
4260 // in young gen is very very slow, and our precleaning
4261 // is running a losing race against a horde of
4262 // mutators intent on flooding us with CMS updates
4263 // (dirty cards).
4264 // One, admittedly dumb, strategy is to give up
4265 // after a certain number of abortable precleaning loops
4266 // or after a certain maximum time. We want to make
4267 // this smarter in the next iteration.
4268 // XXX FIX ME!!! YSR
4269 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4270 while (!(should_abort_preclean() ||
4271 ConcurrentMarkSweepThread::should_terminate())) {
4272 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4273 cumworkdone += workdone;
4274 loops++;
4275 // Voluntarily terminate abortable preclean phase if we have
4276 // been at it for too long.
4277 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4278 loops >= CMSMaxAbortablePrecleanLoops) {
4279 if (PrintGCDetails) {
4280 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4281 }
4282 break;
4283 }
4284 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4285 if (PrintGCDetails) {
4286 gclog_or_tty->print(" CMS: abort preclean due to time ");
4287 }
4288 break;
4289 }
4290 // If we are doing little work each iteration, we should
4291 // take a short break.
4292 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4293 // Sleep for some time, waiting for work to accumulate
4294 stopTimer();
4295 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4296 startTimer();
4297 waited++;
4298 }
4299 }
4300 if (PrintCMSStatistics > 0) {
4301 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4302 loops, waited, cumworkdone);
4303 }
4304 }
4305 CMSTokenSync x(true); // is cms thread
4306 if (_collectorState != Idling) {
4307 assert(_collectorState == AbortablePreclean,
4308 "Spontaneous state transition?");
4309 _collectorState = FinalMarking;
4310 } // Else, a foreground collection completed this CMS cycle.
4311 return;
4312 }
4313
4314 // Respond to an Eden sampling opportunity
4315 void CMSCollector::sample_eden() {
4316 // Make sure a young gc cannot sneak in between our
4317 // reading and recording of a sample.
4318 assert(Thread::current()->is_ConcurrentGC_thread(),
4319 "Only the cms thread may collect Eden samples");
4320 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4321 "Should collect samples while holding CMS token");
4322 if (!_start_sampling) {
4323 return;
4324 }
4325 if (_eden_chunk_array) {
4326 if (_eden_chunk_index < _eden_chunk_capacity) {
4327 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4328 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4329 "Unexpected state of Eden");
4330 // We'd like to check that what we just sampled is an oop-start address;
4331 // however, we cannot do that here since the object may not yet have been
4332 // initialized. So we'll instead do the check when we _use_ this sample
4333 // later.
4334 if (_eden_chunk_index == 0 ||
4335 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4336 _eden_chunk_array[_eden_chunk_index-1])
4337 >= CMSSamplingGrain)) {
4338 _eden_chunk_index++; // commit sample
4339 }
4340 }
4341 }
4342 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4343 size_t used = get_eden_used();
4344 size_t capacity = get_eden_capacity();
4345 assert(used <= capacity, "Unexpected state of Eden");
4346 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4347 _abort_preclean = true;
4348 }
4349 }
4350 }
4351
4352
4353 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4354 assert(_collectorState == Precleaning ||
4355 _collectorState == AbortablePreclean, "incorrect state");
4356 ResourceMark rm;
4357 HandleMark hm;
4358 // Do one pass of scrubbing the discovered reference lists
4359 // to remove any reference objects with strongly-reachable
4360 // referents.
4361 if (clean_refs) {
4362 ReferenceProcessor* rp = ref_processor();
4363 CMSPrecleanRefsYieldClosure yield_cl(this);
4364 assert(rp->span().equals(_span), "Spans should be equal");
4365 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4366 &_markStack);
4367 CMSDrainMarkingStackClosure complete_trace(this,
4368 _span, &_markBitMap, &_markStack,
4369 &keep_alive);
4370
4371 // We don't want this step to interfere with a young
4372 // collection because we don't want to take CPU
4373 // or memory bandwidth away from the young GC threads
4374 // (which may be as many as there are CPUs).
4375 // Note that we don't need to protect ourselves from
4376 // interference with mutators because they can't
4377 // manipulate the discovered reference lists nor affect
4378 // the computed reachability of the referents, the
4379 // only properties manipulated by the precleaning
4380 // of these reference lists.
4381 stopTimer();
4382 CMSTokenSyncWithLocks x(true /* is cms thread */,
4383 bitMapLock());
4384 startTimer();
4385 sample_eden();
4386 // The following will yield to allow foreground
4387 // collection to proceed promptly. XXX YSR:
4388 // The code in this method may need further
4389 // tweaking for better performance and some restructuring
4390 // for cleaner interfaces.
4391 rp->preclean_discovered_references(
4392 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4393 &yield_cl);
4394 }
4395
4396 if (clean_survivor) { // preclean the active survivor space(s)
4397 assert(_young_gen->kind() == Generation::DefNew ||
4398 _young_gen->kind() == Generation::ParNew ||
4399 _young_gen->kind() == Generation::ASParNew,
4400 "incorrect type for cast");
4401 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4402 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4403 &_markBitMap, &_modUnionTable,
4404 &_markStack, &_revisitStack,
4405 true /* precleaning phase */);
4406 stopTimer();
4407 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4408 bitMapLock());
4409 startTimer();
4410 unsigned int before_count =
4411 GenCollectedHeap::heap()->total_collections();
4412 SurvivorSpacePrecleanClosure
4413 sss_cl(this, _span, &_markBitMap, &_markStack,
4414 &pam_cl, before_count, CMSYield);
4415 dng->from()->object_iterate_careful(&sss_cl);
4416 dng->to()->object_iterate_careful(&sss_cl);
4417 }
4418 MarkRefsIntoAndScanClosure
4419 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4420 &_markStack, &_revisitStack, this, CMSYield,
4421 true /* precleaning phase */);
4422 // CAUTION: The following closure has persistent state that may need to
4423 // be reset upon a decrease in the sequence of addresses it
4424 // processes.
4425 ScanMarkedObjectsAgainCarefullyClosure
4426 smoac_cl(this, _span,
4427 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4428
4429 // Preclean dirty cards in ModUnionTable and CardTable using
4430 // appropriate convergence criterion;
4431 // repeat CMSPrecleanIter times unless we find that
4432 // we are losing.
4433 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4434 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4435 "Bad convergence multiplier");
4436 assert(CMSPrecleanThreshold >= 100,
4437 "Unreasonably low CMSPrecleanThreshold");
4438
4439 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4440 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4441 numIter < CMSPrecleanIter;
4442 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4443 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4444 if (CMSPermGenPrecleaningEnabled) {
4445 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4446 }
4447 if (Verbose && PrintGCDetails) {
4448 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4449 }
4450 // Either there are very few dirty cards, so re-mark
4451 // pause will be small anyway, or our pre-cleaning isn't
4452 // that much faster than the rate at which cards are being
4453 // dirtied, so we might as well stop and re-mark since
4454 // precleaning won't improve our re-mark time by much.
4455 if (curNumCards <= CMSPrecleanThreshold ||
4456 (numIter > 0 &&
4457 (curNumCards * CMSPrecleanDenominator >
4458 lastNumCards * CMSPrecleanNumerator))) {
4459 numIter++;
4460 cumNumCards += curNumCards;
4461 break;
4462 }
4463 }
4464 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4465 if (CMSPermGenPrecleaningEnabled) {
4466 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4467 }
4468 cumNumCards += curNumCards;
4469 if (PrintGCDetails && PrintCMSStatistics != 0) {
4470 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4471 curNumCards, cumNumCards, numIter);
4472 }
4473 return cumNumCards; // as a measure of useful work done
4474 }
4475
4476 // PRECLEANING NOTES:
4477 // Precleaning involves:
4478 // . reading the bits of the modUnionTable and clearing the set bits.
4479 // . For the cards corresponding to the set bits, we scan the
4480 // objects on those cards. This means we need the free_list_lock
4481 // so that we can safely iterate over the CMS space when scanning
4482 // for oops.
4483 // . When we scan the objects, we'll be both reading and setting
4484 // marks in the marking bit map, so we'll need the marking bit map.
4485 // . For protecting _collector_state transitions, we take the CGC_lock.
4486 // Note that any races in the reading of of card table entries by the
4487 // CMS thread on the one hand and the clearing of those entries by the
4488 // VM thread or the setting of those entries by the mutator threads on the
4489 // other are quite benign. However, for efficiency it makes sense to keep
4490 // the VM thread from racing with the CMS thread while the latter is
4491 // dirty card info to the modUnionTable. We therefore also use the
4492 // CGC_lock to protect the reading of the card table and the mod union
4493 // table by the CM thread.
4494 // . We run concurrently with mutator updates, so scanning
4495 // needs to be done carefully -- we should not try to scan
4496 // potentially uninitialized objects.
4497 //
4498 // Locking strategy: While holding the CGC_lock, we scan over and
4499 // reset a maximal dirty range of the mod union / card tables, then lock
4500 // the free_list_lock and bitmap lock to do a full marking, then
4501 // release these locks; and repeat the cycle. This allows for a
4502 // certain amount of fairness in the sharing of these locks between
4503 // the CMS collector on the one hand, and the VM thread and the
4504 // mutators on the other.
4505
4506 // NOTE: preclean_mod_union_table() and preclean_card_table()
4507 // further below are largely identical; if you need to modify
4508 // one of these methods, please check the other method too.
4509
4510 size_t CMSCollector::preclean_mod_union_table(
4511 ConcurrentMarkSweepGeneration* gen,
4512 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4513 verify_work_stacks_empty();
4514 verify_overflow_empty();
4515
4516 // strategy: starting with the first card, accumulate contiguous
4517 // ranges of dirty cards; clear these cards, then scan the region
4518 // covered by these cards.
4519
4520 // Since all of the MUT is committed ahead, we can just use
4521 // that, in case the generations expand while we are precleaning.
4522 // It might also be fine to just use the committed part of the
4523 // generation, but we might potentially miss cards when the
4524 // generation is rapidly expanding while we are in the midst
4525 // of precleaning.
4526 HeapWord* startAddr = gen->reserved().start();
4527 HeapWord* endAddr = gen->reserved().end();
4528
4529 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4530
4531 size_t numDirtyCards, cumNumDirtyCards;
4532 HeapWord *nextAddr, *lastAddr;
4533 for (cumNumDirtyCards = numDirtyCards = 0,
4534 nextAddr = lastAddr = startAddr;
4535 nextAddr < endAddr;
4536 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4537
4538 ResourceMark rm;
4539 HandleMark hm;
4540
4541 MemRegion dirtyRegion;
4542 {
4543 stopTimer();
4544 CMSTokenSync ts(true);
4545 startTimer();
4546 sample_eden();
4547 // Get dirty region starting at nextOffset (inclusive),
4548 // simultaneously clearing it.
4549 dirtyRegion =
4550 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4551 assert(dirtyRegion.start() >= nextAddr,
4552 "returned region inconsistent?");
4553 }
4554 // Remember where the next search should begin.
4555 // The returned region (if non-empty) is a right open interval,
4556 // so lastOffset is obtained from the right end of that
4557 // interval.
4558 lastAddr = dirtyRegion.end();
4559 // Should do something more transparent and less hacky XXX
4560 numDirtyCards =
4561 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4562
4563 // We'll scan the cards in the dirty region (with periodic
4564 // yields for foreground GC as needed).
4565 if (!dirtyRegion.is_empty()) {
4566 assert(numDirtyCards > 0, "consistency check");
4567 HeapWord* stop_point = NULL;
4568 {
4569 stopTimer();
4570 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4571 bitMapLock());
4572 startTimer();
4573 verify_work_stacks_empty();
4574 verify_overflow_empty();
4575 sample_eden();
4576 stop_point =
4577 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4578 }
4579 if (stop_point != NULL) {
4580 // The careful iteration stopped early either because it found an
4581 // uninitialized object, or because we were in the midst of an
4582 // "abortable preclean", which should now be aborted. Redirty
4583 // the bits corresponding to the partially-scanned or unscanned
4584 // cards. We'll either restart at the next block boundary or
4585 // abort the preclean.
4586 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4587 (_collectorState == AbortablePreclean && should_abort_preclean()),
4588 "Unparsable objects should only be in perm gen.");
4589
4590 stopTimer();
4591 CMSTokenSyncWithLocks ts(true, bitMapLock());
4592 startTimer();
4593 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4594 if (should_abort_preclean()) {
4595 break; // out of preclean loop
4596 } else {
4597 // Compute the next address at which preclean should pick up;
4598 // might need bitMapLock in order to read P-bits.
4599 lastAddr = next_card_start_after_block(stop_point);
4600 }
4601 }
4602 } else {
4603 assert(lastAddr == endAddr, "consistency check");
4604 assert(numDirtyCards == 0, "consistency check");
4605 break;
4606 }
4607 }
4608 verify_work_stacks_empty();
4609 verify_overflow_empty();
4610 return cumNumDirtyCards;
4611 }
4612
4613 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4614 // below are largely identical; if you need to modify
4615 // one of these methods, please check the other method too.
4616
4617 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4618 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4619 // strategy: it's similar to precleamModUnionTable above, in that
4620 // we accumulate contiguous ranges of dirty cards, mark these cards
4621 // precleaned, then scan the region covered by these cards.
4622 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4623 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4624
4625 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4626
4627 size_t numDirtyCards, cumNumDirtyCards;
4628 HeapWord *lastAddr, *nextAddr;
4629
4630 for (cumNumDirtyCards = numDirtyCards = 0,
4631 nextAddr = lastAddr = startAddr;
4632 nextAddr < endAddr;
4633 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4634
4635 ResourceMark rm;
4636 HandleMark hm;
4637
4638 MemRegion dirtyRegion;
4639 {
4640 // See comments in "Precleaning notes" above on why we
4641 // do this locking. XXX Could the locking overheads be
4642 // too high when dirty cards are sparse? [I don't think so.]
4643 stopTimer();
4644 CMSTokenSync x(true); // is cms thread
4645 startTimer();
4646 sample_eden();
4647 // Get and clear dirty region from card table
4648 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_preclean(
4649 MemRegion(nextAddr, endAddr));
4650 assert(dirtyRegion.start() >= nextAddr,
4651 "returned region inconsistent?");
4652 }
4653 lastAddr = dirtyRegion.end();
4654 numDirtyCards =
4655 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4656
4657 if (!dirtyRegion.is_empty()) {
4658 stopTimer();
4659 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4660 startTimer();
4661 sample_eden();
4662 verify_work_stacks_empty();
4663 verify_overflow_empty();
4664 HeapWord* stop_point =
4665 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4666 if (stop_point != NULL) {
4667 // The careful iteration stopped early because it found an
4668 // uninitialized object. Redirty the bits corresponding to the
4669 // partially-scanned or unscanned cards, and start again at the
4670 // next block boundary.
4671 assert(CMSPermGenPrecleaningEnabled ||
4672 (_collectorState == AbortablePreclean && should_abort_preclean()),
4673 "Unparsable objects should only be in perm gen.");
4674 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4675 if (should_abort_preclean()) {
4676 break; // out of preclean loop
4677 } else {
4678 // Compute the next address at which preclean should pick up.
4679 lastAddr = next_card_start_after_block(stop_point);
4680 }
4681 }
4682 } else {
4683 break;
4684 }
4685 }
4686 verify_work_stacks_empty();
4687 verify_overflow_empty();
4688 return cumNumDirtyCards;
4689 }
4690
4691 void CMSCollector::checkpointRootsFinal(bool asynch,
4692 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4693 assert(_collectorState == FinalMarking, "incorrect state transition?");
4694 check_correct_thread_executing();
4695 // world is stopped at this checkpoint
4696 assert(SafepointSynchronize::is_at_safepoint(),
4697 "world should be stopped");
4698 verify_work_stacks_empty();
4699 verify_overflow_empty();
4700
4701 SpecializationStats::clear();
4702 if (PrintGCDetails) {
4703 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4704 _young_gen->used() / K,
4705 _young_gen->capacity() / K);
4706 }
4707 if (asynch) {
4708 if (CMSScavengeBeforeRemark) {
4709 GenCollectedHeap* gch = GenCollectedHeap::heap();
4710 // Temporarily set flag to false, GCH->do_collection will
4711 // expect it to be false and set to true
4712 FlagSetting fl(gch->_is_gc_active, false);
4713 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4714 PrintGCDetails && Verbose, true, gclog_or_tty);)
4715 int level = _cmsGen->level() - 1;
4716 if (level >= 0) {
4717 gch->do_collection(true, // full (i.e. force, see below)
4718 false, // !clear_all_soft_refs
4719 0, // size
4720 false, // is_tlab
4721 level // max_level
4722 );
4723 }
4724 }
4725 FreelistLocker x(this);
4726 MutexLockerEx y(bitMapLock(),
4727 Mutex::_no_safepoint_check_flag);
4728 assert(!init_mark_was_synchronous, "but that's impossible!");
4729 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4730 } else {
4731 // already have all the locks
4732 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4733 init_mark_was_synchronous);
4734 }
4735 verify_work_stacks_empty();
4736 verify_overflow_empty();
4737 SpecializationStats::print();
4738 }
4739
4740 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4741 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4742
4743 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4744
4745 assert(haveFreelistLocks(), "must have free list locks");
4746 assert_lock_strong(bitMapLock());
4747
4748 if (UseAdaptiveSizePolicy) {
4749 size_policy()->checkpoint_roots_final_begin();
4750 }
4751
4752 ResourceMark rm;
4753 HandleMark hm;
4754
4755 GenCollectedHeap* gch = GenCollectedHeap::heap();
4756
4757 if (should_unload_classes()) {
4758 CodeCache::gc_prologue();
4759 }
4760 assert(haveFreelistLocks(), "must have free list locks");
4761 assert_lock_strong(bitMapLock());
4762
4763 if (!init_mark_was_synchronous) {
4764 // We might assume that we need not fill TLAB's when
4765 // CMSScavengeBeforeRemark is set, because we may have just done
4766 // a scavenge which would have filled all TLAB's -- and besides
4767 // Eden would be empty. This however may not always be the case --
4768 // for instance although we asked for a scavenge, it may not have
4769 // happened because of a JNI critical section. We probably need
4770 // a policy for deciding whether we can in that case wait until
4771 // the critical section releases and then do the remark following
4772 // the scavenge, and skip it here. In the absence of that policy,
4773 // or of an indication of whether the scavenge did indeed occur,
4774 // we cannot rely on TLAB's having been filled and must do
4775 // so here just in case a scavenge did not happen.
4776 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4777 // Update the saved marks which may affect the root scans.
4778 gch->save_marks();
4779
4780 {
4781 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4782
4783 // Note on the role of the mod union table:
4784 // Since the marker in "markFromRoots" marks concurrently with
4785 // mutators, it is possible for some reachable objects not to have been
4786 // scanned. For instance, an only reference to an object A was
4787 // placed in object B after the marker scanned B. Unless B is rescanned,
4788 // A would be collected. Such updates to references in marked objects
4789 // are detected via the mod union table which is the set of all cards
4790 // dirtied since the first checkpoint in this GC cycle and prior to
4791 // the most recent young generation GC, minus those cleaned up by the
4792 // concurrent precleaning.
4793 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
4794 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4795 do_remark_parallel();
4796 } else {
4797 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4798 gclog_or_tty);
4799 do_remark_non_parallel();
4800 }
4801 }
4802 } else {
4803 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4804 // The initial mark was stop-world, so there's no rescanning to
4805 // do; go straight on to the next step below.
4806 }
4807 verify_work_stacks_empty();
4808 verify_overflow_empty();
4809
4810 {
4811 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4812 refProcessingWork(asynch, clear_all_soft_refs);
4813 }
4814 verify_work_stacks_empty();
4815 verify_overflow_empty();
4816
4817 if (should_unload_classes()) {
4818 CodeCache::gc_epilogue();
4819 }
4820
4821 // If we encountered any (marking stack / work queue) overflow
4822 // events during the current CMS cycle, take appropriate
4823 // remedial measures, where possible, so as to try and avoid
4824 // recurrence of that condition.
4825 assert(_markStack.isEmpty(), "No grey objects");
4826 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4827 _ser_kac_ovflw;
4828 if (ser_ovflw > 0) {
4829 if (PrintCMSStatistics != 0) {
4830 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4831 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4832 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4833 _ser_kac_ovflw);
4834 }
4835 _markStack.expand();
4836 _ser_pmc_remark_ovflw = 0;
4837 _ser_pmc_preclean_ovflw = 0;
4838 _ser_kac_ovflw = 0;
4839 }
4840 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4841 if (PrintCMSStatistics != 0) {
4842 gclog_or_tty->print_cr("Work queue overflow (benign) "
4843 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4844 _par_pmc_remark_ovflw, _par_kac_ovflw);
4845 }
4846 _par_pmc_remark_ovflw = 0;
4847 _par_kac_ovflw = 0;
4848 }
4849 if (PrintCMSStatistics != 0) {
4850 if (_markStack._hit_limit > 0) {
4851 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4852 _markStack._hit_limit);
4853 }
4854 if (_markStack._failed_double > 0) {
4855 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4856 " current capacity "SIZE_FORMAT,
4857 _markStack._failed_double,
4858 _markStack.capacity());
4859 }
4860 }
4861 _markStack._hit_limit = 0;
4862 _markStack._failed_double = 0;
4863
4864 if ((VerifyAfterGC || VerifyDuringGC) &&
4865 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4866 verify_after_remark();
4867 }
4868
4869 // Change under the freelistLocks.
4870 _collectorState = Sweeping;
4871 // Call isAllClear() under bitMapLock
4872 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
4873 " final marking");
4874 if (UseAdaptiveSizePolicy) {
4875 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
4876 }
4877 }
4878
4879 // Parallel remark task
4880 class CMSParRemarkTask: public AbstractGangTask {
4881 CMSCollector* _collector;
4882 WorkGang* _workers;
4883 int _n_workers;
4884 CompactibleFreeListSpace* _cms_space;
4885 CompactibleFreeListSpace* _perm_space;
4886
4887 // The per-thread work queues, available here for stealing.
4888 OopTaskQueueSet* _task_queues;
4889 ParallelTaskTerminator _term;
4890
4891 public:
4892 CMSParRemarkTask(CMSCollector* collector,
4893 CompactibleFreeListSpace* cms_space,
4894 CompactibleFreeListSpace* perm_space,
4895 int n_workers, WorkGang* workers,
4896 OopTaskQueueSet* task_queues):
4897 AbstractGangTask("Rescan roots and grey objects in parallel"),
4898 _collector(collector),
4899 _cms_space(cms_space), _perm_space(perm_space),
4900 _n_workers(n_workers),
4901 _workers(workers),
4902 _task_queues(task_queues),
4903 _term(workers->total_workers(), task_queues) { }
4904
4905 OopTaskQueueSet* task_queues() { return _task_queues; }
4906
4907 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4908
4909 ParallelTaskTerminator* terminator() { return &_term; }
4910
4911 void work(int i);
4912
4913 private:
4914 // Work method in support of parallel rescan ... of young gen spaces
4915 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
4916 ContiguousSpace* space,
4917 HeapWord** chunk_array, size_t chunk_top);
4918
4919 // ... of dirty cards in old space
4920 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4921 Par_MarkRefsIntoAndScanClosure* cl);
4922
4923 // ... work stealing for the above
4924 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4925 };
4926
4927 void CMSParRemarkTask::work(int i) {
4928 elapsedTimer _timer;
4929 ResourceMark rm;
4930 HandleMark hm;
4931
4932 // ---------- rescan from roots --------------
4933 _timer.start();
4934 GenCollectedHeap* gch = GenCollectedHeap::heap();
4935 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4936 _collector->_span, _collector->ref_processor(),
4937 &(_collector->_markBitMap),
4938 work_queue(i), &(_collector->_revisitStack));
4939
4940 // Rescan young gen roots first since these are likely
4941 // coarsely partitioned and may, on that account, constitute
4942 // the critical path; thus, it's best to start off that
4943 // work first.
4944 // ---------- young gen roots --------------
4945 {
4946 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
4947 EdenSpace* eden_space = dng->eden();
4948 ContiguousSpace* from_space = dng->from();
4949 ContiguousSpace* to_space = dng->to();
4950
4951 HeapWord** eca = _collector->_eden_chunk_array;
4952 size_t ect = _collector->_eden_chunk_index;
4953 HeapWord** sca = _collector->_survivor_chunk_array;
4954 size_t sct = _collector->_survivor_chunk_index;
4955
4956 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4957 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4958
4959 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
4960 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
4961 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
4962
4963 _timer.stop();
4964 if (PrintCMSStatistics != 0) {
4965 gclog_or_tty->print_cr(
4966 "Finished young gen rescan work in %dth thread: %3.3f sec",
4967 i, _timer.seconds());
4968 }
4969 }
4970
4971 // ---------- remaining roots --------------
4972 _timer.reset();
4973 _timer.start();
4974 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
4975 false, // yg was scanned above
4976 true, // collecting perm gen
4977 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4978 NULL, &par_mrias_cl);
4979 _timer.stop();
4980 if (PrintCMSStatistics != 0) {
4981 gclog_or_tty->print_cr(
4982 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4983 i, _timer.seconds());
4984 }
4985
4986 // ---------- rescan dirty cards ------------
4987 _timer.reset();
4988 _timer.start();
4989
4990 // Do the rescan tasks for each of the two spaces
4991 // (cms_space and perm_space) in turn.
4992 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
4993 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
4994 _timer.stop();
4995 if (PrintCMSStatistics != 0) {
4996 gclog_or_tty->print_cr(
4997 "Finished dirty card rescan work in %dth thread: %3.3f sec",
4998 i, _timer.seconds());
4999 }
5000
5001 // ---------- steal work from other threads ...
5002 // ---------- ... and drain overflow list.
5003 _timer.reset();
5004 _timer.start();
5005 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5006 _timer.stop();
5007 if (PrintCMSStatistics != 0) {
5008 gclog_or_tty->print_cr(
5009 "Finished work stealing in %dth thread: %3.3f sec",
5010 i, _timer.seconds());
5011 }
5012 }
5013
5014 void
5015 CMSParRemarkTask::do_young_space_rescan(int i,
5016 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5017 HeapWord** chunk_array, size_t chunk_top) {
5018 // Until all tasks completed:
5019 // . claim an unclaimed task
5020 // . compute region boundaries corresponding to task claimed
5021 // using chunk_array
5022 // . par_oop_iterate(cl) over that region
5023
5024 ResourceMark rm;
5025 HandleMark hm;
5026
5027 SequentialSubTasksDone* pst = space->par_seq_tasks();
5028 assert(pst->valid(), "Uninitialized use?");
5029
5030 int nth_task = 0;
5031 int n_tasks = pst->n_tasks();
5032
5033 HeapWord *start, *end;
5034 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5035 // We claimed task # nth_task; compute its boundaries.
5036 if (chunk_top == 0) { // no samples were taken
5037 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5038 start = space->bottom();
5039 end = space->top();
5040 } else if (nth_task == 0) {
5041 start = space->bottom();
5042 end = chunk_array[nth_task];
5043 } else if (nth_task < (jint)chunk_top) {
5044 assert(nth_task >= 1, "Control point invariant");
5045 start = chunk_array[nth_task - 1];
5046 end = chunk_array[nth_task];
5047 } else {
5048 assert(nth_task == (jint)chunk_top, "Control point invariant");
5049 start = chunk_array[chunk_top - 1];
5050 end = space->top();
5051 }
5052 MemRegion mr(start, end);
5053 // Verify that mr is in space
5054 assert(mr.is_empty() || space->used_region().contains(mr),
5055 "Should be in space");
5056 // Verify that "start" is an object boundary
5057 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5058 "Should be an oop");
5059 space->par_oop_iterate(mr, cl);
5060 }
5061 pst->all_tasks_completed();
5062 }
5063
5064 void
5065 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5066 CompactibleFreeListSpace* sp, int i,
5067 Par_MarkRefsIntoAndScanClosure* cl) {
5068 // Until all tasks completed:
5069 // . claim an unclaimed task
5070 // . compute region boundaries corresponding to task claimed
5071 // . transfer dirty bits ct->mut for that region
5072 // . apply rescanclosure to dirty mut bits for that region
5073
5074 ResourceMark rm;
5075 HandleMark hm;
5076
5077 OopTaskQueue* work_q = work_queue(i);
5078 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5079 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5080 // CAUTION: This closure has state that persists across calls to
5081 // the work method dirty_range_iterate_clear() in that it has
5082 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5083 // use of that state in the imbedded UpwardsObjectClosure instance
5084 // assumes that the cards are always iterated (even if in parallel
5085 // by several threads) in monotonically increasing order per each
5086 // thread. This is true of the implementation below which picks
5087 // card ranges (chunks) in monotonically increasing order globally
5088 // and, a-fortiori, in monotonically increasing order per thread
5089 // (the latter order being a subsequence of the former).
5090 // If the work code below is ever reorganized into a more chaotic
5091 // work-partitioning form than the current "sequential tasks"
5092 // paradigm, the use of that persistent state will have to be
5093 // revisited and modified appropriately. See also related
5094 // bug 4756801 work on which should examine this code to make
5095 // sure that the changes there do not run counter to the
5096 // assumptions made here and necessary for correctness and
5097 // efficiency. Note also that this code might yield inefficient
5098 // behaviour in the case of very large objects that span one or
5099 // more work chunks. Such objects would potentially be scanned
5100 // several times redundantly. Work on 4756801 should try and
5101 // address that performance anomaly if at all possible. XXX
5102 MemRegion full_span = _collector->_span;
5103 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5104 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5105 MarkFromDirtyCardsClosure
5106 greyRescanClosure(_collector, full_span, // entire span of interest
5107 sp, bm, work_q, rs, cl);
5108
5109 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5110 assert(pst->valid(), "Uninitialized use?");
5111 int nth_task = 0;
5112 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5113 MemRegion span = sp->used_region();
5114 HeapWord* start_addr = span.start();
5115 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5116 alignment);
5117 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5118 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5119 start_addr, "Check alignment");
5120 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5121 chunk_size, "Check alignment");
5122
5123 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5124 // Having claimed the nth_task, compute corresponding mem-region,
5125 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5126 // The alignment restriction ensures that we do not need any
5127 // synchronization with other gang-workers while setting or
5128 // clearing bits in thus chunk of the MUT.
5129 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5130 start_addr + (nth_task+1)*chunk_size);
5131 // The last chunk's end might be way beyond end of the
5132 // used region. In that case pull back appropriately.
5133 if (this_span.end() > end_addr) {
5134 this_span.set_end(end_addr);
5135 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5136 }
5137 // Iterate over the dirty cards covering this chunk, marking them
5138 // precleaned, and setting the corresponding bits in the mod union
5139 // table. Since we have been careful to partition at Card and MUT-word
5140 // boundaries no synchronization is needed between parallel threads.
5141 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5142 &modUnionClosure);
5143
5144 // Having transferred these marks into the modUnionTable,
5145 // rescan the marked objects on the dirty cards in the modUnionTable.
5146 // Even if this is at a synchronous collection, the initial marking
5147 // may have been done during an asynchronous collection so there
5148 // may be dirty bits in the mod-union table.
5149 _collector->_modUnionTable.dirty_range_iterate_clear(
5150 this_span, &greyRescanClosure);
5151 _collector->_modUnionTable.verifyNoOneBitsInRange(
5152 this_span.start(),
5153 this_span.end());
5154 }
5155 pst->all_tasks_completed(); // declare that i am done
5156 }
5157
5158 // . see if we can share work_queues with ParNew? XXX
5159 void
5160 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5161 int* seed) {
5162 OopTaskQueue* work_q = work_queue(i);
5163 NOT_PRODUCT(int num_steals = 0;)
5164 oop obj_to_scan;
5165 CMSBitMap* bm = &(_collector->_markBitMap);
5166 size_t num_from_overflow_list =
5167 MIN2((size_t)work_q->max_elems()/4,
5168 (size_t)ParGCDesiredObjsFromOverflowList);
5169
5170 while (true) {
5171 // Completely finish any left over work from (an) earlier round(s)
5172 cl->trim_queue(0);
5173 // Now check if there's any work in the overflow list
5174 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5175 work_q)) {
5176 // found something in global overflow list;
5177 // not yet ready to go stealing work from others.
5178 // We'd like to assert(work_q->size() != 0, ...)
5179 // because we just took work from the overflow list,
5180 // but of course we can't since all of that could have
5181 // been already stolen from us.
5182 // "He giveth and He taketh away."
5183 continue;
5184 }
5185 // Verify that we have no work before we resort to stealing
5186 assert(work_q->size() == 0, "Have work, shouldn't steal");
5187 // Try to steal from other queues that have work
5188 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5189 NOT_PRODUCT(num_steals++;)
5190 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5191 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5192 // Do scanning work
5193 obj_to_scan->oop_iterate(cl);
5194 // Loop around, finish this work, and try to steal some more
5195 } else if (terminator()->offer_termination()) {
5196 break; // nirvana from the infinite cycle
5197 }
5198 }
5199 NOT_PRODUCT(
5200 if (PrintCMSStatistics != 0) {
5201 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5202 }
5203 )
5204 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5205 "Else our work is not yet done");
5206 }
5207
5208 // Return a thread-local PLAB recording array, as appropriate.
5209 void* CMSCollector::get_data_recorder(int thr_num) {
5210 if (_survivor_plab_array != NULL &&
5211 (CMSPLABRecordAlways ||
5212 (_collectorState > Marking && _collectorState < FinalMarking))) {
5213 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5214 ChunkArray* ca = &_survivor_plab_array[thr_num];
5215 ca->reset(); // clear it so that fresh data is recorded
5216 return (void*) ca;
5217 } else {
5218 return NULL;
5219 }
5220 }
5221
5222 // Reset all the thread-local PLAB recording arrays
5223 void CMSCollector::reset_survivor_plab_arrays() {
5224 for (uint i = 0; i < ParallelGCThreads; i++) {
5225 _survivor_plab_array[i].reset();
5226 }
5227 }
5228
5229 // Merge the per-thread plab arrays into the global survivor chunk
5230 // array which will provide the partitioning of the survivor space
5231 // for CMS rescan.
5232 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
5233 assert(_survivor_plab_array != NULL, "Error");
5234 assert(_survivor_chunk_array != NULL, "Error");
5235 assert(_collectorState == FinalMarking, "Error");
5236 for (uint j = 0; j < ParallelGCThreads; j++) {
5237 _cursor[j] = 0;
5238 }
5239 HeapWord* top = surv->top();
5240 size_t i;
5241 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5242 HeapWord* min_val = top; // Higher than any PLAB address
5243 uint min_tid = 0; // position of min_val this round
5244 for (uint j = 0; j < ParallelGCThreads; j++) {
5245 ChunkArray* cur_sca = &_survivor_plab_array[j];
5246 if (_cursor[j] == cur_sca->end()) {
5247 continue;
5248 }
5249 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5250 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5251 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5252 if (cur_val < min_val) {
5253 min_tid = j;
5254 min_val = cur_val;
5255 } else {
5256 assert(cur_val < top, "All recorded addresses should be less");
5257 }
5258 }
5259 // At this point min_val and min_tid are respectively
5260 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5261 // and the thread (j) that witnesses that address.
5262 // We record this address in the _survivor_chunk_array[i]
5263 // and increment _cursor[min_tid] prior to the next round i.
5264 if (min_val == top) {
5265 break;
5266 }
5267 _survivor_chunk_array[i] = min_val;
5268 _cursor[min_tid]++;
5269 }
5270 // We are all done; record the size of the _survivor_chunk_array
5271 _survivor_chunk_index = i; // exclusive: [0, i)
5272 if (PrintCMSStatistics > 0) {
5273 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5274 }
5275 // Verify that we used up all the recorded entries
5276 #ifdef ASSERT
5277 size_t total = 0;
5278 for (uint j = 0; j < ParallelGCThreads; j++) {
5279 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5280 total += _cursor[j];
5281 }
5282 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5283 // Check that the merged array is in sorted order
5284 if (total > 0) {
5285 for (size_t i = 0; i < total - 1; i++) {
5286 if (PrintCMSStatistics > 0) {
5287 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5288 i, _survivor_chunk_array[i]);
5289 }
5290 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5291 "Not sorted");
5292 }
5293 }
5294 #endif // ASSERT
5295 }
5296
5297 // Set up the space's par_seq_tasks structure for work claiming
5298 // for parallel rescan of young gen.
5299 // See ParRescanTask where this is currently used.
5300 void
5301 CMSCollector::
5302 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5303 assert(n_threads > 0, "Unexpected n_threads argument");
5304 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5305
5306 // Eden space
5307 {
5308 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5309 assert(!pst->valid(), "Clobbering existing data?");
5310 // Each valid entry in [0, _eden_chunk_index) represents a task.
5311 size_t n_tasks = _eden_chunk_index + 1;
5312 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5313 pst->set_par_threads(n_threads);
5314 pst->set_n_tasks((int)n_tasks);
5315 }
5316
5317 // Merge the survivor plab arrays into _survivor_chunk_array
5318 if (_survivor_plab_array != NULL) {
5319 merge_survivor_plab_arrays(dng->from());
5320 } else {
5321 assert(_survivor_chunk_index == 0, "Error");
5322 }
5323
5324 // To space
5325 {
5326 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5327 assert(!pst->valid(), "Clobbering existing data?");
5328 pst->set_par_threads(n_threads);
5329 pst->set_n_tasks(1);
5330 assert(pst->valid(), "Error");
5331 }
5332
5333 // From space
5334 {
5335 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5336 assert(!pst->valid(), "Clobbering existing data?");
5337 size_t n_tasks = _survivor_chunk_index + 1;
5338 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5339 pst->set_par_threads(n_threads);
5340 pst->set_n_tasks((int)n_tasks);
5341 assert(pst->valid(), "Error");
5342 }
5343 }
5344
5345 // Parallel version of remark
5346 void CMSCollector::do_remark_parallel() {
5347 GenCollectedHeap* gch = GenCollectedHeap::heap();
5348 WorkGang* workers = gch->workers();
5349 assert(workers != NULL, "Need parallel worker threads.");
5350 int n_workers = workers->total_workers();
5351 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5352 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5353
5354 CMSParRemarkTask tsk(this,
5355 cms_space, perm_space,
5356 n_workers, workers, task_queues());
5357
5358 // Set up for parallel process_strong_roots work.
5359 gch->set_par_threads(n_workers);
5360 gch->change_strong_roots_parity();
5361 // We won't be iterating over the cards in the card table updating
5362 // the younger_gen cards, so we shouldn't call the following else
5363 // the verification code as well as subsequent younger_refs_iterate
5364 // code would get confused. XXX
5365 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5366
5367 // The young gen rescan work will not be done as part of
5368 // process_strong_roots (which currently doesn't knw how to
5369 // parallelize such a scan), but rather will be broken up into
5370 // a set of parallel tasks (via the sampling that the [abortable]
5371 // preclean phase did of EdenSpace, plus the [two] tasks of
5372 // scanning the [two] survivor spaces. Further fine-grain
5373 // parallelization of the scanning of the survivor spaces
5374 // themselves, and of precleaning of the younger gen itself
5375 // is deferred to the future.
5376 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5377
5378 // The dirty card rescan work is broken up into a "sequence"
5379 // of parallel tasks (per constituent space) that are dynamically
5380 // claimed by the parallel threads.
5381 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5382 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5383
5384 // It turns out that even when we're using 1 thread, doing the work in a
5385 // separate thread causes wide variance in run times. We can't help this
5386 // in the multi-threaded case, but we special-case n=1 here to get
5387 // repeatable measurements of the 1-thread overhead of the parallel code.
5388 if (n_workers > 1) {
5389 // Make refs discovery MT-safe
5390 ReferenceProcessorMTMutator mt(ref_processor(), true);
5391 workers->run_task(&tsk);
5392 } else {
5393 tsk.work(0);
5394 }
5395 gch->set_par_threads(0); // 0 ==> non-parallel.
5396 // restore, single-threaded for now, any preserved marks
5397 // as a result of work_q overflow
5398 restore_preserved_marks_if_any();
5399 }
5400
5401 // Non-parallel version of remark
5402 void CMSCollector::do_remark_non_parallel() {
5403 ResourceMark rm;
5404 HandleMark hm;
5405 GenCollectedHeap* gch = GenCollectedHeap::heap();
5406 MarkRefsIntoAndScanClosure
5407 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5408 &_markStack, &_revisitStack, this,
5409 false /* should_yield */, false /* not precleaning */);
5410 MarkFromDirtyCardsClosure
5411 markFromDirtyCardsClosure(this, _span,
5412 NULL, // space is set further below
5413 &_markBitMap, &_markStack, &_revisitStack,
5414 &mrias_cl);
5415 {
5416 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5417 // Iterate over the dirty cards, marking them precleaned, and
5418 // setting the corresponding bits in the mod union table.
5419 {
5420 ModUnionClosure modUnionClosure(&_modUnionTable);
5421 _ct->ct_bs()->dirty_card_iterate(
5422 _cmsGen->used_region(),
5423 &modUnionClosure);
5424 _ct->ct_bs()->dirty_card_iterate(
5425 _permGen->used_region(),
5426 &modUnionClosure);
5427 }
5428 // Having transferred these marks into the modUnionTable, we just need
5429 // to rescan the marked objects on the dirty cards in the modUnionTable.
5430 // The initial marking may have been done during an asynchronous
5431 // collection so there may be dirty bits in the mod-union table.
5432 const int alignment =
5433 CardTableModRefBS::card_size * BitsPerWord;
5434 {
5435 // ... First handle dirty cards in CMS gen
5436 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5437 MemRegion ur = _cmsGen->used_region();
5438 HeapWord* lb = ur.start();
5439 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5440 MemRegion cms_span(lb, ub);
5441 _modUnionTable.dirty_range_iterate_clear(cms_span,
5442 &markFromDirtyCardsClosure);
5443 verify_work_stacks_empty();
5444 if (PrintCMSStatistics != 0) {
5445 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5446 markFromDirtyCardsClosure.num_dirty_cards());
5447 }
5448 }
5449 {
5450 // .. and then repeat for dirty cards in perm gen
5451 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5452 MemRegion ur = _permGen->used_region();
5453 HeapWord* lb = ur.start();
5454 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5455 MemRegion perm_span(lb, ub);
5456 _modUnionTable.dirty_range_iterate_clear(perm_span,
5457 &markFromDirtyCardsClosure);
5458 verify_work_stacks_empty();
5459 if (PrintCMSStatistics != 0) {
5460 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5461 markFromDirtyCardsClosure.num_dirty_cards());
5462 }
5463 }
5464 }
5465 if (VerifyDuringGC &&
5466 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5467 HandleMark hm; // Discard invalid handles created during verification
5468 Universe::verify(true);
5469 }
5470 {
5471 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5472
5473 verify_work_stacks_empty();
5474
5475 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5476 gch->gen_process_strong_roots(_cmsGen->level(),
5477 true, // younger gens as roots
5478 true, // collecting perm gen
5479 SharedHeap::ScanningOption(roots_scanning_options()),
5480 NULL, &mrias_cl);
5481 }
5482 verify_work_stacks_empty();
5483 // Restore evacuated mark words, if any, used for overflow list links
5484 if (!CMSOverflowEarlyRestoration) {
5485 restore_preserved_marks_if_any();
5486 }
5487 verify_overflow_empty();
5488 }
5489
5490 ////////////////////////////////////////////////////////
5491 // Parallel Reference Processing Task Proxy Class
5492 ////////////////////////////////////////////////////////
5493 class CMSRefProcTaskProxy: public AbstractGangTask {
5494 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5495 CMSCollector* _collector;
5496 CMSBitMap* _mark_bit_map;
5497 const MemRegion _span;
5498 OopTaskQueueSet* _task_queues;
5499 ParallelTaskTerminator _term;
5500 ProcessTask& _task;
5501
5502 public:
5503 CMSRefProcTaskProxy(ProcessTask& task,
5504 CMSCollector* collector,
5505 const MemRegion& span,
5506 CMSBitMap* mark_bit_map,
5507 int total_workers,
5508 OopTaskQueueSet* task_queues):
5509 AbstractGangTask("Process referents by policy in parallel"),
5510 _task(task),
5511 _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
5512 _task_queues(task_queues),
5513 _term(total_workers, task_queues)
5514 {
5515 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5516 "Inconsistency in _span");
5517 }
5518
5519 OopTaskQueueSet* task_queues() { return _task_queues; }
5520
5521 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5522
5523 ParallelTaskTerminator* terminator() { return &_term; }
5524
5525 void do_work_steal(int i,
5526 CMSParDrainMarkingStackClosure* drain,
5527 CMSParKeepAliveClosure* keep_alive,
5528 int* seed);
5529
5530 virtual void work(int i);
5531 };
5532
5533 void CMSRefProcTaskProxy::work(int i) {
5534 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5535 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5536 _mark_bit_map, work_queue(i));
5537 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5538 _mark_bit_map, work_queue(i));
5539 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5540 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5541 if (_task.marks_oops_alive()) {
5542 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5543 _collector->hash_seed(i));
5544 }
5545 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5546 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5547 }
5548
5549 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5550 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5551 EnqueueTask& _task;
5552
5553 public:
5554 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5555 : AbstractGangTask("Enqueue reference objects in parallel"),
5556 _task(task)
5557 { }
5558
5559 virtual void work(int i)
5560 {
5561 _task.work(i);
5562 }
5563 };
5564
5565 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5566 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5567 _collector(collector),
5568 _span(span),
5569 _bit_map(bit_map),
5570 _work_queue(work_queue),
5571 _mark_and_push(collector, span, bit_map, work_queue),
5572 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5573 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5574 { }
5575
5576 // . see if we can share work_queues with ParNew? XXX
5577 void CMSRefProcTaskProxy::do_work_steal(int i,
5578 CMSParDrainMarkingStackClosure* drain,
5579 CMSParKeepAliveClosure* keep_alive,
5580 int* seed) {
5581 OopTaskQueue* work_q = work_queue(i);
5582 NOT_PRODUCT(int num_steals = 0;)
5583 oop obj_to_scan;
5584 size_t num_from_overflow_list =
5585 MIN2((size_t)work_q->max_elems()/4,
5586 (size_t)ParGCDesiredObjsFromOverflowList);
5587
5588 while (true) {
5589 // Completely finish any left over work from (an) earlier round(s)
5590 drain->trim_queue(0);
5591 // Now check if there's any work in the overflow list
5592 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5593 work_q)) {
5594 // Found something in global overflow list;
5595 // not yet ready to go stealing work from others.
5596 // We'd like to assert(work_q->size() != 0, ...)
5597 // because we just took work from the overflow list,
5598 // but of course we can't, since all of that might have
5599 // been already stolen from us.
5600 continue;
5601 }
5602 // Verify that we have no work before we resort to stealing
5603 assert(work_q->size() == 0, "Have work, shouldn't steal");
5604 // Try to steal from other queues that have work
5605 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5606 NOT_PRODUCT(num_steals++;)
5607 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5608 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5609 // Do scanning work
5610 obj_to_scan->oop_iterate(keep_alive);
5611 // Loop around, finish this work, and try to steal some more
5612 } else if (terminator()->offer_termination()) {
5613 break; // nirvana from the infinite cycle
5614 }
5615 }
5616 NOT_PRODUCT(
5617 if (PrintCMSStatistics != 0) {
5618 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5619 }
5620 )
5621 }
5622
5623 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5624 {
5625 GenCollectedHeap* gch = GenCollectedHeap::heap();
5626 WorkGang* workers = gch->workers();
5627 assert(workers != NULL, "Need parallel worker threads.");
5628 int n_workers = workers->total_workers();
5629 CMSRefProcTaskProxy rp_task(task, &_collector,
5630 _collector.ref_processor()->span(),
5631 _collector.markBitMap(),
5632 n_workers, _collector.task_queues());
5633 workers->run_task(&rp_task);
5634 }
5635
5636 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5637 {
5638
5639 GenCollectedHeap* gch = GenCollectedHeap::heap();
5640 WorkGang* workers = gch->workers();
5641 assert(workers != NULL, "Need parallel worker threads.");
5642 CMSRefEnqueueTaskProxy enq_task(task);
5643 workers->run_task(&enq_task);
5644 }
5645
5646 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5647
5648 ResourceMark rm;
5649 HandleMark hm;
5650 ReferencePolicy* soft_ref_policy;
5651
5652 assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete");
5653 // Process weak references.
5654 if (clear_all_soft_refs) {
5655 soft_ref_policy = new AlwaysClearPolicy();
5656 } else {
5657 #ifdef COMPILER2
5658 soft_ref_policy = new LRUMaxHeapPolicy();
5659 #else
5660 soft_ref_policy = new LRUCurrentHeapPolicy();
5661 #endif // COMPILER2
5662 }
5663 verify_work_stacks_empty();
5664
5665 ReferenceProcessor* rp = ref_processor();
5666 assert(rp->span().equals(_span), "Spans should be equal");
5667 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5668 &_markStack);
5669 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5670 _span, &_markBitMap, &_markStack,
5671 &cmsKeepAliveClosure);
5672 {
5673 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5674 if (rp->processing_is_mt()) {
5675 CMSRefProcTaskExecutor task_executor(*this);
5676 rp->process_discovered_references(soft_ref_policy,
5677 &_is_alive_closure,
5678 &cmsKeepAliveClosure,
5679 &cmsDrainMarkingStackClosure,
5680 &task_executor);
5681 } else {
5682 rp->process_discovered_references(soft_ref_policy,
5683 &_is_alive_closure,
5684 &cmsKeepAliveClosure,
5685 &cmsDrainMarkingStackClosure,
5686 NULL);
5687 }
5688 verify_work_stacks_empty();
5689 }
5690
5691 if (should_unload_classes()) {
5692 {
5693 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5694
5695 // Follow SystemDictionary roots and unload classes
5696 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5697
5698 // Follow CodeCache roots and unload any methods marked for unloading
5699 CodeCache::do_unloading(&_is_alive_closure,
5700 &cmsKeepAliveClosure,
5701 purged_class);
5702
5703 cmsDrainMarkingStackClosure.do_void();
5704 verify_work_stacks_empty();
5705
5706 // Update subklass/sibling/implementor links in KlassKlass descendants
5707 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5708 oop k;
5709 while ((k = _revisitStack.pop()) != NULL) {
5710 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5711 &_is_alive_closure,
5712 &cmsKeepAliveClosure);
5713 }
5714 assert(!ClassUnloading ||
5715 (_markStack.isEmpty() && overflow_list_is_empty()),
5716 "Should not have found new reachable objects");
5717 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5718 cmsDrainMarkingStackClosure.do_void();
5719 verify_work_stacks_empty();
5720 }
5721
5722 {
5723 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5724 // Now clean up stale oops in SymbolTable and StringTable
5725 SymbolTable::unlink(&_is_alive_closure);
5726 StringTable::unlink(&_is_alive_closure);
5727 }
5728 }
5729
5730 verify_work_stacks_empty();
5731 // Restore any preserved marks as a result of mark stack or
5732 // work queue overflow
5733 restore_preserved_marks_if_any(); // done single-threaded for now
5734
5735 rp->set_enqueuing_is_done(true);
5736 if (rp->processing_is_mt()) {
5737 CMSRefProcTaskExecutor task_executor(*this);
5738 rp->enqueue_discovered_references(&task_executor);
5739 } else {
5740 rp->enqueue_discovered_references(NULL);
5741 }
5742 rp->verify_no_references_recorded();
5743 assert(!rp->discovery_enabled(), "should have been disabled");
5744
5745 // JVMTI object tagging is based on JNI weak refs. If any of these
5746 // refs were cleared then JVMTI needs to update its maps and
5747 // maybe post ObjectFrees to agents.
5748 JvmtiExport::cms_ref_processing_epilogue();
5749 }
5750
5751 #ifndef PRODUCT
5752 void CMSCollector::check_correct_thread_executing() {
5753 Thread* t = Thread::current();
5754 // Only the VM thread or the CMS thread should be here.
5755 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5756 "Unexpected thread type");
5757 // If this is the vm thread, the foreground process
5758 // should not be waiting. Note that _foregroundGCIsActive is
5759 // true while the foreground collector is waiting.
5760 if (_foregroundGCShouldWait) {
5761 // We cannot be the VM thread
5762 assert(t->is_ConcurrentGC_thread(),
5763 "Should be CMS thread");
5764 } else {
5765 // We can be the CMS thread only if we are in a stop-world
5766 // phase of CMS collection.
5767 if (t->is_ConcurrentGC_thread()) {
5768 assert(_collectorState == InitialMarking ||
5769 _collectorState == FinalMarking,
5770 "Should be a stop-world phase");
5771 // The CMS thread should be holding the CMS_token.
5772 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5773 "Potential interference with concurrently "
5774 "executing VM thread");
5775 }
5776 }
5777 }
5778 #endif
5779
5780 void CMSCollector::sweep(bool asynch) {
5781 assert(_collectorState == Sweeping, "just checking");
5782 check_correct_thread_executing();
5783 verify_work_stacks_empty();
5784 verify_overflow_empty();
5785 incrementSweepCount();
5786 _sweep_timer.stop();
5787 _sweep_estimate.sample(_sweep_timer.seconds());
5788 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5789
5790 // PermGen verification support: If perm gen sweeping is disabled in
5791 // this cycle, we preserve the perm gen object "deadness" information
5792 // in the perm_gen_verify_bit_map. In order to do that we traverse
5793 // all blocks in perm gen and mark all dead objects.
5794 if (verifying() && !should_unload_classes()) {
5795 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
5796 "Should have already been allocated");
5797 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
5798 markBitMap(), perm_gen_verify_bit_map());
5799 if (asynch) {
5800 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5801 bitMapLock());
5802 _permGen->cmsSpace()->blk_iterate(&mdo);
5803 } else {
5804 // In the case of synchronous sweep, we already have
5805 // the requisite locks/tokens.
5806 _permGen->cmsSpace()->blk_iterate(&mdo);
5807 }
5808 }
5809
5810 if (asynch) {
5811 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5812 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5813 // First sweep the old gen then the perm gen
5814 {
5815 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5816 bitMapLock());
5817 sweepWork(_cmsGen, asynch);
5818 }
5819
5820 // Now repeat for perm gen
5821 if (should_unload_classes()) {
5822 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5823 bitMapLock());
5824 sweepWork(_permGen, asynch);
5825 }
5826
5827 // Update Universe::_heap_*_at_gc figures.
5828 // We need all the free list locks to make the abstract state
5829 // transition from Sweeping to Resetting. See detailed note
5830 // further below.
5831 {
5832 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5833 _permGen->freelistLock());
5834 // Update heap occupancy information which is used as
5835 // input to soft ref clearing policy at the next gc.
5836 Universe::update_heap_info_at_gc();
5837 _collectorState = Resizing;
5838 }
5839 } else {
5840 // already have needed locks
5841 sweepWork(_cmsGen, asynch);
5842
5843 if (should_unload_classes()) {
5844 sweepWork(_permGen, asynch);
5845 }
5846 // Update heap occupancy information which is used as
5847 // input to soft ref clearing policy at the next gc.
5848 Universe::update_heap_info_at_gc();
5849 _collectorState = Resizing;
5850 }
5851 verify_work_stacks_empty();
5852 verify_overflow_empty();
5853
5854 _sweep_timer.reset();
5855 _sweep_timer.start();
5856
5857 update_time_of_last_gc(os::javaTimeMillis());
5858
5859 // NOTE on abstract state transitions:
5860 // Mutators allocate-live and/or mark the mod-union table dirty
5861 // based on the state of the collection. The former is done in
5862 // the interval [Marking, Sweeping] and the latter in the interval
5863 // [Marking, Sweeping). Thus the transitions into the Marking state
5864 // and out of the Sweeping state must be synchronously visible
5865 // globally to the mutators.
5866 // The transition into the Marking state happens with the world
5867 // stopped so the mutators will globally see it. Sweeping is
5868 // done asynchronously by the background collector so the transition
5869 // from the Sweeping state to the Resizing state must be done
5870 // under the freelistLock (as is the check for whether to
5871 // allocate-live and whether to dirty the mod-union table).
5872 assert(_collectorState == Resizing, "Change of collector state to"
5873 " Resizing must be done under the freelistLocks (plural)");
5874
5875 // Now that sweeping has been completed, if the GCH's
5876 // incremental_collection_will_fail flag is set, clear it,
5877 // thus inviting a younger gen collection to promote into
5878 // this generation. If such a promotion may still fail,
5879 // the flag will be set again when a young collection is
5880 // attempted.
5881 // I think the incremental_collection_will_fail flag's use
5882 // is specific to a 2 generation collection policy, so i'll
5883 // assert that that's the configuration we are operating within.
5884 // The use of the flag can and should be generalized appropriately
5885 // in the future to deal with a general n-generation system.
5886
5887 GenCollectedHeap* gch = GenCollectedHeap::heap();
5888 assert(gch->collector_policy()->is_two_generation_policy(),
5889 "Resetting of incremental_collection_will_fail flag"
5890 " may be incorrect otherwise");
5891 gch->clear_incremental_collection_will_fail();
5892 gch->update_full_collections_completed(_collection_count_start);
5893 }
5894
5895 // FIX ME!!! Looks like this belongs in CFLSpace, with
5896 // CMSGen merely delegating to it.
5897 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5898 double nearLargestPercent = 0.999;
5899 HeapWord* minAddr = _cmsSpace->bottom();
5900 HeapWord* largestAddr =
5901 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
5902 if (largestAddr == 0) {
5903 // The dictionary appears to be empty. In this case
5904 // try to coalesce at the end of the heap.
5905 largestAddr = _cmsSpace->end();
5906 }
5907 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5908 size_t nearLargestOffset =
5909 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5910 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5911 }
5912
5913 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5914 return addr >= _cmsSpace->nearLargestChunk();
5915 }
5916
5917 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5918 return _cmsSpace->find_chunk_at_end();
5919 }
5920
5921 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5922 bool full) {
5923 // The next lower level has been collected. Gather any statistics
5924 // that are of interest at this point.
5925 if (!full && (current_level + 1) == level()) {
5926 // Gather statistics on the young generation collection.
5927 collector()->stats().record_gc0_end(used());
5928 }
5929 }
5930
5931 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
5932 GenCollectedHeap* gch = GenCollectedHeap::heap();
5933 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
5934 "Wrong type of heap");
5935 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
5936 gch->gen_policy()->size_policy();
5937 assert(sp->is_gc_cms_adaptive_size_policy(),
5938 "Wrong type of size policy");
5939 return sp;
5940 }
5941
5942 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
5943 if (PrintGCDetails && Verbose) {
5944 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
5945 }
5946 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
5947 _debug_collection_type =
5948 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
5949 if (PrintGCDetails && Verbose) {
5950 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
5951 }
5952 }
5953
5954 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
5955 bool asynch) {
5956 // We iterate over the space(s) underlying this generation,
5957 // checking the mark bit map to see if the bits corresponding
5958 // to specific blocks are marked or not. Blocks that are
5959 // marked are live and are not swept up. All remaining blocks
5960 // are swept up, with coalescing on-the-fly as we sweep up
5961 // contiguous free and/or garbage blocks:
5962 // We need to ensure that the sweeper synchronizes with allocators
5963 // and stop-the-world collectors. In particular, the following
5964 // locks are used:
5965 // . CMS token: if this is held, a stop the world collection cannot occur
5966 // . freelistLock: if this is held no allocation can occur from this
5967 // generation by another thread
5968 // . bitMapLock: if this is held, no other thread can access or update
5969 //
5970
5971 // Note that we need to hold the freelistLock if we use
5972 // block iterate below; else the iterator might go awry if
5973 // a mutator (or promotion) causes block contents to change
5974 // (for instance if the allocator divvies up a block).
5975 // If we hold the free list lock, for all practical purposes
5976 // young generation GC's can't occur (they'll usually need to
5977 // promote), so we might as well prevent all young generation
5978 // GC's while we do a sweeping step. For the same reason, we might
5979 // as well take the bit map lock for the entire duration
5980
5981 // check that we hold the requisite locks
5982 assert(have_cms_token(), "Should hold cms token");
5983 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
5984 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
5985 "Should possess CMS token to sweep");
5986 assert_lock_strong(gen->freelistLock());
5987 assert_lock_strong(bitMapLock());
5988
5989 assert(!_sweep_timer.is_active(), "Was switched off in an outer context");
5990 gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
5991 _sweep_estimate.padded_average());
5992 gen->setNearLargestChunk();
5993
5994 {
5995 SweepClosure sweepClosure(this, gen, &_markBitMap,
5996 CMSYield && asynch);
5997 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5998 // We need to free-up/coalesce garbage/blocks from a
5999 // co-terminal free run. This is done in the SweepClosure
6000 // destructor; so, do not remove this scope, else the
6001 // end-of-sweep-census below will be off by a little bit.
6002 }
6003 gen->cmsSpace()->sweep_completed();
6004 gen->cmsSpace()->endSweepFLCensus(sweepCount());
6005 if (should_unload_classes()) { // unloaded classes this cycle,
6006 _concurrent_cycles_since_last_unload = 0; // ... reset count
6007 } else { // did not unload classes,
6008 _concurrent_cycles_since_last_unload++; // ... increment count
6009 }
6010 }
6011
6012 // Reset CMS data structures (for now just the marking bit map)
6013 // preparatory for the next cycle.
6014 void CMSCollector::reset(bool asynch) {
6015 GenCollectedHeap* gch = GenCollectedHeap::heap();
6016 CMSAdaptiveSizePolicy* sp = size_policy();
6017 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6018 if (asynch) {
6019 CMSTokenSyncWithLocks ts(true, bitMapLock());
6020
6021 // If the state is not "Resetting", the foreground thread
6022 // has done a collection and the resetting.
6023 if (_collectorState != Resetting) {
6024 assert(_collectorState == Idling, "The state should only change"
6025 " because the foreground collector has finished the collection");
6026 return;
6027 }
6028
6029 // Clear the mark bitmap (no grey objects to start with)
6030 // for the next cycle.
6031 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6032 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6033
6034 HeapWord* curAddr = _markBitMap.startWord();
6035 while (curAddr < _markBitMap.endWord()) {
6036 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6037 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6038 _markBitMap.clear_large_range(chunk);
6039 if (ConcurrentMarkSweepThread::should_yield() &&
6040 !foregroundGCIsActive() &&
6041 CMSYield) {
6042 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6043 "CMS thread should hold CMS token");
6044 assert_lock_strong(bitMapLock());
6045 bitMapLock()->unlock();
6046 ConcurrentMarkSweepThread::desynchronize(true);
6047 ConcurrentMarkSweepThread::acknowledge_yield_request();
6048 stopTimer();
6049 if (PrintCMSStatistics != 0) {
6050 incrementYields();
6051 }
6052 icms_wait();
6053
6054 // See the comment in coordinator_yield()
6055 for (unsigned i = 0; i < CMSYieldSleepCount &&
6056 ConcurrentMarkSweepThread::should_yield() &&
6057 !CMSCollector::foregroundGCIsActive(); ++i) {
6058 os::sleep(Thread::current(), 1, false);
6059 ConcurrentMarkSweepThread::acknowledge_yield_request();
6060 }
6061
6062 ConcurrentMarkSweepThread::synchronize(true);
6063 bitMapLock()->lock_without_safepoint_check();
6064 startTimer();
6065 }
6066 curAddr = chunk.end();
6067 }
6068 _collectorState = Idling;
6069 } else {
6070 // already have the lock
6071 assert(_collectorState == Resetting, "just checking");
6072 assert_lock_strong(bitMapLock());
6073 _markBitMap.clear_all();
6074 _collectorState = Idling;
6075 }
6076
6077 // Stop incremental mode after a cycle completes, so that any future cycles
6078 // are triggered by allocation.
6079 stop_icms();
6080
6081 NOT_PRODUCT(
6082 if (RotateCMSCollectionTypes) {
6083 _cmsGen->rotate_debug_collection_type();
6084 }
6085 )
6086 }
6087
6088 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6089 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6090 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6091 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6092 TraceCollectorStats tcs(counters());
6093
6094 switch (op) {
6095 case CMS_op_checkpointRootsInitial: {
6096 checkpointRootsInitial(true); // asynch
6097 if (PrintGC) {
6098 _cmsGen->printOccupancy("initial-mark");
6099 }
6100 break;
6101 }
6102 case CMS_op_checkpointRootsFinal: {
6103 checkpointRootsFinal(true, // asynch
6104 false, // !clear_all_soft_refs
6105 false); // !init_mark_was_synchronous
6106 if (PrintGC) {
6107 _cmsGen->printOccupancy("remark");
6108 }
6109 break;
6110 }
6111 default:
6112 fatal("No such CMS_op");
6113 }
6114 }
6115
6116 #ifndef PRODUCT
6117 size_t const CMSCollector::skip_header_HeapWords() {
6118 return FreeChunk::header_size();
6119 }
6120
6121 // Try and collect here conditions that should hold when
6122 // CMS thread is exiting. The idea is that the foreground GC
6123 // thread should not be blocked if it wants to terminate
6124 // the CMS thread and yet continue to run the VM for a while
6125 // after that.
6126 void CMSCollector::verify_ok_to_terminate() const {
6127 assert(Thread::current()->is_ConcurrentGC_thread(),
6128 "should be called by CMS thread");
6129 assert(!_foregroundGCShouldWait, "should be false");
6130 // We could check here that all the various low-level locks
6131 // are not held by the CMS thread, but that is overkill; see
6132 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6133 // is checked.
6134 }
6135 #endif
6136
6137 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6138 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6139 "missing Printezis mark?");
6140 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6141 size_t size = pointer_delta(nextOneAddr + 1, addr);
6142 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6143 "alignment problem");
6144 assert(size >= 3, "Necessary for Printezis marks to work");
6145 return size;
6146 }
6147
6148 // A variant of the above (block_size_using_printezis_bits()) except
6149 // that we return 0 if the P-bits are not yet set.
6150 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6151 if (_markBitMap.isMarked(addr)) {
6152 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
6153 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6154 size_t size = pointer_delta(nextOneAddr + 1, addr);
6155 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6156 "alignment problem");
6157 assert(size >= 3, "Necessary for Printezis marks to work");
6158 return size;
6159 } else {
6160 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
6161 return 0;
6162 }
6163 }
6164
6165 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6166 size_t sz = 0;
6167 oop p = (oop)addr;
6168 if (p->klass() != NULL && p->is_parsable()) {
6169 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6170 } else {
6171 sz = block_size_using_printezis_bits(addr);
6172 }
6173 assert(sz > 0, "size must be nonzero");
6174 HeapWord* next_block = addr + sz;
6175 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6176 CardTableModRefBS::card_size);
6177 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6178 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6179 "must be different cards");
6180 return next_card;
6181 }
6182
6183
6184 // CMS Bit Map Wrapper /////////////////////////////////////////
6185
6186 // Construct a CMS bit map infrastructure, but don't create the
6187 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6188 // further below.
6189 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6190 _bm(NULL,0),
6191 _shifter(shifter),
6192 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6193 {
6194 _bmStartWord = 0;
6195 _bmWordSize = 0;
6196 }
6197
6198 bool CMSBitMap::allocate(MemRegion mr) {
6199 _bmStartWord = mr.start();
6200 _bmWordSize = mr.word_size();
6201 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6202 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6203 if (!brs.is_reserved()) {
6204 warning("CMS bit map allocation failure");
6205 return false;
6206 }
6207 // For now we'll just commit all of the bit map up fromt.
6208 // Later on we'll try to be more parsimonious with swap.
6209 if (!_virtual_space.initialize(brs, brs.size())) {
6210 warning("CMS bit map backing store failure");
6211 return false;
6212 }
6213 assert(_virtual_space.committed_size() == brs.size(),
6214 "didn't reserve backing store for all of CMS bit map?");
6215 _bm.set_map((uintptr_t*)_virtual_space.low());
6216 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6217 _bmWordSize, "inconsistency in bit map sizing");
6218 _bm.set_size(_bmWordSize >> _shifter);
6219
6220 // bm.clear(); // can we rely on getting zero'd memory? verify below
6221 assert(isAllClear(),
6222 "Expected zero'd memory from ReservedSpace constructor");
6223 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6224 "consistency check");
6225 return true;
6226 }
6227
6228 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6229 HeapWord *next_addr, *end_addr, *last_addr;
6230 assert_locked();
6231 assert(covers(mr), "out-of-range error");
6232 // XXX assert that start and end are appropriately aligned
6233 for (next_addr = mr.start(), end_addr = mr.end();
6234 next_addr < end_addr; next_addr = last_addr) {
6235 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6236 last_addr = dirty_region.end();
6237 if (!dirty_region.is_empty()) {
6238 cl->do_MemRegion(dirty_region);
6239 } else {
6240 assert(last_addr == end_addr, "program logic");
6241 return;
6242 }
6243 }
6244 }
6245
6246 #ifndef PRODUCT
6247 void CMSBitMap::assert_locked() const {
6248 CMSLockVerifier::assert_locked(lock());
6249 }
6250
6251 bool CMSBitMap::covers(MemRegion mr) const {
6252 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6253 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6254 "size inconsistency");
6255 return (mr.start() >= _bmStartWord) &&
6256 (mr.end() <= endWord());
6257 }
6258
6259 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6260 return (start >= _bmStartWord && (start + size) <= endWord());
6261 }
6262
6263 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6264 // verify that there are no 1 bits in the interval [left, right)
6265 FalseBitMapClosure falseBitMapClosure;
6266 iterate(&falseBitMapClosure, left, right);
6267 }
6268
6269 void CMSBitMap::region_invariant(MemRegion mr)
6270 {
6271 assert_locked();
6272 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6273 assert(!mr.is_empty(), "unexpected empty region");
6274 assert(covers(mr), "mr should be covered by bit map");
6275 // convert address range into offset range
6276 size_t start_ofs = heapWordToOffset(mr.start());
6277 // Make sure that end() is appropriately aligned
6278 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6279 (1 << (_shifter+LogHeapWordSize))),
6280 "Misaligned mr.end()");
6281 size_t end_ofs = heapWordToOffset(mr.end());
6282 assert(end_ofs > start_ofs, "Should mark at least one bit");
6283 }
6284
6285 #endif
6286
6287 bool CMSMarkStack::allocate(size_t size) {
6288 // allocate a stack of the requisite depth
6289 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6290 size * sizeof(oop)));
6291 if (!rs.is_reserved()) {
6292 warning("CMSMarkStack allocation failure");
6293 return false;
6294 }
6295 if (!_virtual_space.initialize(rs, rs.size())) {
6296 warning("CMSMarkStack backing store failure");
6297 return false;
6298 }
6299 assert(_virtual_space.committed_size() == rs.size(),
6300 "didn't reserve backing store for all of CMS stack?");
6301 _base = (oop*)(_virtual_space.low());
6302 _index = 0;
6303 _capacity = size;
6304 NOT_PRODUCT(_max_depth = 0);
6305 return true;
6306 }
6307
6308 // XXX FIX ME !!! In the MT case we come in here holding a
6309 // leaf lock. For printing we need to take a further lock
6310 // which has lower rank. We need to recallibrate the two
6311 // lock-ranks involved in order to be able to rpint the
6312 // messages below. (Or defer the printing to the caller.
6313 // For now we take the expedient path of just disabling the
6314 // messages for the problematic case.)
6315 void CMSMarkStack::expand() {
6316 assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted");
6317 if (_capacity == CMSMarkStackSizeMax) {
6318 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6319 // We print a warning message only once per CMS cycle.
6320 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6321 }
6322 return;
6323 }
6324 // Double capacity if possible
6325 size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax);
6326 // Do not give up existing stack until we have managed to
6327 // get the double capacity that we desired.
6328 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6329 new_capacity * sizeof(oop)));
6330 if (rs.is_reserved()) {
6331 // Release the backing store associated with old stack
6332 _virtual_space.release();
6333 // Reinitialize virtual space for new stack
6334 if (!_virtual_space.initialize(rs, rs.size())) {
6335 fatal("Not enough swap for expanded marking stack");
6336 }
6337 _base = (oop*)(_virtual_space.low());
6338 _index = 0;
6339 _capacity = new_capacity;
6340 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6341 // Failed to double capacity, continue;
6342 // we print a detail message only once per CMS cycle.
6343 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6344 SIZE_FORMAT"K",
6345 _capacity / K, new_capacity / K);
6346 }
6347 }
6348
6349
6350 // Closures
6351 // XXX: there seems to be a lot of code duplication here;
6352 // should refactor and consolidate common code.
6353
6354 // This closure is used to mark refs into the CMS generation in
6355 // the CMS bit map. Called at the first checkpoint. This closure
6356 // assumes that we do not need to re-mark dirty cards; if the CMS
6357 // generation on which this is used is not an oldest (modulo perm gen)
6358 // generation then this will lose younger_gen cards!
6359
6360 MarkRefsIntoClosure::MarkRefsIntoClosure(
6361 MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods):
6362 _span(span),
6363 _bitMap(bitMap),
6364 _should_do_nmethods(should_do_nmethods)
6365 {
6366 assert(_ref_processor == NULL, "deliberately left NULL");
6367 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6368 }
6369
6370 void MarkRefsIntoClosure::do_oop(oop obj) {
6371 // if p points into _span, then mark corresponding bit in _markBitMap
6372 assert(obj->is_oop(), "expected an oop");
6373 HeapWord* addr = (HeapWord*)obj;
6374 if (_span.contains(addr)) {
6375 // this should be made more efficient
6376 _bitMap->mark(addr);
6377 }
6378 }
6379
6380 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6381 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6382
6383 // A variant of the above, used for CMS marking verification.
6384 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6385 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6386 bool should_do_nmethods):
6387 _span(span),
6388 _verification_bm(verification_bm),
6389 _cms_bm(cms_bm),
6390 _should_do_nmethods(should_do_nmethods) {
6391 assert(_ref_processor == NULL, "deliberately left NULL");
6392 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6393 }
6394
6395 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6396 // if p points into _span, then mark corresponding bit in _markBitMap
6397 assert(obj->is_oop(), "expected an oop");
6398 HeapWord* addr = (HeapWord*)obj;
6399 if (_span.contains(addr)) {
6400 _verification_bm->mark(addr);
6401 if (!_cms_bm->isMarked(addr)) {
6402 oop(addr)->print();
6403 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6404 fatal("... aborting");
6405 }
6406 }
6407 }
6408
6409 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6410 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6411
6412 //////////////////////////////////////////////////
6413 // MarkRefsIntoAndScanClosure
6414 //////////////////////////////////////////////////
6415
6416 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6417 ReferenceProcessor* rp,
6418 CMSBitMap* bit_map,
6419 CMSBitMap* mod_union_table,
6420 CMSMarkStack* mark_stack,
6421 CMSMarkStack* revisit_stack,
6422 CMSCollector* collector,
6423 bool should_yield,
6424 bool concurrent_precleaning):
6425 _collector(collector),
6426 _span(span),
6427 _bit_map(bit_map),
6428 _mark_stack(mark_stack),
6429 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6430 mark_stack, revisit_stack, concurrent_precleaning),
6431 _yield(should_yield),
6432 _concurrent_precleaning(concurrent_precleaning),
6433 _freelistLock(NULL)
6434 {
6435 _ref_processor = rp;
6436 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6437 }
6438
6439 // This closure is used to mark refs into the CMS generation at the
6440 // second (final) checkpoint, and to scan and transitively follow
6441 // the unmarked oops. It is also used during the concurrent precleaning
6442 // phase while scanning objects on dirty cards in the CMS generation.
6443 // The marks are made in the marking bit map and the marking stack is
6444 // used for keeping the (newly) grey objects during the scan.
6445 // The parallel version (Par_...) appears further below.
6446 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6447 if (obj != NULL) {
6448 assert(obj->is_oop(), "expected an oop");
6449 HeapWord* addr = (HeapWord*)obj;
6450 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6451 assert(_collector->overflow_list_is_empty(),
6452 "overflow list should be empty");
6453 if (_span.contains(addr) &&
6454 !_bit_map->isMarked(addr)) {
6455 // mark bit map (object is now grey)
6456 _bit_map->mark(addr);
6457 // push on marking stack (stack should be empty), and drain the
6458 // stack by applying this closure to the oops in the oops popped
6459 // from the stack (i.e. blacken the grey objects)
6460 bool res = _mark_stack->push(obj);
6461 assert(res, "Should have space to push on empty stack");
6462 do {
6463 oop new_oop = _mark_stack->pop();
6464 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6465 assert(new_oop->is_parsable(), "Found unparsable oop");
6466 assert(_bit_map->isMarked((HeapWord*)new_oop),
6467 "only grey objects on this stack");
6468 // iterate over the oops in this oop, marking and pushing
6469 // the ones in CMS heap (i.e. in _span).
6470 new_oop->oop_iterate(&_pushAndMarkClosure);
6471 // check if it's time to yield
6472 do_yield_check();
6473 } while (!_mark_stack->isEmpty() ||
6474 (!_concurrent_precleaning && take_from_overflow_list()));
6475 // if marking stack is empty, and we are not doing this
6476 // during precleaning, then check the overflow list
6477 }
6478 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6479 assert(_collector->overflow_list_is_empty(),
6480 "overflow list was drained above");
6481 // We could restore evacuated mark words, if any, used for
6482 // overflow list links here because the overflow list is
6483 // provably empty here. That would reduce the maximum
6484 // size requirements for preserved_{oop,mark}_stack.
6485 // But we'll just postpone it until we are all done
6486 // so we can just stream through.
6487 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6488 _collector->restore_preserved_marks_if_any();
6489 assert(_collector->no_preserved_marks(), "No preserved marks");
6490 }
6491 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6492 "All preserved marks should have been restored above");
6493 }
6494 }
6495
6496 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6497 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6498
6499 void MarkRefsIntoAndScanClosure::do_yield_work() {
6500 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6501 "CMS thread should hold CMS token");
6502 assert_lock_strong(_freelistLock);
6503 assert_lock_strong(_bit_map->lock());
6504 // relinquish the free_list_lock and bitMaplock()
6505 _bit_map->lock()->unlock();
6506 _freelistLock->unlock();
6507 ConcurrentMarkSweepThread::desynchronize(true);
6508 ConcurrentMarkSweepThread::acknowledge_yield_request();
6509 _collector->stopTimer();
6510 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6511 if (PrintCMSStatistics != 0) {
6512 _collector->incrementYields();
6513 }
6514 _collector->icms_wait();
6515
6516 // See the comment in coordinator_yield()
6517 for (unsigned i = 0;
6518 i < CMSYieldSleepCount &&
6519 ConcurrentMarkSweepThread::should_yield() &&
6520 !CMSCollector::foregroundGCIsActive();
6521 ++i) {
6522 os::sleep(Thread::current(), 1, false);
6523 ConcurrentMarkSweepThread::acknowledge_yield_request();
6524 }
6525
6526 ConcurrentMarkSweepThread::synchronize(true);
6527 _freelistLock->lock_without_safepoint_check();
6528 _bit_map->lock()->lock_without_safepoint_check();
6529 _collector->startTimer();
6530 }
6531
6532 ///////////////////////////////////////////////////////////
6533 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6534 // MarkRefsIntoAndScanClosure
6535 ///////////////////////////////////////////////////////////
6536 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6537 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6538 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6539 _span(span),
6540 _bit_map(bit_map),
6541 _work_queue(work_queue),
6542 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6543 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6544 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6545 revisit_stack)
6546 {
6547 _ref_processor = rp;
6548 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6549 }
6550
6551 // This closure is used to mark refs into the CMS generation at the
6552 // second (final) checkpoint, and to scan and transitively follow
6553 // the unmarked oops. The marks are made in the marking bit map and
6554 // the work_queue is used for keeping the (newly) grey objects during
6555 // the scan phase whence they are also available for stealing by parallel
6556 // threads. Since the marking bit map is shared, updates are
6557 // synchronized (via CAS).
6558 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6559 if (obj != NULL) {
6560 // Ignore mark word because this could be an already marked oop
6561 // that may be chained at the end of the overflow list.
6562 assert(obj->is_oop(), "expected an oop");
6563 HeapWord* addr = (HeapWord*)obj;
6564 if (_span.contains(addr) &&
6565 !_bit_map->isMarked(addr)) {
6566 // mark bit map (object will become grey):
6567 // It is possible for several threads to be
6568 // trying to "claim" this object concurrently;
6569 // the unique thread that succeeds in marking the
6570 // object first will do the subsequent push on
6571 // to the work queue (or overflow list).
6572 if (_bit_map->par_mark(addr)) {
6573 // push on work_queue (which may not be empty), and trim the
6574 // queue to an appropriate length by applying this closure to
6575 // the oops in the oops popped from the stack (i.e. blacken the
6576 // grey objects)
6577 bool res = _work_queue->push(obj);
6578 assert(res, "Low water mark should be less than capacity?");
6579 trim_queue(_low_water_mark);
6580 } // Else, another thread claimed the object
6581 }
6582 }
6583 }
6584
6585 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6586 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6587
6588 // This closure is used to rescan the marked objects on the dirty cards
6589 // in the mod union table and the card table proper.
6590 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6591 oop p, MemRegion mr) {
6592
6593 size_t size = 0;
6594 HeapWord* addr = (HeapWord*)p;
6595 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6596 assert(_span.contains(addr), "we are scanning the CMS generation");
6597 // check if it's time to yield
6598 if (do_yield_check()) {
6599 // We yielded for some foreground stop-world work,
6600 // and we have been asked to abort this ongoing preclean cycle.
6601 return 0;
6602 }
6603 if (_bitMap->isMarked(addr)) {
6604 // it's marked; is it potentially uninitialized?
6605 if (p->klass() != NULL) {
6606 if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) {
6607 // Signal precleaning to redirty the card since
6608 // the klass pointer is already installed.
6609 assert(size == 0, "Initial value");
6610 } else {
6611 assert(p->is_parsable(), "must be parsable.");
6612 // an initialized object; ignore mark word in verification below
6613 // since we are running concurrent with mutators
6614 assert(p->is_oop(true), "should be an oop");
6615 if (p->is_objArray()) {
6616 // objArrays are precisely marked; restrict scanning
6617 // to dirty cards only.
6618 size = p->oop_iterate(_scanningClosure, mr);
6619 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6620 "adjustObjectSize should be the identity for array sizes, "
6621 "which are necessarily larger than minimum object size of "
6622 "two heap words");
6623 } else {
6624 // A non-array may have been imprecisely marked; we need
6625 // to scan object in its entirety.
6626 size = CompactibleFreeListSpace::adjustObjectSize(
6627 p->oop_iterate(_scanningClosure));
6628 }
6629 #ifdef DEBUG
6630 size_t direct_size =
6631 CompactibleFreeListSpace::adjustObjectSize(p->size());
6632 assert(size == direct_size, "Inconsistency in size");
6633 assert(size >= 3, "Necessary for Printezis marks to work");
6634 if (!_bitMap->isMarked(addr+1)) {
6635 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6636 } else {
6637 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6638 assert(_bitMap->isMarked(addr+size-1),
6639 "inconsistent Printezis mark");
6640 }
6641 #endif // DEBUG
6642 }
6643 } else {
6644 // an unitialized object
6645 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6646 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6647 size = pointer_delta(nextOneAddr + 1, addr);
6648 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6649 "alignment problem");
6650 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6651 // will dirty the card when the klass pointer is installed in the
6652 // object (signalling the completion of initialization).
6653 }
6654 } else {
6655 // Either a not yet marked object or an uninitialized object
6656 if (p->klass() == NULL || !p->is_parsable()) {
6657 // An uninitialized object, skip to the next card, since
6658 // we may not be able to read its P-bits yet.
6659 assert(size == 0, "Initial value");
6660 } else {
6661 // An object not (yet) reached by marking: we merely need to
6662 // compute its size so as to go look at the next block.
6663 assert(p->is_oop(true), "should be an oop");
6664 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6665 }
6666 }
6667 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6668 return size;
6669 }
6670
6671 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6672 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6673 "CMS thread should hold CMS token");
6674 assert_lock_strong(_freelistLock);
6675 assert_lock_strong(_bitMap->lock());
6676 // relinquish the free_list_lock and bitMaplock()
6677 _bitMap->lock()->unlock();
6678 _freelistLock->unlock();
6679 ConcurrentMarkSweepThread::desynchronize(true);
6680 ConcurrentMarkSweepThread::acknowledge_yield_request();
6681 _collector->stopTimer();
6682 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6683 if (PrintCMSStatistics != 0) {
6684 _collector->incrementYields();
6685 }
6686 _collector->icms_wait();
6687
6688 // See the comment in coordinator_yield()
6689 for (unsigned i = 0; i < CMSYieldSleepCount &&
6690 ConcurrentMarkSweepThread::should_yield() &&
6691 !CMSCollector::foregroundGCIsActive(); ++i) {
6692 os::sleep(Thread::current(), 1, false);
6693 ConcurrentMarkSweepThread::acknowledge_yield_request();
6694 }
6695
6696 ConcurrentMarkSweepThread::synchronize(true);
6697 _freelistLock->lock_without_safepoint_check();
6698 _bitMap->lock()->lock_without_safepoint_check();
6699 _collector->startTimer();
6700 }
6701
6702
6703 //////////////////////////////////////////////////////////////////
6704 // SurvivorSpacePrecleanClosure
6705 //////////////////////////////////////////////////////////////////
6706 // This (single-threaded) closure is used to preclean the oops in
6707 // the survivor spaces.
6708 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6709
6710 HeapWord* addr = (HeapWord*)p;
6711 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6712 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6713 assert(p->klass() != NULL, "object should be initializd");
6714 assert(p->is_parsable(), "must be parsable.");
6715 // an initialized object; ignore mark word in verification below
6716 // since we are running concurrent with mutators
6717 assert(p->is_oop(true), "should be an oop");
6718 // Note that we do not yield while we iterate over
6719 // the interior oops of p, pushing the relevant ones
6720 // on our marking stack.
6721 size_t size = p->oop_iterate(_scanning_closure);
6722 do_yield_check();
6723 // Observe that below, we do not abandon the preclean
6724 // phase as soon as we should; rather we empty the
6725 // marking stack before returning. This is to satisfy
6726 // some existing assertions. In general, it may be a
6727 // good idea to abort immediately and complete the marking
6728 // from the grey objects at a later time.
6729 while (!_mark_stack->isEmpty()) {
6730 oop new_oop = _mark_stack->pop();
6731 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6732 assert(new_oop->is_parsable(), "Found unparsable oop");
6733 assert(_bit_map->isMarked((HeapWord*)new_oop),
6734 "only grey objects on this stack");
6735 // iterate over the oops in this oop, marking and pushing
6736 // the ones in CMS heap (i.e. in _span).
6737 new_oop->oop_iterate(_scanning_closure);
6738 // check if it's time to yield
6739 do_yield_check();
6740 }
6741 unsigned int after_count =
6742 GenCollectedHeap::heap()->total_collections();
6743 bool abort = (_before_count != after_count) ||
6744 _collector->should_abort_preclean();
6745 return abort ? 0 : size;
6746 }
6747
6748 void SurvivorSpacePrecleanClosure::do_yield_work() {
6749 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6750 "CMS thread should hold CMS token");
6751 assert_lock_strong(_bit_map->lock());
6752 // Relinquish the bit map lock
6753 _bit_map->lock()->unlock();
6754 ConcurrentMarkSweepThread::desynchronize(true);
6755 ConcurrentMarkSweepThread::acknowledge_yield_request();
6756 _collector->stopTimer();
6757 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6758 if (PrintCMSStatistics != 0) {
6759 _collector->incrementYields();
6760 }
6761 _collector->icms_wait();
6762
6763 // See the comment in coordinator_yield()
6764 for (unsigned i = 0; i < CMSYieldSleepCount &&
6765 ConcurrentMarkSweepThread::should_yield() &&
6766 !CMSCollector::foregroundGCIsActive(); ++i) {
6767 os::sleep(Thread::current(), 1, false);
6768 ConcurrentMarkSweepThread::acknowledge_yield_request();
6769 }
6770
6771 ConcurrentMarkSweepThread::synchronize(true);
6772 _bit_map->lock()->lock_without_safepoint_check();
6773 _collector->startTimer();
6774 }
6775
6776 // This closure is used to rescan the marked objects on the dirty cards
6777 // in the mod union table and the card table proper. In the parallel
6778 // case, although the bitMap is shared, we do a single read so the
6779 // isMarked() query is "safe".
6780 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6781 // Ignore mark word because we are running concurrent with mutators
6782 assert(p->is_oop_or_null(true), "expected an oop or null");
6783 HeapWord* addr = (HeapWord*)p;
6784 assert(_span.contains(addr), "we are scanning the CMS generation");
6785 bool is_obj_array = false;
6786 #ifdef DEBUG
6787 if (!_parallel) {
6788 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6789 assert(_collector->overflow_list_is_empty(),
6790 "overflow list should be empty");
6791
6792 }
6793 #endif // DEBUG
6794 if (_bit_map->isMarked(addr)) {
6795 // Obj arrays are precisely marked, non-arrays are not;
6796 // so we scan objArrays precisely and non-arrays in their
6797 // entirety.
6798 if (p->is_objArray()) {
6799 is_obj_array = true;
6800 if (_parallel) {
6801 p->oop_iterate(_par_scan_closure, mr);
6802 } else {
6803 p->oop_iterate(_scan_closure, mr);
6804 }
6805 } else {
6806 if (_parallel) {
6807 p->oop_iterate(_par_scan_closure);
6808 } else {
6809 p->oop_iterate(_scan_closure);
6810 }
6811 }
6812 }
6813 #ifdef DEBUG
6814 if (!_parallel) {
6815 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6816 assert(_collector->overflow_list_is_empty(),
6817 "overflow list should be empty");
6818
6819 }
6820 #endif // DEBUG
6821 return is_obj_array;
6822 }
6823
6824 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6825 MemRegion span,
6826 CMSBitMap* bitMap, CMSMarkStack* markStack,
6827 CMSMarkStack* revisitStack,
6828 bool should_yield, bool verifying):
6829 _collector(collector),
6830 _span(span),
6831 _bitMap(bitMap),
6832 _mut(&collector->_modUnionTable),
6833 _markStack(markStack),
6834 _revisitStack(revisitStack),
6835 _yield(should_yield),
6836 _skipBits(0)
6837 {
6838 assert(_markStack->isEmpty(), "stack should be empty");
6839 _finger = _bitMap->startWord();
6840 _threshold = _finger;
6841 assert(_collector->_restart_addr == NULL, "Sanity check");
6842 assert(_span.contains(_finger), "Out of bounds _finger?");
6843 DEBUG_ONLY(_verifying = verifying;)
6844 }
6845
6846 void MarkFromRootsClosure::reset(HeapWord* addr) {
6847 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6848 assert(_span.contains(addr), "Out of bounds _finger?");
6849 _finger = addr;
6850 _threshold = (HeapWord*)round_to(
6851 (intptr_t)_finger, CardTableModRefBS::card_size);
6852 }
6853
6854 // Should revisit to see if this should be restructured for
6855 // greater efficiency.
6856 void MarkFromRootsClosure::do_bit(size_t offset) {
6857 if (_skipBits > 0) {
6858 _skipBits--;
6859 return;
6860 }
6861 // convert offset into a HeapWord*
6862 HeapWord* addr = _bitMap->startWord() + offset;
6863 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6864 "address out of range");
6865 assert(_bitMap->isMarked(addr), "tautology");
6866 if (_bitMap->isMarked(addr+1)) {
6867 // this is an allocated but not yet initialized object
6868 assert(_skipBits == 0, "tautology");
6869 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6870 oop p = oop(addr);
6871 if (p->klass() == NULL || !p->is_parsable()) {
6872 DEBUG_ONLY(if (!_verifying) {)
6873 // We re-dirty the cards on which this object lies and increase
6874 // the _threshold so that we'll come back to scan this object
6875 // during the preclean or remark phase. (CMSCleanOnEnter)
6876 if (CMSCleanOnEnter) {
6877 size_t sz = _collector->block_size_using_printezis_bits(addr);
6878 HeapWord* start_card_addr = (HeapWord*)round_down(
6879 (intptr_t)addr, CardTableModRefBS::card_size);
6880 HeapWord* end_card_addr = (HeapWord*)round_to(
6881 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6882 MemRegion redirty_range = MemRegion(start_card_addr, end_card_addr);
6883 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6884 // Bump _threshold to end_card_addr; note that
6885 // _threshold cannot possibly exceed end_card_addr, anyhow.
6886 // This prevents future clearing of the card as the scan proceeds
6887 // to the right.
6888 assert(_threshold <= end_card_addr,
6889 "Because we are just scanning into this object");
6890 if (_threshold < end_card_addr) {
6891 _threshold = end_card_addr;
6892 }
6893 if (p->klass() != NULL) {
6894 // Redirty the range of cards...
6895 _mut->mark_range(redirty_range);
6896 } // ...else the setting of klass will dirty the card anyway.
6897 }
6898 DEBUG_ONLY(})
6899 return;
6900 }
6901 }
6902 scanOopsInOop(addr);
6903 }
6904
6905 // We take a break if we've been at this for a while,
6906 // so as to avoid monopolizing the locks involved.
6907 void MarkFromRootsClosure::do_yield_work() {
6908 // First give up the locks, then yield, then re-lock
6909 // We should probably use a constructor/destructor idiom to
6910 // do this unlock/lock or modify the MutexUnlocker class to
6911 // serve our purpose. XXX
6912 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6913 "CMS thread should hold CMS token");
6914 assert_lock_strong(_bitMap->lock());
6915 _bitMap->lock()->unlock();
6916 ConcurrentMarkSweepThread::desynchronize(true);
6917 ConcurrentMarkSweepThread::acknowledge_yield_request();
6918 _collector->stopTimer();
6919 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6920 if (PrintCMSStatistics != 0) {
6921 _collector->incrementYields();
6922 }
6923 _collector->icms_wait();
6924
6925 // See the comment in coordinator_yield()
6926 for (unsigned i = 0; i < CMSYieldSleepCount &&
6927 ConcurrentMarkSweepThread::should_yield() &&
6928 !CMSCollector::foregroundGCIsActive(); ++i) {
6929 os::sleep(Thread::current(), 1, false);
6930 ConcurrentMarkSweepThread::acknowledge_yield_request();
6931 }
6932
6933 ConcurrentMarkSweepThread::synchronize(true);
6934 _bitMap->lock()->lock_without_safepoint_check();
6935 _collector->startTimer();
6936 }
6937
6938 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6939 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6940 assert(_markStack->isEmpty(),
6941 "should drain stack to limit stack usage");
6942 // convert ptr to an oop preparatory to scanning
6943 oop obj = oop(ptr);
6944 // Ignore mark word in verification below, since we
6945 // may be running concurrent with mutators.
6946 assert(obj->is_oop(true), "should be an oop");
6947 assert(_finger <= ptr, "_finger runneth ahead");
6948 // advance the finger to right end of this object
6949 _finger = ptr + obj->size();
6950 assert(_finger > ptr, "we just incremented it above");
6951 // On large heaps, it may take us some time to get through
6952 // the marking phase (especially if running iCMS). During
6953 // this time it's possible that a lot of mutations have
6954 // accumulated in the card table and the mod union table --
6955 // these mutation records are redundant until we have
6956 // actually traced into the corresponding card.
6957 // Here, we check whether advancing the finger would make
6958 // us cross into a new card, and if so clear corresponding
6959 // cards in the MUT (preclean them in the card-table in the
6960 // future).
6961
6962 DEBUG_ONLY(if (!_verifying) {)
6963 // The clean-on-enter optimization is disabled by default,
6964 // until we fix 6178663.
6965 if (CMSCleanOnEnter && (_finger > _threshold)) {
6966 // [_threshold, _finger) represents the interval
6967 // of cards to be cleared in MUT (or precleaned in card table).
6968 // The set of cards to be cleared is all those that overlap
6969 // with the interval [_threshold, _finger); note that
6970 // _threshold is always kept card-aligned but _finger isn't
6971 // always card-aligned.
6972 HeapWord* old_threshold = _threshold;
6973 assert(old_threshold == (HeapWord*)round_to(
6974 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6975 "_threshold should always be card-aligned");
6976 _threshold = (HeapWord*)round_to(
6977 (intptr_t)_finger, CardTableModRefBS::card_size);
6978 MemRegion mr(old_threshold, _threshold);
6979 assert(!mr.is_empty(), "Control point invariant");
6980 assert(_span.contains(mr), "Should clear within span");
6981 // XXX When _finger crosses from old gen into perm gen
6982 // we may be doing unnecessary cleaning; do better in the
6983 // future by detecting that condition and clearing fewer
6984 // MUT/CT entries.
6985 _mut->clear_range(mr);
6986 }
6987 DEBUG_ONLY(})
6988
6989 // Note: the finger doesn't advance while we drain
6990 // the stack below.
6991 PushOrMarkClosure pushOrMarkClosure(_collector,
6992 _span, _bitMap, _markStack,
6993 _revisitStack,
6994 _finger, this);
6995 bool res = _markStack->push(obj);
6996 assert(res, "Empty non-zero size stack should have space for single push");
6997 while (!_markStack->isEmpty()) {
6998 oop new_oop = _markStack->pop();
6999 // Skip verifying header mark word below because we are
7000 // running concurrent with mutators.
7001 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7002 // now scan this oop's oops
7003 new_oop->oop_iterate(&pushOrMarkClosure);
7004 do_yield_check();
7005 }
7006 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7007 }
7008
7009 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7010 CMSCollector* collector, MemRegion span,
7011 CMSBitMap* bit_map,
7012 OopTaskQueue* work_queue,
7013 CMSMarkStack* overflow_stack,
7014 CMSMarkStack* revisit_stack,
7015 bool should_yield):
7016 _collector(collector),
7017 _whole_span(collector->_span),
7018 _span(span),
7019 _bit_map(bit_map),
7020 _mut(&collector->_modUnionTable),
7021 _work_queue(work_queue),
7022 _overflow_stack(overflow_stack),
7023 _revisit_stack(revisit_stack),
7024 _yield(should_yield),
7025 _skip_bits(0),
7026 _task(task)
7027 {
7028 assert(_work_queue->size() == 0, "work_queue should be empty");
7029 _finger = span.start();
7030 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7031 assert(_span.contains(_finger), "Out of bounds _finger?");
7032 }
7033
7034 // Should revisit to see if this should be restructured for
7035 // greater efficiency.
7036 void Par_MarkFromRootsClosure::do_bit(size_t offset) {
7037 if (_skip_bits > 0) {
7038 _skip_bits--;
7039 return;
7040 }
7041 // convert offset into a HeapWord*
7042 HeapWord* addr = _bit_map->startWord() + offset;
7043 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7044 "address out of range");
7045 assert(_bit_map->isMarked(addr), "tautology");
7046 if (_bit_map->isMarked(addr+1)) {
7047 // this is an allocated object that might not yet be initialized
7048 assert(_skip_bits == 0, "tautology");
7049 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7050 oop p = oop(addr);
7051 if (p->klass() == NULL || !p->is_parsable()) {
7052 // in the case of Clean-on-Enter optimization, redirty card
7053 // and avoid clearing card by increasing the threshold.
7054 return;
7055 }
7056 }
7057 scan_oops_in_oop(addr);
7058 }
7059
7060 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7061 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7062 // Should we assert that our work queue is empty or
7063 // below some drain limit?
7064 assert(_work_queue->size() == 0,
7065 "should drain stack to limit stack usage");
7066 // convert ptr to an oop preparatory to scanning
7067 oop obj = oop(ptr);
7068 // Ignore mark word in verification below, since we
7069 // may be running concurrent with mutators.
7070 assert(obj->is_oop(true), "should be an oop");
7071 assert(_finger <= ptr, "_finger runneth ahead");
7072 // advance the finger to right end of this object
7073 _finger = ptr + obj->size();
7074 assert(_finger > ptr, "we just incremented it above");
7075 // On large heaps, it may take us some time to get through
7076 // the marking phase (especially if running iCMS). During
7077 // this time it's possible that a lot of mutations have
7078 // accumulated in the card table and the mod union table --
7079 // these mutation records are redundant until we have
7080 // actually traced into the corresponding card.
7081 // Here, we check whether advancing the finger would make
7082 // us cross into a new card, and if so clear corresponding
7083 // cards in the MUT (preclean them in the card-table in the
7084 // future).
7085
7086 // The clean-on-enter optimization is disabled by default,
7087 // until we fix 6178663.
7088 if (CMSCleanOnEnter && (_finger > _threshold)) {
7089 // [_threshold, _finger) represents the interval
7090 // of cards to be cleared in MUT (or precleaned in card table).
7091 // The set of cards to be cleared is all those that overlap
7092 // with the interval [_threshold, _finger); note that
7093 // _threshold is always kept card-aligned but _finger isn't
7094 // always card-aligned.
7095 HeapWord* old_threshold = _threshold;
7096 assert(old_threshold == (HeapWord*)round_to(
7097 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7098 "_threshold should always be card-aligned");
7099 _threshold = (HeapWord*)round_to(
7100 (intptr_t)_finger, CardTableModRefBS::card_size);
7101 MemRegion mr(old_threshold, _threshold);
7102 assert(!mr.is_empty(), "Control point invariant");
7103 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7104 // XXX When _finger crosses from old gen into perm gen
7105 // we may be doing unnecessary cleaning; do better in the
7106 // future by detecting that condition and clearing fewer
7107 // MUT/CT entries.
7108 _mut->clear_range(mr);
7109 }
7110
7111 // Note: the local finger doesn't advance while we drain
7112 // the stack below, but the global finger sure can and will.
7113 HeapWord** gfa = _task->global_finger_addr();
7114 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7115 _span, _bit_map,
7116 _work_queue,
7117 _overflow_stack,
7118 _revisit_stack,
7119 _finger,
7120 gfa, this);
7121 bool res = _work_queue->push(obj); // overflow could occur here
7122 assert(res, "Will hold once we use workqueues");
7123 while (true) {
7124 oop new_oop;
7125 if (!_work_queue->pop_local(new_oop)) {
7126 // We emptied our work_queue; check if there's stuff that can
7127 // be gotten from the overflow stack.
7128 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7129 _overflow_stack, _work_queue)) {
7130 do_yield_check();
7131 continue;
7132 } else { // done
7133 break;
7134 }
7135 }
7136 // Skip verifying header mark word below because we are
7137 // running concurrent with mutators.
7138 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7139 // now scan this oop's oops
7140 new_oop->oop_iterate(&pushOrMarkClosure);
7141 do_yield_check();
7142 }
7143 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7144 }
7145
7146 // Yield in response to a request from VM Thread or
7147 // from mutators.
7148 void Par_MarkFromRootsClosure::do_yield_work() {
7149 assert(_task != NULL, "sanity");
7150 _task->yield();
7151 }
7152
7153 // A variant of the above used for verifying CMS marking work.
7154 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7155 MemRegion span,
7156 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7157 CMSMarkStack* mark_stack):
7158 _collector(collector),
7159 _span(span),
7160 _verification_bm(verification_bm),
7161 _cms_bm(cms_bm),
7162 _mark_stack(mark_stack),
7163 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7164 mark_stack)
7165 {
7166 assert(_mark_stack->isEmpty(), "stack should be empty");
7167 _finger = _verification_bm->startWord();
7168 assert(_collector->_restart_addr == NULL, "Sanity check");
7169 assert(_span.contains(_finger), "Out of bounds _finger?");
7170 }
7171
7172 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7173 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7174 assert(_span.contains(addr), "Out of bounds _finger?");
7175 _finger = addr;
7176 }
7177
7178 // Should revisit to see if this should be restructured for
7179 // greater efficiency.
7180 void MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7181 // convert offset into a HeapWord*
7182 HeapWord* addr = _verification_bm->startWord() + offset;
7183 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7184 "address out of range");
7185 assert(_verification_bm->isMarked(addr), "tautology");
7186 assert(_cms_bm->isMarked(addr), "tautology");
7187
7188 assert(_mark_stack->isEmpty(),
7189 "should drain stack to limit stack usage");
7190 // convert addr to an oop preparatory to scanning
7191 oop obj = oop(addr);
7192 assert(obj->is_oop(), "should be an oop");
7193 assert(_finger <= addr, "_finger runneth ahead");
7194 // advance the finger to right end of this object
7195 _finger = addr + obj->size();
7196 assert(_finger > addr, "we just incremented it above");
7197 // Note: the finger doesn't advance while we drain
7198 // the stack below.
7199 bool res = _mark_stack->push(obj);
7200 assert(res, "Empty non-zero size stack should have space for single push");
7201 while (!_mark_stack->isEmpty()) {
7202 oop new_oop = _mark_stack->pop();
7203 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7204 // now scan this oop's oops
7205 new_oop->oop_iterate(&_pam_verify_closure);
7206 }
7207 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7208 }
7209
7210 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7211 CMSCollector* collector, MemRegion span,
7212 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7213 CMSMarkStack* mark_stack):
7214 OopClosure(collector->ref_processor()),
7215 _collector(collector),
7216 _span(span),
7217 _verification_bm(verification_bm),
7218 _cms_bm(cms_bm),
7219 _mark_stack(mark_stack)
7220 { }
7221
7222 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7223 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7224
7225 // Upon stack overflow, we discard (part of) the stack,
7226 // remembering the least address amongst those discarded
7227 // in CMSCollector's _restart_address.
7228 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7229 // Remember the least grey address discarded
7230 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7231 _collector->lower_restart_addr(ra);
7232 _mark_stack->reset(); // discard stack contents
7233 _mark_stack->expand(); // expand the stack if possible
7234 }
7235
7236 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7237 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7238 HeapWord* addr = (HeapWord*)obj;
7239 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7240 // Oop lies in _span and isn't yet grey or black
7241 _verification_bm->mark(addr); // now grey
7242 if (!_cms_bm->isMarked(addr)) {
7243 oop(addr)->print();
7244 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7245 addr);
7246 fatal("... aborting");
7247 }
7248
7249 if (!_mark_stack->push(obj)) { // stack overflow
7250 if (PrintCMSStatistics != 0) {
7251 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7252 SIZE_FORMAT, _mark_stack->capacity());
7253 }
7254 assert(_mark_stack->isFull(), "Else push should have succeeded");
7255 handle_stack_overflow(addr);
7256 }
7257 // anything including and to the right of _finger
7258 // will be scanned as we iterate over the remainder of the
7259 // bit map
7260 }
7261 }
7262
7263 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7264 MemRegion span,
7265 CMSBitMap* bitMap, CMSMarkStack* markStack,
7266 CMSMarkStack* revisitStack,
7267 HeapWord* finger, MarkFromRootsClosure* parent) :
7268 OopClosure(collector->ref_processor()),
7269 _collector(collector),
7270 _span(span),
7271 _bitMap(bitMap),
7272 _markStack(markStack),
7273 _revisitStack(revisitStack),
7274 _finger(finger),
7275 _parent(parent),
7276 _should_remember_klasses(collector->should_unload_classes())
7277 { }
7278
7279 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7280 MemRegion span,
7281 CMSBitMap* bit_map,
7282 OopTaskQueue* work_queue,
7283 CMSMarkStack* overflow_stack,
7284 CMSMarkStack* revisit_stack,
7285 HeapWord* finger,
7286 HeapWord** global_finger_addr,
7287 Par_MarkFromRootsClosure* parent) :
7288 OopClosure(collector->ref_processor()),
7289 _collector(collector),
7290 _whole_span(collector->_span),
7291 _span(span),
7292 _bit_map(bit_map),
7293 _work_queue(work_queue),
7294 _overflow_stack(overflow_stack),
7295 _revisit_stack(revisit_stack),
7296 _finger(finger),
7297 _global_finger_addr(global_finger_addr),
7298 _parent(parent),
7299 _should_remember_klasses(collector->should_unload_classes())
7300 { }
7301
7302 void CMSCollector::lower_restart_addr(HeapWord* low) {
7303 assert(_span.contains(low), "Out of bounds addr");
7304 if (_restart_addr == NULL) {
7305 _restart_addr = low;
7306 } else {
7307 _restart_addr = MIN2(_restart_addr, low);
7308 }
7309 }
7310
7311 // Upon stack overflow, we discard (part of) the stack,
7312 // remembering the least address amongst those discarded
7313 // in CMSCollector's _restart_address.
7314 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7315 // Remember the least grey address discarded
7316 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7317 _collector->lower_restart_addr(ra);
7318 _markStack->reset(); // discard stack contents
7319 _markStack->expand(); // expand the stack if possible
7320 }
7321
7322 // Upon stack overflow, we discard (part of) the stack,
7323 // remembering the least address amongst those discarded
7324 // in CMSCollector's _restart_address.
7325 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7326 // We need to do this under a mutex to prevent other
7327 // workers from interfering with the expansion below.
7328 MutexLockerEx ml(_overflow_stack->par_lock(),
7329 Mutex::_no_safepoint_check_flag);
7330 // Remember the least grey address discarded
7331 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7332 _collector->lower_restart_addr(ra);
7333 _overflow_stack->reset(); // discard stack contents
7334 _overflow_stack->expand(); // expand the stack if possible
7335 }
7336
7337 void PushOrMarkClosure::do_oop(oop obj) {
7338 // Ignore mark word because we are running concurrent with mutators.
7339 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7340 HeapWord* addr = (HeapWord*)obj;
7341 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7342 // Oop lies in _span and isn't yet grey or black
7343 _bitMap->mark(addr); // now grey
7344 if (addr < _finger) {
7345 // the bit map iteration has already either passed, or
7346 // sampled, this bit in the bit map; we'll need to
7347 // use the marking stack to scan this oop's oops.
7348 bool simulate_overflow = false;
7349 NOT_PRODUCT(
7350 if (CMSMarkStackOverflowALot &&
7351 _collector->simulate_overflow()) {
7352 // simulate a stack overflow
7353 simulate_overflow = true;
7354 }
7355 )
7356 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7357 if (PrintCMSStatistics != 0) {
7358 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7359 SIZE_FORMAT, _markStack->capacity());
7360 }
7361 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7362 handle_stack_overflow(addr);
7363 }
7364 }
7365 // anything including and to the right of _finger
7366 // will be scanned as we iterate over the remainder of the
7367 // bit map
7368 do_yield_check();
7369 }
7370 }
7371
7372 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7373 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7374
7375 void Par_PushOrMarkClosure::do_oop(oop obj) {
7376 // Ignore mark word because we are running concurrent with mutators.
7377 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7378 HeapWord* addr = (HeapWord*)obj;
7379 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7380 // Oop lies in _span and isn't yet grey or black
7381 // We read the global_finger (volatile read) strictly after marking oop
7382 bool res = _bit_map->par_mark(addr); // now grey
7383 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7384 // Should we push this marked oop on our stack?
7385 // -- if someone else marked it, nothing to do
7386 // -- if target oop is above global finger nothing to do
7387 // -- if target oop is in chunk and above local finger
7388 // then nothing to do
7389 // -- else push on work queue
7390 if ( !res // someone else marked it, they will deal with it
7391 || (addr >= *gfa) // will be scanned in a later task
7392 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7393 return;
7394 }
7395 // the bit map iteration has already either passed, or
7396 // sampled, this bit in the bit map; we'll need to
7397 // use the marking stack to scan this oop's oops.
7398 bool simulate_overflow = false;
7399 NOT_PRODUCT(
7400 if (CMSMarkStackOverflowALot &&
7401 _collector->simulate_overflow()) {
7402 // simulate a stack overflow
7403 simulate_overflow = true;
7404 }
7405 )
7406 if (simulate_overflow ||
7407 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7408 // stack overflow
7409 if (PrintCMSStatistics != 0) {
7410 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7411 SIZE_FORMAT, _overflow_stack->capacity());
7412 }
7413 // We cannot assert that the overflow stack is full because
7414 // it may have been emptied since.
7415 assert(simulate_overflow ||
7416 _work_queue->size() == _work_queue->max_elems(),
7417 "Else push should have succeeded");
7418 handle_stack_overflow(addr);
7419 }
7420 do_yield_check();
7421 }
7422 }
7423
7424 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7425 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7426
7427 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7428 MemRegion span,
7429 ReferenceProcessor* rp,
7430 CMSBitMap* bit_map,
7431 CMSBitMap* mod_union_table,
7432 CMSMarkStack* mark_stack,
7433 CMSMarkStack* revisit_stack,
7434 bool concurrent_precleaning):
7435 OopClosure(rp),
7436 _collector(collector),
7437 _span(span),
7438 _bit_map(bit_map),
7439 _mod_union_table(mod_union_table),
7440 _mark_stack(mark_stack),
7441 _revisit_stack(revisit_stack),
7442 _concurrent_precleaning(concurrent_precleaning),
7443 _should_remember_klasses(collector->should_unload_classes())
7444 {
7445 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7446 }
7447
7448 // Grey object rescan during pre-cleaning and second checkpoint phases --
7449 // the non-parallel version (the parallel version appears further below.)
7450 void PushAndMarkClosure::do_oop(oop obj) {
7451 // If _concurrent_precleaning, ignore mark word verification
7452 assert(obj->is_oop_or_null(_concurrent_precleaning),
7453 "expected an oop or NULL");
7454 HeapWord* addr = (HeapWord*)obj;
7455 // Check if oop points into the CMS generation
7456 // and is not marked
7457 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7458 // a white object ...
7459 _bit_map->mark(addr); // ... now grey
7460 // push on the marking stack (grey set)
7461 bool simulate_overflow = false;
7462 NOT_PRODUCT(
7463 if (CMSMarkStackOverflowALot &&
7464 _collector->simulate_overflow()) {
7465 // simulate a stack overflow
7466 simulate_overflow = true;
7467 }
7468 )
7469 if (simulate_overflow || !_mark_stack->push(obj)) {
7470 if (_concurrent_precleaning) {
7471 // During precleaning we can just dirty the appropriate card
7472 // in the mod union table, thus ensuring that the object remains
7473 // in the grey set and continue. Note that no one can be intefering
7474 // with us in this action of dirtying the mod union table, so
7475 // no locking is required.
7476 _mod_union_table->mark(addr);
7477 _collector->_ser_pmc_preclean_ovflw++;
7478 } else {
7479 // During the remark phase, we need to remember this oop
7480 // in the overflow list.
7481 _collector->push_on_overflow_list(obj);
7482 _collector->_ser_pmc_remark_ovflw++;
7483 }
7484 }
7485 }
7486 }
7487
7488 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7489 MemRegion span,
7490 ReferenceProcessor* rp,
7491 CMSBitMap* bit_map,
7492 OopTaskQueue* work_queue,
7493 CMSMarkStack* revisit_stack):
7494 OopClosure(rp),
7495 _collector(collector),
7496 _span(span),
7497 _bit_map(bit_map),
7498 _work_queue(work_queue),
7499 _revisit_stack(revisit_stack),
7500 _should_remember_klasses(collector->should_unload_classes())
7501 {
7502 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7503 }
7504
7505 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7506 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7507
7508 // Grey object rescan during second checkpoint phase --
7509 // the parallel version.
7510 void Par_PushAndMarkClosure::do_oop(oop obj) {
7511 // In the assert below, we ignore the mark word because
7512 // this oop may point to an already visited object that is
7513 // on the overflow stack (in which case the mark word has
7514 // been hijacked for chaining into the overflow stack --
7515 // if this is the last object in the overflow stack then
7516 // its mark word will be NULL). Because this object may
7517 // have been subsequently popped off the global overflow
7518 // stack, and the mark word possibly restored to the prototypical
7519 // value, by the time we get to examined this failing assert in
7520 // the debugger, is_oop_or_null(false) may subsequently start
7521 // to hold.
7522 assert(obj->is_oop_or_null(true),
7523 "expected an oop or NULL");
7524 HeapWord* addr = (HeapWord*)obj;
7525 // Check if oop points into the CMS generation
7526 // and is not marked
7527 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7528 // a white object ...
7529 // If we manage to "claim" the object, by being the
7530 // first thread to mark it, then we push it on our
7531 // marking stack
7532 if (_bit_map->par_mark(addr)) { // ... now grey
7533 // push on work queue (grey set)
7534 bool simulate_overflow = false;
7535 NOT_PRODUCT(
7536 if (CMSMarkStackOverflowALot &&
7537 _collector->par_simulate_overflow()) {
7538 // simulate a stack overflow
7539 simulate_overflow = true;
7540 }
7541 )
7542 if (simulate_overflow || !_work_queue->push(obj)) {
7543 _collector->par_push_on_overflow_list(obj);
7544 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7545 }
7546 } // Else, some other thread got there first
7547 }
7548 }
7549
7550 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7551 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7552
7553 void PushAndMarkClosure::remember_klass(Klass* k) {
7554 if (!_revisit_stack->push(oop(k))) {
7555 fatal("Revisit stack overflowed in PushAndMarkClosure");
7556 }
7557 }
7558
7559 void Par_PushAndMarkClosure::remember_klass(Klass* k) {
7560 if (!_revisit_stack->par_push(oop(k))) {
7561 fatal("Revist stack overflowed in Par_PushAndMarkClosure");
7562 }
7563 }
7564
7565 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7566 Mutex* bml = _collector->bitMapLock();
7567 assert_lock_strong(bml);
7568 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7569 "CMS thread should hold CMS token");
7570
7571 bml->unlock();
7572 ConcurrentMarkSweepThread::desynchronize(true);
7573
7574 ConcurrentMarkSweepThread::acknowledge_yield_request();
7575
7576 _collector->stopTimer();
7577 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7578 if (PrintCMSStatistics != 0) {
7579 _collector->incrementYields();
7580 }
7581 _collector->icms_wait();
7582
7583 // See the comment in coordinator_yield()
7584 for (unsigned i = 0; i < CMSYieldSleepCount &&
7585 ConcurrentMarkSweepThread::should_yield() &&
7586 !CMSCollector::foregroundGCIsActive(); ++i) {
7587 os::sleep(Thread::current(), 1, false);
7588 ConcurrentMarkSweepThread::acknowledge_yield_request();
7589 }
7590
7591 ConcurrentMarkSweepThread::synchronize(true);
7592 bml->lock();
7593
7594 _collector->startTimer();
7595 }
7596
7597 bool CMSPrecleanRefsYieldClosure::should_return() {
7598 if (ConcurrentMarkSweepThread::should_yield()) {
7599 do_yield_work();
7600 }
7601 return _collector->foregroundGCIsActive();
7602 }
7603
7604 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7605 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7606 "mr should be aligned to start at a card boundary");
7607 // We'd like to assert:
7608 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7609 // "mr should be a range of cards");
7610 // However, that would be too strong in one case -- the last
7611 // partition ends at _unallocated_block which, in general, can be
7612 // an arbitrary boundary, not necessarily card aligned.
7613 if (PrintCMSStatistics != 0) {
7614 _num_dirty_cards +=
7615 mr.word_size()/CardTableModRefBS::card_size_in_words;
7616 }
7617 _space->object_iterate_mem(mr, &_scan_cl);
7618 }
7619
7620 SweepClosure::SweepClosure(CMSCollector* collector,
7621 ConcurrentMarkSweepGeneration* g,
7622 CMSBitMap* bitMap, bool should_yield) :
7623 _collector(collector),
7624 _g(g),
7625 _sp(g->cmsSpace()),
7626 _limit(_sp->sweep_limit()),
7627 _freelistLock(_sp->freelistLock()),
7628 _bitMap(bitMap),
7629 _yield(should_yield),
7630 _inFreeRange(false), // No free range at beginning of sweep
7631 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7632 _lastFreeRangeCoalesced(false),
7633 _freeFinger(g->used_region().start())
7634 {
7635 NOT_PRODUCT(
7636 _numObjectsFreed = 0;
7637 _numWordsFreed = 0;
7638 _numObjectsLive = 0;
7639 _numWordsLive = 0;
7640 _numObjectsAlreadyFree = 0;
7641 _numWordsAlreadyFree = 0;
7642 _last_fc = NULL;
7643
7644 _sp->initializeIndexedFreeListArrayReturnedBytes();
7645 _sp->dictionary()->initializeDictReturnedBytes();
7646 )
7647 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7648 "sweep _limit out of bounds");
7649 if (CMSTraceSweeper) {
7650 gclog_or_tty->print("\n====================\nStarting new sweep\n");
7651 }
7652 }
7653
7654 // We need this destructor to reclaim any space at the end
7655 // of the space, which do_blk below may not have added back to
7656 // the free lists. [basically dealing with the "fringe effect"]
7657 SweepClosure::~SweepClosure() {
7658 assert_lock_strong(_freelistLock);
7659 // this should be treated as the end of a free run if any
7660 // The current free range should be returned to the free lists
7661 // as one coalesced chunk.
7662 if (inFreeRange()) {
7663 flushCurFreeChunk(freeFinger(),
7664 pointer_delta(_limit, freeFinger()));
7665 assert(freeFinger() < _limit, "the finger pointeth off base");
7666 if (CMSTraceSweeper) {
7667 gclog_or_tty->print("destructor:");
7668 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
7669 "[coalesced:"SIZE_FORMAT"]\n",
7670 freeFinger(), pointer_delta(_limit, freeFinger()),
7671 lastFreeRangeCoalesced());
7672 }
7673 }
7674 NOT_PRODUCT(
7675 if (Verbose && PrintGC) {
7676 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7677 SIZE_FORMAT " bytes",
7678 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7679 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7680 SIZE_FORMAT" bytes "
7681 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7682 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7683 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7684 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7685 sizeof(HeapWord);
7686 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7687
7688 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7689 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7690 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7691 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7692 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7693 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7694 indexListReturnedBytes);
7695 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7696 dictReturnedBytes);
7697 }
7698 }
7699 )
7700 // Now, in debug mode, just null out the sweep_limit
7701 NOT_PRODUCT(_sp->clear_sweep_limit();)
7702 if (CMSTraceSweeper) {
7703 gclog_or_tty->print("end of sweep\n================\n");
7704 }
7705 }
7706
7707 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7708 bool freeRangeInFreeLists) {
7709 if (CMSTraceSweeper) {
7710 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
7711 freeFinger, _sp->block_size(freeFinger),
7712 freeRangeInFreeLists);
7713 }
7714 assert(!inFreeRange(), "Trampling existing free range");
7715 set_inFreeRange(true);
7716 set_lastFreeRangeCoalesced(false);
7717
7718 set_freeFinger(freeFinger);
7719 set_freeRangeInFreeLists(freeRangeInFreeLists);
7720 if (CMSTestInFreeList) {
7721 if (freeRangeInFreeLists) {
7722 FreeChunk* fc = (FreeChunk*) freeFinger;
7723 assert(fc->isFree(), "A chunk on the free list should be free.");
7724 assert(fc->size() > 0, "Free range should have a size");
7725 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7726 }
7727 }
7728 }
7729
7730 // Note that the sweeper runs concurrently with mutators. Thus,
7731 // it is possible for direct allocation in this generation to happen
7732 // in the middle of the sweep. Note that the sweeper also coalesces
7733 // contiguous free blocks. Thus, unless the sweeper and the allocator
7734 // synchronize appropriately freshly allocated blocks may get swept up.
7735 // This is accomplished by the sweeper locking the free lists while
7736 // it is sweeping. Thus blocks that are determined to be free are
7737 // indeed free. There is however one additional complication:
7738 // blocks that have been allocated since the final checkpoint and
7739 // mark, will not have been marked and so would be treated as
7740 // unreachable and swept up. To prevent this, the allocator marks
7741 // the bit map when allocating during the sweep phase. This leads,
7742 // however, to a further complication -- objects may have been allocated
7743 // but not yet initialized -- in the sense that the header isn't yet
7744 // installed. The sweeper can not then determine the size of the block
7745 // in order to skip over it. To deal with this case, we use a technique
7746 // (due to Printezis) to encode such uninitialized block sizes in the
7747 // bit map. Since the bit map uses a bit per every HeapWord, but the
7748 // CMS generation has a minimum object size of 3 HeapWords, it follows
7749 // that "normal marks" won't be adjacent in the bit map (there will
7750 // always be at least two 0 bits between successive 1 bits). We make use
7751 // of these "unused" bits to represent uninitialized blocks -- the bit
7752 // corresponding to the start of the uninitialized object and the next
7753 // bit are both set. Finally, a 1 bit marks the end of the object that
7754 // started with the two consecutive 1 bits to indicate its potentially
7755 // uninitialized state.
7756
7757 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7758 FreeChunk* fc = (FreeChunk*)addr;
7759 size_t res;
7760
7761 // check if we are done sweepinrg
7762 if (addr == _limit) { // we have swept up to the limit, do nothing more
7763 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7764 "sweep _limit out of bounds");
7765 // help the closure application finish
7766 return pointer_delta(_sp->end(), _limit);
7767 }
7768 assert(addr <= _limit, "sweep invariant");
7769
7770 // check if we should yield
7771 do_yield_check(addr);
7772 if (fc->isFree()) {
7773 // Chunk that is already free
7774 res = fc->size();
7775 doAlreadyFreeChunk(fc);
7776 debug_only(_sp->verifyFreeLists());
7777 assert(res == fc->size(), "Don't expect the size to change");
7778 NOT_PRODUCT(
7779 _numObjectsAlreadyFree++;
7780 _numWordsAlreadyFree += res;
7781 )
7782 NOT_PRODUCT(_last_fc = fc;)
7783 } else if (!_bitMap->isMarked(addr)) {
7784 // Chunk is fresh garbage
7785 res = doGarbageChunk(fc);
7786 debug_only(_sp->verifyFreeLists());
7787 NOT_PRODUCT(
7788 _numObjectsFreed++;
7789 _numWordsFreed += res;
7790 )
7791 } else {
7792 // Chunk that is alive.
7793 res = doLiveChunk(fc);
7794 debug_only(_sp->verifyFreeLists());
7795 NOT_PRODUCT(
7796 _numObjectsLive++;
7797 _numWordsLive += res;
7798 )
7799 }
7800 return res;
7801 }
7802
7803 // For the smart allocation, record following
7804 // split deaths - a free chunk is removed from its free list because
7805 // it is being split into two or more chunks.
7806 // split birth - a free chunk is being added to its free list because
7807 // a larger free chunk has been split and resulted in this free chunk.
7808 // coal death - a free chunk is being removed from its free list because
7809 // it is being coalesced into a large free chunk.
7810 // coal birth - a free chunk is being added to its free list because
7811 // it was created when two or more free chunks where coalesced into
7812 // this free chunk.
7813 //
7814 // These statistics are used to determine the desired number of free
7815 // chunks of a given size. The desired number is chosen to be relative
7816 // to the end of a CMS sweep. The desired number at the end of a sweep
7817 // is the
7818 // count-at-end-of-previous-sweep (an amount that was enough)
7819 // - count-at-beginning-of-current-sweep (the excess)
7820 // + split-births (gains in this size during interval)
7821 // - split-deaths (demands on this size during interval)
7822 // where the interval is from the end of one sweep to the end of the
7823 // next.
7824 //
7825 // When sweeping the sweeper maintains an accumulated chunk which is
7826 // the chunk that is made up of chunks that have been coalesced. That
7827 // will be termed the left-hand chunk. A new chunk of garbage that
7828 // is being considered for coalescing will be referred to as the
7829 // right-hand chunk.
7830 //
7831 // When making a decision on whether to coalesce a right-hand chunk with
7832 // the current left-hand chunk, the current count vs. the desired count
7833 // of the left-hand chunk is considered. Also if the right-hand chunk
7834 // is near the large chunk at the end of the heap (see
7835 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7836 // left-hand chunk is coalesced.
7837 //
7838 // When making a decision about whether to split a chunk, the desired count
7839 // vs. the current count of the candidate to be split is also considered.
7840 // If the candidate is underpopulated (currently fewer chunks than desired)
7841 // a chunk of an overpopulated (currently more chunks than desired) size may
7842 // be chosen. The "hint" associated with a free list, if non-null, points
7843 // to a free list which may be overpopulated.
7844 //
7845
7846 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
7847 size_t size = fc->size();
7848 // Chunks that cannot be coalesced are not in the
7849 // free lists.
7850 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7851 assert(_sp->verifyChunkInFreeLists(fc),
7852 "free chunk should be in free lists");
7853 }
7854 // a chunk that is already free, should not have been
7855 // marked in the bit map
7856 HeapWord* addr = (HeapWord*) fc;
7857 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7858 // Verify that the bit map has no bits marked between
7859 // addr and purported end of this block.
7860 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7861
7862 // Some chunks cannot be coalesced in under any circumstances.
7863 // See the definition of cantCoalesce().
7864 if (!fc->cantCoalesce()) {
7865 // This chunk can potentially be coalesced.
7866 if (_sp->adaptive_freelists()) {
7867 // All the work is done in
7868 doPostIsFreeOrGarbageChunk(fc, size);
7869 } else { // Not adaptive free lists
7870 // this is a free chunk that can potentially be coalesced by the sweeper;
7871 if (!inFreeRange()) {
7872 // if the next chunk is a free block that can't be coalesced
7873 // it doesn't make sense to remove this chunk from the free lists
7874 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7875 assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
7876 if ((HeapWord*)nextChunk < _limit && // there's a next chunk...
7877 nextChunk->isFree() && // which is free...
7878 nextChunk->cantCoalesce()) { // ... but cant be coalesced
7879 // nothing to do
7880 } else {
7881 // Potentially the start of a new free range:
7882 // Don't eagerly remove it from the free lists.
7883 // No need to remove it if it will just be put
7884 // back again. (Also from a pragmatic point of view
7885 // if it is a free block in a region that is beyond
7886 // any allocated blocks, an assertion will fail)
7887 // Remember the start of a free run.
7888 initialize_free_range(addr, true);
7889 // end - can coalesce with next chunk
7890 }
7891 } else {
7892 // the midst of a free range, we are coalescing
7893 debug_only(record_free_block_coalesced(fc);)
7894 if (CMSTraceSweeper) {
7895 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
7896 }
7897 // remove it from the free lists
7898 _sp->removeFreeChunkFromFreeLists(fc);
7899 set_lastFreeRangeCoalesced(true);
7900 // If the chunk is being coalesced and the current free range is
7901 // in the free lists, remove the current free range so that it
7902 // will be returned to the free lists in its entirety - all
7903 // the coalesced pieces included.
7904 if (freeRangeInFreeLists()) {
7905 FreeChunk* ffc = (FreeChunk*) freeFinger();
7906 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7907 "Size of free range is inconsistent with chunk size.");
7908 if (CMSTestInFreeList) {
7909 assert(_sp->verifyChunkInFreeLists(ffc),
7910 "free range is not in free lists");
7911 }
7912 _sp->removeFreeChunkFromFreeLists(ffc);
7913 set_freeRangeInFreeLists(false);
7914 }
7915 }
7916 }
7917 } else {
7918 // Code path common to both original and adaptive free lists.
7919
7920 // cant coalesce with previous block; this should be treated
7921 // as the end of a free run if any
7922 if (inFreeRange()) {
7923 // we kicked some butt; time to pick up the garbage
7924 assert(freeFinger() < addr, "the finger pointeth off base");
7925 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
7926 }
7927 // else, nothing to do, just continue
7928 }
7929 }
7930
7931 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
7932 // This is a chunk of garbage. It is not in any free list.
7933 // Add it to a free list or let it possibly be coalesced into
7934 // a larger chunk.
7935 HeapWord* addr = (HeapWord*) fc;
7936 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7937
7938 if (_sp->adaptive_freelists()) {
7939 // Verify that the bit map has no bits marked between
7940 // addr and purported end of just dead object.
7941 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7942
7943 doPostIsFreeOrGarbageChunk(fc, size);
7944 } else {
7945 if (!inFreeRange()) {
7946 // start of a new free range
7947 assert(size > 0, "A free range should have a size");
7948 initialize_free_range(addr, false);
7949
7950 } else {
7951 // this will be swept up when we hit the end of the
7952 // free range
7953 if (CMSTraceSweeper) {
7954 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
7955 }
7956 // If the chunk is being coalesced and the current free range is
7957 // in the free lists, remove the current free range so that it
7958 // will be returned to the free lists in its entirety - all
7959 // the coalesced pieces included.
7960 if (freeRangeInFreeLists()) {
7961 FreeChunk* ffc = (FreeChunk*)freeFinger();
7962 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7963 "Size of free range is inconsistent with chunk size.");
7964 if (CMSTestInFreeList) {
7965 assert(_sp->verifyChunkInFreeLists(ffc),
7966 "free range is not in free lists");
7967 }
7968 _sp->removeFreeChunkFromFreeLists(ffc);
7969 set_freeRangeInFreeLists(false);
7970 }
7971 set_lastFreeRangeCoalesced(true);
7972 }
7973 // this will be swept up when we hit the end of the free range
7974
7975 // Verify that the bit map has no bits marked between
7976 // addr and purported end of just dead object.
7977 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7978 }
7979 return size;
7980 }
7981
7982 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
7983 HeapWord* addr = (HeapWord*) fc;
7984 // The sweeper has just found a live object. Return any accumulated
7985 // left hand chunk to the free lists.
7986 if (inFreeRange()) {
7987 if (_sp->adaptive_freelists()) {
7988 flushCurFreeChunk(freeFinger(),
7989 pointer_delta(addr, freeFinger()));
7990 } else { // not adaptive freelists
7991 set_inFreeRange(false);
7992 // Add the free range back to the free list if it is not already
7993 // there.
7994 if (!freeRangeInFreeLists()) {
7995 assert(freeFinger() < addr, "the finger pointeth off base");
7996 if (CMSTraceSweeper) {
7997 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
7998 "[coalesced:%d]\n",
7999 freeFinger(), pointer_delta(addr, freeFinger()),
8000 lastFreeRangeCoalesced());
8001 }
8002 _sp->addChunkAndRepairOffsetTable(freeFinger(),
8003 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
8004 }
8005 }
8006 }
8007
8008 // Common code path for original and adaptive free lists.
8009
8010 // this object is live: we'd normally expect this to be
8011 // an oop, and like to assert the following:
8012 // assert(oop(addr)->is_oop(), "live block should be an oop");
8013 // However, as we commented above, this may be an object whose
8014 // header hasn't yet been initialized.
8015 size_t size;
8016 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8017 if (_bitMap->isMarked(addr + 1)) {
8018 // Determine the size from the bit map, rather than trying to
8019 // compute it from the object header.
8020 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8021 size = pointer_delta(nextOneAddr + 1, addr);
8022 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8023 "alignment problem");
8024
8025 #ifdef DEBUG
8026 if (oop(addr)->klass() != NULL &&
8027 ( !_collector->should_unload_classes()
8028 || oop(addr)->is_parsable())) {
8029 // Ignore mark word because we are running concurrent with mutators
8030 assert(oop(addr)->is_oop(true), "live block should be an oop");
8031 assert(size ==
8032 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8033 "P-mark and computed size do not agree");
8034 }
8035 #endif
8036
8037 } else {
8038 // This should be an initialized object that's alive.
8039 assert(oop(addr)->klass() != NULL &&
8040 (!_collector->should_unload_classes()
8041 || oop(addr)->is_parsable()),
8042 "Should be an initialized object");
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 // Verify that the bit map has no bits marked between
8046 // addr and purported end of this block.
8047 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8048 assert(size >= 3, "Necessary for Printezis marks to work");
8049 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8050 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8051 }
8052 return size;
8053 }
8054
8055 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
8056 size_t chunkSize) {
8057 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
8058 // scheme.
8059 bool fcInFreeLists = fc->isFree();
8060 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8061 assert((HeapWord*)fc <= _limit, "sweep invariant");
8062 if (CMSTestInFreeList && fcInFreeLists) {
8063 assert(_sp->verifyChunkInFreeLists(fc),
8064 "free chunk is not in free lists");
8065 }
8066
8067
8068 if (CMSTraceSweeper) {
8069 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8070 }
8071
8072 HeapWord* addr = (HeapWord*) fc;
8073
8074 bool coalesce;
8075 size_t left = pointer_delta(addr, freeFinger());
8076 size_t right = chunkSize;
8077 switch (FLSCoalescePolicy) {
8078 // numeric value forms a coalition aggressiveness metric
8079 case 0: { // never coalesce
8080 coalesce = false;
8081 break;
8082 }
8083 case 1: { // coalesce if left & right chunks on overpopulated lists
8084 coalesce = _sp->coalOverPopulated(left) &&
8085 _sp->coalOverPopulated(right);
8086 break;
8087 }
8088 case 2: { // coalesce if left chunk on overpopulated list (default)
8089 coalesce = _sp->coalOverPopulated(left);
8090 break;
8091 }
8092 case 3: { // coalesce if left OR right chunk on overpopulated list
8093 coalesce = _sp->coalOverPopulated(left) ||
8094 _sp->coalOverPopulated(right);
8095 break;
8096 }
8097 case 4: { // always coalesce
8098 coalesce = true;
8099 break;
8100 }
8101 default:
8102 ShouldNotReachHere();
8103 }
8104
8105 // Should the current free range be coalesced?
8106 // If the chunk is in a free range and either we decided to coalesce above
8107 // or the chunk is near the large block at the end of the heap
8108 // (isNearLargestChunk() returns true), then coalesce this chunk.
8109 bool doCoalesce = inFreeRange() &&
8110 (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8111 if (doCoalesce) {
8112 // Coalesce the current free range on the left with the new
8113 // chunk on the right. If either is on a free list,
8114 // it must be removed from the list and stashed in the closure.
8115 if (freeRangeInFreeLists()) {
8116 FreeChunk* ffc = (FreeChunk*)freeFinger();
8117 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8118 "Size of free range is inconsistent with chunk size.");
8119 if (CMSTestInFreeList) {
8120 assert(_sp->verifyChunkInFreeLists(ffc),
8121 "Chunk is not in free lists");
8122 }
8123 _sp->coalDeath(ffc->size());
8124 _sp->removeFreeChunkFromFreeLists(ffc);
8125 set_freeRangeInFreeLists(false);
8126 }
8127 if (fcInFreeLists) {
8128 _sp->coalDeath(chunkSize);
8129 assert(fc->size() == chunkSize,
8130 "The chunk has the wrong size or is not in the free lists");
8131 _sp->removeFreeChunkFromFreeLists(fc);
8132 }
8133 set_lastFreeRangeCoalesced(true);
8134 } else { // not in a free range and/or should not coalesce
8135 // Return the current free range and start a new one.
8136 if (inFreeRange()) {
8137 // In a free range but cannot coalesce with the right hand chunk.
8138 // Put the current free range into the free lists.
8139 flushCurFreeChunk(freeFinger(),
8140 pointer_delta(addr, freeFinger()));
8141 }
8142 // Set up for new free range. Pass along whether the right hand
8143 // chunk is in the free lists.
8144 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8145 }
8146 }
8147 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
8148 assert(inFreeRange(), "Should only be called if currently in a free range.");
8149 assert(size > 0,
8150 "A zero sized chunk cannot be added to the free lists.");
8151 if (!freeRangeInFreeLists()) {
8152 if(CMSTestInFreeList) {
8153 FreeChunk* fc = (FreeChunk*) chunk;
8154 fc->setSize(size);
8155 assert(!_sp->verifyChunkInFreeLists(fc),
8156 "chunk should not be in free lists yet");
8157 }
8158 if (CMSTraceSweeper) {
8159 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8160 chunk, size);
8161 }
8162 // A new free range is going to be starting. The current
8163 // free range has not been added to the free lists yet or
8164 // was removed so add it back.
8165 // If the current free range was coalesced, then the death
8166 // of the free range was recorded. Record a birth now.
8167 if (lastFreeRangeCoalesced()) {
8168 _sp->coalBirth(size);
8169 }
8170 _sp->addChunkAndRepairOffsetTable(chunk, size,
8171 lastFreeRangeCoalesced());
8172 }
8173 set_inFreeRange(false);
8174 set_freeRangeInFreeLists(false);
8175 }
8176
8177 // We take a break if we've been at this for a while,
8178 // so as to avoid monopolizing the locks involved.
8179 void SweepClosure::do_yield_work(HeapWord* addr) {
8180 // Return current free chunk being used for coalescing (if any)
8181 // to the appropriate freelist. After yielding, the next
8182 // free block encountered will start a coalescing range of
8183 // free blocks. If the next free block is adjacent to the
8184 // chunk just flushed, they will need to wait for the next
8185 // sweep to be coalesced.
8186 if (inFreeRange()) {
8187 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
8188 }
8189
8190 // First give up the locks, then yield, then re-lock.
8191 // We should probably use a constructor/destructor idiom to
8192 // do this unlock/lock or modify the MutexUnlocker class to
8193 // serve our purpose. XXX
8194 assert_lock_strong(_bitMap->lock());
8195 assert_lock_strong(_freelistLock);
8196 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8197 "CMS thread should hold CMS token");
8198 _bitMap->lock()->unlock();
8199 _freelistLock->unlock();
8200 ConcurrentMarkSweepThread::desynchronize(true);
8201 ConcurrentMarkSweepThread::acknowledge_yield_request();
8202 _collector->stopTimer();
8203 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8204 if (PrintCMSStatistics != 0) {
8205 _collector->incrementYields();
8206 }
8207 _collector->icms_wait();
8208
8209 // See the comment in coordinator_yield()
8210 for (unsigned i = 0; i < CMSYieldSleepCount &&
8211 ConcurrentMarkSweepThread::should_yield() &&
8212 !CMSCollector::foregroundGCIsActive(); ++i) {
8213 os::sleep(Thread::current(), 1, false);
8214 ConcurrentMarkSweepThread::acknowledge_yield_request();
8215 }
8216
8217 ConcurrentMarkSweepThread::synchronize(true);
8218 _freelistLock->lock();
8219 _bitMap->lock()->lock_without_safepoint_check();
8220 _collector->startTimer();
8221 }
8222
8223 #ifndef PRODUCT
8224 // This is actually very useful in a product build if it can
8225 // be called from the debugger. Compile it into the product
8226 // as needed.
8227 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8228 return debug_cms_space->verifyChunkInFreeLists(fc);
8229 }
8230
8231 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8232 if (CMSTraceSweeper) {
8233 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8234 }
8235 }
8236 #endif
8237
8238 // CMSIsAliveClosure
8239 bool CMSIsAliveClosure::do_object_b(oop obj) {
8240 HeapWord* addr = (HeapWord*)obj;
8241 return addr != NULL &&
8242 (!_span.contains(addr) || _bit_map->isMarked(addr));
8243 }
8244
8245 // CMSKeepAliveClosure: the serial version
8246 void CMSKeepAliveClosure::do_oop(oop obj) {
8247 HeapWord* addr = (HeapWord*)obj;
8248 if (_span.contains(addr) &&
8249 !_bit_map->isMarked(addr)) {
8250 _bit_map->mark(addr);
8251 bool simulate_overflow = false;
8252 NOT_PRODUCT(
8253 if (CMSMarkStackOverflowALot &&
8254 _collector->simulate_overflow()) {
8255 // simulate a stack overflow
8256 simulate_overflow = true;
8257 }
8258 )
8259 if (simulate_overflow || !_mark_stack->push(obj)) {
8260 _collector->push_on_overflow_list(obj);
8261 _collector->_ser_kac_ovflw++;
8262 }
8263 }
8264 }
8265
8266 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8267 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8268
8269 // CMSParKeepAliveClosure: a parallel version of the above.
8270 // The work queues are private to each closure (thread),
8271 // but (may be) available for stealing by other threads.
8272 void CMSParKeepAliveClosure::do_oop(oop obj) {
8273 HeapWord* addr = (HeapWord*)obj;
8274 if (_span.contains(addr) &&
8275 !_bit_map->isMarked(addr)) {
8276 // In general, during recursive tracing, several threads
8277 // may be concurrently getting here; the first one to
8278 // "tag" it, claims it.
8279 if (_bit_map->par_mark(addr)) {
8280 bool res = _work_queue->push(obj);
8281 assert(res, "Low water mark should be much less than capacity");
8282 // Do a recursive trim in the hope that this will keep
8283 // stack usage lower, but leave some oops for potential stealers
8284 trim_queue(_low_water_mark);
8285 } // Else, another thread got there first
8286 }
8287 }
8288
8289 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8290 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8291
8292 void CMSParKeepAliveClosure::trim_queue(uint max) {
8293 while (_work_queue->size() > max) {
8294 oop new_oop;
8295 if (_work_queue->pop_local(new_oop)) {
8296 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8297 assert(_bit_map->isMarked((HeapWord*)new_oop),
8298 "no white objects on this stack!");
8299 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8300 // iterate over the oops in this oop, marking and pushing
8301 // the ones in CMS heap (i.e. in _span).
8302 new_oop->oop_iterate(&_mark_and_push);
8303 }
8304 }
8305 }
8306
8307 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8308 HeapWord* addr = (HeapWord*)obj;
8309 if (_span.contains(addr) &&
8310 !_bit_map->isMarked(addr)) {
8311 if (_bit_map->par_mark(addr)) {
8312 bool simulate_overflow = false;
8313 NOT_PRODUCT(
8314 if (CMSMarkStackOverflowALot &&
8315 _collector->par_simulate_overflow()) {
8316 // simulate a stack overflow
8317 simulate_overflow = true;
8318 }
8319 )
8320 if (simulate_overflow || !_work_queue->push(obj)) {
8321 _collector->par_push_on_overflow_list(obj);
8322 _collector->_par_kac_ovflw++;
8323 }
8324 } // Else another thread got there already
8325 }
8326 }
8327
8328 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8329 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8330
8331 //////////////////////////////////////////////////////////////////
8332 // CMSExpansionCause /////////////////////////////
8333 //////////////////////////////////////////////////////////////////
8334 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8335 switch (cause) {
8336 case _no_expansion:
8337 return "No expansion";
8338 case _satisfy_free_ratio:
8339 return "Free ratio";
8340 case _satisfy_promotion:
8341 return "Satisfy promotion";
8342 case _satisfy_allocation:
8343 return "allocation";
8344 case _allocate_par_lab:
8345 return "Par LAB";
8346 case _allocate_par_spooling_space:
8347 return "Par Spooling Space";
8348 case _adaptive_size_policy:
8349 return "Ergonomics";
8350 default:
8351 return "unknown";
8352 }
8353 }
8354
8355 void CMSDrainMarkingStackClosure::do_void() {
8356 // the max number to take from overflow list at a time
8357 const size_t num = _mark_stack->capacity()/4;
8358 while (!_mark_stack->isEmpty() ||
8359 // if stack is empty, check the overflow list
8360 _collector->take_from_overflow_list(num, _mark_stack)) {
8361 oop obj = _mark_stack->pop();
8362 HeapWord* addr = (HeapWord*)obj;
8363 assert(_span.contains(addr), "Should be within span");
8364 assert(_bit_map->isMarked(addr), "Should be marked");
8365 assert(obj->is_oop(), "Should be an oop");
8366 obj->oop_iterate(_keep_alive);
8367 }
8368 }
8369
8370 void CMSParDrainMarkingStackClosure::do_void() {
8371 // drain queue
8372 trim_queue(0);
8373 }
8374
8375 // Trim our work_queue so its length is below max at return
8376 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8377 while (_work_queue->size() > max) {
8378 oop new_oop;
8379 if (_work_queue->pop_local(new_oop)) {
8380 assert(new_oop->is_oop(), "Expected an oop");
8381 assert(_bit_map->isMarked((HeapWord*)new_oop),
8382 "no white objects on this stack!");
8383 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8384 // iterate over the oops in this oop, marking and pushing
8385 // the ones in CMS heap (i.e. in _span).
8386 new_oop->oop_iterate(&_mark_and_push);
8387 }
8388 }
8389 }
8390
8391 ////////////////////////////////////////////////////////////////////
8392 // Support for Marking Stack Overflow list handling and related code
8393 ////////////////////////////////////////////////////////////////////
8394 // Much of the following code is similar in shape and spirit to the
8395 // code used in ParNewGC. We should try and share that code
8396 // as much as possible in the future.
8397
8398 #ifndef PRODUCT
8399 // Debugging support for CMSStackOverflowALot
8400
8401 // It's OK to call this multi-threaded; the worst thing
8402 // that can happen is that we'll get a bunch of closely
8403 // spaced simulated oveflows, but that's OK, in fact
8404 // probably good as it would exercise the overflow code
8405 // under contention.
8406 bool CMSCollector::simulate_overflow() {
8407 if (_overflow_counter-- <= 0) { // just being defensive
8408 _overflow_counter = CMSMarkStackOverflowInterval;
8409 return true;
8410 } else {
8411 return false;
8412 }
8413 }
8414
8415 bool CMSCollector::par_simulate_overflow() {
8416 return simulate_overflow();
8417 }
8418 #endif
8419
8420 // Single-threaded
8421 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8422 assert(stack->isEmpty(), "Expected precondition");
8423 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8424 size_t i = num;
8425 oop cur = _overflow_list;
8426 const markOop proto = markOopDesc::prototype();
8427 NOT_PRODUCT(size_t n = 0;)
8428 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8429 next = oop(cur->mark());
8430 cur->set_mark(proto); // until proven otherwise
8431 assert(cur->is_oop(), "Should be an oop");
8432 bool res = stack->push(cur);
8433 assert(res, "Bit off more than can chew?");
8434 NOT_PRODUCT(n++;)
8435 }
8436 _overflow_list = cur;
8437 #ifndef PRODUCT
8438 assert(_num_par_pushes >= n, "Too many pops?");
8439 _num_par_pushes -=n;
8440 #endif
8441 return !stack->isEmpty();
8442 }
8443
8444 // Multi-threaded; use CAS to break off a prefix
8445 bool CMSCollector::par_take_from_overflow_list(size_t num,
8446 OopTaskQueue* work_q) {
8447 assert(work_q->size() == 0, "That's the current policy");
8448 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8449 if (_overflow_list == NULL) {
8450 return false;
8451 }
8452 // Grab the entire list; we'll put back a suffix
8453 oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
8454 if (prefix == NULL) { // someone grabbed it before we did ...
8455 // ... we could spin for a short while, but for now we don't
8456 return false;
8457 }
8458 size_t i = num;
8459 oop cur = prefix;
8460 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8461 if (cur->mark() != NULL) {
8462 oop suffix_head = cur->mark(); // suffix will be put back on global list
8463 cur->set_mark(NULL); // break off suffix
8464 // Find tail of suffix so we can prepend suffix to global list
8465 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8466 oop suffix_tail = cur;
8467 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8468 "Tautology");
8469 oop observed_overflow_list = _overflow_list;
8470 do {
8471 cur = observed_overflow_list;
8472 suffix_tail->set_mark(markOop(cur));
8473 observed_overflow_list =
8474 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
8475 } while (cur != observed_overflow_list);
8476 }
8477
8478 // Push the prefix elements on work_q
8479 assert(prefix != NULL, "control point invariant");
8480 const markOop proto = markOopDesc::prototype();
8481 oop next;
8482 NOT_PRODUCT(size_t n = 0;)
8483 for (cur = prefix; cur != NULL; cur = next) {
8484 next = oop(cur->mark());
8485 cur->set_mark(proto); // until proven otherwise
8486 assert(cur->is_oop(), "Should be an oop");
8487 bool res = work_q->push(cur);
8488 assert(res, "Bit off more than we can chew?");
8489 NOT_PRODUCT(n++;)
8490 }
8491 #ifndef PRODUCT
8492 assert(_num_par_pushes >= n, "Too many pops?");
8493 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8494 #endif
8495 return true;
8496 }
8497
8498 // Single-threaded
8499 void CMSCollector::push_on_overflow_list(oop p) {
8500 NOT_PRODUCT(_num_par_pushes++;)
8501 assert(p->is_oop(), "Not an oop");
8502 preserve_mark_if_necessary(p);
8503 p->set_mark((markOop)_overflow_list);
8504 _overflow_list = p;
8505 }
8506
8507 // Multi-threaded; use CAS to prepend to overflow list
8508 void CMSCollector::par_push_on_overflow_list(oop p) {
8509 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8510 assert(p->is_oop(), "Not an oop");
8511 par_preserve_mark_if_necessary(p);
8512 oop observed_overflow_list = _overflow_list;
8513 oop cur_overflow_list;
8514 do {
8515 cur_overflow_list = observed_overflow_list;
8516 p->set_mark(markOop(cur_overflow_list));
8517 observed_overflow_list =
8518 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8519 } while (cur_overflow_list != observed_overflow_list);
8520 }
8521
8522 // Single threaded
8523 // General Note on GrowableArray: pushes may silently fail
8524 // because we are (temporarily) out of C-heap for expanding
8525 // the stack. The problem is quite ubiquitous and affects
8526 // a lot of code in the JVM. The prudent thing for GrowableArray
8527 // to do (for now) is to exit with an error. However, that may
8528 // be too draconian in some cases because the caller may be
8529 // able to recover without much harm. For suych cases, we
8530 // should probably introduce a "soft_push" method which returns
8531 // an indication of success or failure with the assumption that
8532 // the caller may be able to recover from a failure; code in
8533 // the VM can then be changed, incrementally, to deal with such
8534 // failures where possible, thus, incrementally hardening the VM
8535 // in such low resource situations.
8536 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8537 int PreserveMarkStackSize = 128;
8538
8539 if (_preserved_oop_stack == NULL) {
8540 assert(_preserved_mark_stack == NULL,
8541 "bijection with preserved_oop_stack");
8542 // Allocate the stacks
8543 _preserved_oop_stack = new (ResourceObj::C_HEAP)
8544 GrowableArray<oop>(PreserveMarkStackSize, true);
8545 _preserved_mark_stack = new (ResourceObj::C_HEAP)
8546 GrowableArray<markOop>(PreserveMarkStackSize, true);
8547 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
8548 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
8549 "Preserved Mark/Oop Stack for CMS (C-heap)");
8550 }
8551 }
8552 _preserved_oop_stack->push(p);
8553 _preserved_mark_stack->push(m);
8554 assert(m == p->mark(), "Mark word changed");
8555 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
8556 "bijection");
8557 }
8558
8559 // Single threaded
8560 void CMSCollector::preserve_mark_if_necessary(oop p) {
8561 markOop m = p->mark();
8562 if (m->must_be_preserved(p)) {
8563 preserve_mark_work(p, m);
8564 }
8565 }
8566
8567 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8568 markOop m = p->mark();
8569 if (m->must_be_preserved(p)) {
8570 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8571 // Even though we read the mark word without holding
8572 // the lock, we are assured that it will not change
8573 // because we "own" this oop, so no other thread can
8574 // be trying to push it on the overflow list; see
8575 // the assertion in preserve_mark_work() that checks
8576 // that m == p->mark().
8577 preserve_mark_work(p, m);
8578 }
8579 }
8580
8581 // We should be able to do this multi-threaded,
8582 // a chunk of stack being a task (this is
8583 // correct because each oop only ever appears
8584 // once in the overflow list. However, it's
8585 // not very easy to completely overlap this with
8586 // other operations, so will generally not be done
8587 // until all work's been completed. Because we
8588 // expect the preserved oop stack (set) to be small,
8589 // it's probably fine to do this single-threaded.
8590 // We can explore cleverer concurrent/overlapped/parallel
8591 // processing of preserved marks if we feel the
8592 // need for this in the future. Stack overflow should
8593 // be so rare in practice and, when it happens, its
8594 // effect on performance so great that this will
8595 // likely just be in the noise anyway.
8596 void CMSCollector::restore_preserved_marks_if_any() {
8597 if (_preserved_oop_stack == NULL) {
8598 assert(_preserved_mark_stack == NULL,
8599 "bijection with preserved_oop_stack");
8600 return;
8601 }
8602
8603 assert(SafepointSynchronize::is_at_safepoint(),
8604 "world should be stopped");
8605 assert(Thread::current()->is_ConcurrentGC_thread() ||
8606 Thread::current()->is_VM_thread(),
8607 "should be single-threaded");
8608
8609 int length = _preserved_oop_stack->length();
8610 assert(_preserved_mark_stack->length() == length, "bijection");
8611 for (int i = 0; i < length; i++) {
8612 oop p = _preserved_oop_stack->at(i);
8613 assert(p->is_oop(), "Should be an oop");
8614 assert(_span.contains(p), "oop should be in _span");
8615 assert(p->mark() == markOopDesc::prototype(),
8616 "Set when taken from overflow list");
8617 markOop m = _preserved_mark_stack->at(i);
8618 p->set_mark(m);
8619 }
8620 _preserved_mark_stack->clear();
8621 _preserved_oop_stack->clear();
8622 assert(_preserved_mark_stack->is_empty() &&
8623 _preserved_oop_stack->is_empty(),
8624 "stacks were cleared above");
8625 }
8626
8627 #ifndef PRODUCT
8628 bool CMSCollector::no_preserved_marks() const {
8629 return ( ( _preserved_mark_stack == NULL
8630 && _preserved_oop_stack == NULL)
8631 || ( _preserved_mark_stack->is_empty()
8632 && _preserved_oop_stack->is_empty()));
8633 }
8634 #endif
8635
8636 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8637 {
8638 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8639 CMSAdaptiveSizePolicy* size_policy =
8640 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8641 assert(size_policy->is_gc_cms_adaptive_size_policy(),
8642 "Wrong type for size policy");
8643 return size_policy;
8644 }
8645
8646 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
8647 size_t desired_promo_size) {
8648 if (cur_promo_size < desired_promo_size) {
8649 size_t expand_bytes = desired_promo_size - cur_promo_size;
8650 if (PrintAdaptiveSizePolicy && Verbose) {
8651 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8652 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
8653 expand_bytes);
8654 }
8655 expand(expand_bytes,
8656 MinHeapDeltaBytes,
8657 CMSExpansionCause::_adaptive_size_policy);
8658 } else if (desired_promo_size < cur_promo_size) {
8659 size_t shrink_bytes = cur_promo_size - desired_promo_size;
8660 if (PrintAdaptiveSizePolicy && Verbose) {
8661 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8662 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
8663 shrink_bytes);
8664 }
8665 shrink(shrink_bytes);
8666 }
8667 }
8668
8669 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
8670 GenCollectedHeap* gch = GenCollectedHeap::heap();
8671 CMSGCAdaptivePolicyCounters* counters =
8672 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
8673 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
8674 "Wrong kind of counters");
8675 return counters;
8676 }
8677
8678
8679 void ASConcurrentMarkSweepGeneration::update_counters() {
8680 if (UsePerfData) {
8681 _space_counters->update_all();
8682 _gen_counters->update_all();
8683 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8684 GenCollectedHeap* gch = GenCollectedHeap::heap();
8685 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8686 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8687 "Wrong gc statistics type");
8688 counters->update_counters(gc_stats_l);
8689 }
8690 }
8691
8692 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
8693 if (UsePerfData) {
8694 _space_counters->update_used(used);
8695 _space_counters->update_capacity();
8696 _gen_counters->update_all();
8697
8698 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8699 GenCollectedHeap* gch = GenCollectedHeap::heap();
8700 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8701 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8702 "Wrong gc statistics type");
8703 counters->update_counters(gc_stats_l);
8704 }
8705 }
8706
8707 // The desired expansion delta is computed so that:
8708 // . desired free percentage or greater is used
8709 void ASConcurrentMarkSweepGeneration::compute_new_size() {
8710 assert_locked_or_safepoint(Heap_lock);
8711
8712 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8713
8714 // If incremental collection failed, we just want to expand
8715 // to the limit.
8716 if (incremental_collection_failed()) {
8717 clear_incremental_collection_failed();
8718 grow_to_reserved();
8719 return;
8720 }
8721
8722 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
8723
8724 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
8725 "Wrong type of heap");
8726 int prev_level = level() - 1;
8727 assert(prev_level >= 0, "The cms generation is the lowest generation");
8728 Generation* prev_gen = gch->get_gen(prev_level);
8729 assert(prev_gen->kind() == Generation::ASParNew,
8730 "Wrong type of young generation");
8731 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
8732 size_t cur_eden = younger_gen->eden()->capacity();
8733 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
8734 size_t cur_promo = free();
8735 size_policy->compute_tenured_generation_free_space(cur_promo,
8736 max_available(),
8737 cur_eden);
8738 resize(cur_promo, size_policy->promo_size());
8739
8740 // Record the new size of the space in the cms generation
8741 // that is available for promotions. This is temporary.
8742 // It should be the desired promo size.
8743 size_policy->avg_cms_promo()->sample(free());
8744 size_policy->avg_old_live()->sample(used());
8745
8746 if (UsePerfData) {
8747 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8748 counters->update_cms_capacity_counter(capacity());
8749 }
8750 }
8751
8752 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
8753 assert_locked_or_safepoint(Heap_lock);
8754 assert_lock_strong(freelistLock());
8755 HeapWord* old_end = _cmsSpace->end();
8756 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
8757 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
8758 FreeChunk* chunk_at_end = find_chunk_at_end();
8759 if (chunk_at_end == NULL) {
8760 // No room to shrink
8761 if (PrintGCDetails && Verbose) {
8762 gclog_or_tty->print_cr("No room to shrink: old_end "
8763 PTR_FORMAT " unallocated_start " PTR_FORMAT
8764 " chunk_at_end " PTR_FORMAT,
8765 old_end, unallocated_start, chunk_at_end);
8766 }
8767 return;
8768 } else {
8769
8770 // Find the chunk at the end of the space and determine
8771 // how much it can be shrunk.
8772 size_t shrinkable_size_in_bytes = chunk_at_end->size();
8773 size_t aligned_shrinkable_size_in_bytes =
8774 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
8775 assert(unallocated_start <= chunk_at_end->end(),
8776 "Inconsistent chunk at end of space");
8777 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
8778 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
8779
8780 // Shrink the underlying space
8781 _virtual_space.shrink_by(bytes);
8782 if (PrintGCDetails && Verbose) {
8783 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
8784 " desired_bytes " SIZE_FORMAT
8785 " shrinkable_size_in_bytes " SIZE_FORMAT
8786 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
8787 " bytes " SIZE_FORMAT,
8788 desired_bytes, shrinkable_size_in_bytes,
8789 aligned_shrinkable_size_in_bytes, bytes);
8790 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
8791 " unallocated_start " SIZE_FORMAT,
8792 old_end, unallocated_start);
8793 }
8794
8795 // If the space did shrink (shrinking is not guaranteed),
8796 // shrink the chunk at the end by the appropriate amount.
8797 if (((HeapWord*)_virtual_space.high()) < old_end) {
8798 size_t new_word_size =
8799 heap_word_size(_virtual_space.committed_size());
8800
8801 // Have to remove the chunk from the dictionary because it is changing
8802 // size and might be someplace elsewhere in the dictionary.
8803
8804 // Get the chunk at end, shrink it, and put it
8805 // back.
8806 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
8807 size_t word_size_change = word_size_before - new_word_size;
8808 size_t chunk_at_end_old_size = chunk_at_end->size();
8809 assert(chunk_at_end_old_size >= word_size_change,
8810 "Shrink is too large");
8811 chunk_at_end->setSize(chunk_at_end_old_size -
8812 word_size_change);
8813 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
8814 word_size_change);
8815
8816 _cmsSpace->returnChunkToDictionary(chunk_at_end);
8817
8818 MemRegion mr(_cmsSpace->bottom(), new_word_size);
8819 _bts->resize(new_word_size); // resize the block offset shared array
8820 Universe::heap()->barrier_set()->resize_covered_region(mr);
8821 _cmsSpace->assert_locked();
8822 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
8823
8824 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
8825
8826 // update the space and generation capacity counters
8827 if (UsePerfData) {
8828 _space_counters->update_capacity();
8829 _gen_counters->update_all();
8830 }
8831
8832 if (Verbose && PrintGCDetails) {
8833 size_t new_mem_size = _virtual_space.committed_size();
8834 size_t old_mem_size = new_mem_size + bytes;
8835 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
8836 name(), old_mem_size/K, bytes/K, new_mem_size/K);
8837 }
8838 }
8839
8840 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
8841 "Inconsistency at end of space");
8842 assert(chunk_at_end->end() == _cmsSpace->end(),
8843 "Shrinking is inconsistent");
8844 return;
8845 }
8846 }
8847
8848 // Transfer some number of overflown objects to usual marking
8849 // stack. Return true if some objects were transferred.
8850 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8851 size_t num = MIN2((size_t)_mark_stack->capacity()/4,
8852 (size_t)ParGCDesiredObjsFromOverflowList);
8853
8854 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8855 assert(_collector->overflow_list_is_empty() || res,
8856 "If list is not empty, we should have taken something");
8857 assert(!res || !_mark_stack->isEmpty(),
8858 "If we took something, it should now be on our stack");
8859 return res;
8860 }
8861
8862 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8863 size_t res = _sp->block_size_no_stall(addr, _collector);
8864 assert(res != 0, "Should always be able to compute a size");
8865 if (_sp->block_is_obj(addr)) {
8866 if (_live_bit_map->isMarked(addr)) {
8867 // It can't have been dead in a previous cycle
8868 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8869 } else {
8870 _dead_bit_map->mark(addr); // mark the dead object
8871 }
8872 }
8873 return res;
8874 }