1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/symbolTable.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "code/codeCache.hpp"
29 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
30 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
37 #include "gc_implementation/parNew/parNewGeneration.hpp"
38 #include "gc_implementation/shared/collectorCounters.hpp"
39 #include "gc_implementation/shared/gcTimer.hpp"
40 #include "gc_implementation/shared/gcTrace.hpp"
41 #include "gc_implementation/shared/gcTraceTime.hpp"
42 #include "gc_implementation/shared/isGCActiveMark.hpp"
43 #include "gc_interface/collectedHeap.inline.hpp"
44 #include "memory/allocation.hpp"
45 #include "memory/cardTableRS.hpp"
46 #include "memory/collectorPolicy.hpp"
47 #include "memory/gcLocker.inline.hpp"
48 #include "memory/genCollectedHeap.hpp"
49 #include "memory/genMarkSweep.hpp"
50 #include "memory/genOopClosures.inline.hpp"
51 #include "memory/iterator.hpp"
52 #include "memory/referencePolicy.hpp"
53 #include "memory/resourceArea.hpp"
54 #include "oops/oop.inline.hpp"
55 #include "prims/jvmtiExport.hpp"
56 #include "runtime/globals_extension.hpp"
57 #include "runtime/handles.inline.hpp"
58 #include "runtime/java.hpp"
59 #include "runtime/vmThread.hpp"
60 #include "services/memoryService.hpp"
61 #include "services/runtimeService.hpp"
62
63 // statics
64 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
65 bool CMSCollector::_full_gc_requested = false;
66
67 //////////////////////////////////////////////////////////////////
68 // In support of CMS/VM thread synchronization
69 //////////////////////////////////////////////////////////////////
70 // We split use of the CGC_lock into 2 "levels".
71 // The low-level locking is of the usual CGC_lock monitor. We introduce
72 // a higher level "token" (hereafter "CMS token") built on top of the
73 // low level monitor (hereafter "CGC lock").
74 // The token-passing protocol gives priority to the VM thread. The
75 // CMS-lock doesn't provide any fairness guarantees, but clients
76 // should ensure that it is only held for very short, bounded
77 // durations.
78 //
79 // When either of the CMS thread or the VM thread is involved in
80 // collection operations during which it does not want the other
81 // thread to interfere, it obtains the CMS token.
82 //
83 // If either thread tries to get the token while the other has
84 // it, that thread waits. However, if the VM thread and CMS thread
85 // both want the token, then the VM thread gets priority while the
86 // CMS thread waits. This ensures, for instance, that the "concurrent"
87 // phases of the CMS thread's work do not block out the VM thread
88 // for long periods of time as the CMS thread continues to hog
89 // the token. (See bug 4616232).
90 //
91 // The baton-passing functions are, however, controlled by the
92 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
93 // and here the low-level CMS lock, not the high level token,
94 // ensures mutual exclusion.
95 //
96 // Two important conditions that we have to satisfy:
97 // 1. if a thread does a low-level wait on the CMS lock, then it
98 // relinquishes the CMS token if it were holding that token
99 // when it acquired the low-level CMS lock.
100 // 2. any low-level notifications on the low-level lock
101 // should only be sent when a thread has relinquished the token.
102 //
103 // In the absence of either property, we'd have potential deadlock.
104 //
105 // We protect each of the CMS (concurrent and sequential) phases
106 // with the CMS _token_, not the CMS _lock_.
107 //
108 // The only code protected by CMS lock is the token acquisition code
109 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
110 // baton-passing code.
111 //
112 // Unfortunately, i couldn't come up with a good abstraction to factor and
113 // hide the naked CGC_lock manipulation in the baton-passing code
114 // further below. That's something we should try to do. Also, the proof
115 // of correctness of this 2-level locking scheme is far from obvious,
116 // and potentially quite slippery. We have an uneasy supsicion, for instance,
117 // that there may be a theoretical possibility of delay/starvation in the
118 // low-level lock/wait/notify scheme used for the baton-passing because of
119 // potential intereference with the priority scheme embodied in the
120 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
121 // invocation further below and marked with "XXX 20011219YSR".
122 // Indeed, as we note elsewhere, this may become yet more slippery
123 // in the presence of multiple CMS and/or multiple VM threads. XXX
124
125 class CMSTokenSync: public StackObj {
126 private:
127 bool _is_cms_thread;
128 public:
129 CMSTokenSync(bool is_cms_thread):
130 _is_cms_thread(is_cms_thread) {
131 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
132 "Incorrect argument to constructor");
133 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
134 }
135
136 ~CMSTokenSync() {
137 assert(_is_cms_thread ?
138 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
139 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
140 "Incorrect state");
141 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
142 }
143 };
144
145 // Convenience class that does a CMSTokenSync, and then acquires
146 // upto three locks.
147 class CMSTokenSyncWithLocks: public CMSTokenSync {
148 private:
149 // Note: locks are acquired in textual declaration order
150 // and released in the opposite order
151 MutexLockerEx _locker1, _locker2, _locker3;
152 public:
153 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
154 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
155 CMSTokenSync(is_cms_thread),
156 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
157 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
158 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
159 { }
160 };
161
162
163 // Wrapper class to temporarily disable icms during a foreground cms collection.
164 class ICMSDisabler: public StackObj {
165 public:
166 // The ctor disables icms and wakes up the thread so it notices the change;
167 // the dtor re-enables icms. Note that the CMSCollector methods will check
168 // CMSIncrementalMode.
169 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
170 ~ICMSDisabler() { CMSCollector::enable_icms(); }
171 };
172
173 //////////////////////////////////////////////////////////////////
174 // Concurrent Mark-Sweep Generation /////////////////////////////
175 //////////////////////////////////////////////////////////////////
176
177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
178
179 // This struct contains per-thread things necessary to support parallel
180 // young-gen collection.
181 class CMSParGCThreadState: public CHeapObj<mtGC> {
182 public:
183 CFLS_LAB lab;
184 PromotionInfo promo;
185
186 // Constructor.
187 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
188 promo.setSpace(cfls);
189 }
190 };
191
192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
193 ReservedSpace rs, size_t initial_byte_size, int level,
194 CardTableRS* ct, bool use_adaptive_freelists,
195 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
196 CardGeneration(rs, initial_byte_size, level, ct),
197 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
198 _debug_collection_type(Concurrent_collection_type)
199 {
200 HeapWord* bottom = (HeapWord*) _virtual_space.low();
201 HeapWord* end = (HeapWord*) _virtual_space.high();
202
203 _direct_allocated_words = 0;
204 NOT_PRODUCT(
205 _numObjectsPromoted = 0;
206 _numWordsPromoted = 0;
207 _numObjectsAllocated = 0;
208 _numWordsAllocated = 0;
209 )
210
211 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
212 use_adaptive_freelists,
213 dictionaryChoice);
214 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
215 if (_cmsSpace == NULL) {
216 vm_exit_during_initialization(
217 "CompactibleFreeListSpace allocation failure");
218 }
219 _cmsSpace->_gen = this;
220
221 _gc_stats = new CMSGCStats();
222
223 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
224 // offsets match. The ability to tell free chunks from objects
225 // depends on this property.
226 debug_only(
227 FreeChunk* junk = NULL;
228 assert(UseCompressedOops ||
229 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
230 "Offset of FreeChunk::_prev within FreeChunk must match"
231 " that of OopDesc::_klass within OopDesc");
232 )
233 if (CollectedHeap::use_parallel_gc_threads()) {
234 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
235 _par_gc_thread_states =
236 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
237 if (_par_gc_thread_states == NULL) {
238 vm_exit_during_initialization("Could not allocate par gc structs");
239 }
240 for (uint i = 0; i < ParallelGCThreads; i++) {
241 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
242 if (_par_gc_thread_states[i] == NULL) {
243 vm_exit_during_initialization("Could not allocate par gc structs");
244 }
245 }
246 } else {
247 _par_gc_thread_states = NULL;
248 }
249 _incremental_collection_failed = false;
250 // The "dilatation_factor" is the expansion that can occur on
251 // account of the fact that the minimum object size in the CMS
252 // generation may be larger than that in, say, a contiguous young
253 // generation.
254 // Ideally, in the calculation below, we'd compute the dilatation
255 // factor as: MinChunkSize/(promoting_gen's min object size)
256 // Since we do not have such a general query interface for the
257 // promoting generation, we'll instead just use the mimimum
258 // object size (which today is a header's worth of space);
259 // note that all arithmetic is in units of HeapWords.
260 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
261 assert(_dilatation_factor >= 1.0, "from previous assert");
262 }
263
264
265 // The field "_initiating_occupancy" represents the occupancy percentage
266 // at which we trigger a new collection cycle. Unless explicitly specified
267 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it
268 // is calculated by:
269 //
270 // Let "f" be MinHeapFreeRatio in
271 //
272 // _intiating_occupancy = 100-f +
273 // f * (CMSTrigger[Perm]Ratio/100)
274 // where CMSTrigger[Perm]Ratio is the argument "tr" below.
275 //
276 // That is, if we assume the heap is at its desired maximum occupancy at the
277 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free
278 // space be allocated before initiating a new collection cycle.
279 //
280 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) {
281 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments");
282 if (io >= 0) {
283 _initiating_occupancy = (double)io / 100.0;
284 } else {
285 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
286 (double)(tr * MinHeapFreeRatio) / 100.0)
287 / 100.0;
288 }
289 }
290
291 void ConcurrentMarkSweepGeneration::ref_processor_init() {
292 assert(collector() != NULL, "no collector");
293 collector()->ref_processor_init();
294 }
295
296 void CMSCollector::ref_processor_init() {
297 if (_ref_processor == NULL) {
298 // Allocate and initialize a reference processor
299 _ref_processor =
300 new ReferenceProcessor(_span, // span
301 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
302 (int) ParallelGCThreads, // mt processing degree
303 _cmsGen->refs_discovery_is_mt(), // mt discovery
304 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
305 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
306 &_is_alive_closure, // closure for liveness info
307 false); // next field updates do not need write barrier
308 // Initialize the _ref_processor field of CMSGen
309 _cmsGen->set_ref_processor(_ref_processor);
310
311 // Allocate a dummy ref processor for perm gen.
312 ReferenceProcessor* rp2 = new ReferenceProcessor();
313 if (rp2 == NULL) {
314 vm_exit_during_initialization("Could not allocate ReferenceProcessor object");
315 }
316 _permGen->set_ref_processor(rp2);
317 }
318 }
319
320 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
321 GenCollectedHeap* gch = GenCollectedHeap::heap();
322 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
323 "Wrong type of heap");
324 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
325 gch->gen_policy()->size_policy();
326 assert(sp->is_gc_cms_adaptive_size_policy(),
327 "Wrong type of size policy");
328 return sp;
329 }
330
331 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
332 CMSGCAdaptivePolicyCounters* results =
333 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
334 assert(
335 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
336 "Wrong gc policy counter kind");
337 return results;
338 }
339
340
341 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
342
343 const char* gen_name = "old";
344
345 // Generation Counters - generation 1, 1 subspace
346 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
347
348 _space_counters = new GSpaceCounters(gen_name, 0,
349 _virtual_space.reserved_size(),
350 this, _gen_counters);
351 }
352
353 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
354 _cms_gen(cms_gen)
355 {
356 assert(alpha <= 100, "bad value");
357 _saved_alpha = alpha;
358
359 // Initialize the alphas to the bootstrap value of 100.
360 _gc0_alpha = _cms_alpha = 100;
361
362 _cms_begin_time.update();
363 _cms_end_time.update();
364
365 _gc0_duration = 0.0;
366 _gc0_period = 0.0;
367 _gc0_promoted = 0;
368
369 _cms_duration = 0.0;
370 _cms_period = 0.0;
371 _cms_allocated = 0;
372
373 _cms_used_at_gc0_begin = 0;
374 _cms_used_at_gc0_end = 0;
375 _allow_duty_cycle_reduction = false;
376 _valid_bits = 0;
377 _icms_duty_cycle = CMSIncrementalDutyCycle;
378 }
379
380 double CMSStats::cms_free_adjustment_factor(size_t free) const {
381 // TBD: CR 6909490
382 return 1.0;
383 }
384
385 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
386 }
387
388 // If promotion failure handling is on use
389 // the padded average size of the promotion for each
390 // young generation collection.
391 double CMSStats::time_until_cms_gen_full() const {
392 size_t cms_free = _cms_gen->cmsSpace()->free();
393 GenCollectedHeap* gch = GenCollectedHeap::heap();
394 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
395 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
396 if (cms_free > expected_promotion) {
397 // Start a cms collection if there isn't enough space to promote
398 // for the next minor collection. Use the padded average as
399 // a safety factor.
400 cms_free -= expected_promotion;
401
402 // Adjust by the safety factor.
403 double cms_free_dbl = (double)cms_free;
404 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
405 // Apply a further correction factor which tries to adjust
406 // for recent occurance of concurrent mode failures.
407 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
408 cms_free_dbl = cms_free_dbl * cms_adjustment;
409
410 if (PrintGCDetails && Verbose) {
411 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
412 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
413 cms_free, expected_promotion);
414 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
415 cms_free_dbl, cms_consumption_rate() + 1.0);
416 }
417 // Add 1 in case the consumption rate goes to zero.
418 return cms_free_dbl / (cms_consumption_rate() + 1.0);
419 }
420 return 0.0;
421 }
422
423 // Compare the duration of the cms collection to the
424 // time remaining before the cms generation is empty.
425 // Note that the time from the start of the cms collection
426 // to the start of the cms sweep (less than the total
427 // duration of the cms collection) can be used. This
428 // has been tried and some applications experienced
429 // promotion failures early in execution. This was
430 // possibly because the averages were not accurate
431 // enough at the beginning.
432 double CMSStats::time_until_cms_start() const {
433 // We add "gc0_period" to the "work" calculation
434 // below because this query is done (mostly) at the
435 // end of a scavenge, so we need to conservatively
436 // account for that much possible delay
437 // in the query so as to avoid concurrent mode failures
438 // due to starting the collection just a wee bit too
439 // late.
440 double work = cms_duration() + gc0_period();
441 double deadline = time_until_cms_gen_full();
442 // If a concurrent mode failure occurred recently, we want to be
443 // more conservative and halve our expected time_until_cms_gen_full()
444 if (work > deadline) {
445 if (Verbose && PrintGCDetails) {
446 gclog_or_tty->print(
447 " CMSCollector: collect because of anticipated promotion "
448 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
449 gc0_period(), time_until_cms_gen_full());
450 }
451 return 0.0;
452 }
453 return work - deadline;
454 }
455
456 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
457 // amount of change to prevent wild oscillation.
458 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
459 unsigned int new_duty_cycle) {
460 assert(old_duty_cycle <= 100, "bad input value");
461 assert(new_duty_cycle <= 100, "bad input value");
462
463 // Note: use subtraction with caution since it may underflow (values are
464 // unsigned). Addition is safe since we're in the range 0-100.
465 unsigned int damped_duty_cycle = new_duty_cycle;
466 if (new_duty_cycle < old_duty_cycle) {
467 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
468 if (new_duty_cycle + largest_delta < old_duty_cycle) {
469 damped_duty_cycle = old_duty_cycle - largest_delta;
470 }
471 } else if (new_duty_cycle > old_duty_cycle) {
472 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
473 if (new_duty_cycle > old_duty_cycle + largest_delta) {
474 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
475 }
476 }
477 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
478
479 if (CMSTraceIncrementalPacing) {
480 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
481 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
482 }
483 return damped_duty_cycle;
484 }
485
486 unsigned int CMSStats::icms_update_duty_cycle_impl() {
487 assert(CMSIncrementalPacing && valid(),
488 "should be handled in icms_update_duty_cycle()");
489
490 double cms_time_so_far = cms_timer().seconds();
491 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
492 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
493
494 // Avoid division by 0.
495 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
496 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
497
498 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
499 if (new_duty_cycle > _icms_duty_cycle) {
500 // Avoid very small duty cycles (1 or 2); 0 is allowed.
501 if (new_duty_cycle > 2) {
502 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
503 new_duty_cycle);
504 }
505 } else if (_allow_duty_cycle_reduction) {
506 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
507 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
508 // Respect the minimum duty cycle.
509 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
510 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
511 }
512
513 if (PrintGCDetails || CMSTraceIncrementalPacing) {
514 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
515 }
516
517 _allow_duty_cycle_reduction = false;
518 return _icms_duty_cycle;
519 }
520
521 #ifndef PRODUCT
522 void CMSStats::print_on(outputStream *st) const {
523 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
524 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
525 gc0_duration(), gc0_period(), gc0_promoted());
526 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
527 cms_duration(), cms_duration_per_mb(),
528 cms_period(), cms_allocated());
529 st->print(",cms_since_beg=%g,cms_since_end=%g",
530 cms_time_since_begin(), cms_time_since_end());
531 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
532 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
533 if (CMSIncrementalMode) {
534 st->print(",dc=%d", icms_duty_cycle());
535 }
536
537 if (valid()) {
538 st->print(",promo_rate=%g,cms_alloc_rate=%g",
539 promotion_rate(), cms_allocation_rate());
540 st->print(",cms_consumption_rate=%g,time_until_full=%g",
541 cms_consumption_rate(), time_until_cms_gen_full());
542 }
543 st->print(" ");
544 }
545 #endif // #ifndef PRODUCT
546
547 CMSCollector::CollectorState CMSCollector::_collectorState =
548 CMSCollector::Idling;
549 bool CMSCollector::_foregroundGCIsActive = false;
550 bool CMSCollector::_foregroundGCShouldWait = false;
551
552 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
553 ConcurrentMarkSweepGeneration* permGen,
554 CardTableRS* ct,
555 ConcurrentMarkSweepPolicy* cp):
556 _cmsGen(cmsGen),
557 _permGen(permGen),
558 _ct(ct),
559 _ref_processor(NULL), // will be set later
560 _conc_workers(NULL), // may be set later
561 _abort_preclean(false),
562 _start_sampling(false),
563 _between_prologue_and_epilogue(false),
564 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
565 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"),
566 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
567 -1 /* lock-free */, "No_lock" /* dummy */),
568 _modUnionClosure(&_modUnionTable),
569 _modUnionClosurePar(&_modUnionTable),
570 // Adjust my span to cover old (cms) gen and perm gen
571 _span(cmsGen->reserved()._union(permGen->reserved())),
572 // Construct the is_alive_closure with _span & markBitMap
573 _is_alive_closure(_span, &_markBitMap),
574 _restart_addr(NULL),
575 _overflow_list(NULL),
576 _stats(cmsGen),
577 _eden_chunk_array(NULL), // may be set in ctor body
578 _eden_chunk_capacity(0), // -- ditto --
579 _eden_chunk_index(0), // -- ditto --
580 _survivor_plab_array(NULL), // -- ditto --
581 _survivor_chunk_array(NULL), // -- ditto --
582 _survivor_chunk_capacity(0), // -- ditto --
583 _survivor_chunk_index(0), // -- ditto --
584 _ser_pmc_preclean_ovflw(0),
585 _ser_kac_preclean_ovflw(0),
586 _ser_pmc_remark_ovflw(0),
587 _par_pmc_remark_ovflw(0),
588 _ser_kac_ovflw(0),
589 _par_kac_ovflw(0),
590 #ifndef PRODUCT
591 _num_par_pushes(0),
592 #endif
593 _collection_count_start(0),
594 _verifying(false),
595 _icms_start_limit(NULL),
596 _icms_stop_limit(NULL),
597 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
598 _completed_initialization(false),
599 _collector_policy(cp),
600 _should_unload_classes(false),
601 _concurrent_cycles_since_last_unload(0),
602 _roots_scanning_options(0),
603 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
604 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
605 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
606 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
607 _cms_start_registered(false)
608 {
609 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
610 ExplicitGCInvokesConcurrent = true;
611 }
612 // Now expand the span and allocate the collection support structures
613 // (MUT, marking bit map etc.) to cover both generations subject to
614 // collection.
615
616 // First check that _permGen is adjacent to _cmsGen and above it.
617 assert( _cmsGen->reserved().word_size() > 0
618 && _permGen->reserved().word_size() > 0,
619 "generations should not be of zero size");
620 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(),
621 "_cmsGen and _permGen should not overlap");
622 assert(_cmsGen->reserved().end() == _permGen->reserved().start(),
623 "_cmsGen->end() different from _permGen->start()");
624
625 // For use by dirty card to oop closures.
626 _cmsGen->cmsSpace()->set_collector(this);
627 _permGen->cmsSpace()->set_collector(this);
628
629 // Allocate MUT and marking bit map
630 {
631 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
632 if (!_markBitMap.allocate(_span)) {
633 warning("Failed to allocate CMS Bit Map");
634 return;
635 }
636 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
637 }
638 {
639 _modUnionTable.allocate(_span);
640 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
641 }
642
643 if (!_markStack.allocate(MarkStackSize)) {
644 warning("Failed to allocate CMS Marking Stack");
645 return;
646 }
647 if (!_revisitStack.allocate(CMSRevisitStackSize)) {
648 warning("Failed to allocate CMS Revisit Stack");
649 return;
650 }
651
652 // Support for multi-threaded concurrent phases
653 if (CMSConcurrentMTEnabled) {
654 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
655 // just for now
656 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
657 }
658 if (ConcGCThreads > 1) {
659 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
660 ConcGCThreads, true);
661 if (_conc_workers == NULL) {
662 warning("GC/CMS: _conc_workers allocation failure: "
663 "forcing -CMSConcurrentMTEnabled");
664 CMSConcurrentMTEnabled = false;
665 } else {
666 _conc_workers->initialize_workers();
667 }
668 } else {
669 CMSConcurrentMTEnabled = false;
670 }
671 }
672 if (!CMSConcurrentMTEnabled) {
673 ConcGCThreads = 0;
674 } else {
675 // Turn off CMSCleanOnEnter optimization temporarily for
676 // the MT case where it's not fixed yet; see 6178663.
677 CMSCleanOnEnter = false;
678 }
679 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
680 "Inconsistency");
681
682 // Parallel task queues; these are shared for the
683 // concurrent and stop-world phases of CMS, but
684 // are not shared with parallel scavenge (ParNew).
685 {
686 uint i;
687 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
688
689 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
690 || ParallelRefProcEnabled)
691 && num_queues > 0) {
692 _task_queues = new OopTaskQueueSet(num_queues);
693 if (_task_queues == NULL) {
694 warning("task_queues allocation failure.");
695 return;
696 }
697 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
698 if (_hash_seed == NULL) {
699 warning("_hash_seed array allocation failure");
700 return;
701 }
702
703 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
704 for (i = 0; i < num_queues; i++) {
705 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
706 if (q == NULL) {
707 warning("work_queue allocation failure.");
708 return;
709 }
710 _task_queues->register_queue(i, q);
711 }
712 for (i = 0; i < num_queues; i++) {
713 _task_queues->queue(i)->initialize();
714 _hash_seed[i] = 17; // copied from ParNew
715 }
716 }
717 }
718
719 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
720 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio);
721
722 // Clip CMSBootstrapOccupancy between 0 and 100.
723 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy)))
724 /(double)100;
725
726 _full_gcs_since_conc_gc = 0;
727
728 // Now tell CMS generations the identity of their collector
729 ConcurrentMarkSweepGeneration::set_collector(this);
730
731 // Create & start a CMS thread for this CMS collector
732 _cmsThread = ConcurrentMarkSweepThread::start(this);
733 assert(cmsThread() != NULL, "CMS Thread should have been created");
734 assert(cmsThread()->collector() == this,
735 "CMS Thread should refer to this gen");
736 assert(CGC_lock != NULL, "Where's the CGC_lock?");
737
738 // Support for parallelizing young gen rescan
739 GenCollectedHeap* gch = GenCollectedHeap::heap();
740 _young_gen = gch->prev_gen(_cmsGen);
741 if (gch->supports_inline_contig_alloc()) {
742 _top_addr = gch->top_addr();
743 _end_addr = gch->end_addr();
744 assert(_young_gen != NULL, "no _young_gen");
745 _eden_chunk_index = 0;
746 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
747 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
748 if (_eden_chunk_array == NULL) {
749 _eden_chunk_capacity = 0;
750 warning("GC/CMS: _eden_chunk_array allocation failure");
751 }
752 }
753 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
754
755 // Support for parallelizing survivor space rescan
756 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) {
757 const size_t max_plab_samples =
758 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
759
760 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
761 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
762 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
763 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
764 || _cursor == NULL) {
765 warning("Failed to allocate survivor plab/chunk array");
766 if (_survivor_plab_array != NULL) {
767 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
768 _survivor_plab_array = NULL;
769 }
770 if (_survivor_chunk_array != NULL) {
771 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
772 _survivor_chunk_array = NULL;
773 }
774 if (_cursor != NULL) {
775 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC);
776 _cursor = NULL;
777 }
778 } else {
779 _survivor_chunk_capacity = 2*max_plab_samples;
780 for (uint i = 0; i < ParallelGCThreads; i++) {
781 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
782 if (vec == NULL) {
783 warning("Failed to allocate survivor plab array");
784 for (int j = i; j > 0; j--) {
785 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC);
786 }
787 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
788 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
789 _survivor_plab_array = NULL;
790 _survivor_chunk_array = NULL;
791 _survivor_chunk_capacity = 0;
792 break;
793 } else {
794 ChunkArray* cur =
795 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
796 max_plab_samples);
797 assert(cur->end() == 0, "Should be 0");
798 assert(cur->array() == vec, "Should be vec");
799 assert(cur->capacity() == max_plab_samples, "Error");
800 }
801 }
802 }
803 }
804 assert( ( _survivor_plab_array != NULL
805 && _survivor_chunk_array != NULL)
806 || ( _survivor_chunk_capacity == 0
807 && _survivor_chunk_index == 0),
808 "Error");
809
810 // Choose what strong roots should be scanned depending on verification options
811 // and perm gen collection mode.
812 if (!CMSClassUnloadingEnabled) {
813 // If class unloading is disabled we want to include all classes into the root set.
814 add_root_scanning_option(SharedHeap::SO_AllClasses);
815 } else {
816 add_root_scanning_option(SharedHeap::SO_SystemClasses);
817 }
818
819 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
820 _gc_counters = new CollectorCounters("CMS", 1);
821 _completed_initialization = true;
822 _inter_sweep_timer.start(); // start of time
823 }
824
825 const char* ConcurrentMarkSweepGeneration::name() const {
826 return "concurrent mark-sweep generation";
827 }
828 void ConcurrentMarkSweepGeneration::update_counters() {
829 if (UsePerfData) {
830 _space_counters->update_all();
831 _gen_counters->update_all();
832 }
833 }
834
835 // this is an optimized version of update_counters(). it takes the
836 // used value as a parameter rather than computing it.
837 //
838 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
839 if (UsePerfData) {
840 _space_counters->update_used(used);
841 _space_counters->update_capacity();
842 _gen_counters->update_all();
843 }
844 }
845
846 void ConcurrentMarkSweepGeneration::print() const {
847 Generation::print();
848 cmsSpace()->print();
849 }
850
851 #ifndef PRODUCT
852 void ConcurrentMarkSweepGeneration::print_statistics() {
853 cmsSpace()->printFLCensus(0);
854 }
855 #endif
856
857 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
858 GenCollectedHeap* gch = GenCollectedHeap::heap();
859 if (PrintGCDetails) {
860 if (Verbose) {
861 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
862 level(), short_name(), s, used(), capacity());
863 } else {
864 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
865 level(), short_name(), s, used() / K, capacity() / K);
866 }
867 }
868 if (Verbose) {
869 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
870 gch->used(), gch->capacity());
871 } else {
872 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
873 gch->used() / K, gch->capacity() / K);
874 }
875 }
876
877 size_t
878 ConcurrentMarkSweepGeneration::contiguous_available() const {
879 // dld proposes an improvement in precision here. If the committed
880 // part of the space ends in a free block we should add that to
881 // uncommitted size in the calculation below. Will make this
882 // change later, staying with the approximation below for the
883 // time being. -- ysr.
884 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
885 }
886
887 size_t
888 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
889 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
890 }
891
892 size_t ConcurrentMarkSweepGeneration::max_available() const {
893 return free() + _virtual_space.uncommitted_size();
894 }
895
896 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
897 size_t available = max_available();
898 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
899 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
900 if (Verbose && PrintGCDetails) {
901 gclog_or_tty->print_cr(
902 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
903 "max_promo("SIZE_FORMAT")",
904 res? "":" not", available, res? ">=":"<",
905 av_promo, max_promotion_in_bytes);
906 }
907 return res;
908 }
909
910 // At a promotion failure dump information on block layout in heap
911 // (cms old generation).
912 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
913 if (CMSDumpAtPromotionFailure) {
914 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
915 }
916 }
917
918 CompactibleSpace*
919 ConcurrentMarkSweepGeneration::first_compaction_space() const {
920 return _cmsSpace;
921 }
922
923 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
924 // Clear the promotion information. These pointers can be adjusted
925 // along with all the other pointers into the heap but
926 // compaction is expected to be a rare event with
927 // a heap using cms so don't do it without seeing the need.
928 if (CollectedHeap::use_parallel_gc_threads()) {
929 for (uint i = 0; i < ParallelGCThreads; i++) {
930 _par_gc_thread_states[i]->promo.reset();
931 }
932 }
933 }
934
935 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
936 blk->do_space(_cmsSpace);
937 }
938
939 void ConcurrentMarkSweepGeneration::compute_new_size() {
940 assert_locked_or_safepoint(Heap_lock);
941
942 // If incremental collection failed, we just want to expand
943 // to the limit.
944 if (incremental_collection_failed()) {
945 clear_incremental_collection_failed();
946 grow_to_reserved();
947 return;
948 }
949
950 size_t expand_bytes = 0;
951 double free_percentage = ((double) free()) / capacity();
952 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
953 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
954
955 // compute expansion delta needed for reaching desired free percentage
956 if (free_percentage < desired_free_percentage) {
957 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
958 assert(desired_capacity >= capacity(), "invalid expansion size");
959 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
960 }
961 if (expand_bytes > 0) {
962 if (PrintGCDetails && Verbose) {
963 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
964 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
965 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
966 gclog_or_tty->print_cr(" Desired free fraction %f",
967 desired_free_percentage);
968 gclog_or_tty->print_cr(" Maximum free fraction %f",
969 maximum_free_percentage);
970 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
971 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
972 desired_capacity/1000);
973 int prev_level = level() - 1;
974 if (prev_level >= 0) {
975 size_t prev_size = 0;
976 GenCollectedHeap* gch = GenCollectedHeap::heap();
977 Generation* prev_gen = gch->_gens[prev_level];
978 prev_size = prev_gen->capacity();
979 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
980 prev_size/1000);
981 }
982 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
983 unsafe_max_alloc_nogc()/1000);
984 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
985 contiguous_available()/1000);
986 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
987 expand_bytes);
988 }
989 // safe if expansion fails
990 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
991 if (PrintGCDetails && Verbose) {
992 gclog_or_tty->print_cr(" Expanded free fraction %f",
993 ((double) free()) / capacity());
994 }
995 }
996 }
997
998 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
999 return cmsSpace()->freelistLock();
1000 }
1001
1002 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1003 bool tlab) {
1004 CMSSynchronousYieldRequest yr;
1005 MutexLockerEx x(freelistLock(),
1006 Mutex::_no_safepoint_check_flag);
1007 return have_lock_and_allocate(size, tlab);
1008 }
1009
1010 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1011 bool tlab /* ignored */) {
1012 assert_lock_strong(freelistLock());
1013 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1014 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1015 // Allocate the object live (grey) if the background collector has
1016 // started marking. This is necessary because the marker may
1017 // have passed this address and consequently this object will
1018 // not otherwise be greyed and would be incorrectly swept up.
1019 // Note that if this object contains references, the writing
1020 // of those references will dirty the card containing this object
1021 // allowing the object to be blackened (and its references scanned)
1022 // either during a preclean phase or at the final checkpoint.
1023 if (res != NULL) {
1024 // We may block here with an uninitialized object with
1025 // its mark-bit or P-bits not yet set. Such objects need
1026 // to be safely navigable by block_start().
1027 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1028 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
1029 collector()->direct_allocated(res, adjustedSize);
1030 _direct_allocated_words += adjustedSize;
1031 // allocation counters
1032 NOT_PRODUCT(
1033 _numObjectsAllocated++;
1034 _numWordsAllocated += (int)adjustedSize;
1035 )
1036 }
1037 return res;
1038 }
1039
1040 // In the case of direct allocation by mutators in a generation that
1041 // is being concurrently collected, the object must be allocated
1042 // live (grey) if the background collector has started marking.
1043 // This is necessary because the marker may
1044 // have passed this address and consequently this object will
1045 // not otherwise be greyed and would be incorrectly swept up.
1046 // Note that if this object contains references, the writing
1047 // of those references will dirty the card containing this object
1048 // allowing the object to be blackened (and its references scanned)
1049 // either during a preclean phase or at the final checkpoint.
1050 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1051 assert(_markBitMap.covers(start, size), "Out of bounds");
1052 if (_collectorState >= Marking) {
1053 MutexLockerEx y(_markBitMap.lock(),
1054 Mutex::_no_safepoint_check_flag);
1055 // [see comments preceding SweepClosure::do_blk() below for details]
1056 // 1. need to mark the object as live so it isn't collected
1057 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1058 // 3. need to mark the end of the object so marking, precleaning or sweeping
1059 // can skip over uninitialized or unparsable objects. An allocated
1060 // object is considered uninitialized for our purposes as long as
1061 // its klass word is NULL. (Unparsable objects are those which are
1062 // initialized in the sense just described, but whose sizes can still
1063 // not be correctly determined. Note that the class of unparsable objects
1064 // can only occur in the perm gen. All old gen objects are parsable
1065 // as soon as they are initialized.)
1066 _markBitMap.mark(start); // object is live
1067 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1068 _markBitMap.mark(start + size - 1);
1069 // mark end of object
1070 }
1071 // check that oop looks uninitialized
1072 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1073 }
1074
1075 void CMSCollector::promoted(bool par, HeapWord* start,
1076 bool is_obj_array, size_t obj_size) {
1077 assert(_markBitMap.covers(start), "Out of bounds");
1078 // See comment in direct_allocated() about when objects should
1079 // be allocated live.
1080 if (_collectorState >= Marking) {
1081 // we already hold the marking bit map lock, taken in
1082 // the prologue
1083 if (par) {
1084 _markBitMap.par_mark(start);
1085 } else {
1086 _markBitMap.mark(start);
1087 }
1088 // We don't need to mark the object as uninitialized (as
1089 // in direct_allocated above) because this is being done with the
1090 // world stopped and the object will be initialized by the
1091 // time the marking, precleaning or sweeping get to look at it.
1092 // But see the code for copying objects into the CMS generation,
1093 // where we need to ensure that concurrent readers of the
1094 // block offset table are able to safely navigate a block that
1095 // is in flux from being free to being allocated (and in
1096 // transition while being copied into) and subsequently
1097 // becoming a bona-fide object when the copy/promotion is complete.
1098 assert(SafepointSynchronize::is_at_safepoint(),
1099 "expect promotion only at safepoints");
1100
1101 if (_collectorState < Sweeping) {
1102 // Mark the appropriate cards in the modUnionTable, so that
1103 // this object gets scanned before the sweep. If this is
1104 // not done, CMS generation references in the object might
1105 // not get marked.
1106 // For the case of arrays, which are otherwise precisely
1107 // marked, we need to dirty the entire array, not just its head.
1108 if (is_obj_array) {
1109 // The [par_]mark_range() method expects mr.end() below to
1110 // be aligned to the granularity of a bit's representation
1111 // in the heap. In the case of the MUT below, that's a
1112 // card size.
1113 MemRegion mr(start,
1114 (HeapWord*)round_to((intptr_t)(start + obj_size),
1115 CardTableModRefBS::card_size /* bytes */));
1116 if (par) {
1117 _modUnionTable.par_mark_range(mr);
1118 } else {
1119 _modUnionTable.mark_range(mr);
1120 }
1121 } else { // not an obj array; we can just mark the head
1122 if (par) {
1123 _modUnionTable.par_mark(start);
1124 } else {
1125 _modUnionTable.mark(start);
1126 }
1127 }
1128 }
1129 }
1130 }
1131
1132 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1133 {
1134 size_t delta = pointer_delta(addr, space->bottom());
1135 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1136 }
1137
1138 void CMSCollector::icms_update_allocation_limits()
1139 {
1140 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1141 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1142
1143 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1144 if (CMSTraceIncrementalPacing) {
1145 stats().print();
1146 }
1147
1148 assert(duty_cycle <= 100, "invalid duty cycle");
1149 if (duty_cycle != 0) {
1150 // The duty_cycle is a percentage between 0 and 100; convert to words and
1151 // then compute the offset from the endpoints of the space.
1152 size_t free_words = eden->free() / HeapWordSize;
1153 double free_words_dbl = (double)free_words;
1154 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1155 size_t offset_words = (free_words - duty_cycle_words) / 2;
1156
1157 _icms_start_limit = eden->top() + offset_words;
1158 _icms_stop_limit = eden->end() - offset_words;
1159
1160 // The limits may be adjusted (shifted to the right) by
1161 // CMSIncrementalOffset, to allow the application more mutator time after a
1162 // young gen gc (when all mutators were stopped) and before CMS starts and
1163 // takes away one or more cpus.
1164 if (CMSIncrementalOffset != 0) {
1165 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1166 size_t adjustment = (size_t)adjustment_dbl;
1167 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1168 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1169 _icms_start_limit += adjustment;
1170 _icms_stop_limit = tmp_stop;
1171 }
1172 }
1173 }
1174 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1175 _icms_start_limit = _icms_stop_limit = eden->end();
1176 }
1177
1178 // Install the new start limit.
1179 eden->set_soft_end(_icms_start_limit);
1180
1181 if (CMSTraceIncrementalMode) {
1182 gclog_or_tty->print(" icms alloc limits: "
1183 PTR_FORMAT "," PTR_FORMAT
1184 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1185 _icms_start_limit, _icms_stop_limit,
1186 percent_of_space(eden, _icms_start_limit),
1187 percent_of_space(eden, _icms_stop_limit));
1188 if (Verbose) {
1189 gclog_or_tty->print("eden: ");
1190 eden->print_on(gclog_or_tty);
1191 }
1192 }
1193 }
1194
1195 // Any changes here should try to maintain the invariant
1196 // that if this method is called with _icms_start_limit
1197 // and _icms_stop_limit both NULL, then it should return NULL
1198 // and not notify the icms thread.
1199 HeapWord*
1200 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1201 size_t word_size)
1202 {
1203 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1204 // nop.
1205 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1206 if (top <= _icms_start_limit) {
1207 if (CMSTraceIncrementalMode) {
1208 space->print_on(gclog_or_tty);
1209 gclog_or_tty->stamp();
1210 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1211 ", new limit=" PTR_FORMAT
1212 " (" SIZE_FORMAT "%%)",
1213 top, _icms_stop_limit,
1214 percent_of_space(space, _icms_stop_limit));
1215 }
1216 ConcurrentMarkSweepThread::start_icms();
1217 assert(top < _icms_stop_limit, "Tautology");
1218 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1219 return _icms_stop_limit;
1220 }
1221
1222 // The allocation will cross both the _start and _stop limits, so do the
1223 // stop notification also and return end().
1224 if (CMSTraceIncrementalMode) {
1225 space->print_on(gclog_or_tty);
1226 gclog_or_tty->stamp();
1227 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1228 ", new limit=" PTR_FORMAT
1229 " (" SIZE_FORMAT "%%)",
1230 top, space->end(),
1231 percent_of_space(space, space->end()));
1232 }
1233 ConcurrentMarkSweepThread::stop_icms();
1234 return space->end();
1235 }
1236
1237 if (top <= _icms_stop_limit) {
1238 if (CMSTraceIncrementalMode) {
1239 space->print_on(gclog_or_tty);
1240 gclog_or_tty->stamp();
1241 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1242 ", new limit=" PTR_FORMAT
1243 " (" SIZE_FORMAT "%%)",
1244 top, space->end(),
1245 percent_of_space(space, space->end()));
1246 }
1247 ConcurrentMarkSweepThread::stop_icms();
1248 return space->end();
1249 }
1250
1251 if (CMSTraceIncrementalMode) {
1252 space->print_on(gclog_or_tty);
1253 gclog_or_tty->stamp();
1254 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1255 ", new limit=" PTR_FORMAT,
1256 top, NULL);
1257 }
1258 }
1259
1260 return NULL;
1261 }
1262
1263 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1264 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1265 // allocate, copy and if necessary update promoinfo --
1266 // delegate to underlying space.
1267 assert_lock_strong(freelistLock());
1268
1269 #ifndef PRODUCT
1270 if (Universe::heap()->promotion_should_fail()) {
1271 return NULL;
1272 }
1273 #endif // #ifndef PRODUCT
1274
1275 oop res = _cmsSpace->promote(obj, obj_size);
1276 if (res == NULL) {
1277 // expand and retry
1278 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1279 expand(s*HeapWordSize, MinHeapDeltaBytes,
1280 CMSExpansionCause::_satisfy_promotion);
1281 // Since there's currently no next generation, we don't try to promote
1282 // into a more senior generation.
1283 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1284 "is made to pass on a possibly failing "
1285 "promotion to next generation");
1286 res = _cmsSpace->promote(obj, obj_size);
1287 }
1288 if (res != NULL) {
1289 // See comment in allocate() about when objects should
1290 // be allocated live.
1291 assert(obj->is_oop(), "Will dereference klass pointer below");
1292 collector()->promoted(false, // Not parallel
1293 (HeapWord*)res, obj->is_objArray(), obj_size);
1294 // promotion counters
1295 NOT_PRODUCT(
1296 _numObjectsPromoted++;
1297 _numWordsPromoted +=
1298 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1299 )
1300 }
1301 return res;
1302 }
1303
1304
1305 HeapWord*
1306 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1307 HeapWord* top,
1308 size_t word_sz)
1309 {
1310 return collector()->allocation_limit_reached(space, top, word_sz);
1311 }
1312
1313 // IMPORTANT: Notes on object size recognition in CMS.
1314 // ---------------------------------------------------
1315 // A block of storage in the CMS generation is always in
1316 // one of three states. A free block (FREE), an allocated
1317 // object (OBJECT) whose size() method reports the correct size,
1318 // and an intermediate state (TRANSIENT) in which its size cannot
1319 // be accurately determined.
1320 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1321 // -----------------------------------------------------
1322 // FREE: klass_word & 1 == 1; mark_word holds block size
1323 //
1324 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1325 // obj->size() computes correct size
1326 // [Perm Gen objects needs to be "parsable" before they can be navigated]
1327 //
1328 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1329 //
1330 // STATE IDENTIFICATION: (64 bit+COOPS)
1331 // ------------------------------------
1332 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1333 //
1334 // OBJECT: klass_word installed; klass_word != 0;
1335 // obj->size() computes correct size
1336 // [Perm Gen comment above continues to hold]
1337 //
1338 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1339 //
1340 //
1341 // STATE TRANSITION DIAGRAM
1342 //
1343 // mut / parnew mut / parnew
1344 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1345 // ^ |
1346 // |------------------------ DEAD <------------------------------------|
1347 // sweep mut
1348 //
1349 // While a block is in TRANSIENT state its size cannot be determined
1350 // so readers will either need to come back later or stall until
1351 // the size can be determined. Note that for the case of direct
1352 // allocation, P-bits, when available, may be used to determine the
1353 // size of an object that may not yet have been initialized.
1354
1355 // Things to support parallel young-gen collection.
1356 oop
1357 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1358 oop old, markOop m,
1359 size_t word_sz) {
1360 #ifndef PRODUCT
1361 if (Universe::heap()->promotion_should_fail()) {
1362 return NULL;
1363 }
1364 #endif // #ifndef PRODUCT
1365
1366 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1367 PromotionInfo* promoInfo = &ps->promo;
1368 // if we are tracking promotions, then first ensure space for
1369 // promotion (including spooling space for saving header if necessary).
1370 // then allocate and copy, then track promoted info if needed.
1371 // When tracking (see PromotionInfo::track()), the mark word may
1372 // be displaced and in this case restoration of the mark word
1373 // occurs in the (oop_since_save_marks_)iterate phase.
1374 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1375 // Out of space for allocating spooling buffers;
1376 // try expanding and allocating spooling buffers.
1377 if (!expand_and_ensure_spooling_space(promoInfo)) {
1378 return NULL;
1379 }
1380 }
1381 assert(promoInfo->has_spooling_space(), "Control point invariant");
1382 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1383 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1384 if (obj_ptr == NULL) {
1385 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1386 if (obj_ptr == NULL) {
1387 return NULL;
1388 }
1389 }
1390 oop obj = oop(obj_ptr);
1391 OrderAccess::storestore();
1392 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1393 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1394 // IMPORTANT: See note on object initialization for CMS above.
1395 // Otherwise, copy the object. Here we must be careful to insert the
1396 // klass pointer last, since this marks the block as an allocated object.
1397 // Except with compressed oops it's the mark word.
1398 HeapWord* old_ptr = (HeapWord*)old;
1399 // Restore the mark word copied above.
1400 obj->set_mark(m);
1401 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1402 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1403 OrderAccess::storestore();
1404
1405 if (UseCompressedOops) {
1406 // Copy gap missed by (aligned) header size calculation below
1407 obj->set_klass_gap(old->klass_gap());
1408 }
1409 if (word_sz > (size_t)oopDesc::header_size()) {
1410 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1411 obj_ptr + oopDesc::header_size(),
1412 word_sz - oopDesc::header_size());
1413 }
1414
1415 // Now we can track the promoted object, if necessary. We take care
1416 // to delay the transition from uninitialized to full object
1417 // (i.e., insertion of klass pointer) until after, so that it
1418 // atomically becomes a promoted object.
1419 if (promoInfo->tracking()) {
1420 promoInfo->track((PromotedObject*)obj, old->klass());
1421 }
1422 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1423 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1424 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1425
1426 // Finally, install the klass pointer (this should be volatile).
1427 OrderAccess::storestore();
1428 obj->set_klass(old->klass());
1429 // We should now be able to calculate the right size for this object
1430 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1431
1432 collector()->promoted(true, // parallel
1433 obj_ptr, old->is_objArray(), word_sz);
1434
1435 NOT_PRODUCT(
1436 Atomic::inc_ptr(&_numObjectsPromoted);
1437 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1438 )
1439
1440 return obj;
1441 }
1442
1443 void
1444 ConcurrentMarkSweepGeneration::
1445 par_promote_alloc_undo(int thread_num,
1446 HeapWord* obj, size_t word_sz) {
1447 // CMS does not support promotion undo.
1448 ShouldNotReachHere();
1449 }
1450
1451 void
1452 ConcurrentMarkSweepGeneration::
1453 par_promote_alloc_done(int thread_num) {
1454 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1455 ps->lab.retire(thread_num);
1456 }
1457
1458 void
1459 ConcurrentMarkSweepGeneration::
1460 par_oop_since_save_marks_iterate_done(int thread_num) {
1461 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1462 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1463 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1464 }
1465
1466 // XXXPERM
1467 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1468 size_t size,
1469 bool tlab)
1470 {
1471 // We allow a STW collection only if a full
1472 // collection was requested.
1473 return full || should_allocate(size, tlab); // FIX ME !!!
1474 // This and promotion failure handling are connected at the
1475 // hip and should be fixed by untying them.
1476 }
1477
1478 bool CMSCollector::shouldConcurrentCollect() {
1479 if (_full_gc_requested) {
1480 if (Verbose && PrintGCDetails) {
1481 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1482 " gc request (or gc_locker)");
1483 }
1484 return true;
1485 }
1486
1487 // For debugging purposes, change the type of collection.
1488 // If the rotation is not on the concurrent collection
1489 // type, don't start a concurrent collection.
1490 NOT_PRODUCT(
1491 if (RotateCMSCollectionTypes &&
1492 (_cmsGen->debug_collection_type() !=
1493 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1494 assert(_cmsGen->debug_collection_type() !=
1495 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1496 "Bad cms collection type");
1497 return false;
1498 }
1499 )
1500
1501 FreelistLocker x(this);
1502 // ------------------------------------------------------------------
1503 // Print out lots of information which affects the initiation of
1504 // a collection.
1505 if (PrintCMSInitiationStatistics && stats().valid()) {
1506 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1507 gclog_or_tty->stamp();
1508 gclog_or_tty->print_cr("");
1509 stats().print_on(gclog_or_tty);
1510 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1511 stats().time_until_cms_gen_full());
1512 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1513 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1514 _cmsGen->contiguous_available());
1515 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1516 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1517 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1518 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1519 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy());
1520 }
1521 // ------------------------------------------------------------------
1522
1523 // If the estimated time to complete a cms collection (cms_duration())
1524 // is less than the estimated time remaining until the cms generation
1525 // is full, start a collection.
1526 if (!UseCMSInitiatingOccupancyOnly) {
1527 if (stats().valid()) {
1528 if (stats().time_until_cms_start() == 0.0) {
1529 return true;
1530 }
1531 } else {
1532 // We want to conservatively collect somewhat early in order
1533 // to try and "bootstrap" our CMS/promotion statistics;
1534 // this branch will not fire after the first successful CMS
1535 // collection because the stats should then be valid.
1536 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1537 if (Verbose && PrintGCDetails) {
1538 gclog_or_tty->print_cr(
1539 " CMSCollector: collect for bootstrapping statistics:"
1540 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1541 _bootstrap_occupancy);
1542 }
1543 return true;
1544 }
1545 }
1546 }
1547
1548 // Otherwise, we start a collection cycle if either the perm gen or
1549 // old gen want a collection cycle started. Each may use
1550 // an appropriate criterion for making this decision.
1551 // XXX We need to make sure that the gen expansion
1552 // criterion dovetails well with this. XXX NEED TO FIX THIS
1553 if (_cmsGen->should_concurrent_collect()) {
1554 if (Verbose && PrintGCDetails) {
1555 gclog_or_tty->print_cr("CMS old gen initiated");
1556 }
1557 return true;
1558 }
1559
1560 // We start a collection if we believe an incremental collection may fail;
1561 // this is not likely to be productive in practice because it's probably too
1562 // late anyway.
1563 GenCollectedHeap* gch = GenCollectedHeap::heap();
1564 assert(gch->collector_policy()->is_two_generation_policy(),
1565 "You may want to check the correctness of the following");
1566 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1567 if (Verbose && PrintGCDetails) {
1568 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1569 }
1570 return true;
1571 }
1572
1573 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) {
1574 bool res = update_should_unload_classes();
1575 if (res) {
1576 if (Verbose && PrintGCDetails) {
1577 gclog_or_tty->print_cr("CMS perm gen initiated");
1578 }
1579 return true;
1580 }
1581 }
1582 return false;
1583 }
1584
1585 // Clear _expansion_cause fields of constituent generations
1586 void CMSCollector::clear_expansion_cause() {
1587 _cmsGen->clear_expansion_cause();
1588 _permGen->clear_expansion_cause();
1589 }
1590
1591 // We should be conservative in starting a collection cycle. To
1592 // start too eagerly runs the risk of collecting too often in the
1593 // extreme. To collect too rarely falls back on full collections,
1594 // which works, even if not optimum in terms of concurrent work.
1595 // As a work around for too eagerly collecting, use the flag
1596 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1597 // giving the user an easily understandable way of controlling the
1598 // collections.
1599 // We want to start a new collection cycle if any of the following
1600 // conditions hold:
1601 // . our current occupancy exceeds the configured initiating occupancy
1602 // for this generation, or
1603 // . we recently needed to expand this space and have not, since that
1604 // expansion, done a collection of this generation, or
1605 // . the underlying space believes that it may be a good idea to initiate
1606 // a concurrent collection (this may be based on criteria such as the
1607 // following: the space uses linear allocation and linear allocation is
1608 // going to fail, or there is believed to be excessive fragmentation in
1609 // the generation, etc... or ...
1610 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1611 // the case of the old generation, not the perm generation; see CR 6543076):
1612 // we may be approaching a point at which allocation requests may fail because
1613 // we will be out of sufficient free space given allocation rate estimates.]
1614 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1615
1616 assert_lock_strong(freelistLock());
1617 if (occupancy() > initiating_occupancy()) {
1618 if (PrintGCDetails && Verbose) {
1619 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1620 short_name(), occupancy(), initiating_occupancy());
1621 }
1622 return true;
1623 }
1624 if (UseCMSInitiatingOccupancyOnly) {
1625 return false;
1626 }
1627 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1628 if (PrintGCDetails && Verbose) {
1629 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1630 short_name());
1631 }
1632 return true;
1633 }
1634 if (_cmsSpace->should_concurrent_collect()) {
1635 if (PrintGCDetails && Verbose) {
1636 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1637 short_name());
1638 }
1639 return true;
1640 }
1641 return false;
1642 }
1643
1644 void ConcurrentMarkSweepGeneration::collect(bool full,
1645 bool clear_all_soft_refs,
1646 size_t size,
1647 bool tlab)
1648 {
1649 collector()->collect(full, clear_all_soft_refs, size, tlab);
1650 }
1651
1652 void CMSCollector::collect(bool full,
1653 bool clear_all_soft_refs,
1654 size_t size,
1655 bool tlab)
1656 {
1657 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1658 // For debugging purposes skip the collection if the state
1659 // is not currently idle
1660 if (TraceCMSState) {
1661 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1662 Thread::current(), full, _collectorState);
1663 }
1664 return;
1665 }
1666
1667 // The following "if" branch is present for defensive reasons.
1668 // In the current uses of this interface, it can be replaced with:
1669 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1670 // But I am not placing that assert here to allow future
1671 // generality in invoking this interface.
1672 if (GC_locker::is_active()) {
1673 // A consistency test for GC_locker
1674 assert(GC_locker::needs_gc(), "Should have been set already");
1675 // Skip this foreground collection, instead
1676 // expanding the heap if necessary.
1677 // Need the free list locks for the call to free() in compute_new_size()
1678 compute_new_size();
1679 return;
1680 }
1681 acquire_control_and_collect(full, clear_all_soft_refs);
1682 _full_gcs_since_conc_gc++;
1683
1684 }
1685
1686 void CMSCollector::request_full_gc(unsigned int full_gc_count) {
1687 GenCollectedHeap* gch = GenCollectedHeap::heap();
1688 unsigned int gc_count = gch->total_full_collections();
1689 if (gc_count == full_gc_count) {
1690 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1691 _full_gc_requested = true;
1692 CGC_lock->notify(); // nudge CMS thread
1693 } else {
1694 assert(gc_count > full_gc_count, "Error: causal loop");
1695 }
1696 }
1697
1698
1699 // The foreground and background collectors need to coordinate in order
1700 // to make sure that they do not mutually interfere with CMS collections.
1701 // When a background collection is active,
1702 // the foreground collector may need to take over (preempt) and
1703 // synchronously complete an ongoing collection. Depending on the
1704 // frequency of the background collections and the heap usage
1705 // of the application, this preemption can be seldom or frequent.
1706 // There are only certain
1707 // points in the background collection that the "collection-baton"
1708 // can be passed to the foreground collector.
1709 //
1710 // The foreground collector will wait for the baton before
1711 // starting any part of the collection. The foreground collector
1712 // will only wait at one location.
1713 //
1714 // The background collector will yield the baton before starting a new
1715 // phase of the collection (e.g., before initial marking, marking from roots,
1716 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1717 // of the loop which switches the phases. The background collector does some
1718 // of the phases (initial mark, final re-mark) with the world stopped.
1719 // Because of locking involved in stopping the world,
1720 // the foreground collector should not block waiting for the background
1721 // collector when it is doing a stop-the-world phase. The background
1722 // collector will yield the baton at an additional point just before
1723 // it enters a stop-the-world phase. Once the world is stopped, the
1724 // background collector checks the phase of the collection. If the
1725 // phase has not changed, it proceeds with the collection. If the
1726 // phase has changed, it skips that phase of the collection. See
1727 // the comments on the use of the Heap_lock in collect_in_background().
1728 //
1729 // Variable used in baton passing.
1730 // _foregroundGCIsActive - Set to true by the foreground collector when
1731 // it wants the baton. The foreground clears it when it has finished
1732 // the collection.
1733 // _foregroundGCShouldWait - Set to true by the background collector
1734 // when it is running. The foreground collector waits while
1735 // _foregroundGCShouldWait is true.
1736 // CGC_lock - monitor used to protect access to the above variables
1737 // and to notify the foreground and background collectors.
1738 // _collectorState - current state of the CMS collection.
1739 //
1740 // The foreground collector
1741 // acquires the CGC_lock
1742 // sets _foregroundGCIsActive
1743 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1744 // various locks acquired in preparation for the collection
1745 // are released so as not to block the background collector
1746 // that is in the midst of a collection
1747 // proceeds with the collection
1748 // clears _foregroundGCIsActive
1749 // returns
1750 //
1751 // The background collector in a loop iterating on the phases of the
1752 // collection
1753 // acquires the CGC_lock
1754 // sets _foregroundGCShouldWait
1755 // if _foregroundGCIsActive is set
1756 // clears _foregroundGCShouldWait, notifies _CGC_lock
1757 // waits on _CGC_lock for _foregroundGCIsActive to become false
1758 // and exits the loop.
1759 // otherwise
1760 // proceed with that phase of the collection
1761 // if the phase is a stop-the-world phase,
1762 // yield the baton once more just before enqueueing
1763 // the stop-world CMS operation (executed by the VM thread).
1764 // returns after all phases of the collection are done
1765 //
1766
1767 void CMSCollector::acquire_control_and_collect(bool full,
1768 bool clear_all_soft_refs) {
1769 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1770 assert(!Thread::current()->is_ConcurrentGC_thread(),
1771 "shouldn't try to acquire control from self!");
1772
1773 // Start the protocol for acquiring control of the
1774 // collection from the background collector (aka CMS thread).
1775 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1776 "VM thread should have CMS token");
1777 // Remember the possibly interrupted state of an ongoing
1778 // concurrent collection
1779 CollectorState first_state = _collectorState;
1780
1781 // Signal to a possibly ongoing concurrent collection that
1782 // we want to do a foreground collection.
1783 _foregroundGCIsActive = true;
1784
1785 // Disable incremental mode during a foreground collection.
1786 ICMSDisabler icms_disabler;
1787
1788 // release locks and wait for a notify from the background collector
1789 // releasing the locks in only necessary for phases which
1790 // do yields to improve the granularity of the collection.
1791 assert_lock_strong(bitMapLock());
1792 // We need to lock the Free list lock for the space that we are
1793 // currently collecting.
1794 assert(haveFreelistLocks(), "Must be holding free list locks");
1795 bitMapLock()->unlock();
1796 releaseFreelistLocks();
1797 {
1798 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1799 if (_foregroundGCShouldWait) {
1800 // We are going to be waiting for action for the CMS thread;
1801 // it had better not be gone (for instance at shutdown)!
1802 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1803 "CMS thread must be running");
1804 // Wait here until the background collector gives us the go-ahead
1805 ConcurrentMarkSweepThread::clear_CMS_flag(
1806 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1807 // Get a possibly blocked CMS thread going:
1808 // Note that we set _foregroundGCIsActive true above,
1809 // without protection of the CGC_lock.
1810 CGC_lock->notify();
1811 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1812 "Possible deadlock");
1813 while (_foregroundGCShouldWait) {
1814 // wait for notification
1815 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1816 // Possibility of delay/starvation here, since CMS token does
1817 // not know to give priority to VM thread? Actually, i think
1818 // there wouldn't be any delay/starvation, but the proof of
1819 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1820 }
1821 ConcurrentMarkSweepThread::set_CMS_flag(
1822 ConcurrentMarkSweepThread::CMS_vm_has_token);
1823 }
1824 }
1825 // The CMS_token is already held. Get back the other locks.
1826 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1827 "VM thread should have CMS token");
1828 getFreelistLocks();
1829 bitMapLock()->lock_without_safepoint_check();
1830 if (TraceCMSState) {
1831 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1832 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1833 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1834 }
1835
1836 // Check if we need to do a compaction, or if not, whether
1837 // we need to start the mark-sweep from scratch.
1838 bool should_compact = false;
1839 bool should_start_over = false;
1840 decide_foreground_collection_type(clear_all_soft_refs,
1841 &should_compact, &should_start_over);
1842
1843 NOT_PRODUCT(
1844 if (RotateCMSCollectionTypes) {
1845 if (_cmsGen->debug_collection_type() ==
1846 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1847 should_compact = true;
1848 } else if (_cmsGen->debug_collection_type() ==
1849 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1850 should_compact = false;
1851 }
1852 }
1853 )
1854
1855 if (PrintGCDetails && first_state > Idling) {
1856 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1857 if (GCCause::is_user_requested_gc(cause) ||
1858 GCCause::is_serviceability_requested_gc(cause)) {
1859 gclog_or_tty->print(" (concurrent mode interrupted)");
1860 } else {
1861 gclog_or_tty->print(" (concurrent mode failure)");
1862 }
1863 }
1864
1865 if (should_compact) {
1866 // If the collection is being acquired from the background
1867 // collector, there may be references on the discovered
1868 // references lists that have NULL referents (being those
1869 // that were concurrently cleared by a mutator) or
1870 // that are no longer active (having been enqueued concurrently
1871 // by the mutator).
1872 // Scrub the list of those references because Mark-Sweep-Compact
1873 // code assumes referents are not NULL and that all discovered
1874 // Reference objects are active.
1875 ref_processor()->clean_up_discovered_references();
1876
1877 do_compaction_work(clear_all_soft_refs);
1878
1879 // Has the GC time limit been exceeded?
1880 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1881 size_t max_eden_size = young_gen->max_capacity() -
1882 young_gen->to()->capacity() -
1883 young_gen->from()->capacity();
1884 GenCollectedHeap* gch = GenCollectedHeap::heap();
1885 GCCause::Cause gc_cause = gch->gc_cause();
1886 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1887 young_gen->eden()->used(),
1888 _cmsGen->max_capacity(),
1889 max_eden_size,
1890 full,
1891 gc_cause,
1892 gch->collector_policy());
1893 } else {
1894 do_mark_sweep_work(clear_all_soft_refs, first_state,
1895 should_start_over);
1896 }
1897 // Reset the expansion cause, now that we just completed
1898 // a collection cycle.
1899 clear_expansion_cause();
1900 _foregroundGCIsActive = false;
1901 return;
1902 }
1903
1904 // Resize the perm generation and the tenured generation
1905 // after obtaining the free list locks for the
1906 // two generations.
1907 void CMSCollector::compute_new_size() {
1908 assert_locked_or_safepoint(Heap_lock);
1909 FreelistLocker z(this);
1910 _permGen->compute_new_size();
1911 _cmsGen->compute_new_size();
1912 }
1913
1914 // A work method used by foreground collection to determine
1915 // what type of collection (compacting or not, continuing or fresh)
1916 // it should do.
1917 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1918 // and CMSCompactWhenClearAllSoftRefs the default in the future
1919 // and do away with the flags after a suitable period.
1920 void CMSCollector::decide_foreground_collection_type(
1921 bool clear_all_soft_refs, bool* should_compact,
1922 bool* should_start_over) {
1923 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1924 // flag is set, and we have either requested a System.gc() or
1925 // the number of full gc's since the last concurrent cycle
1926 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1927 // or if an incremental collection has failed
1928 GenCollectedHeap* gch = GenCollectedHeap::heap();
1929 assert(gch->collector_policy()->is_two_generation_policy(),
1930 "You may want to check the correctness of the following");
1931 // Inform cms gen if this was due to partial collection failing.
1932 // The CMS gen may use this fact to determine its expansion policy.
1933 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1934 assert(!_cmsGen->incremental_collection_failed(),
1935 "Should have been noticed, reacted to and cleared");
1936 _cmsGen->set_incremental_collection_failed();
1937 }
1938 *should_compact =
1939 UseCMSCompactAtFullCollection &&
1940 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1941 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1942 gch->incremental_collection_will_fail(true /* consult_young */));
1943 *should_start_over = false;
1944 if (clear_all_soft_refs && !*should_compact) {
1945 // We are about to do a last ditch collection attempt
1946 // so it would normally make sense to do a compaction
1947 // to reclaim as much space as possible.
1948 if (CMSCompactWhenClearAllSoftRefs) {
1949 // Default: The rationale is that in this case either
1950 // we are past the final marking phase, in which case
1951 // we'd have to start over, or so little has been done
1952 // that there's little point in saving that work. Compaction
1953 // appears to be the sensible choice in either case.
1954 *should_compact = true;
1955 } else {
1956 // We have been asked to clear all soft refs, but not to
1957 // compact. Make sure that we aren't past the final checkpoint
1958 // phase, for that is where we process soft refs. If we are already
1959 // past that phase, we'll need to redo the refs discovery phase and
1960 // if necessary clear soft refs that weren't previously
1961 // cleared. We do so by remembering the phase in which
1962 // we came in, and if we are past the refs processing
1963 // phase, we'll choose to just redo the mark-sweep
1964 // collection from scratch.
1965 if (_collectorState > FinalMarking) {
1966 // We are past the refs processing phase;
1967 // start over and do a fresh synchronous CMS cycle
1968 _collectorState = Resetting; // skip to reset to start new cycle
1969 reset(false /* == !asynch */);
1970 *should_start_over = true;
1971 } // else we can continue a possibly ongoing current cycle
1972 }
1973 }
1974 }
1975
1976 // A work method used by the foreground collector to do
1977 // a mark-sweep-compact.
1978 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1979 GenCollectedHeap* gch = GenCollectedHeap::heap();
1980
1981 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1982 gc_timer->register_gc_start(os::elapsed_counter());
1983
1984 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1985 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1986
1987 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
1988 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
1989 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
1990 "collections passed to foreground collector", _full_gcs_since_conc_gc);
1991 }
1992
1993 // Sample collection interval time and reset for collection pause.
1994 if (UseAdaptiveSizePolicy) {
1995 size_policy()->msc_collection_begin();
1996 }
1997
1998 // Temporarily widen the span of the weak reference processing to
1999 // the entire heap.
2000 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2001 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2002 // Temporarily, clear the "is_alive_non_header" field of the
2003 // reference processor.
2004 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2005 // Temporarily make reference _processing_ single threaded (non-MT).
2006 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2007 // Temporarily make refs discovery atomic
2008 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2009 // Temporarily make reference _discovery_ single threaded (non-MT)
2010 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2011
2012 ref_processor()->set_enqueuing_is_done(false);
2013 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2014 ref_processor()->setup_policy(clear_all_soft_refs);
2015 // If an asynchronous collection finishes, the _modUnionTable is
2016 // all clear. If we are assuming the collection from an asynchronous
2017 // collection, clear the _modUnionTable.
2018 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2019 "_modUnionTable should be clear if the baton was not passed");
2020 _modUnionTable.clear_all();
2021
2022 // We must adjust the allocation statistics being maintained
2023 // in the free list space. We do so by reading and clearing
2024 // the sweep timer and updating the block flux rate estimates below.
2025 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2026 if (_inter_sweep_timer.is_active()) {
2027 _inter_sweep_timer.stop();
2028 // Note that we do not use this sample to update the _inter_sweep_estimate.
2029 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2030 _inter_sweep_estimate.padded_average(),
2031 _intra_sweep_estimate.padded_average());
2032 }
2033
2034 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2035 ref_processor(), clear_all_soft_refs);
2036 #ifdef ASSERT
2037 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2038 size_t free_size = cms_space->free();
2039 assert(free_size ==
2040 pointer_delta(cms_space->end(), cms_space->compaction_top())
2041 * HeapWordSize,
2042 "All the free space should be compacted into one chunk at top");
2043 assert(cms_space->dictionary()->total_chunk_size(
2044 debug_only(cms_space->freelistLock())) == 0 ||
2045 cms_space->totalSizeInIndexedFreeLists() == 0,
2046 "All the free space should be in a single chunk");
2047 size_t num = cms_space->totalCount();
2048 assert((free_size == 0 && num == 0) ||
2049 (free_size > 0 && (num == 1 || num == 2)),
2050 "There should be at most 2 free chunks after compaction");
2051 #endif // ASSERT
2052 _collectorState = Resetting;
2053 assert(_restart_addr == NULL,
2054 "Should have been NULL'd before baton was passed");
2055 reset(false /* == !asynch */);
2056 _cmsGen->reset_after_compaction();
2057 _concurrent_cycles_since_last_unload = 0;
2058
2059 if (verifying() && !should_unload_classes()) {
2060 perm_gen_verify_bit_map()->clear_all();
2061 }
2062
2063 // Clear any data recorded in the PLAB chunk arrays.
2064 if (_survivor_plab_array != NULL) {
2065 reset_survivor_plab_arrays();
2066 }
2067
2068 // Adjust the per-size allocation stats for the next epoch.
2069 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2070 // Restart the "inter sweep timer" for the next epoch.
2071 _inter_sweep_timer.reset();
2072 _inter_sweep_timer.start();
2073
2074 // Sample collection pause time and reset for collection interval.
2075 if (UseAdaptiveSizePolicy) {
2076 size_policy()->msc_collection_end(gch->gc_cause());
2077 }
2078
2079 gc_timer->register_gc_end(os::elapsed_counter());
2080
2081 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2082
2083 // For a mark-sweep-compact, compute_new_size() will be called
2084 // in the heap's do_collection() method.
2085 }
2086
2087 // A work method used by the foreground collector to do
2088 // a mark-sweep, after taking over from a possibly on-going
2089 // concurrent mark-sweep collection.
2090 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2091 CollectorState first_state, bool should_start_over) {
2092 if (PrintGC && Verbose) {
2093 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2094 "collector with count %d",
2095 _full_gcs_since_conc_gc);
2096 }
2097 switch (_collectorState) {
2098 case Idling:
2099 if (first_state == Idling || should_start_over) {
2100 // The background GC was not active, or should
2101 // restarted from scratch; start the cycle.
2102 _collectorState = InitialMarking;
2103 }
2104 // If first_state was not Idling, then a background GC
2105 // was in progress and has now finished. No need to do it
2106 // again. Leave the state as Idling.
2107 break;
2108 case Precleaning:
2109 // In the foreground case don't do the precleaning since
2110 // it is not done concurrently and there is extra work
2111 // required.
2112 _collectorState = FinalMarking;
2113 }
2114 collect_in_foreground(clear_all_soft_refs);
2115
2116 // For a mark-sweep, compute_new_size() will be called
2117 // in the heap's do_collection() method.
2118 }
2119
2120
2121 void CMSCollector::getFreelistLocks() const {
2122 // Get locks for all free lists in all generations that this
2123 // collector is responsible for
2124 _cmsGen->freelistLock()->lock_without_safepoint_check();
2125 _permGen->freelistLock()->lock_without_safepoint_check();
2126 }
2127
2128 void CMSCollector::releaseFreelistLocks() const {
2129 // Release locks for all free lists in all generations that this
2130 // collector is responsible for
2131 _cmsGen->freelistLock()->unlock();
2132 _permGen->freelistLock()->unlock();
2133 }
2134
2135 bool CMSCollector::haveFreelistLocks() const {
2136 // Check locks for all free lists in all generations that this
2137 // collector is responsible for
2138 assert_lock_strong(_cmsGen->freelistLock());
2139 assert_lock_strong(_permGen->freelistLock());
2140 PRODUCT_ONLY(ShouldNotReachHere());
2141 return true;
2142 }
2143
2144 // A utility class that is used by the CMS collector to
2145 // temporarily "release" the foreground collector from its
2146 // usual obligation to wait for the background collector to
2147 // complete an ongoing phase before proceeding.
2148 class ReleaseForegroundGC: public StackObj {
2149 private:
2150 CMSCollector* _c;
2151 public:
2152 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2153 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2154 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2155 // allow a potentially blocked foreground collector to proceed
2156 _c->_foregroundGCShouldWait = false;
2157 if (_c->_foregroundGCIsActive) {
2158 CGC_lock->notify();
2159 }
2160 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2161 "Possible deadlock");
2162 }
2163
2164 ~ReleaseForegroundGC() {
2165 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2166 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2167 _c->_foregroundGCShouldWait = true;
2168 }
2169 };
2170
2171 // There are separate collect_in_background and collect_in_foreground because of
2172 // the different locking requirements of the background collector and the
2173 // foreground collector. There was originally an attempt to share
2174 // one "collect" method between the background collector and the foreground
2175 // collector but the if-then-else required made it cleaner to have
2176 // separate methods.
2177 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2178 assert(Thread::current()->is_ConcurrentGC_thread(),
2179 "A CMS asynchronous collection is only allowed on a CMS thread.");
2180
2181 GenCollectedHeap* gch = GenCollectedHeap::heap();
2182 {
2183 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2184 MutexLockerEx hl(Heap_lock, safepoint_check);
2185 FreelistLocker fll(this);
2186 MutexLockerEx x(CGC_lock, safepoint_check);
2187 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2188 // The foreground collector is active or we're
2189 // not using asynchronous collections. Skip this
2190 // background collection.
2191 assert(!_foregroundGCShouldWait, "Should be clear");
2192 return;
2193 } else {
2194 assert(_collectorState == Idling, "Should be idling before start.");
2195 _collectorState = InitialMarking;
2196 // Reset the expansion cause, now that we are about to begin
2197 // a new cycle.
2198 clear_expansion_cause();
2199 }
2200 // Decide if we want to enable class unloading as part of the
2201 // ensuing concurrent GC cycle.
2202 update_should_unload_classes();
2203 _full_gc_requested = false; // acks all outstanding full gc requests
2204 // Signal that we are about to start a collection
2205 gch->increment_total_full_collections(); // ... starting a collection cycle
2206 _collection_count_start = gch->total_full_collections();
2207 }
2208
2209 // Used for PrintGC
2210 size_t prev_used;
2211 if (PrintGC && Verbose) {
2212 prev_used = _cmsGen->used(); // XXXPERM
2213 }
2214
2215 // The change of the collection state is normally done at this level;
2216 // the exceptions are phases that are executed while the world is
2217 // stopped. For those phases the change of state is done while the
2218 // world is stopped. For baton passing purposes this allows the
2219 // background collector to finish the phase and change state atomically.
2220 // The foreground collector cannot wait on a phase that is done
2221 // while the world is stopped because the foreground collector already
2222 // has the world stopped and would deadlock.
2223 while (_collectorState != Idling) {
2224 if (TraceCMSState) {
2225 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2226 Thread::current(), _collectorState);
2227 }
2228 // The foreground collector
2229 // holds the Heap_lock throughout its collection.
2230 // holds the CMS token (but not the lock)
2231 // except while it is waiting for the background collector to yield.
2232 //
2233 // The foreground collector should be blocked (not for long)
2234 // if the background collector is about to start a phase
2235 // executed with world stopped. If the background
2236 // collector has already started such a phase, the
2237 // foreground collector is blocked waiting for the
2238 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2239 // are executed in the VM thread.
2240 //
2241 // The locking order is
2242 // PendingListLock (PLL) -- if applicable (FinalMarking)
2243 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2244 // CMS token (claimed in
2245 // stop_world_and_do() -->
2246 // safepoint_synchronize() -->
2247 // CMSThread::synchronize())
2248
2249 {
2250 // Check if the FG collector wants us to yield.
2251 CMSTokenSync x(true); // is cms thread
2252 if (waitForForegroundGC()) {
2253 // We yielded to a foreground GC, nothing more to be
2254 // done this round.
2255 assert(_foregroundGCShouldWait == false, "We set it to false in "
2256 "waitForForegroundGC()");
2257 if (TraceCMSState) {
2258 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2259 " exiting collection CMS state %d",
2260 Thread::current(), _collectorState);
2261 }
2262 return;
2263 } else {
2264 // The background collector can run but check to see if the
2265 // foreground collector has done a collection while the
2266 // background collector was waiting to get the CGC_lock
2267 // above. If yes, break so that _foregroundGCShouldWait
2268 // is cleared before returning.
2269 if (_collectorState == Idling) {
2270 break;
2271 }
2272 }
2273 }
2274
2275 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2276 "should be waiting");
2277
2278 switch (_collectorState) {
2279 case InitialMarking:
2280 {
2281 ReleaseForegroundGC x(this);
2282 stats().record_cms_begin();
2283 VM_CMS_Initial_Mark initial_mark_op(this);
2284 VMThread::execute(&initial_mark_op);
2285 }
2286 // The collector state may be any legal state at this point
2287 // since the background collector may have yielded to the
2288 // foreground collector.
2289 break;
2290 case Marking:
2291 // initial marking in checkpointRootsInitialWork has been completed
2292 if (markFromRoots(true)) { // we were successful
2293 assert(_collectorState == Precleaning, "Collector state should "
2294 "have changed");
2295 } else {
2296 assert(_foregroundGCIsActive, "Internal state inconsistency");
2297 }
2298 break;
2299 case Precleaning:
2300 if (UseAdaptiveSizePolicy) {
2301 size_policy()->concurrent_precleaning_begin();
2302 }
2303 // marking from roots in markFromRoots has been completed
2304 preclean();
2305 if (UseAdaptiveSizePolicy) {
2306 size_policy()->concurrent_precleaning_end();
2307 }
2308 assert(_collectorState == AbortablePreclean ||
2309 _collectorState == FinalMarking,
2310 "Collector state should have changed");
2311 break;
2312 case AbortablePreclean:
2313 if (UseAdaptiveSizePolicy) {
2314 size_policy()->concurrent_phases_resume();
2315 }
2316 abortable_preclean();
2317 if (UseAdaptiveSizePolicy) {
2318 size_policy()->concurrent_precleaning_end();
2319 }
2320 assert(_collectorState == FinalMarking, "Collector state should "
2321 "have changed");
2322 break;
2323 case FinalMarking:
2324 {
2325 ReleaseForegroundGC x(this);
2326
2327 VM_CMS_Final_Remark final_remark_op(this);
2328 VMThread::execute(&final_remark_op);
2329 }
2330 assert(_foregroundGCShouldWait, "block post-condition");
2331 break;
2332 case Sweeping:
2333 if (UseAdaptiveSizePolicy) {
2334 size_policy()->concurrent_sweeping_begin();
2335 }
2336 // final marking in checkpointRootsFinal has been completed
2337 sweep(true);
2338 assert(_collectorState == Resizing, "Collector state change "
2339 "to Resizing must be done under the free_list_lock");
2340 _full_gcs_since_conc_gc = 0;
2341
2342 // Stop the timers for adaptive size policy for the concurrent phases
2343 if (UseAdaptiveSizePolicy) {
2344 size_policy()->concurrent_sweeping_end();
2345 size_policy()->concurrent_phases_end(gch->gc_cause(),
2346 gch->prev_gen(_cmsGen)->capacity(),
2347 _cmsGen->free());
2348 }
2349
2350 case Resizing: {
2351 // Sweeping has been completed...
2352 // At this point the background collection has completed.
2353 // Don't move the call to compute_new_size() down
2354 // into code that might be executed if the background
2355 // collection was preempted.
2356 {
2357 ReleaseForegroundGC x(this); // unblock FG collection
2358 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2359 CMSTokenSync z(true); // not strictly needed.
2360 if (_collectorState == Resizing) {
2361 compute_new_size();
2362 _collectorState = Resetting;
2363 } else {
2364 assert(_collectorState == Idling, "The state should only change"
2365 " because the foreground collector has finished the collection");
2366 }
2367 }
2368 break;
2369 }
2370 case Resetting:
2371 // CMS heap resizing has been completed
2372 reset(true);
2373 assert(_collectorState == Idling, "Collector state should "
2374 "have changed");
2375 stats().record_cms_end();
2376 // Don't move the concurrent_phases_end() and compute_new_size()
2377 // calls to here because a preempted background collection
2378 // has it's state set to "Resetting".
2379 break;
2380 case Idling:
2381 default:
2382 ShouldNotReachHere();
2383 break;
2384 }
2385 if (TraceCMSState) {
2386 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2387 Thread::current(), _collectorState);
2388 }
2389 assert(_foregroundGCShouldWait, "block post-condition");
2390 }
2391
2392 // Should this be in gc_epilogue?
2393 collector_policy()->counters()->update_counters();
2394
2395 {
2396 // Clear _foregroundGCShouldWait and, in the event that the
2397 // foreground collector is waiting, notify it, before
2398 // returning.
2399 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2400 _foregroundGCShouldWait = false;
2401 if (_foregroundGCIsActive) {
2402 CGC_lock->notify();
2403 }
2404 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2405 "Possible deadlock");
2406 }
2407 if (TraceCMSState) {
2408 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2409 " exiting collection CMS state %d",
2410 Thread::current(), _collectorState);
2411 }
2412 if (PrintGC && Verbose) {
2413 _cmsGen->print_heap_change(prev_used);
2414 }
2415 }
2416
2417 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2418 _cms_start_registered = true;
2419 _gc_timer_cm->register_gc_start(os::elapsed_counter());
2420 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2421 }
2422
2423 void CMSCollector::register_gc_end() {
2424 if (_cms_start_registered) {
2425 _gc_timer_cm->register_gc_end(os::elapsed_counter());
2426 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2427 _cms_start_registered = false;
2428 }
2429 }
2430
2431 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2432 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2433 "Foreground collector should be waiting, not executing");
2434 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2435 "may only be done by the VM Thread with the world stopped");
2436 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2437 "VM thread should have CMS token");
2438
2439 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2440 true, NULL);)
2441 if (UseAdaptiveSizePolicy) {
2442 size_policy()->ms_collection_begin();
2443 }
2444 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2445
2446 HandleMark hm; // Discard invalid handles created during verification
2447
2448 if (VerifyBeforeGC &&
2449 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2450 Universe::verify();
2451 }
2452
2453 // Snapshot the soft reference policy to be used in this collection cycle.
2454 ref_processor()->setup_policy(clear_all_soft_refs);
2455
2456 bool init_mark_was_synchronous = false; // until proven otherwise
2457 while (_collectorState != Idling) {
2458 if (TraceCMSState) {
2459 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2460 Thread::current(), _collectorState);
2461 }
2462 switch (_collectorState) {
2463 case InitialMarking:
2464 register_gc_start(GenCollectedHeap::heap()->gc_cause());
2465 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2466 checkpointRootsInitial(false);
2467 assert(_collectorState == Marking, "Collector state should have changed"
2468 " within checkpointRootsInitial()");
2469 break;
2470 case Marking:
2471 // initial marking in checkpointRootsInitialWork has been completed
2472 if (VerifyDuringGC &&
2473 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2474 gclog_or_tty->print("Verify before initial mark: ");
2475 Universe::verify();
2476 }
2477 {
2478 bool res = markFromRoots(false);
2479 assert(res && _collectorState == FinalMarking, "Collector state should "
2480 "have changed");
2481 break;
2482 }
2483 case FinalMarking:
2484 if (VerifyDuringGC &&
2485 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2486 gclog_or_tty->print("Verify before re-mark: ");
2487 Universe::verify();
2488 }
2489 checkpointRootsFinal(false, clear_all_soft_refs,
2490 init_mark_was_synchronous);
2491 assert(_collectorState == Sweeping, "Collector state should not "
2492 "have changed within checkpointRootsFinal()");
2493 break;
2494 case Sweeping:
2495 // final marking in checkpointRootsFinal has been completed
2496 if (VerifyDuringGC &&
2497 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2498 gclog_or_tty->print("Verify before sweep: ");
2499 Universe::verify();
2500 }
2501 sweep(false);
2502 assert(_collectorState == Resizing, "Incorrect state");
2503 break;
2504 case Resizing: {
2505 // Sweeping has been completed; the actual resize in this case
2506 // is done separately; nothing to be done in this state.
2507 _collectorState = Resetting;
2508 break;
2509 }
2510 case Resetting:
2511 // The heap has been resized.
2512 if (VerifyDuringGC &&
2513 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2514 gclog_or_tty->print("Verify before reset: ");
2515 Universe::verify();
2516 }
2517 reset(false);
2518 assert(_collectorState == Idling, "Collector state should "
2519 "have changed");
2520 break;
2521 case Precleaning:
2522 case AbortablePreclean:
2523 // Elide the preclean phase
2524 _collectorState = FinalMarking;
2525 break;
2526 default:
2527 ShouldNotReachHere();
2528 }
2529 if (TraceCMSState) {
2530 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2531 Thread::current(), _collectorState);
2532 }
2533 }
2534
2535 if (UseAdaptiveSizePolicy) {
2536 GenCollectedHeap* gch = GenCollectedHeap::heap();
2537 size_policy()->ms_collection_end(gch->gc_cause());
2538 }
2539
2540 if (VerifyAfterGC &&
2541 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2542 Universe::verify();
2543 }
2544 if (TraceCMSState) {
2545 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2546 " exiting collection CMS state %d",
2547 Thread::current(), _collectorState);
2548 }
2549 }
2550
2551 bool CMSCollector::waitForForegroundGC() {
2552 bool res = false;
2553 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2554 "CMS thread should have CMS token");
2555 // Block the foreground collector until the
2556 // background collectors decides whether to
2557 // yield.
2558 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2559 _foregroundGCShouldWait = true;
2560 if (_foregroundGCIsActive) {
2561 // The background collector yields to the
2562 // foreground collector and returns a value
2563 // indicating that it has yielded. The foreground
2564 // collector can proceed.
2565 res = true;
2566 _foregroundGCShouldWait = false;
2567 ConcurrentMarkSweepThread::clear_CMS_flag(
2568 ConcurrentMarkSweepThread::CMS_cms_has_token);
2569 ConcurrentMarkSweepThread::set_CMS_flag(
2570 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2571 // Get a possibly blocked foreground thread going
2572 CGC_lock->notify();
2573 if (TraceCMSState) {
2574 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2575 Thread::current(), _collectorState);
2576 }
2577 while (_foregroundGCIsActive) {
2578 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2579 }
2580 ConcurrentMarkSweepThread::set_CMS_flag(
2581 ConcurrentMarkSweepThread::CMS_cms_has_token);
2582 ConcurrentMarkSweepThread::clear_CMS_flag(
2583 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2584 }
2585 if (TraceCMSState) {
2586 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2587 Thread::current(), _collectorState);
2588 }
2589 return res;
2590 }
2591
2592 // Because of the need to lock the free lists and other structures in
2593 // the collector, common to all the generations that the collector is
2594 // collecting, we need the gc_prologues of individual CMS generations
2595 // delegate to their collector. It may have been simpler had the
2596 // current infrastructure allowed one to call a prologue on a
2597 // collector. In the absence of that we have the generation's
2598 // prologue delegate to the collector, which delegates back
2599 // some "local" work to a worker method in the individual generations
2600 // that it's responsible for collecting, while itself doing any
2601 // work common to all generations it's responsible for. A similar
2602 // comment applies to the gc_epilogue()'s.
2603 // The role of the varaible _between_prologue_and_epilogue is to
2604 // enforce the invocation protocol.
2605 void CMSCollector::gc_prologue(bool full) {
2606 // Call gc_prologue_work() for each CMSGen and PermGen that
2607 // we are responsible for.
2608
2609 // The following locking discipline assumes that we are only called
2610 // when the world is stopped.
2611 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2612
2613 // The CMSCollector prologue must call the gc_prologues for the
2614 // "generations" (including PermGen if any) that it's responsible
2615 // for.
2616
2617 assert( Thread::current()->is_VM_thread()
2618 || ( CMSScavengeBeforeRemark
2619 && Thread::current()->is_ConcurrentGC_thread()),
2620 "Incorrect thread type for prologue execution");
2621
2622 if (_between_prologue_and_epilogue) {
2623 // We have already been invoked; this is a gc_prologue delegation
2624 // from yet another CMS generation that we are responsible for, just
2625 // ignore it since all relevant work has already been done.
2626 return;
2627 }
2628
2629 // set a bit saying prologue has been called; cleared in epilogue
2630 _between_prologue_and_epilogue = true;
2631 // Claim locks for common data structures, then call gc_prologue_work()
2632 // for each CMSGen and PermGen that we are responsible for.
2633
2634 getFreelistLocks(); // gets free list locks on constituent spaces
2635 bitMapLock()->lock_without_safepoint_check();
2636
2637 // Should call gc_prologue_work() for all cms gens we are responsible for
2638 bool registerClosure = _collectorState >= Marking
2639 && _collectorState < Sweeping;
2640 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2641 &_modUnionClosurePar
2642 : &_modUnionClosure;
2643 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2644 _permGen->gc_prologue_work(full, registerClosure, muc);
2645
2646 if (!full) {
2647 stats().record_gc0_begin();
2648 }
2649 }
2650
2651 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2652 // Delegate to CMScollector which knows how to coordinate between
2653 // this and any other CMS generations that it is responsible for
2654 // collecting.
2655 collector()->gc_prologue(full);
2656 }
2657
2658 // This is a "private" interface for use by this generation's CMSCollector.
2659 // Not to be called directly by any other entity (for instance,
2660 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2661 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2662 bool registerClosure, ModUnionClosure* modUnionClosure) {
2663 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2664 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2665 "Should be NULL");
2666 if (registerClosure) {
2667 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2668 }
2669 cmsSpace()->gc_prologue();
2670 // Clear stat counters
2671 NOT_PRODUCT(
2672 assert(_numObjectsPromoted == 0, "check");
2673 assert(_numWordsPromoted == 0, "check");
2674 if (Verbose && PrintGC) {
2675 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2676 SIZE_FORMAT" bytes concurrently",
2677 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2678 }
2679 _numObjectsAllocated = 0;
2680 _numWordsAllocated = 0;
2681 )
2682 }
2683
2684 void CMSCollector::gc_epilogue(bool full) {
2685 // The following locking discipline assumes that we are only called
2686 // when the world is stopped.
2687 assert(SafepointSynchronize::is_at_safepoint(),
2688 "world is stopped assumption");
2689
2690 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2691 // if linear allocation blocks need to be appropriately marked to allow the
2692 // the blocks to be parsable. We also check here whether we need to nudge the
2693 // CMS collector thread to start a new cycle (if it's not already active).
2694 assert( Thread::current()->is_VM_thread()
2695 || ( CMSScavengeBeforeRemark
2696 && Thread::current()->is_ConcurrentGC_thread()),
2697 "Incorrect thread type for epilogue execution");
2698
2699 if (!_between_prologue_and_epilogue) {
2700 // We have already been invoked; this is a gc_epilogue delegation
2701 // from yet another CMS generation that we are responsible for, just
2702 // ignore it since all relevant work has already been done.
2703 return;
2704 }
2705 assert(haveFreelistLocks(), "must have freelist locks");
2706 assert_lock_strong(bitMapLock());
2707
2708 _cmsGen->gc_epilogue_work(full);
2709 _permGen->gc_epilogue_work(full);
2710
2711 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2712 // in case sampling was not already enabled, enable it
2713 _start_sampling = true;
2714 }
2715 // reset _eden_chunk_array so sampling starts afresh
2716 _eden_chunk_index = 0;
2717
2718 size_t cms_used = _cmsGen->cmsSpace()->used();
2719 size_t perm_used = _permGen->cmsSpace()->used();
2720
2721 // update performance counters - this uses a special version of
2722 // update_counters() that allows the utilization to be passed as a
2723 // parameter, avoiding multiple calls to used().
2724 //
2725 _cmsGen->update_counters(cms_used);
2726 _permGen->update_counters(perm_used);
2727
2728 if (CMSIncrementalMode) {
2729 icms_update_allocation_limits();
2730 }
2731
2732 bitMapLock()->unlock();
2733 releaseFreelistLocks();
2734
2735 if (!CleanChunkPoolAsync) {
2736 Chunk::clean_chunk_pool();
2737 }
2738
2739 _between_prologue_and_epilogue = false; // ready for next cycle
2740 }
2741
2742 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2743 collector()->gc_epilogue(full);
2744
2745 // Also reset promotion tracking in par gc thread states.
2746 if (CollectedHeap::use_parallel_gc_threads()) {
2747 for (uint i = 0; i < ParallelGCThreads; i++) {
2748 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2749 }
2750 }
2751 }
2752
2753 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2754 assert(!incremental_collection_failed(), "Should have been cleared");
2755 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2756 cmsSpace()->gc_epilogue();
2757 // Print stat counters
2758 NOT_PRODUCT(
2759 assert(_numObjectsAllocated == 0, "check");
2760 assert(_numWordsAllocated == 0, "check");
2761 if (Verbose && PrintGC) {
2762 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2763 SIZE_FORMAT" bytes",
2764 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2765 }
2766 _numObjectsPromoted = 0;
2767 _numWordsPromoted = 0;
2768 )
2769
2770 if (PrintGC && Verbose) {
2771 // Call down the chain in contiguous_available needs the freelistLock
2772 // so print this out before releasing the freeListLock.
2773 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2774 contiguous_available());
2775 }
2776 }
2777
2778 #ifndef PRODUCT
2779 bool CMSCollector::have_cms_token() {
2780 Thread* thr = Thread::current();
2781 if (thr->is_VM_thread()) {
2782 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2783 } else if (thr->is_ConcurrentGC_thread()) {
2784 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2785 } else if (thr->is_GC_task_thread()) {
2786 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2787 ParGCRareEvent_lock->owned_by_self();
2788 }
2789 return false;
2790 }
2791 #endif
2792
2793 // Check reachability of the given heap address in CMS generation,
2794 // treating all other generations as roots.
2795 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2796 // We could "guarantee" below, rather than assert, but i'll
2797 // leave these as "asserts" so that an adventurous debugger
2798 // could try this in the product build provided some subset of
2799 // the conditions were met, provided they were intersted in the
2800 // results and knew that the computation below wouldn't interfere
2801 // with other concurrent computations mutating the structures
2802 // being read or written.
2803 assert(SafepointSynchronize::is_at_safepoint(),
2804 "Else mutations in object graph will make answer suspect");
2805 assert(have_cms_token(), "Should hold cms token");
2806 assert(haveFreelistLocks(), "must hold free list locks");
2807 assert_lock_strong(bitMapLock());
2808
2809 // Clear the marking bit map array before starting, but, just
2810 // for kicks, first report if the given address is already marked
2811 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2812 _markBitMap.isMarked(addr) ? "" : " not");
2813
2814 if (verify_after_remark()) {
2815 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2816 bool result = verification_mark_bm()->isMarked(addr);
2817 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2818 result ? "IS" : "is NOT");
2819 return result;
2820 } else {
2821 gclog_or_tty->print_cr("Could not compute result");
2822 return false;
2823 }
2824 }
2825
2826 ////////////////////////////////////////////////////////
2827 // CMS Verification Support
2828 ////////////////////////////////////////////////////////
2829 // Following the remark phase, the following invariant
2830 // should hold -- each object in the CMS heap which is
2831 // marked in markBitMap() should be marked in the verification_mark_bm().
2832
2833 class VerifyMarkedClosure: public BitMapClosure {
2834 CMSBitMap* _marks;
2835 bool _failed;
2836
2837 public:
2838 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2839
2840 bool do_bit(size_t offset) {
2841 HeapWord* addr = _marks->offsetToHeapWord(offset);
2842 if (!_marks->isMarked(addr)) {
2843 oop(addr)->print_on(gclog_or_tty);
2844 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2845 _failed = true;
2846 }
2847 return true;
2848 }
2849
2850 bool failed() { return _failed; }
2851 };
2852
2853 bool CMSCollector::verify_after_remark() {
2854 gclog_or_tty->print(" [Verifying CMS Marking... ");
2855 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2856 static bool init = false;
2857
2858 assert(SafepointSynchronize::is_at_safepoint(),
2859 "Else mutations in object graph will make answer suspect");
2860 assert(have_cms_token(),
2861 "Else there may be mutual interference in use of "
2862 " verification data structures");
2863 assert(_collectorState > Marking && _collectorState <= Sweeping,
2864 "Else marking info checked here may be obsolete");
2865 assert(haveFreelistLocks(), "must hold free list locks");
2866 assert_lock_strong(bitMapLock());
2867
2868
2869 // Allocate marking bit map if not already allocated
2870 if (!init) { // first time
2871 if (!verification_mark_bm()->allocate(_span)) {
2872 return false;
2873 }
2874 init = true;
2875 }
2876
2877 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2878
2879 // Turn off refs discovery -- so we will be tracing through refs.
2880 // This is as intended, because by this time
2881 // GC must already have cleared any refs that need to be cleared,
2882 // and traced those that need to be marked; moreover,
2883 // the marking done here is not going to intefere in any
2884 // way with the marking information used by GC.
2885 NoRefDiscovery no_discovery(ref_processor());
2886
2887 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2888
2889 // Clear any marks from a previous round
2890 verification_mark_bm()->clear_all();
2891 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2892 verify_work_stacks_empty();
2893
2894 GenCollectedHeap* gch = GenCollectedHeap::heap();
2895 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2896 // Update the saved marks which may affect the root scans.
2897 gch->save_marks();
2898
2899 if (CMSRemarkVerifyVariant == 1) {
2900 // In this first variant of verification, we complete
2901 // all marking, then check if the new marks-verctor is
2902 // a subset of the CMS marks-vector.
2903 verify_after_remark_work_1();
2904 } else if (CMSRemarkVerifyVariant == 2) {
2905 // In this second variant of verification, we flag an error
2906 // (i.e. an object reachable in the new marks-vector not reachable
2907 // in the CMS marks-vector) immediately, also indicating the
2908 // identify of an object (A) that references the unmarked object (B) --
2909 // presumably, a mutation to A failed to be picked up by preclean/remark?
2910 verify_after_remark_work_2();
2911 } else {
2912 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2913 CMSRemarkVerifyVariant);
2914 }
2915 gclog_or_tty->print(" done] ");
2916 return true;
2917 }
2918
2919 void CMSCollector::verify_after_remark_work_1() {
2920 ResourceMark rm;
2921 HandleMark hm;
2922 GenCollectedHeap* gch = GenCollectedHeap::heap();
2923
2924 // Mark from roots one level into CMS
2925 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2926 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2927
2928 gch->gen_process_strong_roots(_cmsGen->level(),
2929 true, // younger gens are roots
2930 true, // activate StrongRootsScope
2931 true, // collecting perm gen
2932 SharedHeap::ScanningOption(roots_scanning_options()),
2933 ¬Older,
2934 true, // walk code active on stacks
2935 NULL);
2936
2937 // Now mark from the roots
2938 assert(_revisitStack.isEmpty(), "Should be empty");
2939 MarkFromRootsClosure markFromRootsClosure(this, _span,
2940 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2941 false /* don't yield */, true /* verifying */);
2942 assert(_restart_addr == NULL, "Expected pre-condition");
2943 verification_mark_bm()->iterate(&markFromRootsClosure);
2944 while (_restart_addr != NULL) {
2945 // Deal with stack overflow: by restarting at the indicated
2946 // address.
2947 HeapWord* ra = _restart_addr;
2948 markFromRootsClosure.reset(ra);
2949 _restart_addr = NULL;
2950 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2951 }
2952 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2953 verify_work_stacks_empty();
2954 // Should reset the revisit stack above, since no class tree
2955 // surgery is forthcoming.
2956 _revisitStack.reset(); // throwing away all contents
2957
2958 // Marking completed -- now verify that each bit marked in
2959 // verification_mark_bm() is also marked in markBitMap(); flag all
2960 // errors by printing corresponding objects.
2961 VerifyMarkedClosure vcl(markBitMap());
2962 verification_mark_bm()->iterate(&vcl);
2963 if (vcl.failed()) {
2964 gclog_or_tty->print("Verification failed");
2965 Universe::heap()->print_on(gclog_or_tty);
2966 fatal("CMS: failed marking verification after remark");
2967 }
2968 }
2969
2970 void CMSCollector::verify_after_remark_work_2() {
2971 ResourceMark rm;
2972 HandleMark hm;
2973 GenCollectedHeap* gch = GenCollectedHeap::heap();
2974
2975 // Mark from roots one level into CMS
2976 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2977 markBitMap());
2978 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2979 gch->gen_process_strong_roots(_cmsGen->level(),
2980 true, // younger gens are roots
2981 true, // activate StrongRootsScope
2982 true, // collecting perm gen
2983 SharedHeap::ScanningOption(roots_scanning_options()),
2984 ¬Older,
2985 true, // walk code active on stacks
2986 NULL);
2987
2988 // Now mark from the roots
2989 assert(_revisitStack.isEmpty(), "Should be empty");
2990 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2991 verification_mark_bm(), markBitMap(), verification_mark_stack());
2992 assert(_restart_addr == NULL, "Expected pre-condition");
2993 verification_mark_bm()->iterate(&markFromRootsClosure);
2994 while (_restart_addr != NULL) {
2995 // Deal with stack overflow: by restarting at the indicated
2996 // address.
2997 HeapWord* ra = _restart_addr;
2998 markFromRootsClosure.reset(ra);
2999 _restart_addr = NULL;
3000 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3001 }
3002 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3003 verify_work_stacks_empty();
3004 // Should reset the revisit stack above, since no class tree
3005 // surgery is forthcoming.
3006 _revisitStack.reset(); // throwing away all contents
3007
3008 // Marking completed -- now verify that each bit marked in
3009 // verification_mark_bm() is also marked in markBitMap(); flag all
3010 // errors by printing corresponding objects.
3011 VerifyMarkedClosure vcl(markBitMap());
3012 verification_mark_bm()->iterate(&vcl);
3013 assert(!vcl.failed(), "Else verification above should not have succeeded");
3014 }
3015
3016 void ConcurrentMarkSweepGeneration::save_marks() {
3017 // delegate to CMS space
3018 cmsSpace()->save_marks();
3019 for (uint i = 0; i < ParallelGCThreads; i++) {
3020 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3021 }
3022 }
3023
3024 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3025 return cmsSpace()->no_allocs_since_save_marks();
3026 }
3027
3028 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3029 \
3030 void ConcurrentMarkSweepGeneration:: \
3031 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3032 cl->set_generation(this); \
3033 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3034 cl->reset_generation(); \
3035 save_marks(); \
3036 }
3037
3038 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3039
3040 void
3041 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
3042 {
3043 // Not currently implemented; need to do the following. -- ysr.
3044 // dld -- I think that is used for some sort of allocation profiler. So it
3045 // really means the objects allocated by the mutator since the last
3046 // GC. We could potentially implement this cheaply by recording only
3047 // the direct allocations in a side data structure.
3048 //
3049 // I think we probably ought not to be required to support these
3050 // iterations at any arbitrary point; I think there ought to be some
3051 // call to enable/disable allocation profiling in a generation/space,
3052 // and the iterator ought to return the objects allocated in the
3053 // gen/space since the enable call, or the last iterator call (which
3054 // will probably be at a GC.) That way, for gens like CM&S that would
3055 // require some extra data structure to support this, we only pay the
3056 // cost when it's in use...
3057 cmsSpace()->object_iterate_since_last_GC(blk);
3058 }
3059
3060 void
3061 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3062 cl->set_generation(this);
3063 younger_refs_in_space_iterate(_cmsSpace, cl);
3064 cl->reset_generation();
3065 }
3066
3067 void
3068 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
3069 if (freelistLock()->owned_by_self()) {
3070 Generation::oop_iterate(mr, cl);
3071 } else {
3072 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3073 Generation::oop_iterate(mr, cl);
3074 }
3075 }
3076
3077 void
3078 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
3079 if (freelistLock()->owned_by_self()) {
3080 Generation::oop_iterate(cl);
3081 } else {
3082 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3083 Generation::oop_iterate(cl);
3084 }
3085 }
3086
3087 void
3088 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3089 if (freelistLock()->owned_by_self()) {
3090 Generation::object_iterate(cl);
3091 } else {
3092 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3093 Generation::object_iterate(cl);
3094 }
3095 }
3096
3097 void
3098 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3099 if (freelistLock()->owned_by_self()) {
3100 Generation::safe_object_iterate(cl);
3101 } else {
3102 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3103 Generation::safe_object_iterate(cl);
3104 }
3105 }
3106
3107 void
3108 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3109 }
3110
3111 void
3112 ConcurrentMarkSweepGeneration::post_compact() {
3113 }
3114
3115 void
3116 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3117 // Fix the linear allocation blocks to look like free blocks.
3118
3119 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3120 // are not called when the heap is verified during universe initialization and
3121 // at vm shutdown.
3122 if (freelistLock()->owned_by_self()) {
3123 cmsSpace()->prepare_for_verify();
3124 } else {
3125 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3126 cmsSpace()->prepare_for_verify();
3127 }
3128 }
3129
3130 void
3131 ConcurrentMarkSweepGeneration::verify() {
3132 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3133 // are not called when the heap is verified during universe initialization and
3134 // at vm shutdown.
3135 if (freelistLock()->owned_by_self()) {
3136 cmsSpace()->verify();
3137 } else {
3138 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3139 cmsSpace()->verify();
3140 }
3141 }
3142
3143 void CMSCollector::verify() {
3144 _cmsGen->verify();
3145 _permGen->verify();
3146 }
3147
3148 #ifndef PRODUCT
3149 bool CMSCollector::overflow_list_is_empty() const {
3150 assert(_num_par_pushes >= 0, "Inconsistency");
3151 if (_overflow_list == NULL) {
3152 assert(_num_par_pushes == 0, "Inconsistency");
3153 }
3154 return _overflow_list == NULL;
3155 }
3156
3157 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3158 // merely consolidate assertion checks that appear to occur together frequently.
3159 void CMSCollector::verify_work_stacks_empty() const {
3160 assert(_markStack.isEmpty(), "Marking stack should be empty");
3161 assert(overflow_list_is_empty(), "Overflow list should be empty");
3162 }
3163
3164 void CMSCollector::verify_overflow_empty() const {
3165 assert(overflow_list_is_empty(), "Overflow list should be empty");
3166 assert(no_preserved_marks(), "No preserved marks");
3167 }
3168 #endif // PRODUCT
3169
3170 // Decide if we want to enable class unloading as part of the
3171 // ensuing concurrent GC cycle. We will collect the perm gen and
3172 // unload classes if it's the case that:
3173 // (1) an explicit gc request has been made and the flag
3174 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3175 // (2) (a) class unloading is enabled at the command line, and
3176 // (b) (i) perm gen threshold has been crossed, or
3177 // (ii) old gen is getting really full, or
3178 // (iii) the previous N CMS collections did not collect the
3179 // perm gen
3180 // NOTE: Provided there is no change in the state of the heap between
3181 // calls to this method, it should have idempotent results. Moreover,
3182 // its results should be monotonically increasing (i.e. going from 0 to 1,
3183 // but not 1 to 0) between successive calls between which the heap was
3184 // not collected. For the implementation below, it must thus rely on
3185 // the property that concurrent_cycles_since_last_unload()
3186 // will not decrease unless a collection cycle happened and that
3187 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3188 // themselves also monotonic in that sense. See check_monotonicity()
3189 // below.
3190 bool CMSCollector::update_should_unload_classes() {
3191 _should_unload_classes = false;
3192 // Condition 1 above
3193 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3194 _should_unload_classes = true;
3195 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3196 // Disjuncts 2.b.(i,ii,iii) above
3197 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3198 CMSClassUnloadingMaxInterval)
3199 || _permGen->should_concurrent_collect()
3200 || _cmsGen->is_too_full();
3201 }
3202 return _should_unload_classes;
3203 }
3204
3205 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3206 bool res = should_concurrent_collect();
3207 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3208 return res;
3209 }
3210
3211 void CMSCollector::setup_cms_unloading_and_verification_state() {
3212 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3213 || VerifyBeforeExit;
3214 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3215
3216 if (should_unload_classes()) { // Should unload classes this cycle
3217 remove_root_scanning_option(rso); // Shrink the root set appropriately
3218 set_verifying(should_verify); // Set verification state for this cycle
3219 return; // Nothing else needs to be done at this time
3220 }
3221
3222 // Not unloading classes this cycle
3223 assert(!should_unload_classes(), "Inconsitency!");
3224 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3225 // We were not verifying, or we _were_ unloading classes in the last cycle,
3226 // AND some verification options are enabled this cycle; in this case,
3227 // we must make sure that the deadness map is allocated if not already so,
3228 // and cleared (if already allocated previously --
3229 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3230 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3231 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3232 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3233 "permanent generation verification disabled");
3234 return; // Note that we leave verification disabled, so we'll retry this
3235 // allocation next cycle. We _could_ remember this failure
3236 // and skip further attempts and permanently disable verification
3237 // attempts if that is considered more desirable.
3238 }
3239 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3240 "_perm_gen_ver_bit_map inconsistency?");
3241 } else {
3242 perm_gen_verify_bit_map()->clear_all();
3243 }
3244 // Include symbols, strings and code cache elements to prevent their resurrection.
3245 add_root_scanning_option(rso);
3246 set_verifying(true);
3247 } else if (verifying() && !should_verify) {
3248 // We were verifying, but some verification flags got disabled.
3249 set_verifying(false);
3250 // Exclude symbols, strings and code cache elements from root scanning to
3251 // reduce IM and RM pauses.
3252 remove_root_scanning_option(rso);
3253 }
3254 }
3255
3256
3257 #ifndef PRODUCT
3258 HeapWord* CMSCollector::block_start(const void* p) const {
3259 const HeapWord* addr = (HeapWord*)p;
3260 if (_span.contains(p)) {
3261 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3262 return _cmsGen->cmsSpace()->block_start(p);
3263 } else {
3264 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3265 "Inconsistent _span?");
3266 return _permGen->cmsSpace()->block_start(p);
3267 }
3268 }
3269 return NULL;
3270 }
3271 #endif
3272
3273 HeapWord*
3274 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3275 bool tlab,
3276 bool parallel) {
3277 CMSSynchronousYieldRequest yr;
3278 assert(!tlab, "Can't deal with TLAB allocation");
3279 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3280 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3281 CMSExpansionCause::_satisfy_allocation);
3282 if (GCExpandToAllocateDelayMillis > 0) {
3283 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3284 }
3285 return have_lock_and_allocate(word_size, tlab);
3286 }
3287
3288 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3289 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3290 // to CardGeneration and share it...
3291 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3292 return CardGeneration::expand(bytes, expand_bytes);
3293 }
3294
3295 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3296 CMSExpansionCause::Cause cause)
3297 {
3298
3299 bool success = expand(bytes, expand_bytes);
3300
3301 // remember why we expanded; this information is used
3302 // by shouldConcurrentCollect() when making decisions on whether to start
3303 // a new CMS cycle.
3304 if (success) {
3305 set_expansion_cause(cause);
3306 if (PrintGCDetails && Verbose) {
3307 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3308 CMSExpansionCause::to_string(cause));
3309 }
3310 }
3311 }
3312
3313 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3314 HeapWord* res = NULL;
3315 MutexLocker x(ParGCRareEvent_lock);
3316 while (true) {
3317 // Expansion by some other thread might make alloc OK now:
3318 res = ps->lab.alloc(word_sz);
3319 if (res != NULL) return res;
3320 // If there's not enough expansion space available, give up.
3321 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3322 return NULL;
3323 }
3324 // Otherwise, we try expansion.
3325 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3326 CMSExpansionCause::_allocate_par_lab);
3327 // Now go around the loop and try alloc again;
3328 // A competing par_promote might beat us to the expansion space,
3329 // so we may go around the loop again if promotion fails agaion.
3330 if (GCExpandToAllocateDelayMillis > 0) {
3331 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3332 }
3333 }
3334 }
3335
3336
3337 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3338 PromotionInfo* promo) {
3339 MutexLocker x(ParGCRareEvent_lock);
3340 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3341 while (true) {
3342 // Expansion by some other thread might make alloc OK now:
3343 if (promo->ensure_spooling_space()) {
3344 assert(promo->has_spooling_space(),
3345 "Post-condition of successful ensure_spooling_space()");
3346 return true;
3347 }
3348 // If there's not enough expansion space available, give up.
3349 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3350 return false;
3351 }
3352 // Otherwise, we try expansion.
3353 expand(refill_size_bytes, MinHeapDeltaBytes,
3354 CMSExpansionCause::_allocate_par_spooling_space);
3355 // Now go around the loop and try alloc again;
3356 // A competing allocation might beat us to the expansion space,
3357 // so we may go around the loop again if allocation fails again.
3358 if (GCExpandToAllocateDelayMillis > 0) {
3359 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3360 }
3361 }
3362 }
3363
3364
3365
3366 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3367 assert_locked_or_safepoint(Heap_lock);
3368 size_t size = ReservedSpace::page_align_size_down(bytes);
3369 if (size > 0) {
3370 shrink_by(size);
3371 }
3372 }
3373
3374 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3375 assert_locked_or_safepoint(Heap_lock);
3376 bool result = _virtual_space.expand_by(bytes);
3377 if (result) {
3378 HeapWord* old_end = _cmsSpace->end();
3379 size_t new_word_size =
3380 heap_word_size(_virtual_space.committed_size());
3381 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3382 _bts->resize(new_word_size); // resize the block offset shared array
3383 Universe::heap()->barrier_set()->resize_covered_region(mr);
3384 // Hmmmm... why doesn't CFLS::set_end verify locking?
3385 // This is quite ugly; FIX ME XXX
3386 _cmsSpace->assert_locked(freelistLock());
3387 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3388
3389 // update the space and generation capacity counters
3390 if (UsePerfData) {
3391 _space_counters->update_capacity();
3392 _gen_counters->update_all();
3393 }
3394
3395 if (Verbose && PrintGC) {
3396 size_t new_mem_size = _virtual_space.committed_size();
3397 size_t old_mem_size = new_mem_size - bytes;
3398 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3399 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3400 }
3401 }
3402 return result;
3403 }
3404
3405 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3406 assert_locked_or_safepoint(Heap_lock);
3407 bool success = true;
3408 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3409 if (remaining_bytes > 0) {
3410 success = grow_by(remaining_bytes);
3411 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3412 }
3413 return success;
3414 }
3415
3416 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3417 assert_locked_or_safepoint(Heap_lock);
3418 assert_lock_strong(freelistLock());
3419 // XXX Fix when compaction is implemented.
3420 warning("Shrinking of CMS not yet implemented");
3421 return;
3422 }
3423
3424
3425 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3426 // phases.
3427 class CMSPhaseAccounting: public StackObj {
3428 public:
3429 CMSPhaseAccounting(CMSCollector *collector,
3430 const char *phase,
3431 bool print_cr = true);
3432 ~CMSPhaseAccounting();
3433
3434 private:
3435 CMSCollector *_collector;
3436 const char *_phase;
3437 elapsedTimer _wallclock;
3438 bool _print_cr;
3439
3440 public:
3441 // Not MT-safe; so do not pass around these StackObj's
3442 // where they may be accessed by other threads.
3443 jlong wallclock_millis() {
3444 assert(_wallclock.is_active(), "Wall clock should not stop");
3445 _wallclock.stop(); // to record time
3446 jlong ret = _wallclock.milliseconds();
3447 _wallclock.start(); // restart
3448 return ret;
3449 }
3450 };
3451
3452 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3453 const char *phase,
3454 bool print_cr) :
3455 _collector(collector), _phase(phase), _print_cr(print_cr) {
3456
3457 if (PrintCMSStatistics != 0) {
3458 _collector->resetYields();
3459 }
3460 if (PrintGCDetails && PrintGCTimeStamps) {
3461 gclog_or_tty->date_stamp(PrintGCDateStamps);
3462 gclog_or_tty->stamp();
3463 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3464 _collector->cmsGen()->short_name(), _phase);
3465 }
3466 _collector->resetTimer();
3467 _wallclock.start();
3468 _collector->startTimer();
3469 }
3470
3471 CMSPhaseAccounting::~CMSPhaseAccounting() {
3472 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3473 _collector->stopTimer();
3474 _wallclock.stop();
3475 if (PrintGCDetails) {
3476 gclog_or_tty->date_stamp(PrintGCDateStamps);
3477 gclog_or_tty->stamp(PrintGCTimeStamps);
3478 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3479 _collector->cmsGen()->short_name(),
3480 _phase, _collector->timerValue(), _wallclock.seconds());
3481 if (_print_cr) {
3482 gclog_or_tty->print_cr("");
3483 }
3484 if (PrintCMSStatistics != 0) {
3485 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3486 _collector->yields());
3487 }
3488 }
3489 }
3490
3491 // CMS work
3492
3493 // Checkpoint the roots into this generation from outside
3494 // this generation. [Note this initial checkpoint need only
3495 // be approximate -- we'll do a catch up phase subsequently.]
3496 void CMSCollector::checkpointRootsInitial(bool asynch) {
3497 assert(_collectorState == InitialMarking, "Wrong collector state");
3498 check_correct_thread_executing();
3499 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3500
3501 ReferenceProcessor* rp = ref_processor();
3502 SpecializationStats::clear();
3503 assert(_restart_addr == NULL, "Control point invariant");
3504 if (asynch) {
3505 // acquire locks for subsequent manipulations
3506 MutexLockerEx x(bitMapLock(),
3507 Mutex::_no_safepoint_check_flag);
3508 checkpointRootsInitialWork(asynch);
3509 // enable ("weak") refs discovery
3510 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3511 _collectorState = Marking;
3512 } else {
3513 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3514 // which recognizes if we are a CMS generation, and doesn't try to turn on
3515 // discovery; verify that they aren't meddling.
3516 assert(!rp->discovery_is_atomic(),
3517 "incorrect setting of discovery predicate");
3518 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3519 "ref discovery for this generation kind");
3520 // already have locks
3521 checkpointRootsInitialWork(asynch);
3522 // now enable ("weak") refs discovery
3523 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3524 _collectorState = Marking;
3525 }
3526 SpecializationStats::print();
3527 }
3528
3529 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3530 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3531 assert(_collectorState == InitialMarking, "just checking");
3532
3533 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3534 // precede our marking with a collection of all
3535 // younger generations to keep floating garbage to a minimum.
3536 // XXX: we won't do this for now -- it's an optimization to be done later.
3537
3538 // already have locks
3539 assert_lock_strong(bitMapLock());
3540 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3541
3542 // Setup the verification and class unloading state for this
3543 // CMS collection cycle.
3544 setup_cms_unloading_and_verification_state();
3545
3546 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3547 PrintGCDetails && Verbose, true, _gc_timer_cm);)
3548 if (UseAdaptiveSizePolicy) {
3549 size_policy()->checkpoint_roots_initial_begin();
3550 }
3551
3552 // Reset all the PLAB chunk arrays if necessary.
3553 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3554 reset_survivor_plab_arrays();
3555 }
3556
3557 ResourceMark rm;
3558 HandleMark hm;
3559
3560 FalseClosure falseClosure;
3561 // In the case of a synchronous collection, we will elide the
3562 // remark step, so it's important to catch all the nmethod oops
3563 // in this step.
3564 // The final 'true' flag to gen_process_strong_roots will ensure this.
3565 // If 'async' is true, we can relax the nmethod tracing.
3566 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3567 GenCollectedHeap* gch = GenCollectedHeap::heap();
3568
3569 verify_work_stacks_empty();
3570 verify_overflow_empty();
3571
3572 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3573 // Update the saved marks which may affect the root scans.
3574 gch->save_marks();
3575
3576 // weak reference processing has not started yet.
3577 ref_processor()->set_enqueuing_is_done(false);
3578
3579 {
3580 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);)
3581 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3582 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3583 gch->gen_process_strong_roots(_cmsGen->level(),
3584 true, // younger gens are roots
3585 true, // activate StrongRootsScope
3586 true, // collecting perm gen
3587 SharedHeap::ScanningOption(roots_scanning_options()),
3588 ¬Older,
3589 true, // walk all of code cache if (so & SO_CodeCache)
3590 NULL);
3591 }
3592
3593 // Clear mod-union table; it will be dirtied in the prologue of
3594 // CMS generation per each younger generation collection.
3595
3596 assert(_modUnionTable.isAllClear(),
3597 "Was cleared in most recent final checkpoint phase"
3598 " or no bits are set in the gc_prologue before the start of the next "
3599 "subsequent marking phase.");
3600
3601 // Save the end of the used_region of the constituent generations
3602 // to be used to limit the extent of sweep in each generation.
3603 save_sweep_limits();
3604 if (UseAdaptiveSizePolicy) {
3605 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3606 }
3607 verify_overflow_empty();
3608 }
3609
3610 bool CMSCollector::markFromRoots(bool asynch) {
3611 // we might be tempted to assert that:
3612 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3613 // "inconsistent argument?");
3614 // However that wouldn't be right, because it's possible that
3615 // a safepoint is indeed in progress as a younger generation
3616 // stop-the-world GC happens even as we mark in this generation.
3617 assert(_collectorState == Marking, "inconsistent state?");
3618 check_correct_thread_executing();
3619 verify_overflow_empty();
3620
3621 bool res;
3622 if (asynch) {
3623
3624 // Start the timers for adaptive size policy for the concurrent phases
3625 // Do it here so that the foreground MS can use the concurrent
3626 // timer since a foreground MS might has the sweep done concurrently
3627 // or STW.
3628 if (UseAdaptiveSizePolicy) {
3629 size_policy()->concurrent_marking_begin();
3630 }
3631
3632 // Weak ref discovery note: We may be discovering weak
3633 // refs in this generation concurrent (but interleaved) with
3634 // weak ref discovery by a younger generation collector.
3635
3636 CMSTokenSyncWithLocks ts(true, bitMapLock());
3637 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3638 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3639 res = markFromRootsWork(asynch);
3640 if (res) {
3641 _collectorState = Precleaning;
3642 } else { // We failed and a foreground collection wants to take over
3643 assert(_foregroundGCIsActive, "internal state inconsistency");
3644 assert(_restart_addr == NULL, "foreground will restart from scratch");
3645 if (PrintGCDetails) {
3646 gclog_or_tty->print_cr("bailing out to foreground collection");
3647 }
3648 }
3649 if (UseAdaptiveSizePolicy) {
3650 size_policy()->concurrent_marking_end();
3651 }
3652 } else {
3653 assert(SafepointSynchronize::is_at_safepoint(),
3654 "inconsistent with asynch == false");
3655 if (UseAdaptiveSizePolicy) {
3656 size_policy()->ms_collection_marking_begin();
3657 }
3658 // already have locks
3659 res = markFromRootsWork(asynch);
3660 _collectorState = FinalMarking;
3661 if (UseAdaptiveSizePolicy) {
3662 GenCollectedHeap* gch = GenCollectedHeap::heap();
3663 size_policy()->ms_collection_marking_end(gch->gc_cause());
3664 }
3665 }
3666 verify_overflow_empty();
3667 return res;
3668 }
3669
3670 bool CMSCollector::markFromRootsWork(bool asynch) {
3671 // iterate over marked bits in bit map, doing a full scan and mark
3672 // from these roots using the following algorithm:
3673 // . if oop is to the right of the current scan pointer,
3674 // mark corresponding bit (we'll process it later)
3675 // . else (oop is to left of current scan pointer)
3676 // push oop on marking stack
3677 // . drain the marking stack
3678
3679 // Note that when we do a marking step we need to hold the
3680 // bit map lock -- recall that direct allocation (by mutators)
3681 // and promotion (by younger generation collectors) is also
3682 // marking the bit map. [the so-called allocate live policy.]
3683 // Because the implementation of bit map marking is not
3684 // robust wrt simultaneous marking of bits in the same word,
3685 // we need to make sure that there is no such interference
3686 // between concurrent such updates.
3687
3688 // already have locks
3689 assert_lock_strong(bitMapLock());
3690
3691 // Clear the revisit stack, just in case there are any
3692 // obsolete contents from a short-circuited previous CMS cycle.
3693 _revisitStack.reset();
3694 verify_work_stacks_empty();
3695 verify_overflow_empty();
3696 assert(_revisitStack.isEmpty(), "tabula rasa");
3697 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
3698 bool result = false;
3699 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3700 result = do_marking_mt(asynch);
3701 } else {
3702 result = do_marking_st(asynch);
3703 }
3704 return result;
3705 }
3706
3707 // Forward decl
3708 class CMSConcMarkingTask;
3709
3710 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3711 CMSCollector* _collector;
3712 CMSConcMarkingTask* _task;
3713 public:
3714 virtual void yield();
3715
3716 // "n_threads" is the number of threads to be terminated.
3717 // "queue_set" is a set of work queues of other threads.
3718 // "collector" is the CMS collector associated with this task terminator.
3719 // "yield" indicates whether we need the gang as a whole to yield.
3720 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3721 ParallelTaskTerminator(n_threads, queue_set),
3722 _collector(collector) { }
3723
3724 void set_task(CMSConcMarkingTask* task) {
3725 _task = task;
3726 }
3727 };
3728
3729 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3730 CMSConcMarkingTask* _task;
3731 public:
3732 bool should_exit_termination();
3733 void set_task(CMSConcMarkingTask* task) {
3734 _task = task;
3735 }
3736 };
3737
3738 // MT Concurrent Marking Task
3739 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3740 CMSCollector* _collector;
3741 int _n_workers; // requested/desired # workers
3742 bool _asynch;
3743 bool _result;
3744 CompactibleFreeListSpace* _cms_space;
3745 CompactibleFreeListSpace* _perm_space;
3746 char _pad_front[64]; // padding to ...
3747 HeapWord* _global_finger; // ... avoid sharing cache line
3748 char _pad_back[64];
3749 HeapWord* _restart_addr;
3750
3751 // Exposed here for yielding support
3752 Mutex* const _bit_map_lock;
3753
3754 // The per thread work queues, available here for stealing
3755 OopTaskQueueSet* _task_queues;
3756
3757 // Termination (and yielding) support
3758 CMSConcMarkingTerminator _term;
3759 CMSConcMarkingTerminatorTerminator _term_term;
3760
3761 public:
3762 CMSConcMarkingTask(CMSCollector* collector,
3763 CompactibleFreeListSpace* cms_space,
3764 CompactibleFreeListSpace* perm_space,
3765 bool asynch,
3766 YieldingFlexibleWorkGang* workers,
3767 OopTaskQueueSet* task_queues):
3768 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3769 _collector(collector),
3770 _cms_space(cms_space),
3771 _perm_space(perm_space),
3772 _asynch(asynch), _n_workers(0), _result(true),
3773 _task_queues(task_queues),
3774 _term(_n_workers, task_queues, _collector),
3775 _bit_map_lock(collector->bitMapLock())
3776 {
3777 _requested_size = _n_workers;
3778 _term.set_task(this);
3779 _term_term.set_task(this);
3780 assert(_cms_space->bottom() < _perm_space->bottom(),
3781 "Finger incorrectly initialized below");
3782 _restart_addr = _global_finger = _cms_space->bottom();
3783 }
3784
3785
3786 OopTaskQueueSet* task_queues() { return _task_queues; }
3787
3788 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3789
3790 HeapWord** global_finger_addr() { return &_global_finger; }
3791
3792 CMSConcMarkingTerminator* terminator() { return &_term; }
3793
3794 virtual void set_for_termination(int active_workers) {
3795 terminator()->reset_for_reuse(active_workers);
3796 }
3797
3798 void work(uint worker_id);
3799 bool should_yield() {
3800 return ConcurrentMarkSweepThread::should_yield()
3801 && !_collector->foregroundGCIsActive()
3802 && _asynch;
3803 }
3804
3805 virtual void coordinator_yield(); // stuff done by coordinator
3806 bool result() { return _result; }
3807
3808 void reset(HeapWord* ra) {
3809 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3810 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
3811 assert(ra < _perm_space->end(), "ra too large");
3812 _restart_addr = _global_finger = ra;
3813 _term.reset_for_reuse();
3814 }
3815
3816 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3817 OopTaskQueue* work_q);
3818
3819 private:
3820 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3821 void do_work_steal(int i);
3822 void bump_global_finger(HeapWord* f);
3823 };
3824
3825 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3826 assert(_task != NULL, "Error");
3827 return _task->yielding();
3828 // Note that we do not need the disjunct || _task->should_yield() above
3829 // because we want terminating threads to yield only if the task
3830 // is already in the midst of yielding, which happens only after at least one
3831 // thread has yielded.
3832 }
3833
3834 void CMSConcMarkingTerminator::yield() {
3835 if (_task->should_yield()) {
3836 _task->yield();
3837 } else {
3838 ParallelTaskTerminator::yield();
3839 }
3840 }
3841
3842 ////////////////////////////////////////////////////////////////
3843 // Concurrent Marking Algorithm Sketch
3844 ////////////////////////////////////////////////////////////////
3845 // Until all tasks exhausted (both spaces):
3846 // -- claim next available chunk
3847 // -- bump global finger via CAS
3848 // -- find first object that starts in this chunk
3849 // and start scanning bitmap from that position
3850 // -- scan marked objects for oops
3851 // -- CAS-mark target, and if successful:
3852 // . if target oop is above global finger (volatile read)
3853 // nothing to do
3854 // . if target oop is in chunk and above local finger
3855 // then nothing to do
3856 // . else push on work-queue
3857 // -- Deal with possible overflow issues:
3858 // . local work-queue overflow causes stuff to be pushed on
3859 // global (common) overflow queue
3860 // . always first empty local work queue
3861 // . then get a batch of oops from global work queue if any
3862 // . then do work stealing
3863 // -- When all tasks claimed (both spaces)
3864 // and local work queue empty,
3865 // then in a loop do:
3866 // . check global overflow stack; steal a batch of oops and trace
3867 // . try to steal from other threads oif GOS is empty
3868 // . if neither is available, offer termination
3869 // -- Terminate and return result
3870 //
3871 void CMSConcMarkingTask::work(uint worker_id) {
3872 elapsedTimer _timer;
3873 ResourceMark rm;
3874 HandleMark hm;
3875
3876 DEBUG_ONLY(_collector->verify_overflow_empty();)
3877
3878 // Before we begin work, our work queue should be empty
3879 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3880 // Scan the bitmap covering _cms_space, tracing through grey objects.
3881 _timer.start();
3882 do_scan_and_mark(worker_id, _cms_space);
3883 _timer.stop();
3884 if (PrintCMSStatistics != 0) {
3885 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3886 worker_id, _timer.seconds());
3887 // XXX: need xxx/xxx type of notation, two timers
3888 }
3889
3890 // ... do the same for the _perm_space
3891 _timer.reset();
3892 _timer.start();
3893 do_scan_and_mark(worker_id, _perm_space);
3894 _timer.stop();
3895 if (PrintCMSStatistics != 0) {
3896 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3897 worker_id, _timer.seconds());
3898 // XXX: need xxx/xxx type of notation, two timers
3899 }
3900
3901 // ... do work stealing
3902 _timer.reset();
3903 _timer.start();
3904 do_work_steal(worker_id);
3905 _timer.stop();
3906 if (PrintCMSStatistics != 0) {
3907 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3908 worker_id, _timer.seconds());
3909 // XXX: need xxx/xxx type of notation, two timers
3910 }
3911 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3912 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3913 // Note that under the current task protocol, the
3914 // following assertion is true even of the spaces
3915 // expanded since the completion of the concurrent
3916 // marking. XXX This will likely change under a strict
3917 // ABORT semantics.
3918 assert(_global_finger > _cms_space->end() &&
3919 _global_finger >= _perm_space->end(),
3920 "All tasks have been completed");
3921 DEBUG_ONLY(_collector->verify_overflow_empty();)
3922 }
3923
3924 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3925 HeapWord* read = _global_finger;
3926 HeapWord* cur = read;
3927 while (f > read) {
3928 cur = read;
3929 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3930 if (cur == read) {
3931 // our cas succeeded
3932 assert(_global_finger >= f, "protocol consistency");
3933 break;
3934 }
3935 }
3936 }
3937
3938 // This is really inefficient, and should be redone by
3939 // using (not yet available) block-read and -write interfaces to the
3940 // stack and the work_queue. XXX FIX ME !!!
3941 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3942 OopTaskQueue* work_q) {
3943 // Fast lock-free check
3944 if (ovflw_stk->length() == 0) {
3945 return false;
3946 }
3947 assert(work_q->size() == 0, "Shouldn't steal");
3948 MutexLockerEx ml(ovflw_stk->par_lock(),
3949 Mutex::_no_safepoint_check_flag);
3950 // Grab up to 1/4 the size of the work queue
3951 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3952 (size_t)ParGCDesiredObjsFromOverflowList);
3953 num = MIN2(num, ovflw_stk->length());
3954 for (int i = (int) num; i > 0; i--) {
3955 oop cur = ovflw_stk->pop();
3956 assert(cur != NULL, "Counted wrong?");
3957 work_q->push(cur);
3958 }
3959 return num > 0;
3960 }
3961
3962 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3963 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3964 int n_tasks = pst->n_tasks();
3965 // We allow that there may be no tasks to do here because
3966 // we are restarting after a stack overflow.
3967 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3968 uint nth_task = 0;
3969
3970 HeapWord* aligned_start = sp->bottom();
3971 if (sp->used_region().contains(_restart_addr)) {
3972 // Align down to a card boundary for the start of 0th task
3973 // for this space.
3974 aligned_start =
3975 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
3976 CardTableModRefBS::card_size);
3977 }
3978
3979 size_t chunk_size = sp->marking_task_size();
3980 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3981 // Having claimed the nth task in this space,
3982 // compute the chunk that it corresponds to:
3983 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3984 aligned_start + (nth_task+1)*chunk_size);
3985 // Try and bump the global finger via a CAS;
3986 // note that we need to do the global finger bump
3987 // _before_ taking the intersection below, because
3988 // the task corresponding to that region will be
3989 // deemed done even if the used_region() expands
3990 // because of allocation -- as it almost certainly will
3991 // during start-up while the threads yield in the
3992 // closure below.
3993 HeapWord* finger = span.end();
3994 bump_global_finger(finger); // atomically
3995 // There are null tasks here corresponding to chunks
3996 // beyond the "top" address of the space.
3997 span = span.intersection(sp->used_region());
3998 if (!span.is_empty()) { // Non-null task
3999 HeapWord* prev_obj;
4000 assert(!span.contains(_restart_addr) || nth_task == 0,
4001 "Inconsistency");
4002 if (nth_task == 0) {
4003 // For the 0th task, we'll not need to compute a block_start.
4004 if (span.contains(_restart_addr)) {
4005 // In the case of a restart because of stack overflow,
4006 // we might additionally skip a chunk prefix.
4007 prev_obj = _restart_addr;
4008 } else {
4009 prev_obj = span.start();
4010 }
4011 } else {
4012 // We want to skip the first object because
4013 // the protocol is to scan any object in its entirety
4014 // that _starts_ in this span; a fortiori, any
4015 // object starting in an earlier span is scanned
4016 // as part of an earlier claimed task.
4017 // Below we use the "careful" version of block_start
4018 // so we do not try to navigate uninitialized objects.
4019 prev_obj = sp->block_start_careful(span.start());
4020 // Below we use a variant of block_size that uses the
4021 // Printezis bits to avoid waiting for allocated
4022 // objects to become initialized/parsable.
4023 while (prev_obj < span.start()) {
4024 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4025 if (sz > 0) {
4026 prev_obj += sz;
4027 } else {
4028 // In this case we may end up doing a bit of redundant
4029 // scanning, but that appears unavoidable, short of
4030 // locking the free list locks; see bug 6324141.
4031 break;
4032 }
4033 }
4034 }
4035 if (prev_obj < span.end()) {
4036 MemRegion my_span = MemRegion(prev_obj, span.end());
4037 // Do the marking work within a non-empty span --
4038 // the last argument to the constructor indicates whether the
4039 // iteration should be incremental with periodic yields.
4040 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4041 &_collector->_markBitMap,
4042 work_queue(i),
4043 &_collector->_markStack,
4044 &_collector->_revisitStack,
4045 _asynch);
4046 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4047 } // else nothing to do for this task
4048 } // else nothing to do for this task
4049 }
4050 // We'd be tempted to assert here that since there are no
4051 // more tasks left to claim in this space, the global_finger
4052 // must exceed space->top() and a fortiori space->end(). However,
4053 // that would not quite be correct because the bumping of
4054 // global_finger occurs strictly after the claiming of a task,
4055 // so by the time we reach here the global finger may not yet
4056 // have been bumped up by the thread that claimed the last
4057 // task.
4058 pst->all_tasks_completed();
4059 }
4060
4061 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
4062 private:
4063 CMSConcMarkingTask* _task;
4064 MemRegion _span;
4065 CMSBitMap* _bit_map;
4066 CMSMarkStack* _overflow_stack;
4067 OopTaskQueue* _work_queue;
4068 protected:
4069 DO_OOP_WORK_DEFN
4070 public:
4071 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4072 CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
4073 CMSMarkStack* revisit_stack):
4074 Par_KlassRememberingOopClosure(collector, collector->ref_processor(), revisit_stack),
4075 _task(task),
4076 _span(collector->_span),
4077 _work_queue(work_queue),
4078 _bit_map(bit_map),
4079 _overflow_stack(overflow_stack)
4080 { }
4081 virtual void do_oop(oop* p);
4082 virtual void do_oop(narrowOop* p);
4083 void trim_queue(size_t max);
4084 void handle_stack_overflow(HeapWord* lost);
4085 void do_yield_check() {
4086 if (_task->should_yield()) {
4087 _task->yield();
4088 }
4089 }
4090 };
4091
4092 // Grey object scanning during work stealing phase --
4093 // the salient assumption here is that any references
4094 // that are in these stolen objects being scanned must
4095 // already have been initialized (else they would not have
4096 // been published), so we do not need to check for
4097 // uninitialized objects before pushing here.
4098 void Par_ConcMarkingClosure::do_oop(oop obj) {
4099 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4100 HeapWord* addr = (HeapWord*)obj;
4101 // Check if oop points into the CMS generation
4102 // and is not marked
4103 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4104 // a white object ...
4105 // If we manage to "claim" the object, by being the
4106 // first thread to mark it, then we push it on our
4107 // marking stack
4108 if (_bit_map->par_mark(addr)) { // ... now grey
4109 // push on work queue (grey set)
4110 bool simulate_overflow = false;
4111 NOT_PRODUCT(
4112 if (CMSMarkStackOverflowALot &&
4113 _collector->simulate_overflow()) {
4114 // simulate a stack overflow
4115 simulate_overflow = true;
4116 }
4117 )
4118 if (simulate_overflow ||
4119 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4120 // stack overflow
4121 if (PrintCMSStatistics != 0) {
4122 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4123 SIZE_FORMAT, _overflow_stack->capacity());
4124 }
4125 // We cannot assert that the overflow stack is full because
4126 // it may have been emptied since.
4127 assert(simulate_overflow ||
4128 _work_queue->size() == _work_queue->max_elems(),
4129 "Else push should have succeeded");
4130 handle_stack_overflow(addr);
4131 }
4132 } // Else, some other thread got there first
4133 do_yield_check();
4134 }
4135 }
4136
4137 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4138 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4139
4140 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4141 while (_work_queue->size() > max) {
4142 oop new_oop;
4143 if (_work_queue->pop_local(new_oop)) {
4144 assert(new_oop->is_oop(), "Should be an oop");
4145 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4146 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4147 assert(new_oop->is_parsable(), "Should be parsable");
4148 new_oop->oop_iterate(this); // do_oop() above
4149 do_yield_check();
4150 }
4151 }
4152 }
4153
4154 // Upon stack overflow, we discard (part of) the stack,
4155 // remembering the least address amongst those discarded
4156 // in CMSCollector's _restart_address.
4157 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4158 // We need to do this under a mutex to prevent other
4159 // workers from interfering with the work done below.
4160 MutexLockerEx ml(_overflow_stack->par_lock(),
4161 Mutex::_no_safepoint_check_flag);
4162 // Remember the least grey address discarded
4163 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4164 _collector->lower_restart_addr(ra);
4165 _overflow_stack->reset(); // discard stack contents
4166 _overflow_stack->expand(); // expand the stack if possible
4167 }
4168
4169
4170 void CMSConcMarkingTask::do_work_steal(int i) {
4171 OopTaskQueue* work_q = work_queue(i);
4172 oop obj_to_scan;
4173 CMSBitMap* bm = &(_collector->_markBitMap);
4174 CMSMarkStack* ovflw = &(_collector->_markStack);
4175 CMSMarkStack* revisit = &(_collector->_revisitStack);
4176 int* seed = _collector->hash_seed(i);
4177 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
4178 while (true) {
4179 cl.trim_queue(0);
4180 assert(work_q->size() == 0, "Should have been emptied above");
4181 if (get_work_from_overflow_stack(ovflw, work_q)) {
4182 // Can't assert below because the work obtained from the
4183 // overflow stack may already have been stolen from us.
4184 // assert(work_q->size() > 0, "Work from overflow stack");
4185 continue;
4186 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4187 assert(obj_to_scan->is_oop(), "Should be an oop");
4188 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4189 obj_to_scan->oop_iterate(&cl);
4190 } else if (terminator()->offer_termination(&_term_term)) {
4191 assert(work_q->size() == 0, "Impossible!");
4192 break;
4193 } else if (yielding() || should_yield()) {
4194 yield();
4195 }
4196 }
4197 }
4198
4199 // This is run by the CMS (coordinator) thread.
4200 void CMSConcMarkingTask::coordinator_yield() {
4201 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4202 "CMS thread should hold CMS token");
4203 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4204 // First give up the locks, then yield, then re-lock
4205 // We should probably use a constructor/destructor idiom to
4206 // do this unlock/lock or modify the MutexUnlocker class to
4207 // serve our purpose. XXX
4208 assert_lock_strong(_bit_map_lock);
4209 _bit_map_lock->unlock();
4210 ConcurrentMarkSweepThread::desynchronize(true);
4211 ConcurrentMarkSweepThread::acknowledge_yield_request();
4212 _collector->stopTimer();
4213 if (PrintCMSStatistics != 0) {
4214 _collector->incrementYields();
4215 }
4216 _collector->icms_wait();
4217
4218 // It is possible for whichever thread initiated the yield request
4219 // not to get a chance to wake up and take the bitmap lock between
4220 // this thread releasing it and reacquiring it. So, while the
4221 // should_yield() flag is on, let's sleep for a bit to give the
4222 // other thread a chance to wake up. The limit imposed on the number
4223 // of iterations is defensive, to avoid any unforseen circumstances
4224 // putting us into an infinite loop. Since it's always been this
4225 // (coordinator_yield()) method that was observed to cause the
4226 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4227 // which is by default non-zero. For the other seven methods that
4228 // also perform the yield operation, as are using a different
4229 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4230 // can enable the sleeping for those methods too, if necessary.
4231 // See 6442774.
4232 //
4233 // We really need to reconsider the synchronization between the GC
4234 // thread and the yield-requesting threads in the future and we
4235 // should really use wait/notify, which is the recommended
4236 // way of doing this type of interaction. Additionally, we should
4237 // consolidate the eight methods that do the yield operation and they
4238 // are almost identical into one for better maintenability and
4239 // readability. See 6445193.
4240 //
4241 // Tony 2006.06.29
4242 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4243 ConcurrentMarkSweepThread::should_yield() &&
4244 !CMSCollector::foregroundGCIsActive(); ++i) {
4245 os::sleep(Thread::current(), 1, false);
4246 ConcurrentMarkSweepThread::acknowledge_yield_request();
4247 }
4248
4249 ConcurrentMarkSweepThread::synchronize(true);
4250 _bit_map_lock->lock_without_safepoint_check();
4251 _collector->startTimer();
4252 }
4253
4254 bool CMSCollector::do_marking_mt(bool asynch) {
4255 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4256 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4257 conc_workers()->total_workers(),
4258 conc_workers()->active_workers(),
4259 Threads::number_of_non_daemon_threads());
4260 conc_workers()->set_active_workers(num_workers);
4261
4262 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4263 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4264
4265 CMSConcMarkingTask tsk(this,
4266 cms_space,
4267 perm_space,
4268 asynch,
4269 conc_workers(),
4270 task_queues());
4271
4272 // Since the actual number of workers we get may be different
4273 // from the number we requested above, do we need to do anything different
4274 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4275 // class?? XXX
4276 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4277 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4278
4279 // Refs discovery is already non-atomic.
4280 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4281 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4282 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
4283 conc_workers()->start_task(&tsk);
4284 while (tsk.yielded()) {
4285 tsk.coordinator_yield();
4286 conc_workers()->continue_task(&tsk);
4287 }
4288 // If the task was aborted, _restart_addr will be non-NULL
4289 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4290 while (_restart_addr != NULL) {
4291 // XXX For now we do not make use of ABORTED state and have not
4292 // yet implemented the right abort semantics (even in the original
4293 // single-threaded CMS case). That needs some more investigation
4294 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4295 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4296 // If _restart_addr is non-NULL, a marking stack overflow
4297 // occurred; we need to do a fresh marking iteration from the
4298 // indicated restart address.
4299 if (_foregroundGCIsActive && asynch) {
4300 // We may be running into repeated stack overflows, having
4301 // reached the limit of the stack size, while making very
4302 // slow forward progress. It may be best to bail out and
4303 // let the foreground collector do its job.
4304 // Clear _restart_addr, so that foreground GC
4305 // works from scratch. This avoids the headache of
4306 // a "rescan" which would otherwise be needed because
4307 // of the dirty mod union table & card table.
4308 _restart_addr = NULL;
4309 return false;
4310 }
4311 // Adjust the task to restart from _restart_addr
4312 tsk.reset(_restart_addr);
4313 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4314 _restart_addr);
4315 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4316 _restart_addr);
4317 _restart_addr = NULL;
4318 // Get the workers going again
4319 conc_workers()->start_task(&tsk);
4320 while (tsk.yielded()) {
4321 tsk.coordinator_yield();
4322 conc_workers()->continue_task(&tsk);
4323 }
4324 }
4325 assert(tsk.completed(), "Inconsistency");
4326 assert(tsk.result() == true, "Inconsistency");
4327 return true;
4328 }
4329
4330 bool CMSCollector::do_marking_st(bool asynch) {
4331 ResourceMark rm;
4332 HandleMark hm;
4333
4334 // Temporarily make refs discovery single threaded (non-MT)
4335 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4336 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4337 &_markStack, &_revisitStack, CMSYield && asynch);
4338 // the last argument to iterate indicates whether the iteration
4339 // should be incremental with periodic yields.
4340 _markBitMap.iterate(&markFromRootsClosure);
4341 // If _restart_addr is non-NULL, a marking stack overflow
4342 // occurred; we need to do a fresh iteration from the
4343 // indicated restart address.
4344 while (_restart_addr != NULL) {
4345 if (_foregroundGCIsActive && asynch) {
4346 // We may be running into repeated stack overflows, having
4347 // reached the limit of the stack size, while making very
4348 // slow forward progress. It may be best to bail out and
4349 // let the foreground collector do its job.
4350 // Clear _restart_addr, so that foreground GC
4351 // works from scratch. This avoids the headache of
4352 // a "rescan" which would otherwise be needed because
4353 // of the dirty mod union table & card table.
4354 _restart_addr = NULL;
4355 return false; // indicating failure to complete marking
4356 }
4357 // Deal with stack overflow:
4358 // we restart marking from _restart_addr
4359 HeapWord* ra = _restart_addr;
4360 markFromRootsClosure.reset(ra);
4361 _restart_addr = NULL;
4362 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4363 }
4364 return true;
4365 }
4366
4367 void CMSCollector::preclean() {
4368 check_correct_thread_executing();
4369 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4370 verify_work_stacks_empty();
4371 verify_overflow_empty();
4372 _abort_preclean = false;
4373 if (CMSPrecleaningEnabled) {
4374 _eden_chunk_index = 0;
4375 size_t used = get_eden_used();
4376 size_t capacity = get_eden_capacity();
4377 // Don't start sampling unless we will get sufficiently
4378 // many samples.
4379 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4380 * CMSScheduleRemarkEdenPenetration)) {
4381 _start_sampling = true;
4382 } else {
4383 _start_sampling = false;
4384 }
4385 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4386 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4387 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4388 }
4389 CMSTokenSync x(true); // is cms thread
4390 if (CMSPrecleaningEnabled) {
4391 sample_eden();
4392 _collectorState = AbortablePreclean;
4393 } else {
4394 _collectorState = FinalMarking;
4395 }
4396 verify_work_stacks_empty();
4397 verify_overflow_empty();
4398 }
4399
4400 // Try and schedule the remark such that young gen
4401 // occupancy is CMSScheduleRemarkEdenPenetration %.
4402 void CMSCollector::abortable_preclean() {
4403 check_correct_thread_executing();
4404 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4405 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4406
4407 // If Eden's current occupancy is below this threshold,
4408 // immediately schedule the remark; else preclean
4409 // past the next scavenge in an effort to
4410 // schedule the pause as described avove. By choosing
4411 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4412 // we will never do an actual abortable preclean cycle.
4413 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4414 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4415 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4416 // We need more smarts in the abortable preclean
4417 // loop below to deal with cases where allocation
4418 // in young gen is very very slow, and our precleaning
4419 // is running a losing race against a horde of
4420 // mutators intent on flooding us with CMS updates
4421 // (dirty cards).
4422 // One, admittedly dumb, strategy is to give up
4423 // after a certain number of abortable precleaning loops
4424 // or after a certain maximum time. We want to make
4425 // this smarter in the next iteration.
4426 // XXX FIX ME!!! YSR
4427 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4428 while (!(should_abort_preclean() ||
4429 ConcurrentMarkSweepThread::should_terminate())) {
4430 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4431 cumworkdone += workdone;
4432 loops++;
4433 // Voluntarily terminate abortable preclean phase if we have
4434 // been at it for too long.
4435 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4436 loops >= CMSMaxAbortablePrecleanLoops) {
4437 if (PrintGCDetails) {
4438 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4439 }
4440 break;
4441 }
4442 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4443 if (PrintGCDetails) {
4444 gclog_or_tty->print(" CMS: abort preclean due to time ");
4445 }
4446 break;
4447 }
4448 // If we are doing little work each iteration, we should
4449 // take a short break.
4450 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4451 // Sleep for some time, waiting for work to accumulate
4452 stopTimer();
4453 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4454 startTimer();
4455 waited++;
4456 }
4457 }
4458 if (PrintCMSStatistics > 0) {
4459 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4460 loops, waited, cumworkdone);
4461 }
4462 }
4463 CMSTokenSync x(true); // is cms thread
4464 if (_collectorState != Idling) {
4465 assert(_collectorState == AbortablePreclean,
4466 "Spontaneous state transition?");
4467 _collectorState = FinalMarking;
4468 } // Else, a foreground collection completed this CMS cycle.
4469 return;
4470 }
4471
4472 // Respond to an Eden sampling opportunity
4473 void CMSCollector::sample_eden() {
4474 // Make sure a young gc cannot sneak in between our
4475 // reading and recording of a sample.
4476 assert(Thread::current()->is_ConcurrentGC_thread(),
4477 "Only the cms thread may collect Eden samples");
4478 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4479 "Should collect samples while holding CMS token");
4480 if (!_start_sampling) {
4481 return;
4482 }
4483 if (_eden_chunk_array) {
4484 if (_eden_chunk_index < _eden_chunk_capacity) {
4485 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4486 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4487 "Unexpected state of Eden");
4488 // We'd like to check that what we just sampled is an oop-start address;
4489 // however, we cannot do that here since the object may not yet have been
4490 // initialized. So we'll instead do the check when we _use_ this sample
4491 // later.
4492 if (_eden_chunk_index == 0 ||
4493 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4494 _eden_chunk_array[_eden_chunk_index-1])
4495 >= CMSSamplingGrain)) {
4496 _eden_chunk_index++; // commit sample
4497 }
4498 }
4499 }
4500 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4501 size_t used = get_eden_used();
4502 size_t capacity = get_eden_capacity();
4503 assert(used <= capacity, "Unexpected state of Eden");
4504 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4505 _abort_preclean = true;
4506 }
4507 }
4508 }
4509
4510
4511 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4512 assert(_collectorState == Precleaning ||
4513 _collectorState == AbortablePreclean, "incorrect state");
4514 ResourceMark rm;
4515 HandleMark hm;
4516
4517 // Precleaning is currently not MT but the reference processor
4518 // may be set for MT. Disable it temporarily here.
4519 ReferenceProcessor* rp = ref_processor();
4520 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4521
4522 // Do one pass of scrubbing the discovered reference lists
4523 // to remove any reference objects with strongly-reachable
4524 // referents.
4525 if (clean_refs) {
4526 CMSPrecleanRefsYieldClosure yield_cl(this);
4527 assert(rp->span().equals(_span), "Spans should be equal");
4528 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4529 &_markStack, &_revisitStack,
4530 true /* preclean */);
4531 CMSDrainMarkingStackClosure complete_trace(this,
4532 _span, &_markBitMap, &_markStack,
4533 &keep_alive, true /* preclean */);
4534
4535 // We don't want this step to interfere with a young
4536 // collection because we don't want to take CPU
4537 // or memory bandwidth away from the young GC threads
4538 // (which may be as many as there are CPUs).
4539 // Note that we don't need to protect ourselves from
4540 // interference with mutators because they can't
4541 // manipulate the discovered reference lists nor affect
4542 // the computed reachability of the referents, the
4543 // only properties manipulated by the precleaning
4544 // of these reference lists.
4545 stopTimer();
4546 CMSTokenSyncWithLocks x(true /* is cms thread */,
4547 bitMapLock());
4548 startTimer();
4549 sample_eden();
4550
4551 // The following will yield to allow foreground
4552 // collection to proceed promptly. XXX YSR:
4553 // The code in this method may need further
4554 // tweaking for better performance and some restructuring
4555 // for cleaner interfaces.
4556 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4557 rp->preclean_discovered_references(
4558 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4559 &yield_cl, should_unload_classes(), gc_timer);
4560 }
4561
4562 if (clean_survivor) { // preclean the active survivor space(s)
4563 assert(_young_gen->kind() == Generation::DefNew ||
4564 _young_gen->kind() == Generation::ParNew ||
4565 _young_gen->kind() == Generation::ASParNew,
4566 "incorrect type for cast");
4567 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4568 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4569 &_markBitMap, &_modUnionTable,
4570 &_markStack, &_revisitStack,
4571 true /* precleaning phase */);
4572 stopTimer();
4573 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4574 bitMapLock());
4575 startTimer();
4576 unsigned int before_count =
4577 GenCollectedHeap::heap()->total_collections();
4578 SurvivorSpacePrecleanClosure
4579 sss_cl(this, _span, &_markBitMap, &_markStack,
4580 &pam_cl, before_count, CMSYield);
4581 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4582 dng->from()->object_iterate_careful(&sss_cl);
4583 dng->to()->object_iterate_careful(&sss_cl);
4584 }
4585 MarkRefsIntoAndScanClosure
4586 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4587 &_markStack, &_revisitStack, this, CMSYield,
4588 true /* precleaning phase */);
4589 // CAUTION: The following closure has persistent state that may need to
4590 // be reset upon a decrease in the sequence of addresses it
4591 // processes.
4592 ScanMarkedObjectsAgainCarefullyClosure
4593 smoac_cl(this, _span,
4594 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4595
4596 // Preclean dirty cards in ModUnionTable and CardTable using
4597 // appropriate convergence criterion;
4598 // repeat CMSPrecleanIter times unless we find that
4599 // we are losing.
4600 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4601 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4602 "Bad convergence multiplier");
4603 assert(CMSPrecleanThreshold >= 100,
4604 "Unreasonably low CMSPrecleanThreshold");
4605
4606 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4607 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4608 numIter < CMSPrecleanIter;
4609 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4610 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4611 if (CMSPermGenPrecleaningEnabled) {
4612 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4613 }
4614 if (Verbose && PrintGCDetails) {
4615 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4616 }
4617 // Either there are very few dirty cards, so re-mark
4618 // pause will be small anyway, or our pre-cleaning isn't
4619 // that much faster than the rate at which cards are being
4620 // dirtied, so we might as well stop and re-mark since
4621 // precleaning won't improve our re-mark time by much.
4622 if (curNumCards <= CMSPrecleanThreshold ||
4623 (numIter > 0 &&
4624 (curNumCards * CMSPrecleanDenominator >
4625 lastNumCards * CMSPrecleanNumerator))) {
4626 numIter++;
4627 cumNumCards += curNumCards;
4628 break;
4629 }
4630 }
4631 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4632 if (CMSPermGenPrecleaningEnabled) {
4633 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4634 }
4635 cumNumCards += curNumCards;
4636 if (PrintGCDetails && PrintCMSStatistics != 0) {
4637 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4638 curNumCards, cumNumCards, numIter);
4639 }
4640 return cumNumCards; // as a measure of useful work done
4641 }
4642
4643 // PRECLEANING NOTES:
4644 // Precleaning involves:
4645 // . reading the bits of the modUnionTable and clearing the set bits.
4646 // . For the cards corresponding to the set bits, we scan the
4647 // objects on those cards. This means we need the free_list_lock
4648 // so that we can safely iterate over the CMS space when scanning
4649 // for oops.
4650 // . When we scan the objects, we'll be both reading and setting
4651 // marks in the marking bit map, so we'll need the marking bit map.
4652 // . For protecting _collector_state transitions, we take the CGC_lock.
4653 // Note that any races in the reading of of card table entries by the
4654 // CMS thread on the one hand and the clearing of those entries by the
4655 // VM thread or the setting of those entries by the mutator threads on the
4656 // other are quite benign. However, for efficiency it makes sense to keep
4657 // the VM thread from racing with the CMS thread while the latter is
4658 // dirty card info to the modUnionTable. We therefore also use the
4659 // CGC_lock to protect the reading of the card table and the mod union
4660 // table by the CM thread.
4661 // . We run concurrently with mutator updates, so scanning
4662 // needs to be done carefully -- we should not try to scan
4663 // potentially uninitialized objects.
4664 //
4665 // Locking strategy: While holding the CGC_lock, we scan over and
4666 // reset a maximal dirty range of the mod union / card tables, then lock
4667 // the free_list_lock and bitmap lock to do a full marking, then
4668 // release these locks; and repeat the cycle. This allows for a
4669 // certain amount of fairness in the sharing of these locks between
4670 // the CMS collector on the one hand, and the VM thread and the
4671 // mutators on the other.
4672
4673 // NOTE: preclean_mod_union_table() and preclean_card_table()
4674 // further below are largely identical; if you need to modify
4675 // one of these methods, please check the other method too.
4676
4677 size_t CMSCollector::preclean_mod_union_table(
4678 ConcurrentMarkSweepGeneration* gen,
4679 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4680 verify_work_stacks_empty();
4681 verify_overflow_empty();
4682
4683 // Turn off checking for this method but turn it back on
4684 // selectively. There are yield points in this method
4685 // but it is difficult to turn the checking off just around
4686 // the yield points. It is simpler to selectively turn
4687 // it on.
4688 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4689
4690 // strategy: starting with the first card, accumulate contiguous
4691 // ranges of dirty cards; clear these cards, then scan the region
4692 // covered by these cards.
4693
4694 // Since all of the MUT is committed ahead, we can just use
4695 // that, in case the generations expand while we are precleaning.
4696 // It might also be fine to just use the committed part of the
4697 // generation, but we might potentially miss cards when the
4698 // generation is rapidly expanding while we are in the midst
4699 // of precleaning.
4700 HeapWord* startAddr = gen->reserved().start();
4701 HeapWord* endAddr = gen->reserved().end();
4702
4703 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4704
4705 size_t numDirtyCards, cumNumDirtyCards;
4706 HeapWord *nextAddr, *lastAddr;
4707 for (cumNumDirtyCards = numDirtyCards = 0,
4708 nextAddr = lastAddr = startAddr;
4709 nextAddr < endAddr;
4710 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4711
4712 ResourceMark rm;
4713 HandleMark hm;
4714
4715 MemRegion dirtyRegion;
4716 {
4717 stopTimer();
4718 // Potential yield point
4719 CMSTokenSync ts(true);
4720 startTimer();
4721 sample_eden();
4722 // Get dirty region starting at nextOffset (inclusive),
4723 // simultaneously clearing it.
4724 dirtyRegion =
4725 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4726 assert(dirtyRegion.start() >= nextAddr,
4727 "returned region inconsistent?");
4728 }
4729 // Remember where the next search should begin.
4730 // The returned region (if non-empty) is a right open interval,
4731 // so lastOffset is obtained from the right end of that
4732 // interval.
4733 lastAddr = dirtyRegion.end();
4734 // Should do something more transparent and less hacky XXX
4735 numDirtyCards =
4736 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4737
4738 // We'll scan the cards in the dirty region (with periodic
4739 // yields for foreground GC as needed).
4740 if (!dirtyRegion.is_empty()) {
4741 assert(numDirtyCards > 0, "consistency check");
4742 HeapWord* stop_point = NULL;
4743 stopTimer();
4744 // Potential yield point
4745 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4746 bitMapLock());
4747 startTimer();
4748 {
4749 verify_work_stacks_empty();
4750 verify_overflow_empty();
4751 sample_eden();
4752 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4753 stop_point =
4754 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4755 }
4756 if (stop_point != NULL) {
4757 // The careful iteration stopped early either because it found an
4758 // uninitialized object, or because we were in the midst of an
4759 // "abortable preclean", which should now be aborted. Redirty
4760 // the bits corresponding to the partially-scanned or unscanned
4761 // cards. We'll either restart at the next block boundary or
4762 // abort the preclean.
4763 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4764 (_collectorState == AbortablePreclean && should_abort_preclean()),
4765 "Unparsable objects should only be in perm gen.");
4766 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4767 if (should_abort_preclean()) {
4768 break; // out of preclean loop
4769 } else {
4770 // Compute the next address at which preclean should pick up;
4771 // might need bitMapLock in order to read P-bits.
4772 lastAddr = next_card_start_after_block(stop_point);
4773 }
4774 }
4775 } else {
4776 assert(lastAddr == endAddr, "consistency check");
4777 assert(numDirtyCards == 0, "consistency check");
4778 break;
4779 }
4780 }
4781 verify_work_stacks_empty();
4782 verify_overflow_empty();
4783 return cumNumDirtyCards;
4784 }
4785
4786 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4787 // below are largely identical; if you need to modify
4788 // one of these methods, please check the other method too.
4789
4790 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4791 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4792 // strategy: it's similar to precleamModUnionTable above, in that
4793 // we accumulate contiguous ranges of dirty cards, mark these cards
4794 // precleaned, then scan the region covered by these cards.
4795 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4796 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4797
4798 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4799
4800 size_t numDirtyCards, cumNumDirtyCards;
4801 HeapWord *lastAddr, *nextAddr;
4802
4803 for (cumNumDirtyCards = numDirtyCards = 0,
4804 nextAddr = lastAddr = startAddr;
4805 nextAddr < endAddr;
4806 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4807
4808 ResourceMark rm;
4809 HandleMark hm;
4810
4811 MemRegion dirtyRegion;
4812 {
4813 // See comments in "Precleaning notes" above on why we
4814 // do this locking. XXX Could the locking overheads be
4815 // too high when dirty cards are sparse? [I don't think so.]
4816 stopTimer();
4817 CMSTokenSync x(true); // is cms thread
4818 startTimer();
4819 sample_eden();
4820 // Get and clear dirty region from card table
4821 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4822 MemRegion(nextAddr, endAddr),
4823 true,
4824 CardTableModRefBS::precleaned_card_val());
4825
4826 assert(dirtyRegion.start() >= nextAddr,
4827 "returned region inconsistent?");
4828 }
4829 lastAddr = dirtyRegion.end();
4830 numDirtyCards =
4831 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4832
4833 if (!dirtyRegion.is_empty()) {
4834 stopTimer();
4835 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4836 startTimer();
4837 sample_eden();
4838 verify_work_stacks_empty();
4839 verify_overflow_empty();
4840 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4841 HeapWord* stop_point =
4842 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4843 if (stop_point != NULL) {
4844 // The careful iteration stopped early because it found an
4845 // uninitialized object. Redirty the bits corresponding to the
4846 // partially-scanned or unscanned cards, and start again at the
4847 // next block boundary.
4848 assert(CMSPermGenPrecleaningEnabled ||
4849 (_collectorState == AbortablePreclean && should_abort_preclean()),
4850 "Unparsable objects should only be in perm gen.");
4851 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4852 if (should_abort_preclean()) {
4853 break; // out of preclean loop
4854 } else {
4855 // Compute the next address at which preclean should pick up.
4856 lastAddr = next_card_start_after_block(stop_point);
4857 }
4858 }
4859 } else {
4860 break;
4861 }
4862 }
4863 verify_work_stacks_empty();
4864 verify_overflow_empty();
4865 return cumNumDirtyCards;
4866 }
4867
4868 void CMSCollector::checkpointRootsFinal(bool asynch,
4869 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4870 assert(_collectorState == FinalMarking, "incorrect state transition?");
4871 check_correct_thread_executing();
4872 // world is stopped at this checkpoint
4873 assert(SafepointSynchronize::is_at_safepoint(),
4874 "world should be stopped");
4875 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4876
4877 verify_work_stacks_empty();
4878 verify_overflow_empty();
4879
4880 SpecializationStats::clear();
4881 if (PrintGCDetails) {
4882 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4883 _young_gen->used() / K,
4884 _young_gen->capacity() / K);
4885 }
4886 if (asynch) {
4887 if (CMSScavengeBeforeRemark) {
4888 GenCollectedHeap* gch = GenCollectedHeap::heap();
4889 // Temporarily set flag to false, GCH->do_collection will
4890 // expect it to be false and set to true
4891 FlagSetting fl(gch->_is_gc_active, false);
4892 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4893 PrintGCDetails && Verbose, true, _gc_timer_cm);)
4894 int level = _cmsGen->level() - 1;
4895 if (level >= 0) {
4896 gch->do_collection(true, // full (i.e. force, see below)
4897 false, // !clear_all_soft_refs
4898 0, // size
4899 false, // is_tlab
4900 level // max_level
4901 );
4902 }
4903 }
4904 FreelistLocker x(this);
4905 MutexLockerEx y(bitMapLock(),
4906 Mutex::_no_safepoint_check_flag);
4907 assert(!init_mark_was_synchronous, "but that's impossible!");
4908 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4909 } else {
4910 // already have all the locks
4911 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4912 init_mark_was_synchronous);
4913 }
4914 verify_work_stacks_empty();
4915 verify_overflow_empty();
4916 SpecializationStats::print();
4917 }
4918
4919 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4920 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4921
4922 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
4923
4924 assert(haveFreelistLocks(), "must have free list locks");
4925 assert_lock_strong(bitMapLock());
4926
4927 if (UseAdaptiveSizePolicy) {
4928 size_policy()->checkpoint_roots_final_begin();
4929 }
4930
4931 ResourceMark rm;
4932 HandleMark hm;
4933
4934 GenCollectedHeap* gch = GenCollectedHeap::heap();
4935
4936 if (should_unload_classes()) {
4937 CodeCache::gc_prologue();
4938 }
4939 assert(haveFreelistLocks(), "must have free list locks");
4940 assert_lock_strong(bitMapLock());
4941
4942 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());)
4943 if (!init_mark_was_synchronous) {
4944 // We might assume that we need not fill TLAB's when
4945 // CMSScavengeBeforeRemark is set, because we may have just done
4946 // a scavenge which would have filled all TLAB's -- and besides
4947 // Eden would be empty. This however may not always be the case --
4948 // for instance although we asked for a scavenge, it may not have
4949 // happened because of a JNI critical section. We probably need
4950 // a policy for deciding whether we can in that case wait until
4951 // the critical section releases and then do the remark following
4952 // the scavenge, and skip it here. In the absence of that policy,
4953 // or of an indication of whether the scavenge did indeed occur,
4954 // we cannot rely on TLAB's having been filled and must do
4955 // so here just in case a scavenge did not happen.
4956 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4957 // Update the saved marks which may affect the root scans.
4958 gch->save_marks();
4959
4960 {
4961 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4962
4963 // Note on the role of the mod union table:
4964 // Since the marker in "markFromRoots" marks concurrently with
4965 // mutators, it is possible for some reachable objects not to have been
4966 // scanned. For instance, an only reference to an object A was
4967 // placed in object B after the marker scanned B. Unless B is rescanned,
4968 // A would be collected. Such updates to references in marked objects
4969 // are detected via the mod union table which is the set of all cards
4970 // dirtied since the first checkpoint in this GC cycle and prior to
4971 // the most recent young generation GC, minus those cleaned up by the
4972 // concurrent precleaning.
4973 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
4974 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
4975 do_remark_parallel();
4976 } else {
4977 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
4978 _gc_timer_cm);
4979 do_remark_non_parallel();
4980 }
4981 }
4982 } else {
4983 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
4984 // The initial mark was stop-world, so there's no rescanning to
4985 // do; go straight on to the next step below.
4986 }
4987 verify_work_stacks_empty();
4988 verify_overflow_empty();
4989
4990 {
4991 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
4992 refProcessingWork(asynch, clear_all_soft_refs);
4993 }
4994 verify_work_stacks_empty();
4995 verify_overflow_empty();
4996
4997 if (should_unload_classes()) {
4998 CodeCache::gc_epilogue();
4999 }
5000 JvmtiExport::gc_epilogue();
5001
5002 // If we encountered any (marking stack / work queue) overflow
5003 // events during the current CMS cycle, take appropriate
5004 // remedial measures, where possible, so as to try and avoid
5005 // recurrence of that condition.
5006 assert(_markStack.isEmpty(), "No grey objects");
5007 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5008 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5009 if (ser_ovflw > 0) {
5010 if (PrintCMSStatistics != 0) {
5011 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5012 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5013 ", kac_preclean="SIZE_FORMAT")",
5014 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5015 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5016 }
5017 _markStack.expand();
5018 _ser_pmc_remark_ovflw = 0;
5019 _ser_pmc_preclean_ovflw = 0;
5020 _ser_kac_preclean_ovflw = 0;
5021 _ser_kac_ovflw = 0;
5022 }
5023 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5024 if (PrintCMSStatistics != 0) {
5025 gclog_or_tty->print_cr("Work queue overflow (benign) "
5026 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5027 _par_pmc_remark_ovflw, _par_kac_ovflw);
5028 }
5029 _par_pmc_remark_ovflw = 0;
5030 _par_kac_ovflw = 0;
5031 }
5032 if (PrintCMSStatistics != 0) {
5033 if (_markStack._hit_limit > 0) {
5034 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5035 _markStack._hit_limit);
5036 }
5037 if (_markStack._failed_double > 0) {
5038 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5039 " current capacity "SIZE_FORMAT,
5040 _markStack._failed_double,
5041 _markStack.capacity());
5042 }
5043 }
5044 _markStack._hit_limit = 0;
5045 _markStack._failed_double = 0;
5046
5047 // Check that all the klasses have been checked
5048 assert(_revisitStack.isEmpty(), "Not all klasses revisited");
5049
5050 if ((VerifyAfterGC || VerifyDuringGC) &&
5051 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5052 verify_after_remark();
5053 }
5054
5055 // Change under the freelistLocks.
5056 _collectorState = Sweeping;
5057 // Call isAllClear() under bitMapLock
5058 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
5059 " final marking");
5060 if (UseAdaptiveSizePolicy) {
5061 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5062 }
5063 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5064 }
5065
5066 // Parallel remark task
5067 class CMSParRemarkTask: public AbstractGangTask {
5068 CMSCollector* _collector;
5069 int _n_workers;
5070 CompactibleFreeListSpace* _cms_space;
5071 CompactibleFreeListSpace* _perm_space;
5072
5073 // The per-thread work queues, available here for stealing.
5074 OopTaskQueueSet* _task_queues;
5075 ParallelTaskTerminator _term;
5076
5077 public:
5078 // A value of 0 passed to n_workers will cause the number of
5079 // workers to be taken from the active workers in the work gang.
5080 CMSParRemarkTask(CMSCollector* collector,
5081 CompactibleFreeListSpace* cms_space,
5082 CompactibleFreeListSpace* perm_space,
5083 int n_workers, FlexibleWorkGang* workers,
5084 OopTaskQueueSet* task_queues):
5085 AbstractGangTask("Rescan roots and grey objects in parallel"),
5086 _collector(collector),
5087 _cms_space(cms_space), _perm_space(perm_space),
5088 _n_workers(n_workers),
5089 _task_queues(task_queues),
5090 _term(n_workers, task_queues) { }
5091
5092 OopTaskQueueSet* task_queues() { return _task_queues; }
5093
5094 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5095
5096 ParallelTaskTerminator* terminator() { return &_term; }
5097 int n_workers() { return _n_workers; }
5098
5099 void work(uint worker_id);
5100
5101 private:
5102 // Work method in support of parallel rescan ... of young gen spaces
5103 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
5104 ContiguousSpace* space,
5105 HeapWord** chunk_array, size_t chunk_top);
5106
5107 // ... of dirty cards in old space
5108 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5109 Par_MarkRefsIntoAndScanClosure* cl);
5110
5111 // ... work stealing for the above
5112 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5113 };
5114
5115 // work_queue(i) is passed to the closure
5116 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5117 // also is passed to do_dirty_card_rescan_tasks() and to
5118 // do_work_steal() to select the i-th task_queue.
5119
5120 void CMSParRemarkTask::work(uint worker_id) {
5121 elapsedTimer _timer;
5122 ResourceMark rm;
5123 HandleMark hm;
5124
5125 // ---------- rescan from roots --------------
5126 _timer.start();
5127 GenCollectedHeap* gch = GenCollectedHeap::heap();
5128 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5129 _collector->_span, _collector->ref_processor(),
5130 &(_collector->_markBitMap),
5131 work_queue(worker_id), &(_collector->_revisitStack));
5132
5133 // Rescan young gen roots first since these are likely
5134 // coarsely partitioned and may, on that account, constitute
5135 // the critical path; thus, it's best to start off that
5136 // work first.
5137 // ---------- young gen roots --------------
5138 {
5139 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5140 EdenSpace* eden_space = dng->eden();
5141 ContiguousSpace* from_space = dng->from();
5142 ContiguousSpace* to_space = dng->to();
5143
5144 HeapWord** eca = _collector->_eden_chunk_array;
5145 size_t ect = _collector->_eden_chunk_index;
5146 HeapWord** sca = _collector->_survivor_chunk_array;
5147 size_t sct = _collector->_survivor_chunk_index;
5148
5149 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5150 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5151
5152 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0);
5153 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct);
5154 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect);
5155
5156 _timer.stop();
5157 if (PrintCMSStatistics != 0) {
5158 gclog_or_tty->print_cr(
5159 "Finished young gen rescan work in %dth thread: %3.3f sec",
5160 worker_id, _timer.seconds());
5161 }
5162 }
5163
5164 // ---------- remaining roots --------------
5165 _timer.reset();
5166 _timer.start();
5167 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5168 false, // yg was scanned above
5169 false, // this is parallel code
5170 true, // collecting perm gen
5171 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5172 &par_mrias_cl,
5173 true, // walk all of code cache if (so & SO_CodeCache)
5174 NULL);
5175 assert(_collector->should_unload_classes()
5176 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5177 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5178 _timer.stop();
5179 if (PrintCMSStatistics != 0) {
5180 gclog_or_tty->print_cr(
5181 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5182 worker_id, _timer.seconds());
5183 }
5184
5185 // ---------- rescan dirty cards ------------
5186 _timer.reset();
5187 _timer.start();
5188
5189 // Do the rescan tasks for each of the two spaces
5190 // (cms_space and perm_space) in turn.
5191 // "worker_id" is passed to select the task_queue for "worker_id"
5192 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
5193 do_dirty_card_rescan_tasks(_perm_space, worker_id, &par_mrias_cl);
5194 _timer.stop();
5195 if (PrintCMSStatistics != 0) {
5196 gclog_or_tty->print_cr(
5197 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5198 worker_id, _timer.seconds());
5199 }
5200
5201 // ---------- steal work from other threads ...
5202 // ---------- ... and drain overflow list.
5203 _timer.reset();
5204 _timer.start();
5205 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
5206 _timer.stop();
5207 if (PrintCMSStatistics != 0) {
5208 gclog_or_tty->print_cr(
5209 "Finished work stealing in %dth thread: %3.3f sec",
5210 worker_id, _timer.seconds());
5211 }
5212 }
5213
5214 // Note that parameter "i" is not used.
5215 void
5216 CMSParRemarkTask::do_young_space_rescan(int i,
5217 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5218 HeapWord** chunk_array, size_t chunk_top) {
5219 // Until all tasks completed:
5220 // . claim an unclaimed task
5221 // . compute region boundaries corresponding to task claimed
5222 // using chunk_array
5223 // . par_oop_iterate(cl) over that region
5224
5225 ResourceMark rm;
5226 HandleMark hm;
5227
5228 SequentialSubTasksDone* pst = space->par_seq_tasks();
5229 assert(pst->valid(), "Uninitialized use?");
5230
5231 uint nth_task = 0;
5232 uint n_tasks = pst->n_tasks();
5233
5234 HeapWord *start, *end;
5235 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5236 // We claimed task # nth_task; compute its boundaries.
5237 if (chunk_top == 0) { // no samples were taken
5238 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5239 start = space->bottom();
5240 end = space->top();
5241 } else if (nth_task == 0) {
5242 start = space->bottom();
5243 end = chunk_array[nth_task];
5244 } else if (nth_task < (uint)chunk_top) {
5245 assert(nth_task >= 1, "Control point invariant");
5246 start = chunk_array[nth_task - 1];
5247 end = chunk_array[nth_task];
5248 } else {
5249 assert(nth_task == (uint)chunk_top, "Control point invariant");
5250 start = chunk_array[chunk_top - 1];
5251 end = space->top();
5252 }
5253 MemRegion mr(start, end);
5254 // Verify that mr is in space
5255 assert(mr.is_empty() || space->used_region().contains(mr),
5256 "Should be in space");
5257 // Verify that "start" is an object boundary
5258 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5259 "Should be an oop");
5260 space->par_oop_iterate(mr, cl);
5261 }
5262 pst->all_tasks_completed();
5263 }
5264
5265 void
5266 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5267 CompactibleFreeListSpace* sp, int i,
5268 Par_MarkRefsIntoAndScanClosure* cl) {
5269 // Until all tasks completed:
5270 // . claim an unclaimed task
5271 // . compute region boundaries corresponding to task claimed
5272 // . transfer dirty bits ct->mut for that region
5273 // . apply rescanclosure to dirty mut bits for that region
5274
5275 ResourceMark rm;
5276 HandleMark hm;
5277
5278 OopTaskQueue* work_q = work_queue(i);
5279 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5280 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5281 // CAUTION: This closure has state that persists across calls to
5282 // the work method dirty_range_iterate_clear() in that it has
5283 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5284 // use of that state in the imbedded UpwardsObjectClosure instance
5285 // assumes that the cards are always iterated (even if in parallel
5286 // by several threads) in monotonically increasing order per each
5287 // thread. This is true of the implementation below which picks
5288 // card ranges (chunks) in monotonically increasing order globally
5289 // and, a-fortiori, in monotonically increasing order per thread
5290 // (the latter order being a subsequence of the former).
5291 // If the work code below is ever reorganized into a more chaotic
5292 // work-partitioning form than the current "sequential tasks"
5293 // paradigm, the use of that persistent state will have to be
5294 // revisited and modified appropriately. See also related
5295 // bug 4756801 work on which should examine this code to make
5296 // sure that the changes there do not run counter to the
5297 // assumptions made here and necessary for correctness and
5298 // efficiency. Note also that this code might yield inefficient
5299 // behaviour in the case of very large objects that span one or
5300 // more work chunks. Such objects would potentially be scanned
5301 // several times redundantly. Work on 4756801 should try and
5302 // address that performance anomaly if at all possible. XXX
5303 MemRegion full_span = _collector->_span;
5304 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5305 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5306 MarkFromDirtyCardsClosure
5307 greyRescanClosure(_collector, full_span, // entire span of interest
5308 sp, bm, work_q, rs, cl);
5309
5310 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5311 assert(pst->valid(), "Uninitialized use?");
5312 uint nth_task = 0;
5313 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5314 MemRegion span = sp->used_region();
5315 HeapWord* start_addr = span.start();
5316 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5317 alignment);
5318 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5319 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5320 start_addr, "Check alignment");
5321 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5322 chunk_size, "Check alignment");
5323
5324 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5325 // Having claimed the nth_task, compute corresponding mem-region,
5326 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5327 // The alignment restriction ensures that we do not need any
5328 // synchronization with other gang-workers while setting or
5329 // clearing bits in thus chunk of the MUT.
5330 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5331 start_addr + (nth_task+1)*chunk_size);
5332 // The last chunk's end might be way beyond end of the
5333 // used region. In that case pull back appropriately.
5334 if (this_span.end() > end_addr) {
5335 this_span.set_end(end_addr);
5336 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5337 }
5338 // Iterate over the dirty cards covering this chunk, marking them
5339 // precleaned, and setting the corresponding bits in the mod union
5340 // table. Since we have been careful to partition at Card and MUT-word
5341 // boundaries no synchronization is needed between parallel threads.
5342 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5343 &modUnionClosure);
5344
5345 // Having transferred these marks into the modUnionTable,
5346 // rescan the marked objects on the dirty cards in the modUnionTable.
5347 // Even if this is at a synchronous collection, the initial marking
5348 // may have been done during an asynchronous collection so there
5349 // may be dirty bits in the mod-union table.
5350 _collector->_modUnionTable.dirty_range_iterate_clear(
5351 this_span, &greyRescanClosure);
5352 _collector->_modUnionTable.verifyNoOneBitsInRange(
5353 this_span.start(),
5354 this_span.end());
5355 }
5356 pst->all_tasks_completed(); // declare that i am done
5357 }
5358
5359 // . see if we can share work_queues with ParNew? XXX
5360 void
5361 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5362 int* seed) {
5363 OopTaskQueue* work_q = work_queue(i);
5364 NOT_PRODUCT(int num_steals = 0;)
5365 oop obj_to_scan;
5366 CMSBitMap* bm = &(_collector->_markBitMap);
5367
5368 while (true) {
5369 // Completely finish any left over work from (an) earlier round(s)
5370 cl->trim_queue(0);
5371 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5372 (size_t)ParGCDesiredObjsFromOverflowList);
5373 // Now check if there's any work in the overflow list
5374 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5375 // only affects the number of attempts made to get work from the
5376 // overflow list and does not affect the number of workers. Just
5377 // pass ParallelGCThreads so this behavior is unchanged.
5378 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5379 work_q,
5380 ParallelGCThreads)) {
5381 // found something in global overflow list;
5382 // not yet ready to go stealing work from others.
5383 // We'd like to assert(work_q->size() != 0, ...)
5384 // because we just took work from the overflow list,
5385 // but of course we can't since all of that could have
5386 // been already stolen from us.
5387 // "He giveth and He taketh away."
5388 continue;
5389 }
5390 // Verify that we have no work before we resort to stealing
5391 assert(work_q->size() == 0, "Have work, shouldn't steal");
5392 // Try to steal from other queues that have work
5393 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5394 NOT_PRODUCT(num_steals++;)
5395 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5396 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5397 // Do scanning work
5398 obj_to_scan->oop_iterate(cl);
5399 // Loop around, finish this work, and try to steal some more
5400 } else if (terminator()->offer_termination()) {
5401 break; // nirvana from the infinite cycle
5402 }
5403 }
5404 NOT_PRODUCT(
5405 if (PrintCMSStatistics != 0) {
5406 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5407 }
5408 )
5409 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5410 "Else our work is not yet done");
5411 }
5412
5413 // Return a thread-local PLAB recording array, as appropriate.
5414 void* CMSCollector::get_data_recorder(int thr_num) {
5415 if (_survivor_plab_array != NULL &&
5416 (CMSPLABRecordAlways ||
5417 (_collectorState > Marking && _collectorState < FinalMarking))) {
5418 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5419 ChunkArray* ca = &_survivor_plab_array[thr_num];
5420 ca->reset(); // clear it so that fresh data is recorded
5421 return (void*) ca;
5422 } else {
5423 return NULL;
5424 }
5425 }
5426
5427 // Reset all the thread-local PLAB recording arrays
5428 void CMSCollector::reset_survivor_plab_arrays() {
5429 for (uint i = 0; i < ParallelGCThreads; i++) {
5430 _survivor_plab_array[i].reset();
5431 }
5432 }
5433
5434 // Merge the per-thread plab arrays into the global survivor chunk
5435 // array which will provide the partitioning of the survivor space
5436 // for CMS rescan.
5437 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5438 int no_of_gc_threads) {
5439 assert(_survivor_plab_array != NULL, "Error");
5440 assert(_survivor_chunk_array != NULL, "Error");
5441 assert(_collectorState == FinalMarking, "Error");
5442 for (int j = 0; j < no_of_gc_threads; j++) {
5443 _cursor[j] = 0;
5444 }
5445 HeapWord* top = surv->top();
5446 size_t i;
5447 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5448 HeapWord* min_val = top; // Higher than any PLAB address
5449 uint min_tid = 0; // position of min_val this round
5450 for (int j = 0; j < no_of_gc_threads; j++) {
5451 ChunkArray* cur_sca = &_survivor_plab_array[j];
5452 if (_cursor[j] == cur_sca->end()) {
5453 continue;
5454 }
5455 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5456 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5457 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5458 if (cur_val < min_val) {
5459 min_tid = j;
5460 min_val = cur_val;
5461 } else {
5462 assert(cur_val < top, "All recorded addresses should be less");
5463 }
5464 }
5465 // At this point min_val and min_tid are respectively
5466 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5467 // and the thread (j) that witnesses that address.
5468 // We record this address in the _survivor_chunk_array[i]
5469 // and increment _cursor[min_tid] prior to the next round i.
5470 if (min_val == top) {
5471 break;
5472 }
5473 _survivor_chunk_array[i] = min_val;
5474 _cursor[min_tid]++;
5475 }
5476 // We are all done; record the size of the _survivor_chunk_array
5477 _survivor_chunk_index = i; // exclusive: [0, i)
5478 if (PrintCMSStatistics > 0) {
5479 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5480 }
5481 // Verify that we used up all the recorded entries
5482 #ifdef ASSERT
5483 size_t total = 0;
5484 for (int j = 0; j < no_of_gc_threads; j++) {
5485 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5486 total += _cursor[j];
5487 }
5488 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5489 // Check that the merged array is in sorted order
5490 if (total > 0) {
5491 for (size_t i = 0; i < total - 1; i++) {
5492 if (PrintCMSStatistics > 0) {
5493 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5494 i, _survivor_chunk_array[i]);
5495 }
5496 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5497 "Not sorted");
5498 }
5499 }
5500 #endif // ASSERT
5501 }
5502
5503 // Set up the space's par_seq_tasks structure for work claiming
5504 // for parallel rescan of young gen.
5505 // See ParRescanTask where this is currently used.
5506 void
5507 CMSCollector::
5508 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5509 assert(n_threads > 0, "Unexpected n_threads argument");
5510 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5511
5512 // Eden space
5513 {
5514 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5515 assert(!pst->valid(), "Clobbering existing data?");
5516 // Each valid entry in [0, _eden_chunk_index) represents a task.
5517 size_t n_tasks = _eden_chunk_index + 1;
5518 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5519 // Sets the condition for completion of the subtask (how many threads
5520 // need to finish in order to be done).
5521 pst->set_n_threads(n_threads);
5522 pst->set_n_tasks((int)n_tasks);
5523 }
5524
5525 // Merge the survivor plab arrays into _survivor_chunk_array
5526 if (_survivor_plab_array != NULL) {
5527 merge_survivor_plab_arrays(dng->from(), n_threads);
5528 } else {
5529 assert(_survivor_chunk_index == 0, "Error");
5530 }
5531
5532 // To space
5533 {
5534 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5535 assert(!pst->valid(), "Clobbering existing data?");
5536 // Sets the condition for completion of the subtask (how many threads
5537 // need to finish in order to be done).
5538 pst->set_n_threads(n_threads);
5539 pst->set_n_tasks(1);
5540 assert(pst->valid(), "Error");
5541 }
5542
5543 // From space
5544 {
5545 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5546 assert(!pst->valid(), "Clobbering existing data?");
5547 size_t n_tasks = _survivor_chunk_index + 1;
5548 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5549 // Sets the condition for completion of the subtask (how many threads
5550 // need to finish in order to be done).
5551 pst->set_n_threads(n_threads);
5552 pst->set_n_tasks((int)n_tasks);
5553 assert(pst->valid(), "Error");
5554 }
5555 }
5556
5557 // Parallel version of remark
5558 void CMSCollector::do_remark_parallel() {
5559 GenCollectedHeap* gch = GenCollectedHeap::heap();
5560 FlexibleWorkGang* workers = gch->workers();
5561 assert(workers != NULL, "Need parallel worker threads.");
5562 // Choose to use the number of GC workers most recently set
5563 // into "active_workers". If active_workers is not set, set it
5564 // to ParallelGCThreads.
5565 int n_workers = workers->active_workers();
5566 if (n_workers == 0) {
5567 assert(n_workers > 0, "Should have been set during scavenge");
5568 n_workers = ParallelGCThreads;
5569 workers->set_active_workers(n_workers);
5570 }
5571 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5572 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5573
5574 CMSParRemarkTask tsk(this,
5575 cms_space, perm_space,
5576 n_workers, workers, task_queues());
5577
5578 // Set up for parallel process_strong_roots work.
5579 gch->set_par_threads(n_workers);
5580 // We won't be iterating over the cards in the card table updating
5581 // the younger_gen cards, so we shouldn't call the following else
5582 // the verification code as well as subsequent younger_refs_iterate
5583 // code would get confused. XXX
5584 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5585
5586 // The young gen rescan work will not be done as part of
5587 // process_strong_roots (which currently doesn't knw how to
5588 // parallelize such a scan), but rather will be broken up into
5589 // a set of parallel tasks (via the sampling that the [abortable]
5590 // preclean phase did of EdenSpace, plus the [two] tasks of
5591 // scanning the [two] survivor spaces. Further fine-grain
5592 // parallelization of the scanning of the survivor spaces
5593 // themselves, and of precleaning of the younger gen itself
5594 // is deferred to the future.
5595 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5596
5597 // The dirty card rescan work is broken up into a "sequence"
5598 // of parallel tasks (per constituent space) that are dynamically
5599 // claimed by the parallel threads.
5600 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5601 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5602
5603 // It turns out that even when we're using 1 thread, doing the work in a
5604 // separate thread causes wide variance in run times. We can't help this
5605 // in the multi-threaded case, but we special-case n=1 here to get
5606 // repeatable measurements of the 1-thread overhead of the parallel code.
5607 if (n_workers > 1) {
5608 // Make refs discovery MT-safe, if it isn't already: it may not
5609 // necessarily be so, since it's possible that we are doing
5610 // ST marking.
5611 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5612 GenCollectedHeap::StrongRootsScope srs(gch);
5613 workers->run_task(&tsk);
5614 } else {
5615 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5616 GenCollectedHeap::StrongRootsScope srs(gch);
5617 tsk.work(0);
5618 }
5619 gch->set_par_threads(0); // 0 ==> non-parallel.
5620 // restore, single-threaded for now, any preserved marks
5621 // as a result of work_q overflow
5622 restore_preserved_marks_if_any();
5623 }
5624
5625 // Non-parallel version of remark
5626 void CMSCollector::do_remark_non_parallel() {
5627 ResourceMark rm;
5628 HandleMark hm;
5629 GenCollectedHeap* gch = GenCollectedHeap::heap();
5630 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5631
5632 MarkRefsIntoAndScanClosure
5633 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5634 &_markStack, &_revisitStack, this,
5635 false /* should_yield */, false /* not precleaning */);
5636 MarkFromDirtyCardsClosure
5637 markFromDirtyCardsClosure(this, _span,
5638 NULL, // space is set further below
5639 &_markBitMap, &_markStack, &_revisitStack,
5640 &mrias_cl);
5641 {
5642 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
5643 // Iterate over the dirty cards, setting the corresponding bits in the
5644 // mod union table.
5645 {
5646 ModUnionClosure modUnionClosure(&_modUnionTable);
5647 _ct->ct_bs()->dirty_card_iterate(
5648 _cmsGen->used_region(),
5649 &modUnionClosure);
5650 _ct->ct_bs()->dirty_card_iterate(
5651 _permGen->used_region(),
5652 &modUnionClosure);
5653 }
5654 // Having transferred these marks into the modUnionTable, we just need
5655 // to rescan the marked objects on the dirty cards in the modUnionTable.
5656 // The initial marking may have been done during an asynchronous
5657 // collection so there may be dirty bits in the mod-union table.
5658 const int alignment =
5659 CardTableModRefBS::card_size * BitsPerWord;
5660 {
5661 // ... First handle dirty cards in CMS gen
5662 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5663 MemRegion ur = _cmsGen->used_region();
5664 HeapWord* lb = ur.start();
5665 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5666 MemRegion cms_span(lb, ub);
5667 _modUnionTable.dirty_range_iterate_clear(cms_span,
5668 &markFromDirtyCardsClosure);
5669 verify_work_stacks_empty();
5670 if (PrintCMSStatistics != 0) {
5671 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5672 markFromDirtyCardsClosure.num_dirty_cards());
5673 }
5674 }
5675 {
5676 // .. and then repeat for dirty cards in perm gen
5677 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5678 MemRegion ur = _permGen->used_region();
5679 HeapWord* lb = ur.start();
5680 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5681 MemRegion perm_span(lb, ub);
5682 _modUnionTable.dirty_range_iterate_clear(perm_span,
5683 &markFromDirtyCardsClosure);
5684 verify_work_stacks_empty();
5685 if (PrintCMSStatistics != 0) {
5686 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5687 markFromDirtyCardsClosure.num_dirty_cards());
5688 }
5689 }
5690 }
5691 if (VerifyDuringGC &&
5692 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5693 HandleMark hm; // Discard invalid handles created during verification
5694 Universe::verify();
5695 }
5696 {
5697 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
5698
5699 verify_work_stacks_empty();
5700
5701 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5702 GenCollectedHeap::StrongRootsScope srs(gch);
5703 gch->gen_process_strong_roots(_cmsGen->level(),
5704 true, // younger gens as roots
5705 false, // use the local StrongRootsScope
5706 true, // collecting perm gen
5707 SharedHeap::ScanningOption(roots_scanning_options()),
5708 &mrias_cl,
5709 true, // walk code active on stacks
5710 NULL);
5711 assert(should_unload_classes()
5712 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5713 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5714 }
5715 verify_work_stacks_empty();
5716 // Restore evacuated mark words, if any, used for overflow list links
5717 if (!CMSOverflowEarlyRestoration) {
5718 restore_preserved_marks_if_any();
5719 }
5720 verify_overflow_empty();
5721 }
5722
5723 ////////////////////////////////////////////////////////
5724 // Parallel Reference Processing Task Proxy Class
5725 ////////////////////////////////////////////////////////
5726 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5727 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5728 CMSCollector* _collector;
5729 CMSBitMap* _mark_bit_map;
5730 const MemRegion _span;
5731 ProcessTask& _task;
5732
5733 public:
5734 CMSRefProcTaskProxy(ProcessTask& task,
5735 CMSCollector* collector,
5736 const MemRegion& span,
5737 CMSBitMap* mark_bit_map,
5738 AbstractWorkGang* workers,
5739 OopTaskQueueSet* task_queues):
5740 // XXX Should superclass AGTWOQ also know about AWG since it knows
5741 // about the task_queues used by the AWG? Then it could initialize
5742 // the terminator() object. See 6984287. The set_for_termination()
5743 // below is a temporary band-aid for the regression in 6984287.
5744 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5745 task_queues),
5746 _task(task),
5747 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5748 {
5749 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5750 "Inconsistency in _span");
5751 set_for_termination(workers->active_workers());
5752 }
5753
5754 OopTaskQueueSet* task_queues() { return queues(); }
5755
5756 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5757
5758 void do_work_steal(int i,
5759 CMSParDrainMarkingStackClosure* drain,
5760 CMSParKeepAliveClosure* keep_alive,
5761 int* seed);
5762
5763 virtual void work(uint worker_id);
5764 };
5765
5766 void CMSRefProcTaskProxy::work(uint worker_id) {
5767 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5768 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5769 _mark_bit_map,
5770 &_collector->_revisitStack,
5771 work_queue(worker_id));
5772 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5773 _mark_bit_map,
5774 &_collector->_revisitStack,
5775 work_queue(worker_id));
5776 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5777 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5778 if (_task.marks_oops_alive()) {
5779 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5780 _collector->hash_seed(worker_id));
5781 }
5782 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5783 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5784 }
5785
5786 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5787 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5788 EnqueueTask& _task;
5789
5790 public:
5791 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5792 : AbstractGangTask("Enqueue reference objects in parallel"),
5793 _task(task)
5794 { }
5795
5796 virtual void work(uint worker_id)
5797 {
5798 _task.work(worker_id);
5799 }
5800 };
5801
5802 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5803 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack,
5804 OopTaskQueue* work_queue):
5805 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
5806 _span(span),
5807 _bit_map(bit_map),
5808 _work_queue(work_queue),
5809 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue),
5810 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5811 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5812 { }
5813
5814 // . see if we can share work_queues with ParNew? XXX
5815 void CMSRefProcTaskProxy::do_work_steal(int i,
5816 CMSParDrainMarkingStackClosure* drain,
5817 CMSParKeepAliveClosure* keep_alive,
5818 int* seed) {
5819 OopTaskQueue* work_q = work_queue(i);
5820 NOT_PRODUCT(int num_steals = 0;)
5821 oop obj_to_scan;
5822
5823 while (true) {
5824 // Completely finish any left over work from (an) earlier round(s)
5825 drain->trim_queue(0);
5826 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5827 (size_t)ParGCDesiredObjsFromOverflowList);
5828 // Now check if there's any work in the overflow list
5829 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5830 // only affects the number of attempts made to get work from the
5831 // overflow list and does not affect the number of workers. Just
5832 // pass ParallelGCThreads so this behavior is unchanged.
5833 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5834 work_q,
5835 ParallelGCThreads)) {
5836 // Found something in global overflow list;
5837 // not yet ready to go stealing work from others.
5838 // We'd like to assert(work_q->size() != 0, ...)
5839 // because we just took work from the overflow list,
5840 // but of course we can't, since all of that might have
5841 // been already stolen from us.
5842 continue;
5843 }
5844 // Verify that we have no work before we resort to stealing
5845 assert(work_q->size() == 0, "Have work, shouldn't steal");
5846 // Try to steal from other queues that have work
5847 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5848 NOT_PRODUCT(num_steals++;)
5849 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5850 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5851 // Do scanning work
5852 obj_to_scan->oop_iterate(keep_alive);
5853 // Loop around, finish this work, and try to steal some more
5854 } else if (terminator()->offer_termination()) {
5855 break; // nirvana from the infinite cycle
5856 }
5857 }
5858 NOT_PRODUCT(
5859 if (PrintCMSStatistics != 0) {
5860 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5861 }
5862 )
5863 }
5864
5865 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5866 {
5867 GenCollectedHeap* gch = GenCollectedHeap::heap();
5868 FlexibleWorkGang* workers = gch->workers();
5869 assert(workers != NULL, "Need parallel worker threads.");
5870 CMSRefProcTaskProxy rp_task(task, &_collector,
5871 _collector.ref_processor()->span(),
5872 _collector.markBitMap(),
5873 workers, _collector.task_queues());
5874 workers->run_task(&rp_task);
5875 }
5876
5877 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5878 {
5879
5880 GenCollectedHeap* gch = GenCollectedHeap::heap();
5881 FlexibleWorkGang* workers = gch->workers();
5882 assert(workers != NULL, "Need parallel worker threads.");
5883 CMSRefEnqueueTaskProxy enq_task(task);
5884 workers->run_task(&enq_task);
5885 }
5886
5887 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5888
5889 ResourceMark rm;
5890 HandleMark hm;
5891
5892 ReferenceProcessor* rp = ref_processor();
5893 assert(rp->span().equals(_span), "Spans should be equal");
5894 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5895 // Process weak references.
5896 rp->setup_policy(clear_all_soft_refs);
5897 verify_work_stacks_empty();
5898
5899 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5900 &_markStack, &_revisitStack,
5901 false /* !preclean */);
5902 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5903 _span, &_markBitMap, &_markStack,
5904 &cmsKeepAliveClosure, false /* !preclean */);
5905 {
5906 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
5907
5908 ReferenceProcessorStats stats;
5909 if (rp->processing_is_mt()) {
5910 // Set the degree of MT here. If the discovery is done MT, there
5911 // may have been a different number of threads doing the discovery
5912 // and a different number of discovered lists may have Ref objects.
5913 // That is OK as long as the Reference lists are balanced (see
5914 // balance_all_queues() and balance_queues()).
5915 GenCollectedHeap* gch = GenCollectedHeap::heap();
5916 int active_workers = ParallelGCThreads;
5917 FlexibleWorkGang* workers = gch->workers();
5918 if (workers != NULL) {
5919 active_workers = workers->active_workers();
5920 // The expectation is that active_workers will have already
5921 // been set to a reasonable value. If it has not been set,
5922 // investigate.
5923 assert(active_workers > 0, "Should have been set during scavenge");
5924 }
5925 rp->set_active_mt_degree(active_workers);
5926 CMSRefProcTaskExecutor task_executor(*this);
5927 stats = rp->process_discovered_references(&_is_alive_closure,
5928 &cmsKeepAliveClosure,
5929 &cmsDrainMarkingStackClosure,
5930 &task_executor,
5931 _gc_timer_cm);
5932 } else {
5933 stats = rp->process_discovered_references(&_is_alive_closure,
5934 &cmsKeepAliveClosure,
5935 &cmsDrainMarkingStackClosure,
5936 NULL,
5937 _gc_timer_cm);
5938 }
5939 _gc_tracer_cm->report_gc_reference_stats(stats);
5940
5941 verify_work_stacks_empty();
5942 }
5943
5944 if (should_unload_classes()) {
5945 {
5946 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
5947
5948 // Follow SystemDictionary roots and unload classes
5949 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5950
5951 // Follow CodeCache roots and unload any methods marked for unloading
5952 CodeCache::do_unloading(&_is_alive_closure,
5953 &cmsKeepAliveClosure,
5954 purged_class);
5955
5956 cmsDrainMarkingStackClosure.do_void();
5957 verify_work_stacks_empty();
5958
5959 // Update subklass/sibling/implementor links in KlassKlass descendants
5960 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5961 oop k;
5962 while ((k = _revisitStack.pop()) != NULL) {
5963 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
5964 &_is_alive_closure,
5965 &cmsKeepAliveClosure);
5966 }
5967 assert(!ClassUnloading ||
5968 (_markStack.isEmpty() && overflow_list_is_empty()),
5969 "Should not have found new reachable objects");
5970 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
5971 cmsDrainMarkingStackClosure.do_void();
5972 verify_work_stacks_empty();
5973 }
5974
5975 {
5976 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
5977 // Clean up unreferenced symbols in symbol table.
5978 SymbolTable::unlink();
5979 }
5980 }
5981
5982 if (should_unload_classes() || !JavaObjectsInPerm) {
5983 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
5984 // Now clean up stale oops in StringTable
5985 StringTable::unlink(&_is_alive_closure);
5986 }
5987
5988 verify_work_stacks_empty();
5989 // Restore any preserved marks as a result of mark stack or
5990 // work queue overflow
5991 restore_preserved_marks_if_any(); // done single-threaded for now
5992
5993 rp->set_enqueuing_is_done(true);
5994 if (rp->processing_is_mt()) {
5995 rp->balance_all_queues();
5996 CMSRefProcTaskExecutor task_executor(*this);
5997 rp->enqueue_discovered_references(&task_executor);
5998 } else {
5999 rp->enqueue_discovered_references(NULL);
6000 }
6001 rp->verify_no_references_recorded();
6002 assert(!rp->discovery_enabled(), "should have been disabled");
6003 }
6004
6005 #ifndef PRODUCT
6006 void CMSCollector::check_correct_thread_executing() {
6007 Thread* t = Thread::current();
6008 // Only the VM thread or the CMS thread should be here.
6009 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6010 "Unexpected thread type");
6011 // If this is the vm thread, the foreground process
6012 // should not be waiting. Note that _foregroundGCIsActive is
6013 // true while the foreground collector is waiting.
6014 if (_foregroundGCShouldWait) {
6015 // We cannot be the VM thread
6016 assert(t->is_ConcurrentGC_thread(),
6017 "Should be CMS thread");
6018 } else {
6019 // We can be the CMS thread only if we are in a stop-world
6020 // phase of CMS collection.
6021 if (t->is_ConcurrentGC_thread()) {
6022 assert(_collectorState == InitialMarking ||
6023 _collectorState == FinalMarking,
6024 "Should be a stop-world phase");
6025 // The CMS thread should be holding the CMS_token.
6026 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6027 "Potential interference with concurrently "
6028 "executing VM thread");
6029 }
6030 }
6031 }
6032 #endif
6033
6034 void CMSCollector::sweep(bool asynch) {
6035 assert(_collectorState == Sweeping, "just checking");
6036 check_correct_thread_executing();
6037 verify_work_stacks_empty();
6038 verify_overflow_empty();
6039 increment_sweep_count();
6040 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6041
6042 _inter_sweep_timer.stop();
6043 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6044 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6045
6046 // PermGen verification support: If perm gen sweeping is disabled in
6047 // this cycle, we preserve the perm gen object "deadness" information
6048 // in the perm_gen_verify_bit_map. In order to do that we traverse
6049 // all blocks in perm gen and mark all dead objects.
6050 if (verifying() && !should_unload_classes()) {
6051 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
6052 "Should have already been allocated");
6053 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
6054 markBitMap(), perm_gen_verify_bit_map());
6055 if (asynch) {
6056 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6057 bitMapLock());
6058 _permGen->cmsSpace()->blk_iterate(&mdo);
6059 } else {
6060 // In the case of synchronous sweep, we already have
6061 // the requisite locks/tokens.
6062 _permGen->cmsSpace()->blk_iterate(&mdo);
6063 }
6064 }
6065
6066 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6067 _intra_sweep_timer.reset();
6068 _intra_sweep_timer.start();
6069 if (asynch) {
6070 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6071 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6072 // First sweep the old gen then the perm gen
6073 {
6074 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6075 bitMapLock());
6076 sweepWork(_cmsGen, asynch);
6077 }
6078
6079 // Now repeat for perm gen
6080 if (should_unload_classes()) {
6081 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6082 bitMapLock());
6083 sweepWork(_permGen, asynch);
6084 }
6085
6086 // Update Universe::_heap_*_at_gc figures.
6087 // We need all the free list locks to make the abstract state
6088 // transition from Sweeping to Resetting. See detailed note
6089 // further below.
6090 {
6091 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6092 _permGen->freelistLock());
6093 // Update heap occupancy information which is used as
6094 // input to soft ref clearing policy at the next gc.
6095 Universe::update_heap_info_at_gc();
6096 _collectorState = Resizing;
6097 }
6098 } else {
6099 // already have needed locks
6100 sweepWork(_cmsGen, asynch);
6101
6102 if (should_unload_classes()) {
6103 sweepWork(_permGen, asynch);
6104 }
6105 // Update heap occupancy information which is used as
6106 // input to soft ref clearing policy at the next gc.
6107 Universe::update_heap_info_at_gc();
6108 _collectorState = Resizing;
6109 }
6110 verify_work_stacks_empty();
6111 verify_overflow_empty();
6112
6113 _intra_sweep_timer.stop();
6114 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6115
6116 _inter_sweep_timer.reset();
6117 _inter_sweep_timer.start();
6118
6119 // We need to use a monotonically non-deccreasing time in ms
6120 // or we will see time-warp warnings and os::javaTimeMillis()
6121 // does not guarantee monotonicity.
6122 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6123 update_time_of_last_gc(now);
6124
6125 // NOTE on abstract state transitions:
6126 // Mutators allocate-live and/or mark the mod-union table dirty
6127 // based on the state of the collection. The former is done in
6128 // the interval [Marking, Sweeping] and the latter in the interval
6129 // [Marking, Sweeping). Thus the transitions into the Marking state
6130 // and out of the Sweeping state must be synchronously visible
6131 // globally to the mutators.
6132 // The transition into the Marking state happens with the world
6133 // stopped so the mutators will globally see it. Sweeping is
6134 // done asynchronously by the background collector so the transition
6135 // from the Sweeping state to the Resizing state must be done
6136 // under the freelistLock (as is the check for whether to
6137 // allocate-live and whether to dirty the mod-union table).
6138 assert(_collectorState == Resizing, "Change of collector state to"
6139 " Resizing must be done under the freelistLocks (plural)");
6140
6141 // Now that sweeping has been completed, we clear
6142 // the incremental_collection_failed flag,
6143 // thus inviting a younger gen collection to promote into
6144 // this generation. If such a promotion may still fail,
6145 // the flag will be set again when a young collection is
6146 // attempted.
6147 GenCollectedHeap* gch = GenCollectedHeap::heap();
6148 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6149 gch->update_full_collections_completed(_collection_count_start);
6150 }
6151
6152 // FIX ME!!! Looks like this belongs in CFLSpace, with
6153 // CMSGen merely delegating to it.
6154 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6155 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6156 HeapWord* minAddr = _cmsSpace->bottom();
6157 HeapWord* largestAddr =
6158 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6159 if (largestAddr == NULL) {
6160 // The dictionary appears to be empty. In this case
6161 // try to coalesce at the end of the heap.
6162 largestAddr = _cmsSpace->end();
6163 }
6164 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6165 size_t nearLargestOffset =
6166 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6167 if (PrintFLSStatistics != 0) {
6168 gclog_or_tty->print_cr(
6169 "CMS: Large Block: " PTR_FORMAT ";"
6170 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6171 largestAddr,
6172 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6173 }
6174 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6175 }
6176
6177 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6178 return addr >= _cmsSpace->nearLargestChunk();
6179 }
6180
6181 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6182 return _cmsSpace->find_chunk_at_end();
6183 }
6184
6185 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6186 bool full) {
6187 // The next lower level has been collected. Gather any statistics
6188 // that are of interest at this point.
6189 if (!full && (current_level + 1) == level()) {
6190 // Gather statistics on the young generation collection.
6191 collector()->stats().record_gc0_end(used());
6192 }
6193 }
6194
6195 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6196 GenCollectedHeap* gch = GenCollectedHeap::heap();
6197 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6198 "Wrong type of heap");
6199 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6200 gch->gen_policy()->size_policy();
6201 assert(sp->is_gc_cms_adaptive_size_policy(),
6202 "Wrong type of size policy");
6203 return sp;
6204 }
6205
6206 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6207 if (PrintGCDetails && Verbose) {
6208 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6209 }
6210 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6211 _debug_collection_type =
6212 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6213 if (PrintGCDetails && Verbose) {
6214 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6215 }
6216 }
6217
6218 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6219 bool asynch) {
6220 // We iterate over the space(s) underlying this generation,
6221 // checking the mark bit map to see if the bits corresponding
6222 // to specific blocks are marked or not. Blocks that are
6223 // marked are live and are not swept up. All remaining blocks
6224 // are swept up, with coalescing on-the-fly as we sweep up
6225 // contiguous free and/or garbage blocks:
6226 // We need to ensure that the sweeper synchronizes with allocators
6227 // and stop-the-world collectors. In particular, the following
6228 // locks are used:
6229 // . CMS token: if this is held, a stop the world collection cannot occur
6230 // . freelistLock: if this is held no allocation can occur from this
6231 // generation by another thread
6232 // . bitMapLock: if this is held, no other thread can access or update
6233 //
6234
6235 // Note that we need to hold the freelistLock if we use
6236 // block iterate below; else the iterator might go awry if
6237 // a mutator (or promotion) causes block contents to change
6238 // (for instance if the allocator divvies up a block).
6239 // If we hold the free list lock, for all practical purposes
6240 // young generation GC's can't occur (they'll usually need to
6241 // promote), so we might as well prevent all young generation
6242 // GC's while we do a sweeping step. For the same reason, we might
6243 // as well take the bit map lock for the entire duration
6244
6245 // check that we hold the requisite locks
6246 assert(have_cms_token(), "Should hold cms token");
6247 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6248 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6249 "Should possess CMS token to sweep");
6250 assert_lock_strong(gen->freelistLock());
6251 assert_lock_strong(bitMapLock());
6252
6253 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6254 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6255 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6256 _inter_sweep_estimate.padded_average(),
6257 _intra_sweep_estimate.padded_average());
6258 gen->setNearLargestChunk();
6259
6260 {
6261 SweepClosure sweepClosure(this, gen, &_markBitMap,
6262 CMSYield && asynch);
6263 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6264 // We need to free-up/coalesce garbage/blocks from a
6265 // co-terminal free run. This is done in the SweepClosure
6266 // destructor; so, do not remove this scope, else the
6267 // end-of-sweep-census below will be off by a little bit.
6268 }
6269 gen->cmsSpace()->sweep_completed();
6270 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6271 if (should_unload_classes()) { // unloaded classes this cycle,
6272 _concurrent_cycles_since_last_unload = 0; // ... reset count
6273 } else { // did not unload classes,
6274 _concurrent_cycles_since_last_unload++; // ... increment count
6275 }
6276 }
6277
6278 // Reset CMS data structures (for now just the marking bit map)
6279 // preparatory for the next cycle.
6280 void CMSCollector::reset(bool asynch) {
6281 GenCollectedHeap* gch = GenCollectedHeap::heap();
6282 CMSAdaptiveSizePolicy* sp = size_policy();
6283 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6284 if (asynch) {
6285 CMSTokenSyncWithLocks ts(true, bitMapLock());
6286
6287 // If the state is not "Resetting", the foreground thread
6288 // has done a collection and the resetting.
6289 if (_collectorState != Resetting) {
6290 assert(_collectorState == Idling, "The state should only change"
6291 " because the foreground collector has finished the collection");
6292 return;
6293 }
6294
6295 // Clear the mark bitmap (no grey objects to start with)
6296 // for the next cycle.
6297 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6298 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6299
6300 HeapWord* curAddr = _markBitMap.startWord();
6301 while (curAddr < _markBitMap.endWord()) {
6302 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6303 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6304 _markBitMap.clear_large_range(chunk);
6305 if (ConcurrentMarkSweepThread::should_yield() &&
6306 !foregroundGCIsActive() &&
6307 CMSYield) {
6308 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6309 "CMS thread should hold CMS token");
6310 assert_lock_strong(bitMapLock());
6311 bitMapLock()->unlock();
6312 ConcurrentMarkSweepThread::desynchronize(true);
6313 ConcurrentMarkSweepThread::acknowledge_yield_request();
6314 stopTimer();
6315 if (PrintCMSStatistics != 0) {
6316 incrementYields();
6317 }
6318 icms_wait();
6319
6320 // See the comment in coordinator_yield()
6321 for (unsigned i = 0; i < CMSYieldSleepCount &&
6322 ConcurrentMarkSweepThread::should_yield() &&
6323 !CMSCollector::foregroundGCIsActive(); ++i) {
6324 os::sleep(Thread::current(), 1, false);
6325 ConcurrentMarkSweepThread::acknowledge_yield_request();
6326 }
6327
6328 ConcurrentMarkSweepThread::synchronize(true);
6329 bitMapLock()->lock_without_safepoint_check();
6330 startTimer();
6331 }
6332 curAddr = chunk.end();
6333 }
6334 // A successful mostly concurrent collection has been done.
6335 // Because only the full (i.e., concurrent mode failure) collections
6336 // are being measured for gc overhead limits, clean the "near" flag
6337 // and count.
6338 sp->reset_gc_overhead_limit_count();
6339 _collectorState = Idling;
6340 } else {
6341 // already have the lock
6342 assert(_collectorState == Resetting, "just checking");
6343 assert_lock_strong(bitMapLock());
6344 _markBitMap.clear_all();
6345 _collectorState = Idling;
6346 }
6347
6348 // Stop incremental mode after a cycle completes, so that any future cycles
6349 // are triggered by allocation.
6350 stop_icms();
6351
6352 NOT_PRODUCT(
6353 if (RotateCMSCollectionTypes) {
6354 _cmsGen->rotate_debug_collection_type();
6355 }
6356 )
6357
6358 register_gc_end();
6359 }
6360
6361 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6362 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6363 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6364 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
6365 TraceCollectorStats tcs(counters());
6366
6367 switch (op) {
6368 case CMS_op_checkpointRootsInitial: {
6369 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6370 checkpointRootsInitial(true); // asynch
6371 if (PrintGC) {
6372 _cmsGen->printOccupancy("initial-mark");
6373 }
6374 break;
6375 }
6376 case CMS_op_checkpointRootsFinal: {
6377 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6378 checkpointRootsFinal(true, // asynch
6379 false, // !clear_all_soft_refs
6380 false); // !init_mark_was_synchronous
6381 if (PrintGC) {
6382 _cmsGen->printOccupancy("remark");
6383 }
6384 break;
6385 }
6386 default:
6387 fatal("No such CMS_op");
6388 }
6389 }
6390
6391 #ifndef PRODUCT
6392 size_t const CMSCollector::skip_header_HeapWords() {
6393 return FreeChunk::header_size();
6394 }
6395
6396 // Try and collect here conditions that should hold when
6397 // CMS thread is exiting. The idea is that the foreground GC
6398 // thread should not be blocked if it wants to terminate
6399 // the CMS thread and yet continue to run the VM for a while
6400 // after that.
6401 void CMSCollector::verify_ok_to_terminate() const {
6402 assert(Thread::current()->is_ConcurrentGC_thread(),
6403 "should be called by CMS thread");
6404 assert(!_foregroundGCShouldWait, "should be false");
6405 // We could check here that all the various low-level locks
6406 // are not held by the CMS thread, but that is overkill; see
6407 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6408 // is checked.
6409 }
6410 #endif
6411
6412 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6413 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6414 "missing Printezis mark?");
6415 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6416 size_t size = pointer_delta(nextOneAddr + 1, addr);
6417 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6418 "alignment problem");
6419 assert(size >= 3, "Necessary for Printezis marks to work");
6420 return size;
6421 }
6422
6423 // A variant of the above (block_size_using_printezis_bits()) except
6424 // that we return 0 if the P-bits are not yet set.
6425 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6426 if (_markBitMap.isMarked(addr + 1)) {
6427 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6428 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6429 size_t size = pointer_delta(nextOneAddr + 1, addr);
6430 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6431 "alignment problem");
6432 assert(size >= 3, "Necessary for Printezis marks to work");
6433 return size;
6434 }
6435 return 0;
6436 }
6437
6438 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6439 size_t sz = 0;
6440 oop p = (oop)addr;
6441 if (p->klass_or_null() != NULL && p->is_parsable()) {
6442 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6443 } else {
6444 sz = block_size_using_printezis_bits(addr);
6445 }
6446 assert(sz > 0, "size must be nonzero");
6447 HeapWord* next_block = addr + sz;
6448 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6449 CardTableModRefBS::card_size);
6450 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6451 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6452 "must be different cards");
6453 return next_card;
6454 }
6455
6456
6457 // CMS Bit Map Wrapper /////////////////////////////////////////
6458
6459 // Construct a CMS bit map infrastructure, but don't create the
6460 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6461 // further below.
6462 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6463 _bm(),
6464 _shifter(shifter),
6465 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6466 {
6467 _bmStartWord = 0;
6468 _bmWordSize = 0;
6469 }
6470
6471 bool CMSBitMap::allocate(MemRegion mr) {
6472 _bmStartWord = mr.start();
6473 _bmWordSize = mr.word_size();
6474 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6475 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6476 if (!brs.is_reserved()) {
6477 warning("CMS bit map allocation failure");
6478 return false;
6479 }
6480 // For now we'll just commit all of the bit map up fromt.
6481 // Later on we'll try to be more parsimonious with swap.
6482 if (!_virtual_space.initialize(brs, brs.size())) {
6483 warning("CMS bit map backing store failure");
6484 return false;
6485 }
6486 assert(_virtual_space.committed_size() == brs.size(),
6487 "didn't reserve backing store for all of CMS bit map?");
6488 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6489 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6490 _bmWordSize, "inconsistency in bit map sizing");
6491 _bm.set_size(_bmWordSize >> _shifter);
6492
6493 // bm.clear(); // can we rely on getting zero'd memory? verify below
6494 assert(isAllClear(),
6495 "Expected zero'd memory from ReservedSpace constructor");
6496 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6497 "consistency check");
6498 return true;
6499 }
6500
6501 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6502 HeapWord *next_addr, *end_addr, *last_addr;
6503 assert_locked();
6504 assert(covers(mr), "out-of-range error");
6505 // XXX assert that start and end are appropriately aligned
6506 for (next_addr = mr.start(), end_addr = mr.end();
6507 next_addr < end_addr; next_addr = last_addr) {
6508 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6509 last_addr = dirty_region.end();
6510 if (!dirty_region.is_empty()) {
6511 cl->do_MemRegion(dirty_region);
6512 } else {
6513 assert(last_addr == end_addr, "program logic");
6514 return;
6515 }
6516 }
6517 }
6518
6519 #ifndef PRODUCT
6520 void CMSBitMap::assert_locked() const {
6521 CMSLockVerifier::assert_locked(lock());
6522 }
6523
6524 bool CMSBitMap::covers(MemRegion mr) const {
6525 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6526 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6527 "size inconsistency");
6528 return (mr.start() >= _bmStartWord) &&
6529 (mr.end() <= endWord());
6530 }
6531
6532 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6533 return (start >= _bmStartWord && (start + size) <= endWord());
6534 }
6535
6536 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6537 // verify that there are no 1 bits in the interval [left, right)
6538 FalseBitMapClosure falseBitMapClosure;
6539 iterate(&falseBitMapClosure, left, right);
6540 }
6541
6542 void CMSBitMap::region_invariant(MemRegion mr)
6543 {
6544 assert_locked();
6545 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6546 assert(!mr.is_empty(), "unexpected empty region");
6547 assert(covers(mr), "mr should be covered by bit map");
6548 // convert address range into offset range
6549 size_t start_ofs = heapWordToOffset(mr.start());
6550 // Make sure that end() is appropriately aligned
6551 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6552 (1 << (_shifter+LogHeapWordSize))),
6553 "Misaligned mr.end()");
6554 size_t end_ofs = heapWordToOffset(mr.end());
6555 assert(end_ofs > start_ofs, "Should mark at least one bit");
6556 }
6557
6558 #endif
6559
6560 bool CMSMarkStack::allocate(size_t size) {
6561 // allocate a stack of the requisite depth
6562 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6563 size * sizeof(oop)));
6564 if (!rs.is_reserved()) {
6565 warning("CMSMarkStack allocation failure");
6566 return false;
6567 }
6568 if (!_virtual_space.initialize(rs, rs.size())) {
6569 warning("CMSMarkStack backing store failure");
6570 return false;
6571 }
6572 assert(_virtual_space.committed_size() == rs.size(),
6573 "didn't reserve backing store for all of CMS stack?");
6574 _base = (oop*)(_virtual_space.low());
6575 _index = 0;
6576 _capacity = size;
6577 NOT_PRODUCT(_max_depth = 0);
6578 return true;
6579 }
6580
6581 // XXX FIX ME !!! In the MT case we come in here holding a
6582 // leaf lock. For printing we need to take a further lock
6583 // which has lower rank. We need to recallibrate the two
6584 // lock-ranks involved in order to be able to rpint the
6585 // messages below. (Or defer the printing to the caller.
6586 // For now we take the expedient path of just disabling the
6587 // messages for the problematic case.)
6588 void CMSMarkStack::expand() {
6589 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6590 if (_capacity == MarkStackSizeMax) {
6591 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6592 // We print a warning message only once per CMS cycle.
6593 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6594 }
6595 return;
6596 }
6597 // Double capacity if possible
6598 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6599 // Do not give up existing stack until we have managed to
6600 // get the double capacity that we desired.
6601 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6602 new_capacity * sizeof(oop)));
6603 if (rs.is_reserved()) {
6604 // Release the backing store associated with old stack
6605 _virtual_space.release();
6606 // Reinitialize virtual space for new stack
6607 if (!_virtual_space.initialize(rs, rs.size())) {
6608 fatal("Not enough swap for expanded marking stack");
6609 }
6610 _base = (oop*)(_virtual_space.low());
6611 _index = 0;
6612 _capacity = new_capacity;
6613 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6614 // Failed to double capacity, continue;
6615 // we print a detail message only once per CMS cycle.
6616 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6617 SIZE_FORMAT"K",
6618 _capacity / K, new_capacity / K);
6619 }
6620 }
6621
6622
6623 // Closures
6624 // XXX: there seems to be a lot of code duplication here;
6625 // should refactor and consolidate common code.
6626
6627 // This closure is used to mark refs into the CMS generation in
6628 // the CMS bit map. Called at the first checkpoint. This closure
6629 // assumes that we do not need to re-mark dirty cards; if the CMS
6630 // generation on which this is used is not an oldest (modulo perm gen)
6631 // generation then this will lose younger_gen cards!
6632
6633 MarkRefsIntoClosure::MarkRefsIntoClosure(
6634 MemRegion span, CMSBitMap* bitMap):
6635 _span(span),
6636 _bitMap(bitMap)
6637 {
6638 assert(_ref_processor == NULL, "deliberately left NULL");
6639 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6640 }
6641
6642 void MarkRefsIntoClosure::do_oop(oop obj) {
6643 // if p points into _span, then mark corresponding bit in _markBitMap
6644 assert(obj->is_oop(), "expected an oop");
6645 HeapWord* addr = (HeapWord*)obj;
6646 if (_span.contains(addr)) {
6647 // this should be made more efficient
6648 _bitMap->mark(addr);
6649 }
6650 }
6651
6652 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6653 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6654
6655 // A variant of the above, used for CMS marking verification.
6656 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6657 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6658 _span(span),
6659 _verification_bm(verification_bm),
6660 _cms_bm(cms_bm)
6661 {
6662 assert(_ref_processor == NULL, "deliberately left NULL");
6663 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6664 }
6665
6666 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6667 // if p points into _span, then mark corresponding bit in _markBitMap
6668 assert(obj->is_oop(), "expected an oop");
6669 HeapWord* addr = (HeapWord*)obj;
6670 if (_span.contains(addr)) {
6671 _verification_bm->mark(addr);
6672 if (!_cms_bm->isMarked(addr)) {
6673 oop(addr)->print();
6674 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6675 fatal("... aborting");
6676 }
6677 }
6678 }
6679
6680 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6681 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6682
6683 //////////////////////////////////////////////////
6684 // MarkRefsIntoAndScanClosure
6685 //////////////////////////////////////////////////
6686
6687 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6688 ReferenceProcessor* rp,
6689 CMSBitMap* bit_map,
6690 CMSBitMap* mod_union_table,
6691 CMSMarkStack* mark_stack,
6692 CMSMarkStack* revisit_stack,
6693 CMSCollector* collector,
6694 bool should_yield,
6695 bool concurrent_precleaning):
6696 _collector(collector),
6697 _span(span),
6698 _bit_map(bit_map),
6699 _mark_stack(mark_stack),
6700 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6701 mark_stack, revisit_stack, concurrent_precleaning),
6702 _yield(should_yield),
6703 _concurrent_precleaning(concurrent_precleaning),
6704 _freelistLock(NULL)
6705 {
6706 _ref_processor = rp;
6707 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6708 }
6709
6710 // This closure is used to mark refs into the CMS generation at the
6711 // second (final) checkpoint, and to scan and transitively follow
6712 // the unmarked oops. It is also used during the concurrent precleaning
6713 // phase while scanning objects on dirty cards in the CMS generation.
6714 // The marks are made in the marking bit map and the marking stack is
6715 // used for keeping the (newly) grey objects during the scan.
6716 // The parallel version (Par_...) appears further below.
6717 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6718 if (obj != NULL) {
6719 assert(obj->is_oop(), "expected an oop");
6720 HeapWord* addr = (HeapWord*)obj;
6721 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6722 assert(_collector->overflow_list_is_empty(),
6723 "overflow list should be empty");
6724 if (_span.contains(addr) &&
6725 !_bit_map->isMarked(addr)) {
6726 // mark bit map (object is now grey)
6727 _bit_map->mark(addr);
6728 // push on marking stack (stack should be empty), and drain the
6729 // stack by applying this closure to the oops in the oops popped
6730 // from the stack (i.e. blacken the grey objects)
6731 bool res = _mark_stack->push(obj);
6732 assert(res, "Should have space to push on empty stack");
6733 do {
6734 oop new_oop = _mark_stack->pop();
6735 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6736 assert(new_oop->is_parsable(), "Found unparsable oop");
6737 assert(_bit_map->isMarked((HeapWord*)new_oop),
6738 "only grey objects on this stack");
6739 // iterate over the oops in this oop, marking and pushing
6740 // the ones in CMS heap (i.e. in _span).
6741 new_oop->oop_iterate(&_pushAndMarkClosure);
6742 // check if it's time to yield
6743 do_yield_check();
6744 } while (!_mark_stack->isEmpty() ||
6745 (!_concurrent_precleaning && take_from_overflow_list()));
6746 // if marking stack is empty, and we are not doing this
6747 // during precleaning, then check the overflow list
6748 }
6749 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6750 assert(_collector->overflow_list_is_empty(),
6751 "overflow list was drained above");
6752 // We could restore evacuated mark words, if any, used for
6753 // overflow list links here because the overflow list is
6754 // provably empty here. That would reduce the maximum
6755 // size requirements for preserved_{oop,mark}_stack.
6756 // But we'll just postpone it until we are all done
6757 // so we can just stream through.
6758 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6759 _collector->restore_preserved_marks_if_any();
6760 assert(_collector->no_preserved_marks(), "No preserved marks");
6761 }
6762 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6763 "All preserved marks should have been restored above");
6764 }
6765 }
6766
6767 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6768 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6769
6770 void MarkRefsIntoAndScanClosure::do_yield_work() {
6771 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6772 "CMS thread should hold CMS token");
6773 assert_lock_strong(_freelistLock);
6774 assert_lock_strong(_bit_map->lock());
6775 // relinquish the free_list_lock and bitMaplock()
6776 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6777 _bit_map->lock()->unlock();
6778 _freelistLock->unlock();
6779 ConcurrentMarkSweepThread::desynchronize(true);
6780 ConcurrentMarkSweepThread::acknowledge_yield_request();
6781 _collector->stopTimer();
6782 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6783 if (PrintCMSStatistics != 0) {
6784 _collector->incrementYields();
6785 }
6786 _collector->icms_wait();
6787
6788 // See the comment in coordinator_yield()
6789 for (unsigned i = 0;
6790 i < CMSYieldSleepCount &&
6791 ConcurrentMarkSweepThread::should_yield() &&
6792 !CMSCollector::foregroundGCIsActive();
6793 ++i) {
6794 os::sleep(Thread::current(), 1, false);
6795 ConcurrentMarkSweepThread::acknowledge_yield_request();
6796 }
6797
6798 ConcurrentMarkSweepThread::synchronize(true);
6799 _freelistLock->lock_without_safepoint_check();
6800 _bit_map->lock()->lock_without_safepoint_check();
6801 _collector->startTimer();
6802 }
6803
6804 ///////////////////////////////////////////////////////////
6805 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6806 // MarkRefsIntoAndScanClosure
6807 ///////////////////////////////////////////////////////////
6808 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6809 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6810 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6811 _span(span),
6812 _bit_map(bit_map),
6813 _work_queue(work_queue),
6814 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6815 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6816 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6817 revisit_stack)
6818 {
6819 _ref_processor = rp;
6820 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6821 }
6822
6823 // This closure is used to mark refs into the CMS generation at the
6824 // second (final) checkpoint, and to scan and transitively follow
6825 // the unmarked oops. The marks are made in the marking bit map and
6826 // the work_queue is used for keeping the (newly) grey objects during
6827 // the scan phase whence they are also available for stealing by parallel
6828 // threads. Since the marking bit map is shared, updates are
6829 // synchronized (via CAS).
6830 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6831 if (obj != NULL) {
6832 // Ignore mark word because this could be an already marked oop
6833 // that may be chained at the end of the overflow list.
6834 assert(obj->is_oop(true), "expected an oop");
6835 HeapWord* addr = (HeapWord*)obj;
6836 if (_span.contains(addr) &&
6837 !_bit_map->isMarked(addr)) {
6838 // mark bit map (object will become grey):
6839 // It is possible for several threads to be
6840 // trying to "claim" this object concurrently;
6841 // the unique thread that succeeds in marking the
6842 // object first will do the subsequent push on
6843 // to the work queue (or overflow list).
6844 if (_bit_map->par_mark(addr)) {
6845 // push on work_queue (which may not be empty), and trim the
6846 // queue to an appropriate length by applying this closure to
6847 // the oops in the oops popped from the stack (i.e. blacken the
6848 // grey objects)
6849 bool res = _work_queue->push(obj);
6850 assert(res, "Low water mark should be less than capacity?");
6851 trim_queue(_low_water_mark);
6852 } // Else, another thread claimed the object
6853 }
6854 }
6855 }
6856
6857 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6858 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6859
6860 // This closure is used to rescan the marked objects on the dirty cards
6861 // in the mod union table and the card table proper.
6862 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6863 oop p, MemRegion mr) {
6864
6865 size_t size = 0;
6866 HeapWord* addr = (HeapWord*)p;
6867 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6868 assert(_span.contains(addr), "we are scanning the CMS generation");
6869 // check if it's time to yield
6870 if (do_yield_check()) {
6871 // We yielded for some foreground stop-world work,
6872 // and we have been asked to abort this ongoing preclean cycle.
6873 return 0;
6874 }
6875 if (_bitMap->isMarked(addr)) {
6876 // it's marked; is it potentially uninitialized?
6877 if (p->klass_or_null() != NULL) {
6878 // If is_conc_safe is false, the object may be undergoing
6879 // change by the VM outside a safepoint. Don't try to
6880 // scan it, but rather leave it for the remark phase.
6881 if (CMSPermGenPrecleaningEnabled &&
6882 (!p->is_conc_safe() || !p->is_parsable())) {
6883 // Signal precleaning to redirty the card since
6884 // the klass pointer is already installed.
6885 assert(size == 0, "Initial value");
6886 } else {
6887 assert(p->is_parsable(), "must be parsable.");
6888 // an initialized object; ignore mark word in verification below
6889 // since we are running concurrent with mutators
6890 assert(p->is_oop(true), "should be an oop");
6891 if (p->is_objArray()) {
6892 // objArrays are precisely marked; restrict scanning
6893 // to dirty cards only.
6894 size = CompactibleFreeListSpace::adjustObjectSize(
6895 p->oop_iterate(_scanningClosure, mr));
6896 } else {
6897 // A non-array may have been imprecisely marked; we need
6898 // to scan object in its entirety.
6899 size = CompactibleFreeListSpace::adjustObjectSize(
6900 p->oop_iterate(_scanningClosure));
6901 }
6902 #ifdef DEBUG
6903 size_t direct_size =
6904 CompactibleFreeListSpace::adjustObjectSize(p->size());
6905 assert(size == direct_size, "Inconsistency in size");
6906 assert(size >= 3, "Necessary for Printezis marks to work");
6907 if (!_bitMap->isMarked(addr+1)) {
6908 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6909 } else {
6910 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6911 assert(_bitMap->isMarked(addr+size-1),
6912 "inconsistent Printezis mark");
6913 }
6914 #endif // DEBUG
6915 }
6916 } else {
6917 // an unitialized object
6918 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6919 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6920 size = pointer_delta(nextOneAddr + 1, addr);
6921 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6922 "alignment problem");
6923 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6924 // will dirty the card when the klass pointer is installed in the
6925 // object (signalling the completion of initialization).
6926 }
6927 } else {
6928 // Either a not yet marked object or an uninitialized object
6929 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6930 // An uninitialized object, skip to the next card, since
6931 // we may not be able to read its P-bits yet.
6932 assert(size == 0, "Initial value");
6933 } else {
6934 // An object not (yet) reached by marking: we merely need to
6935 // compute its size so as to go look at the next block.
6936 assert(p->is_oop(true), "should be an oop");
6937 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6938 }
6939 }
6940 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6941 return size;
6942 }
6943
6944 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6945 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6946 "CMS thread should hold CMS token");
6947 assert_lock_strong(_freelistLock);
6948 assert_lock_strong(_bitMap->lock());
6949 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6950 // relinquish the free_list_lock and bitMaplock()
6951 _bitMap->lock()->unlock();
6952 _freelistLock->unlock();
6953 ConcurrentMarkSweepThread::desynchronize(true);
6954 ConcurrentMarkSweepThread::acknowledge_yield_request();
6955 _collector->stopTimer();
6956 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6957 if (PrintCMSStatistics != 0) {
6958 _collector->incrementYields();
6959 }
6960 _collector->icms_wait();
6961
6962 // See the comment in coordinator_yield()
6963 for (unsigned i = 0; i < CMSYieldSleepCount &&
6964 ConcurrentMarkSweepThread::should_yield() &&
6965 !CMSCollector::foregroundGCIsActive(); ++i) {
6966 os::sleep(Thread::current(), 1, false);
6967 ConcurrentMarkSweepThread::acknowledge_yield_request();
6968 }
6969
6970 ConcurrentMarkSweepThread::synchronize(true);
6971 _freelistLock->lock_without_safepoint_check();
6972 _bitMap->lock()->lock_without_safepoint_check();
6973 _collector->startTimer();
6974 }
6975
6976
6977 //////////////////////////////////////////////////////////////////
6978 // SurvivorSpacePrecleanClosure
6979 //////////////////////////////////////////////////////////////////
6980 // This (single-threaded) closure is used to preclean the oops in
6981 // the survivor spaces.
6982 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6983
6984 HeapWord* addr = (HeapWord*)p;
6985 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6986 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6987 assert(p->klass_or_null() != NULL, "object should be initializd");
6988 assert(p->is_parsable(), "must be parsable.");
6989 // an initialized object; ignore mark word in verification below
6990 // since we are running concurrent with mutators
6991 assert(p->is_oop(true), "should be an oop");
6992 // Note that we do not yield while we iterate over
6993 // the interior oops of p, pushing the relevant ones
6994 // on our marking stack.
6995 size_t size = p->oop_iterate(_scanning_closure);
6996 do_yield_check();
6997 // Observe that below, we do not abandon the preclean
6998 // phase as soon as we should; rather we empty the
6999 // marking stack before returning. This is to satisfy
7000 // some existing assertions. In general, it may be a
7001 // good idea to abort immediately and complete the marking
7002 // from the grey objects at a later time.
7003 while (!_mark_stack->isEmpty()) {
7004 oop new_oop = _mark_stack->pop();
7005 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7006 assert(new_oop->is_parsable(), "Found unparsable oop");
7007 assert(_bit_map->isMarked((HeapWord*)new_oop),
7008 "only grey objects on this stack");
7009 // iterate over the oops in this oop, marking and pushing
7010 // the ones in CMS heap (i.e. in _span).
7011 new_oop->oop_iterate(_scanning_closure);
7012 // check if it's time to yield
7013 do_yield_check();
7014 }
7015 unsigned int after_count =
7016 GenCollectedHeap::heap()->total_collections();
7017 bool abort = (_before_count != after_count) ||
7018 _collector->should_abort_preclean();
7019 return abort ? 0 : size;
7020 }
7021
7022 void SurvivorSpacePrecleanClosure::do_yield_work() {
7023 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7024 "CMS thread should hold CMS token");
7025 assert_lock_strong(_bit_map->lock());
7026 DEBUG_ONLY(RememberKlassesChecker smx(false);)
7027 // Relinquish the bit map lock
7028 _bit_map->lock()->unlock();
7029 ConcurrentMarkSweepThread::desynchronize(true);
7030 ConcurrentMarkSweepThread::acknowledge_yield_request();
7031 _collector->stopTimer();
7032 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7033 if (PrintCMSStatistics != 0) {
7034 _collector->incrementYields();
7035 }
7036 _collector->icms_wait();
7037
7038 // See the comment in coordinator_yield()
7039 for (unsigned i = 0; i < CMSYieldSleepCount &&
7040 ConcurrentMarkSweepThread::should_yield() &&
7041 !CMSCollector::foregroundGCIsActive(); ++i) {
7042 os::sleep(Thread::current(), 1, false);
7043 ConcurrentMarkSweepThread::acknowledge_yield_request();
7044 }
7045
7046 ConcurrentMarkSweepThread::synchronize(true);
7047 _bit_map->lock()->lock_without_safepoint_check();
7048 _collector->startTimer();
7049 }
7050
7051 // This closure is used to rescan the marked objects on the dirty cards
7052 // in the mod union table and the card table proper. In the parallel
7053 // case, although the bitMap is shared, we do a single read so the
7054 // isMarked() query is "safe".
7055 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7056 // Ignore mark word because we are running concurrent with mutators
7057 assert(p->is_oop_or_null(true), "expected an oop or null");
7058 HeapWord* addr = (HeapWord*)p;
7059 assert(_span.contains(addr), "we are scanning the CMS generation");
7060 bool is_obj_array = false;
7061 #ifdef DEBUG
7062 if (!_parallel) {
7063 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7064 assert(_collector->overflow_list_is_empty(),
7065 "overflow list should be empty");
7066
7067 }
7068 #endif // DEBUG
7069 if (_bit_map->isMarked(addr)) {
7070 // Obj arrays are precisely marked, non-arrays are not;
7071 // so we scan objArrays precisely and non-arrays in their
7072 // entirety.
7073 if (p->is_objArray()) {
7074 is_obj_array = true;
7075 if (_parallel) {
7076 p->oop_iterate(_par_scan_closure, mr);
7077 } else {
7078 p->oop_iterate(_scan_closure, mr);
7079 }
7080 } else {
7081 if (_parallel) {
7082 p->oop_iterate(_par_scan_closure);
7083 } else {
7084 p->oop_iterate(_scan_closure);
7085 }
7086 }
7087 }
7088 #ifdef DEBUG
7089 if (!_parallel) {
7090 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7091 assert(_collector->overflow_list_is_empty(),
7092 "overflow list should be empty");
7093
7094 }
7095 #endif // DEBUG
7096 return is_obj_array;
7097 }
7098
7099 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7100 MemRegion span,
7101 CMSBitMap* bitMap, CMSMarkStack* markStack,
7102 CMSMarkStack* revisitStack,
7103 bool should_yield, bool verifying):
7104 _collector(collector),
7105 _span(span),
7106 _bitMap(bitMap),
7107 _mut(&collector->_modUnionTable),
7108 _markStack(markStack),
7109 _revisitStack(revisitStack),
7110 _yield(should_yield),
7111 _skipBits(0)
7112 {
7113 assert(_markStack->isEmpty(), "stack should be empty");
7114 _finger = _bitMap->startWord();
7115 _threshold = _finger;
7116 assert(_collector->_restart_addr == NULL, "Sanity check");
7117 assert(_span.contains(_finger), "Out of bounds _finger?");
7118 DEBUG_ONLY(_verifying = verifying;)
7119 }
7120
7121 void MarkFromRootsClosure::reset(HeapWord* addr) {
7122 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7123 assert(_span.contains(addr), "Out of bounds _finger?");
7124 _finger = addr;
7125 _threshold = (HeapWord*)round_to(
7126 (intptr_t)_finger, CardTableModRefBS::card_size);
7127 }
7128
7129 // Should revisit to see if this should be restructured for
7130 // greater efficiency.
7131 bool MarkFromRootsClosure::do_bit(size_t offset) {
7132 if (_skipBits > 0) {
7133 _skipBits--;
7134 return true;
7135 }
7136 // convert offset into a HeapWord*
7137 HeapWord* addr = _bitMap->startWord() + offset;
7138 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7139 "address out of range");
7140 assert(_bitMap->isMarked(addr), "tautology");
7141 if (_bitMap->isMarked(addr+1)) {
7142 // this is an allocated but not yet initialized object
7143 assert(_skipBits == 0, "tautology");
7144 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7145 oop p = oop(addr);
7146 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7147 DEBUG_ONLY(if (!_verifying) {)
7148 // We re-dirty the cards on which this object lies and increase
7149 // the _threshold so that we'll come back to scan this object
7150 // during the preclean or remark phase. (CMSCleanOnEnter)
7151 if (CMSCleanOnEnter) {
7152 size_t sz = _collector->block_size_using_printezis_bits(addr);
7153 HeapWord* end_card_addr = (HeapWord*)round_to(
7154 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7155 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7156 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7157 // Bump _threshold to end_card_addr; note that
7158 // _threshold cannot possibly exceed end_card_addr, anyhow.
7159 // This prevents future clearing of the card as the scan proceeds
7160 // to the right.
7161 assert(_threshold <= end_card_addr,
7162 "Because we are just scanning into this object");
7163 if (_threshold < end_card_addr) {
7164 _threshold = end_card_addr;
7165 }
7166 if (p->klass_or_null() != NULL) {
7167 // Redirty the range of cards...
7168 _mut->mark_range(redirty_range);
7169 } // ...else the setting of klass will dirty the card anyway.
7170 }
7171 DEBUG_ONLY(})
7172 return true;
7173 }
7174 }
7175 scanOopsInOop(addr);
7176 return true;
7177 }
7178
7179 // We take a break if we've been at this for a while,
7180 // so as to avoid monopolizing the locks involved.
7181 void MarkFromRootsClosure::do_yield_work() {
7182 // First give up the locks, then yield, then re-lock
7183 // We should probably use a constructor/destructor idiom to
7184 // do this unlock/lock or modify the MutexUnlocker class to
7185 // serve our purpose. XXX
7186 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7187 "CMS thread should hold CMS token");
7188 assert_lock_strong(_bitMap->lock());
7189 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7190 _bitMap->lock()->unlock();
7191 ConcurrentMarkSweepThread::desynchronize(true);
7192 ConcurrentMarkSweepThread::acknowledge_yield_request();
7193 _collector->stopTimer();
7194 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7195 if (PrintCMSStatistics != 0) {
7196 _collector->incrementYields();
7197 }
7198 _collector->icms_wait();
7199
7200 // See the comment in coordinator_yield()
7201 for (unsigned i = 0; i < CMSYieldSleepCount &&
7202 ConcurrentMarkSweepThread::should_yield() &&
7203 !CMSCollector::foregroundGCIsActive(); ++i) {
7204 os::sleep(Thread::current(), 1, false);
7205 ConcurrentMarkSweepThread::acknowledge_yield_request();
7206 }
7207
7208 ConcurrentMarkSweepThread::synchronize(true);
7209 _bitMap->lock()->lock_without_safepoint_check();
7210 _collector->startTimer();
7211 }
7212
7213 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7214 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7215 assert(_markStack->isEmpty(),
7216 "should drain stack to limit stack usage");
7217 // convert ptr to an oop preparatory to scanning
7218 oop obj = oop(ptr);
7219 // Ignore mark word in verification below, since we
7220 // may be running concurrent with mutators.
7221 assert(obj->is_oop(true), "should be an oop");
7222 assert(_finger <= ptr, "_finger runneth ahead");
7223 // advance the finger to right end of this object
7224 _finger = ptr + obj->size();
7225 assert(_finger > ptr, "we just incremented it above");
7226 // On large heaps, it may take us some time to get through
7227 // the marking phase (especially if running iCMS). During
7228 // this time it's possible that a lot of mutations have
7229 // accumulated in the card table and the mod union table --
7230 // these mutation records are redundant until we have
7231 // actually traced into the corresponding card.
7232 // Here, we check whether advancing the finger would make
7233 // us cross into a new card, and if so clear corresponding
7234 // cards in the MUT (preclean them in the card-table in the
7235 // future).
7236
7237 DEBUG_ONLY(if (!_verifying) {)
7238 // The clean-on-enter optimization is disabled by default,
7239 // until we fix 6178663.
7240 if (CMSCleanOnEnter && (_finger > _threshold)) {
7241 // [_threshold, _finger) represents the interval
7242 // of cards to be cleared in MUT (or precleaned in card table).
7243 // The set of cards to be cleared is all those that overlap
7244 // with the interval [_threshold, _finger); note that
7245 // _threshold is always kept card-aligned but _finger isn't
7246 // always card-aligned.
7247 HeapWord* old_threshold = _threshold;
7248 assert(old_threshold == (HeapWord*)round_to(
7249 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7250 "_threshold should always be card-aligned");
7251 _threshold = (HeapWord*)round_to(
7252 (intptr_t)_finger, CardTableModRefBS::card_size);
7253 MemRegion mr(old_threshold, _threshold);
7254 assert(!mr.is_empty(), "Control point invariant");
7255 assert(_span.contains(mr), "Should clear within span");
7256 // XXX When _finger crosses from old gen into perm gen
7257 // we may be doing unnecessary cleaning; do better in the
7258 // future by detecting that condition and clearing fewer
7259 // MUT/CT entries.
7260 _mut->clear_range(mr);
7261 }
7262 DEBUG_ONLY(})
7263 // Note: the finger doesn't advance while we drain
7264 // the stack below.
7265 PushOrMarkClosure pushOrMarkClosure(_collector,
7266 _span, _bitMap, _markStack,
7267 _revisitStack,
7268 _finger, this);
7269 bool res = _markStack->push(obj);
7270 assert(res, "Empty non-zero size stack should have space for single push");
7271 while (!_markStack->isEmpty()) {
7272 oop new_oop = _markStack->pop();
7273 // Skip verifying header mark word below because we are
7274 // running concurrent with mutators.
7275 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7276 // now scan this oop's oops
7277 new_oop->oop_iterate(&pushOrMarkClosure);
7278 do_yield_check();
7279 }
7280 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7281 }
7282
7283 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7284 CMSCollector* collector, MemRegion span,
7285 CMSBitMap* bit_map,
7286 OopTaskQueue* work_queue,
7287 CMSMarkStack* overflow_stack,
7288 CMSMarkStack* revisit_stack,
7289 bool should_yield):
7290 _collector(collector),
7291 _whole_span(collector->_span),
7292 _span(span),
7293 _bit_map(bit_map),
7294 _mut(&collector->_modUnionTable),
7295 _work_queue(work_queue),
7296 _overflow_stack(overflow_stack),
7297 _revisit_stack(revisit_stack),
7298 _yield(should_yield),
7299 _skip_bits(0),
7300 _task(task)
7301 {
7302 assert(_work_queue->size() == 0, "work_queue should be empty");
7303 _finger = span.start();
7304 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7305 assert(_span.contains(_finger), "Out of bounds _finger?");
7306 }
7307
7308 // Should revisit to see if this should be restructured for
7309 // greater efficiency.
7310 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7311 if (_skip_bits > 0) {
7312 _skip_bits--;
7313 return true;
7314 }
7315 // convert offset into a HeapWord*
7316 HeapWord* addr = _bit_map->startWord() + offset;
7317 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7318 "address out of range");
7319 assert(_bit_map->isMarked(addr), "tautology");
7320 if (_bit_map->isMarked(addr+1)) {
7321 // this is an allocated object that might not yet be initialized
7322 assert(_skip_bits == 0, "tautology");
7323 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7324 oop p = oop(addr);
7325 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7326 // in the case of Clean-on-Enter optimization, redirty card
7327 // and avoid clearing card by increasing the threshold.
7328 return true;
7329 }
7330 }
7331 scan_oops_in_oop(addr);
7332 return true;
7333 }
7334
7335 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7336 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7337 // Should we assert that our work queue is empty or
7338 // below some drain limit?
7339 assert(_work_queue->size() == 0,
7340 "should drain stack to limit stack usage");
7341 // convert ptr to an oop preparatory to scanning
7342 oop obj = oop(ptr);
7343 // Ignore mark word in verification below, since we
7344 // may be running concurrent with mutators.
7345 assert(obj->is_oop(true), "should be an oop");
7346 assert(_finger <= ptr, "_finger runneth ahead");
7347 // advance the finger to right end of this object
7348 _finger = ptr + obj->size();
7349 assert(_finger > ptr, "we just incremented it above");
7350 // On large heaps, it may take us some time to get through
7351 // the marking phase (especially if running iCMS). During
7352 // this time it's possible that a lot of mutations have
7353 // accumulated in the card table and the mod union table --
7354 // these mutation records are redundant until we have
7355 // actually traced into the corresponding card.
7356 // Here, we check whether advancing the finger would make
7357 // us cross into a new card, and if so clear corresponding
7358 // cards in the MUT (preclean them in the card-table in the
7359 // future).
7360
7361 // The clean-on-enter optimization is disabled by default,
7362 // until we fix 6178663.
7363 if (CMSCleanOnEnter && (_finger > _threshold)) {
7364 // [_threshold, _finger) represents the interval
7365 // of cards to be cleared in MUT (or precleaned in card table).
7366 // The set of cards to be cleared is all those that overlap
7367 // with the interval [_threshold, _finger); note that
7368 // _threshold is always kept card-aligned but _finger isn't
7369 // always card-aligned.
7370 HeapWord* old_threshold = _threshold;
7371 assert(old_threshold == (HeapWord*)round_to(
7372 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7373 "_threshold should always be card-aligned");
7374 _threshold = (HeapWord*)round_to(
7375 (intptr_t)_finger, CardTableModRefBS::card_size);
7376 MemRegion mr(old_threshold, _threshold);
7377 assert(!mr.is_empty(), "Control point invariant");
7378 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7379 // XXX When _finger crosses from old gen into perm gen
7380 // we may be doing unnecessary cleaning; do better in the
7381 // future by detecting that condition and clearing fewer
7382 // MUT/CT entries.
7383 _mut->clear_range(mr);
7384 }
7385
7386 // Note: the local finger doesn't advance while we drain
7387 // the stack below, but the global finger sure can and will.
7388 HeapWord** gfa = _task->global_finger_addr();
7389 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7390 _span, _bit_map,
7391 _work_queue,
7392 _overflow_stack,
7393 _revisit_stack,
7394 _finger,
7395 gfa, this);
7396 bool res = _work_queue->push(obj); // overflow could occur here
7397 assert(res, "Will hold once we use workqueues");
7398 while (true) {
7399 oop new_oop;
7400 if (!_work_queue->pop_local(new_oop)) {
7401 // We emptied our work_queue; check if there's stuff that can
7402 // be gotten from the overflow stack.
7403 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7404 _overflow_stack, _work_queue)) {
7405 do_yield_check();
7406 continue;
7407 } else { // done
7408 break;
7409 }
7410 }
7411 // Skip verifying header mark word below because we are
7412 // running concurrent with mutators.
7413 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7414 // now scan this oop's oops
7415 new_oop->oop_iterate(&pushOrMarkClosure);
7416 do_yield_check();
7417 }
7418 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7419 }
7420
7421 // Yield in response to a request from VM Thread or
7422 // from mutators.
7423 void Par_MarkFromRootsClosure::do_yield_work() {
7424 assert(_task != NULL, "sanity");
7425 _task->yield();
7426 }
7427
7428 // A variant of the above used for verifying CMS marking work.
7429 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7430 MemRegion span,
7431 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7432 CMSMarkStack* mark_stack):
7433 _collector(collector),
7434 _span(span),
7435 _verification_bm(verification_bm),
7436 _cms_bm(cms_bm),
7437 _mark_stack(mark_stack),
7438 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7439 mark_stack)
7440 {
7441 assert(_mark_stack->isEmpty(), "stack should be empty");
7442 _finger = _verification_bm->startWord();
7443 assert(_collector->_restart_addr == NULL, "Sanity check");
7444 assert(_span.contains(_finger), "Out of bounds _finger?");
7445 }
7446
7447 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7448 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7449 assert(_span.contains(addr), "Out of bounds _finger?");
7450 _finger = addr;
7451 }
7452
7453 // Should revisit to see if this should be restructured for
7454 // greater efficiency.
7455 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7456 // convert offset into a HeapWord*
7457 HeapWord* addr = _verification_bm->startWord() + offset;
7458 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7459 "address out of range");
7460 assert(_verification_bm->isMarked(addr), "tautology");
7461 assert(_cms_bm->isMarked(addr), "tautology");
7462
7463 assert(_mark_stack->isEmpty(),
7464 "should drain stack to limit stack usage");
7465 // convert addr to an oop preparatory to scanning
7466 oop obj = oop(addr);
7467 assert(obj->is_oop(), "should be an oop");
7468 assert(_finger <= addr, "_finger runneth ahead");
7469 // advance the finger to right end of this object
7470 _finger = addr + obj->size();
7471 assert(_finger > addr, "we just incremented it above");
7472 // Note: the finger doesn't advance while we drain
7473 // the stack below.
7474 bool res = _mark_stack->push(obj);
7475 assert(res, "Empty non-zero size stack should have space for single push");
7476 while (!_mark_stack->isEmpty()) {
7477 oop new_oop = _mark_stack->pop();
7478 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7479 // now scan this oop's oops
7480 new_oop->oop_iterate(&_pam_verify_closure);
7481 }
7482 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7483 return true;
7484 }
7485
7486 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7487 CMSCollector* collector, MemRegion span,
7488 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7489 CMSMarkStack* mark_stack):
7490 OopClosure(collector->ref_processor()),
7491 _collector(collector),
7492 _span(span),
7493 _verification_bm(verification_bm),
7494 _cms_bm(cms_bm),
7495 _mark_stack(mark_stack)
7496 { }
7497
7498 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7499 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7500
7501 // Upon stack overflow, we discard (part of) the stack,
7502 // remembering the least address amongst those discarded
7503 // in CMSCollector's _restart_address.
7504 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7505 // Remember the least grey address discarded
7506 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7507 _collector->lower_restart_addr(ra);
7508 _mark_stack->reset(); // discard stack contents
7509 _mark_stack->expand(); // expand the stack if possible
7510 }
7511
7512 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7513 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7514 HeapWord* addr = (HeapWord*)obj;
7515 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7516 // Oop lies in _span and isn't yet grey or black
7517 _verification_bm->mark(addr); // now grey
7518 if (!_cms_bm->isMarked(addr)) {
7519 oop(addr)->print();
7520 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7521 addr);
7522 fatal("... aborting");
7523 }
7524
7525 if (!_mark_stack->push(obj)) { // stack overflow
7526 if (PrintCMSStatistics != 0) {
7527 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7528 SIZE_FORMAT, _mark_stack->capacity());
7529 }
7530 assert(_mark_stack->isFull(), "Else push should have succeeded");
7531 handle_stack_overflow(addr);
7532 }
7533 // anything including and to the right of _finger
7534 // will be scanned as we iterate over the remainder of the
7535 // bit map
7536 }
7537 }
7538
7539 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7540 MemRegion span,
7541 CMSBitMap* bitMap, CMSMarkStack* markStack,
7542 CMSMarkStack* revisitStack,
7543 HeapWord* finger, MarkFromRootsClosure* parent) :
7544 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack),
7545 _span(span),
7546 _bitMap(bitMap),
7547 _markStack(markStack),
7548 _finger(finger),
7549 _parent(parent)
7550 { }
7551
7552 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7553 MemRegion span,
7554 CMSBitMap* bit_map,
7555 OopTaskQueue* work_queue,
7556 CMSMarkStack* overflow_stack,
7557 CMSMarkStack* revisit_stack,
7558 HeapWord* finger,
7559 HeapWord** global_finger_addr,
7560 Par_MarkFromRootsClosure* parent) :
7561 Par_KlassRememberingOopClosure(collector,
7562 collector->ref_processor(),
7563 revisit_stack),
7564 _whole_span(collector->_span),
7565 _span(span),
7566 _bit_map(bit_map),
7567 _work_queue(work_queue),
7568 _overflow_stack(overflow_stack),
7569 _finger(finger),
7570 _global_finger_addr(global_finger_addr),
7571 _parent(parent)
7572 { }
7573
7574 // Assumes thread-safe access by callers, who are
7575 // responsible for mutual exclusion.
7576 void CMSCollector::lower_restart_addr(HeapWord* low) {
7577 assert(_span.contains(low), "Out of bounds addr");
7578 if (_restart_addr == NULL) {
7579 _restart_addr = low;
7580 } else {
7581 _restart_addr = MIN2(_restart_addr, low);
7582 }
7583 }
7584
7585 // Upon stack overflow, we discard (part of) the stack,
7586 // remembering the least address amongst those discarded
7587 // in CMSCollector's _restart_address.
7588 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7589 // Remember the least grey address discarded
7590 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7591 _collector->lower_restart_addr(ra);
7592 _markStack->reset(); // discard stack contents
7593 _markStack->expand(); // expand the stack if possible
7594 }
7595
7596 // Upon stack overflow, we discard (part of) the stack,
7597 // remembering the least address amongst those discarded
7598 // in CMSCollector's _restart_address.
7599 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7600 // We need to do this under a mutex to prevent other
7601 // workers from interfering with the work done below.
7602 MutexLockerEx ml(_overflow_stack->par_lock(),
7603 Mutex::_no_safepoint_check_flag);
7604 // Remember the least grey address discarded
7605 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7606 _collector->lower_restart_addr(ra);
7607 _overflow_stack->reset(); // discard stack contents
7608 _overflow_stack->expand(); // expand the stack if possible
7609 }
7610
7611 void PushOrMarkClosure::do_oop(oop obj) {
7612 // Ignore mark word because we are running concurrent with mutators.
7613 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7614 HeapWord* addr = (HeapWord*)obj;
7615 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7616 // Oop lies in _span and isn't yet grey or black
7617 _bitMap->mark(addr); // now grey
7618 if (addr < _finger) {
7619 // the bit map iteration has already either passed, or
7620 // sampled, this bit in the bit map; we'll need to
7621 // use the marking stack to scan this oop's oops.
7622 bool simulate_overflow = false;
7623 NOT_PRODUCT(
7624 if (CMSMarkStackOverflowALot &&
7625 _collector->simulate_overflow()) {
7626 // simulate a stack overflow
7627 simulate_overflow = true;
7628 }
7629 )
7630 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7631 if (PrintCMSStatistics != 0) {
7632 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7633 SIZE_FORMAT, _markStack->capacity());
7634 }
7635 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7636 handle_stack_overflow(addr);
7637 }
7638 }
7639 // anything including and to the right of _finger
7640 // will be scanned as we iterate over the remainder of the
7641 // bit map
7642 do_yield_check();
7643 }
7644 }
7645
7646 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7647 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7648
7649 void Par_PushOrMarkClosure::do_oop(oop obj) {
7650 // Ignore mark word because we are running concurrent with mutators.
7651 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7652 HeapWord* addr = (HeapWord*)obj;
7653 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7654 // Oop lies in _span and isn't yet grey or black
7655 // We read the global_finger (volatile read) strictly after marking oop
7656 bool res = _bit_map->par_mark(addr); // now grey
7657 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7658 // Should we push this marked oop on our stack?
7659 // -- if someone else marked it, nothing to do
7660 // -- if target oop is above global finger nothing to do
7661 // -- if target oop is in chunk and above local finger
7662 // then nothing to do
7663 // -- else push on work queue
7664 if ( !res // someone else marked it, they will deal with it
7665 || (addr >= *gfa) // will be scanned in a later task
7666 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7667 return;
7668 }
7669 // the bit map iteration has already either passed, or
7670 // sampled, this bit in the bit map; we'll need to
7671 // use the marking stack to scan this oop's oops.
7672 bool simulate_overflow = false;
7673 NOT_PRODUCT(
7674 if (CMSMarkStackOverflowALot &&
7675 _collector->simulate_overflow()) {
7676 // simulate a stack overflow
7677 simulate_overflow = true;
7678 }
7679 )
7680 if (simulate_overflow ||
7681 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7682 // stack overflow
7683 if (PrintCMSStatistics != 0) {
7684 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7685 SIZE_FORMAT, _overflow_stack->capacity());
7686 }
7687 // We cannot assert that the overflow stack is full because
7688 // it may have been emptied since.
7689 assert(simulate_overflow ||
7690 _work_queue->size() == _work_queue->max_elems(),
7691 "Else push should have succeeded");
7692 handle_stack_overflow(addr);
7693 }
7694 do_yield_check();
7695 }
7696 }
7697
7698 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7699 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7700
7701 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector,
7702 ReferenceProcessor* rp,
7703 CMSMarkStack* revisit_stack) :
7704 OopClosure(rp),
7705 _collector(collector),
7706 _revisit_stack(revisit_stack),
7707 _should_remember_klasses(collector->should_unload_classes()) {}
7708
7709 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7710 MemRegion span,
7711 ReferenceProcessor* rp,
7712 CMSBitMap* bit_map,
7713 CMSBitMap* mod_union_table,
7714 CMSMarkStack* mark_stack,
7715 CMSMarkStack* revisit_stack,
7716 bool concurrent_precleaning):
7717 KlassRememberingOopClosure(collector, rp, revisit_stack),
7718 _span(span),
7719 _bit_map(bit_map),
7720 _mod_union_table(mod_union_table),
7721 _mark_stack(mark_stack),
7722 _concurrent_precleaning(concurrent_precleaning)
7723 {
7724 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7725 }
7726
7727 // Grey object rescan during pre-cleaning and second checkpoint phases --
7728 // the non-parallel version (the parallel version appears further below.)
7729 void PushAndMarkClosure::do_oop(oop obj) {
7730 // Ignore mark word verification. If during concurrent precleaning,
7731 // the object monitor may be locked. If during the checkpoint
7732 // phases, the object may already have been reached by a different
7733 // path and may be at the end of the global overflow list (so
7734 // the mark word may be NULL).
7735 assert(obj->is_oop_or_null(true /* ignore mark word */),
7736 "expected an oop or NULL");
7737 HeapWord* addr = (HeapWord*)obj;
7738 // Check if oop points into the CMS generation
7739 // and is not marked
7740 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7741 // a white object ...
7742 _bit_map->mark(addr); // ... now grey
7743 // push on the marking stack (grey set)
7744 bool simulate_overflow = false;
7745 NOT_PRODUCT(
7746 if (CMSMarkStackOverflowALot &&
7747 _collector->simulate_overflow()) {
7748 // simulate a stack overflow
7749 simulate_overflow = true;
7750 }
7751 )
7752 if (simulate_overflow || !_mark_stack->push(obj)) {
7753 if (_concurrent_precleaning) {
7754 // During precleaning we can just dirty the appropriate card(s)
7755 // in the mod union table, thus ensuring that the object remains
7756 // in the grey set and continue. In the case of object arrays
7757 // we need to dirty all of the cards that the object spans,
7758 // since the rescan of object arrays will be limited to the
7759 // dirty cards.
7760 // Note that no one can be intefering with us in this action
7761 // of dirtying the mod union table, so no locking or atomics
7762 // are required.
7763 if (obj->is_objArray()) {
7764 size_t sz = obj->size();
7765 HeapWord* end_card_addr = (HeapWord*)round_to(
7766 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7767 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7768 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7769 _mod_union_table->mark_range(redirty_range);
7770 } else {
7771 _mod_union_table->mark(addr);
7772 }
7773 _collector->_ser_pmc_preclean_ovflw++;
7774 } else {
7775 // During the remark phase, we need to remember this oop
7776 // in the overflow list.
7777 _collector->push_on_overflow_list(obj);
7778 _collector->_ser_pmc_remark_ovflw++;
7779 }
7780 }
7781 }
7782 }
7783
7784 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7785 MemRegion span,
7786 ReferenceProcessor* rp,
7787 CMSBitMap* bit_map,
7788 OopTaskQueue* work_queue,
7789 CMSMarkStack* revisit_stack):
7790 Par_KlassRememberingOopClosure(collector, rp, revisit_stack),
7791 _span(span),
7792 _bit_map(bit_map),
7793 _work_queue(work_queue)
7794 {
7795 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7796 }
7797
7798 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7799 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7800
7801 // Grey object rescan during second checkpoint phase --
7802 // the parallel version.
7803 void Par_PushAndMarkClosure::do_oop(oop obj) {
7804 // In the assert below, we ignore the mark word because
7805 // this oop may point to an already visited object that is
7806 // on the overflow stack (in which case the mark word has
7807 // been hijacked for chaining into the overflow stack --
7808 // if this is the last object in the overflow stack then
7809 // its mark word will be NULL). Because this object may
7810 // have been subsequently popped off the global overflow
7811 // stack, and the mark word possibly restored to the prototypical
7812 // value, by the time we get to examined this failing assert in
7813 // the debugger, is_oop_or_null(false) may subsequently start
7814 // to hold.
7815 assert(obj->is_oop_or_null(true),
7816 "expected an oop or NULL");
7817 HeapWord* addr = (HeapWord*)obj;
7818 // Check if oop points into the CMS generation
7819 // and is not marked
7820 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7821 // a white object ...
7822 // If we manage to "claim" the object, by being the
7823 // first thread to mark it, then we push it on our
7824 // marking stack
7825 if (_bit_map->par_mark(addr)) { // ... now grey
7826 // push on work queue (grey set)
7827 bool simulate_overflow = false;
7828 NOT_PRODUCT(
7829 if (CMSMarkStackOverflowALot &&
7830 _collector->par_simulate_overflow()) {
7831 // simulate a stack overflow
7832 simulate_overflow = true;
7833 }
7834 )
7835 if (simulate_overflow || !_work_queue->push(obj)) {
7836 _collector->par_push_on_overflow_list(obj);
7837 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7838 }
7839 } // Else, some other thread got there first
7840 }
7841 }
7842
7843 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7844 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7845
7846 void PushAndMarkClosure::remember_mdo(DataLayout* v) {
7847 // TBD
7848 }
7849
7850 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) {
7851 // TBD
7852 }
7853
7854 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7855 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7856 Mutex* bml = _collector->bitMapLock();
7857 assert_lock_strong(bml);
7858 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7859 "CMS thread should hold CMS token");
7860
7861 bml->unlock();
7862 ConcurrentMarkSweepThread::desynchronize(true);
7863
7864 ConcurrentMarkSweepThread::acknowledge_yield_request();
7865
7866 _collector->stopTimer();
7867 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7868 if (PrintCMSStatistics != 0) {
7869 _collector->incrementYields();
7870 }
7871 _collector->icms_wait();
7872
7873 // See the comment in coordinator_yield()
7874 for (unsigned i = 0; i < CMSYieldSleepCount &&
7875 ConcurrentMarkSweepThread::should_yield() &&
7876 !CMSCollector::foregroundGCIsActive(); ++i) {
7877 os::sleep(Thread::current(), 1, false);
7878 ConcurrentMarkSweepThread::acknowledge_yield_request();
7879 }
7880
7881 ConcurrentMarkSweepThread::synchronize(true);
7882 bml->lock();
7883
7884 _collector->startTimer();
7885 }
7886
7887 bool CMSPrecleanRefsYieldClosure::should_return() {
7888 if (ConcurrentMarkSweepThread::should_yield()) {
7889 do_yield_work();
7890 }
7891 return _collector->foregroundGCIsActive();
7892 }
7893
7894 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7895 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7896 "mr should be aligned to start at a card boundary");
7897 // We'd like to assert:
7898 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7899 // "mr should be a range of cards");
7900 // However, that would be too strong in one case -- the last
7901 // partition ends at _unallocated_block which, in general, can be
7902 // an arbitrary boundary, not necessarily card aligned.
7903 if (PrintCMSStatistics != 0) {
7904 _num_dirty_cards +=
7905 mr.word_size()/CardTableModRefBS::card_size_in_words;
7906 }
7907 _space->object_iterate_mem(mr, &_scan_cl);
7908 }
7909
7910 SweepClosure::SweepClosure(CMSCollector* collector,
7911 ConcurrentMarkSweepGeneration* g,
7912 CMSBitMap* bitMap, bool should_yield) :
7913 _collector(collector),
7914 _g(g),
7915 _sp(g->cmsSpace()),
7916 _limit(_sp->sweep_limit()),
7917 _freelistLock(_sp->freelistLock()),
7918 _bitMap(bitMap),
7919 _yield(should_yield),
7920 _inFreeRange(false), // No free range at beginning of sweep
7921 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7922 _lastFreeRangeCoalesced(false),
7923 _freeFinger(g->used_region().start())
7924 {
7925 NOT_PRODUCT(
7926 _numObjectsFreed = 0;
7927 _numWordsFreed = 0;
7928 _numObjectsLive = 0;
7929 _numWordsLive = 0;
7930 _numObjectsAlreadyFree = 0;
7931 _numWordsAlreadyFree = 0;
7932 _last_fc = NULL;
7933
7934 _sp->initializeIndexedFreeListArrayReturnedBytes();
7935 _sp->dictionary()->initialize_dict_returned_bytes();
7936 )
7937 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7938 "sweep _limit out of bounds");
7939 if (CMSTraceSweeper) {
7940 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7941 _limit);
7942 }
7943 }
7944
7945 void SweepClosure::print_on(outputStream* st) const {
7946 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7947 _sp->bottom(), _sp->end());
7948 tty->print_cr("_limit = " PTR_FORMAT, _limit);
7949 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
7950 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
7951 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7952 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7953 }
7954
7955 #ifndef PRODUCT
7956 // Assertion checking only: no useful work in product mode --
7957 // however, if any of the flags below become product flags,
7958 // you may need to review this code to see if it needs to be
7959 // enabled in product mode.
7960 SweepClosure::~SweepClosure() {
7961 assert_lock_strong(_freelistLock);
7962 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7963 "sweep _limit out of bounds");
7964 if (inFreeRange()) {
7965 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
7966 print();
7967 ShouldNotReachHere();
7968 }
7969 if (Verbose && PrintGC) {
7970 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
7971 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7972 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
7973 SIZE_FORMAT" bytes "
7974 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
7975 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
7976 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7977 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
7978 * sizeof(HeapWord);
7979 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
7980
7981 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
7982 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7983 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7984 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7985 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
7986 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
7987 indexListReturnedBytes);
7988 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
7989 dict_returned_bytes);
7990 }
7991 }
7992 if (CMSTraceSweeper) {
7993 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
7994 _limit);
7995 }
7996 }
7997 #endif // PRODUCT
7998
7999 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8000 bool freeRangeInFreeLists) {
8001 if (CMSTraceSweeper) {
8002 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
8003 freeFinger, freeRangeInFreeLists);
8004 }
8005 assert(!inFreeRange(), "Trampling existing free range");
8006 set_inFreeRange(true);
8007 set_lastFreeRangeCoalesced(false);
8008
8009 set_freeFinger(freeFinger);
8010 set_freeRangeInFreeLists(freeRangeInFreeLists);
8011 if (CMSTestInFreeList) {
8012 if (freeRangeInFreeLists) {
8013 FreeChunk* fc = (FreeChunk*) freeFinger;
8014 assert(fc->is_free(), "A chunk on the free list should be free.");
8015 assert(fc->size() > 0, "Free range should have a size");
8016 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8017 }
8018 }
8019 }
8020
8021 // Note that the sweeper runs concurrently with mutators. Thus,
8022 // it is possible for direct allocation in this generation to happen
8023 // in the middle of the sweep. Note that the sweeper also coalesces
8024 // contiguous free blocks. Thus, unless the sweeper and the allocator
8025 // synchronize appropriately freshly allocated blocks may get swept up.
8026 // This is accomplished by the sweeper locking the free lists while
8027 // it is sweeping. Thus blocks that are determined to be free are
8028 // indeed free. There is however one additional complication:
8029 // blocks that have been allocated since the final checkpoint and
8030 // mark, will not have been marked and so would be treated as
8031 // unreachable and swept up. To prevent this, the allocator marks
8032 // the bit map when allocating during the sweep phase. This leads,
8033 // however, to a further complication -- objects may have been allocated
8034 // but not yet initialized -- in the sense that the header isn't yet
8035 // installed. The sweeper can not then determine the size of the block
8036 // in order to skip over it. To deal with this case, we use a technique
8037 // (due to Printezis) to encode such uninitialized block sizes in the
8038 // bit map. Since the bit map uses a bit per every HeapWord, but the
8039 // CMS generation has a minimum object size of 3 HeapWords, it follows
8040 // that "normal marks" won't be adjacent in the bit map (there will
8041 // always be at least two 0 bits between successive 1 bits). We make use
8042 // of these "unused" bits to represent uninitialized blocks -- the bit
8043 // corresponding to the start of the uninitialized object and the next
8044 // bit are both set. Finally, a 1 bit marks the end of the object that
8045 // started with the two consecutive 1 bits to indicate its potentially
8046 // uninitialized state.
8047
8048 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8049 FreeChunk* fc = (FreeChunk*)addr;
8050 size_t res;
8051
8052 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8053 // than "addr == _limit" because although _limit was a block boundary when
8054 // we started the sweep, it may no longer be one because heap expansion
8055 // may have caused us to coalesce the block ending at the address _limit
8056 // with a newly expanded chunk (this happens when _limit was set to the
8057 // previous _end of the space), so we may have stepped past _limit:
8058 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8059 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8060 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8061 "sweep _limit out of bounds");
8062 assert(addr < _sp->end(), "addr out of bounds");
8063 // Flush any free range we might be holding as a single
8064 // coalesced chunk to the appropriate free list.
8065 if (inFreeRange()) {
8066 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8067 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8068 flush_cur_free_chunk(freeFinger(),
8069 pointer_delta(addr, freeFinger()));
8070 if (CMSTraceSweeper) {
8071 gclog_or_tty->print("Sweep: last chunk: ");
8072 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8073 "[coalesced:"SIZE_FORMAT"]\n",
8074 freeFinger(), pointer_delta(addr, freeFinger()),
8075 lastFreeRangeCoalesced());
8076 }
8077 }
8078
8079 // help the iterator loop finish
8080 return pointer_delta(_sp->end(), addr);
8081 }
8082
8083 assert(addr < _limit, "sweep invariant");
8084 // check if we should yield
8085 do_yield_check(addr);
8086 if (fc->is_free()) {
8087 // Chunk that is already free
8088 res = fc->size();
8089 do_already_free_chunk(fc);
8090 debug_only(_sp->verifyFreeLists());
8091 // If we flush the chunk at hand in lookahead_and_flush()
8092 // and it's coalesced with a preceding chunk, then the
8093 // process of "mangling" the payload of the coalesced block
8094 // will cause erasure of the size information from the
8095 // (erstwhile) header of all the coalesced blocks but the
8096 // first, so the first disjunct in the assert will not hold
8097 // in that specific case (in which case the second disjunct
8098 // will hold).
8099 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8100 "Otherwise the size info doesn't change at this step");
8101 NOT_PRODUCT(
8102 _numObjectsAlreadyFree++;
8103 _numWordsAlreadyFree += res;
8104 )
8105 NOT_PRODUCT(_last_fc = fc;)
8106 } else if (!_bitMap->isMarked(addr)) {
8107 // Chunk is fresh garbage
8108 res = do_garbage_chunk(fc);
8109 debug_only(_sp->verifyFreeLists());
8110 NOT_PRODUCT(
8111 _numObjectsFreed++;
8112 _numWordsFreed += res;
8113 )
8114 } else {
8115 // Chunk that is alive.
8116 res = do_live_chunk(fc);
8117 debug_only(_sp->verifyFreeLists());
8118 NOT_PRODUCT(
8119 _numObjectsLive++;
8120 _numWordsLive += res;
8121 )
8122 }
8123 return res;
8124 }
8125
8126 // For the smart allocation, record following
8127 // split deaths - a free chunk is removed from its free list because
8128 // it is being split into two or more chunks.
8129 // split birth - a free chunk is being added to its free list because
8130 // a larger free chunk has been split and resulted in this free chunk.
8131 // coal death - a free chunk is being removed from its free list because
8132 // it is being coalesced into a large free chunk.
8133 // coal birth - a free chunk is being added to its free list because
8134 // it was created when two or more free chunks where coalesced into
8135 // this free chunk.
8136 //
8137 // These statistics are used to determine the desired number of free
8138 // chunks of a given size. The desired number is chosen to be relative
8139 // to the end of a CMS sweep. The desired number at the end of a sweep
8140 // is the
8141 // count-at-end-of-previous-sweep (an amount that was enough)
8142 // - count-at-beginning-of-current-sweep (the excess)
8143 // + split-births (gains in this size during interval)
8144 // - split-deaths (demands on this size during interval)
8145 // where the interval is from the end of one sweep to the end of the
8146 // next.
8147 //
8148 // When sweeping the sweeper maintains an accumulated chunk which is
8149 // the chunk that is made up of chunks that have been coalesced. That
8150 // will be termed the left-hand chunk. A new chunk of garbage that
8151 // is being considered for coalescing will be referred to as the
8152 // right-hand chunk.
8153 //
8154 // When making a decision on whether to coalesce a right-hand chunk with
8155 // the current left-hand chunk, the current count vs. the desired count
8156 // of the left-hand chunk is considered. Also if the right-hand chunk
8157 // is near the large chunk at the end of the heap (see
8158 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8159 // left-hand chunk is coalesced.
8160 //
8161 // When making a decision about whether to split a chunk, the desired count
8162 // vs. the current count of the candidate to be split is also considered.
8163 // If the candidate is underpopulated (currently fewer chunks than desired)
8164 // a chunk of an overpopulated (currently more chunks than desired) size may
8165 // be chosen. The "hint" associated with a free list, if non-null, points
8166 // to a free list which may be overpopulated.
8167 //
8168
8169 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8170 const size_t size = fc->size();
8171 // Chunks that cannot be coalesced are not in the
8172 // free lists.
8173 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8174 assert(_sp->verify_chunk_in_free_list(fc),
8175 "free chunk should be in free lists");
8176 }
8177 // a chunk that is already free, should not have been
8178 // marked in the bit map
8179 HeapWord* const addr = (HeapWord*) fc;
8180 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8181 // Verify that the bit map has no bits marked between
8182 // addr and purported end of this block.
8183 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8184
8185 // Some chunks cannot be coalesced under any circumstances.
8186 // See the definition of cantCoalesce().
8187 if (!fc->cantCoalesce()) {
8188 // This chunk can potentially be coalesced.
8189 if (_sp->adaptive_freelists()) {
8190 // All the work is done in
8191 do_post_free_or_garbage_chunk(fc, size);
8192 } else { // Not adaptive free lists
8193 // this is a free chunk that can potentially be coalesced by the sweeper;
8194 if (!inFreeRange()) {
8195 // if the next chunk is a free block that can't be coalesced
8196 // it doesn't make sense to remove this chunk from the free lists
8197 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8198 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8199 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8200 nextChunk->is_free() && // ... which is free...
8201 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8202 // nothing to do
8203 } else {
8204 // Potentially the start of a new free range:
8205 // Don't eagerly remove it from the free lists.
8206 // No need to remove it if it will just be put
8207 // back again. (Also from a pragmatic point of view
8208 // if it is a free block in a region that is beyond
8209 // any allocated blocks, an assertion will fail)
8210 // Remember the start of a free run.
8211 initialize_free_range(addr, true);
8212 // end - can coalesce with next chunk
8213 }
8214 } else {
8215 // the midst of a free range, we are coalescing
8216 print_free_block_coalesced(fc);
8217 if (CMSTraceSweeper) {
8218 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8219 }
8220 // remove it from the free lists
8221 _sp->removeFreeChunkFromFreeLists(fc);
8222 set_lastFreeRangeCoalesced(true);
8223 // If the chunk is being coalesced and the current free range is
8224 // in the free lists, remove the current free range so that it
8225 // will be returned to the free lists in its entirety - all
8226 // the coalesced pieces included.
8227 if (freeRangeInFreeLists()) {
8228 FreeChunk* ffc = (FreeChunk*) freeFinger();
8229 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8230 "Size of free range is inconsistent with chunk size.");
8231 if (CMSTestInFreeList) {
8232 assert(_sp->verify_chunk_in_free_list(ffc),
8233 "free range is not in free lists");
8234 }
8235 _sp->removeFreeChunkFromFreeLists(ffc);
8236 set_freeRangeInFreeLists(false);
8237 }
8238 }
8239 }
8240 // Note that if the chunk is not coalescable (the else arm
8241 // below), we unconditionally flush, without needing to do
8242 // a "lookahead," as we do below.
8243 if (inFreeRange()) lookahead_and_flush(fc, size);
8244 } else {
8245 // Code path common to both original and adaptive free lists.
8246
8247 // cant coalesce with previous block; this should be treated
8248 // as the end of a free run if any
8249 if (inFreeRange()) {
8250 // we kicked some butt; time to pick up the garbage
8251 assert(freeFinger() < addr, "freeFinger points too high");
8252 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8253 }
8254 // else, nothing to do, just continue
8255 }
8256 }
8257
8258 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8259 // This is a chunk of garbage. It is not in any free list.
8260 // Add it to a free list or let it possibly be coalesced into
8261 // a larger chunk.
8262 HeapWord* const addr = (HeapWord*) fc;
8263 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8264
8265 if (_sp->adaptive_freelists()) {
8266 // Verify that the bit map has no bits marked between
8267 // addr and purported end of just dead object.
8268 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8269
8270 do_post_free_or_garbage_chunk(fc, size);
8271 } else {
8272 if (!inFreeRange()) {
8273 // start of a new free range
8274 assert(size > 0, "A free range should have a size");
8275 initialize_free_range(addr, false);
8276 } else {
8277 // this will be swept up when we hit the end of the
8278 // free range
8279 if (CMSTraceSweeper) {
8280 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8281 }
8282 // If the chunk is being coalesced and the current free range is
8283 // in the free lists, remove the current free range so that it
8284 // will be returned to the free lists in its entirety - all
8285 // the coalesced pieces included.
8286 if (freeRangeInFreeLists()) {
8287 FreeChunk* ffc = (FreeChunk*)freeFinger();
8288 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8289 "Size of free range is inconsistent with chunk size.");
8290 if (CMSTestInFreeList) {
8291 assert(_sp->verify_chunk_in_free_list(ffc),
8292 "free range is not in free lists");
8293 }
8294 _sp->removeFreeChunkFromFreeLists(ffc);
8295 set_freeRangeInFreeLists(false);
8296 }
8297 set_lastFreeRangeCoalesced(true);
8298 }
8299 // this will be swept up when we hit the end of the free range
8300
8301 // Verify that the bit map has no bits marked between
8302 // addr and purported end of just dead object.
8303 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8304 }
8305 assert(_limit >= addr + size,
8306 "A freshly garbage chunk can't possibly straddle over _limit");
8307 if (inFreeRange()) lookahead_and_flush(fc, size);
8308 return size;
8309 }
8310
8311 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8312 HeapWord* addr = (HeapWord*) fc;
8313 // The sweeper has just found a live object. Return any accumulated
8314 // left hand chunk to the free lists.
8315 if (inFreeRange()) {
8316 assert(freeFinger() < addr, "freeFinger points too high");
8317 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8318 }
8319
8320 // This object is live: we'd normally expect this to be
8321 // an oop, and like to assert the following:
8322 // assert(oop(addr)->is_oop(), "live block should be an oop");
8323 // However, as we commented above, this may be an object whose
8324 // header hasn't yet been initialized.
8325 size_t size;
8326 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8327 if (_bitMap->isMarked(addr + 1)) {
8328 // Determine the size from the bit map, rather than trying to
8329 // compute it from the object header.
8330 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8331 size = pointer_delta(nextOneAddr + 1, addr);
8332 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8333 "alignment problem");
8334
8335 #ifdef DEBUG
8336 if (oop(addr)->klass_or_null() != NULL &&
8337 ( !_collector->should_unload_classes()
8338 || (oop(addr)->is_parsable()) &&
8339 oop(addr)->is_conc_safe())) {
8340 // Ignore mark word because we are running concurrent with mutators
8341 assert(oop(addr)->is_oop(true), "live block should be an oop");
8342 // is_conc_safe is checked before performing this assertion
8343 // because an object that is not is_conc_safe may yet have
8344 // the return from size() correct.
8345 assert(size ==
8346 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8347 "P-mark and computed size do not agree");
8348 }
8349 #endif
8350
8351 } else {
8352 // This should be an initialized object that's alive.
8353 assert(oop(addr)->klass_or_null() != NULL &&
8354 (!_collector->should_unload_classes()
8355 || oop(addr)->is_parsable()),
8356 "Should be an initialized object");
8357 // Note that there are objects used during class redefinition,
8358 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(),
8359 // which are discarded with their is_conc_safe state still
8360 // false. These object may be floating garbage so may be
8361 // seen here. If they are floating garbage their size
8362 // should be attainable from their klass. Do not that
8363 // is_conc_safe() is true for oop(addr).
8364 // Ignore mark word because we are running concurrent with mutators
8365 assert(oop(addr)->is_oop(true), "live block should be an oop");
8366 // Verify that the bit map has no bits marked between
8367 // addr and purported end of this block.
8368 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8369 assert(size >= 3, "Necessary for Printezis marks to work");
8370 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8371 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8372 }
8373 return size;
8374 }
8375
8376 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8377 size_t chunkSize) {
8378 // do_post_free_or_garbage_chunk() should only be called in the case
8379 // of the adaptive free list allocator.
8380 const bool fcInFreeLists = fc->is_free();
8381 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8382 assert((HeapWord*)fc <= _limit, "sweep invariant");
8383 if (CMSTestInFreeList && fcInFreeLists) {
8384 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8385 }
8386
8387 if (CMSTraceSweeper) {
8388 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8389 }
8390
8391 HeapWord* const fc_addr = (HeapWord*) fc;
8392
8393 bool coalesce;
8394 const size_t left = pointer_delta(fc_addr, freeFinger());
8395 const size_t right = chunkSize;
8396 switch (FLSCoalescePolicy) {
8397 // numeric value forms a coalition aggressiveness metric
8398 case 0: { // never coalesce
8399 coalesce = false;
8400 break;
8401 }
8402 case 1: { // coalesce if left & right chunks on overpopulated lists
8403 coalesce = _sp->coalOverPopulated(left) &&
8404 _sp->coalOverPopulated(right);
8405 break;
8406 }
8407 case 2: { // coalesce if left chunk on overpopulated list (default)
8408 coalesce = _sp->coalOverPopulated(left);
8409 break;
8410 }
8411 case 3: { // coalesce if left OR right chunk on overpopulated list
8412 coalesce = _sp->coalOverPopulated(left) ||
8413 _sp->coalOverPopulated(right);
8414 break;
8415 }
8416 case 4: { // always coalesce
8417 coalesce = true;
8418 break;
8419 }
8420 default:
8421 ShouldNotReachHere();
8422 }
8423
8424 // Should the current free range be coalesced?
8425 // If the chunk is in a free range and either we decided to coalesce above
8426 // or the chunk is near the large block at the end of the heap
8427 // (isNearLargestChunk() returns true), then coalesce this chunk.
8428 const bool doCoalesce = inFreeRange()
8429 && (coalesce || _g->isNearLargestChunk(fc_addr));
8430 if (doCoalesce) {
8431 // Coalesce the current free range on the left with the new
8432 // chunk on the right. If either is on a free list,
8433 // it must be removed from the list and stashed in the closure.
8434 if (freeRangeInFreeLists()) {
8435 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8436 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8437 "Size of free range is inconsistent with chunk size.");
8438 if (CMSTestInFreeList) {
8439 assert(_sp->verify_chunk_in_free_list(ffc),
8440 "Chunk is not in free lists");
8441 }
8442 _sp->coalDeath(ffc->size());
8443 _sp->removeFreeChunkFromFreeLists(ffc);
8444 set_freeRangeInFreeLists(false);
8445 }
8446 if (fcInFreeLists) {
8447 _sp->coalDeath(chunkSize);
8448 assert(fc->size() == chunkSize,
8449 "The chunk has the wrong size or is not in the free lists");
8450 _sp->removeFreeChunkFromFreeLists(fc);
8451 }
8452 set_lastFreeRangeCoalesced(true);
8453 print_free_block_coalesced(fc);
8454 } else { // not in a free range and/or should not coalesce
8455 // Return the current free range and start a new one.
8456 if (inFreeRange()) {
8457 // In a free range but cannot coalesce with the right hand chunk.
8458 // Put the current free range into the free lists.
8459 flush_cur_free_chunk(freeFinger(),
8460 pointer_delta(fc_addr, freeFinger()));
8461 }
8462 // Set up for new free range. Pass along whether the right hand
8463 // chunk is in the free lists.
8464 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8465 }
8466 }
8467
8468 // Lookahead flush:
8469 // If we are tracking a free range, and this is the last chunk that
8470 // we'll look at because its end crosses past _limit, we'll preemptively
8471 // flush it along with any free range we may be holding on to. Note that
8472 // this can be the case only for an already free or freshly garbage
8473 // chunk. If this block is an object, it can never straddle
8474 // over _limit. The "straddling" occurs when _limit is set at
8475 // the previous end of the space when this cycle started, and
8476 // a subsequent heap expansion caused the previously co-terminal
8477 // free block to be coalesced with the newly expanded portion,
8478 // thus rendering _limit a non-block-boundary making it dangerous
8479 // for the sweeper to step over and examine.
8480 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8481 assert(inFreeRange(), "Should only be called if currently in a free range.");
8482 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8483 assert(_sp->used_region().contains(eob - 1),
8484 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8485 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8486 _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8487 if (eob >= _limit) {
8488 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8489 if (CMSTraceSweeper) {
8490 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8491 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8492 "[" PTR_FORMAT "," PTR_FORMAT ")",
8493 _limit, fc, eob, _sp->bottom(), _sp->end());
8494 }
8495 // Return the storage we are tracking back into the free lists.
8496 if (CMSTraceSweeper) {
8497 gclog_or_tty->print_cr("Flushing ... ");
8498 }
8499 assert(freeFinger() < eob, "Error");
8500 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8501 }
8502 }
8503
8504 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8505 assert(inFreeRange(), "Should only be called if currently in a free range.");
8506 assert(size > 0,
8507 "A zero sized chunk cannot be added to the free lists.");
8508 if (!freeRangeInFreeLists()) {
8509 if (CMSTestInFreeList) {
8510 FreeChunk* fc = (FreeChunk*) chunk;
8511 fc->set_size(size);
8512 assert(!_sp->verify_chunk_in_free_list(fc),
8513 "chunk should not be in free lists yet");
8514 }
8515 if (CMSTraceSweeper) {
8516 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8517 chunk, size);
8518 }
8519 // A new free range is going to be starting. The current
8520 // free range has not been added to the free lists yet or
8521 // was removed so add it back.
8522 // If the current free range was coalesced, then the death
8523 // of the free range was recorded. Record a birth now.
8524 if (lastFreeRangeCoalesced()) {
8525 _sp->coalBirth(size);
8526 }
8527 _sp->addChunkAndRepairOffsetTable(chunk, size,
8528 lastFreeRangeCoalesced());
8529 } else if (CMSTraceSweeper) {
8530 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8531 }
8532 set_inFreeRange(false);
8533 set_freeRangeInFreeLists(false);
8534 }
8535
8536 // We take a break if we've been at this for a while,
8537 // so as to avoid monopolizing the locks involved.
8538 void SweepClosure::do_yield_work(HeapWord* addr) {
8539 // Return current free chunk being used for coalescing (if any)
8540 // to the appropriate freelist. After yielding, the next
8541 // free block encountered will start a coalescing range of
8542 // free blocks. If the next free block is adjacent to the
8543 // chunk just flushed, they will need to wait for the next
8544 // sweep to be coalesced.
8545 if (inFreeRange()) {
8546 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8547 }
8548
8549 // First give up the locks, then yield, then re-lock.
8550 // We should probably use a constructor/destructor idiom to
8551 // do this unlock/lock or modify the MutexUnlocker class to
8552 // serve our purpose. XXX
8553 assert_lock_strong(_bitMap->lock());
8554 assert_lock_strong(_freelistLock);
8555 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8556 "CMS thread should hold CMS token");
8557 _bitMap->lock()->unlock();
8558 _freelistLock->unlock();
8559 ConcurrentMarkSweepThread::desynchronize(true);
8560 ConcurrentMarkSweepThread::acknowledge_yield_request();
8561 _collector->stopTimer();
8562 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8563 if (PrintCMSStatistics != 0) {
8564 _collector->incrementYields();
8565 }
8566 _collector->icms_wait();
8567
8568 // See the comment in coordinator_yield()
8569 for (unsigned i = 0; i < CMSYieldSleepCount &&
8570 ConcurrentMarkSweepThread::should_yield() &&
8571 !CMSCollector::foregroundGCIsActive(); ++i) {
8572 os::sleep(Thread::current(), 1, false);
8573 ConcurrentMarkSweepThread::acknowledge_yield_request();
8574 }
8575
8576 ConcurrentMarkSweepThread::synchronize(true);
8577 _freelistLock->lock();
8578 _bitMap->lock()->lock_without_safepoint_check();
8579 _collector->startTimer();
8580 }
8581
8582 #ifndef PRODUCT
8583 // This is actually very useful in a product build if it can
8584 // be called from the debugger. Compile it into the product
8585 // as needed.
8586 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8587 return debug_cms_space->verify_chunk_in_free_list(fc);
8588 }
8589 #endif
8590
8591 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8592 if (CMSTraceSweeper) {
8593 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8594 fc, fc->size());
8595 }
8596 }
8597
8598 // CMSIsAliveClosure
8599 bool CMSIsAliveClosure::do_object_b(oop obj) {
8600 HeapWord* addr = (HeapWord*)obj;
8601 return addr != NULL &&
8602 (!_span.contains(addr) || _bit_map->isMarked(addr));
8603 }
8604
8605 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8606 MemRegion span,
8607 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8608 CMSMarkStack* revisit_stack, bool cpc):
8609 KlassRememberingOopClosure(collector, NULL, revisit_stack),
8610 _span(span),
8611 _bit_map(bit_map),
8612 _mark_stack(mark_stack),
8613 _concurrent_precleaning(cpc) {
8614 assert(!_span.is_empty(), "Empty span could spell trouble");
8615 }
8616
8617
8618 // CMSKeepAliveClosure: the serial version
8619 void CMSKeepAliveClosure::do_oop(oop obj) {
8620 HeapWord* addr = (HeapWord*)obj;
8621 if (_span.contains(addr) &&
8622 !_bit_map->isMarked(addr)) {
8623 _bit_map->mark(addr);
8624 bool simulate_overflow = false;
8625 NOT_PRODUCT(
8626 if (CMSMarkStackOverflowALot &&
8627 _collector->simulate_overflow()) {
8628 // simulate a stack overflow
8629 simulate_overflow = true;
8630 }
8631 )
8632 if (simulate_overflow || !_mark_stack->push(obj)) {
8633 if (_concurrent_precleaning) {
8634 // We dirty the overflown object and let the remark
8635 // phase deal with it.
8636 assert(_collector->overflow_list_is_empty(), "Error");
8637 // In the case of object arrays, we need to dirty all of
8638 // the cards that the object spans. No locking or atomics
8639 // are needed since no one else can be mutating the mod union
8640 // table.
8641 if (obj->is_objArray()) {
8642 size_t sz = obj->size();
8643 HeapWord* end_card_addr =
8644 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8645 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8646 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8647 _collector->_modUnionTable.mark_range(redirty_range);
8648 } else {
8649 _collector->_modUnionTable.mark(addr);
8650 }
8651 _collector->_ser_kac_preclean_ovflw++;
8652 } else {
8653 _collector->push_on_overflow_list(obj);
8654 _collector->_ser_kac_ovflw++;
8655 }
8656 }
8657 }
8658 }
8659
8660 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8661 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8662
8663 // CMSParKeepAliveClosure: a parallel version of the above.
8664 // The work queues are private to each closure (thread),
8665 // but (may be) available for stealing by other threads.
8666 void CMSParKeepAliveClosure::do_oop(oop obj) {
8667 HeapWord* addr = (HeapWord*)obj;
8668 if (_span.contains(addr) &&
8669 !_bit_map->isMarked(addr)) {
8670 // In general, during recursive tracing, several threads
8671 // may be concurrently getting here; the first one to
8672 // "tag" it, claims it.
8673 if (_bit_map->par_mark(addr)) {
8674 bool res = _work_queue->push(obj);
8675 assert(res, "Low water mark should be much less than capacity");
8676 // Do a recursive trim in the hope that this will keep
8677 // stack usage lower, but leave some oops for potential stealers
8678 trim_queue(_low_water_mark);
8679 } // Else, another thread got there first
8680 }
8681 }
8682
8683 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8684 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8685
8686 void CMSParKeepAliveClosure::trim_queue(uint max) {
8687 while (_work_queue->size() > max) {
8688 oop new_oop;
8689 if (_work_queue->pop_local(new_oop)) {
8690 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8691 assert(_bit_map->isMarked((HeapWord*)new_oop),
8692 "no white objects on this stack!");
8693 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8694 // iterate over the oops in this oop, marking and pushing
8695 // the ones in CMS heap (i.e. in _span).
8696 new_oop->oop_iterate(&_mark_and_push);
8697 }
8698 }
8699 }
8700
8701 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8702 CMSCollector* collector,
8703 MemRegion span, CMSBitMap* bit_map,
8704 CMSMarkStack* revisit_stack,
8705 OopTaskQueue* work_queue):
8706 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
8707 _span(span),
8708 _bit_map(bit_map),
8709 _work_queue(work_queue) { }
8710
8711 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8712 HeapWord* addr = (HeapWord*)obj;
8713 if (_span.contains(addr) &&
8714 !_bit_map->isMarked(addr)) {
8715 if (_bit_map->par_mark(addr)) {
8716 bool simulate_overflow = false;
8717 NOT_PRODUCT(
8718 if (CMSMarkStackOverflowALot &&
8719 _collector->par_simulate_overflow()) {
8720 // simulate a stack overflow
8721 simulate_overflow = true;
8722 }
8723 )
8724 if (simulate_overflow || !_work_queue->push(obj)) {
8725 _collector->par_push_on_overflow_list(obj);
8726 _collector->_par_kac_ovflw++;
8727 }
8728 } // Else another thread got there already
8729 }
8730 }
8731
8732 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8733 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8734
8735 //////////////////////////////////////////////////////////////////
8736 // CMSExpansionCause /////////////////////////////
8737 //////////////////////////////////////////////////////////////////
8738 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8739 switch (cause) {
8740 case _no_expansion:
8741 return "No expansion";
8742 case _satisfy_free_ratio:
8743 return "Free ratio";
8744 case _satisfy_promotion:
8745 return "Satisfy promotion";
8746 case _satisfy_allocation:
8747 return "allocation";
8748 case _allocate_par_lab:
8749 return "Par LAB";
8750 case _allocate_par_spooling_space:
8751 return "Par Spooling Space";
8752 case _adaptive_size_policy:
8753 return "Ergonomics";
8754 default:
8755 return "unknown";
8756 }
8757 }
8758
8759 void CMSDrainMarkingStackClosure::do_void() {
8760 // the max number to take from overflow list at a time
8761 const size_t num = _mark_stack->capacity()/4;
8762 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8763 "Overflow list should be NULL during concurrent phases");
8764 while (!_mark_stack->isEmpty() ||
8765 // if stack is empty, check the overflow list
8766 _collector->take_from_overflow_list(num, _mark_stack)) {
8767 oop obj = _mark_stack->pop();
8768 HeapWord* addr = (HeapWord*)obj;
8769 assert(_span.contains(addr), "Should be within span");
8770 assert(_bit_map->isMarked(addr), "Should be marked");
8771 assert(obj->is_oop(), "Should be an oop");
8772 obj->oop_iterate(_keep_alive);
8773 }
8774 }
8775
8776 void CMSParDrainMarkingStackClosure::do_void() {
8777 // drain queue
8778 trim_queue(0);
8779 }
8780
8781 // Trim our work_queue so its length is below max at return
8782 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8783 while (_work_queue->size() > max) {
8784 oop new_oop;
8785 if (_work_queue->pop_local(new_oop)) {
8786 assert(new_oop->is_oop(), "Expected an oop");
8787 assert(_bit_map->isMarked((HeapWord*)new_oop),
8788 "no white objects on this stack!");
8789 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8790 // iterate over the oops in this oop, marking and pushing
8791 // the ones in CMS heap (i.e. in _span).
8792 new_oop->oop_iterate(&_mark_and_push);
8793 }
8794 }
8795 }
8796
8797 ////////////////////////////////////////////////////////////////////
8798 // Support for Marking Stack Overflow list handling and related code
8799 ////////////////////////////////////////////////////////////////////
8800 // Much of the following code is similar in shape and spirit to the
8801 // code used in ParNewGC. We should try and share that code
8802 // as much as possible in the future.
8803
8804 #ifndef PRODUCT
8805 // Debugging support for CMSStackOverflowALot
8806
8807 // It's OK to call this multi-threaded; the worst thing
8808 // that can happen is that we'll get a bunch of closely
8809 // spaced simulated oveflows, but that's OK, in fact
8810 // probably good as it would exercise the overflow code
8811 // under contention.
8812 bool CMSCollector::simulate_overflow() {
8813 if (_overflow_counter-- <= 0) { // just being defensive
8814 _overflow_counter = CMSMarkStackOverflowInterval;
8815 return true;
8816 } else {
8817 return false;
8818 }
8819 }
8820
8821 bool CMSCollector::par_simulate_overflow() {
8822 return simulate_overflow();
8823 }
8824 #endif
8825
8826 // Single-threaded
8827 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8828 assert(stack->isEmpty(), "Expected precondition");
8829 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8830 size_t i = num;
8831 oop cur = _overflow_list;
8832 const markOop proto = markOopDesc::prototype();
8833 NOT_PRODUCT(ssize_t n = 0;)
8834 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8835 next = oop(cur->mark());
8836 cur->set_mark(proto); // until proven otherwise
8837 assert(cur->is_oop(), "Should be an oop");
8838 bool res = stack->push(cur);
8839 assert(res, "Bit off more than can chew?");
8840 NOT_PRODUCT(n++;)
8841 }
8842 _overflow_list = cur;
8843 #ifndef PRODUCT
8844 assert(_num_par_pushes >= n, "Too many pops?");
8845 _num_par_pushes -=n;
8846 #endif
8847 return !stack->isEmpty();
8848 }
8849
8850 #define BUSY (oop(0x1aff1aff))
8851 // (MT-safe) Get a prefix of at most "num" from the list.
8852 // The overflow list is chained through the mark word of
8853 // each object in the list. We fetch the entire list,
8854 // break off a prefix of the right size and return the
8855 // remainder. If other threads try to take objects from
8856 // the overflow list at that time, they will wait for
8857 // some time to see if data becomes available. If (and
8858 // only if) another thread places one or more object(s)
8859 // on the global list before we have returned the suffix
8860 // to the global list, we will walk down our local list
8861 // to find its end and append the global list to
8862 // our suffix before returning it. This suffix walk can
8863 // prove to be expensive (quadratic in the amount of traffic)
8864 // when there are many objects in the overflow list and
8865 // there is much producer-consumer contention on the list.
8866 // *NOTE*: The overflow list manipulation code here and
8867 // in ParNewGeneration:: are very similar in shape,
8868 // except that in the ParNew case we use the old (from/eden)
8869 // copy of the object to thread the list via its klass word.
8870 // Because of the common code, if you make any changes in
8871 // the code below, please check the ParNew version to see if
8872 // similar changes might be needed.
8873 // CR 6797058 has been filed to consolidate the common code.
8874 bool CMSCollector::par_take_from_overflow_list(size_t num,
8875 OopTaskQueue* work_q,
8876 int no_of_gc_threads) {
8877 assert(work_q->size() == 0, "First empty local work queue");
8878 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8879 if (_overflow_list == NULL) {
8880 return false;
8881 }
8882 // Grab the entire list; we'll put back a suffix
8883 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8884 Thread* tid = Thread::current();
8885 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8886 // set to ParallelGCThreads.
8887 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8888 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8889 // If the list is busy, we spin for a short while,
8890 // sleeping between attempts to get the list.
8891 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8892 os::sleep(tid, sleep_time_millis, false);
8893 if (_overflow_list == NULL) {
8894 // Nothing left to take
8895 return false;
8896 } else if (_overflow_list != BUSY) {
8897 // Try and grab the prefix
8898 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8899 }
8900 }
8901 // If the list was found to be empty, or we spun long
8902 // enough, we give up and return empty-handed. If we leave
8903 // the list in the BUSY state below, it must be the case that
8904 // some other thread holds the overflow list and will set it
8905 // to a non-BUSY state in the future.
8906 if (prefix == NULL || prefix == BUSY) {
8907 // Nothing to take or waited long enough
8908 if (prefix == NULL) {
8909 // Write back the NULL in case we overwrote it with BUSY above
8910 // and it is still the same value.
8911 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8912 }
8913 return false;
8914 }
8915 assert(prefix != NULL && prefix != BUSY, "Error");
8916 size_t i = num;
8917 oop cur = prefix;
8918 // Walk down the first "num" objects, unless we reach the end.
8919 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8920 if (cur->mark() == NULL) {
8921 // We have "num" or fewer elements in the list, so there
8922 // is nothing to return to the global list.
8923 // Write back the NULL in lieu of the BUSY we wrote
8924 // above, if it is still the same value.
8925 if (_overflow_list == BUSY) {
8926 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8927 }
8928 } else {
8929 // Chop off the suffix and rerturn it to the global list.
8930 assert(cur->mark() != BUSY, "Error");
8931 oop suffix_head = cur->mark(); // suffix will be put back on global list
8932 cur->set_mark(NULL); // break off suffix
8933 // It's possible that the list is still in the empty(busy) state
8934 // we left it in a short while ago; in that case we may be
8935 // able to place back the suffix without incurring the cost
8936 // of a walk down the list.
8937 oop observed_overflow_list = _overflow_list;
8938 oop cur_overflow_list = observed_overflow_list;
8939 bool attached = false;
8940 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8941 observed_overflow_list =
8942 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8943 if (cur_overflow_list == observed_overflow_list) {
8944 attached = true;
8945 break;
8946 } else cur_overflow_list = observed_overflow_list;
8947 }
8948 if (!attached) {
8949 // Too bad, someone else sneaked in (at least) an element; we'll need
8950 // to do a splice. Find tail of suffix so we can prepend suffix to global
8951 // list.
8952 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8953 oop suffix_tail = cur;
8954 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8955 "Tautology");
8956 observed_overflow_list = _overflow_list;
8957 do {
8958 cur_overflow_list = observed_overflow_list;
8959 if (cur_overflow_list != BUSY) {
8960 // Do the splice ...
8961 suffix_tail->set_mark(markOop(cur_overflow_list));
8962 } else { // cur_overflow_list == BUSY
8963 suffix_tail->set_mark(NULL);
8964 }
8965 // ... and try to place spliced list back on overflow_list ...
8966 observed_overflow_list =
8967 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8968 } while (cur_overflow_list != observed_overflow_list);
8969 // ... until we have succeeded in doing so.
8970 }
8971 }
8972
8973 // Push the prefix elements on work_q
8974 assert(prefix != NULL, "control point invariant");
8975 const markOop proto = markOopDesc::prototype();
8976 oop next;
8977 NOT_PRODUCT(ssize_t n = 0;)
8978 for (cur = prefix; cur != NULL; cur = next) {
8979 next = oop(cur->mark());
8980 cur->set_mark(proto); // until proven otherwise
8981 assert(cur->is_oop(), "Should be an oop");
8982 bool res = work_q->push(cur);
8983 assert(res, "Bit off more than we can chew?");
8984 NOT_PRODUCT(n++;)
8985 }
8986 #ifndef PRODUCT
8987 assert(_num_par_pushes >= n, "Too many pops?");
8988 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
8989 #endif
8990 return true;
8991 }
8992
8993 // Single-threaded
8994 void CMSCollector::push_on_overflow_list(oop p) {
8995 NOT_PRODUCT(_num_par_pushes++;)
8996 assert(p->is_oop(), "Not an oop");
8997 preserve_mark_if_necessary(p);
8998 p->set_mark((markOop)_overflow_list);
8999 _overflow_list = p;
9000 }
9001
9002 // Multi-threaded; use CAS to prepend to overflow list
9003 void CMSCollector::par_push_on_overflow_list(oop p) {
9004 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9005 assert(p->is_oop(), "Not an oop");
9006 par_preserve_mark_if_necessary(p);
9007 oop observed_overflow_list = _overflow_list;
9008 oop cur_overflow_list;
9009 do {
9010 cur_overflow_list = observed_overflow_list;
9011 if (cur_overflow_list != BUSY) {
9012 p->set_mark(markOop(cur_overflow_list));
9013 } else {
9014 p->set_mark(NULL);
9015 }
9016 observed_overflow_list =
9017 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9018 } while (cur_overflow_list != observed_overflow_list);
9019 }
9020 #undef BUSY
9021
9022 // Single threaded
9023 // General Note on GrowableArray: pushes may silently fail
9024 // because we are (temporarily) out of C-heap for expanding
9025 // the stack. The problem is quite ubiquitous and affects
9026 // a lot of code in the JVM. The prudent thing for GrowableArray
9027 // to do (for now) is to exit with an error. However, that may
9028 // be too draconian in some cases because the caller may be
9029 // able to recover without much harm. For such cases, we
9030 // should probably introduce a "soft_push" method which returns
9031 // an indication of success or failure with the assumption that
9032 // the caller may be able to recover from a failure; code in
9033 // the VM can then be changed, incrementally, to deal with such
9034 // failures where possible, thus, incrementally hardening the VM
9035 // in such low resource situations.
9036 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9037 _preserved_oop_stack.push(p);
9038 _preserved_mark_stack.push(m);
9039 assert(m == p->mark(), "Mark word changed");
9040 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9041 "bijection");
9042 }
9043
9044 // Single threaded
9045 void CMSCollector::preserve_mark_if_necessary(oop p) {
9046 markOop m = p->mark();
9047 if (m->must_be_preserved(p)) {
9048 preserve_mark_work(p, m);
9049 }
9050 }
9051
9052 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9053 markOop m = p->mark();
9054 if (m->must_be_preserved(p)) {
9055 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9056 // Even though we read the mark word without holding
9057 // the lock, we are assured that it will not change
9058 // because we "own" this oop, so no other thread can
9059 // be trying to push it on the overflow list; see
9060 // the assertion in preserve_mark_work() that checks
9061 // that m == p->mark().
9062 preserve_mark_work(p, m);
9063 }
9064 }
9065
9066 // We should be able to do this multi-threaded,
9067 // a chunk of stack being a task (this is
9068 // correct because each oop only ever appears
9069 // once in the overflow list. However, it's
9070 // not very easy to completely overlap this with
9071 // other operations, so will generally not be done
9072 // until all work's been completed. Because we
9073 // expect the preserved oop stack (set) to be small,
9074 // it's probably fine to do this single-threaded.
9075 // We can explore cleverer concurrent/overlapped/parallel
9076 // processing of preserved marks if we feel the
9077 // need for this in the future. Stack overflow should
9078 // be so rare in practice and, when it happens, its
9079 // effect on performance so great that this will
9080 // likely just be in the noise anyway.
9081 void CMSCollector::restore_preserved_marks_if_any() {
9082 assert(SafepointSynchronize::is_at_safepoint(),
9083 "world should be stopped");
9084 assert(Thread::current()->is_ConcurrentGC_thread() ||
9085 Thread::current()->is_VM_thread(),
9086 "should be single-threaded");
9087 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9088 "bijection");
9089
9090 while (!_preserved_oop_stack.is_empty()) {
9091 oop p = _preserved_oop_stack.pop();
9092 assert(p->is_oop(), "Should be an oop");
9093 assert(_span.contains(p), "oop should be in _span");
9094 assert(p->mark() == markOopDesc::prototype(),
9095 "Set when taken from overflow list");
9096 markOop m = _preserved_mark_stack.pop();
9097 p->set_mark(m);
9098 }
9099 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9100 "stacks were cleared above");
9101 }
9102
9103 #ifndef PRODUCT
9104 bool CMSCollector::no_preserved_marks() const {
9105 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9106 }
9107 #endif
9108
9109 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9110 {
9111 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9112 CMSAdaptiveSizePolicy* size_policy =
9113 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9114 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9115 "Wrong type for size policy");
9116 return size_policy;
9117 }
9118
9119 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9120 size_t desired_promo_size) {
9121 if (cur_promo_size < desired_promo_size) {
9122 size_t expand_bytes = desired_promo_size - cur_promo_size;
9123 if (PrintAdaptiveSizePolicy && Verbose) {
9124 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9125 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9126 expand_bytes);
9127 }
9128 expand(expand_bytes,
9129 MinHeapDeltaBytes,
9130 CMSExpansionCause::_adaptive_size_policy);
9131 } else if (desired_promo_size < cur_promo_size) {
9132 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9133 if (PrintAdaptiveSizePolicy && Verbose) {
9134 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9135 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9136 shrink_bytes);
9137 }
9138 shrink(shrink_bytes);
9139 }
9140 }
9141
9142 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9143 GenCollectedHeap* gch = GenCollectedHeap::heap();
9144 CMSGCAdaptivePolicyCounters* counters =
9145 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9146 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9147 "Wrong kind of counters");
9148 return counters;
9149 }
9150
9151
9152 void ASConcurrentMarkSweepGeneration::update_counters() {
9153 if (UsePerfData) {
9154 _space_counters->update_all();
9155 _gen_counters->update_all();
9156 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9157 GenCollectedHeap* gch = GenCollectedHeap::heap();
9158 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9159 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9160 "Wrong gc statistics type");
9161 counters->update_counters(gc_stats_l);
9162 }
9163 }
9164
9165 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9166 if (UsePerfData) {
9167 _space_counters->update_used(used);
9168 _space_counters->update_capacity();
9169 _gen_counters->update_all();
9170
9171 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9172 GenCollectedHeap* gch = GenCollectedHeap::heap();
9173 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9174 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9175 "Wrong gc statistics type");
9176 counters->update_counters(gc_stats_l);
9177 }
9178 }
9179
9180 // The desired expansion delta is computed so that:
9181 // . desired free percentage or greater is used
9182 void ASConcurrentMarkSweepGeneration::compute_new_size() {
9183 assert_locked_or_safepoint(Heap_lock);
9184
9185 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9186
9187 // If incremental collection failed, we just want to expand
9188 // to the limit.
9189 if (incremental_collection_failed()) {
9190 clear_incremental_collection_failed();
9191 grow_to_reserved();
9192 return;
9193 }
9194
9195 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
9196
9197 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
9198 "Wrong type of heap");
9199 int prev_level = level() - 1;
9200 assert(prev_level >= 0, "The cms generation is the lowest generation");
9201 Generation* prev_gen = gch->get_gen(prev_level);
9202 assert(prev_gen->kind() == Generation::ASParNew,
9203 "Wrong type of young generation");
9204 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
9205 size_t cur_eden = younger_gen->eden()->capacity();
9206 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
9207 size_t cur_promo = free();
9208 size_policy->compute_tenured_generation_free_space(cur_promo,
9209 max_available(),
9210 cur_eden);
9211 resize(cur_promo, size_policy->promo_size());
9212
9213 // Record the new size of the space in the cms generation
9214 // that is available for promotions. This is temporary.
9215 // It should be the desired promo size.
9216 size_policy->avg_cms_promo()->sample(free());
9217 size_policy->avg_old_live()->sample(used());
9218
9219 if (UsePerfData) {
9220 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9221 counters->update_cms_capacity_counter(capacity());
9222 }
9223 }
9224
9225 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9226 assert_locked_or_safepoint(Heap_lock);
9227 assert_lock_strong(freelistLock());
9228 HeapWord* old_end = _cmsSpace->end();
9229 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9230 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9231 FreeChunk* chunk_at_end = find_chunk_at_end();
9232 if (chunk_at_end == NULL) {
9233 // No room to shrink
9234 if (PrintGCDetails && Verbose) {
9235 gclog_or_tty->print_cr("No room to shrink: old_end "
9236 PTR_FORMAT " unallocated_start " PTR_FORMAT
9237 " chunk_at_end " PTR_FORMAT,
9238 old_end, unallocated_start, chunk_at_end);
9239 }
9240 return;
9241 } else {
9242
9243 // Find the chunk at the end of the space and determine
9244 // how much it can be shrunk.
9245 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9246 size_t aligned_shrinkable_size_in_bytes =
9247 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9248 assert(unallocated_start <= chunk_at_end->end(),
9249 "Inconsistent chunk at end of space");
9250 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9251 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9252
9253 // Shrink the underlying space
9254 _virtual_space.shrink_by(bytes);
9255 if (PrintGCDetails && Verbose) {
9256 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9257 " desired_bytes " SIZE_FORMAT
9258 " shrinkable_size_in_bytes " SIZE_FORMAT
9259 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9260 " bytes " SIZE_FORMAT,
9261 desired_bytes, shrinkable_size_in_bytes,
9262 aligned_shrinkable_size_in_bytes, bytes);
9263 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9264 " unallocated_start " SIZE_FORMAT,
9265 old_end, unallocated_start);
9266 }
9267
9268 // If the space did shrink (shrinking is not guaranteed),
9269 // shrink the chunk at the end by the appropriate amount.
9270 if (((HeapWord*)_virtual_space.high()) < old_end) {
9271 size_t new_word_size =
9272 heap_word_size(_virtual_space.committed_size());
9273
9274 // Have to remove the chunk from the dictionary because it is changing
9275 // size and might be someplace elsewhere in the dictionary.
9276
9277 // Get the chunk at end, shrink it, and put it
9278 // back.
9279 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9280 size_t word_size_change = word_size_before - new_word_size;
9281 size_t chunk_at_end_old_size = chunk_at_end->size();
9282 assert(chunk_at_end_old_size >= word_size_change,
9283 "Shrink is too large");
9284 chunk_at_end->set_size(chunk_at_end_old_size -
9285 word_size_change);
9286 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9287 word_size_change);
9288
9289 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9290
9291 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9292 _bts->resize(new_word_size); // resize the block offset shared array
9293 Universe::heap()->barrier_set()->resize_covered_region(mr);
9294 _cmsSpace->assert_locked();
9295 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9296
9297 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9298
9299 // update the space and generation capacity counters
9300 if (UsePerfData) {
9301 _space_counters->update_capacity();
9302 _gen_counters->update_all();
9303 }
9304
9305 if (Verbose && PrintGCDetails) {
9306 size_t new_mem_size = _virtual_space.committed_size();
9307 size_t old_mem_size = new_mem_size + bytes;
9308 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
9309 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9310 }
9311 }
9312
9313 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9314 "Inconsistency at end of space");
9315 assert(chunk_at_end->end() == _cmsSpace->end(),
9316 "Shrinking is inconsistent");
9317 return;
9318 }
9319 }
9320
9321 // Transfer some number of overflown objects to usual marking
9322 // stack. Return true if some objects were transferred.
9323 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9324 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9325 (size_t)ParGCDesiredObjsFromOverflowList);
9326
9327 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9328 assert(_collector->overflow_list_is_empty() || res,
9329 "If list is not empty, we should have taken something");
9330 assert(!res || !_mark_stack->isEmpty(),
9331 "If we took something, it should now be on our stack");
9332 return res;
9333 }
9334
9335 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9336 size_t res = _sp->block_size_no_stall(addr, _collector);
9337 if (_sp->block_is_obj(addr)) {
9338 if (_live_bit_map->isMarked(addr)) {
9339 // It can't have been dead in a previous cycle
9340 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9341 } else {
9342 _dead_bit_map->mark(addr); // mark the dead object
9343 }
9344 }
9345 // Could be 0, if the block size could not be computed without stalling.
9346 return res;
9347 }
9348
9349 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9350
9351 switch (phase) {
9352 case CMSCollector::InitialMarking:
9353 initialize(true /* fullGC */ ,
9354 cause /* cause of the GC */,
9355 true /* recordGCBeginTime */,
9356 true /* recordPreGCUsage */,
9357 false /* recordPeakUsage */,
9358 false /* recordPostGCusage */,
9359 true /* recordAccumulatedGCTime */,
9360 false /* recordGCEndTime */,
9361 false /* countCollection */ );
9362 break;
9363
9364 case CMSCollector::FinalMarking:
9365 initialize(true /* fullGC */ ,
9366 cause /* cause of the GC */,
9367 false /* recordGCBeginTime */,
9368 false /* recordPreGCUsage */,
9369 false /* recordPeakUsage */,
9370 false /* recordPostGCusage */,
9371 true /* recordAccumulatedGCTime */,
9372 false /* recordGCEndTime */,
9373 false /* countCollection */ );
9374 break;
9375
9376 case CMSCollector::Sweeping:
9377 initialize(true /* fullGC */ ,
9378 cause /* cause of the GC */,
9379 false /* recordGCBeginTime */,
9380 false /* recordPreGCUsage */,
9381 true /* recordPeakUsage */,
9382 true /* recordPostGCusage */,
9383 false /* recordAccumulatedGCTime */,
9384 true /* recordGCEndTime */,
9385 true /* countCollection */ );
9386 break;
9387
9388 default:
9389 ShouldNotReachHere();
9390 }
9391 }
9392