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 bool CMSCollector::is_external_interruption() {
1699 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1700 return GCCause::is_user_requested_gc(cause) ||
1701 GCCause::is_serviceability_requested_gc(cause);
1702 }
1703
1704 void CMSCollector::report_concurrent_mode_interruption() {
1705 if (is_external_interruption()) {
1706 if (PrintGCDetails) {
1707 gclog_or_tty->print(" (concurrent mode interrupted)");
1708 }
1709 } else {
1710 if (PrintGCDetails) {
1711 gclog_or_tty->print(" (concurrent mode failure)");
1712 }
1713 _gc_tracer_cm->report_concurrent_mode_failure();
1714 }
1715 }
1716
1717 // The foreground and background collectors need to coordinate in order
1718 // to make sure that they do not mutually interfere with CMS collections.
1719 // When a background collection is active,
1720 // the foreground collector may need to take over (preempt) and
1721 // synchronously complete an ongoing collection. Depending on the
1722 // frequency of the background collections and the heap usage
1723 // of the application, this preemption can be seldom or frequent.
1724 // There are only certain
1725 // points in the background collection that the "collection-baton"
1726 // can be passed to the foreground collector.
1727 //
1728 // The foreground collector will wait for the baton before
1729 // starting any part of the collection. The foreground collector
1730 // will only wait at one location.
1731 //
1732 // The background collector will yield the baton before starting a new
1733 // phase of the collection (e.g., before initial marking, marking from roots,
1734 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1735 // of the loop which switches the phases. The background collector does some
1736 // of the phases (initial mark, final re-mark) with the world stopped.
1737 // Because of locking involved in stopping the world,
1738 // the foreground collector should not block waiting for the background
1739 // collector when it is doing a stop-the-world phase. The background
1740 // collector will yield the baton at an additional point just before
1741 // it enters a stop-the-world phase. Once the world is stopped, the
1742 // background collector checks the phase of the collection. If the
1743 // phase has not changed, it proceeds with the collection. If the
1744 // phase has changed, it skips that phase of the collection. See
1745 // the comments on the use of the Heap_lock in collect_in_background().
1746 //
1747 // Variable used in baton passing.
1748 // _foregroundGCIsActive - Set to true by the foreground collector when
1749 // it wants the baton. The foreground clears it when it has finished
1750 // the collection.
1751 // _foregroundGCShouldWait - Set to true by the background collector
1752 // when it is running. The foreground collector waits while
1753 // _foregroundGCShouldWait is true.
1754 // CGC_lock - monitor used to protect access to the above variables
1755 // and to notify the foreground and background collectors.
1756 // _collectorState - current state of the CMS collection.
1757 //
1758 // The foreground collector
1759 // acquires the CGC_lock
1760 // sets _foregroundGCIsActive
1761 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1762 // various locks acquired in preparation for the collection
1763 // are released so as not to block the background collector
1764 // that is in the midst of a collection
1765 // proceeds with the collection
1766 // clears _foregroundGCIsActive
1767 // returns
1768 //
1769 // The background collector in a loop iterating on the phases of the
1770 // collection
1771 // acquires the CGC_lock
1772 // sets _foregroundGCShouldWait
1773 // if _foregroundGCIsActive is set
1774 // clears _foregroundGCShouldWait, notifies _CGC_lock
1775 // waits on _CGC_lock for _foregroundGCIsActive to become false
1776 // and exits the loop.
1777 // otherwise
1778 // proceed with that phase of the collection
1779 // if the phase is a stop-the-world phase,
1780 // yield the baton once more just before enqueueing
1781 // the stop-world CMS operation (executed by the VM thread).
1782 // returns after all phases of the collection are done
1783 //
1784
1785 void CMSCollector::acquire_control_and_collect(bool full,
1786 bool clear_all_soft_refs) {
1787 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1788 assert(!Thread::current()->is_ConcurrentGC_thread(),
1789 "shouldn't try to acquire control from self!");
1790
1791 // Start the protocol for acquiring control of the
1792 // collection from the background collector (aka CMS thread).
1793 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1794 "VM thread should have CMS token");
1795 // Remember the possibly interrupted state of an ongoing
1796 // concurrent collection
1797 CollectorState first_state = _collectorState;
1798
1799 // Signal to a possibly ongoing concurrent collection that
1800 // we want to do a foreground collection.
1801 _foregroundGCIsActive = true;
1802
1803 // Disable incremental mode during a foreground collection.
1804 ICMSDisabler icms_disabler;
1805
1806 // release locks and wait for a notify from the background collector
1807 // releasing the locks in only necessary for phases which
1808 // do yields to improve the granularity of the collection.
1809 assert_lock_strong(bitMapLock());
1810 // We need to lock the Free list lock for the space that we are
1811 // currently collecting.
1812 assert(haveFreelistLocks(), "Must be holding free list locks");
1813 bitMapLock()->unlock();
1814 releaseFreelistLocks();
1815 {
1816 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1817 if (_foregroundGCShouldWait) {
1818 // We are going to be waiting for action for the CMS thread;
1819 // it had better not be gone (for instance at shutdown)!
1820 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1821 "CMS thread must be running");
1822 // Wait here until the background collector gives us the go-ahead
1823 ConcurrentMarkSweepThread::clear_CMS_flag(
1824 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1825 // Get a possibly blocked CMS thread going:
1826 // Note that we set _foregroundGCIsActive true above,
1827 // without protection of the CGC_lock.
1828 CGC_lock->notify();
1829 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1830 "Possible deadlock");
1831 while (_foregroundGCShouldWait) {
1832 // wait for notification
1833 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1834 // Possibility of delay/starvation here, since CMS token does
1835 // not know to give priority to VM thread? Actually, i think
1836 // there wouldn't be any delay/starvation, but the proof of
1837 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1838 }
1839 ConcurrentMarkSweepThread::set_CMS_flag(
1840 ConcurrentMarkSweepThread::CMS_vm_has_token);
1841 }
1842 }
1843 // The CMS_token is already held. Get back the other locks.
1844 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1845 "VM thread should have CMS token");
1846 getFreelistLocks();
1847 bitMapLock()->lock_without_safepoint_check();
1848 if (TraceCMSState) {
1849 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1850 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1851 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1852 }
1853
1854 // Check if we need to do a compaction, or if not, whether
1855 // we need to start the mark-sweep from scratch.
1856 bool should_compact = false;
1857 bool should_start_over = false;
1858 decide_foreground_collection_type(clear_all_soft_refs,
1859 &should_compact, &should_start_over);
1860
1861 NOT_PRODUCT(
1862 if (RotateCMSCollectionTypes) {
1863 if (_cmsGen->debug_collection_type() ==
1864 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1865 should_compact = true;
1866 } else if (_cmsGen->debug_collection_type() ==
1867 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1868 should_compact = false;
1869 }
1870 }
1871 )
1872
1873 if (first_state > Idling) {
1874 report_concurrent_mode_interruption();
1875 }
1876
1877 if (should_compact) {
1878 // If the collection is being acquired from the background
1879 // collector, there may be references on the discovered
1880 // references lists that have NULL referents (being those
1881 // that were concurrently cleared by a mutator) or
1882 // that are no longer active (having been enqueued concurrently
1883 // by the mutator).
1884 // Scrub the list of those references because Mark-Sweep-Compact
1885 // code assumes referents are not NULL and that all discovered
1886 // Reference objects are active.
1887 ref_processor()->clean_up_discovered_references();
1888
1889 do_compaction_work(clear_all_soft_refs);
1890
1891 // Has the GC time limit been exceeded?
1892 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1893 size_t max_eden_size = young_gen->max_capacity() -
1894 young_gen->to()->capacity() -
1895 young_gen->from()->capacity();
1896 GenCollectedHeap* gch = GenCollectedHeap::heap();
1897 GCCause::Cause gc_cause = gch->gc_cause();
1898 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1899 young_gen->eden()->used(),
1900 _cmsGen->max_capacity(),
1901 max_eden_size,
1902 full,
1903 gc_cause,
1904 gch->collector_policy());
1905 } else {
1906 do_mark_sweep_work(clear_all_soft_refs, first_state,
1907 should_start_over);
1908 }
1909 // Reset the expansion cause, now that we just completed
1910 // a collection cycle.
1911 clear_expansion_cause();
1912 _foregroundGCIsActive = false;
1913 return;
1914 }
1915
1916 // Resize the perm generation and the tenured generation
1917 // after obtaining the free list locks for the
1918 // two generations.
1919 void CMSCollector::compute_new_size() {
1920 assert_locked_or_safepoint(Heap_lock);
1921 FreelistLocker z(this);
1922 _permGen->compute_new_size();
1923 _cmsGen->compute_new_size();
1924 }
1925
1926 // A work method used by foreground collection to determine
1927 // what type of collection (compacting or not, continuing or fresh)
1928 // it should do.
1929 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1930 // and CMSCompactWhenClearAllSoftRefs the default in the future
1931 // and do away with the flags after a suitable period.
1932 void CMSCollector::decide_foreground_collection_type(
1933 bool clear_all_soft_refs, bool* should_compact,
1934 bool* should_start_over) {
1935 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1936 // flag is set, and we have either requested a System.gc() or
1937 // the number of full gc's since the last concurrent cycle
1938 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1939 // or if an incremental collection has failed
1940 GenCollectedHeap* gch = GenCollectedHeap::heap();
1941 assert(gch->collector_policy()->is_two_generation_policy(),
1942 "You may want to check the correctness of the following");
1943 // Inform cms gen if this was due to partial collection failing.
1944 // The CMS gen may use this fact to determine its expansion policy.
1945 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1946 assert(!_cmsGen->incremental_collection_failed(),
1947 "Should have been noticed, reacted to and cleared");
1948 _cmsGen->set_incremental_collection_failed();
1949 }
1950 *should_compact =
1951 UseCMSCompactAtFullCollection &&
1952 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1953 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1954 gch->incremental_collection_will_fail(true /* consult_young */));
1955 *should_start_over = false;
1956 if (clear_all_soft_refs && !*should_compact) {
1957 // We are about to do a last ditch collection attempt
1958 // so it would normally make sense to do a compaction
1959 // to reclaim as much space as possible.
1960 if (CMSCompactWhenClearAllSoftRefs) {
1961 // Default: The rationale is that in this case either
1962 // we are past the final marking phase, in which case
1963 // we'd have to start over, or so little has been done
1964 // that there's little point in saving that work. Compaction
1965 // appears to be the sensible choice in either case.
1966 *should_compact = true;
1967 } else {
1968 // We have been asked to clear all soft refs, but not to
1969 // compact. Make sure that we aren't past the final checkpoint
1970 // phase, for that is where we process soft refs. If we are already
1971 // past that phase, we'll need to redo the refs discovery phase and
1972 // if necessary clear soft refs that weren't previously
1973 // cleared. We do so by remembering the phase in which
1974 // we came in, and if we are past the refs processing
1975 // phase, we'll choose to just redo the mark-sweep
1976 // collection from scratch.
1977 if (_collectorState > FinalMarking) {
1978 // We are past the refs processing phase;
1979 // start over and do a fresh synchronous CMS cycle
1980 _collectorState = Resetting; // skip to reset to start new cycle
1981 reset(false /* == !asynch */);
1982 *should_start_over = true;
1983 } // else we can continue a possibly ongoing current cycle
1984 }
1985 }
1986 }
1987
1988 // A work method used by the foreground collector to do
1989 // a mark-sweep-compact.
1990 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1991 GenCollectedHeap* gch = GenCollectedHeap::heap();
1992
1993 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1994 gc_timer->register_gc_start(os::elapsed_counter());
1995
1996 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1997 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
1998
1999 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
2000 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
2001 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
2002 "collections passed to foreground collector", _full_gcs_since_conc_gc);
2003 }
2004
2005 // Sample collection interval time and reset for collection pause.
2006 if (UseAdaptiveSizePolicy) {
2007 size_policy()->msc_collection_begin();
2008 }
2009
2010 // Temporarily widen the span of the weak reference processing to
2011 // the entire heap.
2012 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2013 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2014 // Temporarily, clear the "is_alive_non_header" field of the
2015 // reference processor.
2016 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2017 // Temporarily make reference _processing_ single threaded (non-MT).
2018 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2019 // Temporarily make refs discovery atomic
2020 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2021 // Temporarily make reference _discovery_ single threaded (non-MT)
2022 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2023
2024 ref_processor()->set_enqueuing_is_done(false);
2025 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2026 ref_processor()->setup_policy(clear_all_soft_refs);
2027 // If an asynchronous collection finishes, the _modUnionTable is
2028 // all clear. If we are assuming the collection from an asynchronous
2029 // collection, clear the _modUnionTable.
2030 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2031 "_modUnionTable should be clear if the baton was not passed");
2032 _modUnionTable.clear_all();
2033
2034 // We must adjust the allocation statistics being maintained
2035 // in the free list space. We do so by reading and clearing
2036 // the sweep timer and updating the block flux rate estimates below.
2037 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2038 if (_inter_sweep_timer.is_active()) {
2039 _inter_sweep_timer.stop();
2040 // Note that we do not use this sample to update the _inter_sweep_estimate.
2041 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2042 _inter_sweep_estimate.padded_average(),
2043 _intra_sweep_estimate.padded_average());
2044 }
2045
2046 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2047 ref_processor(), clear_all_soft_refs);
2048 #ifdef ASSERT
2049 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2050 size_t free_size = cms_space->free();
2051 assert(free_size ==
2052 pointer_delta(cms_space->end(), cms_space->compaction_top())
2053 * HeapWordSize,
2054 "All the free space should be compacted into one chunk at top");
2055 assert(cms_space->dictionary()->total_chunk_size(
2056 debug_only(cms_space->freelistLock())) == 0 ||
2057 cms_space->totalSizeInIndexedFreeLists() == 0,
2058 "All the free space should be in a single chunk");
2059 size_t num = cms_space->totalCount();
2060 assert((free_size == 0 && num == 0) ||
2061 (free_size > 0 && (num == 1 || num == 2)),
2062 "There should be at most 2 free chunks after compaction");
2063 #endif // ASSERT
2064 _collectorState = Resetting;
2065 assert(_restart_addr == NULL,
2066 "Should have been NULL'd before baton was passed");
2067 reset(false /* == !asynch */);
2068 _cmsGen->reset_after_compaction();
2069 _concurrent_cycles_since_last_unload = 0;
2070
2071 if (verifying() && !should_unload_classes()) {
2072 perm_gen_verify_bit_map()->clear_all();
2073 }
2074
2075 // Clear any data recorded in the PLAB chunk arrays.
2076 if (_survivor_plab_array != NULL) {
2077 reset_survivor_plab_arrays();
2078 }
2079
2080 // Adjust the per-size allocation stats for the next epoch.
2081 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2082 // Restart the "inter sweep timer" for the next epoch.
2083 _inter_sweep_timer.reset();
2084 _inter_sweep_timer.start();
2085
2086 // Sample collection pause time and reset for collection interval.
2087 if (UseAdaptiveSizePolicy) {
2088 size_policy()->msc_collection_end(gch->gc_cause());
2089 }
2090
2091 gc_timer->register_gc_end(os::elapsed_counter());
2092
2093 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2094
2095 // For a mark-sweep-compact, compute_new_size() will be called
2096 // in the heap's do_collection() method.
2097 }
2098
2099 // A work method used by the foreground collector to do
2100 // a mark-sweep, after taking over from a possibly on-going
2101 // concurrent mark-sweep collection.
2102 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2103 CollectorState first_state, bool should_start_over) {
2104 if (PrintGC && Verbose) {
2105 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2106 "collector with count %d",
2107 _full_gcs_since_conc_gc);
2108 }
2109 switch (_collectorState) {
2110 case Idling:
2111 if (first_state == Idling || should_start_over) {
2112 // The background GC was not active, or should
2113 // restarted from scratch; start the cycle.
2114 _collectorState = InitialMarking;
2115 }
2116 // If first_state was not Idling, then a background GC
2117 // was in progress and has now finished. No need to do it
2118 // again. Leave the state as Idling.
2119 break;
2120 case Precleaning:
2121 // In the foreground case don't do the precleaning since
2122 // it is not done concurrently and there is extra work
2123 // required.
2124 _collectorState = FinalMarking;
2125 }
2126 collect_in_foreground(clear_all_soft_refs);
2127
2128 // For a mark-sweep, compute_new_size() will be called
2129 // in the heap's do_collection() method.
2130 }
2131
2132
2133 void CMSCollector::getFreelistLocks() const {
2134 // Get locks for all free lists in all generations that this
2135 // collector is responsible for
2136 _cmsGen->freelistLock()->lock_without_safepoint_check();
2137 _permGen->freelistLock()->lock_without_safepoint_check();
2138 }
2139
2140 void CMSCollector::releaseFreelistLocks() const {
2141 // Release locks for all free lists in all generations that this
2142 // collector is responsible for
2143 _cmsGen->freelistLock()->unlock();
2144 _permGen->freelistLock()->unlock();
2145 }
2146
2147 bool CMSCollector::haveFreelistLocks() const {
2148 // Check locks for all free lists in all generations that this
2149 // collector is responsible for
2150 assert_lock_strong(_cmsGen->freelistLock());
2151 assert_lock_strong(_permGen->freelistLock());
2152 PRODUCT_ONLY(ShouldNotReachHere());
2153 return true;
2154 }
2155
2156 // A utility class that is used by the CMS collector to
2157 // temporarily "release" the foreground collector from its
2158 // usual obligation to wait for the background collector to
2159 // complete an ongoing phase before proceeding.
2160 class ReleaseForegroundGC: public StackObj {
2161 private:
2162 CMSCollector* _c;
2163 public:
2164 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2165 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2166 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2167 // allow a potentially blocked foreground collector to proceed
2168 _c->_foregroundGCShouldWait = false;
2169 if (_c->_foregroundGCIsActive) {
2170 CGC_lock->notify();
2171 }
2172 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2173 "Possible deadlock");
2174 }
2175
2176 ~ReleaseForegroundGC() {
2177 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2178 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2179 _c->_foregroundGCShouldWait = true;
2180 }
2181 };
2182
2183 // There are separate collect_in_background and collect_in_foreground because of
2184 // the different locking requirements of the background collector and the
2185 // foreground collector. There was originally an attempt to share
2186 // one "collect" method between the background collector and the foreground
2187 // collector but the if-then-else required made it cleaner to have
2188 // separate methods.
2189 void CMSCollector::collect_in_background(bool clear_all_soft_refs) {
2190 assert(Thread::current()->is_ConcurrentGC_thread(),
2191 "A CMS asynchronous collection is only allowed on a CMS thread.");
2192
2193 GenCollectedHeap* gch = GenCollectedHeap::heap();
2194 {
2195 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2196 MutexLockerEx hl(Heap_lock, safepoint_check);
2197 FreelistLocker fll(this);
2198 MutexLockerEx x(CGC_lock, safepoint_check);
2199 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2200 // The foreground collector is active or we're
2201 // not using asynchronous collections. Skip this
2202 // background collection.
2203 assert(!_foregroundGCShouldWait, "Should be clear");
2204 return;
2205 } else {
2206 assert(_collectorState == Idling, "Should be idling before start.");
2207 _collectorState = InitialMarking;
2208 register_gc_start(GCCause::_cms_concurrent_mark);
2209 // Reset the expansion cause, now that we are about to begin
2210 // a new cycle.
2211 clear_expansion_cause();
2212 }
2213 // Decide if we want to enable class unloading as part of the
2214 // ensuing concurrent GC cycle.
2215 update_should_unload_classes();
2216 _full_gc_requested = false; // acks all outstanding full gc requests
2217 // Signal that we are about to start a collection
2218 gch->increment_total_full_collections(); // ... starting a collection cycle
2219 _collection_count_start = gch->total_full_collections();
2220 }
2221
2222 // Used for PrintGC
2223 size_t prev_used;
2224 if (PrintGC && Verbose) {
2225 prev_used = _cmsGen->used(); // XXXPERM
2226 }
2227
2228 // The change of the collection state is normally done at this level;
2229 // the exceptions are phases that are executed while the world is
2230 // stopped. For those phases the change of state is done while the
2231 // world is stopped. For baton passing purposes this allows the
2232 // background collector to finish the phase and change state atomically.
2233 // The foreground collector cannot wait on a phase that is done
2234 // while the world is stopped because the foreground collector already
2235 // has the world stopped and would deadlock.
2236 while (_collectorState != Idling) {
2237 if (TraceCMSState) {
2238 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2239 Thread::current(), _collectorState);
2240 }
2241 // The foreground collector
2242 // holds the Heap_lock throughout its collection.
2243 // holds the CMS token (but not the lock)
2244 // except while it is waiting for the background collector to yield.
2245 //
2246 // The foreground collector should be blocked (not for long)
2247 // if the background collector is about to start a phase
2248 // executed with world stopped. If the background
2249 // collector has already started such a phase, the
2250 // foreground collector is blocked waiting for the
2251 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2252 // are executed in the VM thread.
2253 //
2254 // The locking order is
2255 // PendingListLock (PLL) -- if applicable (FinalMarking)
2256 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2257 // CMS token (claimed in
2258 // stop_world_and_do() -->
2259 // safepoint_synchronize() -->
2260 // CMSThread::synchronize())
2261
2262 {
2263 // Check if the FG collector wants us to yield.
2264 CMSTokenSync x(true); // is cms thread
2265 if (waitForForegroundGC()) {
2266 // We yielded to a foreground GC, nothing more to be
2267 // done this round.
2268 assert(_foregroundGCShouldWait == false, "We set it to false in "
2269 "waitForForegroundGC()");
2270 if (TraceCMSState) {
2271 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2272 " exiting collection CMS state %d",
2273 Thread::current(), _collectorState);
2274 }
2275 return;
2276 } else {
2277 // The background collector can run but check to see if the
2278 // foreground collector has done a collection while the
2279 // background collector was waiting to get the CGC_lock
2280 // above. If yes, break so that _foregroundGCShouldWait
2281 // is cleared before returning.
2282 if (_collectorState == Idling) {
2283 break;
2284 }
2285 }
2286 }
2287
2288 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2289 "should be waiting");
2290
2291 switch (_collectorState) {
2292 case InitialMarking:
2293 {
2294 ReleaseForegroundGC x(this);
2295 stats().record_cms_begin();
2296 VM_CMS_Initial_Mark initial_mark_op(this);
2297 VMThread::execute(&initial_mark_op);
2298 }
2299 // The collector state may be any legal state at this point
2300 // since the background collector may have yielded to the
2301 // foreground collector.
2302 break;
2303 case Marking:
2304 // initial marking in checkpointRootsInitialWork has been completed
2305 if (markFromRoots(true)) { // we were successful
2306 assert(_collectorState == Precleaning, "Collector state should "
2307 "have changed");
2308 } else {
2309 assert(_foregroundGCIsActive, "Internal state inconsistency");
2310 }
2311 break;
2312 case Precleaning:
2313 if (UseAdaptiveSizePolicy) {
2314 size_policy()->concurrent_precleaning_begin();
2315 }
2316 // marking from roots in markFromRoots has been completed
2317 preclean();
2318 if (UseAdaptiveSizePolicy) {
2319 size_policy()->concurrent_precleaning_end();
2320 }
2321 assert(_collectorState == AbortablePreclean ||
2322 _collectorState == FinalMarking,
2323 "Collector state should have changed");
2324 break;
2325 case AbortablePreclean:
2326 if (UseAdaptiveSizePolicy) {
2327 size_policy()->concurrent_phases_resume();
2328 }
2329 abortable_preclean();
2330 if (UseAdaptiveSizePolicy) {
2331 size_policy()->concurrent_precleaning_end();
2332 }
2333 assert(_collectorState == FinalMarking, "Collector state should "
2334 "have changed");
2335 break;
2336 case FinalMarking:
2337 {
2338 ReleaseForegroundGC x(this);
2339
2340 VM_CMS_Final_Remark final_remark_op(this);
2341 VMThread::execute(&final_remark_op);
2342 }
2343 assert(_foregroundGCShouldWait, "block post-condition");
2344 break;
2345 case Sweeping:
2346 if (UseAdaptiveSizePolicy) {
2347 size_policy()->concurrent_sweeping_begin();
2348 }
2349 // final marking in checkpointRootsFinal has been completed
2350 sweep(true);
2351 assert(_collectorState == Resizing, "Collector state change "
2352 "to Resizing must be done under the free_list_lock");
2353 _full_gcs_since_conc_gc = 0;
2354
2355 // Stop the timers for adaptive size policy for the concurrent phases
2356 if (UseAdaptiveSizePolicy) {
2357 size_policy()->concurrent_sweeping_end();
2358 size_policy()->concurrent_phases_end(gch->gc_cause(),
2359 gch->prev_gen(_cmsGen)->capacity(),
2360 _cmsGen->free());
2361 }
2362
2363 case Resizing: {
2364 // Sweeping has been completed...
2365 // At this point the background collection has completed.
2366 // Don't move the call to compute_new_size() down
2367 // into code that might be executed if the background
2368 // collection was preempted.
2369 {
2370 ReleaseForegroundGC x(this); // unblock FG collection
2371 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2372 CMSTokenSync z(true); // not strictly needed.
2373 if (_collectorState == Resizing) {
2374 compute_new_size();
2375 save_heap_summary();
2376 _collectorState = Resetting;
2377 } else {
2378 assert(_collectorState == Idling, "The state should only change"
2379 " because the foreground collector has finished the collection");
2380 }
2381 }
2382 break;
2383 }
2384 case Resetting:
2385 // CMS heap resizing has been completed
2386 reset(true);
2387 assert(_collectorState == Idling, "Collector state should "
2388 "have changed");
2389 stats().record_cms_end();
2390 // Don't move the concurrent_phases_end() and compute_new_size()
2391 // calls to here because a preempted background collection
2392 // has it's state set to "Resetting".
2393 break;
2394 case Idling:
2395 default:
2396 ShouldNotReachHere();
2397 break;
2398 }
2399 if (TraceCMSState) {
2400 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2401 Thread::current(), _collectorState);
2402 }
2403 assert(_foregroundGCShouldWait, "block post-condition");
2404 }
2405
2406 // Should this be in gc_epilogue?
2407 collector_policy()->counters()->update_counters();
2408
2409 {
2410 // Clear _foregroundGCShouldWait and, in the event that the
2411 // foreground collector is waiting, notify it, before
2412 // returning.
2413 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2414 _foregroundGCShouldWait = false;
2415 if (_foregroundGCIsActive) {
2416 CGC_lock->notify();
2417 }
2418 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2419 "Possible deadlock");
2420 }
2421 if (TraceCMSState) {
2422 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2423 " exiting collection CMS state %d",
2424 Thread::current(), _collectorState);
2425 }
2426 if (PrintGC && Verbose) {
2427 _cmsGen->print_heap_change(prev_used);
2428 }
2429 }
2430
2431 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
2432 if (!_cms_start_registered) {
2433 register_gc_start(cause);
2434 }
2435 }
2436
2437 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2438 _cms_start_registered = true;
2439 _gc_timer_cm->register_gc_start(os::elapsed_counter());
2440 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2441 }
2442
2443 void CMSCollector::register_gc_end() {
2444 if (_cms_start_registered) {
2445 report_heap_summary(GCWhen::AfterGC);
2446
2447 _gc_timer_cm->register_gc_end(os::elapsed_counter());
2448 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2449 _cms_start_registered = false;
2450 }
2451 }
2452
2453 void CMSCollector::save_heap_summary() {
2454 GenCollectedHeap* gch = GenCollectedHeap::heap();
2455 _last_heap_summary = gch->create_heap_summary();
2456 _last_perm_gen_summary = gch->create_perm_gen_summary();
2457 }
2458
2459 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2460 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_perm_gen_summary);
2461 }
2462
2463 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) {
2464 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2465 "Foreground collector should be waiting, not executing");
2466 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2467 "may only be done by the VM Thread with the world stopped");
2468 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2469 "VM thread should have CMS token");
2470
2471 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2472 true, NULL);)
2473 if (UseAdaptiveSizePolicy) {
2474 size_policy()->ms_collection_begin();
2475 }
2476 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2477
2478 HandleMark hm; // Discard invalid handles created during verification
2479
2480 if (VerifyBeforeGC &&
2481 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2482 Universe::verify();
2483 }
2484
2485 // Snapshot the soft reference policy to be used in this collection cycle.
2486 ref_processor()->setup_policy(clear_all_soft_refs);
2487
2488 bool init_mark_was_synchronous = false; // until proven otherwise
2489 while (_collectorState != Idling) {
2490 if (TraceCMSState) {
2491 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2492 Thread::current(), _collectorState);
2493 }
2494 switch (_collectorState) {
2495 case InitialMarking:
2496 register_foreground_gc_start(GenCollectedHeap::heap()->gc_cause());
2497 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2498 checkpointRootsInitial(false);
2499 assert(_collectorState == Marking, "Collector state should have changed"
2500 " within checkpointRootsInitial()");
2501 break;
2502 case Marking:
2503 // initial marking in checkpointRootsInitialWork has been completed
2504 if (VerifyDuringGC &&
2505 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2506 gclog_or_tty->print("Verify before initial mark: ");
2507 Universe::verify();
2508 }
2509 {
2510 bool res = markFromRoots(false);
2511 assert(res && _collectorState == FinalMarking, "Collector state should "
2512 "have changed");
2513 break;
2514 }
2515 case FinalMarking:
2516 if (VerifyDuringGC &&
2517 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2518 gclog_or_tty->print("Verify before re-mark: ");
2519 Universe::verify();
2520 }
2521 checkpointRootsFinal(false, clear_all_soft_refs,
2522 init_mark_was_synchronous);
2523 assert(_collectorState == Sweeping, "Collector state should not "
2524 "have changed within checkpointRootsFinal()");
2525 break;
2526 case Sweeping:
2527 // final marking in checkpointRootsFinal has been completed
2528 if (VerifyDuringGC &&
2529 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2530 gclog_or_tty->print("Verify before sweep: ");
2531 Universe::verify();
2532 }
2533 sweep(false);
2534 assert(_collectorState == Resizing, "Incorrect state");
2535 break;
2536 case Resizing: {
2537 // Sweeping has been completed; the actual resize in this case
2538 // is done separately; nothing to be done in this state.
2539 _collectorState = Resetting;
2540 break;
2541 }
2542 case Resetting:
2543 // The heap has been resized.
2544 if (VerifyDuringGC &&
2545 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2546 gclog_or_tty->print("Verify before reset: ");
2547 Universe::verify();
2548 }
2549 save_heap_summary();
2550 reset(false);
2551 assert(_collectorState == Idling, "Collector state should "
2552 "have changed");
2553 break;
2554 case Precleaning:
2555 case AbortablePreclean:
2556 // Elide the preclean phase
2557 _collectorState = FinalMarking;
2558 break;
2559 default:
2560 ShouldNotReachHere();
2561 }
2562 if (TraceCMSState) {
2563 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2564 Thread::current(), _collectorState);
2565 }
2566 }
2567
2568 if (UseAdaptiveSizePolicy) {
2569 GenCollectedHeap* gch = GenCollectedHeap::heap();
2570 size_policy()->ms_collection_end(gch->gc_cause());
2571 }
2572
2573 if (VerifyAfterGC &&
2574 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2575 Universe::verify();
2576 }
2577 if (TraceCMSState) {
2578 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2579 " exiting collection CMS state %d",
2580 Thread::current(), _collectorState);
2581 }
2582 }
2583
2584 bool CMSCollector::waitForForegroundGC() {
2585 bool res = false;
2586 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2587 "CMS thread should have CMS token");
2588 // Block the foreground collector until the
2589 // background collectors decides whether to
2590 // yield.
2591 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2592 _foregroundGCShouldWait = true;
2593 if (_foregroundGCIsActive) {
2594 // The background collector yields to the
2595 // foreground collector and returns a value
2596 // indicating that it has yielded. The foreground
2597 // collector can proceed.
2598 res = true;
2599 _foregroundGCShouldWait = false;
2600 ConcurrentMarkSweepThread::clear_CMS_flag(
2601 ConcurrentMarkSweepThread::CMS_cms_has_token);
2602 ConcurrentMarkSweepThread::set_CMS_flag(
2603 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2604 // Get a possibly blocked foreground thread going
2605 CGC_lock->notify();
2606 if (TraceCMSState) {
2607 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2608 Thread::current(), _collectorState);
2609 }
2610 while (_foregroundGCIsActive) {
2611 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2612 }
2613 ConcurrentMarkSweepThread::set_CMS_flag(
2614 ConcurrentMarkSweepThread::CMS_cms_has_token);
2615 ConcurrentMarkSweepThread::clear_CMS_flag(
2616 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2617 }
2618 if (TraceCMSState) {
2619 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2620 Thread::current(), _collectorState);
2621 }
2622 return res;
2623 }
2624
2625 // Because of the need to lock the free lists and other structures in
2626 // the collector, common to all the generations that the collector is
2627 // collecting, we need the gc_prologues of individual CMS generations
2628 // delegate to their collector. It may have been simpler had the
2629 // current infrastructure allowed one to call a prologue on a
2630 // collector. In the absence of that we have the generation's
2631 // prologue delegate to the collector, which delegates back
2632 // some "local" work to a worker method in the individual generations
2633 // that it's responsible for collecting, while itself doing any
2634 // work common to all generations it's responsible for. A similar
2635 // comment applies to the gc_epilogue()'s.
2636 // The role of the varaible _between_prologue_and_epilogue is to
2637 // enforce the invocation protocol.
2638 void CMSCollector::gc_prologue(bool full) {
2639 // Call gc_prologue_work() for each CMSGen and PermGen that
2640 // we are responsible for.
2641
2642 // The following locking discipline assumes that we are only called
2643 // when the world is stopped.
2644 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2645
2646 // The CMSCollector prologue must call the gc_prologues for the
2647 // "generations" (including PermGen if any) that it's responsible
2648 // for.
2649
2650 assert( Thread::current()->is_VM_thread()
2651 || ( CMSScavengeBeforeRemark
2652 && Thread::current()->is_ConcurrentGC_thread()),
2653 "Incorrect thread type for prologue execution");
2654
2655 if (_between_prologue_and_epilogue) {
2656 // We have already been invoked; this is a gc_prologue delegation
2657 // from yet another CMS generation that we are responsible for, just
2658 // ignore it since all relevant work has already been done.
2659 return;
2660 }
2661
2662 // set a bit saying prologue has been called; cleared in epilogue
2663 _between_prologue_and_epilogue = true;
2664 // Claim locks for common data structures, then call gc_prologue_work()
2665 // for each CMSGen and PermGen that we are responsible for.
2666
2667 getFreelistLocks(); // gets free list locks on constituent spaces
2668 bitMapLock()->lock_without_safepoint_check();
2669
2670 // Should call gc_prologue_work() for all cms gens we are responsible for
2671 bool registerClosure = _collectorState >= Marking
2672 && _collectorState < Sweeping;
2673 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2674 &_modUnionClosurePar
2675 : &_modUnionClosure;
2676 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2677 _permGen->gc_prologue_work(full, registerClosure, muc);
2678
2679 if (!full) {
2680 stats().record_gc0_begin();
2681 }
2682 }
2683
2684 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2685 // Delegate to CMScollector which knows how to coordinate between
2686 // this and any other CMS generations that it is responsible for
2687 // collecting.
2688 collector()->gc_prologue(full);
2689 }
2690
2691 // This is a "private" interface for use by this generation's CMSCollector.
2692 // Not to be called directly by any other entity (for instance,
2693 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2694 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2695 bool registerClosure, ModUnionClosure* modUnionClosure) {
2696 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2697 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2698 "Should be NULL");
2699 if (registerClosure) {
2700 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2701 }
2702 cmsSpace()->gc_prologue();
2703 // Clear stat counters
2704 NOT_PRODUCT(
2705 assert(_numObjectsPromoted == 0, "check");
2706 assert(_numWordsPromoted == 0, "check");
2707 if (Verbose && PrintGC) {
2708 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2709 SIZE_FORMAT" bytes concurrently",
2710 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2711 }
2712 _numObjectsAllocated = 0;
2713 _numWordsAllocated = 0;
2714 )
2715 }
2716
2717 void CMSCollector::gc_epilogue(bool full) {
2718 // The following locking discipline assumes that we are only called
2719 // when the world is stopped.
2720 assert(SafepointSynchronize::is_at_safepoint(),
2721 "world is stopped assumption");
2722
2723 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2724 // if linear allocation blocks need to be appropriately marked to allow the
2725 // the blocks to be parsable. We also check here whether we need to nudge the
2726 // CMS collector thread to start a new cycle (if it's not already active).
2727 assert( Thread::current()->is_VM_thread()
2728 || ( CMSScavengeBeforeRemark
2729 && Thread::current()->is_ConcurrentGC_thread()),
2730 "Incorrect thread type for epilogue execution");
2731
2732 if (!_between_prologue_and_epilogue) {
2733 // We have already been invoked; this is a gc_epilogue delegation
2734 // from yet another CMS generation that we are responsible for, just
2735 // ignore it since all relevant work has already been done.
2736 return;
2737 }
2738 assert(haveFreelistLocks(), "must have freelist locks");
2739 assert_lock_strong(bitMapLock());
2740
2741 _cmsGen->gc_epilogue_work(full);
2742 _permGen->gc_epilogue_work(full);
2743
2744 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2745 // in case sampling was not already enabled, enable it
2746 _start_sampling = true;
2747 }
2748 // reset _eden_chunk_array so sampling starts afresh
2749 _eden_chunk_index = 0;
2750
2751 size_t cms_used = _cmsGen->cmsSpace()->used();
2752 size_t perm_used = _permGen->cmsSpace()->used();
2753
2754 // update performance counters - this uses a special version of
2755 // update_counters() that allows the utilization to be passed as a
2756 // parameter, avoiding multiple calls to used().
2757 //
2758 _cmsGen->update_counters(cms_used);
2759 _permGen->update_counters(perm_used);
2760
2761 if (CMSIncrementalMode) {
2762 icms_update_allocation_limits();
2763 }
2764
2765 bitMapLock()->unlock();
2766 releaseFreelistLocks();
2767
2768 if (!CleanChunkPoolAsync) {
2769 Chunk::clean_chunk_pool();
2770 }
2771
2772 _between_prologue_and_epilogue = false; // ready for next cycle
2773 }
2774
2775 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2776 collector()->gc_epilogue(full);
2777
2778 // Also reset promotion tracking in par gc thread states.
2779 if (CollectedHeap::use_parallel_gc_threads()) {
2780 for (uint i = 0; i < ParallelGCThreads; i++) {
2781 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2782 }
2783 }
2784 }
2785
2786 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2787 assert(!incremental_collection_failed(), "Should have been cleared");
2788 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2789 cmsSpace()->gc_epilogue();
2790 // Print stat counters
2791 NOT_PRODUCT(
2792 assert(_numObjectsAllocated == 0, "check");
2793 assert(_numWordsAllocated == 0, "check");
2794 if (Verbose && PrintGC) {
2795 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2796 SIZE_FORMAT" bytes",
2797 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2798 }
2799 _numObjectsPromoted = 0;
2800 _numWordsPromoted = 0;
2801 )
2802
2803 if (PrintGC && Verbose) {
2804 // Call down the chain in contiguous_available needs the freelistLock
2805 // so print this out before releasing the freeListLock.
2806 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2807 contiguous_available());
2808 }
2809 }
2810
2811 #ifndef PRODUCT
2812 bool CMSCollector::have_cms_token() {
2813 Thread* thr = Thread::current();
2814 if (thr->is_VM_thread()) {
2815 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2816 } else if (thr->is_ConcurrentGC_thread()) {
2817 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2818 } else if (thr->is_GC_task_thread()) {
2819 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2820 ParGCRareEvent_lock->owned_by_self();
2821 }
2822 return false;
2823 }
2824 #endif
2825
2826 // Check reachability of the given heap address in CMS generation,
2827 // treating all other generations as roots.
2828 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2829 // We could "guarantee" below, rather than assert, but i'll
2830 // leave these as "asserts" so that an adventurous debugger
2831 // could try this in the product build provided some subset of
2832 // the conditions were met, provided they were intersted in the
2833 // results and knew that the computation below wouldn't interfere
2834 // with other concurrent computations mutating the structures
2835 // being read or written.
2836 assert(SafepointSynchronize::is_at_safepoint(),
2837 "Else mutations in object graph will make answer suspect");
2838 assert(have_cms_token(), "Should hold cms token");
2839 assert(haveFreelistLocks(), "must hold free list locks");
2840 assert_lock_strong(bitMapLock());
2841
2842 // Clear the marking bit map array before starting, but, just
2843 // for kicks, first report if the given address is already marked
2844 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2845 _markBitMap.isMarked(addr) ? "" : " not");
2846
2847 if (verify_after_remark()) {
2848 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2849 bool result = verification_mark_bm()->isMarked(addr);
2850 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2851 result ? "IS" : "is NOT");
2852 return result;
2853 } else {
2854 gclog_or_tty->print_cr("Could not compute result");
2855 return false;
2856 }
2857 }
2858
2859 ////////////////////////////////////////////////////////
2860 // CMS Verification Support
2861 ////////////////////////////////////////////////////////
2862 // Following the remark phase, the following invariant
2863 // should hold -- each object in the CMS heap which is
2864 // marked in markBitMap() should be marked in the verification_mark_bm().
2865
2866 class VerifyMarkedClosure: public BitMapClosure {
2867 CMSBitMap* _marks;
2868 bool _failed;
2869
2870 public:
2871 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2872
2873 bool do_bit(size_t offset) {
2874 HeapWord* addr = _marks->offsetToHeapWord(offset);
2875 if (!_marks->isMarked(addr)) {
2876 oop(addr)->print_on(gclog_or_tty);
2877 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2878 _failed = true;
2879 }
2880 return true;
2881 }
2882
2883 bool failed() { return _failed; }
2884 };
2885
2886 bool CMSCollector::verify_after_remark() {
2887 gclog_or_tty->print(" [Verifying CMS Marking... ");
2888 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2889 static bool init = false;
2890
2891 assert(SafepointSynchronize::is_at_safepoint(),
2892 "Else mutations in object graph will make answer suspect");
2893 assert(have_cms_token(),
2894 "Else there may be mutual interference in use of "
2895 " verification data structures");
2896 assert(_collectorState > Marking && _collectorState <= Sweeping,
2897 "Else marking info checked here may be obsolete");
2898 assert(haveFreelistLocks(), "must hold free list locks");
2899 assert_lock_strong(bitMapLock());
2900
2901
2902 // Allocate marking bit map if not already allocated
2903 if (!init) { // first time
2904 if (!verification_mark_bm()->allocate(_span)) {
2905 return false;
2906 }
2907 init = true;
2908 }
2909
2910 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2911
2912 // Turn off refs discovery -- so we will be tracing through refs.
2913 // This is as intended, because by this time
2914 // GC must already have cleared any refs that need to be cleared,
2915 // and traced those that need to be marked; moreover,
2916 // the marking done here is not going to intefere in any
2917 // way with the marking information used by GC.
2918 NoRefDiscovery no_discovery(ref_processor());
2919
2920 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2921
2922 // Clear any marks from a previous round
2923 verification_mark_bm()->clear_all();
2924 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2925 verify_work_stacks_empty();
2926
2927 GenCollectedHeap* gch = GenCollectedHeap::heap();
2928 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
2929 // Update the saved marks which may affect the root scans.
2930 gch->save_marks();
2931
2932 if (CMSRemarkVerifyVariant == 1) {
2933 // In this first variant of verification, we complete
2934 // all marking, then check if the new marks-verctor is
2935 // a subset of the CMS marks-vector.
2936 verify_after_remark_work_1();
2937 } else if (CMSRemarkVerifyVariant == 2) {
2938 // In this second variant of verification, we flag an error
2939 // (i.e. an object reachable in the new marks-vector not reachable
2940 // in the CMS marks-vector) immediately, also indicating the
2941 // identify of an object (A) that references the unmarked object (B) --
2942 // presumably, a mutation to A failed to be picked up by preclean/remark?
2943 verify_after_remark_work_2();
2944 } else {
2945 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
2946 CMSRemarkVerifyVariant);
2947 }
2948 gclog_or_tty->print(" done] ");
2949 return true;
2950 }
2951
2952 void CMSCollector::verify_after_remark_work_1() {
2953 ResourceMark rm;
2954 HandleMark hm;
2955 GenCollectedHeap* gch = GenCollectedHeap::heap();
2956
2957 // Mark from roots one level into CMS
2958 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2959 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2960
2961 gch->gen_process_strong_roots(_cmsGen->level(),
2962 true, // younger gens are roots
2963 true, // activate StrongRootsScope
2964 true, // collecting perm gen
2965 SharedHeap::ScanningOption(roots_scanning_options()),
2966 ¬Older,
2967 true, // walk code active on stacks
2968 NULL);
2969
2970 // Now mark from the roots
2971 assert(_revisitStack.isEmpty(), "Should be empty");
2972 MarkFromRootsClosure markFromRootsClosure(this, _span,
2973 verification_mark_bm(), verification_mark_stack(), &_revisitStack,
2974 false /* don't yield */, true /* verifying */);
2975 assert(_restart_addr == NULL, "Expected pre-condition");
2976 verification_mark_bm()->iterate(&markFromRootsClosure);
2977 while (_restart_addr != NULL) {
2978 // Deal with stack overflow: by restarting at the indicated
2979 // address.
2980 HeapWord* ra = _restart_addr;
2981 markFromRootsClosure.reset(ra);
2982 _restart_addr = NULL;
2983 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2984 }
2985 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2986 verify_work_stacks_empty();
2987 // Should reset the revisit stack above, since no class tree
2988 // surgery is forthcoming.
2989 _revisitStack.reset(); // throwing away all contents
2990
2991 // Marking completed -- now verify that each bit marked in
2992 // verification_mark_bm() is also marked in markBitMap(); flag all
2993 // errors by printing corresponding objects.
2994 VerifyMarkedClosure vcl(markBitMap());
2995 verification_mark_bm()->iterate(&vcl);
2996 if (vcl.failed()) {
2997 gclog_or_tty->print("Verification failed");
2998 Universe::heap()->print_on(gclog_or_tty);
2999 fatal("CMS: failed marking verification after remark");
3000 }
3001 }
3002
3003 void CMSCollector::verify_after_remark_work_2() {
3004 ResourceMark rm;
3005 HandleMark hm;
3006 GenCollectedHeap* gch = GenCollectedHeap::heap();
3007
3008 // Mark from roots one level into CMS
3009 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
3010 markBitMap());
3011 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3012 gch->gen_process_strong_roots(_cmsGen->level(),
3013 true, // younger gens are roots
3014 true, // activate StrongRootsScope
3015 true, // collecting perm gen
3016 SharedHeap::ScanningOption(roots_scanning_options()),
3017 ¬Older,
3018 true, // walk code active on stacks
3019 NULL);
3020
3021 // Now mark from the roots
3022 assert(_revisitStack.isEmpty(), "Should be empty");
3023 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
3024 verification_mark_bm(), markBitMap(), verification_mark_stack());
3025 assert(_restart_addr == NULL, "Expected pre-condition");
3026 verification_mark_bm()->iterate(&markFromRootsClosure);
3027 while (_restart_addr != NULL) {
3028 // Deal with stack overflow: by restarting at the indicated
3029 // address.
3030 HeapWord* ra = _restart_addr;
3031 markFromRootsClosure.reset(ra);
3032 _restart_addr = NULL;
3033 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3034 }
3035 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3036 verify_work_stacks_empty();
3037 // Should reset the revisit stack above, since no class tree
3038 // surgery is forthcoming.
3039 _revisitStack.reset(); // throwing away all contents
3040
3041 // Marking completed -- now verify that each bit marked in
3042 // verification_mark_bm() is also marked in markBitMap(); flag all
3043 // errors by printing corresponding objects.
3044 VerifyMarkedClosure vcl(markBitMap());
3045 verification_mark_bm()->iterate(&vcl);
3046 assert(!vcl.failed(), "Else verification above should not have succeeded");
3047 }
3048
3049 void ConcurrentMarkSweepGeneration::save_marks() {
3050 // delegate to CMS space
3051 cmsSpace()->save_marks();
3052 for (uint i = 0; i < ParallelGCThreads; i++) {
3053 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3054 }
3055 }
3056
3057 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3058 return cmsSpace()->no_allocs_since_save_marks();
3059 }
3060
3061 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3062 \
3063 void ConcurrentMarkSweepGeneration:: \
3064 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3065 cl->set_generation(this); \
3066 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3067 cl->reset_generation(); \
3068 save_marks(); \
3069 }
3070
3071 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3072
3073 void
3074 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk)
3075 {
3076 // Not currently implemented; need to do the following. -- ysr.
3077 // dld -- I think that is used for some sort of allocation profiler. So it
3078 // really means the objects allocated by the mutator since the last
3079 // GC. We could potentially implement this cheaply by recording only
3080 // the direct allocations in a side data structure.
3081 //
3082 // I think we probably ought not to be required to support these
3083 // iterations at any arbitrary point; I think there ought to be some
3084 // call to enable/disable allocation profiling in a generation/space,
3085 // and the iterator ought to return the objects allocated in the
3086 // gen/space since the enable call, or the last iterator call (which
3087 // will probably be at a GC.) That way, for gens like CM&S that would
3088 // require some extra data structure to support this, we only pay the
3089 // cost when it's in use...
3090 cmsSpace()->object_iterate_since_last_GC(blk);
3091 }
3092
3093 void
3094 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3095 cl->set_generation(this);
3096 younger_refs_in_space_iterate(_cmsSpace, cl);
3097 cl->reset_generation();
3098 }
3099
3100 void
3101 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) {
3102 if (freelistLock()->owned_by_self()) {
3103 Generation::oop_iterate(mr, cl);
3104 } else {
3105 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3106 Generation::oop_iterate(mr, cl);
3107 }
3108 }
3109
3110 void
3111 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) {
3112 if (freelistLock()->owned_by_self()) {
3113 Generation::oop_iterate(cl);
3114 } else {
3115 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3116 Generation::oop_iterate(cl);
3117 }
3118 }
3119
3120 void
3121 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3122 if (freelistLock()->owned_by_self()) {
3123 Generation::object_iterate(cl);
3124 } else {
3125 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3126 Generation::object_iterate(cl);
3127 }
3128 }
3129
3130 void
3131 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3132 if (freelistLock()->owned_by_self()) {
3133 Generation::safe_object_iterate(cl);
3134 } else {
3135 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3136 Generation::safe_object_iterate(cl);
3137 }
3138 }
3139
3140 void
3141 ConcurrentMarkSweepGeneration::pre_adjust_pointers() {
3142 }
3143
3144 void
3145 ConcurrentMarkSweepGeneration::post_compact() {
3146 }
3147
3148 void
3149 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3150 // Fix the linear allocation blocks to look like free blocks.
3151
3152 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3153 // are not called when the heap is verified during universe initialization and
3154 // at vm shutdown.
3155 if (freelistLock()->owned_by_self()) {
3156 cmsSpace()->prepare_for_verify();
3157 } else {
3158 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3159 cmsSpace()->prepare_for_verify();
3160 }
3161 }
3162
3163 void
3164 ConcurrentMarkSweepGeneration::verify() {
3165 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3166 // are not called when the heap is verified during universe initialization and
3167 // at vm shutdown.
3168 if (freelistLock()->owned_by_self()) {
3169 cmsSpace()->verify();
3170 } else {
3171 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3172 cmsSpace()->verify();
3173 }
3174 }
3175
3176 void CMSCollector::verify() {
3177 _cmsGen->verify();
3178 _permGen->verify();
3179 }
3180
3181 #ifndef PRODUCT
3182 bool CMSCollector::overflow_list_is_empty() const {
3183 assert(_num_par_pushes >= 0, "Inconsistency");
3184 if (_overflow_list == NULL) {
3185 assert(_num_par_pushes == 0, "Inconsistency");
3186 }
3187 return _overflow_list == NULL;
3188 }
3189
3190 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3191 // merely consolidate assertion checks that appear to occur together frequently.
3192 void CMSCollector::verify_work_stacks_empty() const {
3193 assert(_markStack.isEmpty(), "Marking stack should be empty");
3194 assert(overflow_list_is_empty(), "Overflow list should be empty");
3195 }
3196
3197 void CMSCollector::verify_overflow_empty() const {
3198 assert(overflow_list_is_empty(), "Overflow list should be empty");
3199 assert(no_preserved_marks(), "No preserved marks");
3200 }
3201 #endif // PRODUCT
3202
3203 // Decide if we want to enable class unloading as part of the
3204 // ensuing concurrent GC cycle. We will collect the perm gen and
3205 // unload classes if it's the case that:
3206 // (1) an explicit gc request has been made and the flag
3207 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3208 // (2) (a) class unloading is enabled at the command line, and
3209 // (b) (i) perm gen threshold has been crossed, or
3210 // (ii) old gen is getting really full, or
3211 // (iii) the previous N CMS collections did not collect the
3212 // perm gen
3213 // NOTE: Provided there is no change in the state of the heap between
3214 // calls to this method, it should have idempotent results. Moreover,
3215 // its results should be monotonically increasing (i.e. going from 0 to 1,
3216 // but not 1 to 0) between successive calls between which the heap was
3217 // not collected. For the implementation below, it must thus rely on
3218 // the property that concurrent_cycles_since_last_unload()
3219 // will not decrease unless a collection cycle happened and that
3220 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are
3221 // themselves also monotonic in that sense. See check_monotonicity()
3222 // below.
3223 bool CMSCollector::update_should_unload_classes() {
3224 _should_unload_classes = false;
3225 // Condition 1 above
3226 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3227 _should_unload_classes = true;
3228 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3229 // Disjuncts 2.b.(i,ii,iii) above
3230 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3231 CMSClassUnloadingMaxInterval)
3232 || _permGen->should_concurrent_collect()
3233 || _cmsGen->is_too_full();
3234 }
3235 return _should_unload_classes;
3236 }
3237
3238 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3239 bool res = should_concurrent_collect();
3240 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3241 return res;
3242 }
3243
3244 void CMSCollector::setup_cms_unloading_and_verification_state() {
3245 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3246 || VerifyBeforeExit;
3247 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3248
3249 if (should_unload_classes()) { // Should unload classes this cycle
3250 remove_root_scanning_option(rso); // Shrink the root set appropriately
3251 set_verifying(should_verify); // Set verification state for this cycle
3252 return; // Nothing else needs to be done at this time
3253 }
3254
3255 // Not unloading classes this cycle
3256 assert(!should_unload_classes(), "Inconsitency!");
3257 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3258 // We were not verifying, or we _were_ unloading classes in the last cycle,
3259 // AND some verification options are enabled this cycle; in this case,
3260 // we must make sure that the deadness map is allocated if not already so,
3261 // and cleared (if already allocated previously --
3262 // CMSBitMap::sizeInBits() is used to determine if it's allocated).
3263 if (perm_gen_verify_bit_map()->sizeInBits() == 0) {
3264 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) {
3265 warning("Failed to allocate permanent generation verification CMS Bit Map;\n"
3266 "permanent generation verification disabled");
3267 return; // Note that we leave verification disabled, so we'll retry this
3268 // allocation next cycle. We _could_ remember this failure
3269 // and skip further attempts and permanently disable verification
3270 // attempts if that is considered more desirable.
3271 }
3272 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()),
3273 "_perm_gen_ver_bit_map inconsistency?");
3274 } else {
3275 perm_gen_verify_bit_map()->clear_all();
3276 }
3277 // Include symbols, strings and code cache elements to prevent their resurrection.
3278 add_root_scanning_option(rso);
3279 set_verifying(true);
3280 } else if (verifying() && !should_verify) {
3281 // We were verifying, but some verification flags got disabled.
3282 set_verifying(false);
3283 // Exclude symbols, strings and code cache elements from root scanning to
3284 // reduce IM and RM pauses.
3285 remove_root_scanning_option(rso);
3286 }
3287 }
3288
3289
3290 #ifndef PRODUCT
3291 HeapWord* CMSCollector::block_start(const void* p) const {
3292 const HeapWord* addr = (HeapWord*)p;
3293 if (_span.contains(p)) {
3294 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3295 return _cmsGen->cmsSpace()->block_start(p);
3296 } else {
3297 assert(_permGen->cmsSpace()->is_in_reserved(addr),
3298 "Inconsistent _span?");
3299 return _permGen->cmsSpace()->block_start(p);
3300 }
3301 }
3302 return NULL;
3303 }
3304 #endif
3305
3306 HeapWord*
3307 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3308 bool tlab,
3309 bool parallel) {
3310 CMSSynchronousYieldRequest yr;
3311 assert(!tlab, "Can't deal with TLAB allocation");
3312 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3313 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3314 CMSExpansionCause::_satisfy_allocation);
3315 if (GCExpandToAllocateDelayMillis > 0) {
3316 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3317 }
3318 return have_lock_and_allocate(word_size, tlab);
3319 }
3320
3321 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3322 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3323 // to CardGeneration and share it...
3324 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3325 return CardGeneration::expand(bytes, expand_bytes);
3326 }
3327
3328 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3329 CMSExpansionCause::Cause cause)
3330 {
3331
3332 bool success = expand(bytes, expand_bytes);
3333
3334 // remember why we expanded; this information is used
3335 // by shouldConcurrentCollect() when making decisions on whether to start
3336 // a new CMS cycle.
3337 if (success) {
3338 set_expansion_cause(cause);
3339 if (PrintGCDetails && Verbose) {
3340 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3341 CMSExpansionCause::to_string(cause));
3342 }
3343 }
3344 }
3345
3346 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3347 HeapWord* res = NULL;
3348 MutexLocker x(ParGCRareEvent_lock);
3349 while (true) {
3350 // Expansion by some other thread might make alloc OK now:
3351 res = ps->lab.alloc(word_sz);
3352 if (res != NULL) return res;
3353 // If there's not enough expansion space available, give up.
3354 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3355 return NULL;
3356 }
3357 // Otherwise, we try expansion.
3358 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3359 CMSExpansionCause::_allocate_par_lab);
3360 // Now go around the loop and try alloc again;
3361 // A competing par_promote might beat us to the expansion space,
3362 // so we may go around the loop again if promotion fails agaion.
3363 if (GCExpandToAllocateDelayMillis > 0) {
3364 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3365 }
3366 }
3367 }
3368
3369
3370 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3371 PromotionInfo* promo) {
3372 MutexLocker x(ParGCRareEvent_lock);
3373 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3374 while (true) {
3375 // Expansion by some other thread might make alloc OK now:
3376 if (promo->ensure_spooling_space()) {
3377 assert(promo->has_spooling_space(),
3378 "Post-condition of successful ensure_spooling_space()");
3379 return true;
3380 }
3381 // If there's not enough expansion space available, give up.
3382 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3383 return false;
3384 }
3385 // Otherwise, we try expansion.
3386 expand(refill_size_bytes, MinHeapDeltaBytes,
3387 CMSExpansionCause::_allocate_par_spooling_space);
3388 // Now go around the loop and try alloc again;
3389 // A competing allocation might beat us to the expansion space,
3390 // so we may go around the loop again if allocation fails again.
3391 if (GCExpandToAllocateDelayMillis > 0) {
3392 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3393 }
3394 }
3395 }
3396
3397
3398
3399 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3400 assert_locked_or_safepoint(Heap_lock);
3401 size_t size = ReservedSpace::page_align_size_down(bytes);
3402 if (size > 0) {
3403 shrink_by(size);
3404 }
3405 }
3406
3407 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3408 assert_locked_or_safepoint(Heap_lock);
3409 bool result = _virtual_space.expand_by(bytes);
3410 if (result) {
3411 HeapWord* old_end = _cmsSpace->end();
3412 size_t new_word_size =
3413 heap_word_size(_virtual_space.committed_size());
3414 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3415 _bts->resize(new_word_size); // resize the block offset shared array
3416 Universe::heap()->barrier_set()->resize_covered_region(mr);
3417 // Hmmmm... why doesn't CFLS::set_end verify locking?
3418 // This is quite ugly; FIX ME XXX
3419 _cmsSpace->assert_locked(freelistLock());
3420 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3421
3422 // update the space and generation capacity counters
3423 if (UsePerfData) {
3424 _space_counters->update_capacity();
3425 _gen_counters->update_all();
3426 }
3427
3428 if (Verbose && PrintGC) {
3429 size_t new_mem_size = _virtual_space.committed_size();
3430 size_t old_mem_size = new_mem_size - bytes;
3431 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK",
3432 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3433 }
3434 }
3435 return result;
3436 }
3437
3438 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3439 assert_locked_or_safepoint(Heap_lock);
3440 bool success = true;
3441 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3442 if (remaining_bytes > 0) {
3443 success = grow_by(remaining_bytes);
3444 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3445 }
3446 return success;
3447 }
3448
3449 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3450 assert_locked_or_safepoint(Heap_lock);
3451 assert_lock_strong(freelistLock());
3452 // XXX Fix when compaction is implemented.
3453 warning("Shrinking of CMS not yet implemented");
3454 return;
3455 }
3456
3457
3458 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3459 // phases.
3460 class CMSPhaseAccounting: public StackObj {
3461 public:
3462 CMSPhaseAccounting(CMSCollector *collector,
3463 const char *phase,
3464 bool print_cr = true);
3465 ~CMSPhaseAccounting();
3466
3467 private:
3468 CMSCollector *_collector;
3469 const char *_phase;
3470 elapsedTimer _wallclock;
3471 bool _print_cr;
3472
3473 public:
3474 // Not MT-safe; so do not pass around these StackObj's
3475 // where they may be accessed by other threads.
3476 jlong wallclock_millis() {
3477 assert(_wallclock.is_active(), "Wall clock should not stop");
3478 _wallclock.stop(); // to record time
3479 jlong ret = _wallclock.milliseconds();
3480 _wallclock.start(); // restart
3481 return ret;
3482 }
3483 };
3484
3485 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3486 const char *phase,
3487 bool print_cr) :
3488 _collector(collector), _phase(phase), _print_cr(print_cr) {
3489
3490 if (PrintCMSStatistics != 0) {
3491 _collector->resetYields();
3492 }
3493 if (PrintGCDetails && PrintGCTimeStamps) {
3494 gclog_or_tty->date_stamp(PrintGCDateStamps);
3495 gclog_or_tty->stamp();
3496 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]",
3497 _collector->cmsGen()->short_name(), _phase);
3498 }
3499 _collector->resetTimer();
3500 _wallclock.start();
3501 _collector->startTimer();
3502 }
3503
3504 CMSPhaseAccounting::~CMSPhaseAccounting() {
3505 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3506 _collector->stopTimer();
3507 _wallclock.stop();
3508 if (PrintGCDetails) {
3509 gclog_or_tty->date_stamp(PrintGCDateStamps);
3510 gclog_or_tty->stamp(PrintGCTimeStamps);
3511 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3512 _collector->cmsGen()->short_name(),
3513 _phase, _collector->timerValue(), _wallclock.seconds());
3514 if (_print_cr) {
3515 gclog_or_tty->print_cr("");
3516 }
3517 if (PrintCMSStatistics != 0) {
3518 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3519 _collector->yields());
3520 }
3521 }
3522 }
3523
3524 // CMS work
3525
3526 // Checkpoint the roots into this generation from outside
3527 // this generation. [Note this initial checkpoint need only
3528 // be approximate -- we'll do a catch up phase subsequently.]
3529 void CMSCollector::checkpointRootsInitial(bool asynch) {
3530 assert(_collectorState == InitialMarking, "Wrong collector state");
3531 check_correct_thread_executing();
3532 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3533
3534 save_heap_summary();
3535 report_heap_summary(GCWhen::BeforeGC);
3536
3537 ReferenceProcessor* rp = ref_processor();
3538 SpecializationStats::clear();
3539 assert(_restart_addr == NULL, "Control point invariant");
3540 if (asynch) {
3541 // acquire locks for subsequent manipulations
3542 MutexLockerEx x(bitMapLock(),
3543 Mutex::_no_safepoint_check_flag);
3544 checkpointRootsInitialWork(asynch);
3545 // enable ("weak") refs discovery
3546 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3547 _collectorState = Marking;
3548 } else {
3549 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3550 // which recognizes if we are a CMS generation, and doesn't try to turn on
3551 // discovery; verify that they aren't meddling.
3552 assert(!rp->discovery_is_atomic(),
3553 "incorrect setting of discovery predicate");
3554 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3555 "ref discovery for this generation kind");
3556 // already have locks
3557 checkpointRootsInitialWork(asynch);
3558 // now enable ("weak") refs discovery
3559 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3560 _collectorState = Marking;
3561 }
3562 SpecializationStats::print();
3563 }
3564
3565 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3566 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3567 assert(_collectorState == InitialMarking, "just checking");
3568
3569 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3570 // precede our marking with a collection of all
3571 // younger generations to keep floating garbage to a minimum.
3572 // XXX: we won't do this for now -- it's an optimization to be done later.
3573
3574 // already have locks
3575 assert_lock_strong(bitMapLock());
3576 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3577
3578 // Setup the verification and class unloading state for this
3579 // CMS collection cycle.
3580 setup_cms_unloading_and_verification_state();
3581
3582 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3583 PrintGCDetails && Verbose, true, _gc_timer_cm);)
3584 if (UseAdaptiveSizePolicy) {
3585 size_policy()->checkpoint_roots_initial_begin();
3586 }
3587
3588 // Reset all the PLAB chunk arrays if necessary.
3589 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3590 reset_survivor_plab_arrays();
3591 }
3592
3593 ResourceMark rm;
3594 HandleMark hm;
3595
3596 FalseClosure falseClosure;
3597 // In the case of a synchronous collection, we will elide the
3598 // remark step, so it's important to catch all the nmethod oops
3599 // in this step.
3600 // The final 'true' flag to gen_process_strong_roots will ensure this.
3601 // If 'async' is true, we can relax the nmethod tracing.
3602 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3603 GenCollectedHeap* gch = GenCollectedHeap::heap();
3604
3605 verify_work_stacks_empty();
3606 verify_overflow_empty();
3607
3608 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3609 // Update the saved marks which may affect the root scans.
3610 gch->save_marks();
3611
3612 // weak reference processing has not started yet.
3613 ref_processor()->set_enqueuing_is_done(false);
3614
3615 {
3616 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);)
3617 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3618 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3619 gch->gen_process_strong_roots(_cmsGen->level(),
3620 true, // younger gens are roots
3621 true, // activate StrongRootsScope
3622 true, // collecting perm gen
3623 SharedHeap::ScanningOption(roots_scanning_options()),
3624 ¬Older,
3625 true, // walk all of code cache if (so & SO_CodeCache)
3626 NULL);
3627 }
3628
3629 // Clear mod-union table; it will be dirtied in the prologue of
3630 // CMS generation per each younger generation collection.
3631
3632 assert(_modUnionTable.isAllClear(),
3633 "Was cleared in most recent final checkpoint phase"
3634 " or no bits are set in the gc_prologue before the start of the next "
3635 "subsequent marking phase.");
3636
3637 // Save the end of the used_region of the constituent generations
3638 // to be used to limit the extent of sweep in each generation.
3639 save_sweep_limits();
3640 if (UseAdaptiveSizePolicy) {
3641 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3642 }
3643 verify_overflow_empty();
3644 }
3645
3646 bool CMSCollector::markFromRoots(bool asynch) {
3647 // we might be tempted to assert that:
3648 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3649 // "inconsistent argument?");
3650 // However that wouldn't be right, because it's possible that
3651 // a safepoint is indeed in progress as a younger generation
3652 // stop-the-world GC happens even as we mark in this generation.
3653 assert(_collectorState == Marking, "inconsistent state?");
3654 check_correct_thread_executing();
3655 verify_overflow_empty();
3656
3657 bool res;
3658 if (asynch) {
3659
3660 // Start the timers for adaptive size policy for the concurrent phases
3661 // Do it here so that the foreground MS can use the concurrent
3662 // timer since a foreground MS might has the sweep done concurrently
3663 // or STW.
3664 if (UseAdaptiveSizePolicy) {
3665 size_policy()->concurrent_marking_begin();
3666 }
3667
3668 // Weak ref discovery note: We may be discovering weak
3669 // refs in this generation concurrent (but interleaved) with
3670 // weak ref discovery by a younger generation collector.
3671
3672 CMSTokenSyncWithLocks ts(true, bitMapLock());
3673 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3674 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3675 res = markFromRootsWork(asynch);
3676 if (res) {
3677 _collectorState = Precleaning;
3678 } else { // We failed and a foreground collection wants to take over
3679 assert(_foregroundGCIsActive, "internal state inconsistency");
3680 assert(_restart_addr == NULL, "foreground will restart from scratch");
3681 if (PrintGCDetails) {
3682 gclog_or_tty->print_cr("bailing out to foreground collection");
3683 }
3684 }
3685 if (UseAdaptiveSizePolicy) {
3686 size_policy()->concurrent_marking_end();
3687 }
3688 } else {
3689 assert(SafepointSynchronize::is_at_safepoint(),
3690 "inconsistent with asynch == false");
3691 if (UseAdaptiveSizePolicy) {
3692 size_policy()->ms_collection_marking_begin();
3693 }
3694 // already have locks
3695 res = markFromRootsWork(asynch);
3696 _collectorState = FinalMarking;
3697 if (UseAdaptiveSizePolicy) {
3698 GenCollectedHeap* gch = GenCollectedHeap::heap();
3699 size_policy()->ms_collection_marking_end(gch->gc_cause());
3700 }
3701 }
3702 verify_overflow_empty();
3703 return res;
3704 }
3705
3706 bool CMSCollector::markFromRootsWork(bool asynch) {
3707 // iterate over marked bits in bit map, doing a full scan and mark
3708 // from these roots using the following algorithm:
3709 // . if oop is to the right of the current scan pointer,
3710 // mark corresponding bit (we'll process it later)
3711 // . else (oop is to left of current scan pointer)
3712 // push oop on marking stack
3713 // . drain the marking stack
3714
3715 // Note that when we do a marking step we need to hold the
3716 // bit map lock -- recall that direct allocation (by mutators)
3717 // and promotion (by younger generation collectors) is also
3718 // marking the bit map. [the so-called allocate live policy.]
3719 // Because the implementation of bit map marking is not
3720 // robust wrt simultaneous marking of bits in the same word,
3721 // we need to make sure that there is no such interference
3722 // between concurrent such updates.
3723
3724 // already have locks
3725 assert_lock_strong(bitMapLock());
3726
3727 // Clear the revisit stack, just in case there are any
3728 // obsolete contents from a short-circuited previous CMS cycle.
3729 _revisitStack.reset();
3730 verify_work_stacks_empty();
3731 verify_overflow_empty();
3732 assert(_revisitStack.isEmpty(), "tabula rasa");
3733 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
3734 bool result = false;
3735 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3736 result = do_marking_mt(asynch);
3737 } else {
3738 result = do_marking_st(asynch);
3739 }
3740 return result;
3741 }
3742
3743 // Forward decl
3744 class CMSConcMarkingTask;
3745
3746 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3747 CMSCollector* _collector;
3748 CMSConcMarkingTask* _task;
3749 public:
3750 virtual void yield();
3751
3752 // "n_threads" is the number of threads to be terminated.
3753 // "queue_set" is a set of work queues of other threads.
3754 // "collector" is the CMS collector associated with this task terminator.
3755 // "yield" indicates whether we need the gang as a whole to yield.
3756 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3757 ParallelTaskTerminator(n_threads, queue_set),
3758 _collector(collector) { }
3759
3760 void set_task(CMSConcMarkingTask* task) {
3761 _task = task;
3762 }
3763 };
3764
3765 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3766 CMSConcMarkingTask* _task;
3767 public:
3768 bool should_exit_termination();
3769 void set_task(CMSConcMarkingTask* task) {
3770 _task = task;
3771 }
3772 };
3773
3774 // MT Concurrent Marking Task
3775 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3776 CMSCollector* _collector;
3777 int _n_workers; // requested/desired # workers
3778 bool _asynch;
3779 bool _result;
3780 CompactibleFreeListSpace* _cms_space;
3781 CompactibleFreeListSpace* _perm_space;
3782 char _pad_front[64]; // padding to ...
3783 HeapWord* _global_finger; // ... avoid sharing cache line
3784 char _pad_back[64];
3785 HeapWord* _restart_addr;
3786
3787 // Exposed here for yielding support
3788 Mutex* const _bit_map_lock;
3789
3790 // The per thread work queues, available here for stealing
3791 OopTaskQueueSet* _task_queues;
3792
3793 // Termination (and yielding) support
3794 CMSConcMarkingTerminator _term;
3795 CMSConcMarkingTerminatorTerminator _term_term;
3796
3797 public:
3798 CMSConcMarkingTask(CMSCollector* collector,
3799 CompactibleFreeListSpace* cms_space,
3800 CompactibleFreeListSpace* perm_space,
3801 bool asynch,
3802 YieldingFlexibleWorkGang* workers,
3803 OopTaskQueueSet* task_queues):
3804 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3805 _collector(collector),
3806 _cms_space(cms_space),
3807 _perm_space(perm_space),
3808 _asynch(asynch), _n_workers(0), _result(true),
3809 _task_queues(task_queues),
3810 _term(_n_workers, task_queues, _collector),
3811 _bit_map_lock(collector->bitMapLock())
3812 {
3813 _requested_size = _n_workers;
3814 _term.set_task(this);
3815 _term_term.set_task(this);
3816 assert(_cms_space->bottom() < _perm_space->bottom(),
3817 "Finger incorrectly initialized below");
3818 _restart_addr = _global_finger = _cms_space->bottom();
3819 }
3820
3821
3822 OopTaskQueueSet* task_queues() { return _task_queues; }
3823
3824 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3825
3826 HeapWord** global_finger_addr() { return &_global_finger; }
3827
3828 CMSConcMarkingTerminator* terminator() { return &_term; }
3829
3830 virtual void set_for_termination(int active_workers) {
3831 terminator()->reset_for_reuse(active_workers);
3832 }
3833
3834 void work(uint worker_id);
3835 bool should_yield() {
3836 return ConcurrentMarkSweepThread::should_yield()
3837 && !_collector->foregroundGCIsActive()
3838 && _asynch;
3839 }
3840
3841 virtual void coordinator_yield(); // stuff done by coordinator
3842 bool result() { return _result; }
3843
3844 void reset(HeapWord* ra) {
3845 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3846 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)");
3847 assert(ra < _perm_space->end(), "ra too large");
3848 _restart_addr = _global_finger = ra;
3849 _term.reset_for_reuse();
3850 }
3851
3852 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3853 OopTaskQueue* work_q);
3854
3855 private:
3856 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3857 void do_work_steal(int i);
3858 void bump_global_finger(HeapWord* f);
3859 };
3860
3861 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3862 assert(_task != NULL, "Error");
3863 return _task->yielding();
3864 // Note that we do not need the disjunct || _task->should_yield() above
3865 // because we want terminating threads to yield only if the task
3866 // is already in the midst of yielding, which happens only after at least one
3867 // thread has yielded.
3868 }
3869
3870 void CMSConcMarkingTerminator::yield() {
3871 if (_task->should_yield()) {
3872 _task->yield();
3873 } else {
3874 ParallelTaskTerminator::yield();
3875 }
3876 }
3877
3878 ////////////////////////////////////////////////////////////////
3879 // Concurrent Marking Algorithm Sketch
3880 ////////////////////////////////////////////////////////////////
3881 // Until all tasks exhausted (both spaces):
3882 // -- claim next available chunk
3883 // -- bump global finger via CAS
3884 // -- find first object that starts in this chunk
3885 // and start scanning bitmap from that position
3886 // -- scan marked objects for oops
3887 // -- CAS-mark target, and if successful:
3888 // . if target oop is above global finger (volatile read)
3889 // nothing to do
3890 // . if target oop is in chunk and above local finger
3891 // then nothing to do
3892 // . else push on work-queue
3893 // -- Deal with possible overflow issues:
3894 // . local work-queue overflow causes stuff to be pushed on
3895 // global (common) overflow queue
3896 // . always first empty local work queue
3897 // . then get a batch of oops from global work queue if any
3898 // . then do work stealing
3899 // -- When all tasks claimed (both spaces)
3900 // and local work queue empty,
3901 // then in a loop do:
3902 // . check global overflow stack; steal a batch of oops and trace
3903 // . try to steal from other threads oif GOS is empty
3904 // . if neither is available, offer termination
3905 // -- Terminate and return result
3906 //
3907 void CMSConcMarkingTask::work(uint worker_id) {
3908 elapsedTimer _timer;
3909 ResourceMark rm;
3910 HandleMark hm;
3911
3912 DEBUG_ONLY(_collector->verify_overflow_empty();)
3913
3914 // Before we begin work, our work queue should be empty
3915 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3916 // Scan the bitmap covering _cms_space, tracing through grey objects.
3917 _timer.start();
3918 do_scan_and_mark(worker_id, _cms_space);
3919 _timer.stop();
3920 if (PrintCMSStatistics != 0) {
3921 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
3922 worker_id, _timer.seconds());
3923 // XXX: need xxx/xxx type of notation, two timers
3924 }
3925
3926 // ... do the same for the _perm_space
3927 _timer.reset();
3928 _timer.start();
3929 do_scan_and_mark(worker_id, _perm_space);
3930 _timer.stop();
3931 if (PrintCMSStatistics != 0) {
3932 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec",
3933 worker_id, _timer.seconds());
3934 // XXX: need xxx/xxx type of notation, two timers
3935 }
3936
3937 // ... do work stealing
3938 _timer.reset();
3939 _timer.start();
3940 do_work_steal(worker_id);
3941 _timer.stop();
3942 if (PrintCMSStatistics != 0) {
3943 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
3944 worker_id, _timer.seconds());
3945 // XXX: need xxx/xxx type of notation, two timers
3946 }
3947 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3948 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3949 // Note that under the current task protocol, the
3950 // following assertion is true even of the spaces
3951 // expanded since the completion of the concurrent
3952 // marking. XXX This will likely change under a strict
3953 // ABORT semantics.
3954 assert(_global_finger > _cms_space->end() &&
3955 _global_finger >= _perm_space->end(),
3956 "All tasks have been completed");
3957 DEBUG_ONLY(_collector->verify_overflow_empty();)
3958 }
3959
3960 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3961 HeapWord* read = _global_finger;
3962 HeapWord* cur = read;
3963 while (f > read) {
3964 cur = read;
3965 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
3966 if (cur == read) {
3967 // our cas succeeded
3968 assert(_global_finger >= f, "protocol consistency");
3969 break;
3970 }
3971 }
3972 }
3973
3974 // This is really inefficient, and should be redone by
3975 // using (not yet available) block-read and -write interfaces to the
3976 // stack and the work_queue. XXX FIX ME !!!
3977 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3978 OopTaskQueue* work_q) {
3979 // Fast lock-free check
3980 if (ovflw_stk->length() == 0) {
3981 return false;
3982 }
3983 assert(work_q->size() == 0, "Shouldn't steal");
3984 MutexLockerEx ml(ovflw_stk->par_lock(),
3985 Mutex::_no_safepoint_check_flag);
3986 // Grab up to 1/4 the size of the work queue
3987 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3988 (size_t)ParGCDesiredObjsFromOverflowList);
3989 num = MIN2(num, ovflw_stk->length());
3990 for (int i = (int) num; i > 0; i--) {
3991 oop cur = ovflw_stk->pop();
3992 assert(cur != NULL, "Counted wrong?");
3993 work_q->push(cur);
3994 }
3995 return num > 0;
3996 }
3997
3998 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3999 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4000 int n_tasks = pst->n_tasks();
4001 // We allow that there may be no tasks to do here because
4002 // we are restarting after a stack overflow.
4003 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
4004 uint nth_task = 0;
4005
4006 HeapWord* aligned_start = sp->bottom();
4007 if (sp->used_region().contains(_restart_addr)) {
4008 // Align down to a card boundary for the start of 0th task
4009 // for this space.
4010 aligned_start =
4011 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
4012 CardTableModRefBS::card_size);
4013 }
4014
4015 size_t chunk_size = sp->marking_task_size();
4016 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4017 // Having claimed the nth task in this space,
4018 // compute the chunk that it corresponds to:
4019 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
4020 aligned_start + (nth_task+1)*chunk_size);
4021 // Try and bump the global finger via a CAS;
4022 // note that we need to do the global finger bump
4023 // _before_ taking the intersection below, because
4024 // the task corresponding to that region will be
4025 // deemed done even if the used_region() expands
4026 // because of allocation -- as it almost certainly will
4027 // during start-up while the threads yield in the
4028 // closure below.
4029 HeapWord* finger = span.end();
4030 bump_global_finger(finger); // atomically
4031 // There are null tasks here corresponding to chunks
4032 // beyond the "top" address of the space.
4033 span = span.intersection(sp->used_region());
4034 if (!span.is_empty()) { // Non-null task
4035 HeapWord* prev_obj;
4036 assert(!span.contains(_restart_addr) || nth_task == 0,
4037 "Inconsistency");
4038 if (nth_task == 0) {
4039 // For the 0th task, we'll not need to compute a block_start.
4040 if (span.contains(_restart_addr)) {
4041 // In the case of a restart because of stack overflow,
4042 // we might additionally skip a chunk prefix.
4043 prev_obj = _restart_addr;
4044 } else {
4045 prev_obj = span.start();
4046 }
4047 } else {
4048 // We want to skip the first object because
4049 // the protocol is to scan any object in its entirety
4050 // that _starts_ in this span; a fortiori, any
4051 // object starting in an earlier span is scanned
4052 // as part of an earlier claimed task.
4053 // Below we use the "careful" version of block_start
4054 // so we do not try to navigate uninitialized objects.
4055 prev_obj = sp->block_start_careful(span.start());
4056 // Below we use a variant of block_size that uses the
4057 // Printezis bits to avoid waiting for allocated
4058 // objects to become initialized/parsable.
4059 while (prev_obj < span.start()) {
4060 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4061 if (sz > 0) {
4062 prev_obj += sz;
4063 } else {
4064 // In this case we may end up doing a bit of redundant
4065 // scanning, but that appears unavoidable, short of
4066 // locking the free list locks; see bug 6324141.
4067 break;
4068 }
4069 }
4070 }
4071 if (prev_obj < span.end()) {
4072 MemRegion my_span = MemRegion(prev_obj, span.end());
4073 // Do the marking work within a non-empty span --
4074 // the last argument to the constructor indicates whether the
4075 // iteration should be incremental with periodic yields.
4076 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4077 &_collector->_markBitMap,
4078 work_queue(i),
4079 &_collector->_markStack,
4080 &_collector->_revisitStack,
4081 _asynch);
4082 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4083 } // else nothing to do for this task
4084 } // else nothing to do for this task
4085 }
4086 // We'd be tempted to assert here that since there are no
4087 // more tasks left to claim in this space, the global_finger
4088 // must exceed space->top() and a fortiori space->end(). However,
4089 // that would not quite be correct because the bumping of
4090 // global_finger occurs strictly after the claiming of a task,
4091 // so by the time we reach here the global finger may not yet
4092 // have been bumped up by the thread that claimed the last
4093 // task.
4094 pst->all_tasks_completed();
4095 }
4096
4097 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
4098 private:
4099 CMSConcMarkingTask* _task;
4100 MemRegion _span;
4101 CMSBitMap* _bit_map;
4102 CMSMarkStack* _overflow_stack;
4103 OopTaskQueue* _work_queue;
4104 protected:
4105 DO_OOP_WORK_DEFN
4106 public:
4107 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4108 CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
4109 CMSMarkStack* revisit_stack):
4110 Par_KlassRememberingOopClosure(collector, collector->ref_processor(), revisit_stack),
4111 _task(task),
4112 _span(collector->_span),
4113 _work_queue(work_queue),
4114 _bit_map(bit_map),
4115 _overflow_stack(overflow_stack)
4116 { }
4117 virtual void do_oop(oop* p);
4118 virtual void do_oop(narrowOop* p);
4119 void trim_queue(size_t max);
4120 void handle_stack_overflow(HeapWord* lost);
4121 void do_yield_check() {
4122 if (_task->should_yield()) {
4123 _task->yield();
4124 }
4125 }
4126 };
4127
4128 // Grey object scanning during work stealing phase --
4129 // the salient assumption here is that any references
4130 // that are in these stolen objects being scanned must
4131 // already have been initialized (else they would not have
4132 // been published), so we do not need to check for
4133 // uninitialized objects before pushing here.
4134 void Par_ConcMarkingClosure::do_oop(oop obj) {
4135 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4136 HeapWord* addr = (HeapWord*)obj;
4137 // Check if oop points into the CMS generation
4138 // and is not marked
4139 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4140 // a white object ...
4141 // If we manage to "claim" the object, by being the
4142 // first thread to mark it, then we push it on our
4143 // marking stack
4144 if (_bit_map->par_mark(addr)) { // ... now grey
4145 // push on work queue (grey set)
4146 bool simulate_overflow = false;
4147 NOT_PRODUCT(
4148 if (CMSMarkStackOverflowALot &&
4149 _collector->simulate_overflow()) {
4150 // simulate a stack overflow
4151 simulate_overflow = true;
4152 }
4153 )
4154 if (simulate_overflow ||
4155 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4156 // stack overflow
4157 if (PrintCMSStatistics != 0) {
4158 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4159 SIZE_FORMAT, _overflow_stack->capacity());
4160 }
4161 // We cannot assert that the overflow stack is full because
4162 // it may have been emptied since.
4163 assert(simulate_overflow ||
4164 _work_queue->size() == _work_queue->max_elems(),
4165 "Else push should have succeeded");
4166 handle_stack_overflow(addr);
4167 }
4168 } // Else, some other thread got there first
4169 do_yield_check();
4170 }
4171 }
4172
4173 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4174 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4175
4176 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4177 while (_work_queue->size() > max) {
4178 oop new_oop;
4179 if (_work_queue->pop_local(new_oop)) {
4180 assert(new_oop->is_oop(), "Should be an oop");
4181 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4182 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4183 assert(new_oop->is_parsable(), "Should be parsable");
4184 new_oop->oop_iterate(this); // do_oop() above
4185 do_yield_check();
4186 }
4187 }
4188 }
4189
4190 // Upon stack overflow, we discard (part of) the stack,
4191 // remembering the least address amongst those discarded
4192 // in CMSCollector's _restart_address.
4193 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4194 // We need to do this under a mutex to prevent other
4195 // workers from interfering with the work done below.
4196 MutexLockerEx ml(_overflow_stack->par_lock(),
4197 Mutex::_no_safepoint_check_flag);
4198 // Remember the least grey address discarded
4199 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4200 _collector->lower_restart_addr(ra);
4201 _overflow_stack->reset(); // discard stack contents
4202 _overflow_stack->expand(); // expand the stack if possible
4203 }
4204
4205
4206 void CMSConcMarkingTask::do_work_steal(int i) {
4207 OopTaskQueue* work_q = work_queue(i);
4208 oop obj_to_scan;
4209 CMSBitMap* bm = &(_collector->_markBitMap);
4210 CMSMarkStack* ovflw = &(_collector->_markStack);
4211 CMSMarkStack* revisit = &(_collector->_revisitStack);
4212 int* seed = _collector->hash_seed(i);
4213 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
4214 while (true) {
4215 cl.trim_queue(0);
4216 assert(work_q->size() == 0, "Should have been emptied above");
4217 if (get_work_from_overflow_stack(ovflw, work_q)) {
4218 // Can't assert below because the work obtained from the
4219 // overflow stack may already have been stolen from us.
4220 // assert(work_q->size() > 0, "Work from overflow stack");
4221 continue;
4222 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4223 assert(obj_to_scan->is_oop(), "Should be an oop");
4224 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4225 obj_to_scan->oop_iterate(&cl);
4226 } else if (terminator()->offer_termination(&_term_term)) {
4227 assert(work_q->size() == 0, "Impossible!");
4228 break;
4229 } else if (yielding() || should_yield()) {
4230 yield();
4231 }
4232 }
4233 }
4234
4235 // This is run by the CMS (coordinator) thread.
4236 void CMSConcMarkingTask::coordinator_yield() {
4237 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4238 "CMS thread should hold CMS token");
4239 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4240 // First give up the locks, then yield, then re-lock
4241 // We should probably use a constructor/destructor idiom to
4242 // do this unlock/lock or modify the MutexUnlocker class to
4243 // serve our purpose. XXX
4244 assert_lock_strong(_bit_map_lock);
4245 _bit_map_lock->unlock();
4246 ConcurrentMarkSweepThread::desynchronize(true);
4247 ConcurrentMarkSweepThread::acknowledge_yield_request();
4248 _collector->stopTimer();
4249 if (PrintCMSStatistics != 0) {
4250 _collector->incrementYields();
4251 }
4252 _collector->icms_wait();
4253
4254 // It is possible for whichever thread initiated the yield request
4255 // not to get a chance to wake up and take the bitmap lock between
4256 // this thread releasing it and reacquiring it. So, while the
4257 // should_yield() flag is on, let's sleep for a bit to give the
4258 // other thread a chance to wake up. The limit imposed on the number
4259 // of iterations is defensive, to avoid any unforseen circumstances
4260 // putting us into an infinite loop. Since it's always been this
4261 // (coordinator_yield()) method that was observed to cause the
4262 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4263 // which is by default non-zero. For the other seven methods that
4264 // also perform the yield operation, as are using a different
4265 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4266 // can enable the sleeping for those methods too, if necessary.
4267 // See 6442774.
4268 //
4269 // We really need to reconsider the synchronization between the GC
4270 // thread and the yield-requesting threads in the future and we
4271 // should really use wait/notify, which is the recommended
4272 // way of doing this type of interaction. Additionally, we should
4273 // consolidate the eight methods that do the yield operation and they
4274 // are almost identical into one for better maintenability and
4275 // readability. See 6445193.
4276 //
4277 // Tony 2006.06.29
4278 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4279 ConcurrentMarkSweepThread::should_yield() &&
4280 !CMSCollector::foregroundGCIsActive(); ++i) {
4281 os::sleep(Thread::current(), 1, false);
4282 ConcurrentMarkSweepThread::acknowledge_yield_request();
4283 }
4284
4285 ConcurrentMarkSweepThread::synchronize(true);
4286 _bit_map_lock->lock_without_safepoint_check();
4287 _collector->startTimer();
4288 }
4289
4290 bool CMSCollector::do_marking_mt(bool asynch) {
4291 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4292 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4293 conc_workers()->total_workers(),
4294 conc_workers()->active_workers(),
4295 Threads::number_of_non_daemon_threads());
4296 conc_workers()->set_active_workers(num_workers);
4297
4298 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4299 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
4300
4301 CMSConcMarkingTask tsk(this,
4302 cms_space,
4303 perm_space,
4304 asynch,
4305 conc_workers(),
4306 task_queues());
4307
4308 // Since the actual number of workers we get may be different
4309 // from the number we requested above, do we need to do anything different
4310 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4311 // class?? XXX
4312 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4313 perm_space->initialize_sequential_subtasks_for_marking(num_workers);
4314
4315 // Refs discovery is already non-atomic.
4316 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4317 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4318 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());)
4319 conc_workers()->start_task(&tsk);
4320 while (tsk.yielded()) {
4321 tsk.coordinator_yield();
4322 conc_workers()->continue_task(&tsk);
4323 }
4324 // If the task was aborted, _restart_addr will be non-NULL
4325 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4326 while (_restart_addr != NULL) {
4327 // XXX For now we do not make use of ABORTED state and have not
4328 // yet implemented the right abort semantics (even in the original
4329 // single-threaded CMS case). That needs some more investigation
4330 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4331 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4332 // If _restart_addr is non-NULL, a marking stack overflow
4333 // occurred; we need to do a fresh marking iteration from the
4334 // indicated restart address.
4335 if (_foregroundGCIsActive && asynch) {
4336 // We may be running into repeated stack overflows, having
4337 // reached the limit of the stack size, while making very
4338 // slow forward progress. It may be best to bail out and
4339 // let the foreground collector do its job.
4340 // Clear _restart_addr, so that foreground GC
4341 // works from scratch. This avoids the headache of
4342 // a "rescan" which would otherwise be needed because
4343 // of the dirty mod union table & card table.
4344 _restart_addr = NULL;
4345 return false;
4346 }
4347 // Adjust the task to restart from _restart_addr
4348 tsk.reset(_restart_addr);
4349 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4350 _restart_addr);
4351 perm_space->initialize_sequential_subtasks_for_marking(num_workers,
4352 _restart_addr);
4353 _restart_addr = NULL;
4354 // Get the workers going again
4355 conc_workers()->start_task(&tsk);
4356 while (tsk.yielded()) {
4357 tsk.coordinator_yield();
4358 conc_workers()->continue_task(&tsk);
4359 }
4360 }
4361 assert(tsk.completed(), "Inconsistency");
4362 assert(tsk.result() == true, "Inconsistency");
4363 return true;
4364 }
4365
4366 bool CMSCollector::do_marking_st(bool asynch) {
4367 ResourceMark rm;
4368 HandleMark hm;
4369
4370 // Temporarily make refs discovery single threaded (non-MT)
4371 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4372 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4373 &_markStack, &_revisitStack, CMSYield && asynch);
4374 // the last argument to iterate indicates whether the iteration
4375 // should be incremental with periodic yields.
4376 _markBitMap.iterate(&markFromRootsClosure);
4377 // If _restart_addr is non-NULL, a marking stack overflow
4378 // occurred; we need to do a fresh iteration from the
4379 // indicated restart address.
4380 while (_restart_addr != NULL) {
4381 if (_foregroundGCIsActive && asynch) {
4382 // We may be running into repeated stack overflows, having
4383 // reached the limit of the stack size, while making very
4384 // slow forward progress. It may be best to bail out and
4385 // let the foreground collector do its job.
4386 // Clear _restart_addr, so that foreground GC
4387 // works from scratch. This avoids the headache of
4388 // a "rescan" which would otherwise be needed because
4389 // of the dirty mod union table & card table.
4390 _restart_addr = NULL;
4391 return false; // indicating failure to complete marking
4392 }
4393 // Deal with stack overflow:
4394 // we restart marking from _restart_addr
4395 HeapWord* ra = _restart_addr;
4396 markFromRootsClosure.reset(ra);
4397 _restart_addr = NULL;
4398 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4399 }
4400 return true;
4401 }
4402
4403 void CMSCollector::preclean() {
4404 check_correct_thread_executing();
4405 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4406 verify_work_stacks_empty();
4407 verify_overflow_empty();
4408 _abort_preclean = false;
4409 if (CMSPrecleaningEnabled) {
4410 _eden_chunk_index = 0;
4411 size_t used = get_eden_used();
4412 size_t capacity = get_eden_capacity();
4413 // Don't start sampling unless we will get sufficiently
4414 // many samples.
4415 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4416 * CMSScheduleRemarkEdenPenetration)) {
4417 _start_sampling = true;
4418 } else {
4419 _start_sampling = false;
4420 }
4421 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4422 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4423 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4424 }
4425 CMSTokenSync x(true); // is cms thread
4426 if (CMSPrecleaningEnabled) {
4427 sample_eden();
4428 _collectorState = AbortablePreclean;
4429 } else {
4430 _collectorState = FinalMarking;
4431 }
4432 verify_work_stacks_empty();
4433 verify_overflow_empty();
4434 }
4435
4436 // Try and schedule the remark such that young gen
4437 // occupancy is CMSScheduleRemarkEdenPenetration %.
4438 void CMSCollector::abortable_preclean() {
4439 check_correct_thread_executing();
4440 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4441 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4442
4443 // If Eden's current occupancy is below this threshold,
4444 // immediately schedule the remark; else preclean
4445 // past the next scavenge in an effort to
4446 // schedule the pause as described avove. By choosing
4447 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4448 // we will never do an actual abortable preclean cycle.
4449 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4450 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4451 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4452 // We need more smarts in the abortable preclean
4453 // loop below to deal with cases where allocation
4454 // in young gen is very very slow, and our precleaning
4455 // is running a losing race against a horde of
4456 // mutators intent on flooding us with CMS updates
4457 // (dirty cards).
4458 // One, admittedly dumb, strategy is to give up
4459 // after a certain number of abortable precleaning loops
4460 // or after a certain maximum time. We want to make
4461 // this smarter in the next iteration.
4462 // XXX FIX ME!!! YSR
4463 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4464 while (!(should_abort_preclean() ||
4465 ConcurrentMarkSweepThread::should_terminate())) {
4466 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4467 cumworkdone += workdone;
4468 loops++;
4469 // Voluntarily terminate abortable preclean phase if we have
4470 // been at it for too long.
4471 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4472 loops >= CMSMaxAbortablePrecleanLoops) {
4473 if (PrintGCDetails) {
4474 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4475 }
4476 break;
4477 }
4478 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4479 if (PrintGCDetails) {
4480 gclog_or_tty->print(" CMS: abort preclean due to time ");
4481 }
4482 break;
4483 }
4484 // If we are doing little work each iteration, we should
4485 // take a short break.
4486 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4487 // Sleep for some time, waiting for work to accumulate
4488 stopTimer();
4489 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4490 startTimer();
4491 waited++;
4492 }
4493 }
4494 if (PrintCMSStatistics > 0) {
4495 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4496 loops, waited, cumworkdone);
4497 }
4498 }
4499 CMSTokenSync x(true); // is cms thread
4500 if (_collectorState != Idling) {
4501 assert(_collectorState == AbortablePreclean,
4502 "Spontaneous state transition?");
4503 _collectorState = FinalMarking;
4504 } // Else, a foreground collection completed this CMS cycle.
4505 return;
4506 }
4507
4508 // Respond to an Eden sampling opportunity
4509 void CMSCollector::sample_eden() {
4510 // Make sure a young gc cannot sneak in between our
4511 // reading and recording of a sample.
4512 assert(Thread::current()->is_ConcurrentGC_thread(),
4513 "Only the cms thread may collect Eden samples");
4514 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4515 "Should collect samples while holding CMS token");
4516 if (!_start_sampling) {
4517 return;
4518 }
4519 if (_eden_chunk_array) {
4520 if (_eden_chunk_index < _eden_chunk_capacity) {
4521 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4522 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4523 "Unexpected state of Eden");
4524 // We'd like to check that what we just sampled is an oop-start address;
4525 // however, we cannot do that here since the object may not yet have been
4526 // initialized. So we'll instead do the check when we _use_ this sample
4527 // later.
4528 if (_eden_chunk_index == 0 ||
4529 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4530 _eden_chunk_array[_eden_chunk_index-1])
4531 >= CMSSamplingGrain)) {
4532 _eden_chunk_index++; // commit sample
4533 }
4534 }
4535 }
4536 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4537 size_t used = get_eden_used();
4538 size_t capacity = get_eden_capacity();
4539 assert(used <= capacity, "Unexpected state of Eden");
4540 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4541 _abort_preclean = true;
4542 }
4543 }
4544 }
4545
4546
4547 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4548 assert(_collectorState == Precleaning ||
4549 _collectorState == AbortablePreclean, "incorrect state");
4550 ResourceMark rm;
4551 HandleMark hm;
4552
4553 // Precleaning is currently not MT but the reference processor
4554 // may be set for MT. Disable it temporarily here.
4555 ReferenceProcessor* rp = ref_processor();
4556 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4557
4558 // Do one pass of scrubbing the discovered reference lists
4559 // to remove any reference objects with strongly-reachable
4560 // referents.
4561 if (clean_refs) {
4562 CMSPrecleanRefsYieldClosure yield_cl(this);
4563 assert(rp->span().equals(_span), "Spans should be equal");
4564 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4565 &_markStack, &_revisitStack,
4566 true /* preclean */);
4567 CMSDrainMarkingStackClosure complete_trace(this,
4568 _span, &_markBitMap, &_markStack,
4569 &keep_alive, true /* preclean */);
4570
4571 // We don't want this step to interfere with a young
4572 // collection because we don't want to take CPU
4573 // or memory bandwidth away from the young GC threads
4574 // (which may be as many as there are CPUs).
4575 // Note that we don't need to protect ourselves from
4576 // interference with mutators because they can't
4577 // manipulate the discovered reference lists nor affect
4578 // the computed reachability of the referents, the
4579 // only properties manipulated by the precleaning
4580 // of these reference lists.
4581 stopTimer();
4582 CMSTokenSyncWithLocks x(true /* is cms thread */,
4583 bitMapLock());
4584 startTimer();
4585 sample_eden();
4586
4587 // The following will yield to allow foreground
4588 // collection to proceed promptly. XXX YSR:
4589 // The code in this method may need further
4590 // tweaking for better performance and some restructuring
4591 // for cleaner interfaces.
4592 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4593 rp->preclean_discovered_references(
4594 rp->is_alive_non_header(), &keep_alive, &complete_trace,
4595 &yield_cl, should_unload_classes(), gc_timer);
4596 }
4597
4598 if (clean_survivor) { // preclean the active survivor space(s)
4599 assert(_young_gen->kind() == Generation::DefNew ||
4600 _young_gen->kind() == Generation::ParNew ||
4601 _young_gen->kind() == Generation::ASParNew,
4602 "incorrect type for cast");
4603 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4604 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4605 &_markBitMap, &_modUnionTable,
4606 &_markStack, &_revisitStack,
4607 true /* precleaning phase */);
4608 stopTimer();
4609 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4610 bitMapLock());
4611 startTimer();
4612 unsigned int before_count =
4613 GenCollectedHeap::heap()->total_collections();
4614 SurvivorSpacePrecleanClosure
4615 sss_cl(this, _span, &_markBitMap, &_markStack,
4616 &pam_cl, before_count, CMSYield);
4617 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4618 dng->from()->object_iterate_careful(&sss_cl);
4619 dng->to()->object_iterate_careful(&sss_cl);
4620 }
4621 MarkRefsIntoAndScanClosure
4622 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4623 &_markStack, &_revisitStack, this, CMSYield,
4624 true /* precleaning phase */);
4625 // CAUTION: The following closure has persistent state that may need to
4626 // be reset upon a decrease in the sequence of addresses it
4627 // processes.
4628 ScanMarkedObjectsAgainCarefullyClosure
4629 smoac_cl(this, _span,
4630 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield);
4631
4632 // Preclean dirty cards in ModUnionTable and CardTable using
4633 // appropriate convergence criterion;
4634 // repeat CMSPrecleanIter times unless we find that
4635 // we are losing.
4636 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4637 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4638 "Bad convergence multiplier");
4639 assert(CMSPrecleanThreshold >= 100,
4640 "Unreasonably low CMSPrecleanThreshold");
4641
4642 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4643 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4644 numIter < CMSPrecleanIter;
4645 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4646 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4647 if (CMSPermGenPrecleaningEnabled) {
4648 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl);
4649 }
4650 if (Verbose && PrintGCDetails) {
4651 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4652 }
4653 // Either there are very few dirty cards, so re-mark
4654 // pause will be small anyway, or our pre-cleaning isn't
4655 // that much faster than the rate at which cards are being
4656 // dirtied, so we might as well stop and re-mark since
4657 // precleaning won't improve our re-mark time by much.
4658 if (curNumCards <= CMSPrecleanThreshold ||
4659 (numIter > 0 &&
4660 (curNumCards * CMSPrecleanDenominator >
4661 lastNumCards * CMSPrecleanNumerator))) {
4662 numIter++;
4663 cumNumCards += curNumCards;
4664 break;
4665 }
4666 }
4667 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4668 if (CMSPermGenPrecleaningEnabled) {
4669 curNumCards += preclean_card_table(_permGen, &smoac_cl);
4670 }
4671 cumNumCards += curNumCards;
4672 if (PrintGCDetails && PrintCMSStatistics != 0) {
4673 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4674 curNumCards, cumNumCards, numIter);
4675 }
4676 return cumNumCards; // as a measure of useful work done
4677 }
4678
4679 // PRECLEANING NOTES:
4680 // Precleaning involves:
4681 // . reading the bits of the modUnionTable and clearing the set bits.
4682 // . For the cards corresponding to the set bits, we scan the
4683 // objects on those cards. This means we need the free_list_lock
4684 // so that we can safely iterate over the CMS space when scanning
4685 // for oops.
4686 // . When we scan the objects, we'll be both reading and setting
4687 // marks in the marking bit map, so we'll need the marking bit map.
4688 // . For protecting _collector_state transitions, we take the CGC_lock.
4689 // Note that any races in the reading of of card table entries by the
4690 // CMS thread on the one hand and the clearing of those entries by the
4691 // VM thread or the setting of those entries by the mutator threads on the
4692 // other are quite benign. However, for efficiency it makes sense to keep
4693 // the VM thread from racing with the CMS thread while the latter is
4694 // dirty card info to the modUnionTable. We therefore also use the
4695 // CGC_lock to protect the reading of the card table and the mod union
4696 // table by the CM thread.
4697 // . We run concurrently with mutator updates, so scanning
4698 // needs to be done carefully -- we should not try to scan
4699 // potentially uninitialized objects.
4700 //
4701 // Locking strategy: While holding the CGC_lock, we scan over and
4702 // reset a maximal dirty range of the mod union / card tables, then lock
4703 // the free_list_lock and bitmap lock to do a full marking, then
4704 // release these locks; and repeat the cycle. This allows for a
4705 // certain amount of fairness in the sharing of these locks between
4706 // the CMS collector on the one hand, and the VM thread and the
4707 // mutators on the other.
4708
4709 // NOTE: preclean_mod_union_table() and preclean_card_table()
4710 // further below are largely identical; if you need to modify
4711 // one of these methods, please check the other method too.
4712
4713 size_t CMSCollector::preclean_mod_union_table(
4714 ConcurrentMarkSweepGeneration* gen,
4715 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4716 verify_work_stacks_empty();
4717 verify_overflow_empty();
4718
4719 // Turn off checking for this method but turn it back on
4720 // selectively. There are yield points in this method
4721 // but it is difficult to turn the checking off just around
4722 // the yield points. It is simpler to selectively turn
4723 // it on.
4724 DEBUG_ONLY(RememberKlassesChecker mux(false);)
4725
4726 // strategy: starting with the first card, accumulate contiguous
4727 // ranges of dirty cards; clear these cards, then scan the region
4728 // covered by these cards.
4729
4730 // Since all of the MUT is committed ahead, we can just use
4731 // that, in case the generations expand while we are precleaning.
4732 // It might also be fine to just use the committed part of the
4733 // generation, but we might potentially miss cards when the
4734 // generation is rapidly expanding while we are in the midst
4735 // of precleaning.
4736 HeapWord* startAddr = gen->reserved().start();
4737 HeapWord* endAddr = gen->reserved().end();
4738
4739 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4740
4741 size_t numDirtyCards, cumNumDirtyCards;
4742 HeapWord *nextAddr, *lastAddr;
4743 for (cumNumDirtyCards = numDirtyCards = 0,
4744 nextAddr = lastAddr = startAddr;
4745 nextAddr < endAddr;
4746 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4747
4748 ResourceMark rm;
4749 HandleMark hm;
4750
4751 MemRegion dirtyRegion;
4752 {
4753 stopTimer();
4754 // Potential yield point
4755 CMSTokenSync ts(true);
4756 startTimer();
4757 sample_eden();
4758 // Get dirty region starting at nextOffset (inclusive),
4759 // simultaneously clearing it.
4760 dirtyRegion =
4761 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4762 assert(dirtyRegion.start() >= nextAddr,
4763 "returned region inconsistent?");
4764 }
4765 // Remember where the next search should begin.
4766 // The returned region (if non-empty) is a right open interval,
4767 // so lastOffset is obtained from the right end of that
4768 // interval.
4769 lastAddr = dirtyRegion.end();
4770 // Should do something more transparent and less hacky XXX
4771 numDirtyCards =
4772 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4773
4774 // We'll scan the cards in the dirty region (with periodic
4775 // yields for foreground GC as needed).
4776 if (!dirtyRegion.is_empty()) {
4777 assert(numDirtyCards > 0, "consistency check");
4778 HeapWord* stop_point = NULL;
4779 stopTimer();
4780 // Potential yield point
4781 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4782 bitMapLock());
4783 startTimer();
4784 {
4785 verify_work_stacks_empty();
4786 verify_overflow_empty();
4787 sample_eden();
4788 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4789 stop_point =
4790 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4791 }
4792 if (stop_point != NULL) {
4793 // The careful iteration stopped early either because it found an
4794 // uninitialized object, or because we were in the midst of an
4795 // "abortable preclean", which should now be aborted. Redirty
4796 // the bits corresponding to the partially-scanned or unscanned
4797 // cards. We'll either restart at the next block boundary or
4798 // abort the preclean.
4799 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) ||
4800 (_collectorState == AbortablePreclean && should_abort_preclean()),
4801 "Unparsable objects should only be in perm gen.");
4802 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4803 if (should_abort_preclean()) {
4804 break; // out of preclean loop
4805 } else {
4806 // Compute the next address at which preclean should pick up;
4807 // might need bitMapLock in order to read P-bits.
4808 lastAddr = next_card_start_after_block(stop_point);
4809 }
4810 }
4811 } else {
4812 assert(lastAddr == endAddr, "consistency check");
4813 assert(numDirtyCards == 0, "consistency check");
4814 break;
4815 }
4816 }
4817 verify_work_stacks_empty();
4818 verify_overflow_empty();
4819 return cumNumDirtyCards;
4820 }
4821
4822 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4823 // below are largely identical; if you need to modify
4824 // one of these methods, please check the other method too.
4825
4826 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4827 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4828 // strategy: it's similar to precleamModUnionTable above, in that
4829 // we accumulate contiguous ranges of dirty cards, mark these cards
4830 // precleaned, then scan the region covered by these cards.
4831 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4832 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4833
4834 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4835
4836 size_t numDirtyCards, cumNumDirtyCards;
4837 HeapWord *lastAddr, *nextAddr;
4838
4839 for (cumNumDirtyCards = numDirtyCards = 0,
4840 nextAddr = lastAddr = startAddr;
4841 nextAddr < endAddr;
4842 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4843
4844 ResourceMark rm;
4845 HandleMark hm;
4846
4847 MemRegion dirtyRegion;
4848 {
4849 // See comments in "Precleaning notes" above on why we
4850 // do this locking. XXX Could the locking overheads be
4851 // too high when dirty cards are sparse? [I don't think so.]
4852 stopTimer();
4853 CMSTokenSync x(true); // is cms thread
4854 startTimer();
4855 sample_eden();
4856 // Get and clear dirty region from card table
4857 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4858 MemRegion(nextAddr, endAddr),
4859 true,
4860 CardTableModRefBS::precleaned_card_val());
4861
4862 assert(dirtyRegion.start() >= nextAddr,
4863 "returned region inconsistent?");
4864 }
4865 lastAddr = dirtyRegion.end();
4866 numDirtyCards =
4867 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4868
4869 if (!dirtyRegion.is_empty()) {
4870 stopTimer();
4871 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4872 startTimer();
4873 sample_eden();
4874 verify_work_stacks_empty();
4875 verify_overflow_empty();
4876 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());)
4877 HeapWord* stop_point =
4878 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4879 if (stop_point != NULL) {
4880 // The careful iteration stopped early because it found an
4881 // uninitialized object. Redirty the bits corresponding to the
4882 // partially-scanned or unscanned cards, and start again at the
4883 // next block boundary.
4884 assert(CMSPermGenPrecleaningEnabled ||
4885 (_collectorState == AbortablePreclean && should_abort_preclean()),
4886 "Unparsable objects should only be in perm gen.");
4887 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4888 if (should_abort_preclean()) {
4889 break; // out of preclean loop
4890 } else {
4891 // Compute the next address at which preclean should pick up.
4892 lastAddr = next_card_start_after_block(stop_point);
4893 }
4894 }
4895 } else {
4896 break;
4897 }
4898 }
4899 verify_work_stacks_empty();
4900 verify_overflow_empty();
4901 return cumNumDirtyCards;
4902 }
4903
4904 void CMSCollector::checkpointRootsFinal(bool asynch,
4905 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4906 assert(_collectorState == FinalMarking, "incorrect state transition?");
4907 check_correct_thread_executing();
4908 // world is stopped at this checkpoint
4909 assert(SafepointSynchronize::is_at_safepoint(),
4910 "world should be stopped");
4911 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
4912
4913 verify_work_stacks_empty();
4914 verify_overflow_empty();
4915
4916 SpecializationStats::clear();
4917 if (PrintGCDetails) {
4918 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
4919 _young_gen->used() / K,
4920 _young_gen->capacity() / K);
4921 }
4922 if (asynch) {
4923 if (CMSScavengeBeforeRemark) {
4924 GenCollectedHeap* gch = GenCollectedHeap::heap();
4925 // Temporarily set flag to false, GCH->do_collection will
4926 // expect it to be false and set to true
4927 FlagSetting fl(gch->_is_gc_active, false);
4928 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
4929 PrintGCDetails && Verbose, true, _gc_timer_cm);)
4930 int level = _cmsGen->level() - 1;
4931 if (level >= 0) {
4932 gch->do_collection(true, // full (i.e. force, see below)
4933 false, // !clear_all_soft_refs
4934 0, // size
4935 false, // is_tlab
4936 level // max_level
4937 );
4938 }
4939 }
4940 FreelistLocker x(this);
4941 MutexLockerEx y(bitMapLock(),
4942 Mutex::_no_safepoint_check_flag);
4943 assert(!init_mark_was_synchronous, "but that's impossible!");
4944 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
4945 } else {
4946 // already have all the locks
4947 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
4948 init_mark_was_synchronous);
4949 }
4950 verify_work_stacks_empty();
4951 verify_overflow_empty();
4952 SpecializationStats::print();
4953 }
4954
4955 void CMSCollector::checkpointRootsFinalWork(bool asynch,
4956 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
4957
4958 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
4959
4960 assert(haveFreelistLocks(), "must have free list locks");
4961 assert_lock_strong(bitMapLock());
4962
4963 if (UseAdaptiveSizePolicy) {
4964 size_policy()->checkpoint_roots_final_begin();
4965 }
4966
4967 ResourceMark rm;
4968 HandleMark hm;
4969
4970 GenCollectedHeap* gch = GenCollectedHeap::heap();
4971
4972 if (should_unload_classes()) {
4973 CodeCache::gc_prologue();
4974 }
4975 assert(haveFreelistLocks(), "must have free list locks");
4976 assert_lock_strong(bitMapLock());
4977
4978 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());)
4979 if (!init_mark_was_synchronous) {
4980 // We might assume that we need not fill TLAB's when
4981 // CMSScavengeBeforeRemark is set, because we may have just done
4982 // a scavenge which would have filled all TLAB's -- and besides
4983 // Eden would be empty. This however may not always be the case --
4984 // for instance although we asked for a scavenge, it may not have
4985 // happened because of a JNI critical section. We probably need
4986 // a policy for deciding whether we can in that case wait until
4987 // the critical section releases and then do the remark following
4988 // the scavenge, and skip it here. In the absence of that policy,
4989 // or of an indication of whether the scavenge did indeed occur,
4990 // we cannot rely on TLAB's having been filled and must do
4991 // so here just in case a scavenge did not happen.
4992 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
4993 // Update the saved marks which may affect the root scans.
4994 gch->save_marks();
4995
4996 {
4997 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
4998
4999 // Note on the role of the mod union table:
5000 // Since the marker in "markFromRoots" marks concurrently with
5001 // mutators, it is possible for some reachable objects not to have been
5002 // scanned. For instance, an only reference to an object A was
5003 // placed in object B after the marker scanned B. Unless B is rescanned,
5004 // A would be collected. Such updates to references in marked objects
5005 // are detected via the mod union table which is the set of all cards
5006 // dirtied since the first checkpoint in this GC cycle and prior to
5007 // the most recent young generation GC, minus those cleaned up by the
5008 // concurrent precleaning.
5009 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
5010 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
5011 do_remark_parallel();
5012 } else {
5013 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
5014 _gc_timer_cm);
5015 do_remark_non_parallel();
5016 }
5017 }
5018 } else {
5019 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
5020 // The initial mark was stop-world, so there's no rescanning to
5021 // do; go straight on to the next step below.
5022 }
5023 verify_work_stacks_empty();
5024 verify_overflow_empty();
5025
5026 {
5027 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
5028 refProcessingWork(asynch, clear_all_soft_refs);
5029 }
5030 verify_work_stacks_empty();
5031 verify_overflow_empty();
5032
5033 if (should_unload_classes()) {
5034 CodeCache::gc_epilogue();
5035 }
5036 JvmtiExport::gc_epilogue();
5037
5038 // If we encountered any (marking stack / work queue) overflow
5039 // events during the current CMS cycle, take appropriate
5040 // remedial measures, where possible, so as to try and avoid
5041 // recurrence of that condition.
5042 assert(_markStack.isEmpty(), "No grey objects");
5043 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5044 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5045 if (ser_ovflw > 0) {
5046 if (PrintCMSStatistics != 0) {
5047 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5048 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5049 ", kac_preclean="SIZE_FORMAT")",
5050 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5051 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5052 }
5053 _markStack.expand();
5054 _ser_pmc_remark_ovflw = 0;
5055 _ser_pmc_preclean_ovflw = 0;
5056 _ser_kac_preclean_ovflw = 0;
5057 _ser_kac_ovflw = 0;
5058 }
5059 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5060 if (PrintCMSStatistics != 0) {
5061 gclog_or_tty->print_cr("Work queue overflow (benign) "
5062 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5063 _par_pmc_remark_ovflw, _par_kac_ovflw);
5064 }
5065 _par_pmc_remark_ovflw = 0;
5066 _par_kac_ovflw = 0;
5067 }
5068 if (PrintCMSStatistics != 0) {
5069 if (_markStack._hit_limit > 0) {
5070 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5071 _markStack._hit_limit);
5072 }
5073 if (_markStack._failed_double > 0) {
5074 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5075 " current capacity "SIZE_FORMAT,
5076 _markStack._failed_double,
5077 _markStack.capacity());
5078 }
5079 }
5080 _markStack._hit_limit = 0;
5081 _markStack._failed_double = 0;
5082
5083 // Check that all the klasses have been checked
5084 assert(_revisitStack.isEmpty(), "Not all klasses revisited");
5085
5086 if ((VerifyAfterGC || VerifyDuringGC) &&
5087 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5088 verify_after_remark();
5089 }
5090
5091 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5092
5093 // Change under the freelistLocks.
5094 _collectorState = Sweeping;
5095 // Call isAllClear() under bitMapLock
5096 assert(_modUnionTable.isAllClear(), "Should be clear by end of the"
5097 " final marking");
5098 if (UseAdaptiveSizePolicy) {
5099 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5100 }
5101 }
5102
5103 // Parallel remark task
5104 class CMSParRemarkTask: public AbstractGangTask {
5105 CMSCollector* _collector;
5106 int _n_workers;
5107 CompactibleFreeListSpace* _cms_space;
5108 CompactibleFreeListSpace* _perm_space;
5109
5110 // The per-thread work queues, available here for stealing.
5111 OopTaskQueueSet* _task_queues;
5112 ParallelTaskTerminator _term;
5113
5114 public:
5115 // A value of 0 passed to n_workers will cause the number of
5116 // workers to be taken from the active workers in the work gang.
5117 CMSParRemarkTask(CMSCollector* collector,
5118 CompactibleFreeListSpace* cms_space,
5119 CompactibleFreeListSpace* perm_space,
5120 int n_workers, FlexibleWorkGang* workers,
5121 OopTaskQueueSet* task_queues):
5122 AbstractGangTask("Rescan roots and grey objects in parallel"),
5123 _collector(collector),
5124 _cms_space(cms_space), _perm_space(perm_space),
5125 _n_workers(n_workers),
5126 _task_queues(task_queues),
5127 _term(n_workers, task_queues) { }
5128
5129 OopTaskQueueSet* task_queues() { return _task_queues; }
5130
5131 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5132
5133 ParallelTaskTerminator* terminator() { return &_term; }
5134 int n_workers() { return _n_workers; }
5135
5136 void work(uint worker_id);
5137
5138 private:
5139 // Work method in support of parallel rescan ... of young gen spaces
5140 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl,
5141 ContiguousSpace* space,
5142 HeapWord** chunk_array, size_t chunk_top);
5143
5144 // ... of dirty cards in old space
5145 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5146 Par_MarkRefsIntoAndScanClosure* cl);
5147
5148 // ... work stealing for the above
5149 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5150 };
5151
5152 // work_queue(i) is passed to the closure
5153 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5154 // also is passed to do_dirty_card_rescan_tasks() and to
5155 // do_work_steal() to select the i-th task_queue.
5156
5157 void CMSParRemarkTask::work(uint worker_id) {
5158 elapsedTimer _timer;
5159 ResourceMark rm;
5160 HandleMark hm;
5161
5162 // ---------- rescan from roots --------------
5163 _timer.start();
5164 GenCollectedHeap* gch = GenCollectedHeap::heap();
5165 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5166 _collector->_span, _collector->ref_processor(),
5167 &(_collector->_markBitMap),
5168 work_queue(worker_id), &(_collector->_revisitStack));
5169
5170 // Rescan young gen roots first since these are likely
5171 // coarsely partitioned and may, on that account, constitute
5172 // the critical path; thus, it's best to start off that
5173 // work first.
5174 // ---------- young gen roots --------------
5175 {
5176 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5177 EdenSpace* eden_space = dng->eden();
5178 ContiguousSpace* from_space = dng->from();
5179 ContiguousSpace* to_space = dng->to();
5180
5181 HeapWord** eca = _collector->_eden_chunk_array;
5182 size_t ect = _collector->_eden_chunk_index;
5183 HeapWord** sca = _collector->_survivor_chunk_array;
5184 size_t sct = _collector->_survivor_chunk_index;
5185
5186 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5187 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5188
5189 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0);
5190 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct);
5191 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect);
5192
5193 _timer.stop();
5194 if (PrintCMSStatistics != 0) {
5195 gclog_or_tty->print_cr(
5196 "Finished young gen rescan work in %dth thread: %3.3f sec",
5197 worker_id, _timer.seconds());
5198 }
5199 }
5200
5201 // ---------- remaining roots --------------
5202 _timer.reset();
5203 _timer.start();
5204 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5205 false, // yg was scanned above
5206 false, // this is parallel code
5207 true, // collecting perm gen
5208 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5209 &par_mrias_cl,
5210 true, // walk all of code cache if (so & SO_CodeCache)
5211 NULL);
5212 assert(_collector->should_unload_classes()
5213 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5214 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5215 _timer.stop();
5216 if (PrintCMSStatistics != 0) {
5217 gclog_or_tty->print_cr(
5218 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5219 worker_id, _timer.seconds());
5220 }
5221
5222 // ---------- rescan dirty cards ------------
5223 _timer.reset();
5224 _timer.start();
5225
5226 // Do the rescan tasks for each of the two spaces
5227 // (cms_space and perm_space) in turn.
5228 // "worker_id" is passed to select the task_queue for "worker_id"
5229 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
5230 do_dirty_card_rescan_tasks(_perm_space, worker_id, &par_mrias_cl);
5231 _timer.stop();
5232 if (PrintCMSStatistics != 0) {
5233 gclog_or_tty->print_cr(
5234 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5235 worker_id, _timer.seconds());
5236 }
5237
5238 // ---------- steal work from other threads ...
5239 // ---------- ... and drain overflow list.
5240 _timer.reset();
5241 _timer.start();
5242 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
5243 _timer.stop();
5244 if (PrintCMSStatistics != 0) {
5245 gclog_or_tty->print_cr(
5246 "Finished work stealing in %dth thread: %3.3f sec",
5247 worker_id, _timer.seconds());
5248 }
5249 }
5250
5251 // Note that parameter "i" is not used.
5252 void
5253 CMSParRemarkTask::do_young_space_rescan(int i,
5254 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
5255 HeapWord** chunk_array, size_t chunk_top) {
5256 // Until all tasks completed:
5257 // . claim an unclaimed task
5258 // . compute region boundaries corresponding to task claimed
5259 // using chunk_array
5260 // . par_oop_iterate(cl) over that region
5261
5262 ResourceMark rm;
5263 HandleMark hm;
5264
5265 SequentialSubTasksDone* pst = space->par_seq_tasks();
5266 assert(pst->valid(), "Uninitialized use?");
5267
5268 uint nth_task = 0;
5269 uint n_tasks = pst->n_tasks();
5270
5271 HeapWord *start, *end;
5272 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5273 // We claimed task # nth_task; compute its boundaries.
5274 if (chunk_top == 0) { // no samples were taken
5275 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5276 start = space->bottom();
5277 end = space->top();
5278 } else if (nth_task == 0) {
5279 start = space->bottom();
5280 end = chunk_array[nth_task];
5281 } else if (nth_task < (uint)chunk_top) {
5282 assert(nth_task >= 1, "Control point invariant");
5283 start = chunk_array[nth_task - 1];
5284 end = chunk_array[nth_task];
5285 } else {
5286 assert(nth_task == (uint)chunk_top, "Control point invariant");
5287 start = chunk_array[chunk_top - 1];
5288 end = space->top();
5289 }
5290 MemRegion mr(start, end);
5291 // Verify that mr is in space
5292 assert(mr.is_empty() || space->used_region().contains(mr),
5293 "Should be in space");
5294 // Verify that "start" is an object boundary
5295 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5296 "Should be an oop");
5297 space->par_oop_iterate(mr, cl);
5298 }
5299 pst->all_tasks_completed();
5300 }
5301
5302 void
5303 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5304 CompactibleFreeListSpace* sp, int i,
5305 Par_MarkRefsIntoAndScanClosure* cl) {
5306 // Until all tasks completed:
5307 // . claim an unclaimed task
5308 // . compute region boundaries corresponding to task claimed
5309 // . transfer dirty bits ct->mut for that region
5310 // . apply rescanclosure to dirty mut bits for that region
5311
5312 ResourceMark rm;
5313 HandleMark hm;
5314
5315 OopTaskQueue* work_q = work_queue(i);
5316 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5317 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5318 // CAUTION: This closure has state that persists across calls to
5319 // the work method dirty_range_iterate_clear() in that it has
5320 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5321 // use of that state in the imbedded UpwardsObjectClosure instance
5322 // assumes that the cards are always iterated (even if in parallel
5323 // by several threads) in monotonically increasing order per each
5324 // thread. This is true of the implementation below which picks
5325 // card ranges (chunks) in monotonically increasing order globally
5326 // and, a-fortiori, in monotonically increasing order per thread
5327 // (the latter order being a subsequence of the former).
5328 // If the work code below is ever reorganized into a more chaotic
5329 // work-partitioning form than the current "sequential tasks"
5330 // paradigm, the use of that persistent state will have to be
5331 // revisited and modified appropriately. See also related
5332 // bug 4756801 work on which should examine this code to make
5333 // sure that the changes there do not run counter to the
5334 // assumptions made here and necessary for correctness and
5335 // efficiency. Note also that this code might yield inefficient
5336 // behaviour in the case of very large objects that span one or
5337 // more work chunks. Such objects would potentially be scanned
5338 // several times redundantly. Work on 4756801 should try and
5339 // address that performance anomaly if at all possible. XXX
5340 MemRegion full_span = _collector->_span;
5341 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5342 CMSMarkStack* rs = &(_collector->_revisitStack); // shared
5343 MarkFromDirtyCardsClosure
5344 greyRescanClosure(_collector, full_span, // entire span of interest
5345 sp, bm, work_q, rs, cl);
5346
5347 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5348 assert(pst->valid(), "Uninitialized use?");
5349 uint nth_task = 0;
5350 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5351 MemRegion span = sp->used_region();
5352 HeapWord* start_addr = span.start();
5353 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5354 alignment);
5355 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5356 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5357 start_addr, "Check alignment");
5358 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5359 chunk_size, "Check alignment");
5360
5361 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5362 // Having claimed the nth_task, compute corresponding mem-region,
5363 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary).
5364 // The alignment restriction ensures that we do not need any
5365 // synchronization with other gang-workers while setting or
5366 // clearing bits in thus chunk of the MUT.
5367 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5368 start_addr + (nth_task+1)*chunk_size);
5369 // The last chunk's end might be way beyond end of the
5370 // used region. In that case pull back appropriately.
5371 if (this_span.end() > end_addr) {
5372 this_span.set_end(end_addr);
5373 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5374 }
5375 // Iterate over the dirty cards covering this chunk, marking them
5376 // precleaned, and setting the corresponding bits in the mod union
5377 // table. Since we have been careful to partition at Card and MUT-word
5378 // boundaries no synchronization is needed between parallel threads.
5379 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5380 &modUnionClosure);
5381
5382 // Having transferred these marks into the modUnionTable,
5383 // rescan the marked objects on the dirty cards in the modUnionTable.
5384 // Even if this is at a synchronous collection, the initial marking
5385 // may have been done during an asynchronous collection so there
5386 // may be dirty bits in the mod-union table.
5387 _collector->_modUnionTable.dirty_range_iterate_clear(
5388 this_span, &greyRescanClosure);
5389 _collector->_modUnionTable.verifyNoOneBitsInRange(
5390 this_span.start(),
5391 this_span.end());
5392 }
5393 pst->all_tasks_completed(); // declare that i am done
5394 }
5395
5396 // . see if we can share work_queues with ParNew? XXX
5397 void
5398 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5399 int* seed) {
5400 OopTaskQueue* work_q = work_queue(i);
5401 NOT_PRODUCT(int num_steals = 0;)
5402 oop obj_to_scan;
5403 CMSBitMap* bm = &(_collector->_markBitMap);
5404
5405 while (true) {
5406 // Completely finish any left over work from (an) earlier round(s)
5407 cl->trim_queue(0);
5408 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5409 (size_t)ParGCDesiredObjsFromOverflowList);
5410 // Now check if there's any work in the overflow list
5411 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5412 // only affects the number of attempts made to get work from the
5413 // overflow list and does not affect the number of workers. Just
5414 // pass ParallelGCThreads so this behavior is unchanged.
5415 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5416 work_q,
5417 ParallelGCThreads)) {
5418 // found something in global overflow list;
5419 // not yet ready to go stealing work from others.
5420 // We'd like to assert(work_q->size() != 0, ...)
5421 // because we just took work from the overflow list,
5422 // but of course we can't since all of that could have
5423 // been already stolen from us.
5424 // "He giveth and He taketh away."
5425 continue;
5426 }
5427 // Verify that we have no work before we resort to stealing
5428 assert(work_q->size() == 0, "Have work, shouldn't steal");
5429 // Try to steal from other queues that have work
5430 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5431 NOT_PRODUCT(num_steals++;)
5432 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5433 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5434 // Do scanning work
5435 obj_to_scan->oop_iterate(cl);
5436 // Loop around, finish this work, and try to steal some more
5437 } else if (terminator()->offer_termination()) {
5438 break; // nirvana from the infinite cycle
5439 }
5440 }
5441 NOT_PRODUCT(
5442 if (PrintCMSStatistics != 0) {
5443 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5444 }
5445 )
5446 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5447 "Else our work is not yet done");
5448 }
5449
5450 // Return a thread-local PLAB recording array, as appropriate.
5451 void* CMSCollector::get_data_recorder(int thr_num) {
5452 if (_survivor_plab_array != NULL &&
5453 (CMSPLABRecordAlways ||
5454 (_collectorState > Marking && _collectorState < FinalMarking))) {
5455 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5456 ChunkArray* ca = &_survivor_plab_array[thr_num];
5457 ca->reset(); // clear it so that fresh data is recorded
5458 return (void*) ca;
5459 } else {
5460 return NULL;
5461 }
5462 }
5463
5464 // Reset all the thread-local PLAB recording arrays
5465 void CMSCollector::reset_survivor_plab_arrays() {
5466 for (uint i = 0; i < ParallelGCThreads; i++) {
5467 _survivor_plab_array[i].reset();
5468 }
5469 }
5470
5471 // Merge the per-thread plab arrays into the global survivor chunk
5472 // array which will provide the partitioning of the survivor space
5473 // for CMS rescan.
5474 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5475 int no_of_gc_threads) {
5476 assert(_survivor_plab_array != NULL, "Error");
5477 assert(_survivor_chunk_array != NULL, "Error");
5478 assert(_collectorState == FinalMarking, "Error");
5479 for (int j = 0; j < no_of_gc_threads; j++) {
5480 _cursor[j] = 0;
5481 }
5482 HeapWord* top = surv->top();
5483 size_t i;
5484 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5485 HeapWord* min_val = top; // Higher than any PLAB address
5486 uint min_tid = 0; // position of min_val this round
5487 for (int j = 0; j < no_of_gc_threads; j++) {
5488 ChunkArray* cur_sca = &_survivor_plab_array[j];
5489 if (_cursor[j] == cur_sca->end()) {
5490 continue;
5491 }
5492 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5493 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5494 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5495 if (cur_val < min_val) {
5496 min_tid = j;
5497 min_val = cur_val;
5498 } else {
5499 assert(cur_val < top, "All recorded addresses should be less");
5500 }
5501 }
5502 // At this point min_val and min_tid are respectively
5503 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5504 // and the thread (j) that witnesses that address.
5505 // We record this address in the _survivor_chunk_array[i]
5506 // and increment _cursor[min_tid] prior to the next round i.
5507 if (min_val == top) {
5508 break;
5509 }
5510 _survivor_chunk_array[i] = min_val;
5511 _cursor[min_tid]++;
5512 }
5513 // We are all done; record the size of the _survivor_chunk_array
5514 _survivor_chunk_index = i; // exclusive: [0, i)
5515 if (PrintCMSStatistics > 0) {
5516 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5517 }
5518 // Verify that we used up all the recorded entries
5519 #ifdef ASSERT
5520 size_t total = 0;
5521 for (int j = 0; j < no_of_gc_threads; j++) {
5522 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5523 total += _cursor[j];
5524 }
5525 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5526 // Check that the merged array is in sorted order
5527 if (total > 0) {
5528 for (size_t i = 0; i < total - 1; i++) {
5529 if (PrintCMSStatistics > 0) {
5530 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5531 i, _survivor_chunk_array[i]);
5532 }
5533 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5534 "Not sorted");
5535 }
5536 }
5537 #endif // ASSERT
5538 }
5539
5540 // Set up the space's par_seq_tasks structure for work claiming
5541 // for parallel rescan of young gen.
5542 // See ParRescanTask where this is currently used.
5543 void
5544 CMSCollector::
5545 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5546 assert(n_threads > 0, "Unexpected n_threads argument");
5547 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5548
5549 // Eden space
5550 {
5551 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5552 assert(!pst->valid(), "Clobbering existing data?");
5553 // Each valid entry in [0, _eden_chunk_index) represents a task.
5554 size_t n_tasks = _eden_chunk_index + 1;
5555 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5556 // Sets the condition for completion of the subtask (how many threads
5557 // need to finish in order to be done).
5558 pst->set_n_threads(n_threads);
5559 pst->set_n_tasks((int)n_tasks);
5560 }
5561
5562 // Merge the survivor plab arrays into _survivor_chunk_array
5563 if (_survivor_plab_array != NULL) {
5564 merge_survivor_plab_arrays(dng->from(), n_threads);
5565 } else {
5566 assert(_survivor_chunk_index == 0, "Error");
5567 }
5568
5569 // To space
5570 {
5571 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5572 assert(!pst->valid(), "Clobbering existing data?");
5573 // Sets the condition for completion of the subtask (how many threads
5574 // need to finish in order to be done).
5575 pst->set_n_threads(n_threads);
5576 pst->set_n_tasks(1);
5577 assert(pst->valid(), "Error");
5578 }
5579
5580 // From space
5581 {
5582 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5583 assert(!pst->valid(), "Clobbering existing data?");
5584 size_t n_tasks = _survivor_chunk_index + 1;
5585 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5586 // Sets the condition for completion of the subtask (how many threads
5587 // need to finish in order to be done).
5588 pst->set_n_threads(n_threads);
5589 pst->set_n_tasks((int)n_tasks);
5590 assert(pst->valid(), "Error");
5591 }
5592 }
5593
5594 // Parallel version of remark
5595 void CMSCollector::do_remark_parallel() {
5596 GenCollectedHeap* gch = GenCollectedHeap::heap();
5597 FlexibleWorkGang* workers = gch->workers();
5598 assert(workers != NULL, "Need parallel worker threads.");
5599 // Choose to use the number of GC workers most recently set
5600 // into "active_workers". If active_workers is not set, set it
5601 // to ParallelGCThreads.
5602 int n_workers = workers->active_workers();
5603 if (n_workers == 0) {
5604 assert(n_workers > 0, "Should have been set during scavenge");
5605 n_workers = ParallelGCThreads;
5606 workers->set_active_workers(n_workers);
5607 }
5608 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5609 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
5610
5611 CMSParRemarkTask tsk(this,
5612 cms_space, perm_space,
5613 n_workers, workers, task_queues());
5614
5615 // Set up for parallel process_strong_roots work.
5616 gch->set_par_threads(n_workers);
5617 // We won't be iterating over the cards in the card table updating
5618 // the younger_gen cards, so we shouldn't call the following else
5619 // the verification code as well as subsequent younger_refs_iterate
5620 // code would get confused. XXX
5621 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5622
5623 // The young gen rescan work will not be done as part of
5624 // process_strong_roots (which currently doesn't knw how to
5625 // parallelize such a scan), but rather will be broken up into
5626 // a set of parallel tasks (via the sampling that the [abortable]
5627 // preclean phase did of EdenSpace, plus the [two] tasks of
5628 // scanning the [two] survivor spaces. Further fine-grain
5629 // parallelization of the scanning of the survivor spaces
5630 // themselves, and of precleaning of the younger gen itself
5631 // is deferred to the future.
5632 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5633
5634 // The dirty card rescan work is broken up into a "sequence"
5635 // of parallel tasks (per constituent space) that are dynamically
5636 // claimed by the parallel threads.
5637 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5638 perm_space->initialize_sequential_subtasks_for_rescan(n_workers);
5639
5640 // It turns out that even when we're using 1 thread, doing the work in a
5641 // separate thread causes wide variance in run times. We can't help this
5642 // in the multi-threaded case, but we special-case n=1 here to get
5643 // repeatable measurements of the 1-thread overhead of the parallel code.
5644 if (n_workers > 1) {
5645 // Make refs discovery MT-safe, if it isn't already: it may not
5646 // necessarily be so, since it's possible that we are doing
5647 // ST marking.
5648 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5649 GenCollectedHeap::StrongRootsScope srs(gch);
5650 workers->run_task(&tsk);
5651 } else {
5652 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5653 GenCollectedHeap::StrongRootsScope srs(gch);
5654 tsk.work(0);
5655 }
5656 gch->set_par_threads(0); // 0 ==> non-parallel.
5657 // restore, single-threaded for now, any preserved marks
5658 // as a result of work_q overflow
5659 restore_preserved_marks_if_any();
5660 }
5661
5662 // Non-parallel version of remark
5663 void CMSCollector::do_remark_non_parallel() {
5664 ResourceMark rm;
5665 HandleMark hm;
5666 GenCollectedHeap* gch = GenCollectedHeap::heap();
5667 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5668
5669 MarkRefsIntoAndScanClosure
5670 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
5671 &_markStack, &_revisitStack, this,
5672 false /* should_yield */, false /* not precleaning */);
5673 MarkFromDirtyCardsClosure
5674 markFromDirtyCardsClosure(this, _span,
5675 NULL, // space is set further below
5676 &_markBitMap, &_markStack, &_revisitStack,
5677 &mrias_cl);
5678 {
5679 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
5680 // Iterate over the dirty cards, setting the corresponding bits in the
5681 // mod union table.
5682 {
5683 ModUnionClosure modUnionClosure(&_modUnionTable);
5684 _ct->ct_bs()->dirty_card_iterate(
5685 _cmsGen->used_region(),
5686 &modUnionClosure);
5687 _ct->ct_bs()->dirty_card_iterate(
5688 _permGen->used_region(),
5689 &modUnionClosure);
5690 }
5691 // Having transferred these marks into the modUnionTable, we just need
5692 // to rescan the marked objects on the dirty cards in the modUnionTable.
5693 // The initial marking may have been done during an asynchronous
5694 // collection so there may be dirty bits in the mod-union table.
5695 const int alignment =
5696 CardTableModRefBS::card_size * BitsPerWord;
5697 {
5698 // ... First handle dirty cards in CMS gen
5699 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5700 MemRegion ur = _cmsGen->used_region();
5701 HeapWord* lb = ur.start();
5702 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5703 MemRegion cms_span(lb, ub);
5704 _modUnionTable.dirty_range_iterate_clear(cms_span,
5705 &markFromDirtyCardsClosure);
5706 verify_work_stacks_empty();
5707 if (PrintCMSStatistics != 0) {
5708 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5709 markFromDirtyCardsClosure.num_dirty_cards());
5710 }
5711 }
5712 {
5713 // .. and then repeat for dirty cards in perm gen
5714 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace());
5715 MemRegion ur = _permGen->used_region();
5716 HeapWord* lb = ur.start();
5717 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5718 MemRegion perm_span(lb, ub);
5719 _modUnionTable.dirty_range_iterate_clear(perm_span,
5720 &markFromDirtyCardsClosure);
5721 verify_work_stacks_empty();
5722 if (PrintCMSStatistics != 0) {
5723 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ",
5724 markFromDirtyCardsClosure.num_dirty_cards());
5725 }
5726 }
5727 }
5728 if (VerifyDuringGC &&
5729 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5730 HandleMark hm; // Discard invalid handles created during verification
5731 Universe::verify();
5732 }
5733 {
5734 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
5735
5736 verify_work_stacks_empty();
5737
5738 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5739 GenCollectedHeap::StrongRootsScope srs(gch);
5740 gch->gen_process_strong_roots(_cmsGen->level(),
5741 true, // younger gens as roots
5742 false, // use the local StrongRootsScope
5743 true, // collecting perm gen
5744 SharedHeap::ScanningOption(roots_scanning_options()),
5745 &mrias_cl,
5746 true, // walk code active on stacks
5747 NULL);
5748 assert(should_unload_classes()
5749 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5750 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5751 }
5752 verify_work_stacks_empty();
5753 // Restore evacuated mark words, if any, used for overflow list links
5754 if (!CMSOverflowEarlyRestoration) {
5755 restore_preserved_marks_if_any();
5756 }
5757 verify_overflow_empty();
5758 }
5759
5760 ////////////////////////////////////////////////////////
5761 // Parallel Reference Processing Task Proxy Class
5762 ////////////////////////////////////////////////////////
5763 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5764 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5765 CMSCollector* _collector;
5766 CMSBitMap* _mark_bit_map;
5767 const MemRegion _span;
5768 ProcessTask& _task;
5769
5770 public:
5771 CMSRefProcTaskProxy(ProcessTask& task,
5772 CMSCollector* collector,
5773 const MemRegion& span,
5774 CMSBitMap* mark_bit_map,
5775 AbstractWorkGang* workers,
5776 OopTaskQueueSet* task_queues):
5777 // XXX Should superclass AGTWOQ also know about AWG since it knows
5778 // about the task_queues used by the AWG? Then it could initialize
5779 // the terminator() object. See 6984287. The set_for_termination()
5780 // below is a temporary band-aid for the regression in 6984287.
5781 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5782 task_queues),
5783 _task(task),
5784 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5785 {
5786 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5787 "Inconsistency in _span");
5788 set_for_termination(workers->active_workers());
5789 }
5790
5791 OopTaskQueueSet* task_queues() { return queues(); }
5792
5793 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5794
5795 void do_work_steal(int i,
5796 CMSParDrainMarkingStackClosure* drain,
5797 CMSParKeepAliveClosure* keep_alive,
5798 int* seed);
5799
5800 virtual void work(uint worker_id);
5801 };
5802
5803 void CMSRefProcTaskProxy::work(uint worker_id) {
5804 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5805 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5806 _mark_bit_map,
5807 &_collector->_revisitStack,
5808 work_queue(worker_id));
5809 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5810 _mark_bit_map,
5811 &_collector->_revisitStack,
5812 work_queue(worker_id));
5813 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5814 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5815 if (_task.marks_oops_alive()) {
5816 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5817 _collector->hash_seed(worker_id));
5818 }
5819 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5820 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5821 }
5822
5823 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5824 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5825 EnqueueTask& _task;
5826
5827 public:
5828 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5829 : AbstractGangTask("Enqueue reference objects in parallel"),
5830 _task(task)
5831 { }
5832
5833 virtual void work(uint worker_id)
5834 {
5835 _task.work(worker_id);
5836 }
5837 };
5838
5839 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5840 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack,
5841 OopTaskQueue* work_queue):
5842 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
5843 _span(span),
5844 _bit_map(bit_map),
5845 _work_queue(work_queue),
5846 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue),
5847 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
5848 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5849 { }
5850
5851 // . see if we can share work_queues with ParNew? XXX
5852 void CMSRefProcTaskProxy::do_work_steal(int i,
5853 CMSParDrainMarkingStackClosure* drain,
5854 CMSParKeepAliveClosure* keep_alive,
5855 int* seed) {
5856 OopTaskQueue* work_q = work_queue(i);
5857 NOT_PRODUCT(int num_steals = 0;)
5858 oop obj_to_scan;
5859
5860 while (true) {
5861 // Completely finish any left over work from (an) earlier round(s)
5862 drain->trim_queue(0);
5863 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5864 (size_t)ParGCDesiredObjsFromOverflowList);
5865 // Now check if there's any work in the overflow list
5866 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5867 // only affects the number of attempts made to get work from the
5868 // overflow list and does not affect the number of workers. Just
5869 // pass ParallelGCThreads so this behavior is unchanged.
5870 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5871 work_q,
5872 ParallelGCThreads)) {
5873 // Found something in global overflow list;
5874 // not yet ready to go stealing work from others.
5875 // We'd like to assert(work_q->size() != 0, ...)
5876 // because we just took work from the overflow list,
5877 // but of course we can't, since all of that might have
5878 // been already stolen from us.
5879 continue;
5880 }
5881 // Verify that we have no work before we resort to stealing
5882 assert(work_q->size() == 0, "Have work, shouldn't steal");
5883 // Try to steal from other queues that have work
5884 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5885 NOT_PRODUCT(num_steals++;)
5886 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5887 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5888 // Do scanning work
5889 obj_to_scan->oop_iterate(keep_alive);
5890 // Loop around, finish this work, and try to steal some more
5891 } else if (terminator()->offer_termination()) {
5892 break; // nirvana from the infinite cycle
5893 }
5894 }
5895 NOT_PRODUCT(
5896 if (PrintCMSStatistics != 0) {
5897 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5898 }
5899 )
5900 }
5901
5902 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5903 {
5904 GenCollectedHeap* gch = GenCollectedHeap::heap();
5905 FlexibleWorkGang* workers = gch->workers();
5906 assert(workers != NULL, "Need parallel worker threads.");
5907 CMSRefProcTaskProxy rp_task(task, &_collector,
5908 _collector.ref_processor()->span(),
5909 _collector.markBitMap(),
5910 workers, _collector.task_queues());
5911 workers->run_task(&rp_task);
5912 }
5913
5914 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5915 {
5916
5917 GenCollectedHeap* gch = GenCollectedHeap::heap();
5918 FlexibleWorkGang* workers = gch->workers();
5919 assert(workers != NULL, "Need parallel worker threads.");
5920 CMSRefEnqueueTaskProxy enq_task(task);
5921 workers->run_task(&enq_task);
5922 }
5923
5924 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
5925
5926 ResourceMark rm;
5927 HandleMark hm;
5928
5929 ReferenceProcessor* rp = ref_processor();
5930 assert(rp->span().equals(_span), "Spans should be equal");
5931 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5932 // Process weak references.
5933 rp->setup_policy(clear_all_soft_refs);
5934 verify_work_stacks_empty();
5935
5936 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5937 &_markStack, &_revisitStack,
5938 false /* !preclean */);
5939 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5940 _span, &_markBitMap, &_markStack,
5941 &cmsKeepAliveClosure, false /* !preclean */);
5942 {
5943 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
5944
5945 ReferenceProcessorStats stats;
5946 if (rp->processing_is_mt()) {
5947 // Set the degree of MT here. If the discovery is done MT, there
5948 // may have been a different number of threads doing the discovery
5949 // and a different number of discovered lists may have Ref objects.
5950 // That is OK as long as the Reference lists are balanced (see
5951 // balance_all_queues() and balance_queues()).
5952 GenCollectedHeap* gch = GenCollectedHeap::heap();
5953 int active_workers = ParallelGCThreads;
5954 FlexibleWorkGang* workers = gch->workers();
5955 if (workers != NULL) {
5956 active_workers = workers->active_workers();
5957 // The expectation is that active_workers will have already
5958 // been set to a reasonable value. If it has not been set,
5959 // investigate.
5960 assert(active_workers > 0, "Should have been set during scavenge");
5961 }
5962 rp->set_active_mt_degree(active_workers);
5963 CMSRefProcTaskExecutor task_executor(*this);
5964 stats = rp->process_discovered_references(&_is_alive_closure,
5965 &cmsKeepAliveClosure,
5966 &cmsDrainMarkingStackClosure,
5967 &task_executor,
5968 _gc_timer_cm);
5969 } else {
5970 stats = rp->process_discovered_references(&_is_alive_closure,
5971 &cmsKeepAliveClosure,
5972 &cmsDrainMarkingStackClosure,
5973 NULL,
5974 _gc_timer_cm);
5975 }
5976 _gc_tracer_cm->report_gc_reference_stats(stats);
5977
5978 verify_work_stacks_empty();
5979 }
5980
5981 if (should_unload_classes()) {
5982 {
5983 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
5984
5985 // Follow SystemDictionary roots and unload classes
5986 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
5987
5988 // Follow CodeCache roots and unload any methods marked for unloading
5989 CodeCache::do_unloading(&_is_alive_closure,
5990 &cmsKeepAliveClosure,
5991 purged_class);
5992
5993 cmsDrainMarkingStackClosure.do_void();
5994 verify_work_stacks_empty();
5995
5996 // Update subklass/sibling/implementor links in KlassKlass descendants
5997 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty");
5998 oop k;
5999 while ((k = _revisitStack.pop()) != NULL) {
6000 ((Klass*)(oopDesc*)k)->follow_weak_klass_links(
6001 &_is_alive_closure,
6002 &cmsKeepAliveClosure);
6003 }
6004 assert(!ClassUnloading ||
6005 (_markStack.isEmpty() && overflow_list_is_empty()),
6006 "Should not have found new reachable objects");
6007 assert(_revisitStack.isEmpty(), "revisit stack should have been drained");
6008 cmsDrainMarkingStackClosure.do_void();
6009 verify_work_stacks_empty();
6010 }
6011
6012 {
6013 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
6014 // Clean up unreferenced symbols in symbol table.
6015 SymbolTable::unlink();
6016 }
6017 }
6018
6019 if (should_unload_classes() || !JavaObjectsInPerm) {
6020 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
6021 // Now clean up stale oops in StringTable
6022 StringTable::unlink(&_is_alive_closure);
6023 }
6024
6025 verify_work_stacks_empty();
6026 // Restore any preserved marks as a result of mark stack or
6027 // work queue overflow
6028 restore_preserved_marks_if_any(); // done single-threaded for now
6029
6030 rp->set_enqueuing_is_done(true);
6031 if (rp->processing_is_mt()) {
6032 rp->balance_all_queues();
6033 CMSRefProcTaskExecutor task_executor(*this);
6034 rp->enqueue_discovered_references(&task_executor);
6035 } else {
6036 rp->enqueue_discovered_references(NULL);
6037 }
6038 rp->verify_no_references_recorded();
6039 assert(!rp->discovery_enabled(), "should have been disabled");
6040 }
6041
6042 #ifndef PRODUCT
6043 void CMSCollector::check_correct_thread_executing() {
6044 Thread* t = Thread::current();
6045 // Only the VM thread or the CMS thread should be here.
6046 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6047 "Unexpected thread type");
6048 // If this is the vm thread, the foreground process
6049 // should not be waiting. Note that _foregroundGCIsActive is
6050 // true while the foreground collector is waiting.
6051 if (_foregroundGCShouldWait) {
6052 // We cannot be the VM thread
6053 assert(t->is_ConcurrentGC_thread(),
6054 "Should be CMS thread");
6055 } else {
6056 // We can be the CMS thread only if we are in a stop-world
6057 // phase of CMS collection.
6058 if (t->is_ConcurrentGC_thread()) {
6059 assert(_collectorState == InitialMarking ||
6060 _collectorState == FinalMarking,
6061 "Should be a stop-world phase");
6062 // The CMS thread should be holding the CMS_token.
6063 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6064 "Potential interference with concurrently "
6065 "executing VM thread");
6066 }
6067 }
6068 }
6069 #endif
6070
6071 void CMSCollector::sweep(bool asynch) {
6072 assert(_collectorState == Sweeping, "just checking");
6073 check_correct_thread_executing();
6074 verify_work_stacks_empty();
6075 verify_overflow_empty();
6076 increment_sweep_count();
6077 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6078
6079 _inter_sweep_timer.stop();
6080 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6081 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6082
6083 // PermGen verification support: If perm gen sweeping is disabled in
6084 // this cycle, we preserve the perm gen object "deadness" information
6085 // in the perm_gen_verify_bit_map. In order to do that we traverse
6086 // all blocks in perm gen and mark all dead objects.
6087 if (verifying() && !should_unload_classes()) {
6088 assert(perm_gen_verify_bit_map()->sizeInBits() != 0,
6089 "Should have already been allocated");
6090 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(),
6091 markBitMap(), perm_gen_verify_bit_map());
6092 if (asynch) {
6093 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6094 bitMapLock());
6095 _permGen->cmsSpace()->blk_iterate(&mdo);
6096 } else {
6097 // In the case of synchronous sweep, we already have
6098 // the requisite locks/tokens.
6099 _permGen->cmsSpace()->blk_iterate(&mdo);
6100 }
6101 }
6102
6103 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6104 _intra_sweep_timer.reset();
6105 _intra_sweep_timer.start();
6106 if (asynch) {
6107 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6108 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6109 // First sweep the old gen then the perm gen
6110 {
6111 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6112 bitMapLock());
6113 sweepWork(_cmsGen, asynch);
6114 }
6115
6116 // Now repeat for perm gen
6117 if (should_unload_classes()) {
6118 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(),
6119 bitMapLock());
6120 sweepWork(_permGen, asynch);
6121 }
6122
6123 // Update Universe::_heap_*_at_gc figures.
6124 // We need all the free list locks to make the abstract state
6125 // transition from Sweeping to Resetting. See detailed note
6126 // further below.
6127 {
6128 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6129 _permGen->freelistLock());
6130 // Update heap occupancy information which is used as
6131 // input to soft ref clearing policy at the next gc.
6132 Universe::update_heap_info_at_gc();
6133 _collectorState = Resizing;
6134 }
6135 } else {
6136 // already have needed locks
6137 sweepWork(_cmsGen, asynch);
6138
6139 if (should_unload_classes()) {
6140 sweepWork(_permGen, asynch);
6141 }
6142 // Update heap occupancy information which is used as
6143 // input to soft ref clearing policy at the next gc.
6144 Universe::update_heap_info_at_gc();
6145 _collectorState = Resizing;
6146 }
6147 verify_work_stacks_empty();
6148 verify_overflow_empty();
6149
6150 _intra_sweep_timer.stop();
6151 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6152
6153 _inter_sweep_timer.reset();
6154 _inter_sweep_timer.start();
6155
6156 // We need to use a monotonically non-deccreasing time in ms
6157 // or we will see time-warp warnings and os::javaTimeMillis()
6158 // does not guarantee monotonicity.
6159 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6160 update_time_of_last_gc(now);
6161
6162 // NOTE on abstract state transitions:
6163 // Mutators allocate-live and/or mark the mod-union table dirty
6164 // based on the state of the collection. The former is done in
6165 // the interval [Marking, Sweeping] and the latter in the interval
6166 // [Marking, Sweeping). Thus the transitions into the Marking state
6167 // and out of the Sweeping state must be synchronously visible
6168 // globally to the mutators.
6169 // The transition into the Marking state happens with the world
6170 // stopped so the mutators will globally see it. Sweeping is
6171 // done asynchronously by the background collector so the transition
6172 // from the Sweeping state to the Resizing state must be done
6173 // under the freelistLock (as is the check for whether to
6174 // allocate-live and whether to dirty the mod-union table).
6175 assert(_collectorState == Resizing, "Change of collector state to"
6176 " Resizing must be done under the freelistLocks (plural)");
6177
6178 // Now that sweeping has been completed, we clear
6179 // the incremental_collection_failed flag,
6180 // thus inviting a younger gen collection to promote into
6181 // this generation. If such a promotion may still fail,
6182 // the flag will be set again when a young collection is
6183 // attempted.
6184 GenCollectedHeap* gch = GenCollectedHeap::heap();
6185 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6186 gch->update_full_collections_completed(_collection_count_start);
6187 }
6188
6189 // FIX ME!!! Looks like this belongs in CFLSpace, with
6190 // CMSGen merely delegating to it.
6191 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6192 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6193 HeapWord* minAddr = _cmsSpace->bottom();
6194 HeapWord* largestAddr =
6195 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6196 if (largestAddr == NULL) {
6197 // The dictionary appears to be empty. In this case
6198 // try to coalesce at the end of the heap.
6199 largestAddr = _cmsSpace->end();
6200 }
6201 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6202 size_t nearLargestOffset =
6203 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6204 if (PrintFLSStatistics != 0) {
6205 gclog_or_tty->print_cr(
6206 "CMS: Large Block: " PTR_FORMAT ";"
6207 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6208 largestAddr,
6209 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6210 }
6211 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6212 }
6213
6214 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6215 return addr >= _cmsSpace->nearLargestChunk();
6216 }
6217
6218 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6219 return _cmsSpace->find_chunk_at_end();
6220 }
6221
6222 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6223 bool full) {
6224 // The next lower level has been collected. Gather any statistics
6225 // that are of interest at this point.
6226 if (!full && (current_level + 1) == level()) {
6227 // Gather statistics on the young generation collection.
6228 collector()->stats().record_gc0_end(used());
6229 }
6230 }
6231
6232 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6233 GenCollectedHeap* gch = GenCollectedHeap::heap();
6234 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6235 "Wrong type of heap");
6236 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6237 gch->gen_policy()->size_policy();
6238 assert(sp->is_gc_cms_adaptive_size_policy(),
6239 "Wrong type of size policy");
6240 return sp;
6241 }
6242
6243 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6244 if (PrintGCDetails && Verbose) {
6245 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6246 }
6247 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6248 _debug_collection_type =
6249 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6250 if (PrintGCDetails && Verbose) {
6251 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6252 }
6253 }
6254
6255 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6256 bool asynch) {
6257 // We iterate over the space(s) underlying this generation,
6258 // checking the mark bit map to see if the bits corresponding
6259 // to specific blocks are marked or not. Blocks that are
6260 // marked are live and are not swept up. All remaining blocks
6261 // are swept up, with coalescing on-the-fly as we sweep up
6262 // contiguous free and/or garbage blocks:
6263 // We need to ensure that the sweeper synchronizes with allocators
6264 // and stop-the-world collectors. In particular, the following
6265 // locks are used:
6266 // . CMS token: if this is held, a stop the world collection cannot occur
6267 // . freelistLock: if this is held no allocation can occur from this
6268 // generation by another thread
6269 // . bitMapLock: if this is held, no other thread can access or update
6270 //
6271
6272 // Note that we need to hold the freelistLock if we use
6273 // block iterate below; else the iterator might go awry if
6274 // a mutator (or promotion) causes block contents to change
6275 // (for instance if the allocator divvies up a block).
6276 // If we hold the free list lock, for all practical purposes
6277 // young generation GC's can't occur (they'll usually need to
6278 // promote), so we might as well prevent all young generation
6279 // GC's while we do a sweeping step. For the same reason, we might
6280 // as well take the bit map lock for the entire duration
6281
6282 // check that we hold the requisite locks
6283 assert(have_cms_token(), "Should hold cms token");
6284 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6285 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6286 "Should possess CMS token to sweep");
6287 assert_lock_strong(gen->freelistLock());
6288 assert_lock_strong(bitMapLock());
6289
6290 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6291 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6292 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6293 _inter_sweep_estimate.padded_average(),
6294 _intra_sweep_estimate.padded_average());
6295 gen->setNearLargestChunk();
6296
6297 {
6298 SweepClosure sweepClosure(this, gen, &_markBitMap,
6299 CMSYield && asynch);
6300 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6301 // We need to free-up/coalesce garbage/blocks from a
6302 // co-terminal free run. This is done in the SweepClosure
6303 // destructor; so, do not remove this scope, else the
6304 // end-of-sweep-census below will be off by a little bit.
6305 }
6306 gen->cmsSpace()->sweep_completed();
6307 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6308 if (should_unload_classes()) { // unloaded classes this cycle,
6309 _concurrent_cycles_since_last_unload = 0; // ... reset count
6310 } else { // did not unload classes,
6311 _concurrent_cycles_since_last_unload++; // ... increment count
6312 }
6313 }
6314
6315 // Reset CMS data structures (for now just the marking bit map)
6316 // preparatory for the next cycle.
6317 void CMSCollector::reset(bool asynch) {
6318 GenCollectedHeap* gch = GenCollectedHeap::heap();
6319 CMSAdaptiveSizePolicy* sp = size_policy();
6320 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6321 if (asynch) {
6322 CMSTokenSyncWithLocks ts(true, bitMapLock());
6323
6324 // If the state is not "Resetting", the foreground thread
6325 // has done a collection and the resetting.
6326 if (_collectorState != Resetting) {
6327 assert(_collectorState == Idling, "The state should only change"
6328 " because the foreground collector has finished the collection");
6329 return;
6330 }
6331
6332 // Clear the mark bitmap (no grey objects to start with)
6333 // for the next cycle.
6334 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6335 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6336
6337 HeapWord* curAddr = _markBitMap.startWord();
6338 while (curAddr < _markBitMap.endWord()) {
6339 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6340 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6341 _markBitMap.clear_large_range(chunk);
6342 if (ConcurrentMarkSweepThread::should_yield() &&
6343 !foregroundGCIsActive() &&
6344 CMSYield) {
6345 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6346 "CMS thread should hold CMS token");
6347 assert_lock_strong(bitMapLock());
6348 bitMapLock()->unlock();
6349 ConcurrentMarkSweepThread::desynchronize(true);
6350 ConcurrentMarkSweepThread::acknowledge_yield_request();
6351 stopTimer();
6352 if (PrintCMSStatistics != 0) {
6353 incrementYields();
6354 }
6355 icms_wait();
6356
6357 // See the comment in coordinator_yield()
6358 for (unsigned i = 0; i < CMSYieldSleepCount &&
6359 ConcurrentMarkSweepThread::should_yield() &&
6360 !CMSCollector::foregroundGCIsActive(); ++i) {
6361 os::sleep(Thread::current(), 1, false);
6362 ConcurrentMarkSweepThread::acknowledge_yield_request();
6363 }
6364
6365 ConcurrentMarkSweepThread::synchronize(true);
6366 bitMapLock()->lock_without_safepoint_check();
6367 startTimer();
6368 }
6369 curAddr = chunk.end();
6370 }
6371 // A successful mostly concurrent collection has been done.
6372 // Because only the full (i.e., concurrent mode failure) collections
6373 // are being measured for gc overhead limits, clean the "near" flag
6374 // and count.
6375 sp->reset_gc_overhead_limit_count();
6376 _collectorState = Idling;
6377 } else {
6378 // already have the lock
6379 assert(_collectorState == Resetting, "just checking");
6380 assert_lock_strong(bitMapLock());
6381 _markBitMap.clear_all();
6382 _collectorState = Idling;
6383 }
6384
6385 // Stop incremental mode after a cycle completes, so that any future cycles
6386 // are triggered by allocation.
6387 stop_icms();
6388
6389 NOT_PRODUCT(
6390 if (RotateCMSCollectionTypes) {
6391 _cmsGen->rotate_debug_collection_type();
6392 }
6393 )
6394
6395 register_gc_end();
6396 }
6397
6398 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6399 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6400 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6401 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
6402 TraceCollectorStats tcs(counters());
6403
6404 switch (op) {
6405 case CMS_op_checkpointRootsInitial: {
6406 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6407 checkpointRootsInitial(true); // asynch
6408 if (PrintGC) {
6409 _cmsGen->printOccupancy("initial-mark");
6410 }
6411 break;
6412 }
6413 case CMS_op_checkpointRootsFinal: {
6414 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6415 checkpointRootsFinal(true, // asynch
6416 false, // !clear_all_soft_refs
6417 false); // !init_mark_was_synchronous
6418 if (PrintGC) {
6419 _cmsGen->printOccupancy("remark");
6420 }
6421 break;
6422 }
6423 default:
6424 fatal("No such CMS_op");
6425 }
6426 }
6427
6428 #ifndef PRODUCT
6429 size_t const CMSCollector::skip_header_HeapWords() {
6430 return FreeChunk::header_size();
6431 }
6432
6433 // Try and collect here conditions that should hold when
6434 // CMS thread is exiting. The idea is that the foreground GC
6435 // thread should not be blocked if it wants to terminate
6436 // the CMS thread and yet continue to run the VM for a while
6437 // after that.
6438 void CMSCollector::verify_ok_to_terminate() const {
6439 assert(Thread::current()->is_ConcurrentGC_thread(),
6440 "should be called by CMS thread");
6441 assert(!_foregroundGCShouldWait, "should be false");
6442 // We could check here that all the various low-level locks
6443 // are not held by the CMS thread, but that is overkill; see
6444 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6445 // is checked.
6446 }
6447 #endif
6448
6449 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6450 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6451 "missing Printezis mark?");
6452 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6453 size_t size = pointer_delta(nextOneAddr + 1, addr);
6454 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6455 "alignment problem");
6456 assert(size >= 3, "Necessary for Printezis marks to work");
6457 return size;
6458 }
6459
6460 // A variant of the above (block_size_using_printezis_bits()) except
6461 // that we return 0 if the P-bits are not yet set.
6462 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6463 if (_markBitMap.isMarked(addr + 1)) {
6464 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6465 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6466 size_t size = pointer_delta(nextOneAddr + 1, addr);
6467 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6468 "alignment problem");
6469 assert(size >= 3, "Necessary for Printezis marks to work");
6470 return size;
6471 }
6472 return 0;
6473 }
6474
6475 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6476 size_t sz = 0;
6477 oop p = (oop)addr;
6478 if (p->klass_or_null() != NULL && p->is_parsable()) {
6479 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6480 } else {
6481 sz = block_size_using_printezis_bits(addr);
6482 }
6483 assert(sz > 0, "size must be nonzero");
6484 HeapWord* next_block = addr + sz;
6485 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6486 CardTableModRefBS::card_size);
6487 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6488 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6489 "must be different cards");
6490 return next_card;
6491 }
6492
6493
6494 // CMS Bit Map Wrapper /////////////////////////////////////////
6495
6496 // Construct a CMS bit map infrastructure, but don't create the
6497 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6498 // further below.
6499 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6500 _bm(),
6501 _shifter(shifter),
6502 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6503 {
6504 _bmStartWord = 0;
6505 _bmWordSize = 0;
6506 }
6507
6508 bool CMSBitMap::allocate(MemRegion mr) {
6509 _bmStartWord = mr.start();
6510 _bmWordSize = mr.word_size();
6511 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6512 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6513 if (!brs.is_reserved()) {
6514 warning("CMS bit map allocation failure");
6515 return false;
6516 }
6517 // For now we'll just commit all of the bit map up fromt.
6518 // Later on we'll try to be more parsimonious with swap.
6519 if (!_virtual_space.initialize(brs, brs.size())) {
6520 warning("CMS bit map backing store failure");
6521 return false;
6522 }
6523 assert(_virtual_space.committed_size() == brs.size(),
6524 "didn't reserve backing store for all of CMS bit map?");
6525 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6526 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6527 _bmWordSize, "inconsistency in bit map sizing");
6528 _bm.set_size(_bmWordSize >> _shifter);
6529
6530 // bm.clear(); // can we rely on getting zero'd memory? verify below
6531 assert(isAllClear(),
6532 "Expected zero'd memory from ReservedSpace constructor");
6533 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6534 "consistency check");
6535 return true;
6536 }
6537
6538 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6539 HeapWord *next_addr, *end_addr, *last_addr;
6540 assert_locked();
6541 assert(covers(mr), "out-of-range error");
6542 // XXX assert that start and end are appropriately aligned
6543 for (next_addr = mr.start(), end_addr = mr.end();
6544 next_addr < end_addr; next_addr = last_addr) {
6545 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6546 last_addr = dirty_region.end();
6547 if (!dirty_region.is_empty()) {
6548 cl->do_MemRegion(dirty_region);
6549 } else {
6550 assert(last_addr == end_addr, "program logic");
6551 return;
6552 }
6553 }
6554 }
6555
6556 #ifndef PRODUCT
6557 void CMSBitMap::assert_locked() const {
6558 CMSLockVerifier::assert_locked(lock());
6559 }
6560
6561 bool CMSBitMap::covers(MemRegion mr) const {
6562 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6563 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6564 "size inconsistency");
6565 return (mr.start() >= _bmStartWord) &&
6566 (mr.end() <= endWord());
6567 }
6568
6569 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6570 return (start >= _bmStartWord && (start + size) <= endWord());
6571 }
6572
6573 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6574 // verify that there are no 1 bits in the interval [left, right)
6575 FalseBitMapClosure falseBitMapClosure;
6576 iterate(&falseBitMapClosure, left, right);
6577 }
6578
6579 void CMSBitMap::region_invariant(MemRegion mr)
6580 {
6581 assert_locked();
6582 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6583 assert(!mr.is_empty(), "unexpected empty region");
6584 assert(covers(mr), "mr should be covered by bit map");
6585 // convert address range into offset range
6586 size_t start_ofs = heapWordToOffset(mr.start());
6587 // Make sure that end() is appropriately aligned
6588 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6589 (1 << (_shifter+LogHeapWordSize))),
6590 "Misaligned mr.end()");
6591 size_t end_ofs = heapWordToOffset(mr.end());
6592 assert(end_ofs > start_ofs, "Should mark at least one bit");
6593 }
6594
6595 #endif
6596
6597 bool CMSMarkStack::allocate(size_t size) {
6598 // allocate a stack of the requisite depth
6599 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6600 size * sizeof(oop)));
6601 if (!rs.is_reserved()) {
6602 warning("CMSMarkStack allocation failure");
6603 return false;
6604 }
6605 if (!_virtual_space.initialize(rs, rs.size())) {
6606 warning("CMSMarkStack backing store failure");
6607 return false;
6608 }
6609 assert(_virtual_space.committed_size() == rs.size(),
6610 "didn't reserve backing store for all of CMS stack?");
6611 _base = (oop*)(_virtual_space.low());
6612 _index = 0;
6613 _capacity = size;
6614 NOT_PRODUCT(_max_depth = 0);
6615 return true;
6616 }
6617
6618 // XXX FIX ME !!! In the MT case we come in here holding a
6619 // leaf lock. For printing we need to take a further lock
6620 // which has lower rank. We need to recallibrate the two
6621 // lock-ranks involved in order to be able to rpint the
6622 // messages below. (Or defer the printing to the caller.
6623 // For now we take the expedient path of just disabling the
6624 // messages for the problematic case.)
6625 void CMSMarkStack::expand() {
6626 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6627 if (_capacity == MarkStackSizeMax) {
6628 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6629 // We print a warning message only once per CMS cycle.
6630 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6631 }
6632 return;
6633 }
6634 // Double capacity if possible
6635 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6636 // Do not give up existing stack until we have managed to
6637 // get the double capacity that we desired.
6638 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6639 new_capacity * sizeof(oop)));
6640 if (rs.is_reserved()) {
6641 // Release the backing store associated with old stack
6642 _virtual_space.release();
6643 // Reinitialize virtual space for new stack
6644 if (!_virtual_space.initialize(rs, rs.size())) {
6645 fatal("Not enough swap for expanded marking stack");
6646 }
6647 _base = (oop*)(_virtual_space.low());
6648 _index = 0;
6649 _capacity = new_capacity;
6650 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6651 // Failed to double capacity, continue;
6652 // we print a detail message only once per CMS cycle.
6653 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6654 SIZE_FORMAT"K",
6655 _capacity / K, new_capacity / K);
6656 }
6657 }
6658
6659
6660 // Closures
6661 // XXX: there seems to be a lot of code duplication here;
6662 // should refactor and consolidate common code.
6663
6664 // This closure is used to mark refs into the CMS generation in
6665 // the CMS bit map. Called at the first checkpoint. This closure
6666 // assumes that we do not need to re-mark dirty cards; if the CMS
6667 // generation on which this is used is not an oldest (modulo perm gen)
6668 // generation then this will lose younger_gen cards!
6669
6670 MarkRefsIntoClosure::MarkRefsIntoClosure(
6671 MemRegion span, CMSBitMap* bitMap):
6672 _span(span),
6673 _bitMap(bitMap)
6674 {
6675 assert(_ref_processor == NULL, "deliberately left NULL");
6676 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6677 }
6678
6679 void MarkRefsIntoClosure::do_oop(oop obj) {
6680 // if p points into _span, then mark corresponding bit in _markBitMap
6681 assert(obj->is_oop(), "expected an oop");
6682 HeapWord* addr = (HeapWord*)obj;
6683 if (_span.contains(addr)) {
6684 // this should be made more efficient
6685 _bitMap->mark(addr);
6686 }
6687 }
6688
6689 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6690 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6691
6692 // A variant of the above, used for CMS marking verification.
6693 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6694 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6695 _span(span),
6696 _verification_bm(verification_bm),
6697 _cms_bm(cms_bm)
6698 {
6699 assert(_ref_processor == NULL, "deliberately left NULL");
6700 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6701 }
6702
6703 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6704 // if p points into _span, then mark corresponding bit in _markBitMap
6705 assert(obj->is_oop(), "expected an oop");
6706 HeapWord* addr = (HeapWord*)obj;
6707 if (_span.contains(addr)) {
6708 _verification_bm->mark(addr);
6709 if (!_cms_bm->isMarked(addr)) {
6710 oop(addr)->print();
6711 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6712 fatal("... aborting");
6713 }
6714 }
6715 }
6716
6717 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6718 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6719
6720 //////////////////////////////////////////////////
6721 // MarkRefsIntoAndScanClosure
6722 //////////////////////////////////////////////////
6723
6724 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6725 ReferenceProcessor* rp,
6726 CMSBitMap* bit_map,
6727 CMSBitMap* mod_union_table,
6728 CMSMarkStack* mark_stack,
6729 CMSMarkStack* revisit_stack,
6730 CMSCollector* collector,
6731 bool should_yield,
6732 bool concurrent_precleaning):
6733 _collector(collector),
6734 _span(span),
6735 _bit_map(bit_map),
6736 _mark_stack(mark_stack),
6737 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6738 mark_stack, revisit_stack, concurrent_precleaning),
6739 _yield(should_yield),
6740 _concurrent_precleaning(concurrent_precleaning),
6741 _freelistLock(NULL)
6742 {
6743 _ref_processor = rp;
6744 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6745 }
6746
6747 // This closure is used to mark refs into the CMS generation at the
6748 // second (final) checkpoint, and to scan and transitively follow
6749 // the unmarked oops. It is also used during the concurrent precleaning
6750 // phase while scanning objects on dirty cards in the CMS generation.
6751 // The marks are made in the marking bit map and the marking stack is
6752 // used for keeping the (newly) grey objects during the scan.
6753 // The parallel version (Par_...) appears further below.
6754 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6755 if (obj != NULL) {
6756 assert(obj->is_oop(), "expected an oop");
6757 HeapWord* addr = (HeapWord*)obj;
6758 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6759 assert(_collector->overflow_list_is_empty(),
6760 "overflow list should be empty");
6761 if (_span.contains(addr) &&
6762 !_bit_map->isMarked(addr)) {
6763 // mark bit map (object is now grey)
6764 _bit_map->mark(addr);
6765 // push on marking stack (stack should be empty), and drain the
6766 // stack by applying this closure to the oops in the oops popped
6767 // from the stack (i.e. blacken the grey objects)
6768 bool res = _mark_stack->push(obj);
6769 assert(res, "Should have space to push on empty stack");
6770 do {
6771 oop new_oop = _mark_stack->pop();
6772 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
6773 assert(new_oop->is_parsable(), "Found unparsable oop");
6774 assert(_bit_map->isMarked((HeapWord*)new_oop),
6775 "only grey objects on this stack");
6776 // iterate over the oops in this oop, marking and pushing
6777 // the ones in CMS heap (i.e. in _span).
6778 new_oop->oop_iterate(&_pushAndMarkClosure);
6779 // check if it's time to yield
6780 do_yield_check();
6781 } while (!_mark_stack->isEmpty() ||
6782 (!_concurrent_precleaning && take_from_overflow_list()));
6783 // if marking stack is empty, and we are not doing this
6784 // during precleaning, then check the overflow list
6785 }
6786 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6787 assert(_collector->overflow_list_is_empty(),
6788 "overflow list was drained above");
6789 // We could restore evacuated mark words, if any, used for
6790 // overflow list links here because the overflow list is
6791 // provably empty here. That would reduce the maximum
6792 // size requirements for preserved_{oop,mark}_stack.
6793 // But we'll just postpone it until we are all done
6794 // so we can just stream through.
6795 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
6796 _collector->restore_preserved_marks_if_any();
6797 assert(_collector->no_preserved_marks(), "No preserved marks");
6798 }
6799 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
6800 "All preserved marks should have been restored above");
6801 }
6802 }
6803
6804 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6805 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
6806
6807 void MarkRefsIntoAndScanClosure::do_yield_work() {
6808 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6809 "CMS thread should hold CMS token");
6810 assert_lock_strong(_freelistLock);
6811 assert_lock_strong(_bit_map->lock());
6812 // relinquish the free_list_lock and bitMaplock()
6813 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6814 _bit_map->lock()->unlock();
6815 _freelistLock->unlock();
6816 ConcurrentMarkSweepThread::desynchronize(true);
6817 ConcurrentMarkSweepThread::acknowledge_yield_request();
6818 _collector->stopTimer();
6819 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6820 if (PrintCMSStatistics != 0) {
6821 _collector->incrementYields();
6822 }
6823 _collector->icms_wait();
6824
6825 // See the comment in coordinator_yield()
6826 for (unsigned i = 0;
6827 i < CMSYieldSleepCount &&
6828 ConcurrentMarkSweepThread::should_yield() &&
6829 !CMSCollector::foregroundGCIsActive();
6830 ++i) {
6831 os::sleep(Thread::current(), 1, false);
6832 ConcurrentMarkSweepThread::acknowledge_yield_request();
6833 }
6834
6835 ConcurrentMarkSweepThread::synchronize(true);
6836 _freelistLock->lock_without_safepoint_check();
6837 _bit_map->lock()->lock_without_safepoint_check();
6838 _collector->startTimer();
6839 }
6840
6841 ///////////////////////////////////////////////////////////
6842 // Par_MarkRefsIntoAndScanClosure: a parallel version of
6843 // MarkRefsIntoAndScanClosure
6844 ///////////////////////////////////////////////////////////
6845 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
6846 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
6847 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack):
6848 _span(span),
6849 _bit_map(bit_map),
6850 _work_queue(work_queue),
6851 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6852 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6853 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue,
6854 revisit_stack)
6855 {
6856 _ref_processor = rp;
6857 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6858 }
6859
6860 // This closure is used to mark refs into the CMS generation at the
6861 // second (final) checkpoint, and to scan and transitively follow
6862 // the unmarked oops. The marks are made in the marking bit map and
6863 // the work_queue is used for keeping the (newly) grey objects during
6864 // the scan phase whence they are also available for stealing by parallel
6865 // threads. Since the marking bit map is shared, updates are
6866 // synchronized (via CAS).
6867 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6868 if (obj != NULL) {
6869 // Ignore mark word because this could be an already marked oop
6870 // that may be chained at the end of the overflow list.
6871 assert(obj->is_oop(true), "expected an oop");
6872 HeapWord* addr = (HeapWord*)obj;
6873 if (_span.contains(addr) &&
6874 !_bit_map->isMarked(addr)) {
6875 // mark bit map (object will become grey):
6876 // It is possible for several threads to be
6877 // trying to "claim" this object concurrently;
6878 // the unique thread that succeeds in marking the
6879 // object first will do the subsequent push on
6880 // to the work queue (or overflow list).
6881 if (_bit_map->par_mark(addr)) {
6882 // push on work_queue (which may not be empty), and trim the
6883 // queue to an appropriate length by applying this closure to
6884 // the oops in the oops popped from the stack (i.e. blacken the
6885 // grey objects)
6886 bool res = _work_queue->push(obj);
6887 assert(res, "Low water mark should be less than capacity?");
6888 trim_queue(_low_water_mark);
6889 } // Else, another thread claimed the object
6890 }
6891 }
6892 }
6893
6894 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6895 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
6896
6897 // This closure is used to rescan the marked objects on the dirty cards
6898 // in the mod union table and the card table proper.
6899 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6900 oop p, MemRegion mr) {
6901
6902 size_t size = 0;
6903 HeapWord* addr = (HeapWord*)p;
6904 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6905 assert(_span.contains(addr), "we are scanning the CMS generation");
6906 // check if it's time to yield
6907 if (do_yield_check()) {
6908 // We yielded for some foreground stop-world work,
6909 // and we have been asked to abort this ongoing preclean cycle.
6910 return 0;
6911 }
6912 if (_bitMap->isMarked(addr)) {
6913 // it's marked; is it potentially uninitialized?
6914 if (p->klass_or_null() != NULL) {
6915 // If is_conc_safe is false, the object may be undergoing
6916 // change by the VM outside a safepoint. Don't try to
6917 // scan it, but rather leave it for the remark phase.
6918 if (CMSPermGenPrecleaningEnabled &&
6919 (!p->is_conc_safe() || !p->is_parsable())) {
6920 // Signal precleaning to redirty the card since
6921 // the klass pointer is already installed.
6922 assert(size == 0, "Initial value");
6923 } else {
6924 assert(p->is_parsable(), "must be parsable.");
6925 // an initialized object; ignore mark word in verification below
6926 // since we are running concurrent with mutators
6927 assert(p->is_oop(true), "should be an oop");
6928 if (p->is_objArray()) {
6929 // objArrays are precisely marked; restrict scanning
6930 // to dirty cards only.
6931 size = CompactibleFreeListSpace::adjustObjectSize(
6932 p->oop_iterate(_scanningClosure, mr));
6933 } else {
6934 // A non-array may have been imprecisely marked; we need
6935 // to scan object in its entirety.
6936 size = CompactibleFreeListSpace::adjustObjectSize(
6937 p->oop_iterate(_scanningClosure));
6938 }
6939 #ifdef DEBUG
6940 size_t direct_size =
6941 CompactibleFreeListSpace::adjustObjectSize(p->size());
6942 assert(size == direct_size, "Inconsistency in size");
6943 assert(size >= 3, "Necessary for Printezis marks to work");
6944 if (!_bitMap->isMarked(addr+1)) {
6945 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
6946 } else {
6947 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
6948 assert(_bitMap->isMarked(addr+size-1),
6949 "inconsistent Printezis mark");
6950 }
6951 #endif // DEBUG
6952 }
6953 } else {
6954 // an unitialized object
6955 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6956 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6957 size = pointer_delta(nextOneAddr + 1, addr);
6958 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6959 "alignment problem");
6960 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6961 // will dirty the card when the klass pointer is installed in the
6962 // object (signalling the completion of initialization).
6963 }
6964 } else {
6965 // Either a not yet marked object or an uninitialized object
6966 if (p->klass_or_null() == NULL || !p->is_parsable()) {
6967 // An uninitialized object, skip to the next card, since
6968 // we may not be able to read its P-bits yet.
6969 assert(size == 0, "Initial value");
6970 } else {
6971 // An object not (yet) reached by marking: we merely need to
6972 // compute its size so as to go look at the next block.
6973 assert(p->is_oop(true), "should be an oop");
6974 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6975 }
6976 }
6977 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6978 return size;
6979 }
6980
6981 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6982 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6983 "CMS thread should hold CMS token");
6984 assert_lock_strong(_freelistLock);
6985 assert_lock_strong(_bitMap->lock());
6986 DEBUG_ONLY(RememberKlassesChecker mux(false);)
6987 // relinquish the free_list_lock and bitMaplock()
6988 _bitMap->lock()->unlock();
6989 _freelistLock->unlock();
6990 ConcurrentMarkSweepThread::desynchronize(true);
6991 ConcurrentMarkSweepThread::acknowledge_yield_request();
6992 _collector->stopTimer();
6993 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
6994 if (PrintCMSStatistics != 0) {
6995 _collector->incrementYields();
6996 }
6997 _collector->icms_wait();
6998
6999 // See the comment in coordinator_yield()
7000 for (unsigned i = 0; i < CMSYieldSleepCount &&
7001 ConcurrentMarkSweepThread::should_yield() &&
7002 !CMSCollector::foregroundGCIsActive(); ++i) {
7003 os::sleep(Thread::current(), 1, false);
7004 ConcurrentMarkSweepThread::acknowledge_yield_request();
7005 }
7006
7007 ConcurrentMarkSweepThread::synchronize(true);
7008 _freelistLock->lock_without_safepoint_check();
7009 _bitMap->lock()->lock_without_safepoint_check();
7010 _collector->startTimer();
7011 }
7012
7013
7014 //////////////////////////////////////////////////////////////////
7015 // SurvivorSpacePrecleanClosure
7016 //////////////////////////////////////////////////////////////////
7017 // This (single-threaded) closure is used to preclean the oops in
7018 // the survivor spaces.
7019 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
7020
7021 HeapWord* addr = (HeapWord*)p;
7022 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7023 assert(!_span.contains(addr), "we are scanning the survivor spaces");
7024 assert(p->klass_or_null() != NULL, "object should be initializd");
7025 assert(p->is_parsable(), "must be parsable.");
7026 // an initialized object; ignore mark word in verification below
7027 // since we are running concurrent with mutators
7028 assert(p->is_oop(true), "should be an oop");
7029 // Note that we do not yield while we iterate over
7030 // the interior oops of p, pushing the relevant ones
7031 // on our marking stack.
7032 size_t size = p->oop_iterate(_scanning_closure);
7033 do_yield_check();
7034 // Observe that below, we do not abandon the preclean
7035 // phase as soon as we should; rather we empty the
7036 // marking stack before returning. This is to satisfy
7037 // some existing assertions. In general, it may be a
7038 // good idea to abort immediately and complete the marking
7039 // from the grey objects at a later time.
7040 while (!_mark_stack->isEmpty()) {
7041 oop new_oop = _mark_stack->pop();
7042 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7043 assert(new_oop->is_parsable(), "Found unparsable oop");
7044 assert(_bit_map->isMarked((HeapWord*)new_oop),
7045 "only grey objects on this stack");
7046 // iterate over the oops in this oop, marking and pushing
7047 // the ones in CMS heap (i.e. in _span).
7048 new_oop->oop_iterate(_scanning_closure);
7049 // check if it's time to yield
7050 do_yield_check();
7051 }
7052 unsigned int after_count =
7053 GenCollectedHeap::heap()->total_collections();
7054 bool abort = (_before_count != after_count) ||
7055 _collector->should_abort_preclean();
7056 return abort ? 0 : size;
7057 }
7058
7059 void SurvivorSpacePrecleanClosure::do_yield_work() {
7060 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7061 "CMS thread should hold CMS token");
7062 assert_lock_strong(_bit_map->lock());
7063 DEBUG_ONLY(RememberKlassesChecker smx(false);)
7064 // Relinquish the bit map lock
7065 _bit_map->lock()->unlock();
7066 ConcurrentMarkSweepThread::desynchronize(true);
7067 ConcurrentMarkSweepThread::acknowledge_yield_request();
7068 _collector->stopTimer();
7069 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7070 if (PrintCMSStatistics != 0) {
7071 _collector->incrementYields();
7072 }
7073 _collector->icms_wait();
7074
7075 // See the comment in coordinator_yield()
7076 for (unsigned i = 0; i < CMSYieldSleepCount &&
7077 ConcurrentMarkSweepThread::should_yield() &&
7078 !CMSCollector::foregroundGCIsActive(); ++i) {
7079 os::sleep(Thread::current(), 1, false);
7080 ConcurrentMarkSweepThread::acknowledge_yield_request();
7081 }
7082
7083 ConcurrentMarkSweepThread::synchronize(true);
7084 _bit_map->lock()->lock_without_safepoint_check();
7085 _collector->startTimer();
7086 }
7087
7088 // This closure is used to rescan the marked objects on the dirty cards
7089 // in the mod union table and the card table proper. In the parallel
7090 // case, although the bitMap is shared, we do a single read so the
7091 // isMarked() query is "safe".
7092 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7093 // Ignore mark word because we are running concurrent with mutators
7094 assert(p->is_oop_or_null(true), "expected an oop or null");
7095 HeapWord* addr = (HeapWord*)p;
7096 assert(_span.contains(addr), "we are scanning the CMS generation");
7097 bool is_obj_array = false;
7098 #ifdef DEBUG
7099 if (!_parallel) {
7100 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7101 assert(_collector->overflow_list_is_empty(),
7102 "overflow list should be empty");
7103
7104 }
7105 #endif // DEBUG
7106 if (_bit_map->isMarked(addr)) {
7107 // Obj arrays are precisely marked, non-arrays are not;
7108 // so we scan objArrays precisely and non-arrays in their
7109 // entirety.
7110 if (p->is_objArray()) {
7111 is_obj_array = true;
7112 if (_parallel) {
7113 p->oop_iterate(_par_scan_closure, mr);
7114 } else {
7115 p->oop_iterate(_scan_closure, mr);
7116 }
7117 } else {
7118 if (_parallel) {
7119 p->oop_iterate(_par_scan_closure);
7120 } else {
7121 p->oop_iterate(_scan_closure);
7122 }
7123 }
7124 }
7125 #ifdef DEBUG
7126 if (!_parallel) {
7127 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7128 assert(_collector->overflow_list_is_empty(),
7129 "overflow list should be empty");
7130
7131 }
7132 #endif // DEBUG
7133 return is_obj_array;
7134 }
7135
7136 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7137 MemRegion span,
7138 CMSBitMap* bitMap, CMSMarkStack* markStack,
7139 CMSMarkStack* revisitStack,
7140 bool should_yield, bool verifying):
7141 _collector(collector),
7142 _span(span),
7143 _bitMap(bitMap),
7144 _mut(&collector->_modUnionTable),
7145 _markStack(markStack),
7146 _revisitStack(revisitStack),
7147 _yield(should_yield),
7148 _skipBits(0)
7149 {
7150 assert(_markStack->isEmpty(), "stack should be empty");
7151 _finger = _bitMap->startWord();
7152 _threshold = _finger;
7153 assert(_collector->_restart_addr == NULL, "Sanity check");
7154 assert(_span.contains(_finger), "Out of bounds _finger?");
7155 DEBUG_ONLY(_verifying = verifying;)
7156 }
7157
7158 void MarkFromRootsClosure::reset(HeapWord* addr) {
7159 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7160 assert(_span.contains(addr), "Out of bounds _finger?");
7161 _finger = addr;
7162 _threshold = (HeapWord*)round_to(
7163 (intptr_t)_finger, CardTableModRefBS::card_size);
7164 }
7165
7166 // Should revisit to see if this should be restructured for
7167 // greater efficiency.
7168 bool MarkFromRootsClosure::do_bit(size_t offset) {
7169 if (_skipBits > 0) {
7170 _skipBits--;
7171 return true;
7172 }
7173 // convert offset into a HeapWord*
7174 HeapWord* addr = _bitMap->startWord() + offset;
7175 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7176 "address out of range");
7177 assert(_bitMap->isMarked(addr), "tautology");
7178 if (_bitMap->isMarked(addr+1)) {
7179 // this is an allocated but not yet initialized object
7180 assert(_skipBits == 0, "tautology");
7181 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7182 oop p = oop(addr);
7183 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7184 DEBUG_ONLY(if (!_verifying) {)
7185 // We re-dirty the cards on which this object lies and increase
7186 // the _threshold so that we'll come back to scan this object
7187 // during the preclean or remark phase. (CMSCleanOnEnter)
7188 if (CMSCleanOnEnter) {
7189 size_t sz = _collector->block_size_using_printezis_bits(addr);
7190 HeapWord* end_card_addr = (HeapWord*)round_to(
7191 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7192 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7193 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7194 // Bump _threshold to end_card_addr; note that
7195 // _threshold cannot possibly exceed end_card_addr, anyhow.
7196 // This prevents future clearing of the card as the scan proceeds
7197 // to the right.
7198 assert(_threshold <= end_card_addr,
7199 "Because we are just scanning into this object");
7200 if (_threshold < end_card_addr) {
7201 _threshold = end_card_addr;
7202 }
7203 if (p->klass_or_null() != NULL) {
7204 // Redirty the range of cards...
7205 _mut->mark_range(redirty_range);
7206 } // ...else the setting of klass will dirty the card anyway.
7207 }
7208 DEBUG_ONLY(})
7209 return true;
7210 }
7211 }
7212 scanOopsInOop(addr);
7213 return true;
7214 }
7215
7216 // We take a break if we've been at this for a while,
7217 // so as to avoid monopolizing the locks involved.
7218 void MarkFromRootsClosure::do_yield_work() {
7219 // First give up the locks, then yield, then re-lock
7220 // We should probably use a constructor/destructor idiom to
7221 // do this unlock/lock or modify the MutexUnlocker class to
7222 // serve our purpose. XXX
7223 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7224 "CMS thread should hold CMS token");
7225 assert_lock_strong(_bitMap->lock());
7226 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7227 _bitMap->lock()->unlock();
7228 ConcurrentMarkSweepThread::desynchronize(true);
7229 ConcurrentMarkSweepThread::acknowledge_yield_request();
7230 _collector->stopTimer();
7231 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7232 if (PrintCMSStatistics != 0) {
7233 _collector->incrementYields();
7234 }
7235 _collector->icms_wait();
7236
7237 // See the comment in coordinator_yield()
7238 for (unsigned i = 0; i < CMSYieldSleepCount &&
7239 ConcurrentMarkSweepThread::should_yield() &&
7240 !CMSCollector::foregroundGCIsActive(); ++i) {
7241 os::sleep(Thread::current(), 1, false);
7242 ConcurrentMarkSweepThread::acknowledge_yield_request();
7243 }
7244
7245 ConcurrentMarkSweepThread::synchronize(true);
7246 _bitMap->lock()->lock_without_safepoint_check();
7247 _collector->startTimer();
7248 }
7249
7250 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7251 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7252 assert(_markStack->isEmpty(),
7253 "should drain stack to limit stack usage");
7254 // convert ptr to an oop preparatory to scanning
7255 oop obj = oop(ptr);
7256 // Ignore mark word in verification below, since we
7257 // may be running concurrent with mutators.
7258 assert(obj->is_oop(true), "should be an oop");
7259 assert(_finger <= ptr, "_finger runneth ahead");
7260 // advance the finger to right end of this object
7261 _finger = ptr + obj->size();
7262 assert(_finger > ptr, "we just incremented it above");
7263 // On large heaps, it may take us some time to get through
7264 // the marking phase (especially if running iCMS). During
7265 // this time it's possible that a lot of mutations have
7266 // accumulated in the card table and the mod union table --
7267 // these mutation records are redundant until we have
7268 // actually traced into the corresponding card.
7269 // Here, we check whether advancing the finger would make
7270 // us cross into a new card, and if so clear corresponding
7271 // cards in the MUT (preclean them in the card-table in the
7272 // future).
7273
7274 DEBUG_ONLY(if (!_verifying) {)
7275 // The clean-on-enter optimization is disabled by default,
7276 // until we fix 6178663.
7277 if (CMSCleanOnEnter && (_finger > _threshold)) {
7278 // [_threshold, _finger) represents the interval
7279 // of cards to be cleared in MUT (or precleaned in card table).
7280 // The set of cards to be cleared is all those that overlap
7281 // with the interval [_threshold, _finger); note that
7282 // _threshold is always kept card-aligned but _finger isn't
7283 // always card-aligned.
7284 HeapWord* old_threshold = _threshold;
7285 assert(old_threshold == (HeapWord*)round_to(
7286 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7287 "_threshold should always be card-aligned");
7288 _threshold = (HeapWord*)round_to(
7289 (intptr_t)_finger, CardTableModRefBS::card_size);
7290 MemRegion mr(old_threshold, _threshold);
7291 assert(!mr.is_empty(), "Control point invariant");
7292 assert(_span.contains(mr), "Should clear within span");
7293 // XXX When _finger crosses from old gen into perm gen
7294 // we may be doing unnecessary cleaning; do better in the
7295 // future by detecting that condition and clearing fewer
7296 // MUT/CT entries.
7297 _mut->clear_range(mr);
7298 }
7299 DEBUG_ONLY(})
7300 // Note: the finger doesn't advance while we drain
7301 // the stack below.
7302 PushOrMarkClosure pushOrMarkClosure(_collector,
7303 _span, _bitMap, _markStack,
7304 _revisitStack,
7305 _finger, this);
7306 bool res = _markStack->push(obj);
7307 assert(res, "Empty non-zero size stack should have space for single push");
7308 while (!_markStack->isEmpty()) {
7309 oop new_oop = _markStack->pop();
7310 // Skip verifying header mark word below because we are
7311 // running concurrent with mutators.
7312 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7313 // now scan this oop's oops
7314 new_oop->oop_iterate(&pushOrMarkClosure);
7315 do_yield_check();
7316 }
7317 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7318 }
7319
7320 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7321 CMSCollector* collector, MemRegion span,
7322 CMSBitMap* bit_map,
7323 OopTaskQueue* work_queue,
7324 CMSMarkStack* overflow_stack,
7325 CMSMarkStack* revisit_stack,
7326 bool should_yield):
7327 _collector(collector),
7328 _whole_span(collector->_span),
7329 _span(span),
7330 _bit_map(bit_map),
7331 _mut(&collector->_modUnionTable),
7332 _work_queue(work_queue),
7333 _overflow_stack(overflow_stack),
7334 _revisit_stack(revisit_stack),
7335 _yield(should_yield),
7336 _skip_bits(0),
7337 _task(task)
7338 {
7339 assert(_work_queue->size() == 0, "work_queue should be empty");
7340 _finger = span.start();
7341 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7342 assert(_span.contains(_finger), "Out of bounds _finger?");
7343 }
7344
7345 // Should revisit to see if this should be restructured for
7346 // greater efficiency.
7347 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7348 if (_skip_bits > 0) {
7349 _skip_bits--;
7350 return true;
7351 }
7352 // convert offset into a HeapWord*
7353 HeapWord* addr = _bit_map->startWord() + offset;
7354 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7355 "address out of range");
7356 assert(_bit_map->isMarked(addr), "tautology");
7357 if (_bit_map->isMarked(addr+1)) {
7358 // this is an allocated object that might not yet be initialized
7359 assert(_skip_bits == 0, "tautology");
7360 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7361 oop p = oop(addr);
7362 if (p->klass_or_null() == NULL || !p->is_parsable()) {
7363 // in the case of Clean-on-Enter optimization, redirty card
7364 // and avoid clearing card by increasing the threshold.
7365 return true;
7366 }
7367 }
7368 scan_oops_in_oop(addr);
7369 return true;
7370 }
7371
7372 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7373 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7374 // Should we assert that our work queue is empty or
7375 // below some drain limit?
7376 assert(_work_queue->size() == 0,
7377 "should drain stack to limit stack usage");
7378 // convert ptr to an oop preparatory to scanning
7379 oop obj = oop(ptr);
7380 // Ignore mark word in verification below, since we
7381 // may be running concurrent with mutators.
7382 assert(obj->is_oop(true), "should be an oop");
7383 assert(_finger <= ptr, "_finger runneth ahead");
7384 // advance the finger to right end of this object
7385 _finger = ptr + obj->size();
7386 assert(_finger > ptr, "we just incremented it above");
7387 // On large heaps, it may take us some time to get through
7388 // the marking phase (especially if running iCMS). During
7389 // this time it's possible that a lot of mutations have
7390 // accumulated in the card table and the mod union table --
7391 // these mutation records are redundant until we have
7392 // actually traced into the corresponding card.
7393 // Here, we check whether advancing the finger would make
7394 // us cross into a new card, and if so clear corresponding
7395 // cards in the MUT (preclean them in the card-table in the
7396 // future).
7397
7398 // The clean-on-enter optimization is disabled by default,
7399 // until we fix 6178663.
7400 if (CMSCleanOnEnter && (_finger > _threshold)) {
7401 // [_threshold, _finger) represents the interval
7402 // of cards to be cleared in MUT (or precleaned in card table).
7403 // The set of cards to be cleared is all those that overlap
7404 // with the interval [_threshold, _finger); note that
7405 // _threshold is always kept card-aligned but _finger isn't
7406 // always card-aligned.
7407 HeapWord* old_threshold = _threshold;
7408 assert(old_threshold == (HeapWord*)round_to(
7409 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7410 "_threshold should always be card-aligned");
7411 _threshold = (HeapWord*)round_to(
7412 (intptr_t)_finger, CardTableModRefBS::card_size);
7413 MemRegion mr(old_threshold, _threshold);
7414 assert(!mr.is_empty(), "Control point invariant");
7415 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7416 // XXX When _finger crosses from old gen into perm gen
7417 // we may be doing unnecessary cleaning; do better in the
7418 // future by detecting that condition and clearing fewer
7419 // MUT/CT entries.
7420 _mut->clear_range(mr);
7421 }
7422
7423 // Note: the local finger doesn't advance while we drain
7424 // the stack below, but the global finger sure can and will.
7425 HeapWord** gfa = _task->global_finger_addr();
7426 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7427 _span, _bit_map,
7428 _work_queue,
7429 _overflow_stack,
7430 _revisit_stack,
7431 _finger,
7432 gfa, this);
7433 bool res = _work_queue->push(obj); // overflow could occur here
7434 assert(res, "Will hold once we use workqueues");
7435 while (true) {
7436 oop new_oop;
7437 if (!_work_queue->pop_local(new_oop)) {
7438 // We emptied our work_queue; check if there's stuff that can
7439 // be gotten from the overflow stack.
7440 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7441 _overflow_stack, _work_queue)) {
7442 do_yield_check();
7443 continue;
7444 } else { // done
7445 break;
7446 }
7447 }
7448 // Skip verifying header mark word below because we are
7449 // running concurrent with mutators.
7450 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7451 // now scan this oop's oops
7452 new_oop->oop_iterate(&pushOrMarkClosure);
7453 do_yield_check();
7454 }
7455 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7456 }
7457
7458 // Yield in response to a request from VM Thread or
7459 // from mutators.
7460 void Par_MarkFromRootsClosure::do_yield_work() {
7461 assert(_task != NULL, "sanity");
7462 _task->yield();
7463 }
7464
7465 // A variant of the above used for verifying CMS marking work.
7466 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7467 MemRegion span,
7468 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7469 CMSMarkStack* mark_stack):
7470 _collector(collector),
7471 _span(span),
7472 _verification_bm(verification_bm),
7473 _cms_bm(cms_bm),
7474 _mark_stack(mark_stack),
7475 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7476 mark_stack)
7477 {
7478 assert(_mark_stack->isEmpty(), "stack should be empty");
7479 _finger = _verification_bm->startWord();
7480 assert(_collector->_restart_addr == NULL, "Sanity check");
7481 assert(_span.contains(_finger), "Out of bounds _finger?");
7482 }
7483
7484 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7485 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7486 assert(_span.contains(addr), "Out of bounds _finger?");
7487 _finger = addr;
7488 }
7489
7490 // Should revisit to see if this should be restructured for
7491 // greater efficiency.
7492 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7493 // convert offset into a HeapWord*
7494 HeapWord* addr = _verification_bm->startWord() + offset;
7495 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7496 "address out of range");
7497 assert(_verification_bm->isMarked(addr), "tautology");
7498 assert(_cms_bm->isMarked(addr), "tautology");
7499
7500 assert(_mark_stack->isEmpty(),
7501 "should drain stack to limit stack usage");
7502 // convert addr to an oop preparatory to scanning
7503 oop obj = oop(addr);
7504 assert(obj->is_oop(), "should be an oop");
7505 assert(_finger <= addr, "_finger runneth ahead");
7506 // advance the finger to right end of this object
7507 _finger = addr + obj->size();
7508 assert(_finger > addr, "we just incremented it above");
7509 // Note: the finger doesn't advance while we drain
7510 // the stack below.
7511 bool res = _mark_stack->push(obj);
7512 assert(res, "Empty non-zero size stack should have space for single push");
7513 while (!_mark_stack->isEmpty()) {
7514 oop new_oop = _mark_stack->pop();
7515 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7516 // now scan this oop's oops
7517 new_oop->oop_iterate(&_pam_verify_closure);
7518 }
7519 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7520 return true;
7521 }
7522
7523 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7524 CMSCollector* collector, MemRegion span,
7525 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7526 CMSMarkStack* mark_stack):
7527 OopClosure(collector->ref_processor()),
7528 _collector(collector),
7529 _span(span),
7530 _verification_bm(verification_bm),
7531 _cms_bm(cms_bm),
7532 _mark_stack(mark_stack)
7533 { }
7534
7535 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7536 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7537
7538 // Upon stack overflow, we discard (part of) the stack,
7539 // remembering the least address amongst those discarded
7540 // in CMSCollector's _restart_address.
7541 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7542 // Remember the least grey address discarded
7543 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7544 _collector->lower_restart_addr(ra);
7545 _mark_stack->reset(); // discard stack contents
7546 _mark_stack->expand(); // expand the stack if possible
7547 }
7548
7549 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7550 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7551 HeapWord* addr = (HeapWord*)obj;
7552 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7553 // Oop lies in _span and isn't yet grey or black
7554 _verification_bm->mark(addr); // now grey
7555 if (!_cms_bm->isMarked(addr)) {
7556 oop(addr)->print();
7557 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7558 addr);
7559 fatal("... aborting");
7560 }
7561
7562 if (!_mark_stack->push(obj)) { // stack overflow
7563 if (PrintCMSStatistics != 0) {
7564 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7565 SIZE_FORMAT, _mark_stack->capacity());
7566 }
7567 assert(_mark_stack->isFull(), "Else push should have succeeded");
7568 handle_stack_overflow(addr);
7569 }
7570 // anything including and to the right of _finger
7571 // will be scanned as we iterate over the remainder of the
7572 // bit map
7573 }
7574 }
7575
7576 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7577 MemRegion span,
7578 CMSBitMap* bitMap, CMSMarkStack* markStack,
7579 CMSMarkStack* revisitStack,
7580 HeapWord* finger, MarkFromRootsClosure* parent) :
7581 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack),
7582 _span(span),
7583 _bitMap(bitMap),
7584 _markStack(markStack),
7585 _finger(finger),
7586 _parent(parent)
7587 { }
7588
7589 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7590 MemRegion span,
7591 CMSBitMap* bit_map,
7592 OopTaskQueue* work_queue,
7593 CMSMarkStack* overflow_stack,
7594 CMSMarkStack* revisit_stack,
7595 HeapWord* finger,
7596 HeapWord** global_finger_addr,
7597 Par_MarkFromRootsClosure* parent) :
7598 Par_KlassRememberingOopClosure(collector,
7599 collector->ref_processor(),
7600 revisit_stack),
7601 _whole_span(collector->_span),
7602 _span(span),
7603 _bit_map(bit_map),
7604 _work_queue(work_queue),
7605 _overflow_stack(overflow_stack),
7606 _finger(finger),
7607 _global_finger_addr(global_finger_addr),
7608 _parent(parent)
7609 { }
7610
7611 // Assumes thread-safe access by callers, who are
7612 // responsible for mutual exclusion.
7613 void CMSCollector::lower_restart_addr(HeapWord* low) {
7614 assert(_span.contains(low), "Out of bounds addr");
7615 if (_restart_addr == NULL) {
7616 _restart_addr = low;
7617 } else {
7618 _restart_addr = MIN2(_restart_addr, low);
7619 }
7620 }
7621
7622 // Upon stack overflow, we discard (part of) the stack,
7623 // remembering the least address amongst those discarded
7624 // in CMSCollector's _restart_address.
7625 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7626 // Remember the least grey address discarded
7627 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7628 _collector->lower_restart_addr(ra);
7629 _markStack->reset(); // discard stack contents
7630 _markStack->expand(); // expand the stack if possible
7631 }
7632
7633 // Upon stack overflow, we discard (part of) the stack,
7634 // remembering the least address amongst those discarded
7635 // in CMSCollector's _restart_address.
7636 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7637 // We need to do this under a mutex to prevent other
7638 // workers from interfering with the work done below.
7639 MutexLockerEx ml(_overflow_stack->par_lock(),
7640 Mutex::_no_safepoint_check_flag);
7641 // Remember the least grey address discarded
7642 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7643 _collector->lower_restart_addr(ra);
7644 _overflow_stack->reset(); // discard stack contents
7645 _overflow_stack->expand(); // expand the stack if possible
7646 }
7647
7648 void PushOrMarkClosure::do_oop(oop obj) {
7649 // Ignore mark word because we are running concurrent with mutators.
7650 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7651 HeapWord* addr = (HeapWord*)obj;
7652 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7653 // Oop lies in _span and isn't yet grey or black
7654 _bitMap->mark(addr); // now grey
7655 if (addr < _finger) {
7656 // the bit map iteration has already either passed, or
7657 // sampled, this bit in the bit map; we'll need to
7658 // use the marking stack to scan this oop's oops.
7659 bool simulate_overflow = false;
7660 NOT_PRODUCT(
7661 if (CMSMarkStackOverflowALot &&
7662 _collector->simulate_overflow()) {
7663 // simulate a stack overflow
7664 simulate_overflow = true;
7665 }
7666 )
7667 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7668 if (PrintCMSStatistics != 0) {
7669 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7670 SIZE_FORMAT, _markStack->capacity());
7671 }
7672 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7673 handle_stack_overflow(addr);
7674 }
7675 }
7676 // anything including and to the right of _finger
7677 // will be scanned as we iterate over the remainder of the
7678 // bit map
7679 do_yield_check();
7680 }
7681 }
7682
7683 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7684 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7685
7686 void Par_PushOrMarkClosure::do_oop(oop obj) {
7687 // Ignore mark word because we are running concurrent with mutators.
7688 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7689 HeapWord* addr = (HeapWord*)obj;
7690 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7691 // Oop lies in _span and isn't yet grey or black
7692 // We read the global_finger (volatile read) strictly after marking oop
7693 bool res = _bit_map->par_mark(addr); // now grey
7694 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7695 // Should we push this marked oop on our stack?
7696 // -- if someone else marked it, nothing to do
7697 // -- if target oop is above global finger nothing to do
7698 // -- if target oop is in chunk and above local finger
7699 // then nothing to do
7700 // -- else push on work queue
7701 if ( !res // someone else marked it, they will deal with it
7702 || (addr >= *gfa) // will be scanned in a later task
7703 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7704 return;
7705 }
7706 // the bit map iteration has already either passed, or
7707 // sampled, this bit in the bit map; we'll need to
7708 // use the marking stack to scan this oop's oops.
7709 bool simulate_overflow = false;
7710 NOT_PRODUCT(
7711 if (CMSMarkStackOverflowALot &&
7712 _collector->simulate_overflow()) {
7713 // simulate a stack overflow
7714 simulate_overflow = true;
7715 }
7716 )
7717 if (simulate_overflow ||
7718 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7719 // stack overflow
7720 if (PrintCMSStatistics != 0) {
7721 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7722 SIZE_FORMAT, _overflow_stack->capacity());
7723 }
7724 // We cannot assert that the overflow stack is full because
7725 // it may have been emptied since.
7726 assert(simulate_overflow ||
7727 _work_queue->size() == _work_queue->max_elems(),
7728 "Else push should have succeeded");
7729 handle_stack_overflow(addr);
7730 }
7731 do_yield_check();
7732 }
7733 }
7734
7735 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7736 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7737
7738 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector,
7739 ReferenceProcessor* rp,
7740 CMSMarkStack* revisit_stack) :
7741 OopClosure(rp),
7742 _collector(collector),
7743 _revisit_stack(revisit_stack),
7744 _should_remember_klasses(collector->should_unload_classes()) {}
7745
7746 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7747 MemRegion span,
7748 ReferenceProcessor* rp,
7749 CMSBitMap* bit_map,
7750 CMSBitMap* mod_union_table,
7751 CMSMarkStack* mark_stack,
7752 CMSMarkStack* revisit_stack,
7753 bool concurrent_precleaning):
7754 KlassRememberingOopClosure(collector, rp, revisit_stack),
7755 _span(span),
7756 _bit_map(bit_map),
7757 _mod_union_table(mod_union_table),
7758 _mark_stack(mark_stack),
7759 _concurrent_precleaning(concurrent_precleaning)
7760 {
7761 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7762 }
7763
7764 // Grey object rescan during pre-cleaning and second checkpoint phases --
7765 // the non-parallel version (the parallel version appears further below.)
7766 void PushAndMarkClosure::do_oop(oop obj) {
7767 // Ignore mark word verification. If during concurrent precleaning,
7768 // the object monitor may be locked. If during the checkpoint
7769 // phases, the object may already have been reached by a different
7770 // path and may be at the end of the global overflow list (so
7771 // the mark word may be NULL).
7772 assert(obj->is_oop_or_null(true /* ignore mark word */),
7773 "expected an oop or NULL");
7774 HeapWord* addr = (HeapWord*)obj;
7775 // Check if oop points into the CMS generation
7776 // and is not marked
7777 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7778 // a white object ...
7779 _bit_map->mark(addr); // ... now grey
7780 // push on the marking stack (grey set)
7781 bool simulate_overflow = false;
7782 NOT_PRODUCT(
7783 if (CMSMarkStackOverflowALot &&
7784 _collector->simulate_overflow()) {
7785 // simulate a stack overflow
7786 simulate_overflow = true;
7787 }
7788 )
7789 if (simulate_overflow || !_mark_stack->push(obj)) {
7790 if (_concurrent_precleaning) {
7791 // During precleaning we can just dirty the appropriate card(s)
7792 // in the mod union table, thus ensuring that the object remains
7793 // in the grey set and continue. In the case of object arrays
7794 // we need to dirty all of the cards that the object spans,
7795 // since the rescan of object arrays will be limited to the
7796 // dirty cards.
7797 // Note that no one can be intefering with us in this action
7798 // of dirtying the mod union table, so no locking or atomics
7799 // are required.
7800 if (obj->is_objArray()) {
7801 size_t sz = obj->size();
7802 HeapWord* end_card_addr = (HeapWord*)round_to(
7803 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7804 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7805 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7806 _mod_union_table->mark_range(redirty_range);
7807 } else {
7808 _mod_union_table->mark(addr);
7809 }
7810 _collector->_ser_pmc_preclean_ovflw++;
7811 } else {
7812 // During the remark phase, we need to remember this oop
7813 // in the overflow list.
7814 _collector->push_on_overflow_list(obj);
7815 _collector->_ser_pmc_remark_ovflw++;
7816 }
7817 }
7818 }
7819 }
7820
7821 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
7822 MemRegion span,
7823 ReferenceProcessor* rp,
7824 CMSBitMap* bit_map,
7825 OopTaskQueue* work_queue,
7826 CMSMarkStack* revisit_stack):
7827 Par_KlassRememberingOopClosure(collector, rp, revisit_stack),
7828 _span(span),
7829 _bit_map(bit_map),
7830 _work_queue(work_queue)
7831 {
7832 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7833 }
7834
7835 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
7836 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
7837
7838 // Grey object rescan during second checkpoint phase --
7839 // the parallel version.
7840 void Par_PushAndMarkClosure::do_oop(oop obj) {
7841 // In the assert below, we ignore the mark word because
7842 // this oop may point to an already visited object that is
7843 // on the overflow stack (in which case the mark word has
7844 // been hijacked for chaining into the overflow stack --
7845 // if this is the last object in the overflow stack then
7846 // its mark word will be NULL). Because this object may
7847 // have been subsequently popped off the global overflow
7848 // stack, and the mark word possibly restored to the prototypical
7849 // value, by the time we get to examined this failing assert in
7850 // the debugger, is_oop_or_null(false) may subsequently start
7851 // to hold.
7852 assert(obj->is_oop_or_null(true),
7853 "expected an oop or NULL");
7854 HeapWord* addr = (HeapWord*)obj;
7855 // Check if oop points into the CMS generation
7856 // and is not marked
7857 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7858 // a white object ...
7859 // If we manage to "claim" the object, by being the
7860 // first thread to mark it, then we push it on our
7861 // marking stack
7862 if (_bit_map->par_mark(addr)) { // ... now grey
7863 // push on work queue (grey set)
7864 bool simulate_overflow = false;
7865 NOT_PRODUCT(
7866 if (CMSMarkStackOverflowALot &&
7867 _collector->par_simulate_overflow()) {
7868 // simulate a stack overflow
7869 simulate_overflow = true;
7870 }
7871 )
7872 if (simulate_overflow || !_work_queue->push(obj)) {
7873 _collector->par_push_on_overflow_list(obj);
7874 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
7875 }
7876 } // Else, some other thread got there first
7877 }
7878 }
7879
7880 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7881 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
7882
7883 void PushAndMarkClosure::remember_mdo(DataLayout* v) {
7884 // TBD
7885 }
7886
7887 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) {
7888 // TBD
7889 }
7890
7891 void CMSPrecleanRefsYieldClosure::do_yield_work() {
7892 DEBUG_ONLY(RememberKlassesChecker mux(false);)
7893 Mutex* bml = _collector->bitMapLock();
7894 assert_lock_strong(bml);
7895 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7896 "CMS thread should hold CMS token");
7897
7898 bml->unlock();
7899 ConcurrentMarkSweepThread::desynchronize(true);
7900
7901 ConcurrentMarkSweepThread::acknowledge_yield_request();
7902
7903 _collector->stopTimer();
7904 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7905 if (PrintCMSStatistics != 0) {
7906 _collector->incrementYields();
7907 }
7908 _collector->icms_wait();
7909
7910 // See the comment in coordinator_yield()
7911 for (unsigned i = 0; i < CMSYieldSleepCount &&
7912 ConcurrentMarkSweepThread::should_yield() &&
7913 !CMSCollector::foregroundGCIsActive(); ++i) {
7914 os::sleep(Thread::current(), 1, false);
7915 ConcurrentMarkSweepThread::acknowledge_yield_request();
7916 }
7917
7918 ConcurrentMarkSweepThread::synchronize(true);
7919 bml->lock();
7920
7921 _collector->startTimer();
7922 }
7923
7924 bool CMSPrecleanRefsYieldClosure::should_return() {
7925 if (ConcurrentMarkSweepThread::should_yield()) {
7926 do_yield_work();
7927 }
7928 return _collector->foregroundGCIsActive();
7929 }
7930
7931 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7932 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
7933 "mr should be aligned to start at a card boundary");
7934 // We'd like to assert:
7935 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
7936 // "mr should be a range of cards");
7937 // However, that would be too strong in one case -- the last
7938 // partition ends at _unallocated_block which, in general, can be
7939 // an arbitrary boundary, not necessarily card aligned.
7940 if (PrintCMSStatistics != 0) {
7941 _num_dirty_cards +=
7942 mr.word_size()/CardTableModRefBS::card_size_in_words;
7943 }
7944 _space->object_iterate_mem(mr, &_scan_cl);
7945 }
7946
7947 SweepClosure::SweepClosure(CMSCollector* collector,
7948 ConcurrentMarkSweepGeneration* g,
7949 CMSBitMap* bitMap, bool should_yield) :
7950 _collector(collector),
7951 _g(g),
7952 _sp(g->cmsSpace()),
7953 _limit(_sp->sweep_limit()),
7954 _freelistLock(_sp->freelistLock()),
7955 _bitMap(bitMap),
7956 _yield(should_yield),
7957 _inFreeRange(false), // No free range at beginning of sweep
7958 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7959 _lastFreeRangeCoalesced(false),
7960 _freeFinger(g->used_region().start())
7961 {
7962 NOT_PRODUCT(
7963 _numObjectsFreed = 0;
7964 _numWordsFreed = 0;
7965 _numObjectsLive = 0;
7966 _numWordsLive = 0;
7967 _numObjectsAlreadyFree = 0;
7968 _numWordsAlreadyFree = 0;
7969 _last_fc = NULL;
7970
7971 _sp->initializeIndexedFreeListArrayReturnedBytes();
7972 _sp->dictionary()->initialize_dict_returned_bytes();
7973 )
7974 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7975 "sweep _limit out of bounds");
7976 if (CMSTraceSweeper) {
7977 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
7978 _limit);
7979 }
7980 }
7981
7982 void SweepClosure::print_on(outputStream* st) const {
7983 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7984 _sp->bottom(), _sp->end());
7985 tty->print_cr("_limit = " PTR_FORMAT, _limit);
7986 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
7987 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
7988 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7989 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7990 }
7991
7992 #ifndef PRODUCT
7993 // Assertion checking only: no useful work in product mode --
7994 // however, if any of the flags below become product flags,
7995 // you may need to review this code to see if it needs to be
7996 // enabled in product mode.
7997 SweepClosure::~SweepClosure() {
7998 assert_lock_strong(_freelistLock);
7999 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8000 "sweep _limit out of bounds");
8001 if (inFreeRange()) {
8002 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
8003 print();
8004 ShouldNotReachHere();
8005 }
8006 if (Verbose && PrintGC) {
8007 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
8008 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
8009 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
8010 SIZE_FORMAT" bytes "
8011 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
8012 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
8013 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
8014 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
8015 * sizeof(HeapWord);
8016 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
8017
8018 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
8019 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
8020 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
8021 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
8022 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
8023 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
8024 indexListReturnedBytes);
8025 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
8026 dict_returned_bytes);
8027 }
8028 }
8029 if (CMSTraceSweeper) {
8030 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
8031 _limit);
8032 }
8033 }
8034 #endif // PRODUCT
8035
8036 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8037 bool freeRangeInFreeLists) {
8038 if (CMSTraceSweeper) {
8039 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
8040 freeFinger, freeRangeInFreeLists);
8041 }
8042 assert(!inFreeRange(), "Trampling existing free range");
8043 set_inFreeRange(true);
8044 set_lastFreeRangeCoalesced(false);
8045
8046 set_freeFinger(freeFinger);
8047 set_freeRangeInFreeLists(freeRangeInFreeLists);
8048 if (CMSTestInFreeList) {
8049 if (freeRangeInFreeLists) {
8050 FreeChunk* fc = (FreeChunk*) freeFinger;
8051 assert(fc->is_free(), "A chunk on the free list should be free.");
8052 assert(fc->size() > 0, "Free range should have a size");
8053 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8054 }
8055 }
8056 }
8057
8058 // Note that the sweeper runs concurrently with mutators. Thus,
8059 // it is possible for direct allocation in this generation to happen
8060 // in the middle of the sweep. Note that the sweeper also coalesces
8061 // contiguous free blocks. Thus, unless the sweeper and the allocator
8062 // synchronize appropriately freshly allocated blocks may get swept up.
8063 // This is accomplished by the sweeper locking the free lists while
8064 // it is sweeping. Thus blocks that are determined to be free are
8065 // indeed free. There is however one additional complication:
8066 // blocks that have been allocated since the final checkpoint and
8067 // mark, will not have been marked and so would be treated as
8068 // unreachable and swept up. To prevent this, the allocator marks
8069 // the bit map when allocating during the sweep phase. This leads,
8070 // however, to a further complication -- objects may have been allocated
8071 // but not yet initialized -- in the sense that the header isn't yet
8072 // installed. The sweeper can not then determine the size of the block
8073 // in order to skip over it. To deal with this case, we use a technique
8074 // (due to Printezis) to encode such uninitialized block sizes in the
8075 // bit map. Since the bit map uses a bit per every HeapWord, but the
8076 // CMS generation has a minimum object size of 3 HeapWords, it follows
8077 // that "normal marks" won't be adjacent in the bit map (there will
8078 // always be at least two 0 bits between successive 1 bits). We make use
8079 // of these "unused" bits to represent uninitialized blocks -- the bit
8080 // corresponding to the start of the uninitialized object and the next
8081 // bit are both set. Finally, a 1 bit marks the end of the object that
8082 // started with the two consecutive 1 bits to indicate its potentially
8083 // uninitialized state.
8084
8085 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8086 FreeChunk* fc = (FreeChunk*)addr;
8087 size_t res;
8088
8089 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8090 // than "addr == _limit" because although _limit was a block boundary when
8091 // we started the sweep, it may no longer be one because heap expansion
8092 // may have caused us to coalesce the block ending at the address _limit
8093 // with a newly expanded chunk (this happens when _limit was set to the
8094 // previous _end of the space), so we may have stepped past _limit:
8095 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8096 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8097 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8098 "sweep _limit out of bounds");
8099 assert(addr < _sp->end(), "addr out of bounds");
8100 // Flush any free range we might be holding as a single
8101 // coalesced chunk to the appropriate free list.
8102 if (inFreeRange()) {
8103 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8104 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8105 flush_cur_free_chunk(freeFinger(),
8106 pointer_delta(addr, freeFinger()));
8107 if (CMSTraceSweeper) {
8108 gclog_or_tty->print("Sweep: last chunk: ");
8109 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8110 "[coalesced:"SIZE_FORMAT"]\n",
8111 freeFinger(), pointer_delta(addr, freeFinger()),
8112 lastFreeRangeCoalesced());
8113 }
8114 }
8115
8116 // help the iterator loop finish
8117 return pointer_delta(_sp->end(), addr);
8118 }
8119
8120 assert(addr < _limit, "sweep invariant");
8121 // check if we should yield
8122 do_yield_check(addr);
8123 if (fc->is_free()) {
8124 // Chunk that is already free
8125 res = fc->size();
8126 do_already_free_chunk(fc);
8127 debug_only(_sp->verifyFreeLists());
8128 // If we flush the chunk at hand in lookahead_and_flush()
8129 // and it's coalesced with a preceding chunk, then the
8130 // process of "mangling" the payload of the coalesced block
8131 // will cause erasure of the size information from the
8132 // (erstwhile) header of all the coalesced blocks but the
8133 // first, so the first disjunct in the assert will not hold
8134 // in that specific case (in which case the second disjunct
8135 // will hold).
8136 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8137 "Otherwise the size info doesn't change at this step");
8138 NOT_PRODUCT(
8139 _numObjectsAlreadyFree++;
8140 _numWordsAlreadyFree += res;
8141 )
8142 NOT_PRODUCT(_last_fc = fc;)
8143 } else if (!_bitMap->isMarked(addr)) {
8144 // Chunk is fresh garbage
8145 res = do_garbage_chunk(fc);
8146 debug_only(_sp->verifyFreeLists());
8147 NOT_PRODUCT(
8148 _numObjectsFreed++;
8149 _numWordsFreed += res;
8150 )
8151 } else {
8152 // Chunk that is alive.
8153 res = do_live_chunk(fc);
8154 debug_only(_sp->verifyFreeLists());
8155 NOT_PRODUCT(
8156 _numObjectsLive++;
8157 _numWordsLive += res;
8158 )
8159 }
8160 return res;
8161 }
8162
8163 // For the smart allocation, record following
8164 // split deaths - a free chunk is removed from its free list because
8165 // it is being split into two or more chunks.
8166 // split birth - a free chunk is being added to its free list because
8167 // a larger free chunk has been split and resulted in this free chunk.
8168 // coal death - a free chunk is being removed from its free list because
8169 // it is being coalesced into a large free chunk.
8170 // coal birth - a free chunk is being added to its free list because
8171 // it was created when two or more free chunks where coalesced into
8172 // this free chunk.
8173 //
8174 // These statistics are used to determine the desired number of free
8175 // chunks of a given size. The desired number is chosen to be relative
8176 // to the end of a CMS sweep. The desired number at the end of a sweep
8177 // is the
8178 // count-at-end-of-previous-sweep (an amount that was enough)
8179 // - count-at-beginning-of-current-sweep (the excess)
8180 // + split-births (gains in this size during interval)
8181 // - split-deaths (demands on this size during interval)
8182 // where the interval is from the end of one sweep to the end of the
8183 // next.
8184 //
8185 // When sweeping the sweeper maintains an accumulated chunk which is
8186 // the chunk that is made up of chunks that have been coalesced. That
8187 // will be termed the left-hand chunk. A new chunk of garbage that
8188 // is being considered for coalescing will be referred to as the
8189 // right-hand chunk.
8190 //
8191 // When making a decision on whether to coalesce a right-hand chunk with
8192 // the current left-hand chunk, the current count vs. the desired count
8193 // of the left-hand chunk is considered. Also if the right-hand chunk
8194 // is near the large chunk at the end of the heap (see
8195 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8196 // left-hand chunk is coalesced.
8197 //
8198 // When making a decision about whether to split a chunk, the desired count
8199 // vs. the current count of the candidate to be split is also considered.
8200 // If the candidate is underpopulated (currently fewer chunks than desired)
8201 // a chunk of an overpopulated (currently more chunks than desired) size may
8202 // be chosen. The "hint" associated with a free list, if non-null, points
8203 // to a free list which may be overpopulated.
8204 //
8205
8206 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8207 const size_t size = fc->size();
8208 // Chunks that cannot be coalesced are not in the
8209 // free lists.
8210 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8211 assert(_sp->verify_chunk_in_free_list(fc),
8212 "free chunk should be in free lists");
8213 }
8214 // a chunk that is already free, should not have been
8215 // marked in the bit map
8216 HeapWord* const addr = (HeapWord*) fc;
8217 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8218 // Verify that the bit map has no bits marked between
8219 // addr and purported end of this block.
8220 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8221
8222 // Some chunks cannot be coalesced under any circumstances.
8223 // See the definition of cantCoalesce().
8224 if (!fc->cantCoalesce()) {
8225 // This chunk can potentially be coalesced.
8226 if (_sp->adaptive_freelists()) {
8227 // All the work is done in
8228 do_post_free_or_garbage_chunk(fc, size);
8229 } else { // Not adaptive free lists
8230 // this is a free chunk that can potentially be coalesced by the sweeper;
8231 if (!inFreeRange()) {
8232 // if the next chunk is a free block that can't be coalesced
8233 // it doesn't make sense to remove this chunk from the free lists
8234 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8235 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8236 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8237 nextChunk->is_free() && // ... which is free...
8238 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8239 // nothing to do
8240 } else {
8241 // Potentially the start of a new free range:
8242 // Don't eagerly remove it from the free lists.
8243 // No need to remove it if it will just be put
8244 // back again. (Also from a pragmatic point of view
8245 // if it is a free block in a region that is beyond
8246 // any allocated blocks, an assertion will fail)
8247 // Remember the start of a free run.
8248 initialize_free_range(addr, true);
8249 // end - can coalesce with next chunk
8250 }
8251 } else {
8252 // the midst of a free range, we are coalescing
8253 print_free_block_coalesced(fc);
8254 if (CMSTraceSweeper) {
8255 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8256 }
8257 // remove it from the free lists
8258 _sp->removeFreeChunkFromFreeLists(fc);
8259 set_lastFreeRangeCoalesced(true);
8260 // If the chunk is being coalesced and the current free range is
8261 // in the free lists, remove the current free range so that it
8262 // will be returned to the free lists in its entirety - all
8263 // the coalesced pieces included.
8264 if (freeRangeInFreeLists()) {
8265 FreeChunk* ffc = (FreeChunk*) freeFinger();
8266 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8267 "Size of free range is inconsistent with chunk size.");
8268 if (CMSTestInFreeList) {
8269 assert(_sp->verify_chunk_in_free_list(ffc),
8270 "free range is not in free lists");
8271 }
8272 _sp->removeFreeChunkFromFreeLists(ffc);
8273 set_freeRangeInFreeLists(false);
8274 }
8275 }
8276 }
8277 // Note that if the chunk is not coalescable (the else arm
8278 // below), we unconditionally flush, without needing to do
8279 // a "lookahead," as we do below.
8280 if (inFreeRange()) lookahead_and_flush(fc, size);
8281 } else {
8282 // Code path common to both original and adaptive free lists.
8283
8284 // cant coalesce with previous block; this should be treated
8285 // as the end of a free run if any
8286 if (inFreeRange()) {
8287 // we kicked some butt; time to pick up the garbage
8288 assert(freeFinger() < addr, "freeFinger points too high");
8289 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8290 }
8291 // else, nothing to do, just continue
8292 }
8293 }
8294
8295 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8296 // This is a chunk of garbage. It is not in any free list.
8297 // Add it to a free list or let it possibly be coalesced into
8298 // a larger chunk.
8299 HeapWord* const addr = (HeapWord*) fc;
8300 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8301
8302 if (_sp->adaptive_freelists()) {
8303 // Verify that the bit map has no bits marked between
8304 // addr and purported end of just dead object.
8305 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8306
8307 do_post_free_or_garbage_chunk(fc, size);
8308 } else {
8309 if (!inFreeRange()) {
8310 // start of a new free range
8311 assert(size > 0, "A free range should have a size");
8312 initialize_free_range(addr, false);
8313 } else {
8314 // this will be swept up when we hit the end of the
8315 // free range
8316 if (CMSTraceSweeper) {
8317 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8318 }
8319 // If the chunk is being coalesced and the current free range is
8320 // in the free lists, remove the current free range so that it
8321 // will be returned to the free lists in its entirety - all
8322 // the coalesced pieces included.
8323 if (freeRangeInFreeLists()) {
8324 FreeChunk* ffc = (FreeChunk*)freeFinger();
8325 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8326 "Size of free range is inconsistent with chunk size.");
8327 if (CMSTestInFreeList) {
8328 assert(_sp->verify_chunk_in_free_list(ffc),
8329 "free range is not in free lists");
8330 }
8331 _sp->removeFreeChunkFromFreeLists(ffc);
8332 set_freeRangeInFreeLists(false);
8333 }
8334 set_lastFreeRangeCoalesced(true);
8335 }
8336 // this will be swept up when we hit the end of the free range
8337
8338 // Verify that the bit map has no bits marked between
8339 // addr and purported end of just dead object.
8340 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8341 }
8342 assert(_limit >= addr + size,
8343 "A freshly garbage chunk can't possibly straddle over _limit");
8344 if (inFreeRange()) lookahead_and_flush(fc, size);
8345 return size;
8346 }
8347
8348 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8349 HeapWord* addr = (HeapWord*) fc;
8350 // The sweeper has just found a live object. Return any accumulated
8351 // left hand chunk to the free lists.
8352 if (inFreeRange()) {
8353 assert(freeFinger() < addr, "freeFinger points too high");
8354 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8355 }
8356
8357 // This object is live: we'd normally expect this to be
8358 // an oop, and like to assert the following:
8359 // assert(oop(addr)->is_oop(), "live block should be an oop");
8360 // However, as we commented above, this may be an object whose
8361 // header hasn't yet been initialized.
8362 size_t size;
8363 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8364 if (_bitMap->isMarked(addr + 1)) {
8365 // Determine the size from the bit map, rather than trying to
8366 // compute it from the object header.
8367 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8368 size = pointer_delta(nextOneAddr + 1, addr);
8369 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8370 "alignment problem");
8371
8372 #ifdef DEBUG
8373 if (oop(addr)->klass_or_null() != NULL &&
8374 ( !_collector->should_unload_classes()
8375 || (oop(addr)->is_parsable()) &&
8376 oop(addr)->is_conc_safe())) {
8377 // Ignore mark word because we are running concurrent with mutators
8378 assert(oop(addr)->is_oop(true), "live block should be an oop");
8379 // is_conc_safe is checked before performing this assertion
8380 // because an object that is not is_conc_safe may yet have
8381 // the return from size() correct.
8382 assert(size ==
8383 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8384 "P-mark and computed size do not agree");
8385 }
8386 #endif
8387
8388 } else {
8389 // This should be an initialized object that's alive.
8390 assert(oop(addr)->klass_or_null() != NULL &&
8391 (!_collector->should_unload_classes()
8392 || oop(addr)->is_parsable()),
8393 "Should be an initialized object");
8394 // Note that there are objects used during class redefinition,
8395 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(),
8396 // which are discarded with their is_conc_safe state still
8397 // false. These object may be floating garbage so may be
8398 // seen here. If they are floating garbage their size
8399 // should be attainable from their klass. Do not that
8400 // is_conc_safe() is true for oop(addr).
8401 // Ignore mark word because we are running concurrent with mutators
8402 assert(oop(addr)->is_oop(true), "live block should be an oop");
8403 // Verify that the bit map has no bits marked between
8404 // addr and purported end of this block.
8405 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8406 assert(size >= 3, "Necessary for Printezis marks to work");
8407 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8408 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8409 }
8410 return size;
8411 }
8412
8413 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8414 size_t chunkSize) {
8415 // do_post_free_or_garbage_chunk() should only be called in the case
8416 // of the adaptive free list allocator.
8417 const bool fcInFreeLists = fc->is_free();
8418 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8419 assert((HeapWord*)fc <= _limit, "sweep invariant");
8420 if (CMSTestInFreeList && fcInFreeLists) {
8421 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8422 }
8423
8424 if (CMSTraceSweeper) {
8425 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8426 }
8427
8428 HeapWord* const fc_addr = (HeapWord*) fc;
8429
8430 bool coalesce;
8431 const size_t left = pointer_delta(fc_addr, freeFinger());
8432 const size_t right = chunkSize;
8433 switch (FLSCoalescePolicy) {
8434 // numeric value forms a coalition aggressiveness metric
8435 case 0: { // never coalesce
8436 coalesce = false;
8437 break;
8438 }
8439 case 1: { // coalesce if left & right chunks on overpopulated lists
8440 coalesce = _sp->coalOverPopulated(left) &&
8441 _sp->coalOverPopulated(right);
8442 break;
8443 }
8444 case 2: { // coalesce if left chunk on overpopulated list (default)
8445 coalesce = _sp->coalOverPopulated(left);
8446 break;
8447 }
8448 case 3: { // coalesce if left OR right chunk on overpopulated list
8449 coalesce = _sp->coalOverPopulated(left) ||
8450 _sp->coalOverPopulated(right);
8451 break;
8452 }
8453 case 4: { // always coalesce
8454 coalesce = true;
8455 break;
8456 }
8457 default:
8458 ShouldNotReachHere();
8459 }
8460
8461 // Should the current free range be coalesced?
8462 // If the chunk is in a free range and either we decided to coalesce above
8463 // or the chunk is near the large block at the end of the heap
8464 // (isNearLargestChunk() returns true), then coalesce this chunk.
8465 const bool doCoalesce = inFreeRange()
8466 && (coalesce || _g->isNearLargestChunk(fc_addr));
8467 if (doCoalesce) {
8468 // Coalesce the current free range on the left with the new
8469 // chunk on the right. If either is on a free list,
8470 // it must be removed from the list and stashed in the closure.
8471 if (freeRangeInFreeLists()) {
8472 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8473 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8474 "Size of free range is inconsistent with chunk size.");
8475 if (CMSTestInFreeList) {
8476 assert(_sp->verify_chunk_in_free_list(ffc),
8477 "Chunk is not in free lists");
8478 }
8479 _sp->coalDeath(ffc->size());
8480 _sp->removeFreeChunkFromFreeLists(ffc);
8481 set_freeRangeInFreeLists(false);
8482 }
8483 if (fcInFreeLists) {
8484 _sp->coalDeath(chunkSize);
8485 assert(fc->size() == chunkSize,
8486 "The chunk has the wrong size or is not in the free lists");
8487 _sp->removeFreeChunkFromFreeLists(fc);
8488 }
8489 set_lastFreeRangeCoalesced(true);
8490 print_free_block_coalesced(fc);
8491 } else { // not in a free range and/or should not coalesce
8492 // Return the current free range and start a new one.
8493 if (inFreeRange()) {
8494 // In a free range but cannot coalesce with the right hand chunk.
8495 // Put the current free range into the free lists.
8496 flush_cur_free_chunk(freeFinger(),
8497 pointer_delta(fc_addr, freeFinger()));
8498 }
8499 // Set up for new free range. Pass along whether the right hand
8500 // chunk is in the free lists.
8501 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8502 }
8503 }
8504
8505 // Lookahead flush:
8506 // If we are tracking a free range, and this is the last chunk that
8507 // we'll look at because its end crosses past _limit, we'll preemptively
8508 // flush it along with any free range we may be holding on to. Note that
8509 // this can be the case only for an already free or freshly garbage
8510 // chunk. If this block is an object, it can never straddle
8511 // over _limit. The "straddling" occurs when _limit is set at
8512 // the previous end of the space when this cycle started, and
8513 // a subsequent heap expansion caused the previously co-terminal
8514 // free block to be coalesced with the newly expanded portion,
8515 // thus rendering _limit a non-block-boundary making it dangerous
8516 // for the sweeper to step over and examine.
8517 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8518 assert(inFreeRange(), "Should only be called if currently in a free range.");
8519 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8520 assert(_sp->used_region().contains(eob - 1),
8521 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8522 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8523 _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8524 if (eob >= _limit) {
8525 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8526 if (CMSTraceSweeper) {
8527 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8528 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8529 "[" PTR_FORMAT "," PTR_FORMAT ")",
8530 _limit, fc, eob, _sp->bottom(), _sp->end());
8531 }
8532 // Return the storage we are tracking back into the free lists.
8533 if (CMSTraceSweeper) {
8534 gclog_or_tty->print_cr("Flushing ... ");
8535 }
8536 assert(freeFinger() < eob, "Error");
8537 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8538 }
8539 }
8540
8541 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8542 assert(inFreeRange(), "Should only be called if currently in a free range.");
8543 assert(size > 0,
8544 "A zero sized chunk cannot be added to the free lists.");
8545 if (!freeRangeInFreeLists()) {
8546 if (CMSTestInFreeList) {
8547 FreeChunk* fc = (FreeChunk*) chunk;
8548 fc->set_size(size);
8549 assert(!_sp->verify_chunk_in_free_list(fc),
8550 "chunk should not be in free lists yet");
8551 }
8552 if (CMSTraceSweeper) {
8553 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8554 chunk, size);
8555 }
8556 // A new free range is going to be starting. The current
8557 // free range has not been added to the free lists yet or
8558 // was removed so add it back.
8559 // If the current free range was coalesced, then the death
8560 // of the free range was recorded. Record a birth now.
8561 if (lastFreeRangeCoalesced()) {
8562 _sp->coalBirth(size);
8563 }
8564 _sp->addChunkAndRepairOffsetTable(chunk, size,
8565 lastFreeRangeCoalesced());
8566 } else if (CMSTraceSweeper) {
8567 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8568 }
8569 set_inFreeRange(false);
8570 set_freeRangeInFreeLists(false);
8571 }
8572
8573 // We take a break if we've been at this for a while,
8574 // so as to avoid monopolizing the locks involved.
8575 void SweepClosure::do_yield_work(HeapWord* addr) {
8576 // Return current free chunk being used for coalescing (if any)
8577 // to the appropriate freelist. After yielding, the next
8578 // free block encountered will start a coalescing range of
8579 // free blocks. If the next free block is adjacent to the
8580 // chunk just flushed, they will need to wait for the next
8581 // sweep to be coalesced.
8582 if (inFreeRange()) {
8583 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8584 }
8585
8586 // First give up the locks, then yield, then re-lock.
8587 // We should probably use a constructor/destructor idiom to
8588 // do this unlock/lock or modify the MutexUnlocker class to
8589 // serve our purpose. XXX
8590 assert_lock_strong(_bitMap->lock());
8591 assert_lock_strong(_freelistLock);
8592 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8593 "CMS thread should hold CMS token");
8594 _bitMap->lock()->unlock();
8595 _freelistLock->unlock();
8596 ConcurrentMarkSweepThread::desynchronize(true);
8597 ConcurrentMarkSweepThread::acknowledge_yield_request();
8598 _collector->stopTimer();
8599 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8600 if (PrintCMSStatistics != 0) {
8601 _collector->incrementYields();
8602 }
8603 _collector->icms_wait();
8604
8605 // See the comment in coordinator_yield()
8606 for (unsigned i = 0; i < CMSYieldSleepCount &&
8607 ConcurrentMarkSweepThread::should_yield() &&
8608 !CMSCollector::foregroundGCIsActive(); ++i) {
8609 os::sleep(Thread::current(), 1, false);
8610 ConcurrentMarkSweepThread::acknowledge_yield_request();
8611 }
8612
8613 ConcurrentMarkSweepThread::synchronize(true);
8614 _freelistLock->lock();
8615 _bitMap->lock()->lock_without_safepoint_check();
8616 _collector->startTimer();
8617 }
8618
8619 #ifndef PRODUCT
8620 // This is actually very useful in a product build if it can
8621 // be called from the debugger. Compile it into the product
8622 // as needed.
8623 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8624 return debug_cms_space->verify_chunk_in_free_list(fc);
8625 }
8626 #endif
8627
8628 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8629 if (CMSTraceSweeper) {
8630 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8631 fc, fc->size());
8632 }
8633 }
8634
8635 // CMSIsAliveClosure
8636 bool CMSIsAliveClosure::do_object_b(oop obj) {
8637 HeapWord* addr = (HeapWord*)obj;
8638 return addr != NULL &&
8639 (!_span.contains(addr) || _bit_map->isMarked(addr));
8640 }
8641
8642 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8643 MemRegion span,
8644 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8645 CMSMarkStack* revisit_stack, bool cpc):
8646 KlassRememberingOopClosure(collector, NULL, revisit_stack),
8647 _span(span),
8648 _bit_map(bit_map),
8649 _mark_stack(mark_stack),
8650 _concurrent_precleaning(cpc) {
8651 assert(!_span.is_empty(), "Empty span could spell trouble");
8652 }
8653
8654
8655 // CMSKeepAliveClosure: the serial version
8656 void CMSKeepAliveClosure::do_oop(oop obj) {
8657 HeapWord* addr = (HeapWord*)obj;
8658 if (_span.contains(addr) &&
8659 !_bit_map->isMarked(addr)) {
8660 _bit_map->mark(addr);
8661 bool simulate_overflow = false;
8662 NOT_PRODUCT(
8663 if (CMSMarkStackOverflowALot &&
8664 _collector->simulate_overflow()) {
8665 // simulate a stack overflow
8666 simulate_overflow = true;
8667 }
8668 )
8669 if (simulate_overflow || !_mark_stack->push(obj)) {
8670 if (_concurrent_precleaning) {
8671 // We dirty the overflown object and let the remark
8672 // phase deal with it.
8673 assert(_collector->overflow_list_is_empty(), "Error");
8674 // In the case of object arrays, we need to dirty all of
8675 // the cards that the object spans. No locking or atomics
8676 // are needed since no one else can be mutating the mod union
8677 // table.
8678 if (obj->is_objArray()) {
8679 size_t sz = obj->size();
8680 HeapWord* end_card_addr =
8681 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8682 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8683 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8684 _collector->_modUnionTable.mark_range(redirty_range);
8685 } else {
8686 _collector->_modUnionTable.mark(addr);
8687 }
8688 _collector->_ser_kac_preclean_ovflw++;
8689 } else {
8690 _collector->push_on_overflow_list(obj);
8691 _collector->_ser_kac_ovflw++;
8692 }
8693 }
8694 }
8695 }
8696
8697 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8698 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8699
8700 // CMSParKeepAliveClosure: a parallel version of the above.
8701 // The work queues are private to each closure (thread),
8702 // but (may be) available for stealing by other threads.
8703 void CMSParKeepAliveClosure::do_oop(oop obj) {
8704 HeapWord* addr = (HeapWord*)obj;
8705 if (_span.contains(addr) &&
8706 !_bit_map->isMarked(addr)) {
8707 // In general, during recursive tracing, several threads
8708 // may be concurrently getting here; the first one to
8709 // "tag" it, claims it.
8710 if (_bit_map->par_mark(addr)) {
8711 bool res = _work_queue->push(obj);
8712 assert(res, "Low water mark should be much less than capacity");
8713 // Do a recursive trim in the hope that this will keep
8714 // stack usage lower, but leave some oops for potential stealers
8715 trim_queue(_low_water_mark);
8716 } // Else, another thread got there first
8717 }
8718 }
8719
8720 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8721 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8722
8723 void CMSParKeepAliveClosure::trim_queue(uint max) {
8724 while (_work_queue->size() > max) {
8725 oop new_oop;
8726 if (_work_queue->pop_local(new_oop)) {
8727 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8728 assert(_bit_map->isMarked((HeapWord*)new_oop),
8729 "no white objects on this stack!");
8730 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8731 // iterate over the oops in this oop, marking and pushing
8732 // the ones in CMS heap (i.e. in _span).
8733 new_oop->oop_iterate(&_mark_and_push);
8734 }
8735 }
8736 }
8737
8738 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8739 CMSCollector* collector,
8740 MemRegion span, CMSBitMap* bit_map,
8741 CMSMarkStack* revisit_stack,
8742 OopTaskQueue* work_queue):
8743 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
8744 _span(span),
8745 _bit_map(bit_map),
8746 _work_queue(work_queue) { }
8747
8748 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8749 HeapWord* addr = (HeapWord*)obj;
8750 if (_span.contains(addr) &&
8751 !_bit_map->isMarked(addr)) {
8752 if (_bit_map->par_mark(addr)) {
8753 bool simulate_overflow = false;
8754 NOT_PRODUCT(
8755 if (CMSMarkStackOverflowALot &&
8756 _collector->par_simulate_overflow()) {
8757 // simulate a stack overflow
8758 simulate_overflow = true;
8759 }
8760 )
8761 if (simulate_overflow || !_work_queue->push(obj)) {
8762 _collector->par_push_on_overflow_list(obj);
8763 _collector->_par_kac_ovflw++;
8764 }
8765 } // Else another thread got there already
8766 }
8767 }
8768
8769 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8770 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8771
8772 //////////////////////////////////////////////////////////////////
8773 // CMSExpansionCause /////////////////////////////
8774 //////////////////////////////////////////////////////////////////
8775 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8776 switch (cause) {
8777 case _no_expansion:
8778 return "No expansion";
8779 case _satisfy_free_ratio:
8780 return "Free ratio";
8781 case _satisfy_promotion:
8782 return "Satisfy promotion";
8783 case _satisfy_allocation:
8784 return "allocation";
8785 case _allocate_par_lab:
8786 return "Par LAB";
8787 case _allocate_par_spooling_space:
8788 return "Par Spooling Space";
8789 case _adaptive_size_policy:
8790 return "Ergonomics";
8791 default:
8792 return "unknown";
8793 }
8794 }
8795
8796 void CMSDrainMarkingStackClosure::do_void() {
8797 // the max number to take from overflow list at a time
8798 const size_t num = _mark_stack->capacity()/4;
8799 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8800 "Overflow list should be NULL during concurrent phases");
8801 while (!_mark_stack->isEmpty() ||
8802 // if stack is empty, check the overflow list
8803 _collector->take_from_overflow_list(num, _mark_stack)) {
8804 oop obj = _mark_stack->pop();
8805 HeapWord* addr = (HeapWord*)obj;
8806 assert(_span.contains(addr), "Should be within span");
8807 assert(_bit_map->isMarked(addr), "Should be marked");
8808 assert(obj->is_oop(), "Should be an oop");
8809 obj->oop_iterate(_keep_alive);
8810 }
8811 }
8812
8813 void CMSParDrainMarkingStackClosure::do_void() {
8814 // drain queue
8815 trim_queue(0);
8816 }
8817
8818 // Trim our work_queue so its length is below max at return
8819 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8820 while (_work_queue->size() > max) {
8821 oop new_oop;
8822 if (_work_queue->pop_local(new_oop)) {
8823 assert(new_oop->is_oop(), "Expected an oop");
8824 assert(_bit_map->isMarked((HeapWord*)new_oop),
8825 "no white objects on this stack!");
8826 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8827 // iterate over the oops in this oop, marking and pushing
8828 // the ones in CMS heap (i.e. in _span).
8829 new_oop->oop_iterate(&_mark_and_push);
8830 }
8831 }
8832 }
8833
8834 ////////////////////////////////////////////////////////////////////
8835 // Support for Marking Stack Overflow list handling and related code
8836 ////////////////////////////////////////////////////////////////////
8837 // Much of the following code is similar in shape and spirit to the
8838 // code used in ParNewGC. We should try and share that code
8839 // as much as possible in the future.
8840
8841 #ifndef PRODUCT
8842 // Debugging support for CMSStackOverflowALot
8843
8844 // It's OK to call this multi-threaded; the worst thing
8845 // that can happen is that we'll get a bunch of closely
8846 // spaced simulated oveflows, but that's OK, in fact
8847 // probably good as it would exercise the overflow code
8848 // under contention.
8849 bool CMSCollector::simulate_overflow() {
8850 if (_overflow_counter-- <= 0) { // just being defensive
8851 _overflow_counter = CMSMarkStackOverflowInterval;
8852 return true;
8853 } else {
8854 return false;
8855 }
8856 }
8857
8858 bool CMSCollector::par_simulate_overflow() {
8859 return simulate_overflow();
8860 }
8861 #endif
8862
8863 // Single-threaded
8864 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
8865 assert(stack->isEmpty(), "Expected precondition");
8866 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
8867 size_t i = num;
8868 oop cur = _overflow_list;
8869 const markOop proto = markOopDesc::prototype();
8870 NOT_PRODUCT(ssize_t n = 0;)
8871 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
8872 next = oop(cur->mark());
8873 cur->set_mark(proto); // until proven otherwise
8874 assert(cur->is_oop(), "Should be an oop");
8875 bool res = stack->push(cur);
8876 assert(res, "Bit off more than can chew?");
8877 NOT_PRODUCT(n++;)
8878 }
8879 _overflow_list = cur;
8880 #ifndef PRODUCT
8881 assert(_num_par_pushes >= n, "Too many pops?");
8882 _num_par_pushes -=n;
8883 #endif
8884 return !stack->isEmpty();
8885 }
8886
8887 #define BUSY (oop(0x1aff1aff))
8888 // (MT-safe) Get a prefix of at most "num" from the list.
8889 // The overflow list is chained through the mark word of
8890 // each object in the list. We fetch the entire list,
8891 // break off a prefix of the right size and return the
8892 // remainder. If other threads try to take objects from
8893 // the overflow list at that time, they will wait for
8894 // some time to see if data becomes available. If (and
8895 // only if) another thread places one or more object(s)
8896 // on the global list before we have returned the suffix
8897 // to the global list, we will walk down our local list
8898 // to find its end and append the global list to
8899 // our suffix before returning it. This suffix walk can
8900 // prove to be expensive (quadratic in the amount of traffic)
8901 // when there are many objects in the overflow list and
8902 // there is much producer-consumer contention on the list.
8903 // *NOTE*: The overflow list manipulation code here and
8904 // in ParNewGeneration:: are very similar in shape,
8905 // except that in the ParNew case we use the old (from/eden)
8906 // copy of the object to thread the list via its klass word.
8907 // Because of the common code, if you make any changes in
8908 // the code below, please check the ParNew version to see if
8909 // similar changes might be needed.
8910 // CR 6797058 has been filed to consolidate the common code.
8911 bool CMSCollector::par_take_from_overflow_list(size_t num,
8912 OopTaskQueue* work_q,
8913 int no_of_gc_threads) {
8914 assert(work_q->size() == 0, "First empty local work queue");
8915 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
8916 if (_overflow_list == NULL) {
8917 return false;
8918 }
8919 // Grab the entire list; we'll put back a suffix
8920 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8921 Thread* tid = Thread::current();
8922 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
8923 // set to ParallelGCThreads.
8924 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
8925 size_t sleep_time_millis = MAX2((size_t)1, num/100);
8926 // If the list is busy, we spin for a short while,
8927 // sleeping between attempts to get the list.
8928 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
8929 os::sleep(tid, sleep_time_millis, false);
8930 if (_overflow_list == NULL) {
8931 // Nothing left to take
8932 return false;
8933 } else if (_overflow_list != BUSY) {
8934 // Try and grab the prefix
8935 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
8936 }
8937 }
8938 // If the list was found to be empty, or we spun long
8939 // enough, we give up and return empty-handed. If we leave
8940 // the list in the BUSY state below, it must be the case that
8941 // some other thread holds the overflow list and will set it
8942 // to a non-BUSY state in the future.
8943 if (prefix == NULL || prefix == BUSY) {
8944 // Nothing to take or waited long enough
8945 if (prefix == NULL) {
8946 // Write back the NULL in case we overwrote it with BUSY above
8947 // and it is still the same value.
8948 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8949 }
8950 return false;
8951 }
8952 assert(prefix != NULL && prefix != BUSY, "Error");
8953 size_t i = num;
8954 oop cur = prefix;
8955 // Walk down the first "num" objects, unless we reach the end.
8956 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
8957 if (cur->mark() == NULL) {
8958 // We have "num" or fewer elements in the list, so there
8959 // is nothing to return to the global list.
8960 // Write back the NULL in lieu of the BUSY we wrote
8961 // above, if it is still the same value.
8962 if (_overflow_list == BUSY) {
8963 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
8964 }
8965 } else {
8966 // Chop off the suffix and rerturn it to the global list.
8967 assert(cur->mark() != BUSY, "Error");
8968 oop suffix_head = cur->mark(); // suffix will be put back on global list
8969 cur->set_mark(NULL); // break off suffix
8970 // It's possible that the list is still in the empty(busy) state
8971 // we left it in a short while ago; in that case we may be
8972 // able to place back the suffix without incurring the cost
8973 // of a walk down the list.
8974 oop observed_overflow_list = _overflow_list;
8975 oop cur_overflow_list = observed_overflow_list;
8976 bool attached = false;
8977 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
8978 observed_overflow_list =
8979 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
8980 if (cur_overflow_list == observed_overflow_list) {
8981 attached = true;
8982 break;
8983 } else cur_overflow_list = observed_overflow_list;
8984 }
8985 if (!attached) {
8986 // Too bad, someone else sneaked in (at least) an element; we'll need
8987 // to do a splice. Find tail of suffix so we can prepend suffix to global
8988 // list.
8989 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
8990 oop suffix_tail = cur;
8991 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
8992 "Tautology");
8993 observed_overflow_list = _overflow_list;
8994 do {
8995 cur_overflow_list = observed_overflow_list;
8996 if (cur_overflow_list != BUSY) {
8997 // Do the splice ...
8998 suffix_tail->set_mark(markOop(cur_overflow_list));
8999 } else { // cur_overflow_list == BUSY
9000 suffix_tail->set_mark(NULL);
9001 }
9002 // ... and try to place spliced list back on overflow_list ...
9003 observed_overflow_list =
9004 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9005 } while (cur_overflow_list != observed_overflow_list);
9006 // ... until we have succeeded in doing so.
9007 }
9008 }
9009
9010 // Push the prefix elements on work_q
9011 assert(prefix != NULL, "control point invariant");
9012 const markOop proto = markOopDesc::prototype();
9013 oop next;
9014 NOT_PRODUCT(ssize_t n = 0;)
9015 for (cur = prefix; cur != NULL; cur = next) {
9016 next = oop(cur->mark());
9017 cur->set_mark(proto); // until proven otherwise
9018 assert(cur->is_oop(), "Should be an oop");
9019 bool res = work_q->push(cur);
9020 assert(res, "Bit off more than we can chew?");
9021 NOT_PRODUCT(n++;)
9022 }
9023 #ifndef PRODUCT
9024 assert(_num_par_pushes >= n, "Too many pops?");
9025 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
9026 #endif
9027 return true;
9028 }
9029
9030 // Single-threaded
9031 void CMSCollector::push_on_overflow_list(oop p) {
9032 NOT_PRODUCT(_num_par_pushes++;)
9033 assert(p->is_oop(), "Not an oop");
9034 preserve_mark_if_necessary(p);
9035 p->set_mark((markOop)_overflow_list);
9036 _overflow_list = p;
9037 }
9038
9039 // Multi-threaded; use CAS to prepend to overflow list
9040 void CMSCollector::par_push_on_overflow_list(oop p) {
9041 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9042 assert(p->is_oop(), "Not an oop");
9043 par_preserve_mark_if_necessary(p);
9044 oop observed_overflow_list = _overflow_list;
9045 oop cur_overflow_list;
9046 do {
9047 cur_overflow_list = observed_overflow_list;
9048 if (cur_overflow_list != BUSY) {
9049 p->set_mark(markOop(cur_overflow_list));
9050 } else {
9051 p->set_mark(NULL);
9052 }
9053 observed_overflow_list =
9054 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9055 } while (cur_overflow_list != observed_overflow_list);
9056 }
9057 #undef BUSY
9058
9059 // Single threaded
9060 // General Note on GrowableArray: pushes may silently fail
9061 // because we are (temporarily) out of C-heap for expanding
9062 // the stack. The problem is quite ubiquitous and affects
9063 // a lot of code in the JVM. The prudent thing for GrowableArray
9064 // to do (for now) is to exit with an error. However, that may
9065 // be too draconian in some cases because the caller may be
9066 // able to recover without much harm. For such cases, we
9067 // should probably introduce a "soft_push" method which returns
9068 // an indication of success or failure with the assumption that
9069 // the caller may be able to recover from a failure; code in
9070 // the VM can then be changed, incrementally, to deal with such
9071 // failures where possible, thus, incrementally hardening the VM
9072 // in such low resource situations.
9073 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9074 _preserved_oop_stack.push(p);
9075 _preserved_mark_stack.push(m);
9076 assert(m == p->mark(), "Mark word changed");
9077 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9078 "bijection");
9079 }
9080
9081 // Single threaded
9082 void CMSCollector::preserve_mark_if_necessary(oop p) {
9083 markOop m = p->mark();
9084 if (m->must_be_preserved(p)) {
9085 preserve_mark_work(p, m);
9086 }
9087 }
9088
9089 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9090 markOop m = p->mark();
9091 if (m->must_be_preserved(p)) {
9092 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9093 // Even though we read the mark word without holding
9094 // the lock, we are assured that it will not change
9095 // because we "own" this oop, so no other thread can
9096 // be trying to push it on the overflow list; see
9097 // the assertion in preserve_mark_work() that checks
9098 // that m == p->mark().
9099 preserve_mark_work(p, m);
9100 }
9101 }
9102
9103 // We should be able to do this multi-threaded,
9104 // a chunk of stack being a task (this is
9105 // correct because each oop only ever appears
9106 // once in the overflow list. However, it's
9107 // not very easy to completely overlap this with
9108 // other operations, so will generally not be done
9109 // until all work's been completed. Because we
9110 // expect the preserved oop stack (set) to be small,
9111 // it's probably fine to do this single-threaded.
9112 // We can explore cleverer concurrent/overlapped/parallel
9113 // processing of preserved marks if we feel the
9114 // need for this in the future. Stack overflow should
9115 // be so rare in practice and, when it happens, its
9116 // effect on performance so great that this will
9117 // likely just be in the noise anyway.
9118 void CMSCollector::restore_preserved_marks_if_any() {
9119 assert(SafepointSynchronize::is_at_safepoint(),
9120 "world should be stopped");
9121 assert(Thread::current()->is_ConcurrentGC_thread() ||
9122 Thread::current()->is_VM_thread(),
9123 "should be single-threaded");
9124 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9125 "bijection");
9126
9127 while (!_preserved_oop_stack.is_empty()) {
9128 oop p = _preserved_oop_stack.pop();
9129 assert(p->is_oop(), "Should be an oop");
9130 assert(_span.contains(p), "oop should be in _span");
9131 assert(p->mark() == markOopDesc::prototype(),
9132 "Set when taken from overflow list");
9133 markOop m = _preserved_mark_stack.pop();
9134 p->set_mark(m);
9135 }
9136 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9137 "stacks were cleared above");
9138 }
9139
9140 #ifndef PRODUCT
9141 bool CMSCollector::no_preserved_marks() const {
9142 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9143 }
9144 #endif
9145
9146 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9147 {
9148 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9149 CMSAdaptiveSizePolicy* size_policy =
9150 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9151 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9152 "Wrong type for size policy");
9153 return size_policy;
9154 }
9155
9156 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9157 size_t desired_promo_size) {
9158 if (cur_promo_size < desired_promo_size) {
9159 size_t expand_bytes = desired_promo_size - cur_promo_size;
9160 if (PrintAdaptiveSizePolicy && Verbose) {
9161 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9162 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9163 expand_bytes);
9164 }
9165 expand(expand_bytes,
9166 MinHeapDeltaBytes,
9167 CMSExpansionCause::_adaptive_size_policy);
9168 } else if (desired_promo_size < cur_promo_size) {
9169 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9170 if (PrintAdaptiveSizePolicy && Verbose) {
9171 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9172 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9173 shrink_bytes);
9174 }
9175 shrink(shrink_bytes);
9176 }
9177 }
9178
9179 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9180 GenCollectedHeap* gch = GenCollectedHeap::heap();
9181 CMSGCAdaptivePolicyCounters* counters =
9182 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9183 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9184 "Wrong kind of counters");
9185 return counters;
9186 }
9187
9188
9189 void ASConcurrentMarkSweepGeneration::update_counters() {
9190 if (UsePerfData) {
9191 _space_counters->update_all();
9192 _gen_counters->update_all();
9193 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9194 GenCollectedHeap* gch = GenCollectedHeap::heap();
9195 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9196 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9197 "Wrong gc statistics type");
9198 counters->update_counters(gc_stats_l);
9199 }
9200 }
9201
9202 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9203 if (UsePerfData) {
9204 _space_counters->update_used(used);
9205 _space_counters->update_capacity();
9206 _gen_counters->update_all();
9207
9208 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9209 GenCollectedHeap* gch = GenCollectedHeap::heap();
9210 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9211 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9212 "Wrong gc statistics type");
9213 counters->update_counters(gc_stats_l);
9214 }
9215 }
9216
9217 // The desired expansion delta is computed so that:
9218 // . desired free percentage or greater is used
9219 void ASConcurrentMarkSweepGeneration::compute_new_size() {
9220 assert_locked_or_safepoint(Heap_lock);
9221
9222 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9223
9224 // If incremental collection failed, we just want to expand
9225 // to the limit.
9226 if (incremental_collection_failed()) {
9227 clear_incremental_collection_failed();
9228 grow_to_reserved();
9229 return;
9230 }
9231
9232 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing");
9233
9234 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
9235 "Wrong type of heap");
9236 int prev_level = level() - 1;
9237 assert(prev_level >= 0, "The cms generation is the lowest generation");
9238 Generation* prev_gen = gch->get_gen(prev_level);
9239 assert(prev_gen->kind() == Generation::ASParNew,
9240 "Wrong type of young generation");
9241 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen;
9242 size_t cur_eden = younger_gen->eden()->capacity();
9243 CMSAdaptiveSizePolicy* size_policy = cms_size_policy();
9244 size_t cur_promo = free();
9245 size_policy->compute_tenured_generation_free_space(cur_promo,
9246 max_available(),
9247 cur_eden);
9248 resize(cur_promo, size_policy->promo_size());
9249
9250 // Record the new size of the space in the cms generation
9251 // that is available for promotions. This is temporary.
9252 // It should be the desired promo size.
9253 size_policy->avg_cms_promo()->sample(free());
9254 size_policy->avg_old_live()->sample(used());
9255
9256 if (UsePerfData) {
9257 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9258 counters->update_cms_capacity_counter(capacity());
9259 }
9260 }
9261
9262 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9263 assert_locked_or_safepoint(Heap_lock);
9264 assert_lock_strong(freelistLock());
9265 HeapWord* old_end = _cmsSpace->end();
9266 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9267 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9268 FreeChunk* chunk_at_end = find_chunk_at_end();
9269 if (chunk_at_end == NULL) {
9270 // No room to shrink
9271 if (PrintGCDetails && Verbose) {
9272 gclog_or_tty->print_cr("No room to shrink: old_end "
9273 PTR_FORMAT " unallocated_start " PTR_FORMAT
9274 " chunk_at_end " PTR_FORMAT,
9275 old_end, unallocated_start, chunk_at_end);
9276 }
9277 return;
9278 } else {
9279
9280 // Find the chunk at the end of the space and determine
9281 // how much it can be shrunk.
9282 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9283 size_t aligned_shrinkable_size_in_bytes =
9284 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9285 assert(unallocated_start <= chunk_at_end->end(),
9286 "Inconsistent chunk at end of space");
9287 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9288 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9289
9290 // Shrink the underlying space
9291 _virtual_space.shrink_by(bytes);
9292 if (PrintGCDetails && Verbose) {
9293 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9294 " desired_bytes " SIZE_FORMAT
9295 " shrinkable_size_in_bytes " SIZE_FORMAT
9296 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9297 " bytes " SIZE_FORMAT,
9298 desired_bytes, shrinkable_size_in_bytes,
9299 aligned_shrinkable_size_in_bytes, bytes);
9300 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9301 " unallocated_start " SIZE_FORMAT,
9302 old_end, unallocated_start);
9303 }
9304
9305 // If the space did shrink (shrinking is not guaranteed),
9306 // shrink the chunk at the end by the appropriate amount.
9307 if (((HeapWord*)_virtual_space.high()) < old_end) {
9308 size_t new_word_size =
9309 heap_word_size(_virtual_space.committed_size());
9310
9311 // Have to remove the chunk from the dictionary because it is changing
9312 // size and might be someplace elsewhere in the dictionary.
9313
9314 // Get the chunk at end, shrink it, and put it
9315 // back.
9316 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9317 size_t word_size_change = word_size_before - new_word_size;
9318 size_t chunk_at_end_old_size = chunk_at_end->size();
9319 assert(chunk_at_end_old_size >= word_size_change,
9320 "Shrink is too large");
9321 chunk_at_end->set_size(chunk_at_end_old_size -
9322 word_size_change);
9323 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9324 word_size_change);
9325
9326 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9327
9328 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9329 _bts->resize(new_word_size); // resize the block offset shared array
9330 Universe::heap()->barrier_set()->resize_covered_region(mr);
9331 _cmsSpace->assert_locked();
9332 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9333
9334 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9335
9336 // update the space and generation capacity counters
9337 if (UsePerfData) {
9338 _space_counters->update_capacity();
9339 _gen_counters->update_all();
9340 }
9341
9342 if (Verbose && PrintGCDetails) {
9343 size_t new_mem_size = _virtual_space.committed_size();
9344 size_t old_mem_size = new_mem_size + bytes;
9345 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK",
9346 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9347 }
9348 }
9349
9350 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9351 "Inconsistency at end of space");
9352 assert(chunk_at_end->end() == _cmsSpace->end(),
9353 "Shrinking is inconsistent");
9354 return;
9355 }
9356 }
9357
9358 // Transfer some number of overflown objects to usual marking
9359 // stack. Return true if some objects were transferred.
9360 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9361 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9362 (size_t)ParGCDesiredObjsFromOverflowList);
9363
9364 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9365 assert(_collector->overflow_list_is_empty() || res,
9366 "If list is not empty, we should have taken something");
9367 assert(!res || !_mark_stack->isEmpty(),
9368 "If we took something, it should now be on our stack");
9369 return res;
9370 }
9371
9372 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9373 size_t res = _sp->block_size_no_stall(addr, _collector);
9374 if (_sp->block_is_obj(addr)) {
9375 if (_live_bit_map->isMarked(addr)) {
9376 // It can't have been dead in a previous cycle
9377 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9378 } else {
9379 _dead_bit_map->mark(addr); // mark the dead object
9380 }
9381 }
9382 // Could be 0, if the block size could not be computed without stalling.
9383 return res;
9384 }
9385
9386 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9387
9388 switch (phase) {
9389 case CMSCollector::InitialMarking:
9390 initialize(true /* fullGC */ ,
9391 cause /* cause of the GC */,
9392 true /* recordGCBeginTime */,
9393 true /* recordPreGCUsage */,
9394 false /* recordPeakUsage */,
9395 false /* recordPostGCusage */,
9396 true /* recordAccumulatedGCTime */,
9397 false /* recordGCEndTime */,
9398 false /* countCollection */ );
9399 break;
9400
9401 case CMSCollector::FinalMarking:
9402 initialize(true /* fullGC */ ,
9403 cause /* cause of the GC */,
9404 false /* recordGCBeginTime */,
9405 false /* recordPreGCUsage */,
9406 false /* recordPeakUsage */,
9407 false /* recordPostGCusage */,
9408 true /* recordAccumulatedGCTime */,
9409 false /* recordGCEndTime */,
9410 false /* countCollection */ );
9411 break;
9412
9413 case CMSCollector::Sweeping:
9414 initialize(true /* fullGC */ ,
9415 cause /* cause of the GC */,
9416 false /* recordGCBeginTime */,
9417 false /* recordPreGCUsage */,
9418 true /* recordPeakUsage */,
9419 true /* recordPostGCusage */,
9420 false /* recordAccumulatedGCTime */,
9421 true /* recordGCEndTime */,
9422 true /* countCollection */ );
9423 break;
9424
9425 default:
9426 ShouldNotReachHere();
9427 }
9428 }
9429