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(UseCompressedOops ||
194 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
195 "Offset of FreeChunk::_prev within FreeChunk must match"
196 " that of OopDesc::_klass within OopDesc");
197 )
198 if (ParallelGCThreads > 0) {
199 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
200 _par_gc_thread_states =
201 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
202 if (_par_gc_thread_states == NULL) {
203 vm_exit_during_initialization("Could not allocate par gc structs");
204 }
205 for (uint i = 0; i < ParallelGCThreads; i++) {
206 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
207 if (_par_gc_thread_states[i] == NULL) {
208 vm_exit_during_initialization("Could not allocate par gc structs");
209 }
210 }
211 } else {
212 _par_gc_thread_states = NULL;
213 }
214 _incremental_collection_failed = false;
215 // The "dilatation_factor" is the expansion that can occur on
216 // account of the fact that the minimum object size in the CMS
217 // generation may be larger than that in, say, a contiguous young
218 // generation.
219 // Ideally, in the calculation below, we'd compute the dilatation
220 // factor as: MinChunkSize/(promoting_gen's min object size)
221 // Since we do not have such a general query interface for the
222 // promoting generation, we'll instead just use the mimimum
223 // object size (which today is a header's worth of space);
224 // note that all arithmetic is in units of HeapWords.
225 assert(MinChunkSize >= oopDesc::header_size(), "just checking");
226 assert(_dilatation_factor >= 1.0, "from previous assert");
227 }
228
229
230 // The field "_initiating_occupancy" represents the occupancy percentage
231 // at which we trigger a new collection cycle. Unless explicitly specified
232 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
233 // is calculated by:
234 //
235 // Let "f" be MinHeapFreeRatio in
236 //
237 // _intiating_occupancy = 100-f +
238 // f * (CMSTrigger[Perm]Ratio/100)
239 // where CMSTrigger[Perm]Ratio is the argument "tr" below.
240 //
241 // That is, if we assume the heap is at its desired maximum occupancy at the
242 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
243 // space be allocated before initiating a new collection cycle.
244 //
245 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
246 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
247 if (io >= 0) {
248 _initiating_occupancy = (double)io / 100.0;
249 } else {
250 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
251 (double)(tr * MinHeapFreeRatio) / 100.0)
252 / 100.0;
253 }
254 }
255
256
257 void ConcurrentMarkSweepGeneration::ref_processor_init() {
258 assert(collector() != NULL, "no collector");
259 collector()->ref_processor_init();
260 }
261
262 void CMSCollector::ref_processor_init() {
263 if (_ref_processor == NULL) {
264 // Allocate and initialize a reference processor
265 _ref_processor = ReferenceProcessor::create_ref_processor(
266 _span, // span
267 _cmsGen->refs_discovery_is_atomic(), // atomic_discovery
268 _cmsGen->refs_discovery_is_mt(), // mt_discovery
269 &_is_alive_closure,
270 ParallelGCThreads,
271 ParallelRefProcEnabled);
272 // Initialize the _ref_processor field of CMSGen
273 _cmsGen->set_ref_processor(_ref_processor);
274
275 // Allocate a dummy ref processor for perm gen.
276 ReferenceProcessor* rp2 = new ReferenceProcessor();
277 if (rp2 == NULL) {
278 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
279 }
280 _permGen->set_ref_processor(rp2);
281 }
282 }
283
284 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
285 GenCollectedHeap* gch = GenCollectedHeap::heap();
286 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
287 "Wrong type of heap");
288 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
289 gch->gen_policy()->size_policy();
290 assert(sp->is_gc_cms_adaptive_size_policy(),
291 "Wrong type of size policy");
292 return sp;
293 }
294
295 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
296 CMSGCAdaptivePolicyCounters* results =
297 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
298 assert(
299 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
300 "Wrong gc policy counter kind");
301 return results;
302 }
303
304
305 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
306
307 const char* gen_name = "old";
308
309 // Generation Counters - generation 1, 1 subspace
310 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
311
312 _space_counters = new GSpaceCounters(gen_name, 0,
313 _virtual_space.reserved_size(),
314 this, _gen_counters);
315 }
316
317 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
318 _cms_gen(cms_gen)
319 {
320 assert(alpha <= 100, "bad value");
321 _saved_alpha = alpha;
322
323 // Initialize the alphas to the bootstrap value of 100.
324 _gc0_alpha = _cms_alpha = 100;
325
326 _cms_begin_time.update();
327 _cms_end_time.update();
328
329 _gc0_duration = 0.0;
330 _gc0_period = 0.0;
331 _gc0_promoted = 0;
332
333 _cms_duration = 0.0;
334 _cms_period = 0.0;
335 _cms_allocated = 0;
336
337 _cms_used_at_gc0_begin = 0;
338 _cms_used_at_gc0_end = 0;
339 _allow_duty_cycle_reduction = false;
340 _valid_bits = 0;
341 _icms_duty_cycle = CMSIncrementalDutyCycle;
342 }
343
344 // If promotion failure handling is on use
345 // the padded average size of the promotion for each
346 // young generation collection.
347 double CMSStats::time_until_cms_gen_full() const {
348 size_t cms_free = _cms_gen->cmsSpace()->free();
349 GenCollectedHeap* gch = GenCollectedHeap::heap();
350 size_t expected_promotion = gch->get_gen(0)->capacity();
351 if (HandlePromotionFailure) {
352 expected_promotion = MIN2(
353 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(),
354 expected_promotion);
355 }
356 if (cms_free > expected_promotion) {
357 // Start a cms collection if there isn't enough space to promote
358 // for the next minor collection. Use the padded average as
359 // a safety factor.
360 cms_free -= expected_promotion;
361
362 // Adjust by the safety factor.
363 double cms_free_dbl = (double)cms_free;
364 cms_free_dbl = cms_free_dbl * (100.0 - CMSIncrementalSafetyFactor) / 100.0;
365
366 if (PrintGCDetails && Verbose) {
367 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
368 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
369 cms_free, expected_promotion);
370 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
371 cms_free_dbl, cms_consumption_rate() + 1.0);
372 }
373 // Add 1 in case the consumption rate goes to zero.
374 return cms_free_dbl / (cms_consumption_rate() + 1.0);
375 }
376 return 0.0;
377 }
378
379 // Compare the duration of the cms collection to the
380 // time remaining before the cms generation is empty.
381 // Note that the time from the start of the cms collection
382 // to the start of the cms sweep (less than the total
383 // duration of the cms collection) can be used. This
384 // has been tried and some applications experienced
385 // promotion failures early in execution. This was
386 // possibly because the averages were not accurate
387 // enough at the beginning.
388 double CMSStats::time_until_cms_start() const {
389 // We add "gc0_period" to the "work" calculation
390 // below because this query is done (mostly) at the
391 // end of a scavenge, so we need to conservatively
392 // account for that much possible delay
393 // in the query so as to avoid concurrent mode failures
394 // due to starting the collection just a wee bit too
395 // late.
396 double work = cms_duration() + gc0_period();
397 double deadline = time_until_cms_gen_full();
398 if (work > deadline) {
399 if (Verbose && PrintGCDetails) {
400 gclog_or_tty->print(
401 " CMSCollector: collect because of anticipated promotion "
402 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
403 gc0_period(), time_until_cms_gen_full());
404 }
405 return 0.0;
406 }
407 return work - deadline;
408 }
409
410 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
411 // amount of change to prevent wild oscillation.
412 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
413 unsigned int new_duty_cycle) {
414 assert(old_duty_cycle <= 100, "bad input value");
415 assert(new_duty_cycle <= 100, "bad input value");
416
417 // Note: use subtraction with caution since it may underflow (values are
418 // unsigned). Addition is safe since we're in the range 0-100.
419 unsigned int damped_duty_cycle = new_duty_cycle;
420 if (new_duty_cycle < old_duty_cycle) {
421 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
422 if (new_duty_cycle + largest_delta < old_duty_cycle) {
423 damped_duty_cycle = old_duty_cycle - largest_delta;
424 }
425 } else if (new_duty_cycle > old_duty_cycle) {
426 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
427 if (new_duty_cycle > old_duty_cycle + largest_delta) {
428 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
429 }
430 }
431 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
432
433 if (CMSTraceIncrementalPacing) {
434 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
435 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
436 }
437 return damped_duty_cycle;
438 }
439
440 unsigned int CMSStats::icms_update_duty_cycle_impl() {
441 assert(CMSIncrementalPacing && valid(),
442 "should be handled in icms_update_duty_cycle()");
443
444 double cms_time_so_far = cms_timer().seconds();
445 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
446 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
447
448 // Avoid division by 0.
449 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
450 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
451
452 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
453 if (new_duty_cycle > _icms_duty_cycle) {
454 // Avoid very small duty cycles (1 or 2); 0 is allowed.
455 if (new_duty_cycle > 2) {
456 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
457 new_duty_cycle);
458 }
459 } else if (_allow_duty_cycle_reduction) {
460 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
461 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
462 // Respect the minimum duty cycle.
463 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
464 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
465 }
466
467 if (PrintGCDetails || CMSTraceIncrementalPacing) {
468 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
469 }
470
471 _allow_duty_cycle_reduction = false;
472 return _icms_duty_cycle;
473 }
474
475 #ifndef PRODUCT
476 void CMSStats::print_on(outputStream *st) const {
477 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
478 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
479 gc0_duration(), gc0_period(), gc0_promoted());
480 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
481 cms_duration(), cms_duration_per_mb(),
482 cms_period(), cms_allocated());
483 st->print(",cms_since_beg=%g,cms_since_end=%g",
484 cms_time_since_begin(), cms_time_since_end());
485 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
486 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
487 if (CMSIncrementalMode) {
488 st->print(",dc=%d", icms_duty_cycle());
489 }
490
491 if (valid()) {
492 st->print(",promo_rate=%g,cms_alloc_rate=%g",
493 promotion_rate(), cms_allocation_rate());
494 st->print(",cms_consumption_rate=%g,time_until_full=%g",
495 cms_consumption_rate(), time_until_cms_gen_full());
496 }
497 st->print(" ");
498 }
499 #endif // #ifndef PRODUCT
500
501 CMSCollector::CollectorState CMSCollector::_collectorState =
502 CMSCollector::Idling;
503 bool CMSCollector::_foregroundGCIsActive = false;
504 bool CMSCollector::_foregroundGCShouldWait = false;
505
506 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
507 ConcurrentMarkSweepGeneration* permGen,
508 CardTableRS* ct,
509 ConcurrentMarkSweepPolicy* cp):
510 _cmsGen(cmsGen),
511 _permGen(permGen),
512 _ct(ct),
513 _ref_processor(NULL), // will be set later
514 _conc_workers(NULL), // may be set later
515 _abort_preclean(false),
516 _start_sampling(false),
517 _between_prologue_and_epilogue(false),
518 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
519 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
520 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
521 -1 /* lock-free */, "No_lock" /* dummy */),
522 _modUnionClosure(&_modUnionTable),
523 _modUnionClosurePar(&_modUnionTable),
524 // Adjust my span to cover old (cms) gen and perm gen
525 _span(cmsGen->reserved()._union(permGen->reserved())),
526 // Construct the is_alive_closure with _span & markBitMap
527 _is_alive_closure(_span, &_markBitMap),
528 _restart_addr(NULL),
529 _overflow_list(NULL),
530 _preserved_oop_stack(NULL),
531 _preserved_mark_stack(NULL),
532 _stats(cmsGen),
533 _eden_chunk_array(NULL), // may be set in ctor body
534 _eden_chunk_capacity(0), // -- ditto --
535 _eden_chunk_index(0), // -- ditto --
536 _survivor_plab_array(NULL), // -- ditto --
537 _survivor_chunk_array(NULL), // -- ditto --
538 _survivor_chunk_capacity(0), // -- ditto --
539 _survivor_chunk_index(0), // -- ditto --
540 _ser_pmc_preclean_ovflw(0),
541 _ser_pmc_remark_ovflw(0),
542 _par_pmc_remark_ovflw(0),
543 _ser_kac_ovflw(0),
544 _par_kac_ovflw(0),
545 #ifndef PRODUCT
546 _num_par_pushes(0),
547 #endif
548 _collection_count_start(0),
549 _verifying(false),
550 _icms_start_limit(NULL),
551 _icms_stop_limit(NULL),
552 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
553 _completed_initialization(false),
554 _collector_policy(cp),
555 _should_unload_classes(false),
556 _concurrent_cycles_since_last_unload(0),
557 _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding)
558 {
559 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
560 ExplicitGCInvokesConcurrent = true;
561 }
562 // Now expand the span and allocate the collection support structures
563 // (MUT, marking bit map etc.) to cover both generations subject to
564 // collection.
565
566 // First check that _permGen is adjacent to _cmsGen and above it.
567 assert( _cmsGen->reserved().word_size() > 0
568 && _permGen->reserved().word_size() > 0,
569 "generations should not be of zero size");
570 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
571 "_cmsGen and _permGen should not overlap");
572 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
573 "_cmsGen->end() different from _permGen->start()");
574
575 // For use by dirty card to oop closures.
576 _cmsGen->cmsSpace()->set_collector(this);
577 _permGen->cmsSpace()->set_collector(this);
578
579 // Allocate MUT and marking bit map
580 {
581 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
582 if (!_markBitMap.allocate(_span)) {
583 warning("Failed to allocate CMS Bit Map");
584 return;
585 }
586 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
587 }
588 {
589 _modUnionTable.allocate(_span);
590 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
591 }
592
593 if (!_markStack.allocate(CMSMarkStackSize)) {
594 warning("Failed to allocate CMS Marking Stack");
595 return;
596 }
597 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
598 warning("Failed to allocate CMS Revisit Stack");
599 return;
600 }
601
602 // Support for multi-threaded concurrent phases
603 if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) {
604 if (FLAG_IS_DEFAULT(ParallelCMSThreads)) {
605 // just for now
606 FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4);
607 }
608 if (ParallelCMSThreads > 1) {
609 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
610 ParallelCMSThreads, true);
611 if (_conc_workers == NULL) {
612 warning("GC/CMS: _conc_workers allocation failure: "
613 "forcing -CMSConcurrentMTEnabled");
614 CMSConcurrentMTEnabled = false;
615 }
616 } else {
617 CMSConcurrentMTEnabled = false;
618 }
619 }
620 if (!CMSConcurrentMTEnabled) {
621 ParallelCMSThreads = 0;
622 } else {
623 // Turn off CMSCleanOnEnter optimization temporarily for
624 // the MT case where it's not fixed yet; see 6178663.
625 CMSCleanOnEnter = false;
626 }
627 assert((_conc_workers != NULL) == (ParallelCMSThreads > 1),
628 "Inconsistency");
629
630 // Parallel task queues; these are shared for the
631 // concurrent and stop-world phases of CMS, but
632 // are not shared with parallel scavenge (ParNew).
633 {
634 uint i;
635 uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads);
636
637 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
638 || ParallelRefProcEnabled)
639 && num_queues > 0) {
640 _task_queues = new OopTaskQueueSet(num_queues);
641 if (_task_queues == NULL) {
642 warning("task_queues allocation failure.");
643 return;
644 }
645 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues);
646 if (_hash_seed == NULL) {
647 warning("_hash_seed array allocation failure");
648 return;
649 }
650
651 // XXX use a global constant instead of 64!
652 typedef struct OopTaskQueuePadded {
653 OopTaskQueue work_queue;
654 char pad[64 - sizeof(OopTaskQueue)]; // prevent false sharing
655 } OopTaskQueuePadded;
656
657 for (i = 0; i < num_queues; i++) {
658 OopTaskQueuePadded *q_padded = new OopTaskQueuePadded();
659 if (q_padded == NULL) {
660 warning("work_queue allocation failure.");
661 return;
662 }
663 _task_queues->register_queue(i, &q_padded->work_queue);
664 }
665 for (i = 0; i < num_queues; i++) {
666 _task_queues->queue(i)->initialize();
667 _hash_seed[i] = 17; // copied from ParNew
668 }
669 }
670 }
671
672 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
673 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
674
675 // Clip CMSBootstrapOccupancy between 0 and 100.
676 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
677 /(double)100;
678
679 _full_gcs_since_conc_gc = 0;
680
681 // Now tell CMS generations the identity of their collector
682 ConcurrentMarkSweepGeneration::set_collector(this);
683
684 // Create & start a CMS thread for this CMS collector
685 _cmsThread = ConcurrentMarkSweepThread::start(this);
686 assert(cmsThread() != NULL, "CMS Thread should have been created");
687 assert(cmsThread()->collector() == this,
688 "CMS Thread should refer to this gen");
689 assert(CGC_lock != NULL, "Where's the CGC_lock?");
690
691 // Support for parallelizing young gen rescan
692 GenCollectedHeap* gch = GenCollectedHeap::heap();
693 _young_gen = gch->prev_gen(_cmsGen);
694 if (gch->supports_inline_contig_alloc()) {
695 _top_addr = gch->top_addr();
696 _end_addr = gch->end_addr();
697 assert(_young_gen != NULL, "no _young_gen");
698 _eden_chunk_index = 0;
699 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
700 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity);
701 if (_eden_chunk_array == NULL) {
702 _eden_chunk_capacity = 0;
703 warning("GC/CMS: _eden_chunk_array allocation failure");
704 }
705 }
706 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
707
708 // Support for parallelizing survivor space rescan
709 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
710 size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize);
711 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads);
712 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples);
713 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads);
714 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
715 || _cursor == NULL) {
716 warning("Failed to allocate survivor plab/chunk array");
717 if (_survivor_plab_array != NULL) {
718 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
719 _survivor_plab_array = NULL;
720 }
721 if (_survivor_chunk_array != NULL) {
722 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
723 _survivor_chunk_array = NULL;
724 }
725 if (_cursor != NULL) {
726 FREE_C_HEAP_ARRAY(size_t, _cursor);
727 _cursor = NULL;
728 }
729 } else {
730 _survivor_chunk_capacity = 2*max_plab_samples;
731 for (uint i = 0; i < ParallelGCThreads; i++) {
732 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples);
733 if (vec == NULL) {
734 warning("Failed to allocate survivor plab array");
735 for (int j = i; j > 0; j--) {
736 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array());
737 }
738 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array);
739 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array);
740 _survivor_plab_array = NULL;
741 _survivor_chunk_array = NULL;
742 _survivor_chunk_capacity = 0;
743 break;
744 } else {
745 ChunkArray* cur =
746 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
747 max_plab_samples);
748 assert(cur->end() == 0, "Should be 0");
749 assert(cur->array() == vec, "Should be vec");
750 assert(cur->capacity() == max_plab_samples, "Error");
751 }
752 }
753 }
754 }
755 assert( ( _survivor_plab_array != NULL
756 && _survivor_chunk_array != NULL)
757 || ( _survivor_chunk_capacity == 0
758 && _survivor_chunk_index == 0),
759 "Error");
760
761 // Choose what strong roots should be scanned depending on verification options
762 // and perm gen collection mode.
763 if (!CMSClassUnloadingEnabled) {
764 // If class unloading is disabled we want to include all classes into the root set.
765 add_root_scanning_option(SharedHeap::SO_AllClasses);
766 } else {
767 add_root_scanning_option(SharedHeap::SO_SystemClasses);
768 }
769
770 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
771 _gc_counters = new CollectorCounters("CMS", 1);
772 _completed_initialization = true;
773 _sweep_timer.start(); // start of time
774 }
775
776 const char* ConcurrentMarkSweepGeneration::name() const {
777 return "concurrent mark-sweep generation";
778 }
779 void ConcurrentMarkSweepGeneration::update_counters() {
780 if (UsePerfData) {
781 _space_counters->update_all();
782 _gen_counters->update_all();
783 }
784 }
785
786 // this is an optimized version of update_counters(). it takes the
787 // used value as a parameter rather than computing it.
788 //
789 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
790 if (UsePerfData) {
791 _space_counters->update_used(used);
792 _space_counters->update_capacity();
793 _gen_counters->update_all();
794 }
795 }
796
797 void ConcurrentMarkSweepGeneration::print() const {
798 Generation::print();
799 cmsSpace()->print();
800 }
801
802 #ifndef PRODUCT
803 void ConcurrentMarkSweepGeneration::print_statistics() {
804 cmsSpace()->printFLCensus(0);
805 }
806 #endif
807
808 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
809 GenCollectedHeap* gch = GenCollectedHeap::heap();
810 if (PrintGCDetails) {
811 if (Verbose) {
812 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
813 level(), short_name(), s, used(), capacity());
814 } else {
815 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
816 level(), short_name(), s, used() / K, capacity() / K);
817 }
818 }
819 if (Verbose) {
820 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
821 gch->used(), gch->capacity());
822 } else {
823 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
824 gch->used() / K, gch->capacity() / K);
825 }
826 }
827
828 size_t
829 ConcurrentMarkSweepGeneration::contiguous_available() const {
830 // dld proposes an improvement in precision here. If the committed
831 // part of the space ends in a free block we should add that to
832 // uncommitted size in the calculation below. Will make this
833 // change later, staying with the approximation below for the
834 // time being. -- ysr.
835 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
836 }
837
838 size_t
839 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
840 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
841 }
842
843 size_t ConcurrentMarkSweepGeneration::max_available() const {
844 return free() + _virtual_space.uncommitted_size();
845 }
846
847 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(
848 size_t max_promotion_in_bytes,
849 bool younger_handles_promotion_failure) const {
850
851 // This is the most conservative test. Full promotion is
852 // guaranteed if this is used. The multiplicative factor is to
853 // account for the worst case "dilatation".
854 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes;
855 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t
856 adjusted_max_promo_bytes = (double)max_uintx;
857 }
858 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes);
859
860 if (younger_handles_promotion_failure && !result) {
861 // Full promotion is not guaranteed because fragmentation
862 // of the cms generation can prevent the full promotion.
863 result = (max_available() >= (size_t)adjusted_max_promo_bytes);
864
865 if (!result) {
866 // With promotion failure handling the test for the ability
867 // to support the promotion does not have to be guaranteed.
868 // Use an average of the amount promoted.
869 result = max_available() >= (size_t)
870 gc_stats()->avg_promoted()->padded_average();
871 if (PrintGC && Verbose && result) {
872 gclog_or_tty->print_cr(
873 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
874 " max_available: " SIZE_FORMAT
875 " avg_promoted: " SIZE_FORMAT,
876 max_available(), (size_t)
877 gc_stats()->avg_promoted()->padded_average());
878 }
879 } else {
880 if (PrintGC && Verbose) {
881 gclog_or_tty->print_cr(
882 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
883 " max_available: " SIZE_FORMAT
884 " adj_max_promo_bytes: " SIZE_FORMAT,
885 max_available(), (size_t)adjusted_max_promo_bytes);
886 }
887 }
888 } else {
889 if (PrintGC && Verbose) {
890 gclog_or_tty->print_cr(
891 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe"
892 " contiguous_available: " SIZE_FORMAT
893 " adj_max_promo_bytes: " SIZE_FORMAT,
894 max_contiguous_available(), (size_t)adjusted_max_promo_bytes);
895 }
896 }
897 return result;
898 }
899
900 CompactibleSpace*
901 ConcurrentMarkSweepGeneration::first_compaction_space() const {
902 return _cmsSpace;
903 }
904
905 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
906 // Clear the promotion information. These pointers can be adjusted
907 // along with all the other pointers into the heap but
908 // compaction is expected to be a rare event with
909 // a heap using cms so don't do it without seeing the need.
910 if (ParallelGCThreads > 0) {
911 for (uint i = 0; i < ParallelGCThreads; i++) {
912 _par_gc_thread_states[i]->promo.reset();
913 }
914 }
915 }
916
917 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
918 blk->do_space(_cmsSpace);
919 }
920
921 void ConcurrentMarkSweepGeneration::compute_new_size() {
922 assert_locked_or_safepoint(Heap_lock);
923
924 // If incremental collection failed, we just want to expand
925 // to the limit.
926 if (incremental_collection_failed()) {
927 clear_incremental_collection_failed();
928 grow_to_reserved();
929 return;
930 }
931
932 size_t expand_bytes = 0;
933 double free_percentage = ((double) free()) / capacity();
934 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
935 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
936
937 // compute expansion delta needed for reaching desired free percentage
938 if (free_percentage < desired_free_percentage) {
939 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
940 assert(desired_capacity >= capacity(), "invalid expansion size");
941 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
942 }
943 if (expand_bytes > 0) {
944 if (PrintGCDetails && Verbose) {
945 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
946 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
947 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
948 gclog_or_tty->print_cr(" Desired free fraction %f",
949 desired_free_percentage);
950 gclog_or_tty->print_cr(" Maximum free fraction %f",
951 maximum_free_percentage);
952 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
953 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
954 desired_capacity/1000);
955 int prev_level = level() - 1;
956 if (prev_level >= 0) {
957 size_t prev_size = 0;
958 GenCollectedHeap* gch = GenCollectedHeap::heap();
959 Generation* prev_gen = gch->_gens[prev_level];
960 prev_size = prev_gen->capacity();
961 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
962 prev_size/1000);
963 }
964 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
965 unsafe_max_alloc_nogc()/1000);
966 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
967 contiguous_available()/1000);
968 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
969 expand_bytes);
970 }
971 // safe if expansion fails
972 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
973 if (PrintGCDetails && Verbose) {
974 gclog_or_tty->print_cr(" Expanded free fraction %f",
975 ((double) free()) / capacity());
976 }
977 }
978 }
979
980 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
981 return cmsSpace()->freelistLock();
982 }
983
984 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
985 bool tlab) {
986 CMSSynchronousYieldRequest yr;
987 MutexLockerEx x(freelistLock(),
988 Mutex::_no_safepoint_check_flag);
989 return have_lock_and_allocate(size, tlab);
990 }
991
992 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
993 bool tlab) {
994 assert_lock_strong(freelistLock());
995 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
996 HeapWord* res = cmsSpace()->allocate(adjustedSize);
997 // Allocate the object live (grey) if the background collector has
998 // started marking. This is necessary because the marker may
999 // have passed this address and consequently this object will
1000 // not otherwise be greyed and would be incorrectly swept up.
1001 // Note that if this object contains references, the writing
1002 // of those references will dirty the card containing this object
1003 // allowing the object to be blackened (and its references scanned)
1004 // either during a preclean phase or at the final checkpoint.
1005 if (res != NULL) {
1006 collector()->direct_allocated(res, adjustedSize);
1007 _direct_allocated_words += adjustedSize;
1008 // allocation counters
1009 NOT_PRODUCT(
1010 _numObjectsAllocated++;
1011 _numWordsAllocated += (int)adjustedSize;
1012 )
1013 }
1014 return res;
1015 }
1016
1017 // In the case of direct allocation by mutators in a generation that
1018 // is being concurrently collected, the object must be allocated
1019 // live (grey) if the background collector has started marking.
1020 // This is necessary because the marker may
1021 // have passed this address and consequently this object will
1022 // not otherwise be greyed and would be incorrectly swept up.
1023 // Note that if this object contains references, the writing
1024 // of those references will dirty the card containing this object
1025 // allowing the object to be blackened (and its references scanned)
1026 // either during a preclean phase or at the final checkpoint.
1027 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1028 assert(_markBitMap.covers(start, size), "Out of bounds");
1029 if (_collectorState >= Marking) {
1030 MutexLockerEx y(_markBitMap.lock(),
1031 Mutex::_no_safepoint_check_flag);
1032 // [see comments preceding SweepClosure::do_blk() below for details]
1033 // 1. need to mark the object as live so it isn't collected
1034 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1035 // 3. need to mark the end of the object so sweeper can skip over it
1036 // if it's uninitialized when the sweeper reaches it.
1037 _markBitMap.mark(start); // object is live
1038 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1039 _markBitMap.mark(start + size - 1);
1040 // mark end of object
1041 }
1042 // check that oop looks uninitialized
1043 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1044 }
1045
1046 void CMSCollector::promoted(bool par, HeapWord* start,
1047 bool is_obj_array, size_t obj_size) {
1048 assert(_markBitMap.covers(start), "Out of bounds");
1049 // See comment in direct_allocated() about when objects should
1050 // be allocated live.
1051 if (_collectorState >= Marking) {
1052 // we already hold the marking bit map lock, taken in
1053 // the prologue
1054 if (par) {
1055 _markBitMap.par_mark(start);
1056 } else {
1057 _markBitMap.mark(start);
1058 }
1059 // We don't need to mark the object as uninitialized (as
1060 // in direct_allocated above) because this is being done with the
1061 // world stopped and the object will be initialized by the
1062 // time the sweeper gets to look at it.
1063 assert(SafepointSynchronize::is_at_safepoint(),
1064 "expect promotion only at safepoints");
1065
1066 if (_collectorState < Sweeping) {
1067 // Mark the appropriate cards in the modUnionTable, so that
1068 // this object gets scanned before the sweep. If this is
1069 // not done, CMS generation references in the object might
1070 // not get marked.
1071 // For the case of arrays, which are otherwise precisely
1072 // marked, we need to dirty the entire array, not just its head.
1073 if (is_obj_array) {
1074 // The [par_]mark_range() method expects mr.end() below to
1075 // be aligned to the granularity of a bit's representation
1076 // in the heap. In the case of the MUT below, that's a
1077 // card size.
1078 MemRegion mr(start,
1079 (HeapWord*)round_to((intptr_t)(start + obj_size),
1080 CardTableModRefBS::card_size /* bytes */));
1081 if (par) {
1082 _modUnionTable.par_mark_range(mr);
1083 } else {
1084 _modUnionTable.mark_range(mr);
1085 }
1086 } else { // not an obj array; we can just mark the head
1087 if (par) {
1088 _modUnionTable.par_mark(start);
1089 } else {
1090 _modUnionTable.mark(start);
1091 }
1092 }
1093 }
1094 }
1095 }
1096
1097 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1098 {
1099 size_t delta = pointer_delta(addr, space->bottom());
1100 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1101 }
1102
1103 void CMSCollector::icms_update_allocation_limits()
1104 {
1105 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1106 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1107
1108 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1109 if (CMSTraceIncrementalPacing) {
1110 stats().print();
1111 }
1112
1113 assert(duty_cycle <= 100, "invalid duty cycle");
1114 if (duty_cycle != 0) {
1115 // The duty_cycle is a percentage between 0 and 100; convert to words and
1116 // then compute the offset from the endpoints of the space.
1117 size_t free_words = eden->free() / HeapWordSize;
1118 double free_words_dbl = (double)free_words;
1119 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1120 size_t offset_words = (free_words - duty_cycle_words) / 2;
1121
1122 _icms_start_limit = eden->top() + offset_words;
1123 _icms_stop_limit = eden->end() - offset_words;
1124
1125 // The limits may be adjusted (shifted to the right) by
1126 // CMSIncrementalOffset, to allow the application more mutator time after a
1127 // young gen gc (when all mutators were stopped) and before CMS starts and
1128 // takes away one or more cpus.
1129 if (CMSIncrementalOffset != 0) {
1130 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1131 size_t adjustment = (size_t)adjustment_dbl;
1132 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1133 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1134 _icms_start_limit += adjustment;
1135 _icms_stop_limit = tmp_stop;
1136 }
1137 }
1138 }
1139 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1140 _icms_start_limit = _icms_stop_limit = eden->end();
1141 }
1142
1143 // Install the new start limit.
1144 eden->set_soft_end(_icms_start_limit);
1145
1146 if (CMSTraceIncrementalMode) {
1147 gclog_or_tty->print(" icms alloc limits: "
1148 PTR_FORMAT "," PTR_FORMAT
1149 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1150 _icms_start_limit, _icms_stop_limit,
1151 percent_of_space(eden, _icms_start_limit),
1152 percent_of_space(eden, _icms_stop_limit));
1153 if (Verbose) {
1154 gclog_or_tty->print("eden: ");
1155 eden->print_on(gclog_or_tty);
1156 }
1157 }
1158 }
1159
1160 // Any changes here should try to maintain the invariant
1161 // that if this method is called with _icms_start_limit
1162 // and _icms_stop_limit both NULL, then it should return NULL
1163 // and not notify the icms thread.
1164 HeapWord*
1165 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1166 size_t word_size)
1167 {
1168 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1169 // nop.
1170 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1171 if (top <= _icms_start_limit) {
1172 if (CMSTraceIncrementalMode) {
1173 space->print_on(gclog_or_tty);
1174 gclog_or_tty->stamp();
1175 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1176 ", new limit=" PTR_FORMAT
1177 " (" SIZE_FORMAT "%%)",
1178 top, _icms_stop_limit,
1179 percent_of_space(space, _icms_stop_limit));
1180 }
1181 ConcurrentMarkSweepThread::start_icms();
1182 assert(top < _icms_stop_limit, "Tautology");
1183 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1184 return _icms_stop_limit;
1185 }
1186
1187 // The allocation will cross both the _start and _stop limits, so do the
1188 // stop notification also and return end().
1189 if (CMSTraceIncrementalMode) {
1190 space->print_on(gclog_or_tty);
1191 gclog_or_tty->stamp();
1192 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1193 ", new limit=" PTR_FORMAT
1194 " (" SIZE_FORMAT "%%)",
1195 top, space->end(),
1196 percent_of_space(space, space->end()));
1197 }
1198 ConcurrentMarkSweepThread::stop_icms();
1199 return space->end();
1200 }
1201
1202 if (top <= _icms_stop_limit) {
1203 if (CMSTraceIncrementalMode) {
1204 space->print_on(gclog_or_tty);
1205 gclog_or_tty->stamp();
1206 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1207 ", new limit=" PTR_FORMAT
1208 " (" SIZE_FORMAT "%%)",
1209 top, space->end(),
1210 percent_of_space(space, space->end()));
1211 }
1212 ConcurrentMarkSweepThread::stop_icms();
1213 return space->end();
1214 }
1215
1216 if (CMSTraceIncrementalMode) {
1217 space->print_on(gclog_or_tty);
1218 gclog_or_tty->stamp();
1219 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1220 ", new limit=" PTR_FORMAT,
1221 top, NULL);
1222 }
1223 }
1224
1225 return NULL;
1226 }
1227
1228 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1229 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1230 // allocate, copy and if necessary update promoinfo --
1231 // delegate to underlying space.
1232 assert_lock_strong(freelistLock());
1233
1234 #ifndef PRODUCT
1235 if (Universe::heap()->promotion_should_fail()) {
1236 return NULL;
1237 }
1238 #endif // #ifndef PRODUCT
1239
1240 oop res = _cmsSpace->promote(obj, obj_size);
1241 if (res == NULL) {
1242 // expand and retry
1243 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1244 expand(s*HeapWordSize, MinHeapDeltaBytes,
1245 CMSExpansionCause::_satisfy_promotion);
1246 // Since there's currently no next generation, we don't try to promote
1247 // into a more senior generation.
1248 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1249 "is made to pass on a possibly failing "
1250 "promotion to next generation");
1251 res = _cmsSpace->promote(obj, obj_size);
1252 }
1253 if (res != NULL) {
1254 // See comment in allocate() about when objects should
1255 // be allocated live.
1256 assert(obj->is_oop(), "Will dereference klass pointer below");
1257 collector()->promoted(false, // Not parallel
1258 (HeapWord*)res, obj->is_objArray(), obj_size);
1259 // promotion counters
1260 NOT_PRODUCT(
1261 _numObjectsPromoted++;
1262 _numWordsPromoted +=
1263 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1264 )
1265 }
1266 return res;
1267 }
1268
1269
1270 HeapWord*
1271 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1272 HeapWord* top,
1273 size_t word_sz)
1274 {
1275 return collector()->allocation_limit_reached(space, top, word_sz);
1276 }
1277
1278 // Things to support parallel young-gen collection.
1279 oop
1280 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1281 oop old, markOop m,
1282 size_t word_sz) {
1283 #ifndef PRODUCT
1284 if (Universe::heap()->promotion_should_fail()) {
1285 return NULL;
1286 }
1287 #endif // #ifndef PRODUCT
1288
1289 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1290 PromotionInfo* promoInfo = &ps->promo;
1291 // if we are tracking promotions, then first ensure space for
1292 // promotion (including spooling space for saving header if necessary).
1293 // then allocate and copy, then track promoted info if needed.
1294 // When tracking (see PromotionInfo::track()), the mark word may
1295 // be displaced and in this case restoration of the mark word
1296 // occurs in the (oop_since_save_marks_)iterate phase.
1297 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1298 // Out of space for allocating spooling buffers;
1299 // try expanding and allocating spooling buffers.
1300 if (!expand_and_ensure_spooling_space(promoInfo)) {
1301 return NULL;
1302 }
1303 }
1304 assert(promoInfo->has_spooling_space(), "Control point invariant");
1305 HeapWord* obj_ptr = ps->lab.alloc(word_sz);
1306 if (obj_ptr == NULL) {
1307 obj_ptr = expand_and_par_lab_allocate(ps, word_sz);
1308 if (obj_ptr == NULL) {
1309 return NULL;
1310 }
1311 }
1312 oop obj = oop(obj_ptr);
1313 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1314 // Otherwise, copy the object. Here we must be careful to insert the
1315 // klass pointer last, since this marks the block as an allocated object.
1316 // Except with compressed oops it's the mark word.
1317 HeapWord* old_ptr = (HeapWord*)old;
1318 if (word_sz > (size_t)oopDesc::header_size()) {
1319 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1320 obj_ptr + oopDesc::header_size(),
1321 word_sz - oopDesc::header_size());
1322 }
1323
1324 if (UseCompressedOops) {
1325 // Copy gap missed by (aligned) header size calculation above
1326 obj->set_klass_gap(old->klass_gap());
1327 }
1328
1329 // Restore the mark word copied above.
1330 obj->set_mark(m);
1331
1332 // Now we can track the promoted object, if necessary. We take care
1333 // To delay the transition from uninitialized to full object
1334 // (i.e., insertion of klass pointer) until after, so that it
1335 // atomically becomes a promoted object.
1336 if (promoInfo->tracking()) {
1337 promoInfo->track((PromotedObject*)obj, old->klass());
1338 }
1339
1340 // Finally, install the klass pointer (this should be volatile).
1341 obj->set_klass(old->klass());
1342
1343 assert(old->is_oop(), "Will dereference klass ptr below");
1344 collector()->promoted(true, // parallel
1345 obj_ptr, old->is_objArray(), word_sz);
1346
1347 NOT_PRODUCT(
1348 Atomic::inc(&_numObjectsPromoted);
1349 Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()),
1350 &_numWordsPromoted);
1351 )
1352
1353 return obj;
1354 }
1355
1356 void
1357 ConcurrentMarkSweepGeneration::
1358 par_promote_alloc_undo(int thread_num,
1359 HeapWord* obj, size_t word_sz) {
1360 // CMS does not support promotion undo.
1361 ShouldNotReachHere();
1362 }
1363
1364 void
1365 ConcurrentMarkSweepGeneration::
1366 par_promote_alloc_done(int thread_num) {
1367 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1368 ps->lab.retire();
1369 #if CFLS_LAB_REFILL_STATS
1370 if (thread_num == 0) {
1371 _cmsSpace->print_par_alloc_stats();
1372 }
1373 #endif
1374 }
1375
1376 void
1377 ConcurrentMarkSweepGeneration::
1378 par_oop_since_save_marks_iterate_done(int thread_num) {
1379 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1380 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1381 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1382 }
1383
1384 // XXXPERM
1385 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1386 size_t size,
1387 bool tlab)
1388 {
1389 // We allow a STW collection only if a full
1390 // collection was requested.
1391 return full || should_allocate(size, tlab); // FIX ME !!!
1392 // This and promotion failure handling are connected at the
1393 // hip and should be fixed by untying them.
1394 }
1395
1396 bool CMSCollector::shouldConcurrentCollect() {
1397 if (_full_gc_requested) {
1398 assert(ExplicitGCInvokesConcurrent, "Unexpected state");
1399 if (Verbose && PrintGCDetails) {
1400 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1401 " gc request");
1402 }
1403 return true;
1404 }
1405
1406 // For debugging purposes, change the type of collection.
1407 // If the rotation is not on the concurrent collection
1408 // type, don't start a concurrent collection.
1409 NOT_PRODUCT(
1410 if (RotateCMSCollectionTypes &&
1411 (_cmsGen->debug_collection_type() !=
1412 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1413 assert(_cmsGen->debug_collection_type() !=
1414 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1415 "Bad cms collection type");
1416 return false;
1417 }
1418 )
1419
1420 FreelistLocker x(this);
1421 // ------------------------------------------------------------------
1422 // Print out lots of information which affects the initiation of
1423 // a collection.
1424 if (PrintCMSInitiationStatistics && stats().valid()) {
1425 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1426 gclog_or_tty->stamp();
1427 gclog_or_tty->print_cr("");
1428 stats().print_on(gclog_or_tty);
1429 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1430 stats().time_until_cms_gen_full());
1431 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1432 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1433 _cmsGen->contiguous_available());
1434 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1435 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1436 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1437 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1438 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1439 }
1440 // ------------------------------------------------------------------
1441
1442 // If the estimated time to complete a cms collection (cms_duration())
1443 // is less than the estimated time remaining until the cms generation
1444 // is full, start a collection.
1445 if (!UseCMSInitiatingOccupancyOnly) {
1446 if (stats().valid()) {
1447 if (stats().time_until_cms_start() == 0.0) {
1448 return true;
1449 }
1450 } else {
1451 // We want to conservatively collect somewhat early in order
1452 // to try and "bootstrap" our CMS/promotion statistics;
1453 // this branch will not fire after the first successful CMS
1454 // collection because the stats should then be valid.
1455 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1456 if (Verbose && PrintGCDetails) {
1457 gclog_or_tty->print_cr(
1458 " CMSCollector: collect for bootstrapping statistics:"
1459 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1460 _bootstrap_occupancy);
1461 }
1462 return true;
1463 }
1464 }
1465 }
1466
1467 // Otherwise, we start a collection cycle if either the perm gen or
1468 // old gen want a collection cycle started. Each may use
1469 // an appropriate criterion for making this decision.
1470 // XXX We need to make sure that the gen expansion
1471 // criterion dovetails well with this. XXX NEED TO FIX THIS
1472 if (_cmsGen->should_concurrent_collect()) {
1473 if (Verbose && PrintGCDetails) {
1474 gclog_or_tty->print_cr("CMS old gen initiated");
1475 }
1476 return true;
1477 }
1478
1479 // We start a collection if we believe an incremental collection may fail;
1480 // this is not likely to be productive in practice because it's probably too
1481 // late anyway.
1482 GenCollectedHeap* gch = GenCollectedHeap::heap();
1483 assert(gch->collector_policy()->is_two_generation_policy(),
1484 "You may want to check the correctness of the following");
1485 if (gch->incremental_collection_will_fail()) {
1486 if (PrintGCDetails && Verbose) {
1487 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1488 }
1489 return true;
1490 }
1491
1492 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1493 bool res = update_should_unload_classes();
1494 if (res) {
1495 if (Verbose && PrintGCDetails) {
1496 gclog_or_tty->print_cr("CMS perm gen initiated");
1497 }
1498 return true;
1499 }
1500 }
1501 return false;
1502 }
1503
1504 // Clear _expansion_cause fields of constituent generations
1505 void CMSCollector::clear_expansion_cause() {
1506 _cmsGen->clear_expansion_cause();
1507 _permGen->clear_expansion_cause();
1508 }
1509
1510 // We should be conservative in starting a collection cycle. To
1511 // start too eagerly runs the risk of collecting too often in the
1512 // extreme. To collect too rarely falls back on full collections,
1513 // which works, even if not optimum in terms of concurrent work.
1514 // As a work around for too eagerly collecting, use the flag
1515 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1516 // giving the user an easily understandable way of controlling the
1517 // collections.
1518 // We want to start a new collection cycle if any of the following
1519 // conditions hold:
1520 // . our current occupancy exceeds the configured initiating occupancy
1521 // for this generation, or
1522 // . we recently needed to expand this space and have not, since that
1523 // expansion, done a collection of this generation, or
1524 // . the underlying space believes that it may be a good idea to initiate
1525 // a concurrent collection (this may be based on criteria such as the
1526 // following: the space uses linear allocation and linear allocation is
1527 // going to fail, or there is believed to be excessive fragmentation in
1528 // the generation, etc... or ...
1529 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1530 // the case of the old generation, not the perm generation; see CR 6543076):
1531 // we may be approaching a point at which allocation requests may fail because
1532 // we will be out of sufficient free space given allocation rate estimates.]
1533 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1534
1535 assert_lock_strong(freelistLock());
1536 if (occupancy() > initiating_occupancy()) {
1537 if (PrintGCDetails && Verbose) {
1538 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1539 short_name(), occupancy(), initiating_occupancy());
1540 }
1541 return true;
1542 }
1543 if (UseCMSInitiatingOccupancyOnly) {
1544 return false;
1545 }
1546 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1547 if (PrintGCDetails && Verbose) {
1548 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1549 short_name());
1550 }
1551 return true;
1552 }
1553 if (_cmsSpace->should_concurrent_collect()) {
1554 if (PrintGCDetails && Verbose) {
1555 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1556 short_name());
1557 }
1558 return true;
1559 }
1560 return false;
1561 }
1562
1563 void ConcurrentMarkSweepGeneration::collect(bool full,
1564 bool clear_all_soft_refs,
1565 size_t size,
1566 bool tlab)
1567 {
1568 collector()->collect(full, clear_all_soft_refs, size, tlab);
1569 }
1570
1571 void CMSCollector::collect(bool full,
1572 bool clear_all_soft_refs,
1573 size_t size,
1574 bool tlab)
1575 {
1576 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1577 // For debugging purposes skip the collection if the state
1578 // is not currently idle
1579 if (TraceCMSState) {
1580 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1581 Thread::current(), full, _collectorState);
1582 }
1583 return;
1584 }
1585
1586 // The following "if" branch is present for defensive reasons.
1587 // In the current uses of this interface, it can be replaced with:
1588 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1589 // But I am not placing that assert here to allow future
1590 // generality in invoking this interface.
1591 if (GC_locker::is_active()) {
1592 // A consistency test for GC_locker
1593 assert(GC_locker::needs_gc(), "Should have been set already");
1594 // Skip this foreground collection, instead
1595 // expanding the heap if necessary.
1596 // Need the free list locks for the call to free() in compute_new_size()
1597 compute_new_size();
1598 return;
1599 }
1600 acquire_control_and_collect(full, clear_all_soft_refs);
1601 _full_gcs_since_conc_gc++;
1602
1603 }
1604
1605 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1606 GenCollectedHeap* gch = GenCollectedHeap::heap();
1607 unsigned int gc_count = gch->total_full_collections();
1608 if (gc_count == full_gc_count) {
1609 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1610 _full_gc_requested = true;
1611 CGC_lock->notify(); // nudge CMS thread
1612 }
1613 }
1614
1615
1616 // The foreground and background collectors need to coordinate in order
1617 // to make sure that they do not mutually interfere with CMS collections.
1618 // When a background collection is active,
1619 // the foreground collector may need to take over (preempt) and
1620 // synchronously complete an ongoing collection. Depending on the
1621 // frequency of the background collections and the heap usage
1622 // of the application, this preemption can be seldom or frequent.
1623 // There are only certain
1624 // points in the background collection that the "collection-baton"
1625 // can be passed to the foreground collector.
1626 //
1627 // The foreground collector will wait for the baton before
1628 // starting any part of the collection. The foreground collector
1629 // will only wait at one location.
1630 //
1631 // The background collector will yield the baton before starting a new
1632 // phase of the collection (e.g., before initial marking, marking from roots,
1633 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1634 // of the loop which switches the phases. The background collector does some
1635 // of the phases (initial mark, final re-mark) with the world stopped.
1636 // Because of locking involved in stopping the world,
1637 // the foreground collector should not block waiting for the background
1638 // collector when it is doing a stop-the-world phase. The background
1639 // collector will yield the baton at an additional point just before
1640 // it enters a stop-the-world phase. Once the world is stopped, the
1641 // background collector checks the phase of the collection. If the
1642 // phase has not changed, it proceeds with the collection. If the
1643 // phase has changed, it skips that phase of the collection. See
1644 // the comments on the use of the Heap_lock in collect_in_background().
1645 //
1646 // Variable used in baton passing.
1647 // _foregroundGCIsActive - Set to true by the foreground collector when
1648 // it wants the baton. The foreground clears it when it has finished
1649 // the collection.
1650 // _foregroundGCShouldWait - Set to true by the background collector
1651 // when it is running. The foreground collector waits while
1652 // _foregroundGCShouldWait is true.
1653 // CGC_lock - monitor used to protect access to the above variables
1654 // and to notify the foreground and background collectors.
1655 // _collectorState - current state of the CMS collection.
1656 //
1657 // The foreground collector
1658 // acquires the CGC_lock
1659 // sets _foregroundGCIsActive
1660 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1661 // various locks acquired in preparation for the collection
1662 // are released so as not to block the background collector
1663 // that is in the midst of a collection
1664 // proceeds with the collection
1665 // clears _foregroundGCIsActive
1666 // returns
1667 //
1668 // The background collector in a loop iterating on the phases of the
1669 // collection
1670 // acquires the CGC_lock
1671 // sets _foregroundGCShouldWait
1672 // if _foregroundGCIsActive is set
1673 // clears _foregroundGCShouldWait, notifies _CGC_lock
1674 // waits on _CGC_lock for _foregroundGCIsActive to become false
1675 // and exits the loop.
1676 // otherwise
1677 // proceed with that phase of the collection
1678 // if the phase is a stop-the-world phase,
1679 // yield the baton once more just before enqueueing
1680 // the stop-world CMS operation (executed by the VM thread).
1681 // returns after all phases of the collection are done
1682 //
1683
1684 void CMSCollector::acquire_control_and_collect(bool full,
1685 bool clear_all_soft_refs) {
1686 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1687 assert(!Thread::current()->is_ConcurrentGC_thread(),
1688 "shouldn't try to acquire control from self!");
1689
1690 // Start the protocol for acquiring control of the
1691 // collection from the background collector (aka CMS thread).
1692 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1693 "VM thread should have CMS token");
1694 // Remember the possibly interrupted state of an ongoing
1695 // concurrent collection
1696 CollectorState first_state = _collectorState;
1697
1698 // Signal to a possibly ongoing concurrent collection that
1699 // we want to do a foreground collection.
1700 _foregroundGCIsActive = true;
1701
1702 // Disable incremental mode during a foreground collection.
1703 ICMSDisabler icms_disabler;
1704
1705 // release locks and wait for a notify from the background collector
1706 // releasing the locks in only necessary for phases which
1707 // do yields to improve the granularity of the collection.
1708 assert_lock_strong(bitMapLock());
1709 // We need to lock the Free list lock for the space that we are
1710 // currently collecting.
1711 assert(haveFreelistLocks(), "Must be holding free list locks");
1712 bitMapLock()->unlock();
1713 releaseFreelistLocks();
1714 {
1715 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1716 if (_foregroundGCShouldWait) {
1717 // We are going to be waiting for action for the CMS thread;
1718 // it had better not be gone (for instance at shutdown)!
1719 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1720 "CMS thread must be running");
1721 // Wait here until the background collector gives us the go-ahead
1722 ConcurrentMarkSweepThread::clear_CMS_flag(
1723 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1724 // Get a possibly blocked CMS thread going:
1725 // Note that we set _foregroundGCIsActive true above,
1726 // without protection of the CGC_lock.
1727 CGC_lock->notify();
1728 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1729 "Possible deadlock");
1730 while (_foregroundGCShouldWait) {
1731 // wait for notification
1732 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1733 // Possibility of delay/starvation here, since CMS token does
1734 // not know to give priority to VM thread? Actually, i think
1735 // there wouldn't be any delay/starvation, but the proof of
1736 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1737 }
1738 ConcurrentMarkSweepThread::set_CMS_flag(
1739 ConcurrentMarkSweepThread::CMS_vm_has_token);
1740 }
1741 }
1742 // The CMS_token is already held. Get back the other locks.
1743 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1744 "VM thread should have CMS token");
1745 getFreelistLocks();
1746 bitMapLock()->lock_without_safepoint_check();
1747 if (TraceCMSState) {
1748 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1749 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1750 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1751 }
1752
1753 // Check if we need to do a compaction, or if not, whether
1754 // we need to start the mark-sweep from scratch.
1755 bool should_compact = false;
1756 bool should_start_over = false;
1757 decide_foreground_collection_type(clear_all_soft_refs,
1758 &should_compact, &should_start_over);
1759
1760 NOT_PRODUCT(
1761 if (RotateCMSCollectionTypes) {
1762 if (_cmsGen->debug_collection_type() ==
1763 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1764 should_compact = true;
1765 } else if (_cmsGen->debug_collection_type() ==
1766 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1767 should_compact = false;
1768 }
1769 }
1770 )
1771
1772 if (PrintGCDetails && first_state > Idling) {
1773 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1774 if (GCCause::is_user_requested_gc(cause) ||
1775 GCCause::is_serviceability_requested_gc(cause)) {
1776 gclog_or_tty->print(" (concurrent mode interrupted)");
1777 } else {
1778 gclog_or_tty->print(" (concurrent mode failure)");
1779 }
1780 }
1781
1782 if (should_compact) {
1783 // If the collection is being acquired from the background
1784 // collector, there may be references on the discovered
1785 // references lists that have NULL referents (being those
1786 // that were concurrently cleared by a mutator) or
1787 // that are no longer active (having been enqueued concurrently
1788 // by the mutator).
1789 // Scrub the list of those references because Mark-Sweep-Compact
1790 // code assumes referents are not NULL and that all discovered
1791 // Reference objects are active.
1792 ref_processor()->clean_up_discovered_references();
1793
1794 do_compaction_work(clear_all_soft_refs);
1795
1796 // Has the GC time limit been exceeded?
1797 check_gc_time_limit();
1798
1799 } else {
1800 do_mark_sweep_work(clear_all_soft_refs, first_state,
1801 should_start_over);
1802 }
1803 // Reset the expansion cause, now that we just completed
1804 // a collection cycle.
1805 clear_expansion_cause();
1806 _foregroundGCIsActive = false;
1807 return;
1808 }
1809
1810 void CMSCollector::check_gc_time_limit() {
1811
1812 // Ignore explicit GC's. Exiting here does not set the flag and
1813 // does not reset the count. Updating of the averages for system
1814 // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC.
1815 GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause();
1816 if (GCCause::is_user_requested_gc(gc_cause) ||
1817 GCCause::is_serviceability_requested_gc(gc_cause)) {
1818 return;
1819 }
1820
1821 // Calculate the fraction of the CMS generation was freed during
1822 // the last collection.
1823 // Only consider the STW compacting cost for now.
1824 //
1825 // Note that the gc time limit test only works for the collections
1826 // of the young gen + tenured gen and not for collections of the
1827 // permanent gen. That is because the calculation of the space
1828 // freed by the collection is the free space in the young gen +
1829 // tenured gen.
1830
1831 double fraction_free =
1832 ((double)_cmsGen->free())/((double)_cmsGen->max_capacity());
1833 if ((100.0 * size_policy()->compacting_gc_cost()) >
1834 ((double) GCTimeLimit) &&
1835 ((fraction_free * 100) < GCHeapFreeLimit)) {
1836 size_policy()->inc_gc_time_limit_count();
1837 if (UseGCOverheadLimit &&
1838 (size_policy()->gc_time_limit_count() >
1839 AdaptiveSizePolicyGCTimeLimitThreshold)) {
1840 size_policy()->set_gc_time_limit_exceeded(true);
1841 // Avoid consecutive OOM due to the gc time limit by resetting
1842 // the counter.
1843 size_policy()->reset_gc_time_limit_count();
1844 if (PrintGCDetails) {
1845 gclog_or_tty->print_cr(" GC is exceeding overhead limit "
1846 "of %d%%", GCTimeLimit);
1847 }
1848 } else {
1849 if (PrintGCDetails) {
1850 gclog_or_tty->print_cr(" GC would exceed overhead limit "
1851 "of %d%%", GCTimeLimit);
1852 }
1853 }
1854 } else {
1855 size_policy()->reset_gc_time_limit_count();
1856 }
1857 }
1858
1859 // Resize the perm generation and the tenured generation
1860 // after obtaining the free list locks for the
1861 // two generations.
1862 void CMSCollector::compute_new_size() {
1863 assert_locked_or_safepoint(Heap_lock);
1864 FreelistLocker z(this);
1865 _permGen->compute_new_size();
1866 _cmsGen->compute_new_size();
1867 }
1868
1869 // A work method used by foreground collection to determine
1870 // what type of collection (compacting or not, continuing or fresh)
1871 // it should do.
1872 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1873 // and CMSCompactWhenClearAllSoftRefs the default in the future
1874 // and do away with the flags after a suitable period.
1875 void CMSCollector::decide_foreground_collection_type(
1876 bool clear_all_soft_refs, bool* should_compact,
1877 bool* should_start_over) {
1878 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1879 // flag is set, and we have either requested a System.gc() or
1880 // the number of full gc's since the last concurrent cycle
1881 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1882 // or if an incremental collection has failed
1883 GenCollectedHeap* gch = GenCollectedHeap::heap();
1884 assert(gch->collector_policy()->is_two_generation_policy(),
1885 "You may want to check the correctness of the following");
1886 // Inform cms gen if this was due to partial collection failing.
1887 // The CMS gen may use this fact to determine its expansion policy.
1888 if (gch->incremental_collection_will_fail()) {
1889 assert(!_cmsGen->incremental_collection_failed(),
1890 "Should have been noticed, reacted to and cleared");
1891 _cmsGen->set_incremental_collection_failed();
1892 }
1893 *should_compact =
1894 UseCMSCompactAtFullCollection &&
1895 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1896 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1897 gch->incremental_collection_will_fail());
1898 *should_start_over = false;
1899 if (clear_all_soft_refs && !*should_compact) {
1900 // We are about to do a last ditch collection attempt
1901 // so it would normally make sense to do a compaction
1902 // to reclaim as much space as possible.
1903 if (CMSCompactWhenClearAllSoftRefs) {
1904 // Default: The rationale is that in this case either
1905 // we are past the final marking phase, in which case
1906 // we'd have to start over, or so little has been done
1907 // that there's little point in saving that work. Compaction
1908 // appears to be the sensible choice in either case.
1909 *should_compact = true;
1910 } else {
1911 // We have been asked to clear all soft refs, but not to
1912 // compact. Make sure that we aren't past the final checkpoint
1913 // phase, for that is where we process soft refs. If we are already
1914 // past that phase, we'll need to redo the refs discovery phase and
1915 // if necessary clear soft refs that weren't previously
1916 // cleared. We do so by remembering the phase in which
1917 // we came in, and if we are past the refs processing
1918 // phase, we'll choose to just redo the mark-sweep
1919 // collection from scratch.
1920 if (_collectorState > FinalMarking) {
1921 // We are past the refs processing phase;
1922 // start over and do a fresh synchronous CMS cycle
1923 _collectorState = Resetting; // skip to reset to start new cycle
1924 reset(false /* == !asynch */);
1925 *should_start_over = true;
1926 } // else we can continue a possibly ongoing current cycle
1927 }
1928 }
1929 }
1930
1931 // A work method used by the foreground collector to do
1932 // a mark-sweep-compact.
1933 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1934 GenCollectedHeap* gch = GenCollectedHeap::heap();
1935 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty);
1936 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1937 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1938 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1939 }
1940
1941 // Sample collection interval time and reset for collection pause.
1942 if (UseAdaptiveSizePolicy) {
1943 size_policy()->msc_collection_begin();
1944 }
1945
1946 // Temporarily widen the span of the weak reference processing to
1947 // the entire heap.
1948 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
1949 ReferenceProcessorSpanMutator x(ref_processor(), new_span);
1950
1951 // Temporarily, clear the "is_alive_non_header" field of the
1952 // reference processor.
1953 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL);
1954
1955 // Temporarily make reference _processing_ single threaded (non-MT).
1956 ReferenceProcessorMTProcMutator z(ref_processor(), false);
1957
1958 // Temporarily make refs discovery atomic
1959 ReferenceProcessorAtomicMutator w(ref_processor(), true);
1960
1961 ref_processor()->set_enqueuing_is_done(false);
1962 ref_processor()->enable_discovery();
1963 // If an asynchronous collection finishes, the _modUnionTable is
1964 // all clear. If we are assuming the collection from an asynchronous
1965 // collection, clear the _modUnionTable.
1966 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1967 "_modUnionTable should be clear if the baton was not passed");
1968 _modUnionTable.clear_all();
1969
1970 // We must adjust the allocation statistics being maintained
1971 // in the free list space. We do so by reading and clearing
1972 // the sweep timer and updating the block flux rate estimates below.
1973 assert(_sweep_timer.is_active(), "We should never see the timer inactive");
1974 _sweep_timer.stop();
1975 // Note that we do not use this sample to update the _sweep_estimate.
1976 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
1977 _sweep_estimate.padded_average());
1978
1979 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
1980 ref_processor(), clear_all_soft_refs);
1981 #ifdef ASSERT
1982 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1983 size_t free_size = cms_space->free();
1984 assert(free_size ==
1985 pointer_delta(cms_space->end(), cms_space->compaction_top())
1986 * HeapWordSize,
1987 "All the free space should be compacted into one chunk at top");
1988 assert(cms_space->dictionary()->totalChunkSize(
1989 debug_only(cms_space->freelistLock())) == 0 ||
1990 cms_space->totalSizeInIndexedFreeLists() == 0,
1991 "All the free space should be in a single chunk");
1992 size_t num = cms_space->totalCount();
1993 assert((free_size == 0 && num == 0) ||
1994 (free_size > 0 && (num == 1 || num == 2)),
1995 "There should be at most 2 free chunks after compaction");
1996 #endif // ASSERT
1997 _collectorState = Resetting;
1998 assert(_restart_addr == NULL,
1999 "Should have been NULL'd before baton was passed");
2000 reset(false /* == !asynch */);
2001 _cmsGen->reset_after_compaction();
2002 _concurrent_cycles_since_last_unload = 0;
2003
2004 if (verifying() && !should_unload_classes()) {
2005 perm_gen_verify_bit_map()->clear_all();
2006 }
2007
2008 // Clear any data recorded in the PLAB chunk arrays.
2009 if (_survivor_plab_array != NULL) {
2010 reset_survivor_plab_arrays();
2011 }
2012
2013 // Adjust the per-size allocation stats for the next epoch.
2014 _cmsGen->cmsSpace()->endSweepFLCensus(sweepCount() /* fake */);
2015 // Restart the "sweep timer" for next epoch.
2016 _sweep_timer.reset();
2017 _sweep_timer.start();
2018
2019 // Sample collection pause time and reset for collection interval.
2020 if (UseAdaptiveSizePolicy) {
2021 size_policy()->msc_collection_end(gch->gc_cause());
2022 }
2023
2024 // For a mark-sweep-compact, compute_new_size() will be called
2025 // in the heap's do_collection() method.
2026 }
2027
2028 // A work method used by the foreground collector to do
2029 // a mark-sweep, after taking over from a possibly on-going
2030 // concurrent mark-sweep collection.
2031 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2032 CollectorState first_state, bool should_start_over) {
2033 if (PrintGC && Verbose) {
2034 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2035 "collector with count %d",
2036 _full_gcs_since_conc_gc);
2037 }
2038 switch (_collectorState) {
2039 case Idling:
2040 if (first_state == Idling || should_start_over) {
2041 // The background GC was not active, or should
2042 // restarted from scratch; start the cycle.
2043 _collectorState = InitialMarking;
2044 }
2045 // If first_state was not Idling, then a background GC
2046 // was in progress and has now finished. No need to do it
2047 // again. Leave the state as Idling.
2048 break;
2049 case Precleaning:
2050 // In the foreground case don't do the precleaning since
2051 // it is not done concurrently and there is extra work
2052 // required.
2053 _collectorState = FinalMarking;
2054 }
2055 if (PrintGCDetails &&
2056 (_collectorState > Idling ||
2057 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) {
2058 gclog_or_tty->print(" (concurrent mode failure)");
2059 }
2060 collect_in_foreground(clear_all_soft_refs);
2061
2062 // For a mark-sweep, compute_new_size() will be called
2063 // in the heap's do_collection() method.
2064 }
2065
2066
2067 void CMSCollector::getFreelistLocks() const {
2068 // Get locks for all free lists in all generations that this
2069 // collector is responsible for
2070 _cmsGen->freelistLock()->lock_without_safepoint_check();
2071 _permGen->freelistLock()->lock_without_safepoint_check();
2072 }
2073
2074 void CMSCollector::releaseFreelistLocks() const {
2075 // Release locks for all free lists in all generations that this
2076 // collector is responsible for
2077 _cmsGen->freelistLock()->unlock();
2078 _permGen->freelistLock()->unlock();
2079 }
2080
2081 bool CMSCollector::haveFreelistLocks() const {
2082 // Check locks for all free lists in all generations that this
2083 // collector is responsible for
2084 assert_lock_strong(_cmsGen->freelistLock());
2085 assert_lock_strong(_permGen->freelistLock());
2086 PRODUCT_ONLY(ShouldNotReachHere());
2087 return true;
2088 }
2089
2090 // A utility class that is used by the CMS collector to
2091 // temporarily "release" the foreground collector from its
2092 // usual obligation to wait for the background collector to
2093 // complete an ongoing phase before proceeding.
2094 class ReleaseForegroundGC: public StackObj {
2095 private:
2096 CMSCollector* _c;
2097 public:
2098 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2099 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2100 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2101 // allow a potentially blocked foreground collector to proceed
2102 _c->_foregroundGCShouldWait = false;
2103 if (_c->_foregroundGCIsActive) {
2104 CGC_lock->notify();
2105 }
2106 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2107 "Possible deadlock");
2108 }
2109
2110 ~ReleaseForegroundGC() {
2111 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2112 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2113 _c->_foregroundGCShouldWait = true;
2114 }
2115 };
2116
2117 // There are separate collect_in_background and collect_in_foreground because of
2118 // the different locking requirements of the background collector and the
2119 // foreground collector. There was originally an attempt to share
2120 // one "collect" method between the background collector and the foreground
2121 // collector but the if-then-else required made it cleaner to have
2122 // separate methods.
2123 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2124 assert(Thread::current()->is_ConcurrentGC_thread(),
2125 "A CMS asynchronous collection is only allowed on a CMS thread.");
2126
2127 GenCollectedHeap* gch = GenCollectedHeap::heap();
2128 {
2129 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2130 MutexLockerEx hl(Heap_lock, safepoint_check);
2131 FreelistLocker fll(this);
2132 MutexLockerEx x(CGC_lock, safepoint_check);
2133 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2134 // The foreground collector is active or we're
2135 // not using asynchronous collections. Skip this
2136 // background collection.
2137 assert(!_foregroundGCShouldWait, "Should be clear");
2138 return;
2139 } else {
2140 assert(_collectorState == Idling, "Should be idling before start.");
2141 _collectorState = InitialMarking;
2142 // Reset the expansion cause, now that we are about to begin
2143 // a new cycle.
2144 clear_expansion_cause();
2145 }
2146 // Decide if we want to enable class unloading as part of the
2147 // ensuing concurrent GC cycle.
2148 update_should_unload_classes();
2149 _full_gc_requested = false; // acks all outstanding full gc requests
2150 // Signal that we are about to start a collection
2151 gch->increment_total_full_collections(); // ... starting a collection cycle
2152 _collection_count_start = gch->total_full_collections();
2153 }
2154
2155 // Used for PrintGC
2156 size_t prev_used;
2157 if (PrintGC && Verbose) {
2158 prev_used = _cmsGen->used(); // XXXPERM
2159 }
2160
2161 // The change of the collection state is normally done at this level;
2162 // the exceptions are phases that are executed while the world is
2163 // stopped. For those phases the change of state is done while the
2164 // world is stopped. For baton passing purposes this allows the
2165 // background collector to finish the phase and change state atomically.
2166 // The foreground collector cannot wait on a phase that is done
2167 // while the world is stopped because the foreground collector already
2168 // has the world stopped and would deadlock.
2169 while (_collectorState != Idling) {
2170 if (TraceCMSState) {
2171 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2172 Thread::current(), _collectorState);
2173 }
2174 // The foreground collector
2175 // holds the Heap_lock throughout its collection.
2176 // holds the CMS token (but not the lock)
2177 // except while it is waiting for the background collector to yield.
2178 //
2179 // The foreground collector should be blocked (not for long)
2180 // if the background collector is about to start a phase
2181 // executed with world stopped. If the background
2182 // collector has already started such a phase, the
2183 // foreground collector is blocked waiting for the
2184 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2185 // are executed in the VM thread.
2186 //
2187 // The locking order is
2188 // PendingListLock (PLL) -- if applicable (FinalMarking)
2189 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2190 // CMS token (claimed in
2191 // stop_world_and_do() -->
2192 // safepoint_synchronize() -->
2193 // CMSThread::synchronize())
2194
2195 {
2196 // Check if the FG collector wants us to yield.
2197 CMSTokenSync x(true); // is cms thread
2198 if (waitForForegroundGC()) {
2199 // We yielded to a foreground GC, nothing more to be
2200 // done this round.
2201 assert(_foregroundGCShouldWait == false, "We set it to false in "
2202 "waitForForegroundGC()");
2203 if (TraceCMSState) {
2204 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2205 " exiting collection CMS state %d",
2206 Thread::current(), _collectorState);
2207 }
2208 return;
2209 } else {
2210 // The background collector can run but check to see if the
2211 // foreground collector has done a collection while the
2212 // background collector was waiting to get the CGC_lock
2213 // above. If yes, break so that _foregroundGCShouldWait
2214 // is cleared before returning.
2215 if (_collectorState == Idling) {
2216 break;
2217 }
2218 }
2219 }
2220
2221 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2222 "should be waiting");
2223
2224 switch (_collectorState) {
2225 case InitialMarking:
2226 {
2227 ReleaseForegroundGC x(this);
2228 stats().record_cms_begin();
2229
2230 VM_CMS_Initial_Mark initial_mark_op(this);
2231 VMThread::execute(&initial_mark_op);
2232 }
2233 // The collector state may be any legal state at this point
2234 // since the background collector may have yielded to the
2235 // foreground collector.
2236 break;
2237 case Marking:
2238 // initial marking in checkpointRootsInitialWork has been completed
2239 if (markFromRoots(true)) { // we were successful
2240 assert(_collectorState == Precleaning, "Collector state should "
2241 "have changed");
2242 } else {
2243 assert(_foregroundGCIsActive, "Internal state inconsistency");
2244 }
2245 break;
2246 case Precleaning:
2247 if (UseAdaptiveSizePolicy) {
2248 size_policy()->concurrent_precleaning_begin();
2249 }
2250 // marking from roots in markFromRoots has been completed
2251 preclean();
2252 if (UseAdaptiveSizePolicy) {
2253 size_policy()->concurrent_precleaning_end();
2254 }
2255 assert(_collectorState == AbortablePreclean ||
2256 _collectorState == FinalMarking,
2257 "Collector state should have changed");
2258 break;
2259 case AbortablePreclean:
2260 if (UseAdaptiveSizePolicy) {
2261 size_policy()->concurrent_phases_resume();
2262 }
2263 abortable_preclean();
2264 if (UseAdaptiveSizePolicy) {
2265 size_policy()->concurrent_precleaning_end();
2266 }
2267 assert(_collectorState == FinalMarking, "Collector state should "
2268 "have changed");
2269 break;
2270 case FinalMarking:
2271 {
2272 ReleaseForegroundGC x(this);
2273
2274 VM_CMS_Final_Remark final_remark_op(this);
2275 VMThread::execute(&final_remark_op);
2276 }
2277 assert(_foregroundGCShouldWait, "block post-condition");
2278 break;
2279 case Sweeping:
2280 if (UseAdaptiveSizePolicy) {
2281 size_policy()->concurrent_sweeping_begin();
2282 }
2283 // final marking in checkpointRootsFinal has been completed
2284 sweep(true);
2285 assert(_collectorState == Resizing, "Collector state change "
2286 "to Resizing must be done under the free_list_lock");
2287 _full_gcs_since_conc_gc = 0;
2288
2289 // Stop the timers for adaptive size policy for the concurrent phases
2290 if (UseAdaptiveSizePolicy) {
2291 size_policy()->concurrent_sweeping_end();
2292 size_policy()->concurrent_phases_end(gch->gc_cause(),
2293 gch->prev_gen(_cmsGen)->capacity(),
2294 _cmsGen->free());
2295 }
2296
2297 case Resizing: {
2298 // Sweeping has been completed...
2299 // At this point the background collection has completed.
2300 // Don't move the call to compute_new_size() down
2301 // into code that might be executed if the background
2302 // collection was preempted.
2303 {
2304 ReleaseForegroundGC x(this); // unblock FG collection
2305 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2306 CMSTokenSync z(true); // not strictly needed.
2307 if (_collectorState == Resizing) {
2308 compute_new_size();
2309 _collectorState = Resetting;
2310 } else {
2311 assert(_collectorState == Idling, "The state should only change"
2312 " because the foreground collector has finished the collection");
2313 }
2314 }
2315 break;
2316 }
2317 case Resetting:
2318 // CMS heap resizing has been completed
2319 reset(true);
2320 assert(_collectorState == Idling, "Collector state should "
2321 "have changed");
2322 stats().record_cms_end();
2323 // Don't move the concurrent_phases_end() and compute_new_size()
2324 // calls to here because a preempted background collection
2325 // has it's state set to "Resetting".
2326 break;
2327 case Idling:
2328 default:
2329 ShouldNotReachHere();
2330 break;
2331 }
2332 if (TraceCMSState) {
2333 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2334 Thread::current(), _collectorState);
2335 }
2336 assert(_foregroundGCShouldWait, "block post-condition");
2337 }
2338
2339 // Should this be in gc_epilogue?
2340 collector_policy()->counters()->update_counters();
2341
2342 {
2343 // Clear _foregroundGCShouldWait and, in the event that the
2344 // foreground collector is waiting, notify it, before
2345 // returning.
2346 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2347 _foregroundGCShouldWait = false;
2348 if (_foregroundGCIsActive) {
2349 CGC_lock->notify();
2350 }
2351 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2352 "Possible deadlock");
2353 }
2354 if (TraceCMSState) {
2355 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2356 " exiting collection CMS state %d",
2357 Thread::current(), _collectorState);
2358 }
2359 if (PrintGC && Verbose) {
2360 _cmsGen->print_heap_change(prev_used);
2361 }
2362 }
2363
2364 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2365 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2366 "Foreground collector should be waiting, not executing");
2367 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2368 "may only be done by the VM Thread with the world stopped");
2369 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2370 "VM thread should have CMS token");
2371
2372 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2373 true, gclog_or_tty);)
2374 if (UseAdaptiveSizePolicy) {
2375 size_policy()->ms_collection_begin();
2376 }
2377 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2378
2379 HandleMark hm; // Discard invalid handles created during verification
2380
2381 if (VerifyBeforeGC &&
2382 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2383 Universe::verify(true);
2384 }
2385
2386 bool init_mark_was_synchronous = false; // until proven otherwise
2387 while (_collectorState != Idling) {
2388 if (TraceCMSState) {
2389 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2390 Thread::current(), _collectorState);
2391 }
2392 switch (_collectorState) {
2393 case InitialMarking:
2394 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2395 checkpointRootsInitial(false);
2396 assert(_collectorState == Marking, "Collector state should have changed"
2397 " within checkpointRootsInitial()");
2398 break;
2399 case Marking:
2400 // initial marking in checkpointRootsInitialWork has been completed
2401 if (VerifyDuringGC &&
2402 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2403 gclog_or_tty->print("Verify before initial mark: ");
2404 Universe::verify(true);
2405 }
2406 {
2407 bool res = markFromRoots(false);
2408 assert(res && _collectorState == FinalMarking, "Collector state should "
2409 "have changed");
2410 break;
2411 }
2412 case FinalMarking:
2413 if (VerifyDuringGC &&
2414 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2415 gclog_or_tty->print("Verify before re-mark: ");
2416 Universe::verify(true);
2417 }
2418 checkpointRootsFinal(false, clear_all_soft_refs,
2419 init_mark_was_synchronous);
2420 assert(_collectorState == Sweeping, "Collector state should not "
2421 "have changed within checkpointRootsFinal()");
2422 break;
2423 case Sweeping:
2424 // final marking in checkpointRootsFinal has been completed
2425 if (VerifyDuringGC &&
2426 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2427 gclog_or_tty->print("Verify before sweep: ");
2428 Universe::verify(true);
2429 }
2430 sweep(false);
2431 assert(_collectorState == Resizing, "Incorrect state");
2432 break;
2433 case Resizing: {
2434 // Sweeping has been completed; the actual resize in this case
2435 // is done separately; nothing to be done in this state.
2436 _collectorState = Resetting;
2437 break;
2438 }
2439 case Resetting:
2440 // The heap has been resized.
2441 if (VerifyDuringGC &&
2442 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2443 gclog_or_tty->print("Verify before reset: ");
2444 Universe::verify(true);
2445 }
2446 reset(false);
2447 assert(_collectorState == Idling, "Collector state should "
2448 "have changed");
2449 break;
2450 case Precleaning:
2451 case AbortablePreclean:
2452 // Elide the preclean phase
2453 _collectorState = FinalMarking;
2454 break;
2455 default:
2456 ShouldNotReachHere();
2457 }
2458 if (TraceCMSState) {
2459 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2460 Thread::current(), _collectorState);
2461 }
2462 }
2463
2464 if (UseAdaptiveSizePolicy) {
2465 GenCollectedHeap* gch = GenCollectedHeap::heap();
2466 size_policy()->ms_collection_end(gch->gc_cause());
2467 }
2468
2469 if (VerifyAfterGC &&
2470 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2471 Universe::verify(true);
2472 }
2473 if (TraceCMSState) {
2474 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2475 " exiting collection CMS state %d",
2476 Thread::current(), _collectorState);
2477 }
2478 }
2479
2480 bool CMSCollector::waitForForegroundGC() {
2481 bool res = false;
2482 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2483 "CMS thread should have CMS token");
2484 // Block the foreground collector until the
2485 // background collectors decides whether to
2486 // yield.
2487 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2488 _foregroundGCShouldWait = true;
2489 if (_foregroundGCIsActive) {
2490 // The background collector yields to the
2491 // foreground collector and returns a value
2492 // indicating that it has yielded. The foreground
2493 // collector can proceed.
2494 res = true;
2495 _foregroundGCShouldWait = false;
2496 ConcurrentMarkSweepThread::clear_CMS_flag(
2497 ConcurrentMarkSweepThread::CMS_cms_has_token);
2498 ConcurrentMarkSweepThread::set_CMS_flag(
2499 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2500 // Get a possibly blocked foreground thread going
2501 CGC_lock->notify();
2502 if (TraceCMSState) {
2503 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2504 Thread::current(), _collectorState);
2505 }
2506 while (_foregroundGCIsActive) {
2507 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2508 }
2509 ConcurrentMarkSweepThread::set_CMS_flag(
2510 ConcurrentMarkSweepThread::CMS_cms_has_token);
2511 ConcurrentMarkSweepThread::clear_CMS_flag(
2512 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2513 }
2514 if (TraceCMSState) {
2515 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2516 Thread::current(), _collectorState);
2517 }
2518 return res;
2519 }
2520
2521 // Because of the need to lock the free lists and other structures in
2522 // the collector, common to all the generations that the collector is
2523 // collecting, we need the gc_prologues of individual CMS generations
2524 // delegate to their collector. It may have been simpler had the
2525 // current infrastructure allowed one to call a prologue on a
2526 // collector. In the absence of that we have the generation's
2527 // prologue delegate to the collector, which delegates back
2528 // some "local" work to a worker method in the individual generations
2529 // that it's responsible for collecting, while itself doing any
2530 // work common to all generations it's responsible for. A similar
2531 // comment applies to the gc_epilogue()'s.
2532 // The role of the varaible _between_prologue_and_epilogue is to
2533 // enforce the invocation protocol.
2534 void CMSCollector::gc_prologue(bool full) {
2535 // Call gc_prologue_work() for each CMSGen and PermGen that
2536 // we are responsible for.
2537
2538 // The following locking discipline assumes that we are only called
2539 // when the world is stopped.
2540 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2541
2542 // The CMSCollector prologue must call the gc_prologues for the
2543 // "generations" (including PermGen if any) that it's responsible
2544 // for.
2545
2546 assert( Thread::current()->is_VM_thread()
2547 || ( CMSScavengeBeforeRemark
2548 && Thread::current()->is_ConcurrentGC_thread()),
2549 "Incorrect thread type for prologue execution");
2550
2551 if (_between_prologue_and_epilogue) {
2552 // We have already been invoked; this is a gc_prologue delegation
2553 // from yet another CMS generation that we are responsible for, just
2554 // ignore it since all relevant work has already been done.
2555 return;
2556 }
2557
2558 // set a bit saying prologue has been called; cleared in epilogue
2559 _between_prologue_and_epilogue = true;
2560 // Claim locks for common data structures, then call gc_prologue_work()
2561 // for each CMSGen and PermGen that we are responsible for.
2562
2563 getFreelistLocks(); // gets free list locks on constituent spaces
2564 bitMapLock()->lock_without_safepoint_check();
2565
2566 // Should call gc_prologue_work() for all cms gens we are responsible for
2567 bool registerClosure = _collectorState >= Marking
2568 && _collectorState < Sweeping;
2569 ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar
2570 : &_modUnionClosure;
2571 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2572 _permGen->gc_prologue_work(full, registerClosure, muc);
2573
2574 if (!full) {
2575 stats().record_gc0_begin();
2576 }
2577 }
2578
2579 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2580 // Delegate to CMScollector which knows how to coordinate between
2581 // this and any other CMS generations that it is responsible for
2582 // collecting.
2583 collector()->gc_prologue(full);
2584 }
2585
2586 // This is a "private" interface for use by this generation's CMSCollector.
2587 // Not to be called directly by any other entity (for instance,
2588 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2589 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2590 bool registerClosure, ModUnionClosure* modUnionClosure) {
2591 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2592 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2593 "Should be NULL");
2594 if (registerClosure) {
2595 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2596 }
2597 cmsSpace()->gc_prologue();
2598 // Clear stat counters
2599 NOT_PRODUCT(
2600 assert(_numObjectsPromoted == 0, "check");
2601 assert(_numWordsPromoted == 0, "check");
2602 if (Verbose && PrintGC) {
2603 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2604 SIZE_FORMAT" bytes concurrently",
2605 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2606 }
2607 _numObjectsAllocated = 0;
2608 _numWordsAllocated = 0;
2609 )
2610 }
2611
2612 void CMSCollector::gc_epilogue(bool full) {
2613 // The following locking discipline assumes that we are only called
2614 // when the world is stopped.
2615 assert(SafepointSynchronize::is_at_safepoint(),
2616 "world is stopped assumption");
2617
2618 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2619 // if linear allocation blocks need to be appropriately marked to allow the
2620 // the blocks to be parsable. We also check here whether we need to nudge the
2621 // CMS collector thread to start a new cycle (if it's not already active).
2622 assert( Thread::current()->is_VM_thread()
2623 || ( CMSScavengeBeforeRemark
2624 && Thread::current()->is_ConcurrentGC_thread()),
2625 "Incorrect thread type for epilogue execution");
2626
2627 if (!_between_prologue_and_epilogue) {
2628 // We have already been invoked; this is a gc_epilogue delegation
2629 // from yet another CMS generation that we are responsible for, just
2630 // ignore it since all relevant work has already been done.
2631 return;
2632 }
2633 assert(haveFreelistLocks(), "must have freelist locks");
2634 assert_lock_strong(bitMapLock());
2635
2636 _cmsGen->gc_epilogue_work(full);
2637 _permGen->gc_epilogue_work(full);
2638
2639 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2640 // in case sampling was not already enabled, enable it
2641 _start_sampling = true;
2642 }
2643 // reset _eden_chunk_array so sampling starts afresh
2644 _eden_chunk_index = 0;
2645
2646 size_t cms_used = _cmsGen->cmsSpace()->used();
2647 size_t perm_used = _permGen->cmsSpace()->used();
2648
2649 // update performance counters - this uses a special version of
2650 // update_counters() that allows the utilization to be passed as a
2651 // parameter, avoiding multiple calls to used().
2652 //
2653 _cmsGen->update_counters(cms_used);
2654 _permGen->update_counters(perm_used);
2655
2656 if (CMSIncrementalMode) {
2657 icms_update_allocation_limits();
2658 }
2659
2660 bitMapLock()->unlock();
2661 releaseFreelistLocks();
2662
2663 _between_prologue_and_epilogue = false; // ready for next cycle
2664 }
2665
2666 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2667 collector()->gc_epilogue(full);
2668
2669 // Also reset promotion tracking in par gc thread states.
2670 if (ParallelGCThreads > 0) {
2671 for (uint i = 0; i < ParallelGCThreads; i++) {
2672 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2673 }
2674 }
2675 }
2676
2677 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2678 assert(!incremental_collection_failed(), "Should have been cleared");
2679 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2680 cmsSpace()->gc_epilogue();
2681 // Print stat counters
2682 NOT_PRODUCT(
2683 assert(_numObjectsAllocated == 0, "check");
2684 assert(_numWordsAllocated == 0, "check");
2685 if (Verbose && PrintGC) {
2686 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2687 SIZE_FORMAT" bytes",
2688 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2689 }
2690 _numObjectsPromoted = 0;
2691 _numWordsPromoted = 0;
2692 )
2693
2694 if (PrintGC && Verbose) {
2695 // Call down the chain in contiguous_available needs the freelistLock
2696 // so print this out before releasing the freeListLock.
2697 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2698 contiguous_available());
2699 }
2700 }
2701
2702 #ifndef PRODUCT
2703 bool CMSCollector::have_cms_token() {
2704 Thread* thr = Thread::current();
2705 if (thr->is_VM_thread()) {
2706 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2707 } else if (thr->is_ConcurrentGC_thread()) {
2708 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2709 } else if (thr->is_GC_task_thread()) {
2710 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2711 ParGCRareEvent_lock->owned_by_self();
2712 }
2713 return false;
2714 }
2715 #endif
2716
2717 // Check reachability of the given heap address in CMS generation,
2718 // treating all other generations as roots.
2719 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2720 // We could "guarantee" below, rather than assert, but i'll
2721 // leave these as "asserts" so that an adventurous debugger
2722 // could try this in the product build provided some subset of
2723 // the conditions were met, provided they were intersted in the
2724 // results and knew that the computation below wouldn't interfere
2725 // with other concurrent computations mutating the structures
2726 // being read or written.
2727 assert(SafepointSynchronize::is_at_safepoint(),
2728 "Else mutations in object graph will make answer suspect");
2729 assert(have_cms_token(), "Should hold cms token");
2730 assert(haveFreelistLocks(), "must hold free list locks");
2731 assert_lock_strong(bitMapLock());
2732
2733 // Clear the marking bit map array before starting, but, just
2734 // for kicks, first report if the given address is already marked
2735 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2736 _markBitMap.isMarked(addr) ? "" : " not");
2737
2738 if (verify_after_remark()) {
2739 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2740 bool result = verification_mark_bm()->isMarked(addr);
2741 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2742 result ? "IS" : "is NOT");
2743 return result;
2744 } else {
2745 gclog_or_tty->print_cr("Could not compute result");
2746 return false;
2747 }
2748 }
2749
2750 ////////////////////////////////////////////////////////
2751 // CMS Verification Support
2752 ////////////////////////////////////////////////////////
2753 // Following the remark phase, the following invariant
2754 // should hold -- each object in the CMS heap which is
2755 // marked in markBitMap() should be marked in the verification_mark_bm().
2756
2757 class VerifyMarkedClosure: public BitMapClosure {
2758 CMSBitMap* _marks;
2759 bool _failed;
2760
2761 public:
2762 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2763
2764 void do_bit(size_t offset) {
2765 HeapWord* addr = _marks->offsetToHeapWord(offset);
2766 if (!_marks->isMarked(addr)) {
2767 oop(addr)->print();
2768 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2769 _failed = true;
2770 }
2771 }
2772
2773 bool failed() { return _failed; }
2774 };
2775
2776 bool CMSCollector::verify_after_remark() {
2777 gclog_or_tty->print(" [Verifying CMS Marking... ");
2778 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2779 static bool init = false;
2780
2781 assert(SafepointSynchronize::is_at_safepoint(),
2782 "Else mutations in object graph will make answer suspect");
2783 assert(have_cms_token(),
2784 "Else there may be mutual interference in use of "
2785 " verification data structures");
2786 assert(_collectorState > Marking && _collectorState <= Sweeping,
2787 "Else marking info checked here may be obsolete");
2788 assert(haveFreelistLocks(), "must hold free list locks");
2789 assert_lock_strong(bitMapLock());
2790
2791
2792 // Allocate marking bit map if not already allocated
2793 if (!init) { // first time
2794 if (!verification_mark_bm()->allocate(_span)) {
2795 return false;
2796 }
2797 init = true;
2798 }
2799
2800 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2801
2802 // Turn off refs discovery -- so we will be tracing through refs.
2803 // This is as intended, because by this time
2804 // GC must already have cleared any refs that need to be cleared,
2805 // and traced those that need to be marked; moreover,
2806 // the marking done here is not going to intefere in any
2807 // way with the marking information used by GC.
2808 NoRefDiscovery no_discovery(ref_processor());
2809
2810 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2811
2812 // Clear any marks from a previous round
2813 verification_mark_bm()->clear_all();
2814 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2815 assert(overflow_list_is_empty(), "overflow list should be empty");
2816
2817 GenCollectedHeap* gch = GenCollectedHeap::heap();
2818 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2819 // Update the saved marks which may affect the root scans.
2820 gch->save_marks();
2821
2822 if (CMSRemarkVerifyVariant == 1) {
2823 // In this first variant of verification, we complete
2824 // all marking, then check if the new marks-verctor is
2825 // a subset of the CMS marks-vector.
2826 verify_after_remark_work_1();
2827 } else if (CMSRemarkVerifyVariant == 2) {
2828 // In this second variant of verification, we flag an error
2829 // (i.e. an object reachable in the new marks-vector not reachable
2830 // in the CMS marks-vector) immediately, also indicating the
2831 // identify of an object (A) that references the unmarked object (B) --
2832 // presumably, a mutation to A failed to be picked up by preclean/remark?
2833 verify_after_remark_work_2();
2834 } else {
2835 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2836 CMSRemarkVerifyVariant);
2837 }
2838 gclog_or_tty->print(" done] ");
2839 return true;
2840 }
2841
2842 void CMSCollector::verify_after_remark_work_1() {
2843 ResourceMark rm;
2844 HandleMark hm;
2845 GenCollectedHeap* gch = GenCollectedHeap::heap();
2846
2847 // Mark from roots one level into CMS
2848 MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */);
2849 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2850
2851 gch->gen_process_strong_roots(_cmsGen->level(),
2852 true, // younger gens are roots
2853 true, // collecting perm gen
2854 SharedHeap::ScanningOption(roots_scanning_options()),
2855 NULL, ¬Older);
2856
2857 // Now mark from the roots
2858 assert(_revisitStack.isEmpty(), "Should be empty");
2859 MarkFromRootsClosure markFromRootsClosure(this, _span,
2860 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2861 false /* don't yield */, true /* verifying */);
2862 assert(_restart_addr == NULL, "Expected pre-condition");
2863 verification_mark_bm()->iterate(&markFromRootsClosure);
2864 while (_restart_addr != NULL) {
2865 // Deal with stack overflow: by restarting at the indicated
2866 // address.
2867 HeapWord* ra = _restart_addr;
2868 markFromRootsClosure.reset(ra);
2869 _restart_addr = NULL;
2870 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2871 }
2872 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2873 verify_work_stacks_empty();
2874 // Should reset the revisit stack above, since no class tree
2875 // surgery is forthcoming.
2876 _revisitStack.reset(); // throwing away all contents
2877
2878 // Marking completed -- now verify that each bit marked in
2879 // verification_mark_bm() is also marked in markBitMap(); flag all
2880 // errors by printing corresponding objects.
2881 VerifyMarkedClosure vcl(markBitMap());
2882 verification_mark_bm()->iterate(&vcl);
2883 if (vcl.failed()) {
2884 gclog_or_tty->print("Verification failed");
2885 Universe::heap()->print();
2886 fatal(" ... aborting");
2887 }
2888 }
2889
2890 void CMSCollector::verify_after_remark_work_2() {
2891 ResourceMark rm;
2892 HandleMark hm;
2893 GenCollectedHeap* gch = GenCollectedHeap::heap();
2894
2895 // Mark from roots one level into CMS
2896 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2897 markBitMap(), true /* nmethods */);
2898 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2899 gch->gen_process_strong_roots(_cmsGen->level(),
2900 true, // younger gens are roots
2901 true, // collecting perm gen
2902 SharedHeap::ScanningOption(roots_scanning_options()),
2903 NULL, ¬Older);
2904
2905 // Now mark from the roots
2906 assert(_revisitStack.isEmpty(), "Should be empty");
2907 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2908 verification_mark_bm(), markBitMap(), verification_mark_stack());
2909 assert(_restart_addr == NULL, "Expected pre-condition");
2910 verification_mark_bm()->iterate(&markFromRootsClosure);
2911 while (_restart_addr != NULL) {
2912 // Deal with stack overflow: by restarting at the indicated
2913 // address.
2914 HeapWord* ra = _restart_addr;
2915 markFromRootsClosure.reset(ra);
2916 _restart_addr = NULL;
2917 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2918 }
2919 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2920 verify_work_stacks_empty();
2921 // Should reset the revisit stack above, since no class tree
2922 // surgery is forthcoming.
2923 _revisitStack.reset(); // throwing away all contents
2924
2925 // Marking completed -- now verify that each bit marked in
2926 // verification_mark_bm() is also marked in markBitMap(); flag all
2927 // errors by printing corresponding objects.
2928 VerifyMarkedClosure vcl(markBitMap());
2929 verification_mark_bm()->iterate(&vcl);
2930 assert(!vcl.failed(), "Else verification above should not have succeeded");
2931 }
2932
2933 void ConcurrentMarkSweepGeneration::save_marks() {
2934 // delegate to CMS space
2935 cmsSpace()->save_marks();
2936 for (uint i = 0; i < ParallelGCThreads; i++) {
2937 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2938 }
2939 }
2940
2941 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2942 return cmsSpace()->no_allocs_since_save_marks();
2943 }
2944
2945 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2946 \
2947 void ConcurrentMarkSweepGeneration:: \
2948 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2949 cl->set_generation(this); \
2950 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2951 cl->reset_generation(); \
2952 save_marks(); \
2953 }
2954
2955 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2956
2957 void
2958 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
2959 {
2960 // Not currently implemented; need to do the following. -- ysr.
2961 // dld -- I think that is used for some sort of allocation profiler. So it
2962 // really means the objects allocated by the mutator since the last
2963 // GC. We could potentially implement this cheaply by recording only
2964 // the direct allocations in a side data structure.
2965 //
2966 // I think we probably ought not to be required to support these
2967 // iterations at any arbitrary point; I think there ought to be some
2968 // call to enable/disable allocation profiling in a generation/space,
2969 // and the iterator ought to return the objects allocated in the
2970 // gen/space since the enable call, or the last iterator call (which
2971 // will probably be at a GC.) That way, for gens like CM&S that would
2972 // require some extra data structure to support this, we only pay the
2973 // cost when it's in use...
2974 cmsSpace()->object_iterate_since_last_GC(blk);
2975 }
2976
2977 void
2978 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
2979 cl->set_generation(this);
2980 younger_refs_in_space_iterate(_cmsSpace, cl);
2981 cl->reset_generation();
2982 }
2983
2984 void
2985 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
2986 if (freelistLock()->owned_by_self()) {
2987 Generation::oop_iterate(mr, cl);
2988 } else {
2989 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2990 Generation::oop_iterate(mr, cl);
2991 }
2992 }
2993
2994 void
2995 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
2996 if (freelistLock()->owned_by_self()) {
2997 Generation::oop_iterate(cl);
2998 } else {
2999 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3000 Generation::oop_iterate(cl);
3001 }
3002 }
3003
3004 void
3005 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3006 if (freelistLock()->owned_by_self()) {
3007 Generation::object_iterate(cl);
3008 } else {
3009 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3010 Generation::object_iterate(cl);
3011 }
3012 }
3013
3014 void
3015 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3016 }
3017
3018 void
3019 ConcurrentMarkSweepGeneration::post_compact() {
3020 }
3021
3022 void
3023 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3024 // Fix the linear allocation blocks to look like free blocks.
3025
3026 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3027 // are not called when the heap is verified during universe initialization and
3028 // at vm shutdown.
3029 if (freelistLock()->owned_by_self()) {
3030 cmsSpace()->prepare_for_verify();
3031 } else {
3032 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3033 cmsSpace()->prepare_for_verify();
3034 }
3035 }
3036
3037 void
3038 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) {
3039 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3040 // are not called when the heap is verified during universe initialization and
3041 // at vm shutdown.
3042 if (freelistLock()->owned_by_self()) {
3043 cmsSpace()->verify(false /* ignored */);
3044 } else {
3045 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3046 cmsSpace()->verify(false /* ignored */);
3047 }
3048 }
3049
3050 void CMSCollector::verify(bool allow_dirty /* ignored */) {
3051 _cmsGen->verify(allow_dirty);
3052 _permGen->verify(allow_dirty);
3053 }
3054
3055 #ifndef PRODUCT
3056 bool CMSCollector::overflow_list_is_empty() const {
3057 assert(_num_par_pushes >= 0, "Inconsistency");
3058 if (_overflow_list == NULL) {
3059 assert(_num_par_pushes == 0, "Inconsistency");
3060 }
3061 return _overflow_list == NULL;
3062 }
3063
3064 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3065 // merely consolidate assertion checks that appear to occur together frequently.
3066 void CMSCollector::verify_work_stacks_empty() const {
3067 assert(_markStack.isEmpty(), "Marking stack should be empty");
3068 assert(overflow_list_is_empty(), "Overflow list should be empty");
3069 }
3070
3071 void CMSCollector::verify_overflow_empty() const {
3072 assert(overflow_list_is_empty(), "Overflow list should be empty");
3073 assert(no_preserved_marks(), "No preserved marks");
3074 }
3075 #endif // PRODUCT
3076
3077 // Decide if we want to enable class unloading as part of the
3078 // ensuing concurrent GC cycle. We will collect the perm gen and
3079 // unload classes if it's the case that:
3080 // (1) an explicit gc request has been made and the flag
3081 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3082 // (2) (a) class unloading is enabled at the command line, and
3083 // (b) (i) perm gen threshold has been crossed, or
3084 // (ii) old gen is getting really full, or
3085 // (iii) the previous N CMS collections did not collect the
3086 // perm gen
3087 // NOTE: Provided there is no change in the state of the heap between
3088 // calls to this method, it should have idempotent results. Moreover,
3089 // its results should be monotonically increasing (i.e. going from 0 to 1,
3090 // but not 1 to 0) between successive calls between which the heap was
3091 // not collected. For the implementation below, it must thus rely on
3092 // the property that concurrent_cycles_since_last_unload()
3093 // will not decrease unless a collection cycle happened and that
3094 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3095 // themselves also monotonic in that sense. See check_monotonicity()
3096 // below.
3097 bool CMSCollector::update_should_unload_classes() {
3098 _should_unload_classes = false;
3099 // Condition 1 above
3100 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3101 _should_unload_classes = true;
3102 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3103 // Disjuncts 2.b.(i,ii,iii) above
3104 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3105 CMSClassUnloadingMaxInterval)
3106 || _permGen->should_concurrent_collect()
3107 || _cmsGen->is_too_full();
3108 }
3109 return _should_unload_classes;
3110 }
3111
3112 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3113 bool res = should_concurrent_collect();
3114 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3115 return res;
3116 }
3117
3118 void CMSCollector::setup_cms_unloading_and_verification_state() {
3119 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3120 || VerifyBeforeExit;
3121 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings
3122 | SharedHeap::SO_CodeCache;
3123
3124 if (should_unload_classes()) { // Should unload classes this cycle
3125 remove_root_scanning_option(rso); // Shrink the root set appropriately
3126 set_verifying(should_verify); // Set verification state for this cycle
3127 return; // Nothing else needs to be done at this time
3128 }
3129
3130 // Not unloading classes this cycle
3131 assert(!should_unload_classes(), "Inconsitency!");
3132 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3133 // We were not verifying, or we _were_ unloading classes in the last cycle,
3134 // AND some verification options are enabled this cycle; in this case,
3135 // we must make sure that the deadness map is allocated if not already so,
3136 // and cleared (if already allocated previously --
3137 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3138 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3139 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3140 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3141 "permanent generation verification disabled");
3142 return; // Note that we leave verification disabled, so we'll retry this
3143 // allocation next cycle. We _could_ remember this failure
3144 // and skip further attempts and permanently disable verification
3145 // attempts if that is considered more desirable.
3146 }
3147 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3148 "_perm_gen_ver_bit_map inconsistency?");
3149 } else {
3150 perm_gen_verify_bit_map()->clear_all();
3151 }
3152 // Include symbols, strings and code cache elements to prevent their resurrection.
3153 add_root_scanning_option(rso);
3154 set_verifying(true);
3155 } else if (verifying() && !should_verify) {
3156 // We were verifying, but some verification flags got disabled.
3157 set_verifying(false);
3158 // Exclude symbols, strings and code cache elements from root scanning to
3159 // reduce IM and RM pauses.
3160 remove_root_scanning_option(rso);
3161 }
3162 }
3163
3164
3165 #ifndef PRODUCT
3166 HeapWord* CMSCollector::block_start(const void* p) const {
3167 const HeapWord* addr = (HeapWord*)p;
3168 if (_span.contains(p)) {
3169 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3170 return _cmsGen->cmsSpace()->block_start(p);
3171 } else {
3172 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3173 "Inconsistent _span?");
3174 return _permGen->cmsSpace()->block_start(p);
3175 }
3176 }
3177 return NULL;
3178 }
3179 #endif
3180
3181 HeapWord*
3182 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3183 bool tlab,
3184 bool parallel) {
3185 assert(!tlab, "Can't deal with TLAB allocation");
3186 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3187 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3188 CMSExpansionCause::_satisfy_allocation);
3189 if (GCExpandToAllocateDelayMillis > 0) {
3190 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3191 }
3192 return have_lock_and_allocate(word_size, tlab);
3193 }
3194
3195 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3196 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3197 // to CardGeneration and share it...
3198 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3199 CMSExpansionCause::Cause cause)
3200 {
3201 assert_locked_or_safepoint(Heap_lock);
3202
3203 size_t aligned_bytes = ReservedSpace::page_align_size_up(bytes);
3204 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
3205 bool success = false;
3206 if (aligned_expand_bytes > aligned_bytes) {
3207 success = grow_by(aligned_expand_bytes);
3208 }
3209 if (!success) {
3210 success = grow_by(aligned_bytes);
3211 }
3212 if (!success) {
3213 size_t remaining_bytes = _virtual_space.uncommitted_size();
3214 if (remaining_bytes > 0) {
3215 success = grow_by(remaining_bytes);
3216 }
3217 }
3218 if (GC_locker::is_active()) {
3219 if (PrintGC && Verbose) {
3220 gclog_or_tty->print_cr("Garbage collection disabled, expanded heap instead");
3221 }
3222 }
3223 // remember why we expanded; this information is used
3224 // by shouldConcurrentCollect() when making decisions on whether to start
3225 // a new CMS cycle.
3226 if (success) {
3227 set_expansion_cause(cause);
3228 if (PrintGCDetails && Verbose) {
3229 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3230 CMSExpansionCause::to_string(cause));
3231 }
3232 }
3233 }
3234
3235 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3236 HeapWord* res = NULL;
3237 MutexLocker x(ParGCRareEvent_lock);
3238 while (true) {
3239 // Expansion by some other thread might make alloc OK now:
3240 res = ps->lab.alloc(word_sz);
3241 if (res != NULL) return res;
3242 // If there's not enough expansion space available, give up.
3243 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3244 return NULL;
3245 }
3246 // Otherwise, we try expansion.
3247 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3248 CMSExpansionCause::_allocate_par_lab);
3249 // Now go around the loop and try alloc again;
3250 // A competing par_promote might beat us to the expansion space,
3251 // so we may go around the loop again if promotion fails agaion.
3252 if (GCExpandToAllocateDelayMillis > 0) {
3253 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3254 }
3255 }
3256 }
3257
3258
3259 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3260 PromotionInfo* promo) {
3261 MutexLocker x(ParGCRareEvent_lock);
3262 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3263 while (true) {
3264 // Expansion by some other thread might make alloc OK now:
3265 if (promo->ensure_spooling_space()) {
3266 assert(promo->has_spooling_space(),
3267 "Post-condition of successful ensure_spooling_space()");
3268 return true;
3269 }
3270 // If there's not enough expansion space available, give up.
3271 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3272 return false;
3273 }
3274 // Otherwise, we try expansion.
3275 expand(refill_size_bytes, MinHeapDeltaBytes,
3276 CMSExpansionCause::_allocate_par_spooling_space);
3277 // Now go around the loop and try alloc again;
3278 // A competing allocation might beat us to the expansion space,
3279 // so we may go around the loop again if allocation fails again.
3280 if (GCExpandToAllocateDelayMillis > 0) {
3281 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3282 }
3283 }
3284 }
3285
3286
3287
3288 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3289 assert_locked_or_safepoint(Heap_lock);
3290 size_t size = ReservedSpace::page_align_size_down(bytes);
3291 if (size > 0) {
3292 shrink_by(size);
3293 }
3294 }
3295
3296 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3297 assert_locked_or_safepoint(Heap_lock);
3298 bool result = _virtual_space.expand_by(bytes);
3299 if (result) {
3300 HeapWord* old_end = _cmsSpace->end();
3301 size_t new_word_size =
3302 heap_word_size(_virtual_space.committed_size());
3303 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3304 _bts->resize(new_word_size); // resize the block offset shared array
3305 Universe::heap()->barrier_set()->resize_covered_region(mr);
3306 // Hmmmm... why doesn't CFLS::set_end verify locking?
3307 // This is quite ugly; FIX ME XXX
3308 _cmsSpace->assert_locked();
3309 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3310
3311 // update the space and generation capacity counters
3312 if (UsePerfData) {
3313 _space_counters->update_capacity();
3314 _gen_counters->update_all();
3315 }
3316
3317 if (Verbose && PrintGC) {
3318 size_t new_mem_size = _virtual_space.committed_size();
3319 size_t old_mem_size = new_mem_size - bytes;
3320 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3321 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3322 }
3323 }
3324 return result;
3325 }
3326
3327 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3328 assert_locked_or_safepoint(Heap_lock);
3329 bool success = true;
3330 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3331 if (remaining_bytes > 0) {
3332 success = grow_by(remaining_bytes);
3333 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3334 }
3335 return success;
3336 }
3337
3338 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3339 assert_locked_or_safepoint(Heap_lock);
3340 assert_lock_strong(freelistLock());
3341 // XXX Fix when compaction is implemented.
3342 warning("Shrinking of CMS not yet implemented");
3343 return;
3344 }
3345
3346
3347 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3348 // phases.
3349 class CMSPhaseAccounting: public StackObj {
3350 public:
3351 CMSPhaseAccounting(CMSCollector *collector,
3352 const char *phase,
3353 bool print_cr = true);
3354 ~CMSPhaseAccounting();
3355
3356 private:
3357 CMSCollector *_collector;
3358 const char *_phase;
3359 elapsedTimer _wallclock;
3360 bool _print_cr;
3361
3362 public:
3363 // Not MT-safe; so do not pass around these StackObj's
3364 // where they may be accessed by other threads.
3365 jlong wallclock_millis() {
3366 assert(_wallclock.is_active(), "Wall clock should not stop");
3367 _wallclock.stop(); // to record time
3368 jlong ret = _wallclock.milliseconds();
3369 _wallclock.start(); // restart
3370 return ret;
3371 }
3372 };
3373
3374 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3375 const char *phase,
3376 bool print_cr) :
3377 _collector(collector), _phase(phase), _print_cr(print_cr) {
3378
3379 if (PrintCMSStatistics != 0) {
3380 _collector->resetYields();
3381 }
3382 if (PrintGCDetails && PrintGCTimeStamps) {
3383 gclog_or_tty->date_stamp(PrintGCDateStamps);
3384 gclog_or_tty->stamp();
3385 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3386 _collector->cmsGen()->short_name(), _phase);
3387 }
3388 _collector->resetTimer();
3389 _wallclock.start();
3390 _collector->startTimer();
3391 }
3392
3393 CMSPhaseAccounting::~CMSPhaseAccounting() {
3394 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3395 _collector->stopTimer();
3396 _wallclock.stop();
3397 if (PrintGCDetails) {
3398 gclog_or_tty->date_stamp(PrintGCDateStamps);
3399 if (PrintGCTimeStamps) {
3400 gclog_or_tty->stamp();
3401 gclog_or_tty->print(": ");
3402 }
3403 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3404 _collector->cmsGen()->short_name(),
3405 _phase, _collector->timerValue(), _wallclock.seconds());
3406 if (_print_cr) {
3407 gclog_or_tty->print_cr("");
3408 }
3409 if (PrintCMSStatistics != 0) {
3410 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3411 _collector->yields());
3412 }
3413 }
3414 }
3415
3416 // CMS work
3417
3418 // Checkpoint the roots into this generation from outside
3419 // this generation. [Note this initial checkpoint need only
3420 // be approximate -- we'll do a catch up phase subsequently.]
3421 void CMSCollector::checkpointRootsInitial(bool asynch) {
3422 assert(_collectorState == InitialMarking, "Wrong collector state");
3423 check_correct_thread_executing();
3424 ReferenceProcessor* rp = ref_processor();
3425 SpecializationStats::clear();
3426 assert(_restart_addr == NULL, "Control point invariant");
3427 if (asynch) {
3428 // acquire locks for subsequent manipulations
3429 MutexLockerEx x(bitMapLock(),
3430 Mutex::_no_safepoint_check_flag);
3431 checkpointRootsInitialWork(asynch);
3432 rp->verify_no_references_recorded();
3433 rp->enable_discovery(); // enable ("weak") refs discovery
3434 _collectorState = Marking;
3435 } else {
3436 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3437 // which recognizes if we are a CMS generation, and doesn't try to turn on
3438 // discovery; verify that they aren't meddling.
3439 assert(!rp->discovery_is_atomic(),
3440 "incorrect setting of discovery predicate");
3441 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3442 "ref discovery for this generation kind");
3443 // already have locks
3444 checkpointRootsInitialWork(asynch);
3445 rp->enable_discovery(); // now enable ("weak") refs discovery
3446 _collectorState = Marking;
3447 }
3448 SpecializationStats::print();
3449 }
3450
3451 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3452 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3453 assert(_collectorState == InitialMarking, "just checking");
3454
3455 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3456 // precede our marking with a collection of all
3457 // younger generations to keep floating garbage to a minimum.
3458 // XXX: we won't do this for now -- it's an optimization to be done later.
3459
3460 // already have locks
3461 assert_lock_strong(bitMapLock());
3462 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3463
3464 // Setup the verification and class unloading state for this
3465 // CMS collection cycle.
3466 setup_cms_unloading_and_verification_state();
3467
3468 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork",
3469 PrintGCDetails && Verbose, true, gclog_or_tty);)
3470 if (UseAdaptiveSizePolicy) {
3471 size_policy()->checkpoint_roots_initial_begin();
3472 }
3473
3474 // Reset all the PLAB chunk arrays if necessary.
3475 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3476 reset_survivor_plab_arrays();
3477 }
3478
3479 ResourceMark rm;
3480 HandleMark hm;
3481
3482 FalseClosure falseClosure;
3483 // In the case of a synchronous collection, we will elide the
3484 // remark step, so it's important to catch all the nmethod oops
3485 // in this step; hence the last argument to the constrcutor below.
3486 MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */);
3487 GenCollectedHeap* gch = GenCollectedHeap::heap();
3488
3489 verify_work_stacks_empty();
3490 verify_overflow_empty();
3491
3492 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3493 // Update the saved marks which may affect the root scans.
3494 gch->save_marks();
3495
3496 // weak reference processing has not started yet.
3497 ref_processor()->set_enqueuing_is_done(false);
3498
3499 {
3500 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3501 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3502 gch->gen_process_strong_roots(_cmsGen->level(),
3503 true, // younger gens are roots
3504 true, // collecting perm gen
3505 SharedHeap::ScanningOption(roots_scanning_options()),
3506 NULL, ¬Older);
3507 }
3508
3509 // Clear mod-union table; it will be dirtied in the prologue of
3510 // CMS generation per each younger generation collection.
3511
3512 assert(_modUnionTable.isAllClear(),
3513 "Was cleared in most recent final checkpoint phase"
3514 " or no bits are set in the gc_prologue before the start of the next "
3515 "subsequent marking phase.");
3516
3517 // Temporarily disabled, since pre/post-consumption closures don't
3518 // care about precleaned cards
3519 #if 0
3520 {
3521 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(),
3522 (HeapWord*)_virtual_space.high());
3523 _ct->ct_bs()->preclean_dirty_cards(mr);
3524 }
3525 #endif
3526
3527 // Save the end of the used_region of the constituent generations
3528 // to be used to limit the extent of sweep in each generation.
3529 save_sweep_limits();
3530 if (UseAdaptiveSizePolicy) {
3531 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3532 }
3533 verify_overflow_empty();
3534 }
3535
3536 bool CMSCollector::markFromRoots(bool asynch) {
3537 // we might be tempted to assert that:
3538 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3539 // "inconsistent argument?");
3540 // However that wouldn't be right, because it's possible that
3541 // a safepoint is indeed in progress as a younger generation
3542 // stop-the-world GC happens even as we mark in this generation.
3543 assert(_collectorState == Marking, "inconsistent state?");
3544 check_correct_thread_executing();
3545 verify_overflow_empty();
3546
3547 bool res;
3548 if (asynch) {
3549
3550 // Start the timers for adaptive size policy for the concurrent phases
3551 // Do it here so that the foreground MS can use the concurrent
3552 // timer since a foreground MS might has the sweep done concurrently
3553 // or STW.
3554 if (UseAdaptiveSizePolicy) {
3555 size_policy()->concurrent_marking_begin();
3556 }
3557
3558 // Weak ref discovery note: We may be discovering weak
3559 // refs in this generation concurrent (but interleaved) with
3560 // weak ref discovery by a younger generation collector.
3561
3562 CMSTokenSyncWithLocks ts(true, bitMapLock());
3563 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3564 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3565 res = markFromRootsWork(asynch);
3566 if (res) {
3567 _collectorState = Precleaning;
3568 } else { // We failed and a foreground collection wants to take over
3569 assert(_foregroundGCIsActive, "internal state inconsistency");
3570 assert(_restart_addr == NULL, "foreground will restart from scratch");
3571 if (PrintGCDetails) {
3572 gclog_or_tty->print_cr("bailing out to foreground collection");
3573 }
3574 }
3575 if (UseAdaptiveSizePolicy) {
3576 size_policy()->concurrent_marking_end();
3577 }
3578 } else {
3579 assert(SafepointSynchronize::is_at_safepoint(),
3580 "inconsistent with asynch == false");
3581 if (UseAdaptiveSizePolicy) {
3582 size_policy()->ms_collection_marking_begin();
3583 }
3584 // already have locks
3585 res = markFromRootsWork(asynch);
3586 _collectorState = FinalMarking;
3587 if (UseAdaptiveSizePolicy) {
3588 GenCollectedHeap* gch = GenCollectedHeap::heap();
3589 size_policy()->ms_collection_marking_end(gch->gc_cause());
3590 }
3591 }
3592 verify_overflow_empty();
3593 return res;
3594 }
3595
3596 bool CMSCollector::markFromRootsWork(bool asynch) {
3597 // iterate over marked bits in bit map, doing a full scan and mark
3598 // from these roots using the following algorithm:
3599 // . if oop is to the right of the current scan pointer,
3600 // mark corresponding bit (we'll process it later)
3601 // . else (oop is to left of current scan pointer)
3602 // push oop on marking stack
3603 // . drain the marking stack
3604
3605 // Note that when we do a marking step we need to hold the
3606 // bit map lock -- recall that direct allocation (by mutators)
3607 // and promotion (by younger generation collectors) is also
3608 // marking the bit map. [the so-called allocate live policy.]
3609 // Because the implementation of bit map marking is not
3610 // robust wrt simultaneous marking of bits in the same word,
3611 // we need to make sure that there is no such interference
3612 // between concurrent such updates.
3613
3614 // already have locks
3615 assert_lock_strong(bitMapLock());
3616
3617 // Clear the revisit stack, just in case there are any
3618 // obsolete contents from a short-circuited previous CMS cycle.
3619 _revisitStack.reset();
3620 verify_work_stacks_empty();
3621 verify_overflow_empty();
3622 assert(_revisitStack.isEmpty(), "tabula rasa");
3623
3624 bool result = false;
3625 if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) {
3626 result = do_marking_mt(asynch);
3627 } else {
3628 result = do_marking_st(asynch);
3629 }
3630 return result;
3631 }
3632
3633 // Forward decl
3634 class CMSConcMarkingTask;
3635
3636 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3637 CMSCollector* _collector;
3638 CMSConcMarkingTask* _task;
3639 bool _yield;
3640 protected:
3641 virtual void yield();
3642 public:
3643 // "n_threads" is the number of threads to be terminated.
3644 // "queue_set" is a set of work queues of other threads.
3645 // "collector" is the CMS collector associated with this task terminator.
3646 // "yield" indicates whether we need the gang as a whole to yield.
3647 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
3648 CMSCollector* collector, bool yield) :
3649 ParallelTaskTerminator(n_threads, queue_set),
3650 _collector(collector),
3651 _yield(yield) { }
3652
3653 void set_task(CMSConcMarkingTask* task) {
3654 _task = task;
3655 }
3656 };
3657
3658 // MT Concurrent Marking Task
3659 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3660 CMSCollector* _collector;
3661 YieldingFlexibleWorkGang* _workers; // the whole gang
3662 int _n_workers; // requested/desired # workers
3663 bool _asynch;
3664 bool _result;
3665 CompactibleFreeListSpace* _cms_space;
3666 CompactibleFreeListSpace* _perm_space;
3667 HeapWord* _global_finger;
3668
3669 // Exposed here for yielding support
3670 Mutex* const _bit_map_lock;
3671
3672 // The per thread work queues, available here for stealing
3673 OopTaskQueueSet* _task_queues;
3674 CMSConcMarkingTerminator _term;
3675
3676 public:
3677 CMSConcMarkingTask(CMSCollector* collector,
3678 CompactibleFreeListSpace* cms_space,
3679 CompactibleFreeListSpace* perm_space,
3680 bool asynch, int n_workers,
3681 YieldingFlexibleWorkGang* workers,
3682 OopTaskQueueSet* task_queues):
3683 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3684 _collector(collector),
3685 _cms_space(cms_space),
3686 _perm_space(perm_space),
3687 _asynch(asynch), _n_workers(n_workers), _result(true),
3688 _workers(workers), _task_queues(task_queues),
3689 _term(n_workers, task_queues, _collector, asynch),
3690 _bit_map_lock(collector->bitMapLock())
3691 {
3692 assert(n_workers <= workers->total_workers(),
3693 "Else termination won't work correctly today"); // XXX FIX ME!
3694 _requested_size = n_workers;
3695 _term.set_task(this);
3696 assert(_cms_space->bottom() < _perm_space->bottom(),
3697 "Finger incorrectly initialized below");
3698 _global_finger = _cms_space->bottom();
3699 }
3700
3701
3702 OopTaskQueueSet* task_queues() { return _task_queues; }
3703
3704 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3705
3706 HeapWord** global_finger_addr() { return &_global_finger; }
3707
3708 CMSConcMarkingTerminator* terminator() { return &_term; }
3709
3710 void work(int i);
3711
3712 virtual void coordinator_yield(); // stuff done by coordinator
3713 bool result() { return _result; }
3714
3715 void reset(HeapWord* ra) {
3716 _term.reset_for_reuse();
3717 }
3718
3719 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3720 OopTaskQueue* work_q);
3721
3722 private:
3723 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3724 void do_work_steal(int i);
3725 void bump_global_finger(HeapWord* f);
3726 };
3727
3728 void CMSConcMarkingTerminator::yield() {
3729 if (ConcurrentMarkSweepThread::should_yield() &&
3730 !_collector->foregroundGCIsActive() &&
3731 _yield) {
3732 _task->yield();
3733 } else {
3734 ParallelTaskTerminator::yield();
3735 }
3736 }
3737
3738 ////////////////////////////////////////////////////////////////
3739 // Concurrent Marking Algorithm Sketch
3740 ////////////////////////////////////////////////////////////////
3741 // Until all tasks exhausted (both spaces):
3742 // -- claim next available chunk
3743 // -- bump global finger via CAS
3744 // -- find first object that starts in this chunk
3745 // and start scanning bitmap from that position
3746 // -- scan marked objects for oops
3747 // -- CAS-mark target, and if successful:
3748 // . if target oop is above global finger (volatile read)
3749 // nothing to do
3750 // . if target oop is in chunk and above local finger
3751 // then nothing to do
3752 // . else push on work-queue
3753 // -- Deal with possible overflow issues:
3754 // . local work-queue overflow causes stuff to be pushed on
3755 // global (common) overflow queue
3756 // . always first empty local work queue
3757 // . then get a batch of oops from global work queue if any
3758 // . then do work stealing
3759 // -- When all tasks claimed (both spaces)
3760 // and local work queue empty,
3761 // then in a loop do:
3762 // . check global overflow stack; steal a batch of oops and trace
3763 // . try to steal from other threads oif GOS is empty
3764 // . if neither is available, offer termination
3765 // -- Terminate and return result
3766 //
3767 void CMSConcMarkingTask::work(int i) {
3768 elapsedTimer _timer;
3769 ResourceMark rm;
3770 HandleMark hm;
3771
3772 DEBUG_ONLY(_collector->verify_overflow_empty();)
3773
3774 // Before we begin work, our work queue should be empty
3775 assert(work_queue(i)->size() == 0, "Expected to be empty");
3776 // Scan the bitmap covering _cms_space, tracing through grey objects.
3777 _timer.start();
3778 do_scan_and_mark(i, _cms_space);
3779 _timer.stop();
3780 if (PrintCMSStatistics != 0) {
3781 gclog_or_tty->print_cr("Finished cms 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 the same for the _perm_space
3786 _timer.reset();
3787 _timer.start();
3788 do_scan_and_mark(i, _perm_space);
3789 _timer.stop();
3790 if (PrintCMSStatistics != 0) {
3791 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3792 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3793 }
3794
3795 // ... do work stealing
3796 _timer.reset();
3797 _timer.start();
3798 do_work_steal(i);
3799 _timer.stop();
3800 if (PrintCMSStatistics != 0) {
3801 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3802 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers
3803 }
3804 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3805 assert(work_queue(i)->size() == 0, "Should have been emptied");
3806 // Note that under the current task protocol, the
3807 // following assertion is true even of the spaces
3808 // expanded since the completion of the concurrent
3809 // marking. XXX This will likely change under a strict
3810 // ABORT semantics.
3811 assert(_global_finger > _cms_space->end() &&
3812 _global_finger >= _perm_space->end(),
3813 "All tasks have been completed");
3814 DEBUG_ONLY(_collector->verify_overflow_empty();)
3815 }
3816
3817 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3818 HeapWord* read = _global_finger;
3819 HeapWord* cur = read;
3820 while (f > read) {
3821 cur = read;
3822 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3823 if (cur == read) {
3824 // our cas succeeded
3825 assert(_global_finger >= f, "protocol consistency");
3826 break;
3827 }
3828 }
3829 }
3830
3831 // This is really inefficient, and should be redone by
3832 // using (not yet available) block-read and -write interfaces to the
3833 // stack and the work_queue. XXX FIX ME !!!
3834 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3835 OopTaskQueue* work_q) {
3836 // Fast lock-free check
3837 if (ovflw_stk->length() == 0) {
3838 return false;
3839 }
3840 assert(work_q->size() == 0, "Shouldn't steal");
3841 MutexLockerEx ml(ovflw_stk->par_lock(),
3842 Mutex::_no_safepoint_check_flag);
3843 // Grab up to 1/4 the size of the work queue
3844 size_t num = MIN2((size_t)work_q->max_elems()/4,
3845 (size_t)ParGCDesiredObjsFromOverflowList);
3846 num = MIN2(num, ovflw_stk->length());
3847 for (int i = (int) num; i > 0; i--) {
3848 oop cur = ovflw_stk->pop();
3849 assert(cur != NULL, "Counted wrong?");
3850 work_q->push(cur);
3851 }
3852 return num > 0;
3853 }
3854
3855 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3856 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3857 int n_tasks = pst->n_tasks();
3858 // We allow that there may be no tasks to do here because
3859 // we are restarting after a stack overflow.
3860 assert(pst->valid() || n_tasks == 0, "Uninitializd use?");
3861 int nth_task = 0;
3862
3863 HeapWord* start = sp->bottom();
3864 size_t chunk_size = sp->marking_task_size();
3865 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3866 // Having claimed the nth task in this space,
3867 // compute the chunk that it corresponds to:
3868 MemRegion span = MemRegion(start + nth_task*chunk_size,
3869 start + (nth_task+1)*chunk_size);
3870 // Try and bump the global finger via a CAS;
3871 // note that we need to do the global finger bump
3872 // _before_ taking the intersection below, because
3873 // the task corresponding to that region will be
3874 // deemed done even if the used_region() expands
3875 // because of allocation -- as it almost certainly will
3876 // during start-up while the threads yield in the
3877 // closure below.
3878 HeapWord* finger = span.end();
3879 bump_global_finger(finger); // atomically
3880 // There are null tasks here corresponding to chunks
3881 // beyond the "top" address of the space.
3882 span = span.intersection(sp->used_region());
3883 if (!span.is_empty()) { // Non-null task
3884 // We want to skip the first object because
3885 // the protocol is to scan any object in its entirety
3886 // that _starts_ in this span; a fortiori, any
3887 // object starting in an earlier span is scanned
3888 // as part of an earlier claimed task.
3889 // Below we use the "careful" version of block_start
3890 // so we do not try to navigate uninitialized objects.
3891 HeapWord* prev_obj = sp->block_start_careful(span.start());
3892 // Below we use a variant of block_size that uses the
3893 // Printezis bits to avoid waiting for allocated
3894 // objects to become initialized/parsable.
3895 while (prev_obj < span.start()) {
3896 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3897 if (sz > 0) {
3898 prev_obj += sz;
3899 } else {
3900 // In this case we may end up doing a bit of redundant
3901 // scanning, but that appears unavoidable, short of
3902 // locking the free list locks; see bug 6324141.
3903 break;
3904 }
3905 }
3906 if (prev_obj < span.end()) {
3907 MemRegion my_span = MemRegion(prev_obj, span.end());
3908 // Do the marking work within a non-empty span --
3909 // the last argument to the constructor indicates whether the
3910 // iteration should be incremental with periodic yields.
3911 Par_MarkFromRootsClosure cl(this, _collector, my_span,
3912 &_collector->_markBitMap,
3913 work_queue(i),
3914 &_collector->_markStack,
3915 &_collector->_revisitStack,
3916 _asynch);
3917 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3918 } // else nothing to do for this task
3919 } // else nothing to do for this task
3920 }
3921 // We'd be tempted to assert here that since there are no
3922 // more tasks left to claim in this space, the global_finger
3923 // must exceed space->top() and a fortiori space->end(). However,
3924 // that would not quite be correct because the bumping of
3925 // global_finger occurs strictly after the claiming of a task,
3926 // so by the time we reach here the global finger may not yet
3927 // have been bumped up by the thread that claimed the last
3928 // task.
3929 pst->all_tasks_completed();
3930 }
3931
3932 class Par_ConcMarkingClosure: public OopClosure {
3933 private:
3934 CMSCollector* _collector;
3935 MemRegion _span;
3936 CMSBitMap* _bit_map;
3937 CMSMarkStack* _overflow_stack;
3938 CMSMarkStack* _revisit_stack; // XXXXXX Check proper use
3939 OopTaskQueue* _work_queue;
3940 protected:
3941 DO_OOP_WORK_DEFN
3942 public:
3943 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
3944 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3945 _collector(collector),
3946 _span(_collector->_span),
3947 _work_queue(work_queue),
3948 _bit_map(bit_map),
3949 _overflow_stack(overflow_stack) { } // need to initialize revisit stack etc.
3950 virtual void do_oop(oop* p);
3951 virtual void do_oop(narrowOop* p);
3952 void trim_queue(size_t max);
3953 void handle_stack_overflow(HeapWord* lost);
3954 };
3955
3956 // Grey object rescan during work stealing phase --
3957 // the salient assumption here is that stolen oops must
3958 // always be initialized, so we do not need to check for
3959 // uninitialized objects before scanning here.
3960 void Par_ConcMarkingClosure::do_oop(oop obj) {
3961 assert(obj->is_oop_or_null(), "expected an oop or NULL");
3962 HeapWord* addr = (HeapWord*)obj;
3963 // Check if oop points into the CMS generation
3964 // and is not marked
3965 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3966 // a white object ...
3967 // If we manage to "claim" the object, by being the
3968 // first thread to mark it, then we push it on our
3969 // marking stack
3970 if (_bit_map->par_mark(addr)) { // ... now grey
3971 // push on work queue (grey set)
3972 bool simulate_overflow = false;
3973 NOT_PRODUCT(
3974 if (CMSMarkStackOverflowALot &&
3975 _collector->simulate_overflow()) {
3976 // simulate a stack overflow
3977 simulate_overflow = true;
3978 }
3979 )
3980 if (simulate_overflow ||
3981 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3982 // stack overflow
3983 if (PrintCMSStatistics != 0) {
3984 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
3985 SIZE_FORMAT, _overflow_stack->capacity());
3986 }
3987 // We cannot assert that the overflow stack is full because
3988 // it may have been emptied since.
3989 assert(simulate_overflow ||
3990 _work_queue->size() == _work_queue->max_elems(),
3991 "Else push should have succeeded");
3992 handle_stack_overflow(addr);
3993 }
3994 } // Else, some other thread got there first
3995 }
3996 }
3997
3998 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
3999 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4000
4001 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4002 while (_work_queue->size() > max) {
4003 oop new_oop;
4004 if (_work_queue->pop_local(new_oop)) {
4005 assert(new_oop->is_oop(), "Should be an oop");
4006 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4007 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4008 assert(new_oop->is_parsable(), "Should be parsable");
4009 new_oop->oop_iterate(this); // do_oop() above
4010 }
4011 }
4012 }
4013
4014 // Upon stack overflow, we discard (part of) the stack,
4015 // remembering the least address amongst those discarded
4016 // in CMSCollector's _restart_address.
4017 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4018 // We need to do this under a mutex to prevent other
4019 // workers from interfering with the expansion below.
4020 MutexLockerEx ml(_overflow_stack->par_lock(),
4021 Mutex::_no_safepoint_check_flag);
4022 // Remember the least grey address discarded
4023 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4024 _collector->lower_restart_addr(ra);
4025 _overflow_stack->reset(); // discard stack contents
4026 _overflow_stack->expand(); // expand the stack if possible
4027 }
4028
4029
4030 void CMSConcMarkingTask::do_work_steal(int i) {
4031 OopTaskQueue* work_q = work_queue(i);
4032 oop obj_to_scan;
4033 CMSBitMap* bm = &(_collector->_markBitMap);
4034 CMSMarkStack* ovflw = &(_collector->_markStack);
4035 int* seed = _collector->hash_seed(i);
4036 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw);
4037 while (true) {
4038 cl.trim_queue(0);
4039 assert(work_q->size() == 0, "Should have been emptied above");
4040 if (get_work_from_overflow_stack(ovflw, work_q)) {
4041 // Can't assert below because the work obtained from the
4042 // overflow stack may already have been stolen from us.
4043 // assert(work_q->size() > 0, "Work from overflow stack");
4044 continue;
4045 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4046 assert(obj_to_scan->is_oop(), "Should be an oop");
4047 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4048 obj_to_scan->oop_iterate(&cl);
4049 } else if (terminator()->offer_termination()) {
4050 assert(work_q->size() == 0, "Impossible!");
4051 break;
4052 }
4053 }
4054 }
4055
4056 // This is run by the CMS (coordinator) thread.
4057 void CMSConcMarkingTask::coordinator_yield() {
4058 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4059 "CMS thread should hold CMS token");
4060
4061 // First give up the locks, then yield, then re-lock
4062 // We should probably use a constructor/destructor idiom to
4063 // do this unlock/lock or modify the MutexUnlocker class to
4064 // serve our purpose. XXX
4065 assert_lock_strong(_bit_map_lock);
4066 _bit_map_lock->unlock();
4067 ConcurrentMarkSweepThread::desynchronize(true);
4068 ConcurrentMarkSweepThread::acknowledge_yield_request();
4069 _collector->stopTimer();
4070 if (PrintCMSStatistics != 0) {
4071 _collector->incrementYields();
4072 }
4073 _collector->icms_wait();
4074
4075 // It is possible for whichever thread initiated the yield request
4076 // not to get a chance to wake up and take the bitmap lock between
4077 // this thread releasing it and reacquiring it. So, while the
4078 // should_yield() flag is on, let's sleep for a bit to give the
4079 // other thread a chance to wake up. The limit imposed on the number
4080 // of iterations is defensive, to avoid any unforseen circumstances
4081 // putting us into an infinite loop. Since it's always been this
4082 // (coordinator_yield()) method that was observed to cause the
4083 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4084 // which is by default non-zero. For the other seven methods that
4085 // also perform the yield operation, as are using a different
4086 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4087 // can enable the sleeping for those methods too, if necessary.
4088 // See 6442774.
4089 //
4090 // We really need to reconsider the synchronization between the GC
4091 // thread and the yield-requesting threads in the future and we
4092 // should really use wait/notify, which is the recommended
4093 // way of doing this type of interaction. Additionally, we should
4094 // consolidate the eight methods that do the yield operation and they
4095 // are almost identical into one for better maintenability and
4096 // readability. See 6445193.
4097 //
4098 // Tony 2006.06.29
4099 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4100 ConcurrentMarkSweepThread::should_yield() &&
4101 !CMSCollector::foregroundGCIsActive(); ++i) {
4102 os::sleep(Thread::current(), 1, false);
4103 ConcurrentMarkSweepThread::acknowledge_yield_request();
4104 }
4105
4106 ConcurrentMarkSweepThread::synchronize(true);
4107 _bit_map_lock->lock_without_safepoint_check();
4108 _collector->startTimer();
4109 }
4110
4111 bool CMSCollector::do_marking_mt(bool asynch) {
4112 assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition");
4113 // In the future this would be determined ergonomically, based
4114 // on #cpu's, # active mutator threads (and load), and mutation rate.
4115 int num_workers = ParallelCMSThreads;
4116
4117 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4118 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4119
4120 CMSConcMarkingTask tsk(this, cms_space, perm_space,
4121 asynch, num_workers /* number requested XXX */,
4122 conc_workers(), task_queues());
4123
4124 // Since the actual number of workers we get may be different
4125 // from the number we requested above, do we need to do anything different
4126 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4127 // class?? XXX
4128 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4129 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4130
4131 // Refs discovery is already non-atomic.
4132 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4133 // Mutate the Refs discovery so it is MT during the
4134 // multi-threaded marking phase.
4135 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1);
4136
4137 conc_workers()->start_task(&tsk);
4138 while (tsk.yielded()) {
4139 tsk.coordinator_yield();
4140 conc_workers()->continue_task(&tsk);
4141 }
4142 // If the task was aborted, _restart_addr will be non-NULL
4143 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4144 while (_restart_addr != NULL) {
4145 // XXX For now we do not make use of ABORTED state and have not
4146 // yet implemented the right abort semantics (even in the original
4147 // single-threaded CMS case). That needs some more investigation
4148 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4149 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4150 // If _restart_addr is non-NULL, a marking stack overflow
4151 // occured; we need to do a fresh marking iteration from the
4152 // indicated restart address.
4153 if (_foregroundGCIsActive && asynch) {
4154 // We may be running into repeated stack overflows, having
4155 // reached the limit of the stack size, while making very
4156 // slow forward progress. It may be best to bail out and
4157 // let the foreground collector do its job.
4158 // Clear _restart_addr, so that foreground GC
4159 // works from scratch. This avoids the headache of
4160 // a "rescan" which would otherwise be needed because
4161 // of the dirty mod union table & card table.
4162 _restart_addr = NULL;
4163 return false;
4164 }
4165 // Adjust the task to restart from _restart_addr
4166 tsk.reset(_restart_addr);
4167 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4168 _restart_addr);
4169 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4170 _restart_addr);
4171 _restart_addr = NULL;
4172 // Get the workers going again
4173 conc_workers()->start_task(&tsk);
4174 while (tsk.yielded()) {
4175 tsk.coordinator_yield();
4176 conc_workers()->continue_task(&tsk);
4177 }
4178 }
4179 assert(tsk.completed(), "Inconsistency");
4180 assert(tsk.result() == true, "Inconsistency");
4181 return true;
4182 }
4183
4184 bool CMSCollector::do_marking_st(bool asynch) {
4185 ResourceMark rm;
4186 HandleMark hm;
4187
4188 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4189 &_markStack, &_revisitStack, CMSYield && asynch);
4190 // the last argument to iterate indicates whether the iteration
4191 // should be incremental with periodic yields.
4192 _markBitMap.iterate(&markFromRootsClosure);
4193 // If _restart_addr is non-NULL, a marking stack overflow
4194 // occured; we need to do a fresh iteration from the
4195 // indicated restart address.
4196 while (_restart_addr != NULL) {
4197 if (_foregroundGCIsActive && asynch) {
4198 // We may be running into repeated stack overflows, having
4199 // reached the limit of the stack size, while making very
4200 // slow forward progress. It may be best to bail out and
4201 // let the foreground collector do its job.
4202 // Clear _restart_addr, so that foreground GC
4203 // works from scratch. This avoids the headache of
4204 // a "rescan" which would otherwise be needed because
4205 // of the dirty mod union table & card table.
4206 _restart_addr = NULL;
4207 return false; // indicating failure to complete marking
4208 }
4209 // Deal with stack overflow:
4210 // we restart marking from _restart_addr
4211 HeapWord* ra = _restart_addr;
4212 markFromRootsClosure.reset(ra);
4213 _restart_addr = NULL;
4214 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4215 }
4216 return true;
4217 }
4218
4219 void CMSCollector::preclean() {
4220 check_correct_thread_executing();
4221 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4222 verify_work_stacks_empty();
4223 verify_overflow_empty();
4224 _abort_preclean = false;
4225 if (CMSPrecleaningEnabled) {
4226 _eden_chunk_index = 0;
4227 size_t used = get_eden_used();
4228 size_t capacity = get_eden_capacity();
4229 // Don't start sampling unless we will get sufficiently
4230 // many samples.
4231 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4232 * CMSScheduleRemarkEdenPenetration)) {
4233 _start_sampling = true;
4234 } else {
4235 _start_sampling = false;
4236 }
4237 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4238 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4239 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4240 }
4241 CMSTokenSync x(true); // is cms thread
4242 if (CMSPrecleaningEnabled) {
4243 sample_eden();
4244 _collectorState = AbortablePreclean;
4245 } else {
4246 _collectorState = FinalMarking;
4247 }
4248 verify_work_stacks_empty();
4249 verify_overflow_empty();
4250 }
4251
4252 // Try and schedule the remark such that young gen
4253 // occupancy is CMSScheduleRemarkEdenPenetration %.
4254 void CMSCollector::abortable_preclean() {
4255 check_correct_thread_executing();
4256 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4257 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4258
4259 // If Eden's current occupancy is below this threshold,
4260 // immediately schedule the remark; else preclean
4261 // past the next scavenge in an effort to
4262 // schedule the pause as described avove. By choosing
4263 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4264 // we will never do an actual abortable preclean cycle.
4265 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4266 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4267 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4268 // We need more smarts in the abortable preclean
4269 // loop below to deal with cases where allocation
4270 // in young gen is very very slow, and our precleaning
4271 // is running a losing race against a horde of
4272 // mutators intent on flooding us with CMS updates
4273 // (dirty cards).
4274 // One, admittedly dumb, strategy is to give up
4275 // after a certain number of abortable precleaning loops
4276 // or after a certain maximum time. We want to make
4277 // this smarter in the next iteration.
4278 // XXX FIX ME!!! YSR
4279 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4280 while (!(should_abort_preclean() ||
4281 ConcurrentMarkSweepThread::should_terminate())) {
4282 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4283 cumworkdone += workdone;
4284 loops++;
4285 // Voluntarily terminate abortable preclean phase if we have
4286 // been at it for too long.
4287 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4288 loops >= CMSMaxAbortablePrecleanLoops) {
4289 if (PrintGCDetails) {
4290 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4291 }
4292 break;
4293 }
4294 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4295 if (PrintGCDetails) {
4296 gclog_or_tty->print(" CMS: abort preclean due to time ");
4297 }
4298 break;
4299 }
4300 // If we are doing little work each iteration, we should
4301 // take a short break.
4302 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4303 // Sleep for some time, waiting for work to accumulate
4304 stopTimer();
4305 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4306 startTimer();
4307 waited++;
4308 }
4309 }
4310 if (PrintCMSStatistics > 0) {
4311 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4312 loops, waited, cumworkdone);
4313 }
4314 }
4315 CMSTokenSync x(true); // is cms thread
4316 if (_collectorState != Idling) {
4317 assert(_collectorState == AbortablePreclean,
4318 "Spontaneous state transition?");
4319 _collectorState = FinalMarking;
4320 } // Else, a foreground collection completed this CMS cycle.
4321 return;
4322 }
4323
4324 // Respond to an Eden sampling opportunity
4325 void CMSCollector::sample_eden() {
4326 // Make sure a young gc cannot sneak in between our
4327 // reading and recording of a sample.
4328 assert(Thread::current()->is_ConcurrentGC_thread(),
4329 "Only the cms thread may collect Eden samples");
4330 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4331 "Should collect samples while holding CMS token");
4332 if (!_start_sampling) {
4333 return;
4334 }
4335 if (_eden_chunk_array) {
4336 if (_eden_chunk_index < _eden_chunk_capacity) {
4337 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4338 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4339 "Unexpected state of Eden");
4340 // We'd like to check that what we just sampled is an oop-start address;
4341 // however, we cannot do that here since the object may not yet have been
4342 // initialized. So we'll instead do the check when we _use_ this sample
4343 // later.
4344 if (_eden_chunk_index == 0 ||
4345 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4346 _eden_chunk_array[_eden_chunk_index-1])
4347 >= CMSSamplingGrain)) {
4348 _eden_chunk_index++; // commit sample
4349 }
4350 }
4351 }
4352 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4353 size_t used = get_eden_used();
4354 size_t capacity = get_eden_capacity();
4355 assert(used <= capacity, "Unexpected state of Eden");
4356 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4357 _abort_preclean = true;
4358 }
4359 }
4360 }
4361
4362
4363 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4364 assert(_collectorState == Precleaning ||
4365 _collectorState == AbortablePreclean, "incorrect state");
4366 ResourceMark rm;
4367 HandleMark hm;
4368 // Do one pass of scrubbing the discovered reference lists
4369 // to remove any reference objects with strongly-reachable
4370 // referents.
4371 if (clean_refs) {
4372 ReferenceProcessor* rp = ref_processor();
4373 CMSPrecleanRefsYieldClosure yield_cl(this);
4374 assert(rp->span().equals(_span), "Spans should be equal");
4375 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4376 &_markStack);
4377 CMSDrainMarkingStackClosure complete_trace(this,
4378 _span, &_markBitMap, &_markStack,
4379 &keep_alive);
4380
4381 // We don't want this step to interfere with a young
4382 // collection because we don't want to take CPU
4383 // or memory bandwidth away from the young GC threads
4384 // (which may be as many as there are CPUs).
4385 // Note that we don't need to protect ourselves from
4386 // interference with mutators because they can't
4387 // manipulate the discovered reference lists nor affect
4388 // the computed reachability of the referents, the
4389 // only properties manipulated by the precleaning
4390 // of these reference lists.
4391 stopTimer();
4392 CMSTokenSyncWithLocks x(true /* is cms thread */,
4393 bitMapLock());
4394 startTimer();
4395 sample_eden();
4396 // The following will yield to allow foreground
4397 // collection to proceed promptly. XXX YSR:
4398 // The code in this method may need further
4399 // tweaking for better performance and some restructuring
4400 // for cleaner interfaces.
4401 rp->preclean_discovered_references(
4402 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4403 &yield_cl);
4404 }
4405
4406 if (clean_survivor) { // preclean the active survivor space(s)
4407 assert(_young_gen->kind() == Generation::DefNew ||
4408 _young_gen->kind() == Generation::ParNew ||
4409 _young_gen->kind() == Generation::ASParNew,
4410 "incorrect type for cast");
4411 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4412 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4413 &_markBitMap, &_modUnionTable,
4414 &_markStack, &_revisitStack,
4415 true /* precleaning phase */);
4416 stopTimer();
4417 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4418 bitMapLock());
4419 startTimer();
4420 unsigned int before_count =
4421 GenCollectedHeap::heap()->total_collections();
4422 SurvivorSpacePrecleanClosure
4423 sss_cl(this, _span, &_markBitMap, &_markStack,
4424 &pam_cl, before_count, CMSYield);
4425 dng->from()->object_iterate_careful(&sss_cl);
4426 dng->to()->object_iterate_careful(&sss_cl);
4427 }
4428 MarkRefsIntoAndScanClosure
4429 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4430 &_markStack, &_revisitStack, this, CMSYield,
4431 true /* precleaning phase */);
4432 // CAUTION: The following closure has persistent state that may need to
4433 // be reset upon a decrease in the sequence of addresses it
4434 // processes.
4435 ScanMarkedObjectsAgainCarefullyClosure
4436 smoac_cl(this, _span,
4437 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4438
4439 // Preclean dirty cards in ModUnionTable and CardTable using
4440 // appropriate convergence criterion;
4441 // repeat CMSPrecleanIter times unless we find that
4442 // we are losing.
4443 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4444 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4445 "Bad convergence multiplier");
4446 assert(CMSPrecleanThreshold >= 100,
4447 "Unreasonably low CMSPrecleanThreshold");
4448
4449 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4450 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4451 numIter < CMSPrecleanIter;
4452 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4453 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4454 if (CMSPermGenPrecleaningEnabled) {
4455 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4456 }
4457 if (Verbose && PrintGCDetails) {
4458 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4459 }
4460 // Either there are very few dirty cards, so re-mark
4461 // pause will be small anyway, or our pre-cleaning isn't
4462 // that much faster than the rate at which cards are being
4463 // dirtied, so we might as well stop and re-mark since
4464 // precleaning won't improve our re-mark time by much.
4465 if (curNumCards <= CMSPrecleanThreshold ||
4466 (numIter > 0 &&
4467 (curNumCards * CMSPrecleanDenominator >
4468 lastNumCards * CMSPrecleanNumerator))) {
4469 numIter++;
4470 cumNumCards += curNumCards;
4471 break;
4472 }
4473 }
4474 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4475 if (CMSPermGenPrecleaningEnabled) {
4476 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4477 }
4478 cumNumCards += curNumCards;
4479 if (PrintGCDetails && PrintCMSStatistics != 0) {
4480 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4481 curNumCards, cumNumCards, numIter);
4482 }
4483 return cumNumCards; // as a measure of useful work done
4484 }
4485
4486 // PRECLEANING NOTES:
4487 // Precleaning involves:
4488 // . reading the bits of the modUnionTable and clearing the set bits.
4489 // . For the cards corresponding to the set bits, we scan the
4490 // objects on those cards. This means we need the free_list_lock
4491 // so that we can safely iterate over the CMS space when scanning
4492 // for oops.
4493 // . When we scan the objects, we'll be both reading and setting
4494 // marks in the marking bit map, so we'll need the marking bit map.
4495 // . For protecting _collector_state transitions, we take the CGC_lock.
4496 // Note that any races in the reading of of card table entries by the
4497 // CMS thread on the one hand and the clearing of those entries by the
4498 // VM thread or the setting of those entries by the mutator threads on the
4499 // other are quite benign. However, for efficiency it makes sense to keep
4500 // the VM thread from racing with the CMS thread while the latter is
4501 // dirty card info to the modUnionTable. We therefore also use the
4502 // CGC_lock to protect the reading of the card table and the mod union
4503 // table by the CM thread.
4504 // . We run concurrently with mutator updates, so scanning
4505 // needs to be done carefully -- we should not try to scan
4506 // potentially uninitialized objects.
4507 //
4508 // Locking strategy: While holding the CGC_lock, we scan over and
4509 // reset a maximal dirty range of the mod union / card tables, then lock
4510 // the free_list_lock and bitmap lock to do a full marking, then
4511 // release these locks; and repeat the cycle. This allows for a
4512 // certain amount of fairness in the sharing of these locks between
4513 // the CMS collector on the one hand, and the VM thread and the
4514 // mutators on the other.
4515
4516 // NOTE: preclean_mod_union_table() and preclean_card_table()
4517 // further below are largely identical; if you need to modify
4518 // one of these methods, please check the other method too.
4519
4520 size_t CMSCollector::preclean_mod_union_table(
4521 ConcurrentMarkSweepGeneration* gen,
4522 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4523 verify_work_stacks_empty();
4524 verify_overflow_empty();
4525
4526 // strategy: starting with the first card, accumulate contiguous
4527 // ranges of dirty cards; clear these cards, then scan the region
4528 // covered by these cards.
4529
4530 // Since all of the MUT is committed ahead, we can just use
4531 // that, in case the generations expand while we are precleaning.
4532 // It might also be fine to just use the committed part of the
4533 // generation, but we might potentially miss cards when the
4534 // generation is rapidly expanding while we are in the midst
4535 // of precleaning.
4536 HeapWord* startAddr = gen->reserved().start();
4537 HeapWord* endAddr = gen->reserved().end();
4538
4539 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4540
4541 size_t numDirtyCards, cumNumDirtyCards;
4542 HeapWord *nextAddr, *lastAddr;
4543 for (cumNumDirtyCards = numDirtyCards = 0,
4544 nextAddr = lastAddr = startAddr;
4545 nextAddr < endAddr;
4546 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4547
4548 ResourceMark rm;
4549 HandleMark hm;
4550
4551 MemRegion dirtyRegion;
4552 {
4553 stopTimer();
4554 CMSTokenSync ts(true);
4555 startTimer();
4556 sample_eden();
4557 // Get dirty region starting at nextOffset (inclusive),
4558 // simultaneously clearing it.
4559 dirtyRegion =
4560 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4561 assert(dirtyRegion.start() >= nextAddr,
4562 "returned region inconsistent?");
4563 }
4564 // Remember where the next search should begin.
4565 // The returned region (if non-empty) is a right open interval,
4566 // so lastOffset is obtained from the right end of that
4567 // interval.
4568 lastAddr = dirtyRegion.end();
4569 // Should do something more transparent and less hacky XXX
4570 numDirtyCards =
4571 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4572
4573 // We'll scan the cards in the dirty region (with periodic
4574 // yields for foreground GC as needed).
4575 if (!dirtyRegion.is_empty()) {
4576 assert(numDirtyCards > 0, "consistency check");
4577 HeapWord* stop_point = NULL;
4578 {
4579 stopTimer();
4580 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4581 bitMapLock());
4582 startTimer();
4583 verify_work_stacks_empty();
4584 verify_overflow_empty();
4585 sample_eden();
4586 stop_point =
4587 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4588 }
4589 if (stop_point != NULL) {
4590 // The careful iteration stopped early either because it found an
4591 // uninitialized object, or because we were in the midst of an
4592 // "abortable preclean", which should now be aborted. Redirty
4593 // the bits corresponding to the partially-scanned or unscanned
4594 // cards. We'll either restart at the next block boundary or
4595 // abort the preclean.
4596 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4597 (_collectorState == AbortablePreclean && should_abort_preclean()),
4598 "Unparsable objects should only be in perm gen.");
4599
4600 stopTimer();
4601 CMSTokenSyncWithLocks ts(true, bitMapLock());
4602 startTimer();
4603 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4604 if (should_abort_preclean()) {
4605 break; // out of preclean loop
4606 } else {
4607 // Compute the next address at which preclean should pick up;
4608 // might need bitMapLock in order to read P-bits.
4609 lastAddr = next_card_start_after_block(stop_point);
4610 }
4611 }
4612 } else {
4613 assert(lastAddr == endAddr, "consistency check");
4614 assert(numDirtyCards == 0, "consistency check");
4615 break;
4616 }
4617 }
4618 verify_work_stacks_empty();
4619 verify_overflow_empty();
4620 return cumNumDirtyCards;
4621 }
4622
4623 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4624 // below are largely identical; if you need to modify
4625 // one of these methods, please check the other method too.
4626
4627 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4628 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4629 // strategy: it's similar to precleamModUnionTable above, in that
4630 // we accumulate contiguous ranges of dirty cards, mark these cards
4631 // precleaned, then scan the region covered by these cards.
4632 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4633 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4634
4635 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4636
4637 size_t numDirtyCards, cumNumDirtyCards;
4638 HeapWord *lastAddr, *nextAddr;
4639
4640 for (cumNumDirtyCards = numDirtyCards = 0,
4641 nextAddr = lastAddr = startAddr;
4642 nextAddr < endAddr;
4643 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4644
4645 ResourceMark rm;
4646 HandleMark hm;
4647
4648 MemRegion dirtyRegion;
4649 {
4650 // See comments in "Precleaning notes" above on why we
4651 // do this locking. XXX Could the locking overheads be
4652 // too high when dirty cards are sparse? [I don't think so.]
4653 stopTimer();
4654 CMSTokenSync x(true); // is cms thread
4655 startTimer();
4656 sample_eden();
4657 // Get and clear dirty region from card table
4658 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_preclean(
4659 MemRegion(nextAddr, endAddr));
4660 assert(dirtyRegion.start() >= nextAddr,
4661 "returned region inconsistent?");
4662 }
4663 lastAddr = dirtyRegion.end();
4664 numDirtyCards =
4665 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4666
4667 if (!dirtyRegion.is_empty()) {
4668 stopTimer();
4669 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4670 startTimer();
4671 sample_eden();
4672 verify_work_stacks_empty();
4673 verify_overflow_empty();
4674 HeapWord* stop_point =
4675 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4676 if (stop_point != NULL) {
4677 // The careful iteration stopped early because it found an
4678 // uninitialized object. Redirty the bits corresponding to the
4679 // partially-scanned or unscanned cards, and start again at the
4680 // next block boundary.
4681 assert(CMSPermGenPrecleaningEnabled ||
4682 (_collectorState == AbortablePreclean && should_abort_preclean()),
4683 "Unparsable objects should only be in perm gen.");
4684 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4685 if (should_abort_preclean()) {
4686 break; // out of preclean loop
4687 } else {
4688 // Compute the next address at which preclean should pick up.
4689 lastAddr = next_card_start_after_block(stop_point);
4690 }
4691 }
4692 } else {
4693 break;
4694 }
4695 }
4696 verify_work_stacks_empty();
4697 verify_overflow_empty();
4698 return cumNumDirtyCards;
4699 }
4700
4701 void CMSCollector::checkpointRootsFinal(bool asynch,
4702 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4703 assert(_collectorState == FinalMarking, "incorrect state transition?");
4704 check_correct_thread_executing();
4705 // world is stopped at this checkpoint
4706 assert(SafepointSynchronize::is_at_safepoint(),
4707 "world should be stopped");
4708 verify_work_stacks_empty();
4709 verify_overflow_empty();
4710
4711 SpecializationStats::clear();
4712 if (PrintGCDetails) {
4713 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4714 _young_gen->used() / K,
4715 _young_gen->capacity() / K);
4716 }
4717 if (asynch) {
4718 if (CMSScavengeBeforeRemark) {
4719 GenCollectedHeap* gch = GenCollectedHeap::heap();
4720 // Temporarily set flag to false, GCH->do_collection will
4721 // expect it to be false and set to true
4722 FlagSetting fl(gch->_is_gc_active, false);
4723 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark",
4724 PrintGCDetails && Verbose, true, gclog_or_tty);)
4725 int level = _cmsGen->level() - 1;
4726 if (level >= 0) {
4727 gch->do_collection(true, // full (i.e. force, see below)
4728 false, // !clear_all_soft_refs
4729 0, // size
4730 false, // is_tlab
4731 level // max_level
4732 );
4733 }
4734 }
4735 FreelistLocker x(this);
4736 MutexLockerEx y(bitMapLock(),
4737 Mutex::_no_safepoint_check_flag);
4738 assert(!init_mark_was_synchronous, "but that's impossible!");
4739 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4740 } else {
4741 // already have all the locks
4742 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4743 init_mark_was_synchronous);
4744 }
4745 verify_work_stacks_empty();
4746 verify_overflow_empty();
4747 SpecializationStats::print();
4748 }
4749
4750 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4751 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4752
4753 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);)
4754
4755 assert(haveFreelistLocks(), "must have free list locks");
4756 assert_lock_strong(bitMapLock());
4757
4758 if (UseAdaptiveSizePolicy) {
4759 size_policy()->checkpoint_roots_final_begin();
4760 }
4761
4762 ResourceMark rm;
4763 HandleMark hm;
4764
4765 GenCollectedHeap* gch = GenCollectedHeap::heap();
4766
4767 if (should_unload_classes()) {
4768 CodeCache::gc_prologue();
4769 }
4770 assert(haveFreelistLocks(), "must have free list locks");
4771 assert_lock_strong(bitMapLock());
4772
4773 if (!init_mark_was_synchronous) {
4774 // We might assume that we need not fill TLAB's when
4775 // CMSScavengeBeforeRemark is set, because we may have just done
4776 // a scavenge which would have filled all TLAB's -- and besides
4777 // Eden would be empty. This however may not always be the case --
4778 // for instance although we asked for a scavenge, it may not have
4779 // happened because of a JNI critical section. We probably need
4780 // a policy for deciding whether we can in that case wait until
4781 // the critical section releases and then do the remark following
4782 // the scavenge, and skip it here. In the absence of that policy,
4783 // or of an indication of whether the scavenge did indeed occur,
4784 // we cannot rely on TLAB's having been filled and must do
4785 // so here just in case a scavenge did not happen.
4786 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4787 // Update the saved marks which may affect the root scans.
4788 gch->save_marks();
4789
4790 {
4791 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4792
4793 // Note on the role of the mod union table:
4794 // Since the marker in "markFromRoots" marks concurrently with
4795 // mutators, it is possible for some reachable objects not to have been
4796 // scanned. For instance, an only reference to an object A was
4797 // placed in object B after the marker scanned B. Unless B is rescanned,
4798 // A would be collected. Such updates to references in marked objects
4799 // are detected via the mod union table which is the set of all cards
4800 // dirtied since the first checkpoint in this GC cycle and prior to
4801 // the most recent young generation GC, minus those cleaned up by the
4802 // concurrent precleaning.
4803 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
4804 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
4805 do_remark_parallel();
4806 } else {
4807 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4808 gclog_or_tty);
4809 do_remark_non_parallel();
4810 }
4811 }
4812 } else {
4813 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4814 // The initial mark was stop-world, so there's no rescanning to
4815 // do; go straight on to the next step below.
4816 }
4817 verify_work_stacks_empty();
4818 verify_overflow_empty();
4819
4820 {
4821 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);)
4822 refProcessingWork(asynch, clear_all_soft_refs);
4823 }
4824 verify_work_stacks_empty();
4825 verify_overflow_empty();
4826
4827 if (should_unload_classes()) {
4828 CodeCache::gc_epilogue();
4829 }
4830
4831 // If we encountered any (marking stack / work queue) overflow
4832 // events during the current CMS cycle, take appropriate
4833 // remedial measures, where possible, so as to try and avoid
4834 // recurrence of that condition.
4835 assert(_markStack.isEmpty(), "No grey objects");
4836 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4837 _ser_kac_ovflw;
4838 if (ser_ovflw > 0) {
4839 if (PrintCMSStatistics != 0) {
4840 gclog_or_tty->print_cr("Marking stack overflow (benign) "
4841 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4842 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
4843 _ser_kac_ovflw);
4844 }
4845 _markStack.expand();
4846 _ser_pmc_remark_ovflw = 0;
4847 _ser_pmc_preclean_ovflw = 0;
4848 _ser_kac_ovflw = 0;
4849 }
4850 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4851 if (PrintCMSStatistics != 0) {
4852 gclog_or_tty->print_cr("Work queue overflow (benign) "
4853 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
4854 _par_pmc_remark_ovflw, _par_kac_ovflw);
4855 }
4856 _par_pmc_remark_ovflw = 0;
4857 _par_kac_ovflw = 0;
4858 }
4859 if (PrintCMSStatistics != 0) {
4860 if (_markStack._hit_limit > 0) {
4861 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
4862 _markStack._hit_limit);
4863 }
4864 if (_markStack._failed_double > 0) {
4865 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
4866 " current capacity "SIZE_FORMAT,
4867 _markStack._failed_double,
4868 _markStack.capacity());
4869 }
4870 }
4871 _markStack._hit_limit = 0;
4872 _markStack._failed_double = 0;
4873
4874 if ((VerifyAfterGC || VerifyDuringGC) &&
4875 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
4876 verify_after_remark();
4877 }
4878
4879 // Change under the freelistLocks.
4880 _collectorState = Sweeping;
4881 // Call isAllClear() under bitMapLock
4882 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
4883 " final marking");
4884 if (UseAdaptiveSizePolicy) {
4885 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
4886 }
4887 }
4888
4889 // Parallel remark task
4890 class CMSParRemarkTask: public AbstractGangTask {
4891 CMSCollector* _collector;
4892 WorkGang* _workers;
4893 int _n_workers;
4894 CompactibleFreeListSpace* _cms_space;
4895 CompactibleFreeListSpace* _perm_space;
4896
4897 // The per-thread work queues, available here for stealing.
4898 OopTaskQueueSet* _task_queues;
4899 ParallelTaskTerminator _term;
4900
4901 public:
4902 CMSParRemarkTask(CMSCollector* collector,
4903 CompactibleFreeListSpace* cms_space,
4904 CompactibleFreeListSpace* perm_space,
4905 int n_workers, WorkGang* workers,
4906 OopTaskQueueSet* task_queues):
4907 AbstractGangTask("Rescan roots and grey objects in parallel"),
4908 _collector(collector),
4909 _cms_space(cms_space), _perm_space(perm_space),
4910 _n_workers(n_workers),
4911 _workers(workers),
4912 _task_queues(task_queues),
4913 _term(workers->total_workers(), task_queues) { }
4914
4915 OopTaskQueueSet* task_queues() { return _task_queues; }
4916
4917 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4918
4919 ParallelTaskTerminator* terminator() { return &_term; }
4920
4921 void work(int i);
4922
4923 private:
4924 // Work method in support of parallel rescan ... of young gen spaces
4925 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
4926 ContiguousSpace* space,
4927 HeapWord** chunk_array, size_t chunk_top);
4928
4929 // ... of dirty cards in old space
4930 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4931 Par_MarkRefsIntoAndScanClosure* cl);
4932
4933 // ... work stealing for the above
4934 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
4935 };
4936
4937 void CMSParRemarkTask::work(int i) {
4938 elapsedTimer _timer;
4939 ResourceMark rm;
4940 HandleMark hm;
4941
4942 // ---------- rescan from roots --------------
4943 _timer.start();
4944 GenCollectedHeap* gch = GenCollectedHeap::heap();
4945 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4946 _collector->_span, _collector->ref_processor(),
4947 &(_collector->_markBitMap),
4948 work_queue(i), &(_collector->_revisitStack));
4949
4950 // Rescan young gen roots first since these are likely
4951 // coarsely partitioned and may, on that account, constitute
4952 // the critical path; thus, it's best to start off that
4953 // work first.
4954 // ---------- young gen roots --------------
4955 {
4956 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
4957 EdenSpace* eden_space = dng->eden();
4958 ContiguousSpace* from_space = dng->from();
4959 ContiguousSpace* to_space = dng->to();
4960
4961 HeapWord** eca = _collector->_eden_chunk_array;
4962 size_t ect = _collector->_eden_chunk_index;
4963 HeapWord** sca = _collector->_survivor_chunk_array;
4964 size_t sct = _collector->_survivor_chunk_index;
4965
4966 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4967 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4968
4969 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0);
4970 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct);
4971 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect);
4972
4973 _timer.stop();
4974 if (PrintCMSStatistics != 0) {
4975 gclog_or_tty->print_cr(
4976 "Finished young gen rescan work in %dth thread: %3.3f sec",
4977 i, _timer.seconds());
4978 }
4979 }
4980
4981 // ---------- remaining roots --------------
4982 _timer.reset();
4983 _timer.start();
4984 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
4985 false, // yg was scanned above
4986 true, // collecting perm gen
4987 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4988 NULL, &par_mrias_cl);
4989 _timer.stop();
4990 if (PrintCMSStatistics != 0) {
4991 gclog_or_tty->print_cr(
4992 "Finished remaining root rescan work in %dth thread: %3.3f sec",
4993 i, _timer.seconds());
4994 }
4995
4996 // ---------- rescan dirty cards ------------
4997 _timer.reset();
4998 _timer.start();
4999
5000 // Do the rescan tasks for each of the two spaces
5001 // (cms_space and perm_space) in turn.
5002 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
5003 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
5004 _timer.stop();
5005 if (PrintCMSStatistics != 0) {
5006 gclog_or_tty->print_cr(
5007 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5008 i, _timer.seconds());
5009 }
5010
5011 // ---------- steal work from other threads ...
5012 // ---------- ... and drain overflow list.
5013 _timer.reset();
5014 _timer.start();
5015 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i));
5016 _timer.stop();
5017 if (PrintCMSStatistics != 0) {
5018 gclog_or_tty->print_cr(
5019 "Finished work stealing in %dth thread: %3.3f sec",
5020 i, _timer.seconds());
5021 }
5022 }
5023
5024 void
5025 CMSParRemarkTask::do_young_space_rescan(int i,
5026 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5027 HeapWord** chunk_array, size_t chunk_top) {
5028 // Until all tasks completed:
5029 // . claim an unclaimed task
5030 // . compute region boundaries corresponding to task claimed
5031 // using chunk_array
5032 // . par_oop_iterate(cl) over that region
5033
5034 ResourceMark rm;
5035 HandleMark hm;
5036
5037 SequentialSubTasksDone* pst = space->par_seq_tasks();
5038 assert(pst->valid(), "Uninitialized use?");
5039
5040 int nth_task = 0;
5041 int n_tasks = pst->n_tasks();
5042
5043 HeapWord *start, *end;
5044 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5045 // We claimed task # nth_task; compute its boundaries.
5046 if (chunk_top == 0) { // no samples were taken
5047 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5048 start = space->bottom();
5049 end = space->top();
5050 } else if (nth_task == 0) {
5051 start = space->bottom();
5052 end = chunk_array[nth_task];
5053 } else if (nth_task < (jint)chunk_top) {
5054 assert(nth_task >= 1, "Control point invariant");
5055 start = chunk_array[nth_task - 1];
5056 end = chunk_array[nth_task];
5057 } else {
5058 assert(nth_task == (jint)chunk_top, "Control point invariant");
5059 start = chunk_array[chunk_top - 1];
5060 end = space->top();
5061 }
5062 MemRegion mr(start, end);
5063 // Verify that mr is in space
5064 assert(mr.is_empty() || space->used_region().contains(mr),
5065 "Should be in space");
5066 // Verify that "start" is an object boundary
5067 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5068 "Should be an oop");
5069 space->par_oop_iterate(mr, cl);
5070 }
5071 pst->all_tasks_completed();
5072 }
5073
5074 void
5075 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5076 CompactibleFreeListSpace* sp, int i,
5077 Par_MarkRefsIntoAndScanClosure* cl) {
5078 // Until all tasks completed:
5079 // . claim an unclaimed task
5080 // . compute region boundaries corresponding to task claimed
5081 // . transfer dirty bits ct->mut for that region
5082 // . apply rescanclosure to dirty mut bits for that region
5083
5084 ResourceMark rm;
5085 HandleMark hm;
5086
5087 OopTaskQueue* work_q = work_queue(i);
5088 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5089 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5090 // CAUTION: This closure has state that persists across calls to
5091 // the work method dirty_range_iterate_clear() in that it has
5092 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5093 // use of that state in the imbedded UpwardsObjectClosure instance
5094 // assumes that the cards are always iterated (even if in parallel
5095 // by several threads) in monotonically increasing order per each
5096 // thread. This is true of the implementation below which picks
5097 // card ranges (chunks) in monotonically increasing order globally
5098 // and, a-fortiori, in monotonically increasing order per thread
5099 // (the latter order being a subsequence of the former).
5100 // If the work code below is ever reorganized into a more chaotic
5101 // work-partitioning form than the current "sequential tasks"
5102 // paradigm, the use of that persistent state will have to be
5103 // revisited and modified appropriately. See also related
5104 // bug 4756801 work on which should examine this code to make
5105 // sure that the changes there do not run counter to the
5106 // assumptions made here and necessary for correctness and
5107 // efficiency. Note also that this code might yield inefficient
5108 // behaviour in the case of very large objects that span one or
5109 // more work chunks. Such objects would potentially be scanned
5110 // several times redundantly. Work on 4756801 should try and
5111 // address that performance anomaly if at all possible. XXX
5112 MemRegion full_span = _collector->_span;
5113 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5114 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5115 MarkFromDirtyCardsClosure
5116 greyRescanClosure(_collector, full_span, // entire span of interest
5117 sp, bm, work_q, rs, cl);
5118
5119 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5120 assert(pst->valid(), "Uninitialized use?");
5121 int nth_task = 0;
5122 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5123 MemRegion span = sp->used_region();
5124 HeapWord* start_addr = span.start();
5125 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5126 alignment);
5127 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5128 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5129 start_addr, "Check alignment");
5130 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5131 chunk_size, "Check alignment");
5132
5133 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5134 // Having claimed the nth_task, compute corresponding mem-region,
5135 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5136 // The alignment restriction ensures that we do not need any
5137 // synchronization with other gang-workers while setting or
5138 // clearing bits in thus chunk of the MUT.
5139 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5140 start_addr + (nth_task+1)*chunk_size);
5141 // The last chunk's end might be way beyond end of the
5142 // used region. In that case pull back appropriately.
5143 if (this_span.end() > end_addr) {
5144 this_span.set_end(end_addr);
5145 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5146 }
5147 // Iterate over the dirty cards covering this chunk, marking them
5148 // precleaned, and setting the corresponding bits in the mod union
5149 // table. Since we have been careful to partition at Card and MUT-word
5150 // boundaries no synchronization is needed between parallel threads.
5151 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5152 &modUnionClosure);
5153
5154 // Having transferred these marks into the modUnionTable,
5155 // rescan the marked objects on the dirty cards in the modUnionTable.
5156 // Even if this is at a synchronous collection, the initial marking
5157 // may have been done during an asynchronous collection so there
5158 // may be dirty bits in the mod-union table.
5159 _collector->_modUnionTable.dirty_range_iterate_clear(
5160 this_span, &greyRescanClosure);
5161 _collector->_modUnionTable.verifyNoOneBitsInRange(
5162 this_span.start(),
5163 this_span.end());
5164 }
5165 pst->all_tasks_completed(); // declare that i am done
5166 }
5167
5168 // . see if we can share work_queues with ParNew? XXX
5169 void
5170 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5171 int* seed) {
5172 OopTaskQueue* work_q = work_queue(i);
5173 NOT_PRODUCT(int num_steals = 0;)
5174 oop obj_to_scan;
5175 CMSBitMap* bm = &(_collector->_markBitMap);
5176 size_t num_from_overflow_list =
5177 MIN2((size_t)work_q->max_elems()/4,
5178 (size_t)ParGCDesiredObjsFromOverflowList);
5179
5180 while (true) {
5181 // Completely finish any left over work from (an) earlier round(s)
5182 cl->trim_queue(0);
5183 // Now check if there's any work in the overflow list
5184 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5185 work_q)) {
5186 // found something in global overflow list;
5187 // not yet ready to go stealing work from others.
5188 // We'd like to assert(work_q->size() != 0, ...)
5189 // because we just took work from the overflow list,
5190 // but of course we can't since all of that could have
5191 // been already stolen from us.
5192 // "He giveth and He taketh away."
5193 continue;
5194 }
5195 // Verify that we have no work before we resort to stealing
5196 assert(work_q->size() == 0, "Have work, shouldn't steal");
5197 // Try to steal from other queues that have work
5198 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5199 NOT_PRODUCT(num_steals++;)
5200 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5201 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5202 // Do scanning work
5203 obj_to_scan->oop_iterate(cl);
5204 // Loop around, finish this work, and try to steal some more
5205 } else if (terminator()->offer_termination()) {
5206 break; // nirvana from the infinite cycle
5207 }
5208 }
5209 NOT_PRODUCT(
5210 if (PrintCMSStatistics != 0) {
5211 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5212 }
5213 )
5214 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5215 "Else our work is not yet done");
5216 }
5217
5218 // Return a thread-local PLAB recording array, as appropriate.
5219 void* CMSCollector::get_data_recorder(int thr_num) {
5220 if (_survivor_plab_array != NULL &&
5221 (CMSPLABRecordAlways ||
5222 (_collectorState > Marking && _collectorState < FinalMarking))) {
5223 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5224 ChunkArray* ca = &_survivor_plab_array[thr_num];
5225 ca->reset(); // clear it so that fresh data is recorded
5226 return (void*) ca;
5227 } else {
5228 return NULL;
5229 }
5230 }
5231
5232 // Reset all the thread-local PLAB recording arrays
5233 void CMSCollector::reset_survivor_plab_arrays() {
5234 for (uint i = 0; i < ParallelGCThreads; i++) {
5235 _survivor_plab_array[i].reset();
5236 }
5237 }
5238
5239 // Merge the per-thread plab arrays into the global survivor chunk
5240 // array which will provide the partitioning of the survivor space
5241 // for CMS rescan.
5242 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
5243 assert(_survivor_plab_array != NULL, "Error");
5244 assert(_survivor_chunk_array != NULL, "Error");
5245 assert(_collectorState == FinalMarking, "Error");
5246 for (uint j = 0; j < ParallelGCThreads; j++) {
5247 _cursor[j] = 0;
5248 }
5249 HeapWord* top = surv->top();
5250 size_t i;
5251 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5252 HeapWord* min_val = top; // Higher than any PLAB address
5253 uint min_tid = 0; // position of min_val this round
5254 for (uint j = 0; j < ParallelGCThreads; j++) {
5255 ChunkArray* cur_sca = &_survivor_plab_array[j];
5256 if (_cursor[j] == cur_sca->end()) {
5257 continue;
5258 }
5259 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5260 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5261 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5262 if (cur_val < min_val) {
5263 min_tid = j;
5264 min_val = cur_val;
5265 } else {
5266 assert(cur_val < top, "All recorded addresses should be less");
5267 }
5268 }
5269 // At this point min_val and min_tid are respectively
5270 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5271 // and the thread (j) that witnesses that address.
5272 // We record this address in the _survivor_chunk_array[i]
5273 // and increment _cursor[min_tid] prior to the next round i.
5274 if (min_val == top) {
5275 break;
5276 }
5277 _survivor_chunk_array[i] = min_val;
5278 _cursor[min_tid]++;
5279 }
5280 // We are all done; record the size of the _survivor_chunk_array
5281 _survivor_chunk_index = i; // exclusive: [0, i)
5282 if (PrintCMSStatistics > 0) {
5283 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5284 }
5285 // Verify that we used up all the recorded entries
5286 #ifdef ASSERT
5287 size_t total = 0;
5288 for (uint j = 0; j < ParallelGCThreads; j++) {
5289 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5290 total += _cursor[j];
5291 }
5292 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5293 // Check that the merged array is in sorted order
5294 if (total > 0) {
5295 for (size_t i = 0; i < total - 1; i++) {
5296 if (PrintCMSStatistics > 0) {
5297 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5298 i, _survivor_chunk_array[i]);
5299 }
5300 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5301 "Not sorted");
5302 }
5303 }
5304 #endif // ASSERT
5305 }
5306
5307 // Set up the space's par_seq_tasks structure for work claiming
5308 // for parallel rescan of young gen.
5309 // See ParRescanTask where this is currently used.
5310 void
5311 CMSCollector::
5312 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5313 assert(n_threads > 0, "Unexpected n_threads argument");
5314 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5315
5316 // Eden space
5317 {
5318 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5319 assert(!pst->valid(), "Clobbering existing data?");
5320 // Each valid entry in [0, _eden_chunk_index) represents a task.
5321 size_t n_tasks = _eden_chunk_index + 1;
5322 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5323 pst->set_par_threads(n_threads);
5324 pst->set_n_tasks((int)n_tasks);
5325 }
5326
5327 // Merge the survivor plab arrays into _survivor_chunk_array
5328 if (_survivor_plab_array != NULL) {
5329 merge_survivor_plab_arrays(dng->from());
5330 } else {
5331 assert(_survivor_chunk_index == 0, "Error");
5332 }
5333
5334 // To space
5335 {
5336 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5337 assert(!pst->valid(), "Clobbering existing data?");
5338 pst->set_par_threads(n_threads);
5339 pst->set_n_tasks(1);
5340 assert(pst->valid(), "Error");
5341 }
5342
5343 // From space
5344 {
5345 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5346 assert(!pst->valid(), "Clobbering existing data?");
5347 size_t n_tasks = _survivor_chunk_index + 1;
5348 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5349 pst->set_par_threads(n_threads);
5350 pst->set_n_tasks((int)n_tasks);
5351 assert(pst->valid(), "Error");
5352 }
5353 }
5354
5355 // Parallel version of remark
5356 void CMSCollector::do_remark_parallel() {
5357 GenCollectedHeap* gch = GenCollectedHeap::heap();
5358 WorkGang* workers = gch->workers();
5359 assert(workers != NULL, "Need parallel worker threads.");
5360 int n_workers = workers->total_workers();
5361 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5362 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5363
5364 CMSParRemarkTask tsk(this,
5365 cms_space, perm_space,
5366 n_workers, workers, task_queues());
5367
5368 // Set up for parallel process_strong_roots work.
5369 gch->set_par_threads(n_workers);
5370 gch->change_strong_roots_parity();
5371 // We won't be iterating over the cards in the card table updating
5372 // the younger_gen cards, so we shouldn't call the following else
5373 // the verification code as well as subsequent younger_refs_iterate
5374 // code would get confused. XXX
5375 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5376
5377 // The young gen rescan work will not be done as part of
5378 // process_strong_roots (which currently doesn't knw how to
5379 // parallelize such a scan), but rather will be broken up into
5380 // a set of parallel tasks (via the sampling that the [abortable]
5381 // preclean phase did of EdenSpace, plus the [two] tasks of
5382 // scanning the [two] survivor spaces. Further fine-grain
5383 // parallelization of the scanning of the survivor spaces
5384 // themselves, and of precleaning of the younger gen itself
5385 // is deferred to the future.
5386 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5387
5388 // The dirty card rescan work is broken up into a "sequence"
5389 // of parallel tasks (per constituent space) that are dynamically
5390 // claimed by the parallel threads.
5391 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5392 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5393
5394 // It turns out that even when we're using 1 thread, doing the work in a
5395 // separate thread causes wide variance in run times. We can't help this
5396 // in the multi-threaded case, but we special-case n=1 here to get
5397 // repeatable measurements of the 1-thread overhead of the parallel code.
5398 if (n_workers > 1) {
5399 // Make refs discovery MT-safe
5400 ReferenceProcessorMTMutator mt(ref_processor(), true);
5401 workers->run_task(&tsk);
5402 } else {
5403 tsk.work(0);
5404 }
5405 gch->set_par_threads(0); // 0 ==> non-parallel.
5406 // restore, single-threaded for now, any preserved marks
5407 // as a result of work_q overflow
5408 restore_preserved_marks_if_any();
5409 }
5410
5411 // Non-parallel version of remark
5412 void CMSCollector::do_remark_non_parallel() {
5413 ResourceMark rm;
5414 HandleMark hm;
5415 GenCollectedHeap* gch = GenCollectedHeap::heap();
5416 MarkRefsIntoAndScanClosure
5417 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5418 &_markStack, &_revisitStack, this,
5419 false /* should_yield */, false /* not precleaning */);
5420 MarkFromDirtyCardsClosure
5421 markFromDirtyCardsClosure(this, _span,
5422 NULL, // space is set further below
5423 &_markBitMap, &_markStack, &_revisitStack,
5424 &mrias_cl);
5425 {
5426 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty);
5427 // Iterate over the dirty cards, marking them precleaned, and
5428 // setting the corresponding bits in the mod union table.
5429 {
5430 ModUnionClosure modUnionClosure(&_modUnionTable);
5431 _ct->ct_bs()->dirty_card_iterate(
5432 _cmsGen->used_region(),
5433 &modUnionClosure);
5434 _ct->ct_bs()->dirty_card_iterate(
5435 _permGen->used_region(),
5436 &modUnionClosure);
5437 }
5438 // Having transferred these marks into the modUnionTable, we just need
5439 // to rescan the marked objects on the dirty cards in the modUnionTable.
5440 // The initial marking may have been done during an asynchronous
5441 // collection so there may be dirty bits in the mod-union table.
5442 const int alignment =
5443 CardTableModRefBS::card_size * BitsPerWord;
5444 {
5445 // ... First handle dirty cards in CMS gen
5446 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5447 MemRegion ur = _cmsGen->used_region();
5448 HeapWord* lb = ur.start();
5449 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5450 MemRegion cms_span(lb, ub);
5451 _modUnionTable.dirty_range_iterate_clear(cms_span,
5452 &markFromDirtyCardsClosure);
5453 verify_work_stacks_empty();
5454 if (PrintCMSStatistics != 0) {
5455 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5456 markFromDirtyCardsClosure.num_dirty_cards());
5457 }
5458 }
5459 {
5460 // .. and then repeat for dirty cards in perm gen
5461 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5462 MemRegion ur = _permGen->used_region();
5463 HeapWord* lb = ur.start();
5464 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5465 MemRegion perm_span(lb, ub);
5466 _modUnionTable.dirty_range_iterate_clear(perm_span,
5467 &markFromDirtyCardsClosure);
5468 verify_work_stacks_empty();
5469 if (PrintCMSStatistics != 0) {
5470 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5471 markFromDirtyCardsClosure.num_dirty_cards());
5472 }
5473 }
5474 }
5475 if (VerifyDuringGC &&
5476 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5477 HandleMark hm; // Discard invalid handles created during verification
5478 Universe::verify(true);
5479 }
5480 {
5481 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty);
5482
5483 verify_work_stacks_empty();
5484
5485 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5486 gch->gen_process_strong_roots(_cmsGen->level(),
5487 true, // younger gens as roots
5488 true, // collecting perm gen
5489 SharedHeap::ScanningOption(roots_scanning_options()),
5490 NULL, &mrias_cl);
5491 }
5492 verify_work_stacks_empty();
5493 // Restore evacuated mark words, if any, used for overflow list links
5494 if (!CMSOverflowEarlyRestoration) {
5495 restore_preserved_marks_if_any();
5496 }
5497 verify_overflow_empty();
5498 }
5499
5500 ////////////////////////////////////////////////////////
5501 // Parallel Reference Processing Task Proxy Class
5502 ////////////////////////////////////////////////////////
5503 class CMSRefProcTaskProxy: public AbstractGangTask {
5504 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5505 CMSCollector* _collector;
5506 CMSBitMap* _mark_bit_map;
5507 const MemRegion _span;
5508 OopTaskQueueSet* _task_queues;
5509 ParallelTaskTerminator _term;
5510 ProcessTask& _task;
5511
5512 public:
5513 CMSRefProcTaskProxy(ProcessTask& task,
5514 CMSCollector* collector,
5515 const MemRegion& span,
5516 CMSBitMap* mark_bit_map,
5517 int total_workers,
5518 OopTaskQueueSet* task_queues):
5519 AbstractGangTask("Process referents by policy in parallel"),
5520 _task(task),
5521 _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
5522 _task_queues(task_queues),
5523 _term(total_workers, task_queues)
5524 {
5525 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5526 "Inconsistency in _span");
5527 }
5528
5529 OopTaskQueueSet* task_queues() { return _task_queues; }
5530
5531 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5532
5533 ParallelTaskTerminator* terminator() { return &_term; }
5534
5535 void do_work_steal(int i,
5536 CMSParDrainMarkingStackClosure* drain,
5537 CMSParKeepAliveClosure* keep_alive,
5538 int* seed);
5539
5540 virtual void work(int i);
5541 };
5542
5543 void CMSRefProcTaskProxy::work(int i) {
5544 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5545 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5546 _mark_bit_map, work_queue(i));
5547 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5548 _mark_bit_map, work_queue(i));
5549 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5550 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack);
5551 if (_task.marks_oops_alive()) {
5552 do_work_steal(i, &par_drain_stack, &par_keep_alive,
5553 _collector->hash_seed(i));
5554 }
5555 assert(work_queue(i)->size() == 0, "work_queue should be empty");
5556 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5557 }
5558
5559 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5560 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5561 EnqueueTask& _task;
5562
5563 public:
5564 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5565 : AbstractGangTask("Enqueue reference objects in parallel"),
5566 _task(task)
5567 { }
5568
5569 virtual void work(int i)
5570 {
5571 _task.work(i);
5572 }
5573 };
5574
5575 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5576 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5577 _collector(collector),
5578 _span(span),
5579 _bit_map(bit_map),
5580 _work_queue(work_queue),
5581 _mark_and_push(collector, span, bit_map, work_queue),
5582 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5583 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5584 { }
5585
5586 // . see if we can share work_queues with ParNew? XXX
5587 void CMSRefProcTaskProxy::do_work_steal(int i,
5588 CMSParDrainMarkingStackClosure* drain,
5589 CMSParKeepAliveClosure* keep_alive,
5590 int* seed) {
5591 OopTaskQueue* work_q = work_queue(i);
5592 NOT_PRODUCT(int num_steals = 0;)
5593 oop obj_to_scan;
5594 size_t num_from_overflow_list =
5595 MIN2((size_t)work_q->max_elems()/4,
5596 (size_t)ParGCDesiredObjsFromOverflowList);
5597
5598 while (true) {
5599 // Completely finish any left over work from (an) earlier round(s)
5600 drain->trim_queue(0);
5601 // Now check if there's any work in the overflow list
5602 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5603 work_q)) {
5604 // Found something in global overflow list;
5605 // not yet ready to go stealing work from others.
5606 // We'd like to assert(work_q->size() != 0, ...)
5607 // because we just took work from the overflow list,
5608 // but of course we can't, since all of that might have
5609 // been already stolen from us.
5610 continue;
5611 }
5612 // Verify that we have no work before we resort to stealing
5613 assert(work_q->size() == 0, "Have work, shouldn't steal");
5614 // Try to steal from other queues that have work
5615 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5616 NOT_PRODUCT(num_steals++;)
5617 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5618 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5619 // Do scanning work
5620 obj_to_scan->oop_iterate(keep_alive);
5621 // Loop around, finish this work, and try to steal some more
5622 } else if (terminator()->offer_termination()) {
5623 break; // nirvana from the infinite cycle
5624 }
5625 }
5626 NOT_PRODUCT(
5627 if (PrintCMSStatistics != 0) {
5628 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5629 }
5630 )
5631 }
5632
5633 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5634 {
5635 GenCollectedHeap* gch = GenCollectedHeap::heap();
5636 WorkGang* workers = gch->workers();
5637 assert(workers != NULL, "Need parallel worker threads.");
5638 int n_workers = workers->total_workers();
5639 CMSRefProcTaskProxy rp_task(task, &_collector,
5640 _collector.ref_processor()->span(),
5641 _collector.markBitMap(),
5642 n_workers, _collector.task_queues());
5643 workers->run_task(&rp_task);
5644 }
5645
5646 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5647 {
5648
5649 GenCollectedHeap* gch = GenCollectedHeap::heap();
5650 WorkGang* workers = gch->workers();
5651 assert(workers != NULL, "Need parallel worker threads.");
5652 CMSRefEnqueueTaskProxy enq_task(task);
5653 workers->run_task(&enq_task);
5654 }
5655
5656 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5657
5658 ResourceMark rm;
5659 HandleMark hm;
5660 ReferencePolicy* soft_ref_policy;
5661
5662 assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete");
5663 // Process weak references.
5664 if (clear_all_soft_refs) {
5665 soft_ref_policy = new AlwaysClearPolicy();
5666 } else {
5667 #ifdef COMPILER2
5668 soft_ref_policy = new LRUMaxHeapPolicy();
5669 #else
5670 soft_ref_policy = new LRUCurrentHeapPolicy();
5671 #endif // COMPILER2
5672 }
5673 verify_work_stacks_empty();
5674
5675 ReferenceProcessor* rp = ref_processor();
5676 assert(rp->span().equals(_span), "Spans should be equal");
5677 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5678 &_markStack);
5679 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5680 _span, &_markBitMap, &_markStack,
5681 &cmsKeepAliveClosure);
5682 {
5683 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
5684 if (rp->processing_is_mt()) {
5685 CMSRefProcTaskExecutor task_executor(*this);
5686 rp->process_discovered_references(soft_ref_policy,
5687 &_is_alive_closure,
5688 &cmsKeepAliveClosure,
5689 &cmsDrainMarkingStackClosure,
5690 &task_executor);
5691 } else {
5692 rp->process_discovered_references(soft_ref_policy,
5693 &_is_alive_closure,
5694 &cmsKeepAliveClosure,
5695 &cmsDrainMarkingStackClosure,
5696 NULL);
5697 }
5698 verify_work_stacks_empty();
5699 }
5700
5701 if (should_unload_classes()) {
5702 {
5703 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty);
5704
5705 // Follow SystemDictionary roots and unload classes
5706 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5707
5708 // Follow CodeCache roots and unload any methods marked for unloading
5709 CodeCache::do_unloading(&_is_alive_closure,
5710 &cmsKeepAliveClosure,
5711 purged_class);
5712
5713 cmsDrainMarkingStackClosure.do_void();
5714 verify_work_stacks_empty();
5715
5716 // Update subklass/sibling/implementor links in KlassKlass descendants
5717 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5718 oop k;
5719 while ((k = _revisitStack.pop()) != NULL) {
5720 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5721 &_is_alive_closure,
5722 &cmsKeepAliveClosure);
5723 }
5724 assert(!ClassUnloading ||
5725 (_markStack.isEmpty() && overflow_list_is_empty()),
5726 "Should not have found new reachable objects");
5727 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5728 cmsDrainMarkingStackClosure.do_void();
5729 verify_work_stacks_empty();
5730 }
5731
5732 {
5733 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty);
5734 // Now clean up stale oops in SymbolTable and StringTable
5735 SymbolTable::unlink(&_is_alive_closure);
5736 StringTable::unlink(&_is_alive_closure);
5737 }
5738 }
5739
5740 verify_work_stacks_empty();
5741 // Restore any preserved marks as a result of mark stack or
5742 // work queue overflow
5743 restore_preserved_marks_if_any(); // done single-threaded for now
5744
5745 rp->set_enqueuing_is_done(true);
5746 if (rp->processing_is_mt()) {
5747 CMSRefProcTaskExecutor task_executor(*this);
5748 rp->enqueue_discovered_references(&task_executor);
5749 } else {
5750 rp->enqueue_discovered_references(NULL);
5751 }
5752 rp->verify_no_references_recorded();
5753 assert(!rp->discovery_enabled(), "should have been disabled");
5754
5755 // JVMTI object tagging is based on JNI weak refs. If any of these
5756 // refs were cleared then JVMTI needs to update its maps and
5757 // maybe post ObjectFrees to agents.
5758 JvmtiExport::cms_ref_processing_epilogue();
5759 }
5760
5761 #ifndef PRODUCT
5762 void CMSCollector::check_correct_thread_executing() {
5763 Thread* t = Thread::current();
5764 // Only the VM thread or the CMS thread should be here.
5765 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5766 "Unexpected thread type");
5767 // If this is the vm thread, the foreground process
5768 // should not be waiting. Note that _foregroundGCIsActive is
5769 // true while the foreground collector is waiting.
5770 if (_foregroundGCShouldWait) {
5771 // We cannot be the VM thread
5772 assert(t->is_ConcurrentGC_thread(),
5773 "Should be CMS thread");
5774 } else {
5775 // We can be the CMS thread only if we are in a stop-world
5776 // phase of CMS collection.
5777 if (t->is_ConcurrentGC_thread()) {
5778 assert(_collectorState == InitialMarking ||
5779 _collectorState == FinalMarking,
5780 "Should be a stop-world phase");
5781 // The CMS thread should be holding the CMS_token.
5782 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5783 "Potential interference with concurrently "
5784 "executing VM thread");
5785 }
5786 }
5787 }
5788 #endif
5789
5790 void CMSCollector::sweep(bool asynch) {
5791 assert(_collectorState == Sweeping, "just checking");
5792 check_correct_thread_executing();
5793 verify_work_stacks_empty();
5794 verify_overflow_empty();
5795 incrementSweepCount();
5796 _sweep_timer.stop();
5797 _sweep_estimate.sample(_sweep_timer.seconds());
5798 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
5799
5800 // PermGen verification support: If perm gen sweeping is disabled in
5801 // this cycle, we preserve the perm gen object "deadness" information
5802 // in the perm_gen_verify_bit_map. In order to do that we traverse
5803 // all blocks in perm gen and mark all dead objects.
5804 if (verifying() && !should_unload_classes()) {
5805 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
5806 "Should have already been allocated");
5807 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
5808 markBitMap(), perm_gen_verify_bit_map());
5809 if (asynch) {
5810 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5811 bitMapLock());
5812 _permGen->cmsSpace()->blk_iterate(&mdo);
5813 } else {
5814 // In the case of synchronous sweep, we already have
5815 // the requisite locks/tokens.
5816 _permGen->cmsSpace()->blk_iterate(&mdo);
5817 }
5818 }
5819
5820 if (asynch) {
5821 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
5822 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
5823 // First sweep the old gen then the perm gen
5824 {
5825 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5826 bitMapLock());
5827 sweepWork(_cmsGen, asynch);
5828 }
5829
5830 // Now repeat for perm gen
5831 if (should_unload_classes()) {
5832 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
5833 bitMapLock());
5834 sweepWork(_permGen, asynch);
5835 }
5836
5837 // Update Universe::_heap_*_at_gc figures.
5838 // We need all the free list locks to make the abstract state
5839 // transition from Sweeping to Resetting. See detailed note
5840 // further below.
5841 {
5842 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5843 _permGen->freelistLock());
5844 // Update heap occupancy information which is used as
5845 // input to soft ref clearing policy at the next gc.
5846 Universe::update_heap_info_at_gc();
5847 _collectorState = Resizing;
5848 }
5849 } else {
5850 // already have needed locks
5851 sweepWork(_cmsGen, asynch);
5852
5853 if (should_unload_classes()) {
5854 sweepWork(_permGen, asynch);
5855 }
5856 // Update heap occupancy information which is used as
5857 // input to soft ref clearing policy at the next gc.
5858 Universe::update_heap_info_at_gc();
5859 _collectorState = Resizing;
5860 }
5861 verify_work_stacks_empty();
5862 verify_overflow_empty();
5863
5864 _sweep_timer.reset();
5865 _sweep_timer.start();
5866
5867 update_time_of_last_gc(os::javaTimeMillis());
5868
5869 // NOTE on abstract state transitions:
5870 // Mutators allocate-live and/or mark the mod-union table dirty
5871 // based on the state of the collection. The former is done in
5872 // the interval [Marking, Sweeping] and the latter in the interval
5873 // [Marking, Sweeping). Thus the transitions into the Marking state
5874 // and out of the Sweeping state must be synchronously visible
5875 // globally to the mutators.
5876 // The transition into the Marking state happens with the world
5877 // stopped so the mutators will globally see it. Sweeping is
5878 // done asynchronously by the background collector so the transition
5879 // from the Sweeping state to the Resizing state must be done
5880 // under the freelistLock (as is the check for whether to
5881 // allocate-live and whether to dirty the mod-union table).
5882 assert(_collectorState == Resizing, "Change of collector state to"
5883 " Resizing must be done under the freelistLocks (plural)");
5884
5885 // Now that sweeping has been completed, if the GCH's
5886 // incremental_collection_will_fail flag is set, clear it,
5887 // thus inviting a younger gen collection to promote into
5888 // this generation. If such a promotion may still fail,
5889 // the flag will be set again when a young collection is
5890 // attempted.
5891 // I think the incremental_collection_will_fail flag's use
5892 // is specific to a 2 generation collection policy, so i'll
5893 // assert that that's the configuration we are operating within.
5894 // The use of the flag can and should be generalized appropriately
5895 // in the future to deal with a general n-generation system.
5896
5897 GenCollectedHeap* gch = GenCollectedHeap::heap();
5898 assert(gch->collector_policy()->is_two_generation_policy(),
5899 "Resetting of incremental_collection_will_fail flag"
5900 " may be incorrect otherwise");
5901 gch->clear_incremental_collection_will_fail();
5902 gch->update_full_collections_completed(_collection_count_start);
5903 }
5904
5905 // FIX ME!!! Looks like this belongs in CFLSpace, with
5906 // CMSGen merely delegating to it.
5907 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5908 double nearLargestPercent = 0.999;
5909 HeapWord* minAddr = _cmsSpace->bottom();
5910 HeapWord* largestAddr =
5911 (HeapWord*) _cmsSpace->dictionary()->findLargestDict();
5912 if (largestAddr == 0) {
5913 // The dictionary appears to be empty. In this case
5914 // try to coalesce at the end of the heap.
5915 largestAddr = _cmsSpace->end();
5916 }
5917 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5918 size_t nearLargestOffset =
5919 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5920 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5921 }
5922
5923 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5924 return addr >= _cmsSpace->nearLargestChunk();
5925 }
5926
5927 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5928 return _cmsSpace->find_chunk_at_end();
5929 }
5930
5931 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
5932 bool full) {
5933 // The next lower level has been collected. Gather any statistics
5934 // that are of interest at this point.
5935 if (!full && (current_level + 1) == level()) {
5936 // Gather statistics on the young generation collection.
5937 collector()->stats().record_gc0_end(used());
5938 }
5939 }
5940
5941 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
5942 GenCollectedHeap* gch = GenCollectedHeap::heap();
5943 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
5944 "Wrong type of heap");
5945 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
5946 gch->gen_policy()->size_policy();
5947 assert(sp->is_gc_cms_adaptive_size_policy(),
5948 "Wrong type of size policy");
5949 return sp;
5950 }
5951
5952 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
5953 if (PrintGCDetails && Verbose) {
5954 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
5955 }
5956 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
5957 _debug_collection_type =
5958 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
5959 if (PrintGCDetails && Verbose) {
5960 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
5961 }
5962 }
5963
5964 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
5965 bool asynch) {
5966 // We iterate over the space(s) underlying this generation,
5967 // checking the mark bit map to see if the bits corresponding
5968 // to specific blocks are marked or not. Blocks that are
5969 // marked are live and are not swept up. All remaining blocks
5970 // are swept up, with coalescing on-the-fly as we sweep up
5971 // contiguous free and/or garbage blocks:
5972 // We need to ensure that the sweeper synchronizes with allocators
5973 // and stop-the-world collectors. In particular, the following
5974 // locks are used:
5975 // . CMS token: if this is held, a stop the world collection cannot occur
5976 // . freelistLock: if this is held no allocation can occur from this
5977 // generation by another thread
5978 // . bitMapLock: if this is held, no other thread can access or update
5979 //
5980
5981 // Note that we need to hold the freelistLock if we use
5982 // block iterate below; else the iterator might go awry if
5983 // a mutator (or promotion) causes block contents to change
5984 // (for instance if the allocator divvies up a block).
5985 // If we hold the free list lock, for all practical purposes
5986 // young generation GC's can't occur (they'll usually need to
5987 // promote), so we might as well prevent all young generation
5988 // GC's while we do a sweeping step. For the same reason, we might
5989 // as well take the bit map lock for the entire duration
5990
5991 // check that we hold the requisite locks
5992 assert(have_cms_token(), "Should hold cms token");
5993 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
5994 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
5995 "Should possess CMS token to sweep");
5996 assert_lock_strong(gen->freelistLock());
5997 assert_lock_strong(bitMapLock());
5998
5999 assert(!_sweep_timer.is_active(), "Was switched off in an outer context");
6000 gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()),
6001 _sweep_estimate.padded_average());
6002 gen->setNearLargestChunk();
6003
6004 {
6005 SweepClosure sweepClosure(this, gen, &_markBitMap,
6006 CMSYield && asynch);
6007 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6008 // We need to free-up/coalesce garbage/blocks from a
6009 // co-terminal free run. This is done in the SweepClosure
6010 // destructor; so, do not remove this scope, else the
6011 // end-of-sweep-census below will be off by a little bit.
6012 }
6013 gen->cmsSpace()->sweep_completed();
6014 gen->cmsSpace()->endSweepFLCensus(sweepCount());
6015 if (should_unload_classes()) { // unloaded classes this cycle,
6016 _concurrent_cycles_since_last_unload = 0; // ... reset count
6017 } else { // did not unload classes,
6018 _concurrent_cycles_since_last_unload++; // ... increment count
6019 }
6020 }
6021
6022 // Reset CMS data structures (for now just the marking bit map)
6023 // preparatory for the next cycle.
6024 void CMSCollector::reset(bool asynch) {
6025 GenCollectedHeap* gch = GenCollectedHeap::heap();
6026 CMSAdaptiveSizePolicy* sp = size_policy();
6027 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6028 if (asynch) {
6029 CMSTokenSyncWithLocks ts(true, bitMapLock());
6030
6031 // If the state is not "Resetting", the foreground thread
6032 // has done a collection and the resetting.
6033 if (_collectorState != Resetting) {
6034 assert(_collectorState == Idling, "The state should only change"
6035 " because the foreground collector has finished the collection");
6036 return;
6037 }
6038
6039 // Clear the mark bitmap (no grey objects to start with)
6040 // for the next cycle.
6041 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6042 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6043
6044 HeapWord* curAddr = _markBitMap.startWord();
6045 while (curAddr < _markBitMap.endWord()) {
6046 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6047 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6048 _markBitMap.clear_large_range(chunk);
6049 if (ConcurrentMarkSweepThread::should_yield() &&
6050 !foregroundGCIsActive() &&
6051 CMSYield) {
6052 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6053 "CMS thread should hold CMS token");
6054 assert_lock_strong(bitMapLock());
6055 bitMapLock()->unlock();
6056 ConcurrentMarkSweepThread::desynchronize(true);
6057 ConcurrentMarkSweepThread::acknowledge_yield_request();
6058 stopTimer();
6059 if (PrintCMSStatistics != 0) {
6060 incrementYields();
6061 }
6062 icms_wait();
6063
6064 // See the comment in coordinator_yield()
6065 for (unsigned i = 0; i < CMSYieldSleepCount &&
6066 ConcurrentMarkSweepThread::should_yield() &&
6067 !CMSCollector::foregroundGCIsActive(); ++i) {
6068 os::sleep(Thread::current(), 1, false);
6069 ConcurrentMarkSweepThread::acknowledge_yield_request();
6070 }
6071
6072 ConcurrentMarkSweepThread::synchronize(true);
6073 bitMapLock()->lock_without_safepoint_check();
6074 startTimer();
6075 }
6076 curAddr = chunk.end();
6077 }
6078 _collectorState = Idling;
6079 } else {
6080 // already have the lock
6081 assert(_collectorState == Resetting, "just checking");
6082 assert_lock_strong(bitMapLock());
6083 _markBitMap.clear_all();
6084 _collectorState = Idling;
6085 }
6086
6087 // Stop incremental mode after a cycle completes, so that any future cycles
6088 // are triggered by allocation.
6089 stop_icms();
6090
6091 NOT_PRODUCT(
6092 if (RotateCMSCollectionTypes) {
6093 _cmsGen->rotate_debug_collection_type();
6094 }
6095 )
6096 }
6097
6098 void CMSCollector::do_CMS_operation(CMS_op_type op) {
6099 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6100 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6101 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty);
6102 TraceCollectorStats tcs(counters());
6103
6104 switch (op) {
6105 case CMS_op_checkpointRootsInitial: {
6106 checkpointRootsInitial(true); // asynch
6107 if (PrintGC) {
6108 _cmsGen->printOccupancy("initial-mark");
6109 }
6110 break;
6111 }
6112 case CMS_op_checkpointRootsFinal: {
6113 checkpointRootsFinal(true, // asynch
6114 false, // !clear_all_soft_refs
6115 false); // !init_mark_was_synchronous
6116 if (PrintGC) {
6117 _cmsGen->printOccupancy("remark");
6118 }
6119 break;
6120 }
6121 default:
6122 fatal("No such CMS_op");
6123 }
6124 }
6125
6126 #ifndef PRODUCT
6127 size_t const CMSCollector::skip_header_HeapWords() {
6128 return FreeChunk::header_size();
6129 }
6130
6131 // Try and collect here conditions that should hold when
6132 // CMS thread is exiting. The idea is that the foreground GC
6133 // thread should not be blocked if it wants to terminate
6134 // the CMS thread and yet continue to run the VM for a while
6135 // after that.
6136 void CMSCollector::verify_ok_to_terminate() const {
6137 assert(Thread::current()->is_ConcurrentGC_thread(),
6138 "should be called by CMS thread");
6139 assert(!_foregroundGCShouldWait, "should be false");
6140 // We could check here that all the various low-level locks
6141 // are not held by the CMS thread, but that is overkill; see
6142 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6143 // is checked.
6144 }
6145 #endif
6146
6147 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6148 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6149 "missing Printezis mark?");
6150 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6151 size_t size = pointer_delta(nextOneAddr + 1, addr);
6152 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6153 "alignment problem");
6154 assert(size >= 3, "Necessary for Printezis marks to work");
6155 return size;
6156 }
6157
6158 // A variant of the above (block_size_using_printezis_bits()) except
6159 // that we return 0 if the P-bits are not yet set.
6160 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6161 if (_markBitMap.isMarked(addr)) {
6162 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?");
6163 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6164 size_t size = pointer_delta(nextOneAddr + 1, addr);
6165 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6166 "alignment problem");
6167 assert(size >= 3, "Necessary for Printezis marks to work");
6168 return size;
6169 } else {
6170 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?");
6171 return 0;
6172 }
6173 }
6174
6175 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6176 size_t sz = 0;
6177 oop p = (oop)addr;
6178 if (p->klass_or_null() != NULL && p->is_parsable()) {
6179 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6180 } else {
6181 sz = block_size_using_printezis_bits(addr);
6182 }
6183 assert(sz > 0, "size must be nonzero");
6184 HeapWord* next_block = addr + sz;
6185 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6186 CardTableModRefBS::card_size);
6187 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6188 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6189 "must be different cards");
6190 return next_card;
6191 }
6192
6193
6194 // CMS Bit Map Wrapper /////////////////////////////////////////
6195
6196 // Construct a CMS bit map infrastructure, but don't create the
6197 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6198 // further below.
6199 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6200 _bm(NULL,0),
6201 _shifter(shifter),
6202 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6203 {
6204 _bmStartWord = 0;
6205 _bmWordSize = 0;
6206 }
6207
6208 bool CMSBitMap::allocate(MemRegion mr) {
6209 _bmStartWord = mr.start();
6210 _bmWordSize = mr.word_size();
6211 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6212 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6213 if (!brs.is_reserved()) {
6214 warning("CMS bit map allocation failure");
6215 return false;
6216 }
6217 // For now we'll just commit all of the bit map up fromt.
6218 // Later on we'll try to be more parsimonious with swap.
6219 if (!_virtual_space.initialize(brs, brs.size())) {
6220 warning("CMS bit map backing store failure");
6221 return false;
6222 }
6223 assert(_virtual_space.committed_size() == brs.size(),
6224 "didn't reserve backing store for all of CMS bit map?");
6225 _bm.set_map((uintptr_t*)_virtual_space.low());
6226 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6227 _bmWordSize, "inconsistency in bit map sizing");
6228 _bm.set_size(_bmWordSize >> _shifter);
6229
6230 // bm.clear(); // can we rely on getting zero'd memory? verify below
6231 assert(isAllClear(),
6232 "Expected zero'd memory from ReservedSpace constructor");
6233 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6234 "consistency check");
6235 return true;
6236 }
6237
6238 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6239 HeapWord *next_addr, *end_addr, *last_addr;
6240 assert_locked();
6241 assert(covers(mr), "out-of-range error");
6242 // XXX assert that start and end are appropriately aligned
6243 for (next_addr = mr.start(), end_addr = mr.end();
6244 next_addr < end_addr; next_addr = last_addr) {
6245 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6246 last_addr = dirty_region.end();
6247 if (!dirty_region.is_empty()) {
6248 cl->do_MemRegion(dirty_region);
6249 } else {
6250 assert(last_addr == end_addr, "program logic");
6251 return;
6252 }
6253 }
6254 }
6255
6256 #ifndef PRODUCT
6257 void CMSBitMap::assert_locked() const {
6258 CMSLockVerifier::assert_locked(lock());
6259 }
6260
6261 bool CMSBitMap::covers(MemRegion mr) const {
6262 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6263 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6264 "size inconsistency");
6265 return (mr.start() >= _bmStartWord) &&
6266 (mr.end() <= endWord());
6267 }
6268
6269 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6270 return (start >= _bmStartWord && (start + size) <= endWord());
6271 }
6272
6273 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6274 // verify that there are no 1 bits in the interval [left, right)
6275 FalseBitMapClosure falseBitMapClosure;
6276 iterate(&falseBitMapClosure, left, right);
6277 }
6278
6279 void CMSBitMap::region_invariant(MemRegion mr)
6280 {
6281 assert_locked();
6282 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6283 assert(!mr.is_empty(), "unexpected empty region");
6284 assert(covers(mr), "mr should be covered by bit map");
6285 // convert address range into offset range
6286 size_t start_ofs = heapWordToOffset(mr.start());
6287 // Make sure that end() is appropriately aligned
6288 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6289 (1 << (_shifter+LogHeapWordSize))),
6290 "Misaligned mr.end()");
6291 size_t end_ofs = heapWordToOffset(mr.end());
6292 assert(end_ofs > start_ofs, "Should mark at least one bit");
6293 }
6294
6295 #endif
6296
6297 bool CMSMarkStack::allocate(size_t size) {
6298 // allocate a stack of the requisite depth
6299 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6300 size * sizeof(oop)));
6301 if (!rs.is_reserved()) {
6302 warning("CMSMarkStack allocation failure");
6303 return false;
6304 }
6305 if (!_virtual_space.initialize(rs, rs.size())) {
6306 warning("CMSMarkStack backing store failure");
6307 return false;
6308 }
6309 assert(_virtual_space.committed_size() == rs.size(),
6310 "didn't reserve backing store for all of CMS stack?");
6311 _base = (oop*)(_virtual_space.low());
6312 _index = 0;
6313 _capacity = size;
6314 NOT_PRODUCT(_max_depth = 0);
6315 return true;
6316 }
6317
6318 // XXX FIX ME !!! In the MT case we come in here holding a
6319 // leaf lock. For printing we need to take a further lock
6320 // which has lower rank. We need to recallibrate the two
6321 // lock-ranks involved in order to be able to rpint the
6322 // messages below. (Or defer the printing to the caller.
6323 // For now we take the expedient path of just disabling the
6324 // messages for the problematic case.)
6325 void CMSMarkStack::expand() {
6326 assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted");
6327 if (_capacity == CMSMarkStackSizeMax) {
6328 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6329 // We print a warning message only once per CMS cycle.
6330 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6331 }
6332 return;
6333 }
6334 // Double capacity if possible
6335 size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax);
6336 // Do not give up existing stack until we have managed to
6337 // get the double capacity that we desired.
6338 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6339 new_capacity * sizeof(oop)));
6340 if (rs.is_reserved()) {
6341 // Release the backing store associated with old stack
6342 _virtual_space.release();
6343 // Reinitialize virtual space for new stack
6344 if (!_virtual_space.initialize(rs, rs.size())) {
6345 fatal("Not enough swap for expanded marking stack");
6346 }
6347 _base = (oop*)(_virtual_space.low());
6348 _index = 0;
6349 _capacity = new_capacity;
6350 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6351 // Failed to double capacity, continue;
6352 // we print a detail message only once per CMS cycle.
6353 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6354 SIZE_FORMAT"K",
6355 _capacity / K, new_capacity / K);
6356 }
6357 }
6358
6359
6360 // Closures
6361 // XXX: there seems to be a lot of code duplication here;
6362 // should refactor and consolidate common code.
6363
6364 // This closure is used to mark refs into the CMS generation in
6365 // the CMS bit map. Called at the first checkpoint. This closure
6366 // assumes that we do not need to re-mark dirty cards; if the CMS
6367 // generation on which this is used is not an oldest (modulo perm gen)
6368 // generation then this will lose younger_gen cards!
6369
6370 MarkRefsIntoClosure::MarkRefsIntoClosure(
6371 MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods):
6372 _span(span),
6373 _bitMap(bitMap),
6374 _should_do_nmethods(should_do_nmethods)
6375 {
6376 assert(_ref_processor == NULL, "deliberately left NULL");
6377 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6378 }
6379
6380 void MarkRefsIntoClosure::do_oop(oop obj) {
6381 // if p points into _span, then mark corresponding bit in _markBitMap
6382 assert(obj->is_oop(), "expected an oop");
6383 HeapWord* addr = (HeapWord*)obj;
6384 if (_span.contains(addr)) {
6385 // this should be made more efficient
6386 _bitMap->mark(addr);
6387 }
6388 }
6389
6390 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6391 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6392
6393 // A variant of the above, used for CMS marking verification.
6394 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6395 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6396 bool should_do_nmethods):
6397 _span(span),
6398 _verification_bm(verification_bm),
6399 _cms_bm(cms_bm),
6400 _should_do_nmethods(should_do_nmethods) {
6401 assert(_ref_processor == NULL, "deliberately left NULL");
6402 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6403 }
6404
6405 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6406 // if p points into _span, then mark corresponding bit in _markBitMap
6407 assert(obj->is_oop(), "expected an oop");
6408 HeapWord* addr = (HeapWord*)obj;
6409 if (_span.contains(addr)) {
6410 _verification_bm->mark(addr);
6411 if (!_cms_bm->isMarked(addr)) {
6412 oop(addr)->print();
6413 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6414 fatal("... aborting");
6415 }
6416 }
6417 }
6418
6419 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6420 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6421
6422 //////////////////////////////////////////////////
6423 // MarkRefsIntoAndScanClosure
6424 //////////////////////////////////////////////////
6425
6426 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6427 ReferenceProcessor* rp,
6428 CMSBitMap* bit_map,
6429 CMSBitMap* mod_union_table,
6430 CMSMarkStack* mark_stack,
6431 CMSMarkStack* revisit_stack,
6432 CMSCollector* collector,
6433 bool should_yield,
6434 bool concurrent_precleaning):
6435 _collector(collector),
6436 _span(span),
6437 _bit_map(bit_map),
6438 _mark_stack(mark_stack),
6439 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6440 mark_stack, revisit_stack, concurrent_precleaning),
6441 _yield(should_yield),
6442 _concurrent_precleaning(concurrent_precleaning),
6443 _freelistLock(NULL)
6444 {
6445 _ref_processor = rp;
6446 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6447 }
6448
6449 // This closure is used to mark refs into the CMS generation at the
6450 // second (final) checkpoint, and to scan and transitively follow
6451 // the unmarked oops. It is also used during the concurrent precleaning
6452 // phase while scanning objects on dirty cards in the CMS generation.
6453 // The marks are made in the marking bit map and the marking stack is
6454 // used for keeping the (newly) grey objects during the scan.
6455 // The parallel version (Par_...) appears further below.
6456 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6457 if (obj != NULL) {
6458 assert(obj->is_oop(), "expected an oop");
6459 HeapWord* addr = (HeapWord*)obj;
6460 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6461 assert(_collector->overflow_list_is_empty(),
6462 "overflow list should be empty");
6463 if (_span.contains(addr) &&
6464 !_bit_map->isMarked(addr)) {
6465 // mark bit map (object is now grey)
6466 _bit_map->mark(addr);
6467 // push on marking stack (stack should be empty), and drain the
6468 // stack by applying this closure to the oops in the oops popped
6469 // from the stack (i.e. blacken the grey objects)
6470 bool res = _mark_stack->push(obj);
6471 assert(res, "Should have space to push on empty stack");
6472 do {
6473 oop new_oop = _mark_stack->pop();
6474 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6475 assert(new_oop->is_parsable(), "Found unparsable oop");
6476 assert(_bit_map->isMarked((HeapWord*)new_oop),
6477 "only grey objects on this stack");
6478 // iterate over the oops in this oop, marking and pushing
6479 // the ones in CMS heap (i.e. in _span).
6480 new_oop->oop_iterate(&_pushAndMarkClosure);
6481 // check if it's time to yield
6482 do_yield_check();
6483 } while (!_mark_stack->isEmpty() ||
6484 (!_concurrent_precleaning && take_from_overflow_list()));
6485 // if marking stack is empty, and we are not doing this
6486 // during precleaning, then check the overflow list
6487 }
6488 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6489 assert(_collector->overflow_list_is_empty(),
6490 "overflow list was drained above");
6491 // We could restore evacuated mark words, if any, used for
6492 // overflow list links here because the overflow list is
6493 // provably empty here. That would reduce the maximum
6494 // size requirements for preserved_{oop,mark}_stack.
6495 // But we'll just postpone it until we are all done
6496 // so we can just stream through.
6497 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6498 _collector->restore_preserved_marks_if_any();
6499 assert(_collector->no_preserved_marks(), "No preserved marks");
6500 }
6501 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6502 "All preserved marks should have been restored above");
6503 }
6504 }
6505
6506 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6507 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6508
6509 void MarkRefsIntoAndScanClosure::do_yield_work() {
6510 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6511 "CMS thread should hold CMS token");
6512 assert_lock_strong(_freelistLock);
6513 assert_lock_strong(_bit_map->lock());
6514 // relinquish the free_list_lock and bitMaplock()
6515 _bit_map->lock()->unlock();
6516 _freelistLock->unlock();
6517 ConcurrentMarkSweepThread::desynchronize(true);
6518 ConcurrentMarkSweepThread::acknowledge_yield_request();
6519 _collector->stopTimer();
6520 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6521 if (PrintCMSStatistics != 0) {
6522 _collector->incrementYields();
6523 }
6524 _collector->icms_wait();
6525
6526 // See the comment in coordinator_yield()
6527 for (unsigned i = 0;
6528 i < CMSYieldSleepCount &&
6529 ConcurrentMarkSweepThread::should_yield() &&
6530 !CMSCollector::foregroundGCIsActive();
6531 ++i) {
6532 os::sleep(Thread::current(), 1, false);
6533 ConcurrentMarkSweepThread::acknowledge_yield_request();
6534 }
6535
6536 ConcurrentMarkSweepThread::synchronize(true);
6537 _freelistLock->lock_without_safepoint_check();
6538 _bit_map->lock()->lock_without_safepoint_check();
6539 _collector->startTimer();
6540 }
6541
6542 ///////////////////////////////////////////////////////////
6543 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6544 // MarkRefsIntoAndScanClosure
6545 ///////////////////////////////////////////////////////////
6546 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6547 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6548 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6549 _span(span),
6550 _bit_map(bit_map),
6551 _work_queue(work_queue),
6552 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6553 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6554 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6555 revisit_stack)
6556 {
6557 _ref_processor = rp;
6558 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6559 }
6560
6561 // This closure is used to mark refs into the CMS generation at the
6562 // second (final) checkpoint, and to scan and transitively follow
6563 // the unmarked oops. The marks are made in the marking bit map and
6564 // the work_queue is used for keeping the (newly) grey objects during
6565 // the scan phase whence they are also available for stealing by parallel
6566 // threads. Since the marking bit map is shared, updates are
6567 // synchronized (via CAS).
6568 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6569 if (obj != NULL) {
6570 // Ignore mark word because this could be an already marked oop
6571 // that may be chained at the end of the overflow list.
6572 assert(obj->is_oop(), "expected an oop");
6573 HeapWord* addr = (HeapWord*)obj;
6574 if (_span.contains(addr) &&
6575 !_bit_map->isMarked(addr)) {
6576 // mark bit map (object will become grey):
6577 // It is possible for several threads to be
6578 // trying to "claim" this object concurrently;
6579 // the unique thread that succeeds in marking the
6580 // object first will do the subsequent push on
6581 // to the work queue (or overflow list).
6582 if (_bit_map->par_mark(addr)) {
6583 // push on work_queue (which may not be empty), and trim the
6584 // queue to an appropriate length by applying this closure to
6585 // the oops in the oops popped from the stack (i.e. blacken the
6586 // grey objects)
6587 bool res = _work_queue->push(obj);
6588 assert(res, "Low water mark should be less than capacity?");
6589 trim_queue(_low_water_mark);
6590 } // Else, another thread claimed the object
6591 }
6592 }
6593 }
6594
6595 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6596 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6597
6598 // This closure is used to rescan the marked objects on the dirty cards
6599 // in the mod union table and the card table proper.
6600 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6601 oop p, MemRegion mr) {
6602
6603 size_t size = 0;
6604 HeapWord* addr = (HeapWord*)p;
6605 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6606 assert(_span.contains(addr), "we are scanning the CMS generation");
6607 // check if it's time to yield
6608 if (do_yield_check()) {
6609 // We yielded for some foreground stop-world work,
6610 // and we have been asked to abort this ongoing preclean cycle.
6611 return 0;
6612 }
6613 if (_bitMap->isMarked(addr)) {
6614 // it's marked; is it potentially uninitialized?
6615 if (p->klass_or_null() != NULL) {
6616 if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) {
6617 // Signal precleaning to redirty the card since
6618 // the klass pointer is already installed.
6619 assert(size == 0, "Initial value");
6620 } else {
6621 assert(p->is_parsable(), "must be parsable.");
6622 // an initialized object; ignore mark word in verification below
6623 // since we are running concurrent with mutators
6624 assert(p->is_oop(true), "should be an oop");
6625 if (p->is_objArray()) {
6626 // objArrays are precisely marked; restrict scanning
6627 // to dirty cards only.
6628 size = CompactibleFreeListSpace::adjustObjectSize(
6629 p->oop_iterate(_scanningClosure, mr));
6630 } else {
6631 // A non-array may have been imprecisely marked; we need
6632 // to scan object in its entirety.
6633 size = CompactibleFreeListSpace::adjustObjectSize(
6634 p->oop_iterate(_scanningClosure));
6635 }
6636 #ifdef DEBUG
6637 size_t direct_size =
6638 CompactibleFreeListSpace::adjustObjectSize(p->size());
6639 assert(size == direct_size, "Inconsistency in size");
6640 assert(size >= 3, "Necessary for Printezis marks to work");
6641 if (!_bitMap->isMarked(addr+1)) {
6642 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6643 } else {
6644 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6645 assert(_bitMap->isMarked(addr+size-1),
6646 "inconsistent Printezis mark");
6647 }
6648 #endif // DEBUG
6649 }
6650 } else {
6651 // an unitialized object
6652 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6653 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6654 size = pointer_delta(nextOneAddr + 1, addr);
6655 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6656 "alignment problem");
6657 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6658 // will dirty the card when the klass pointer is installed in the
6659 // object (signalling the completion of initialization).
6660 }
6661 } else {
6662 // Either a not yet marked object or an uninitialized object
6663 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6664 // An uninitialized object, skip to the next card, since
6665 // we may not be able to read its P-bits yet.
6666 assert(size == 0, "Initial value");
6667 } else {
6668 // An object not (yet) reached by marking: we merely need to
6669 // compute its size so as to go look at the next block.
6670 assert(p->is_oop(true), "should be an oop");
6671 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6672 }
6673 }
6674 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6675 return size;
6676 }
6677
6678 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6679 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6680 "CMS thread should hold CMS token");
6681 assert_lock_strong(_freelistLock);
6682 assert_lock_strong(_bitMap->lock());
6683 // relinquish the free_list_lock and bitMaplock()
6684 _bitMap->lock()->unlock();
6685 _freelistLock->unlock();
6686 ConcurrentMarkSweepThread::desynchronize(true);
6687 ConcurrentMarkSweepThread::acknowledge_yield_request();
6688 _collector->stopTimer();
6689 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6690 if (PrintCMSStatistics != 0) {
6691 _collector->incrementYields();
6692 }
6693 _collector->icms_wait();
6694
6695 // See the comment in coordinator_yield()
6696 for (unsigned i = 0; i < CMSYieldSleepCount &&
6697 ConcurrentMarkSweepThread::should_yield() &&
6698 !CMSCollector::foregroundGCIsActive(); ++i) {
6699 os::sleep(Thread::current(), 1, false);
6700 ConcurrentMarkSweepThread::acknowledge_yield_request();
6701 }
6702
6703 ConcurrentMarkSweepThread::synchronize(true);
6704 _freelistLock->lock_without_safepoint_check();
6705 _bitMap->lock()->lock_without_safepoint_check();
6706 _collector->startTimer();
6707 }
6708
6709
6710 //////////////////////////////////////////////////////////////////
6711 // SurvivorSpacePrecleanClosure
6712 //////////////////////////////////////////////////////////////////
6713 // This (single-threaded) closure is used to preclean the oops in
6714 // the survivor spaces.
6715 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6716
6717 HeapWord* addr = (HeapWord*)p;
6718 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6719 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6720 assert(p->klass_or_null() != NULL, "object should be initializd");
6721 assert(p->is_parsable(), "must be parsable.");
6722 // an initialized object; ignore mark word in verification below
6723 // since we are running concurrent with mutators
6724 assert(p->is_oop(true), "should be an oop");
6725 // Note that we do not yield while we iterate over
6726 // the interior oops of p, pushing the relevant ones
6727 // on our marking stack.
6728 size_t size = p->oop_iterate(_scanning_closure);
6729 do_yield_check();
6730 // Observe that below, we do not abandon the preclean
6731 // phase as soon as we should; rather we empty the
6732 // marking stack before returning. This is to satisfy
6733 // some existing assertions. In general, it may be a
6734 // good idea to abort immediately and complete the marking
6735 // from the grey objects at a later time.
6736 while (!_mark_stack->isEmpty()) {
6737 oop new_oop = _mark_stack->pop();
6738 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6739 assert(new_oop->is_parsable(), "Found unparsable oop");
6740 assert(_bit_map->isMarked((HeapWord*)new_oop),
6741 "only grey objects on this stack");
6742 // iterate over the oops in this oop, marking and pushing
6743 // the ones in CMS heap (i.e. in _span).
6744 new_oop->oop_iterate(_scanning_closure);
6745 // check if it's time to yield
6746 do_yield_check();
6747 }
6748 unsigned int after_count =
6749 GenCollectedHeap::heap()->total_collections();
6750 bool abort = (_before_count != after_count) ||
6751 _collector->should_abort_preclean();
6752 return abort ? 0 : size;
6753 }
6754
6755 void SurvivorSpacePrecleanClosure::do_yield_work() {
6756 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6757 "CMS thread should hold CMS token");
6758 assert_lock_strong(_bit_map->lock());
6759 // Relinquish the bit map lock
6760 _bit_map->lock()->unlock();
6761 ConcurrentMarkSweepThread::desynchronize(true);
6762 ConcurrentMarkSweepThread::acknowledge_yield_request();
6763 _collector->stopTimer();
6764 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6765 if (PrintCMSStatistics != 0) {
6766 _collector->incrementYields();
6767 }
6768 _collector->icms_wait();
6769
6770 // See the comment in coordinator_yield()
6771 for (unsigned i = 0; i < CMSYieldSleepCount &&
6772 ConcurrentMarkSweepThread::should_yield() &&
6773 !CMSCollector::foregroundGCIsActive(); ++i) {
6774 os::sleep(Thread::current(), 1, false);
6775 ConcurrentMarkSweepThread::acknowledge_yield_request();
6776 }
6777
6778 ConcurrentMarkSweepThread::synchronize(true);
6779 _bit_map->lock()->lock_without_safepoint_check();
6780 _collector->startTimer();
6781 }
6782
6783 // This closure is used to rescan the marked objects on the dirty cards
6784 // in the mod union table and the card table proper. In the parallel
6785 // case, although the bitMap is shared, we do a single read so the
6786 // isMarked() query is "safe".
6787 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6788 // Ignore mark word because we are running concurrent with mutators
6789 assert(p->is_oop_or_null(true), "expected an oop or null");
6790 HeapWord* addr = (HeapWord*)p;
6791 assert(_span.contains(addr), "we are scanning the CMS generation");
6792 bool is_obj_array = false;
6793 #ifdef DEBUG
6794 if (!_parallel) {
6795 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6796 assert(_collector->overflow_list_is_empty(),
6797 "overflow list should be empty");
6798
6799 }
6800 #endif // DEBUG
6801 if (_bit_map->isMarked(addr)) {
6802 // Obj arrays are precisely marked, non-arrays are not;
6803 // so we scan objArrays precisely and non-arrays in their
6804 // entirety.
6805 if (p->is_objArray()) {
6806 is_obj_array = true;
6807 if (_parallel) {
6808 p->oop_iterate(_par_scan_closure, mr);
6809 } else {
6810 p->oop_iterate(_scan_closure, mr);
6811 }
6812 } else {
6813 if (_parallel) {
6814 p->oop_iterate(_par_scan_closure);
6815 } else {
6816 p->oop_iterate(_scan_closure);
6817 }
6818 }
6819 }
6820 #ifdef DEBUG
6821 if (!_parallel) {
6822 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6823 assert(_collector->overflow_list_is_empty(),
6824 "overflow list should be empty");
6825
6826 }
6827 #endif // DEBUG
6828 return is_obj_array;
6829 }
6830
6831 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6832 MemRegion span,
6833 CMSBitMap* bitMap, CMSMarkStack* markStack,
6834 CMSMarkStack* revisitStack,
6835 bool should_yield, bool verifying):
6836 _collector(collector),
6837 _span(span),
6838 _bitMap(bitMap),
6839 _mut(&collector->_modUnionTable),
6840 _markStack(markStack),
6841 _revisitStack(revisitStack),
6842 _yield(should_yield),
6843 _skipBits(0)
6844 {
6845 assert(_markStack->isEmpty(), "stack should be empty");
6846 _finger = _bitMap->startWord();
6847 _threshold = _finger;
6848 assert(_collector->_restart_addr == NULL, "Sanity check");
6849 assert(_span.contains(_finger), "Out of bounds _finger?");
6850 DEBUG_ONLY(_verifying = verifying;)
6851 }
6852
6853 void MarkFromRootsClosure::reset(HeapWord* addr) {
6854 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6855 assert(_span.contains(addr), "Out of bounds _finger?");
6856 _finger = addr;
6857 _threshold = (HeapWord*)round_to(
6858 (intptr_t)_finger, CardTableModRefBS::card_size);
6859 }
6860
6861 // Should revisit to see if this should be restructured for
6862 // greater efficiency.
6863 void MarkFromRootsClosure::do_bit(size_t offset) {
6864 if (_skipBits > 0) {
6865 _skipBits--;
6866 return;
6867 }
6868 // convert offset into a HeapWord*
6869 HeapWord* addr = _bitMap->startWord() + offset;
6870 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6871 "address out of range");
6872 assert(_bitMap->isMarked(addr), "tautology");
6873 if (_bitMap->isMarked(addr+1)) {
6874 // this is an allocated but not yet initialized object
6875 assert(_skipBits == 0, "tautology");
6876 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6877 oop p = oop(addr);
6878 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6879 DEBUG_ONLY(if (!_verifying) {)
6880 // We re-dirty the cards on which this object lies and increase
6881 // the _threshold so that we'll come back to scan this object
6882 // during the preclean or remark phase. (CMSCleanOnEnter)
6883 if (CMSCleanOnEnter) {
6884 size_t sz = _collector->block_size_using_printezis_bits(addr);
6885 HeapWord* start_card_addr = (HeapWord*)round_down(
6886 (intptr_t)addr, CardTableModRefBS::card_size);
6887 HeapWord* end_card_addr = (HeapWord*)round_to(
6888 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
6889 MemRegion redirty_range = MemRegion(start_card_addr, end_card_addr);
6890 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6891 // Bump _threshold to end_card_addr; note that
6892 // _threshold cannot possibly exceed end_card_addr, anyhow.
6893 // This prevents future clearing of the card as the scan proceeds
6894 // to the right.
6895 assert(_threshold <= end_card_addr,
6896 "Because we are just scanning into this object");
6897 if (_threshold < end_card_addr) {
6898 _threshold = end_card_addr;
6899 }
6900 if (p->klass_or_null() != NULL) {
6901 // Redirty the range of cards...
6902 _mut->mark_range(redirty_range);
6903 } // ...else the setting of klass will dirty the card anyway.
6904 }
6905 DEBUG_ONLY(})
6906 return;
6907 }
6908 }
6909 scanOopsInOop(addr);
6910 }
6911
6912 // We take a break if we've been at this for a while,
6913 // so as to avoid monopolizing the locks involved.
6914 void MarkFromRootsClosure::do_yield_work() {
6915 // First give up the locks, then yield, then re-lock
6916 // We should probably use a constructor/destructor idiom to
6917 // do this unlock/lock or modify the MutexUnlocker class to
6918 // serve our purpose. XXX
6919 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6920 "CMS thread should hold CMS token");
6921 assert_lock_strong(_bitMap->lock());
6922 _bitMap->lock()->unlock();
6923 ConcurrentMarkSweepThread::desynchronize(true);
6924 ConcurrentMarkSweepThread::acknowledge_yield_request();
6925 _collector->stopTimer();
6926 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6927 if (PrintCMSStatistics != 0) {
6928 _collector->incrementYields();
6929 }
6930 _collector->icms_wait();
6931
6932 // See the comment in coordinator_yield()
6933 for (unsigned i = 0; i < CMSYieldSleepCount &&
6934 ConcurrentMarkSweepThread::should_yield() &&
6935 !CMSCollector::foregroundGCIsActive(); ++i) {
6936 os::sleep(Thread::current(), 1, false);
6937 ConcurrentMarkSweepThread::acknowledge_yield_request();
6938 }
6939
6940 ConcurrentMarkSweepThread::synchronize(true);
6941 _bitMap->lock()->lock_without_safepoint_check();
6942 _collector->startTimer();
6943 }
6944
6945 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6946 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6947 assert(_markStack->isEmpty(),
6948 "should drain stack to limit stack usage");
6949 // convert ptr to an oop preparatory to scanning
6950 oop obj = oop(ptr);
6951 // Ignore mark word in verification below, since we
6952 // may be running concurrent with mutators.
6953 assert(obj->is_oop(true), "should be an oop");
6954 assert(_finger <= ptr, "_finger runneth ahead");
6955 // advance the finger to right end of this object
6956 _finger = ptr + obj->size();
6957 assert(_finger > ptr, "we just incremented it above");
6958 // On large heaps, it may take us some time to get through
6959 // the marking phase (especially if running iCMS). During
6960 // this time it's possible that a lot of mutations have
6961 // accumulated in the card table and the mod union table --
6962 // these mutation records are redundant until we have
6963 // actually traced into the corresponding card.
6964 // Here, we check whether advancing the finger would make
6965 // us cross into a new card, and if so clear corresponding
6966 // cards in the MUT (preclean them in the card-table in the
6967 // future).
6968
6969 DEBUG_ONLY(if (!_verifying) {)
6970 // The clean-on-enter optimization is disabled by default,
6971 // until we fix 6178663.
6972 if (CMSCleanOnEnter && (_finger > _threshold)) {
6973 // [_threshold, _finger) represents the interval
6974 // of cards to be cleared in MUT (or precleaned in card table).
6975 // The set of cards to be cleared is all those that overlap
6976 // with the interval [_threshold, _finger); note that
6977 // _threshold is always kept card-aligned but _finger isn't
6978 // always card-aligned.
6979 HeapWord* old_threshold = _threshold;
6980 assert(old_threshold == (HeapWord*)round_to(
6981 (intptr_t)old_threshold, CardTableModRefBS::card_size),
6982 "_threshold should always be card-aligned");
6983 _threshold = (HeapWord*)round_to(
6984 (intptr_t)_finger, CardTableModRefBS::card_size);
6985 MemRegion mr(old_threshold, _threshold);
6986 assert(!mr.is_empty(), "Control point invariant");
6987 assert(_span.contains(mr), "Should clear within span");
6988 // XXX When _finger crosses from old gen into perm gen
6989 // we may be doing unnecessary cleaning; do better in the
6990 // future by detecting that condition and clearing fewer
6991 // MUT/CT entries.
6992 _mut->clear_range(mr);
6993 }
6994 DEBUG_ONLY(})
6995
6996 // Note: the finger doesn't advance while we drain
6997 // the stack below.
6998 PushOrMarkClosure pushOrMarkClosure(_collector,
6999 _span, _bitMap, _markStack,
7000 _revisitStack,
7001 _finger, this);
7002 bool res = _markStack->push(obj);
7003 assert(res, "Empty non-zero size stack should have space for single push");
7004 while (!_markStack->isEmpty()) {
7005 oop new_oop = _markStack->pop();
7006 // Skip verifying header mark word below because we are
7007 // running concurrent with mutators.
7008 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7009 // now scan this oop's oops
7010 new_oop->oop_iterate(&pushOrMarkClosure);
7011 do_yield_check();
7012 }
7013 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7014 }
7015
7016 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7017 CMSCollector* collector, MemRegion span,
7018 CMSBitMap* bit_map,
7019 OopTaskQueue* work_queue,
7020 CMSMarkStack* overflow_stack,
7021 CMSMarkStack* revisit_stack,
7022 bool should_yield):
7023 _collector(collector),
7024 _whole_span(collector->_span),
7025 _span(span),
7026 _bit_map(bit_map),
7027 _mut(&collector->_modUnionTable),
7028 _work_queue(work_queue),
7029 _overflow_stack(overflow_stack),
7030 _revisit_stack(revisit_stack),
7031 _yield(should_yield),
7032 _skip_bits(0),
7033 _task(task)
7034 {
7035 assert(_work_queue->size() == 0, "work_queue should be empty");
7036 _finger = span.start();
7037 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7038 assert(_span.contains(_finger), "Out of bounds _finger?");
7039 }
7040
7041 // Should revisit to see if this should be restructured for
7042 // greater efficiency.
7043 void Par_MarkFromRootsClosure::do_bit(size_t offset) {
7044 if (_skip_bits > 0) {
7045 _skip_bits--;
7046 return;
7047 }
7048 // convert offset into a HeapWord*
7049 HeapWord* addr = _bit_map->startWord() + offset;
7050 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7051 "address out of range");
7052 assert(_bit_map->isMarked(addr), "tautology");
7053 if (_bit_map->isMarked(addr+1)) {
7054 // this is an allocated object that might not yet be initialized
7055 assert(_skip_bits == 0, "tautology");
7056 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7057 oop p = oop(addr);
7058 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7059 // in the case of Clean-on-Enter optimization, redirty card
7060 // and avoid clearing card by increasing the threshold.
7061 return;
7062 }
7063 }
7064 scan_oops_in_oop(addr);
7065 }
7066
7067 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7068 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7069 // Should we assert that our work queue is empty or
7070 // below some drain limit?
7071 assert(_work_queue->size() == 0,
7072 "should drain stack to limit stack usage");
7073 // convert ptr to an oop preparatory to scanning
7074 oop obj = oop(ptr);
7075 // Ignore mark word in verification below, since we
7076 // may be running concurrent with mutators.
7077 assert(obj->is_oop(true), "should be an oop");
7078 assert(_finger <= ptr, "_finger runneth ahead");
7079 // advance the finger to right end of this object
7080 _finger = ptr + obj->size();
7081 assert(_finger > ptr, "we just incremented it above");
7082 // On large heaps, it may take us some time to get through
7083 // the marking phase (especially if running iCMS). During
7084 // this time it's possible that a lot of mutations have
7085 // accumulated in the card table and the mod union table --
7086 // these mutation records are redundant until we have
7087 // actually traced into the corresponding card.
7088 // Here, we check whether advancing the finger would make
7089 // us cross into a new card, and if so clear corresponding
7090 // cards in the MUT (preclean them in the card-table in the
7091 // future).
7092
7093 // The clean-on-enter optimization is disabled by default,
7094 // until we fix 6178663.
7095 if (CMSCleanOnEnter && (_finger > _threshold)) {
7096 // [_threshold, _finger) represents the interval
7097 // of cards to be cleared in MUT (or precleaned in card table).
7098 // The set of cards to be cleared is all those that overlap
7099 // with the interval [_threshold, _finger); note that
7100 // _threshold is always kept card-aligned but _finger isn't
7101 // always card-aligned.
7102 HeapWord* old_threshold = _threshold;
7103 assert(old_threshold == (HeapWord*)round_to(
7104 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7105 "_threshold should always be card-aligned");
7106 _threshold = (HeapWord*)round_to(
7107 (intptr_t)_finger, CardTableModRefBS::card_size);
7108 MemRegion mr(old_threshold, _threshold);
7109 assert(!mr.is_empty(), "Control point invariant");
7110 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7111 // XXX When _finger crosses from old gen into perm gen
7112 // we may be doing unnecessary cleaning; do better in the
7113 // future by detecting that condition and clearing fewer
7114 // MUT/CT entries.
7115 _mut->clear_range(mr);
7116 }
7117
7118 // Note: the local finger doesn't advance while we drain
7119 // the stack below, but the global finger sure can and will.
7120 HeapWord** gfa = _task->global_finger_addr();
7121 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7122 _span, _bit_map,
7123 _work_queue,
7124 _overflow_stack,
7125 _revisit_stack,
7126 _finger,
7127 gfa, this);
7128 bool res = _work_queue->push(obj); // overflow could occur here
7129 assert(res, "Will hold once we use workqueues");
7130 while (true) {
7131 oop new_oop;
7132 if (!_work_queue->pop_local(new_oop)) {
7133 // We emptied our work_queue; check if there's stuff that can
7134 // be gotten from the overflow stack.
7135 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7136 _overflow_stack, _work_queue)) {
7137 do_yield_check();
7138 continue;
7139 } else { // done
7140 break;
7141 }
7142 }
7143 // Skip verifying header mark word below because we are
7144 // running concurrent with mutators.
7145 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7146 // now scan this oop's oops
7147 new_oop->oop_iterate(&pushOrMarkClosure);
7148 do_yield_check();
7149 }
7150 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7151 }
7152
7153 // Yield in response to a request from VM Thread or
7154 // from mutators.
7155 void Par_MarkFromRootsClosure::do_yield_work() {
7156 assert(_task != NULL, "sanity");
7157 _task->yield();
7158 }
7159
7160 // A variant of the above used for verifying CMS marking work.
7161 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7162 MemRegion span,
7163 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7164 CMSMarkStack* mark_stack):
7165 _collector(collector),
7166 _span(span),
7167 _verification_bm(verification_bm),
7168 _cms_bm(cms_bm),
7169 _mark_stack(mark_stack),
7170 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7171 mark_stack)
7172 {
7173 assert(_mark_stack->isEmpty(), "stack should be empty");
7174 _finger = _verification_bm->startWord();
7175 assert(_collector->_restart_addr == NULL, "Sanity check");
7176 assert(_span.contains(_finger), "Out of bounds _finger?");
7177 }
7178
7179 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7180 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7181 assert(_span.contains(addr), "Out of bounds _finger?");
7182 _finger = addr;
7183 }
7184
7185 // Should revisit to see if this should be restructured for
7186 // greater efficiency.
7187 void MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7188 // convert offset into a HeapWord*
7189 HeapWord* addr = _verification_bm->startWord() + offset;
7190 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7191 "address out of range");
7192 assert(_verification_bm->isMarked(addr), "tautology");
7193 assert(_cms_bm->isMarked(addr), "tautology");
7194
7195 assert(_mark_stack->isEmpty(),
7196 "should drain stack to limit stack usage");
7197 // convert addr to an oop preparatory to scanning
7198 oop obj = oop(addr);
7199 assert(obj->is_oop(), "should be an oop");
7200 assert(_finger <= addr, "_finger runneth ahead");
7201 // advance the finger to right end of this object
7202 _finger = addr + obj->size();
7203 assert(_finger > addr, "we just incremented it above");
7204 // Note: the finger doesn't advance while we drain
7205 // the stack below.
7206 bool res = _mark_stack->push(obj);
7207 assert(res, "Empty non-zero size stack should have space for single push");
7208 while (!_mark_stack->isEmpty()) {
7209 oop new_oop = _mark_stack->pop();
7210 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7211 // now scan this oop's oops
7212 new_oop->oop_iterate(&_pam_verify_closure);
7213 }
7214 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7215 }
7216
7217 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7218 CMSCollector* collector, MemRegion span,
7219 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7220 CMSMarkStack* mark_stack):
7221 OopClosure(collector->ref_processor()),
7222 _collector(collector),
7223 _span(span),
7224 _verification_bm(verification_bm),
7225 _cms_bm(cms_bm),
7226 _mark_stack(mark_stack)
7227 { }
7228
7229 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7230 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7231
7232 // Upon stack overflow, we discard (part of) the stack,
7233 // remembering the least address amongst those discarded
7234 // in CMSCollector's _restart_address.
7235 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7236 // Remember the least grey address discarded
7237 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7238 _collector->lower_restart_addr(ra);
7239 _mark_stack->reset(); // discard stack contents
7240 _mark_stack->expand(); // expand the stack if possible
7241 }
7242
7243 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7244 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7245 HeapWord* addr = (HeapWord*)obj;
7246 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7247 // Oop lies in _span and isn't yet grey or black
7248 _verification_bm->mark(addr); // now grey
7249 if (!_cms_bm->isMarked(addr)) {
7250 oop(addr)->print();
7251 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7252 addr);
7253 fatal("... aborting");
7254 }
7255
7256 if (!_mark_stack->push(obj)) { // stack overflow
7257 if (PrintCMSStatistics != 0) {
7258 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7259 SIZE_FORMAT, _mark_stack->capacity());
7260 }
7261 assert(_mark_stack->isFull(), "Else push should have succeeded");
7262 handle_stack_overflow(addr);
7263 }
7264 // anything including and to the right of _finger
7265 // will be scanned as we iterate over the remainder of the
7266 // bit map
7267 }
7268 }
7269
7270 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7271 MemRegion span,
7272 CMSBitMap* bitMap, CMSMarkStack* markStack,
7273 CMSMarkStack* revisitStack,
7274 HeapWord* finger, MarkFromRootsClosure* parent) :
7275 OopClosure(collector->ref_processor()),
7276 _collector(collector),
7277 _span(span),
7278 _bitMap(bitMap),
7279 _markStack(markStack),
7280 _revisitStack(revisitStack),
7281 _finger(finger),
7282 _parent(parent),
7283 _should_remember_klasses(collector->should_unload_classes())
7284 { }
7285
7286 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7287 MemRegion span,
7288 CMSBitMap* bit_map,
7289 OopTaskQueue* work_queue,
7290 CMSMarkStack* overflow_stack,
7291 CMSMarkStack* revisit_stack,
7292 HeapWord* finger,
7293 HeapWord** global_finger_addr,
7294 Par_MarkFromRootsClosure* parent) :
7295 OopClosure(collector->ref_processor()),
7296 _collector(collector),
7297 _whole_span(collector->_span),
7298 _span(span),
7299 _bit_map(bit_map),
7300 _work_queue(work_queue),
7301 _overflow_stack(overflow_stack),
7302 _revisit_stack(revisit_stack),
7303 _finger(finger),
7304 _global_finger_addr(global_finger_addr),
7305 _parent(parent),
7306 _should_remember_klasses(collector->should_unload_classes())
7307 { }
7308
7309 void CMSCollector::lower_restart_addr(HeapWord* low) {
7310 assert(_span.contains(low), "Out of bounds addr");
7311 if (_restart_addr == NULL) {
7312 _restart_addr = low;
7313 } else {
7314 _restart_addr = MIN2(_restart_addr, low);
7315 }
7316 }
7317
7318 // Upon stack overflow, we discard (part of) the stack,
7319 // remembering the least address amongst those discarded
7320 // in CMSCollector's _restart_address.
7321 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7322 // Remember the least grey address discarded
7323 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7324 _collector->lower_restart_addr(ra);
7325 _markStack->reset(); // discard stack contents
7326 _markStack->expand(); // expand the stack if possible
7327 }
7328
7329 // Upon stack overflow, we discard (part of) the stack,
7330 // remembering the least address amongst those discarded
7331 // in CMSCollector's _restart_address.
7332 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7333 // We need to do this under a mutex to prevent other
7334 // workers from interfering with the expansion below.
7335 MutexLockerEx ml(_overflow_stack->par_lock(),
7336 Mutex::_no_safepoint_check_flag);
7337 // Remember the least grey address discarded
7338 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7339 _collector->lower_restart_addr(ra);
7340 _overflow_stack->reset(); // discard stack contents
7341 _overflow_stack->expand(); // expand the stack if possible
7342 }
7343
7344 void PushOrMarkClosure::do_oop(oop obj) {
7345 // Ignore mark word because we are running concurrent with mutators.
7346 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7347 HeapWord* addr = (HeapWord*)obj;
7348 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7349 // Oop lies in _span and isn't yet grey or black
7350 _bitMap->mark(addr); // now grey
7351 if (addr < _finger) {
7352 // the bit map iteration has already either passed, or
7353 // sampled, this bit in the bit map; we'll need to
7354 // use the marking stack to scan this oop's oops.
7355 bool simulate_overflow = false;
7356 NOT_PRODUCT(
7357 if (CMSMarkStackOverflowALot &&
7358 _collector->simulate_overflow()) {
7359 // simulate a stack overflow
7360 simulate_overflow = true;
7361 }
7362 )
7363 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7364 if (PrintCMSStatistics != 0) {
7365 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7366 SIZE_FORMAT, _markStack->capacity());
7367 }
7368 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7369 handle_stack_overflow(addr);
7370 }
7371 }
7372 // anything including and to the right of _finger
7373 // will be scanned as we iterate over the remainder of the
7374 // bit map
7375 do_yield_check();
7376 }
7377 }
7378
7379 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7380 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7381
7382 void Par_PushOrMarkClosure::do_oop(oop obj) {
7383 // Ignore mark word because we are running concurrent with mutators.
7384 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7385 HeapWord* addr = (HeapWord*)obj;
7386 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7387 // Oop lies in _span and isn't yet grey or black
7388 // We read the global_finger (volatile read) strictly after marking oop
7389 bool res = _bit_map->par_mark(addr); // now grey
7390 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7391 // Should we push this marked oop on our stack?
7392 // -- if someone else marked it, nothing to do
7393 // -- if target oop is above global finger nothing to do
7394 // -- if target oop is in chunk and above local finger
7395 // then nothing to do
7396 // -- else push on work queue
7397 if ( !res // someone else marked it, they will deal with it
7398 || (addr >= *gfa) // will be scanned in a later task
7399 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7400 return;
7401 }
7402 // the bit map iteration has already either passed, or
7403 // sampled, this bit in the bit map; we'll need to
7404 // use the marking stack to scan this oop's oops.
7405 bool simulate_overflow = false;
7406 NOT_PRODUCT(
7407 if (CMSMarkStackOverflowALot &&
7408 _collector->simulate_overflow()) {
7409 // simulate a stack overflow
7410 simulate_overflow = true;
7411 }
7412 )
7413 if (simulate_overflow ||
7414 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7415 // stack overflow
7416 if (PrintCMSStatistics != 0) {
7417 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7418 SIZE_FORMAT, _overflow_stack->capacity());
7419 }
7420 // We cannot assert that the overflow stack is full because
7421 // it may have been emptied since.
7422 assert(simulate_overflow ||
7423 _work_queue->size() == _work_queue->max_elems(),
7424 "Else push should have succeeded");
7425 handle_stack_overflow(addr);
7426 }
7427 do_yield_check();
7428 }
7429 }
7430
7431 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7432 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7433
7434 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7435 MemRegion span,
7436 ReferenceProcessor* rp,
7437 CMSBitMap* bit_map,
7438 CMSBitMap* mod_union_table,
7439 CMSMarkStack* mark_stack,
7440 CMSMarkStack* revisit_stack,
7441 bool concurrent_precleaning):
7442 OopClosure(rp),
7443 _collector(collector),
7444 _span(span),
7445 _bit_map(bit_map),
7446 _mod_union_table(mod_union_table),
7447 _mark_stack(mark_stack),
7448 _revisit_stack(revisit_stack),
7449 _concurrent_precleaning(concurrent_precleaning),
7450 _should_remember_klasses(collector->should_unload_classes())
7451 {
7452 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7453 }
7454
7455 // Grey object rescan during pre-cleaning and second checkpoint phases --
7456 // the non-parallel version (the parallel version appears further below.)
7457 void PushAndMarkClosure::do_oop(oop obj) {
7458 // If _concurrent_precleaning, ignore mark word verification
7459 assert(obj->is_oop_or_null(_concurrent_precleaning),
7460 "expected an oop or NULL");
7461 HeapWord* addr = (HeapWord*)obj;
7462 // Check if oop points into the CMS generation
7463 // and is not marked
7464 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7465 // a white object ...
7466 _bit_map->mark(addr); // ... now grey
7467 // push on the marking stack (grey set)
7468 bool simulate_overflow = false;
7469 NOT_PRODUCT(
7470 if (CMSMarkStackOverflowALot &&
7471 _collector->simulate_overflow()) {
7472 // simulate a stack overflow
7473 simulate_overflow = true;
7474 }
7475 )
7476 if (simulate_overflow || !_mark_stack->push(obj)) {
7477 if (_concurrent_precleaning) {
7478 // During precleaning we can just dirty the appropriate card
7479 // in the mod union table, thus ensuring that the object remains
7480 // in the grey set and continue. Note that no one can be intefering
7481 // with us in this action of dirtying the mod union table, so
7482 // no locking is required.
7483 _mod_union_table->mark(addr);
7484 _collector->_ser_pmc_preclean_ovflw++;
7485 } else {
7486 // During the remark phase, we need to remember this oop
7487 // in the overflow list.
7488 _collector->push_on_overflow_list(obj);
7489 _collector->_ser_pmc_remark_ovflw++;
7490 }
7491 }
7492 }
7493 }
7494
7495 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7496 MemRegion span,
7497 ReferenceProcessor* rp,
7498 CMSBitMap* bit_map,
7499 OopTaskQueue* work_queue,
7500 CMSMarkStack* revisit_stack):
7501 OopClosure(rp),
7502 _collector(collector),
7503 _span(span),
7504 _bit_map(bit_map),
7505 _work_queue(work_queue),
7506 _revisit_stack(revisit_stack),
7507 _should_remember_klasses(collector->should_unload_classes())
7508 {
7509 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7510 }
7511
7512 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7513 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7514
7515 // Grey object rescan during second checkpoint phase --
7516 // the parallel version.
7517 void Par_PushAndMarkClosure::do_oop(oop obj) {
7518 // In the assert below, we ignore the mark word because
7519 // this oop may point to an already visited object that is
7520 // on the overflow stack (in which case the mark word has
7521 // been hijacked for chaining into the overflow stack --
7522 // if this is the last object in the overflow stack then
7523 // its mark word will be NULL). Because this object may
7524 // have been subsequently popped off the global overflow
7525 // stack, and the mark word possibly restored to the prototypical
7526 // value, by the time we get to examined this failing assert in
7527 // the debugger, is_oop_or_null(false) may subsequently start
7528 // to hold.
7529 assert(obj->is_oop_or_null(true),
7530 "expected an oop or NULL");
7531 HeapWord* addr = (HeapWord*)obj;
7532 // Check if oop points into the CMS generation
7533 // and is not marked
7534 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7535 // a white object ...
7536 // If we manage to "claim" the object, by being the
7537 // first thread to mark it, then we push it on our
7538 // marking stack
7539 if (_bit_map->par_mark(addr)) { // ... now grey
7540 // push on work queue (grey set)
7541 bool simulate_overflow = false;
7542 NOT_PRODUCT(
7543 if (CMSMarkStackOverflowALot &&
7544 _collector->par_simulate_overflow()) {
7545 // simulate a stack overflow
7546 simulate_overflow = true;
7547 }
7548 )
7549 if (simulate_overflow || !_work_queue->push(obj)) {
7550 _collector->par_push_on_overflow_list(obj);
7551 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7552 }
7553 } // Else, some other thread got there first
7554 }
7555 }
7556
7557 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7558 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7559
7560 void PushAndMarkClosure::remember_klass(Klass* k) {
7561 if (!_revisit_stack->push(oop(k))) {
7562 fatal("Revisit stack overflowed in PushAndMarkClosure");
7563 }
7564 }
7565
7566 void Par_PushAndMarkClosure::remember_klass(Klass* k) {
7567 if (!_revisit_stack->par_push(oop(k))) {
7568 fatal("Revist stack overflowed in Par_PushAndMarkClosure");
7569 }
7570 }
7571
7572 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7573 Mutex* bml = _collector->bitMapLock();
7574 assert_lock_strong(bml);
7575 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7576 "CMS thread should hold CMS token");
7577
7578 bml->unlock();
7579 ConcurrentMarkSweepThread::desynchronize(true);
7580
7581 ConcurrentMarkSweepThread::acknowledge_yield_request();
7582
7583 _collector->stopTimer();
7584 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7585 if (PrintCMSStatistics != 0) {
7586 _collector->incrementYields();
7587 }
7588 _collector->icms_wait();
7589
7590 // See the comment in coordinator_yield()
7591 for (unsigned i = 0; i < CMSYieldSleepCount &&
7592 ConcurrentMarkSweepThread::should_yield() &&
7593 !CMSCollector::foregroundGCIsActive(); ++i) {
7594 os::sleep(Thread::current(), 1, false);
7595 ConcurrentMarkSweepThread::acknowledge_yield_request();
7596 }
7597
7598 ConcurrentMarkSweepThread::synchronize(true);
7599 bml->lock();
7600
7601 _collector->startTimer();
7602 }
7603
7604 bool CMSPrecleanRefsYieldClosure::should_return() {
7605 if (ConcurrentMarkSweepThread::should_yield()) {
7606 do_yield_work();
7607 }
7608 return _collector->foregroundGCIsActive();
7609 }
7610
7611 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7612 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7613 "mr should be aligned to start at a card boundary");
7614 // We'd like to assert:
7615 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7616 // "mr should be a range of cards");
7617 // However, that would be too strong in one case -- the last
7618 // partition ends at _unallocated_block which, in general, can be
7619 // an arbitrary boundary, not necessarily card aligned.
7620 if (PrintCMSStatistics != 0) {
7621 _num_dirty_cards +=
7622 mr.word_size()/CardTableModRefBS::card_size_in_words;
7623 }
7624 _space->object_iterate_mem(mr, &_scan_cl);
7625 }
7626
7627 SweepClosure::SweepClosure(CMSCollector* collector,
7628 ConcurrentMarkSweepGeneration* g,
7629 CMSBitMap* bitMap, bool should_yield) :
7630 _collector(collector),
7631 _g(g),
7632 _sp(g->cmsSpace()),
7633 _limit(_sp->sweep_limit()),
7634 _freelistLock(_sp->freelistLock()),
7635 _bitMap(bitMap),
7636 _yield(should_yield),
7637 _inFreeRange(false), // No free range at beginning of sweep
7638 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7639 _lastFreeRangeCoalesced(false),
7640 _freeFinger(g->used_region().start())
7641 {
7642 NOT_PRODUCT(
7643 _numObjectsFreed = 0;
7644 _numWordsFreed = 0;
7645 _numObjectsLive = 0;
7646 _numWordsLive = 0;
7647 _numObjectsAlreadyFree = 0;
7648 _numWordsAlreadyFree = 0;
7649 _last_fc = NULL;
7650
7651 _sp->initializeIndexedFreeListArrayReturnedBytes();
7652 _sp->dictionary()->initializeDictReturnedBytes();
7653 )
7654 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7655 "sweep _limit out of bounds");
7656 if (CMSTraceSweeper) {
7657 gclog_or_tty->print("\n====================\nStarting new sweep\n");
7658 }
7659 }
7660
7661 // We need this destructor to reclaim any space at the end
7662 // of the space, which do_blk below may not have added back to
7663 // the free lists. [basically dealing with the "fringe effect"]
7664 SweepClosure::~SweepClosure() {
7665 assert_lock_strong(_freelistLock);
7666 // this should be treated as the end of a free run if any
7667 // The current free range should be returned to the free lists
7668 // as one coalesced chunk.
7669 if (inFreeRange()) {
7670 flushCurFreeChunk(freeFinger(),
7671 pointer_delta(_limit, freeFinger()));
7672 assert(freeFinger() < _limit, "the finger pointeth off base");
7673 if (CMSTraceSweeper) {
7674 gclog_or_tty->print("destructor:");
7675 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") "
7676 "[coalesced:"SIZE_FORMAT"]\n",
7677 freeFinger(), pointer_delta(_limit, freeFinger()),
7678 lastFreeRangeCoalesced());
7679 }
7680 }
7681 NOT_PRODUCT(
7682 if (Verbose && PrintGC) {
7683 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, "
7684 SIZE_FORMAT " bytes",
7685 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7686 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7687 SIZE_FORMAT" bytes "
7688 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7689 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7690 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7691 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) *
7692 sizeof(HeapWord);
7693 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7694
7695 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7696 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7697 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes();
7698 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes;
7699 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes);
7700 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7701 indexListReturnedBytes);
7702 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7703 dictReturnedBytes);
7704 }
7705 }
7706 )
7707 // Now, in debug mode, just null out the sweep_limit
7708 NOT_PRODUCT(_sp->clear_sweep_limit();)
7709 if (CMSTraceSweeper) {
7710 gclog_or_tty->print("end of sweep\n================\n");
7711 }
7712 }
7713
7714 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7715 bool freeRangeInFreeLists) {
7716 if (CMSTraceSweeper) {
7717 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n",
7718 freeFinger, _sp->block_size(freeFinger),
7719 freeRangeInFreeLists);
7720 }
7721 assert(!inFreeRange(), "Trampling existing free range");
7722 set_inFreeRange(true);
7723 set_lastFreeRangeCoalesced(false);
7724
7725 set_freeFinger(freeFinger);
7726 set_freeRangeInFreeLists(freeRangeInFreeLists);
7727 if (CMSTestInFreeList) {
7728 if (freeRangeInFreeLists) {
7729 FreeChunk* fc = (FreeChunk*) freeFinger;
7730 assert(fc->isFree(), "A chunk on the free list should be free.");
7731 assert(fc->size() > 0, "Free range should have a size");
7732 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists");
7733 }
7734 }
7735 }
7736
7737 // Note that the sweeper runs concurrently with mutators. Thus,
7738 // it is possible for direct allocation in this generation to happen
7739 // in the middle of the sweep. Note that the sweeper also coalesces
7740 // contiguous free blocks. Thus, unless the sweeper and the allocator
7741 // synchronize appropriately freshly allocated blocks may get swept up.
7742 // This is accomplished by the sweeper locking the free lists while
7743 // it is sweeping. Thus blocks that are determined to be free are
7744 // indeed free. There is however one additional complication:
7745 // blocks that have been allocated since the final checkpoint and
7746 // mark, will not have been marked and so would be treated as
7747 // unreachable and swept up. To prevent this, the allocator marks
7748 // the bit map when allocating during the sweep phase. This leads,
7749 // however, to a further complication -- objects may have been allocated
7750 // but not yet initialized -- in the sense that the header isn't yet
7751 // installed. The sweeper can not then determine the size of the block
7752 // in order to skip over it. To deal with this case, we use a technique
7753 // (due to Printezis) to encode such uninitialized block sizes in the
7754 // bit map. Since the bit map uses a bit per every HeapWord, but the
7755 // CMS generation has a minimum object size of 3 HeapWords, it follows
7756 // that "normal marks" won't be adjacent in the bit map (there will
7757 // always be at least two 0 bits between successive 1 bits). We make use
7758 // of these "unused" bits to represent uninitialized blocks -- the bit
7759 // corresponding to the start of the uninitialized object and the next
7760 // bit are both set. Finally, a 1 bit marks the end of the object that
7761 // started with the two consecutive 1 bits to indicate its potentially
7762 // uninitialized state.
7763
7764 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7765 FreeChunk* fc = (FreeChunk*)addr;
7766 size_t res;
7767
7768 // check if we are done sweepinrg
7769 if (addr == _limit) { // we have swept up to the limit, do nothing more
7770 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7771 "sweep _limit out of bounds");
7772 // help the closure application finish
7773 return pointer_delta(_sp->end(), _limit);
7774 }
7775 assert(addr <= _limit, "sweep invariant");
7776
7777 // check if we should yield
7778 do_yield_check(addr);
7779 if (fc->isFree()) {
7780 // Chunk that is already free
7781 res = fc->size();
7782 doAlreadyFreeChunk(fc);
7783 debug_only(_sp->verifyFreeLists());
7784 assert(res == fc->size(), "Don't expect the size to change");
7785 NOT_PRODUCT(
7786 _numObjectsAlreadyFree++;
7787 _numWordsAlreadyFree += res;
7788 )
7789 NOT_PRODUCT(_last_fc = fc;)
7790 } else if (!_bitMap->isMarked(addr)) {
7791 // Chunk is fresh garbage
7792 res = doGarbageChunk(fc);
7793 debug_only(_sp->verifyFreeLists());
7794 NOT_PRODUCT(
7795 _numObjectsFreed++;
7796 _numWordsFreed += res;
7797 )
7798 } else {
7799 // Chunk that is alive.
7800 res = doLiveChunk(fc);
7801 debug_only(_sp->verifyFreeLists());
7802 NOT_PRODUCT(
7803 _numObjectsLive++;
7804 _numWordsLive += res;
7805 )
7806 }
7807 return res;
7808 }
7809
7810 // For the smart allocation, record following
7811 // split deaths - a free chunk is removed from its free list because
7812 // it is being split into two or more chunks.
7813 // split birth - a free chunk is being added to its free list because
7814 // a larger free chunk has been split and resulted in this free chunk.
7815 // coal death - a free chunk is being removed from its free list because
7816 // it is being coalesced into a large free chunk.
7817 // coal birth - a free chunk is being added to its free list because
7818 // it was created when two or more free chunks where coalesced into
7819 // this free chunk.
7820 //
7821 // These statistics are used to determine the desired number of free
7822 // chunks of a given size. The desired number is chosen to be relative
7823 // to the end of a CMS sweep. The desired number at the end of a sweep
7824 // is the
7825 // count-at-end-of-previous-sweep (an amount that was enough)
7826 // - count-at-beginning-of-current-sweep (the excess)
7827 // + split-births (gains in this size during interval)
7828 // - split-deaths (demands on this size during interval)
7829 // where the interval is from the end of one sweep to the end of the
7830 // next.
7831 //
7832 // When sweeping the sweeper maintains an accumulated chunk which is
7833 // the chunk that is made up of chunks that have been coalesced. That
7834 // will be termed the left-hand chunk. A new chunk of garbage that
7835 // is being considered for coalescing will be referred to as the
7836 // right-hand chunk.
7837 //
7838 // When making a decision on whether to coalesce a right-hand chunk with
7839 // the current left-hand chunk, the current count vs. the desired count
7840 // of the left-hand chunk is considered. Also if the right-hand chunk
7841 // is near the large chunk at the end of the heap (see
7842 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7843 // left-hand chunk is coalesced.
7844 //
7845 // When making a decision about whether to split a chunk, the desired count
7846 // vs. the current count of the candidate to be split is also considered.
7847 // If the candidate is underpopulated (currently fewer chunks than desired)
7848 // a chunk of an overpopulated (currently more chunks than desired) size may
7849 // be chosen. The "hint" associated with a free list, if non-null, points
7850 // to a free list which may be overpopulated.
7851 //
7852
7853 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) {
7854 size_t size = fc->size();
7855 // Chunks that cannot be coalesced are not in the
7856 // free lists.
7857 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7858 assert(_sp->verifyChunkInFreeLists(fc),
7859 "free chunk should be in free lists");
7860 }
7861 // a chunk that is already free, should not have been
7862 // marked in the bit map
7863 HeapWord* addr = (HeapWord*) fc;
7864 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7865 // Verify that the bit map has no bits marked between
7866 // addr and purported end of this block.
7867 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7868
7869 // Some chunks cannot be coalesced in under any circumstances.
7870 // See the definition of cantCoalesce().
7871 if (!fc->cantCoalesce()) {
7872 // This chunk can potentially be coalesced.
7873 if (_sp->adaptive_freelists()) {
7874 // All the work is done in
7875 doPostIsFreeOrGarbageChunk(fc, size);
7876 } else { // Not adaptive free lists
7877 // this is a free chunk that can potentially be coalesced by the sweeper;
7878 if (!inFreeRange()) {
7879 // if the next chunk is a free block that can't be coalesced
7880 // it doesn't make sense to remove this chunk from the free lists
7881 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
7882 assert((HeapWord*)nextChunk <= _limit, "sweep invariant");
7883 if ((HeapWord*)nextChunk < _limit && // there's a next chunk...
7884 nextChunk->isFree() && // which is free...
7885 nextChunk->cantCoalesce()) { // ... but cant be coalesced
7886 // nothing to do
7887 } else {
7888 // Potentially the start of a new free range:
7889 // Don't eagerly remove it from the free lists.
7890 // No need to remove it if it will just be put
7891 // back again. (Also from a pragmatic point of view
7892 // if it is a free block in a region that is beyond
7893 // any allocated blocks, an assertion will fail)
7894 // Remember the start of a free run.
7895 initialize_free_range(addr, true);
7896 // end - can coalesce with next chunk
7897 }
7898 } else {
7899 // the midst of a free range, we are coalescing
7900 debug_only(record_free_block_coalesced(fc);)
7901 if (CMSTraceSweeper) {
7902 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
7903 }
7904 // remove it from the free lists
7905 _sp->removeFreeChunkFromFreeLists(fc);
7906 set_lastFreeRangeCoalesced(true);
7907 // If the chunk is being coalesced and the current free range is
7908 // in the free lists, remove the current free range so that it
7909 // will be returned to the free lists in its entirety - all
7910 // the coalesced pieces included.
7911 if (freeRangeInFreeLists()) {
7912 FreeChunk* ffc = (FreeChunk*) freeFinger();
7913 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7914 "Size of free range is inconsistent with chunk size.");
7915 if (CMSTestInFreeList) {
7916 assert(_sp->verifyChunkInFreeLists(ffc),
7917 "free range is not in free lists");
7918 }
7919 _sp->removeFreeChunkFromFreeLists(ffc);
7920 set_freeRangeInFreeLists(false);
7921 }
7922 }
7923 }
7924 } else {
7925 // Code path common to both original and adaptive free lists.
7926
7927 // cant coalesce with previous block; this should be treated
7928 // as the end of a free run if any
7929 if (inFreeRange()) {
7930 // we kicked some butt; time to pick up the garbage
7931 assert(freeFinger() < addr, "the finger pointeth off base");
7932 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
7933 }
7934 // else, nothing to do, just continue
7935 }
7936 }
7937
7938 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) {
7939 // This is a chunk of garbage. It is not in any free list.
7940 // Add it to a free list or let it possibly be coalesced into
7941 // a larger chunk.
7942 HeapWord* addr = (HeapWord*) fc;
7943 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7944
7945 if (_sp->adaptive_freelists()) {
7946 // Verify that the bit map has no bits marked between
7947 // addr and purported end of just dead object.
7948 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7949
7950 doPostIsFreeOrGarbageChunk(fc, size);
7951 } else {
7952 if (!inFreeRange()) {
7953 // start of a new free range
7954 assert(size > 0, "A free range should have a size");
7955 initialize_free_range(addr, false);
7956
7957 } else {
7958 // this will be swept up when we hit the end of the
7959 // free range
7960 if (CMSTraceSweeper) {
7961 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
7962 }
7963 // If the chunk is being coalesced and the current free range is
7964 // in the free lists, remove the current free range so that it
7965 // will be returned to the free lists in its entirety - all
7966 // the coalesced pieces included.
7967 if (freeRangeInFreeLists()) {
7968 FreeChunk* ffc = (FreeChunk*)freeFinger();
7969 assert(ffc->size() == pointer_delta(addr, freeFinger()),
7970 "Size of free range is inconsistent with chunk size.");
7971 if (CMSTestInFreeList) {
7972 assert(_sp->verifyChunkInFreeLists(ffc),
7973 "free range is not in free lists");
7974 }
7975 _sp->removeFreeChunkFromFreeLists(ffc);
7976 set_freeRangeInFreeLists(false);
7977 }
7978 set_lastFreeRangeCoalesced(true);
7979 }
7980 // this will be swept up when we hit the end of the free range
7981
7982 // Verify that the bit map has no bits marked between
7983 // addr and purported end of just dead object.
7984 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7985 }
7986 return size;
7987 }
7988
7989 size_t SweepClosure::doLiveChunk(FreeChunk* fc) {
7990 HeapWord* addr = (HeapWord*) fc;
7991 // The sweeper has just found a live object. Return any accumulated
7992 // left hand chunk to the free lists.
7993 if (inFreeRange()) {
7994 if (_sp->adaptive_freelists()) {
7995 flushCurFreeChunk(freeFinger(),
7996 pointer_delta(addr, freeFinger()));
7997 } else { // not adaptive freelists
7998 set_inFreeRange(false);
7999 // Add the free range back to the free list if it is not already
8000 // there.
8001 if (!freeRangeInFreeLists()) {
8002 assert(freeFinger() < addr, "the finger pointeth off base");
8003 if (CMSTraceSweeper) {
8004 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) "
8005 "[coalesced:%d]\n",
8006 freeFinger(), pointer_delta(addr, freeFinger()),
8007 lastFreeRangeCoalesced());
8008 }
8009 _sp->addChunkAndRepairOffsetTable(freeFinger(),
8010 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced());
8011 }
8012 }
8013 }
8014
8015 // Common code path for original and adaptive free lists.
8016
8017 // this object is live: we'd normally expect this to be
8018 // an oop, and like to assert the following:
8019 // assert(oop(addr)->is_oop(), "live block should be an oop");
8020 // However, as we commented above, this may be an object whose
8021 // header hasn't yet been initialized.
8022 size_t size;
8023 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8024 if (_bitMap->isMarked(addr + 1)) {
8025 // Determine the size from the bit map, rather than trying to
8026 // compute it from the object header.
8027 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8028 size = pointer_delta(nextOneAddr + 1, addr);
8029 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8030 "alignment problem");
8031
8032 #ifdef DEBUG
8033 if (oop(addr)->klass_or_null() != NULL &&
8034 ( !_collector->should_unload_classes()
8035 || oop(addr)->is_parsable())) {
8036 // Ignore mark word because we are running concurrent with mutators
8037 assert(oop(addr)->is_oop(true), "live block should be an oop");
8038 assert(size ==
8039 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8040 "P-mark and computed size do not agree");
8041 }
8042 #endif
8043
8044 } else {
8045 // This should be an initialized object that's alive.
8046 assert(oop(addr)->klass_or_null() != NULL &&
8047 (!_collector->should_unload_classes()
8048 || oop(addr)->is_parsable()),
8049 "Should be an initialized object");
8050 // Ignore mark word because we are running concurrent with mutators
8051 assert(oop(addr)->is_oop(true), "live block should be an oop");
8052 // Verify that the bit map has no bits marked between
8053 // addr and purported end of this block.
8054 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8055 assert(size >= 3, "Necessary for Printezis marks to work");
8056 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8057 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8058 }
8059 return size;
8060 }
8061
8062 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc,
8063 size_t chunkSize) {
8064 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation
8065 // scheme.
8066 bool fcInFreeLists = fc->isFree();
8067 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8068 assert((HeapWord*)fc <= _limit, "sweep invariant");
8069 if (CMSTestInFreeList && fcInFreeLists) {
8070 assert(_sp->verifyChunkInFreeLists(fc),
8071 "free chunk is not in free lists");
8072 }
8073
8074
8075 if (CMSTraceSweeper) {
8076 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8077 }
8078
8079 HeapWord* addr = (HeapWord*) fc;
8080
8081 bool coalesce;
8082 size_t left = pointer_delta(addr, freeFinger());
8083 size_t right = chunkSize;
8084 switch (FLSCoalescePolicy) {
8085 // numeric value forms a coalition aggressiveness metric
8086 case 0: { // never coalesce
8087 coalesce = false;
8088 break;
8089 }
8090 case 1: { // coalesce if left & right chunks on overpopulated lists
8091 coalesce = _sp->coalOverPopulated(left) &&
8092 _sp->coalOverPopulated(right);
8093 break;
8094 }
8095 case 2: { // coalesce if left chunk on overpopulated list (default)
8096 coalesce = _sp->coalOverPopulated(left);
8097 break;
8098 }
8099 case 3: { // coalesce if left OR right chunk on overpopulated list
8100 coalesce = _sp->coalOverPopulated(left) ||
8101 _sp->coalOverPopulated(right);
8102 break;
8103 }
8104 case 4: { // always coalesce
8105 coalesce = true;
8106 break;
8107 }
8108 default:
8109 ShouldNotReachHere();
8110 }
8111
8112 // Should the current free range be coalesced?
8113 // If the chunk is in a free range and either we decided to coalesce above
8114 // or the chunk is near the large block at the end of the heap
8115 // (isNearLargestChunk() returns true), then coalesce this chunk.
8116 bool doCoalesce = inFreeRange() &&
8117 (coalesce || _g->isNearLargestChunk((HeapWord*)fc));
8118 if (doCoalesce) {
8119 // Coalesce the current free range on the left with the new
8120 // chunk on the right. If either is on a free list,
8121 // it must be removed from the list and stashed in the closure.
8122 if (freeRangeInFreeLists()) {
8123 FreeChunk* ffc = (FreeChunk*)freeFinger();
8124 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8125 "Size of free range is inconsistent with chunk size.");
8126 if (CMSTestInFreeList) {
8127 assert(_sp->verifyChunkInFreeLists(ffc),
8128 "Chunk is not in free lists");
8129 }
8130 _sp->coalDeath(ffc->size());
8131 _sp->removeFreeChunkFromFreeLists(ffc);
8132 set_freeRangeInFreeLists(false);
8133 }
8134 if (fcInFreeLists) {
8135 _sp->coalDeath(chunkSize);
8136 assert(fc->size() == chunkSize,
8137 "The chunk has the wrong size or is not in the free lists");
8138 _sp->removeFreeChunkFromFreeLists(fc);
8139 }
8140 set_lastFreeRangeCoalesced(true);
8141 } else { // not in a free range and/or should not coalesce
8142 // Return the current free range and start a new one.
8143 if (inFreeRange()) {
8144 // In a free range but cannot coalesce with the right hand chunk.
8145 // Put the current free range into the free lists.
8146 flushCurFreeChunk(freeFinger(),
8147 pointer_delta(addr, freeFinger()));
8148 }
8149 // Set up for new free range. Pass along whether the right hand
8150 // chunk is in the free lists.
8151 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8152 }
8153 }
8154 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) {
8155 assert(inFreeRange(), "Should only be called if currently in a free range.");
8156 assert(size > 0,
8157 "A zero sized chunk cannot be added to the free lists.");
8158 if (!freeRangeInFreeLists()) {
8159 if(CMSTestInFreeList) {
8160 FreeChunk* fc = (FreeChunk*) chunk;
8161 fc->setSize(size);
8162 assert(!_sp->verifyChunkInFreeLists(fc),
8163 "chunk should not be in free lists yet");
8164 }
8165 if (CMSTraceSweeper) {
8166 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8167 chunk, size);
8168 }
8169 // A new free range is going to be starting. The current
8170 // free range has not been added to the free lists yet or
8171 // was removed so add it back.
8172 // If the current free range was coalesced, then the death
8173 // of the free range was recorded. Record a birth now.
8174 if (lastFreeRangeCoalesced()) {
8175 _sp->coalBirth(size);
8176 }
8177 _sp->addChunkAndRepairOffsetTable(chunk, size,
8178 lastFreeRangeCoalesced());
8179 }
8180 set_inFreeRange(false);
8181 set_freeRangeInFreeLists(false);
8182 }
8183
8184 // We take a break if we've been at this for a while,
8185 // so as to avoid monopolizing the locks involved.
8186 void SweepClosure::do_yield_work(HeapWord* addr) {
8187 // Return current free chunk being used for coalescing (if any)
8188 // to the appropriate freelist. After yielding, the next
8189 // free block encountered will start a coalescing range of
8190 // free blocks. If the next free block is adjacent to the
8191 // chunk just flushed, they will need to wait for the next
8192 // sweep to be coalesced.
8193 if (inFreeRange()) {
8194 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger()));
8195 }
8196
8197 // First give up the locks, then yield, then re-lock.
8198 // We should probably use a constructor/destructor idiom to
8199 // do this unlock/lock or modify the MutexUnlocker class to
8200 // serve our purpose. XXX
8201 assert_lock_strong(_bitMap->lock());
8202 assert_lock_strong(_freelistLock);
8203 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8204 "CMS thread should hold CMS token");
8205 _bitMap->lock()->unlock();
8206 _freelistLock->unlock();
8207 ConcurrentMarkSweepThread::desynchronize(true);
8208 ConcurrentMarkSweepThread::acknowledge_yield_request();
8209 _collector->stopTimer();
8210 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8211 if (PrintCMSStatistics != 0) {
8212 _collector->incrementYields();
8213 }
8214 _collector->icms_wait();
8215
8216 // See the comment in coordinator_yield()
8217 for (unsigned i = 0; i < CMSYieldSleepCount &&
8218 ConcurrentMarkSweepThread::should_yield() &&
8219 !CMSCollector::foregroundGCIsActive(); ++i) {
8220 os::sleep(Thread::current(), 1, false);
8221 ConcurrentMarkSweepThread::acknowledge_yield_request();
8222 }
8223
8224 ConcurrentMarkSweepThread::synchronize(true);
8225 _freelistLock->lock();
8226 _bitMap->lock()->lock_without_safepoint_check();
8227 _collector->startTimer();
8228 }
8229
8230 #ifndef PRODUCT
8231 // This is actually very useful in a product build if it can
8232 // be called from the debugger. Compile it into the product
8233 // as needed.
8234 bool debug_verifyChunkInFreeLists(FreeChunk* fc) {
8235 return debug_cms_space->verifyChunkInFreeLists(fc);
8236 }
8237
8238 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const {
8239 if (CMSTraceSweeper) {
8240 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size());
8241 }
8242 }
8243 #endif
8244
8245 // CMSIsAliveClosure
8246 bool CMSIsAliveClosure::do_object_b(oop obj) {
8247 HeapWord* addr = (HeapWord*)obj;
8248 return addr != NULL &&
8249 (!_span.contains(addr) || _bit_map->isMarked(addr));
8250 }
8251
8252 // CMSKeepAliveClosure: the serial version
8253 void CMSKeepAliveClosure::do_oop(oop obj) {
8254 HeapWord* addr = (HeapWord*)obj;
8255 if (_span.contains(addr) &&
8256 !_bit_map->isMarked(addr)) {
8257 _bit_map->mark(addr);
8258 bool simulate_overflow = false;
8259 NOT_PRODUCT(
8260 if (CMSMarkStackOverflowALot &&
8261 _collector->simulate_overflow()) {
8262 // simulate a stack overflow
8263 simulate_overflow = true;
8264 }
8265 )
8266 if (simulate_overflow || !_mark_stack->push(obj)) {
8267 _collector->push_on_overflow_list(obj);
8268 _collector->_ser_kac_ovflw++;
8269 }
8270 }
8271 }
8272
8273 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8274 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8275
8276 // CMSParKeepAliveClosure: a parallel version of the above.
8277 // The work queues are private to each closure (thread),
8278 // but (may be) available for stealing by other threads.
8279 void CMSParKeepAliveClosure::do_oop(oop obj) {
8280 HeapWord* addr = (HeapWord*)obj;
8281 if (_span.contains(addr) &&
8282 !_bit_map->isMarked(addr)) {
8283 // In general, during recursive tracing, several threads
8284 // may be concurrently getting here; the first one to
8285 // "tag" it, claims it.
8286 if (_bit_map->par_mark(addr)) {
8287 bool res = _work_queue->push(obj);
8288 assert(res, "Low water mark should be much less than capacity");
8289 // Do a recursive trim in the hope that this will keep
8290 // stack usage lower, but leave some oops for potential stealers
8291 trim_queue(_low_water_mark);
8292 } // Else, another thread got there first
8293 }
8294 }
8295
8296 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8297 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8298
8299 void CMSParKeepAliveClosure::trim_queue(uint max) {
8300 while (_work_queue->size() > max) {
8301 oop new_oop;
8302 if (_work_queue->pop_local(new_oop)) {
8303 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8304 assert(_bit_map->isMarked((HeapWord*)new_oop),
8305 "no white objects on this stack!");
8306 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8307 // iterate over the oops in this oop, marking and pushing
8308 // the ones in CMS heap (i.e. in _span).
8309 new_oop->oop_iterate(&_mark_and_push);
8310 }
8311 }
8312 }
8313
8314 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8315 HeapWord* addr = (HeapWord*)obj;
8316 if (_span.contains(addr) &&
8317 !_bit_map->isMarked(addr)) {
8318 if (_bit_map->par_mark(addr)) {
8319 bool simulate_overflow = false;
8320 NOT_PRODUCT(
8321 if (CMSMarkStackOverflowALot &&
8322 _collector->par_simulate_overflow()) {
8323 // simulate a stack overflow
8324 simulate_overflow = true;
8325 }
8326 )
8327 if (simulate_overflow || !_work_queue->push(obj)) {
8328 _collector->par_push_on_overflow_list(obj);
8329 _collector->_par_kac_ovflw++;
8330 }
8331 } // Else another thread got there already
8332 }
8333 }
8334
8335 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8336 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8337
8338 //////////////////////////////////////////////////////////////////
8339 // CMSExpansionCause /////////////////////////////
8340 //////////////////////////////////////////////////////////////////
8341 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8342 switch (cause) {
8343 case _no_expansion:
8344 return "No expansion";
8345 case _satisfy_free_ratio:
8346 return "Free ratio";
8347 case _satisfy_promotion:
8348 return "Satisfy promotion";
8349 case _satisfy_allocation:
8350 return "allocation";
8351 case _allocate_par_lab:
8352 return "Par LAB";
8353 case _allocate_par_spooling_space:
8354 return "Par Spooling Space";
8355 case _adaptive_size_policy:
8356 return "Ergonomics";
8357 default:
8358 return "unknown";
8359 }
8360 }
8361
8362 void CMSDrainMarkingStackClosure::do_void() {
8363 // the max number to take from overflow list at a time
8364 const size_t num = _mark_stack->capacity()/4;
8365 while (!_mark_stack->isEmpty() ||
8366 // if stack is empty, check the overflow list
8367 _collector->take_from_overflow_list(num, _mark_stack)) {
8368 oop obj = _mark_stack->pop();
8369 HeapWord* addr = (HeapWord*)obj;
8370 assert(_span.contains(addr), "Should be within span");
8371 assert(_bit_map->isMarked(addr), "Should be marked");
8372 assert(obj->is_oop(), "Should be an oop");
8373 obj->oop_iterate(_keep_alive);
8374 }
8375 }
8376
8377 void CMSParDrainMarkingStackClosure::do_void() {
8378 // drain queue
8379 trim_queue(0);
8380 }
8381
8382 // Trim our work_queue so its length is below max at return
8383 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8384 while (_work_queue->size() > max) {
8385 oop new_oop;
8386 if (_work_queue->pop_local(new_oop)) {
8387 assert(new_oop->is_oop(), "Expected an oop");
8388 assert(_bit_map->isMarked((HeapWord*)new_oop),
8389 "no white objects on this stack!");
8390 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8391 // iterate over the oops in this oop, marking and pushing
8392 // the ones in CMS heap (i.e. in _span).
8393 new_oop->oop_iterate(&_mark_and_push);
8394 }
8395 }
8396 }
8397
8398 ////////////////////////////////////////////////////////////////////
8399 // Support for Marking Stack Overflow list handling and related code
8400 ////////////////////////////////////////////////////////////////////
8401 // Much of the following code is similar in shape and spirit to the
8402 // code used in ParNewGC. We should try and share that code
8403 // as much as possible in the future.
8404
8405 #ifndef PRODUCT
8406 // Debugging support for CMSStackOverflowALot
8407
8408 // It's OK to call this multi-threaded; the worst thing
8409 // that can happen is that we'll get a bunch of closely
8410 // spaced simulated oveflows, but that's OK, in fact
8411 // probably good as it would exercise the overflow code
8412 // under contention.
8413 bool CMSCollector::simulate_overflow() {
8414 if (_overflow_counter-- <= 0) { // just being defensive
8415 _overflow_counter = CMSMarkStackOverflowInterval;
8416 return true;
8417 } else {
8418 return false;
8419 }
8420 }
8421
8422 bool CMSCollector::par_simulate_overflow() {
8423 return simulate_overflow();
8424 }
8425 #endif
8426
8427 // Single-threaded
8428 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8429 assert(stack->isEmpty(), "Expected precondition");
8430 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8431 size_t i = num;
8432 oop cur = _overflow_list;
8433 const markOop proto = markOopDesc::prototype();
8434 NOT_PRODUCT(size_t n = 0;)
8435 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8436 next = oop(cur->mark());
8437 cur->set_mark(proto); // until proven otherwise
8438 assert(cur->is_oop(), "Should be an oop");
8439 bool res = stack->push(cur);
8440 assert(res, "Bit off more than can chew?");
8441 NOT_PRODUCT(n++;)
8442 }
8443 _overflow_list = cur;
8444 #ifndef PRODUCT
8445 assert(_num_par_pushes >= n, "Too many pops?");
8446 _num_par_pushes -=n;
8447 #endif
8448 return !stack->isEmpty();
8449 }
8450
8451 // Multi-threaded; use CAS to break off a prefix
8452 bool CMSCollector::par_take_from_overflow_list(size_t num,
8453 OopTaskQueue* work_q) {
8454 assert(work_q->size() == 0, "That's the current policy");
8455 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8456 if (_overflow_list == NULL) {
8457 return false;
8458 }
8459 // Grab the entire list; we'll put back a suffix
8460 oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
8461 if (prefix == NULL) { // someone grabbed it before we did ...
8462 // ... we could spin for a short while, but for now we don't
8463 return false;
8464 }
8465 size_t i = num;
8466 oop cur = prefix;
8467 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8468 if (cur->mark() != NULL) {
8469 oop suffix_head = cur->mark(); // suffix will be put back on global list
8470 cur->set_mark(NULL); // break off suffix
8471 // Find tail of suffix so we can prepend suffix to global list
8472 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8473 oop suffix_tail = cur;
8474 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8475 "Tautology");
8476 oop observed_overflow_list = _overflow_list;
8477 do {
8478 cur = observed_overflow_list;
8479 suffix_tail->set_mark(markOop(cur));
8480 observed_overflow_list =
8481 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
8482 } while (cur != observed_overflow_list);
8483 }
8484
8485 // Push the prefix elements on work_q
8486 assert(prefix != NULL, "control point invariant");
8487 const markOop proto = markOopDesc::prototype();
8488 oop next;
8489 NOT_PRODUCT(size_t n = 0;)
8490 for (cur = prefix; cur != NULL; cur = next) {
8491 next = oop(cur->mark());
8492 cur->set_mark(proto); // until proven otherwise
8493 assert(cur->is_oop(), "Should be an oop");
8494 bool res = work_q->push(cur);
8495 assert(res, "Bit off more than we can chew?");
8496 NOT_PRODUCT(n++;)
8497 }
8498 #ifndef PRODUCT
8499 assert(_num_par_pushes >= n, "Too many pops?");
8500 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8501 #endif
8502 return true;
8503 }
8504
8505 // Single-threaded
8506 void CMSCollector::push_on_overflow_list(oop p) {
8507 NOT_PRODUCT(_num_par_pushes++;)
8508 assert(p->is_oop(), "Not an oop");
8509 preserve_mark_if_necessary(p);
8510 p->set_mark((markOop)_overflow_list);
8511 _overflow_list = p;
8512 }
8513
8514 // Multi-threaded; use CAS to prepend to overflow list
8515 void CMSCollector::par_push_on_overflow_list(oop p) {
8516 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
8517 assert(p->is_oop(), "Not an oop");
8518 par_preserve_mark_if_necessary(p);
8519 oop observed_overflow_list = _overflow_list;
8520 oop cur_overflow_list;
8521 do {
8522 cur_overflow_list = observed_overflow_list;
8523 p->set_mark(markOop(cur_overflow_list));
8524 observed_overflow_list =
8525 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
8526 } while (cur_overflow_list != observed_overflow_list);
8527 }
8528
8529 // Single threaded
8530 // General Note on GrowableArray: pushes may silently fail
8531 // because we are (temporarily) out of C-heap for expanding
8532 // the stack. The problem is quite ubiquitous and affects
8533 // a lot of code in the JVM. The prudent thing for GrowableArray
8534 // to do (for now) is to exit with an error. However, that may
8535 // be too draconian in some cases because the caller may be
8536 // able to recover without much harm. For suych cases, we
8537 // should probably introduce a "soft_push" method which returns
8538 // an indication of success or failure with the assumption that
8539 // the caller may be able to recover from a failure; code in
8540 // the VM can then be changed, incrementally, to deal with such
8541 // failures where possible, thus, incrementally hardening the VM
8542 // in such low resource situations.
8543 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8544 int PreserveMarkStackSize = 128;
8545
8546 if (_preserved_oop_stack == NULL) {
8547 assert(_preserved_mark_stack == NULL,
8548 "bijection with preserved_oop_stack");
8549 // Allocate the stacks
8550 _preserved_oop_stack = new (ResourceObj::C_HEAP)
8551 GrowableArray<oop>(PreserveMarkStackSize, true);
8552 _preserved_mark_stack = new (ResourceObj::C_HEAP)
8553 GrowableArray<markOop>(PreserveMarkStackSize, true);
8554 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
8555 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
8556 "Preserved Mark/Oop Stack for CMS (C-heap)");
8557 }
8558 }
8559 _preserved_oop_stack->push(p);
8560 _preserved_mark_stack->push(m);
8561 assert(m == p->mark(), "Mark word changed");
8562 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
8563 "bijection");
8564 }
8565
8566 // Single threaded
8567 void CMSCollector::preserve_mark_if_necessary(oop p) {
8568 markOop m = p->mark();
8569 if (m->must_be_preserved(p)) {
8570 preserve_mark_work(p, m);
8571 }
8572 }
8573
8574 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8575 markOop m = p->mark();
8576 if (m->must_be_preserved(p)) {
8577 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8578 // Even though we read the mark word without holding
8579 // the lock, we are assured that it will not change
8580 // because we "own" this oop, so no other thread can
8581 // be trying to push it on the overflow list; see
8582 // the assertion in preserve_mark_work() that checks
8583 // that m == p->mark().
8584 preserve_mark_work(p, m);
8585 }
8586 }
8587
8588 // We should be able to do this multi-threaded,
8589 // a chunk of stack being a task (this is
8590 // correct because each oop only ever appears
8591 // once in the overflow list. However, it's
8592 // not very easy to completely overlap this with
8593 // other operations, so will generally not be done
8594 // until all work's been completed. Because we
8595 // expect the preserved oop stack (set) to be small,
8596 // it's probably fine to do this single-threaded.
8597 // We can explore cleverer concurrent/overlapped/parallel
8598 // processing of preserved marks if we feel the
8599 // need for this in the future. Stack overflow should
8600 // be so rare in practice and, when it happens, its
8601 // effect on performance so great that this will
8602 // likely just be in the noise anyway.
8603 void CMSCollector::restore_preserved_marks_if_any() {
8604 if (_preserved_oop_stack == NULL) {
8605 assert(_preserved_mark_stack == NULL,
8606 "bijection with preserved_oop_stack");
8607 return;
8608 }
8609
8610 assert(SafepointSynchronize::is_at_safepoint(),
8611 "world should be stopped");
8612 assert(Thread::current()->is_ConcurrentGC_thread() ||
8613 Thread::current()->is_VM_thread(),
8614 "should be single-threaded");
8615
8616 int length = _preserved_oop_stack->length();
8617 assert(_preserved_mark_stack->length() == length, "bijection");
8618 for (int i = 0; i < length; i++) {
8619 oop p = _preserved_oop_stack->at(i);
8620 assert(p->is_oop(), "Should be an oop");
8621 assert(_span.contains(p), "oop should be in _span");
8622 assert(p->mark() == markOopDesc::prototype(),
8623 "Set when taken from overflow list");
8624 markOop m = _preserved_mark_stack->at(i);
8625 p->set_mark(m);
8626 }
8627 _preserved_mark_stack->clear();
8628 _preserved_oop_stack->clear();
8629 assert(_preserved_mark_stack->is_empty() &&
8630 _preserved_oop_stack->is_empty(),
8631 "stacks were cleared above");
8632 }
8633
8634 #ifndef PRODUCT
8635 bool CMSCollector::no_preserved_marks() const {
8636 return ( ( _preserved_mark_stack == NULL
8637 && _preserved_oop_stack == NULL)
8638 || ( _preserved_mark_stack->is_empty()
8639 && _preserved_oop_stack->is_empty()));
8640 }
8641 #endif
8642
8643 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
8644 {
8645 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8646 CMSAdaptiveSizePolicy* size_policy =
8647 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
8648 assert(size_policy->is_gc_cms_adaptive_size_policy(),
8649 "Wrong type for size policy");
8650 return size_policy;
8651 }
8652
8653 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
8654 size_t desired_promo_size) {
8655 if (cur_promo_size < desired_promo_size) {
8656 size_t expand_bytes = desired_promo_size - cur_promo_size;
8657 if (PrintAdaptiveSizePolicy && Verbose) {
8658 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8659 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
8660 expand_bytes);
8661 }
8662 expand(expand_bytes,
8663 MinHeapDeltaBytes,
8664 CMSExpansionCause::_adaptive_size_policy);
8665 } else if (desired_promo_size < cur_promo_size) {
8666 size_t shrink_bytes = cur_promo_size - desired_promo_size;
8667 if (PrintAdaptiveSizePolicy && Verbose) {
8668 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
8669 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
8670 shrink_bytes);
8671 }
8672 shrink(shrink_bytes);
8673 }
8674 }
8675
8676 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
8677 GenCollectedHeap* gch = GenCollectedHeap::heap();
8678 CMSGCAdaptivePolicyCounters* counters =
8679 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
8680 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
8681 "Wrong kind of counters");
8682 return counters;
8683 }
8684
8685
8686 void ASConcurrentMarkSweepGeneration::update_counters() {
8687 if (UsePerfData) {
8688 _space_counters->update_all();
8689 _gen_counters->update_all();
8690 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8691 GenCollectedHeap* gch = GenCollectedHeap::heap();
8692 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8693 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8694 "Wrong gc statistics type");
8695 counters->update_counters(gc_stats_l);
8696 }
8697 }
8698
8699 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
8700 if (UsePerfData) {
8701 _space_counters->update_used(used);
8702 _space_counters->update_capacity();
8703 _gen_counters->update_all();
8704
8705 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8706 GenCollectedHeap* gch = GenCollectedHeap::heap();
8707 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
8708 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
8709 "Wrong gc statistics type");
8710 counters->update_counters(gc_stats_l);
8711 }
8712 }
8713
8714 // The desired expansion delta is computed so that:
8715 // . desired free percentage or greater is used
8716 void ASConcurrentMarkSweepGeneration::compute_new_size() {
8717 assert_locked_or_safepoint(Heap_lock);
8718
8719 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
8720
8721 // If incremental collection failed, we just want to expand
8722 // to the limit.
8723 if (incremental_collection_failed()) {
8724 clear_incremental_collection_failed();
8725 grow_to_reserved();
8726 return;
8727 }
8728
8729 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
8730
8731 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
8732 "Wrong type of heap");
8733 int prev_level = level() - 1;
8734 assert(prev_level >= 0, "The cms generation is the lowest generation");
8735 Generation* prev_gen = gch->get_gen(prev_level);
8736 assert(prev_gen->kind() == Generation::ASParNew,
8737 "Wrong type of young generation");
8738 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
8739 size_t cur_eden = younger_gen->eden()->capacity();
8740 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
8741 size_t cur_promo = free();
8742 size_policy->compute_tenured_generation_free_space(cur_promo,
8743 max_available(),
8744 cur_eden);
8745 resize(cur_promo, size_policy->promo_size());
8746
8747 // Record the new size of the space in the cms generation
8748 // that is available for promotions. This is temporary.
8749 // It should be the desired promo size.
8750 size_policy->avg_cms_promo()->sample(free());
8751 size_policy->avg_old_live()->sample(used());
8752
8753 if (UsePerfData) {
8754 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
8755 counters->update_cms_capacity_counter(capacity());
8756 }
8757 }
8758
8759 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
8760 assert_locked_or_safepoint(Heap_lock);
8761 assert_lock_strong(freelistLock());
8762 HeapWord* old_end = _cmsSpace->end();
8763 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
8764 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
8765 FreeChunk* chunk_at_end = find_chunk_at_end();
8766 if (chunk_at_end == NULL) {
8767 // No room to shrink
8768 if (PrintGCDetails && Verbose) {
8769 gclog_or_tty->print_cr("No room to shrink: old_end "
8770 PTR_FORMAT " unallocated_start " PTR_FORMAT
8771 " chunk_at_end " PTR_FORMAT,
8772 old_end, unallocated_start, chunk_at_end);
8773 }
8774 return;
8775 } else {
8776
8777 // Find the chunk at the end of the space and determine
8778 // how much it can be shrunk.
8779 size_t shrinkable_size_in_bytes = chunk_at_end->size();
8780 size_t aligned_shrinkable_size_in_bytes =
8781 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
8782 assert(unallocated_start <= chunk_at_end->end(),
8783 "Inconsistent chunk at end of space");
8784 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
8785 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
8786
8787 // Shrink the underlying space
8788 _virtual_space.shrink_by(bytes);
8789 if (PrintGCDetails && Verbose) {
8790 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
8791 " desired_bytes " SIZE_FORMAT
8792 " shrinkable_size_in_bytes " SIZE_FORMAT
8793 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
8794 " bytes " SIZE_FORMAT,
8795 desired_bytes, shrinkable_size_in_bytes,
8796 aligned_shrinkable_size_in_bytes, bytes);
8797 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
8798 " unallocated_start " SIZE_FORMAT,
8799 old_end, unallocated_start);
8800 }
8801
8802 // If the space did shrink (shrinking is not guaranteed),
8803 // shrink the chunk at the end by the appropriate amount.
8804 if (((HeapWord*)_virtual_space.high()) < old_end) {
8805 size_t new_word_size =
8806 heap_word_size(_virtual_space.committed_size());
8807
8808 // Have to remove the chunk from the dictionary because it is changing
8809 // size and might be someplace elsewhere in the dictionary.
8810
8811 // Get the chunk at end, shrink it, and put it
8812 // back.
8813 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
8814 size_t word_size_change = word_size_before - new_word_size;
8815 size_t chunk_at_end_old_size = chunk_at_end->size();
8816 assert(chunk_at_end_old_size >= word_size_change,
8817 "Shrink is too large");
8818 chunk_at_end->setSize(chunk_at_end_old_size -
8819 word_size_change);
8820 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
8821 word_size_change);
8822
8823 _cmsSpace->returnChunkToDictionary(chunk_at_end);
8824
8825 MemRegion mr(_cmsSpace->bottom(), new_word_size);
8826 _bts->resize(new_word_size); // resize the block offset shared array
8827 Universe::heap()->barrier_set()->resize_covered_region(mr);
8828 _cmsSpace->assert_locked();
8829 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
8830
8831 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
8832
8833 // update the space and generation capacity counters
8834 if (UsePerfData) {
8835 _space_counters->update_capacity();
8836 _gen_counters->update_all();
8837 }
8838
8839 if (Verbose && PrintGCDetails) {
8840 size_t new_mem_size = _virtual_space.committed_size();
8841 size_t old_mem_size = new_mem_size + bytes;
8842 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
8843 name(), old_mem_size/K, bytes/K, new_mem_size/K);
8844 }
8845 }
8846
8847 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
8848 "Inconsistency at end of space");
8849 assert(chunk_at_end->end() == _cmsSpace->end(),
8850 "Shrinking is inconsistent");
8851 return;
8852 }
8853 }
8854
8855 // Transfer some number of overflown objects to usual marking
8856 // stack. Return true if some objects were transferred.
8857 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8858 size_t num = MIN2((size_t)_mark_stack->capacity()/4,
8859 (size_t)ParGCDesiredObjsFromOverflowList);
8860
8861 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8862 assert(_collector->overflow_list_is_empty() || res,
8863 "If list is not empty, we should have taken something");
8864 assert(!res || !_mark_stack->isEmpty(),
8865 "If we took something, it should now be on our stack");
8866 return res;
8867 }
8868
8869 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8870 size_t res = _sp->block_size_no_stall(addr, _collector);
8871 assert(res != 0, "Should always be able to compute a size");
8872 if (_sp->block_is_obj(addr)) {
8873 if (_live_bit_map->isMarked(addr)) {
8874 // It can't have been dead in a previous cycle
8875 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8876 } else {
8877 _dead_bit_map->mark(addr); // mark the dead object
8878 }
8879 }
8880 return res;
8881 }