1 /*
2 * Copyright (c) 2001, 2013, 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/classLoaderData.hpp"
27 #include "classfile/symbolTable.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "code/codeCache.hpp"
30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp"
32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp"
34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp"
36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp"
37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp"
38 #include "gc_implementation/parNew/parNewGeneration.hpp"
39 #include "gc_implementation/shared/collectorCounters.hpp"
40 #include "gc_implementation/shared/gcTimer.hpp"
41 #include "gc_implementation/shared/gcTrace.hpp"
42 #include "gc_implementation/shared/gcTraceTime.hpp"
43 #include "gc_implementation/shared/isGCActiveMark.hpp"
44 #include "gc_interface/collectedHeap.inline.hpp"
45 #include "memory/allocation.hpp"
46 #include "memory/cardTableRS.hpp"
47 #include "memory/collectorPolicy.hpp"
48 #include "memory/gcLocker.inline.hpp"
49 #include "memory/genCollectedHeap.hpp"
50 #include "memory/genMarkSweep.hpp"
51 #include "memory/genOopClosures.inline.hpp"
52 #include "memory/iterator.hpp"
53 #include "memory/padded.hpp"
54 #include "memory/referencePolicy.hpp"
55 #include "memory/resourceArea.hpp"
56 #include "memory/tenuredGeneration.hpp"
57 #include "oops/oop.inline.hpp"
58 #include "prims/jvmtiExport.hpp"
59 #include "runtime/globals_extension.hpp"
60 #include "runtime/handles.inline.hpp"
61 #include "runtime/java.hpp"
62 #include "runtime/vmThread.hpp"
63 #include "services/memoryService.hpp"
64 #include "services/runtimeService.hpp"
65
66 // statics
67 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
68 bool CMSCollector::_full_gc_requested = false;
69 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
70
71 //////////////////////////////////////////////////////////////////
72 // In support of CMS/VM thread synchronization
73 //////////////////////////////////////////////////////////////////
74 // We split use of the CGC_lock into 2 "levels".
75 // The low-level locking is of the usual CGC_lock monitor. We introduce
76 // a higher level "token" (hereafter "CMS token") built on top of the
77 // low level monitor (hereafter "CGC lock").
78 // The token-passing protocol gives priority to the VM thread. The
79 // CMS-lock doesn't provide any fairness guarantees, but clients
80 // should ensure that it is only held for very short, bounded
81 // durations.
82 //
83 // When either of the CMS thread or the VM thread is involved in
84 // collection operations during which it does not want the other
85 // thread to interfere, it obtains the CMS token.
86 //
87 // If either thread tries to get the token while the other has
88 // it, that thread waits. However, if the VM thread and CMS thread
89 // both want the token, then the VM thread gets priority while the
90 // CMS thread waits. This ensures, for instance, that the "concurrent"
91 // phases of the CMS thread's work do not block out the VM thread
92 // for long periods of time as the CMS thread continues to hog
93 // the token. (See bug 4616232).
94 //
95 // The baton-passing functions are, however, controlled by the
96 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
97 // and here the low-level CMS lock, not the high level token,
98 // ensures mutual exclusion.
99 //
100 // Two important conditions that we have to satisfy:
101 // 1. if a thread does a low-level wait on the CMS lock, then it
102 // relinquishes the CMS token if it were holding that token
103 // when it acquired the low-level CMS lock.
104 // 2. any low-level notifications on the low-level lock
105 // should only be sent when a thread has relinquished the token.
106 //
107 // In the absence of either property, we'd have potential deadlock.
108 //
109 // We protect each of the CMS (concurrent and sequential) phases
110 // with the CMS _token_, not the CMS _lock_.
111 //
112 // The only code protected by CMS lock is the token acquisition code
113 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
114 // baton-passing code.
115 //
116 // Unfortunately, i couldn't come up with a good abstraction to factor and
117 // hide the naked CGC_lock manipulation in the baton-passing code
118 // further below. That's something we should try to do. Also, the proof
119 // of correctness of this 2-level locking scheme is far from obvious,
120 // and potentially quite slippery. We have an uneasy suspicion, for instance,
121 // that there may be a theoretical possibility of delay/starvation in the
122 // low-level lock/wait/notify scheme used for the baton-passing because of
123 // potential interference with the priority scheme embodied in the
124 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
125 // invocation further below and marked with "XXX 20011219YSR".
126 // Indeed, as we note elsewhere, this may become yet more slippery
127 // in the presence of multiple CMS and/or multiple VM threads. XXX
128
129 class CMSTokenSync: public StackObj {
130 private:
131 bool _is_cms_thread;
132 public:
133 CMSTokenSync(bool is_cms_thread):
134 _is_cms_thread(is_cms_thread) {
135 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
136 "Incorrect argument to constructor");
137 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
138 }
139
140 ~CMSTokenSync() {
141 assert(_is_cms_thread ?
142 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
143 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
144 "Incorrect state");
145 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
146 }
147 };
148
149 // Convenience class that does a CMSTokenSync, and then acquires
150 // upto three locks.
151 class CMSTokenSyncWithLocks: public CMSTokenSync {
152 private:
153 // Note: locks are acquired in textual declaration order
154 // and released in the opposite order
155 MutexLockerEx _locker1, _locker2, _locker3;
156 public:
157 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
158 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
159 CMSTokenSync(is_cms_thread),
160 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
161 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
162 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
163 { }
164 };
165
166
167 // Wrapper class to temporarily disable icms during a foreground cms collection.
168 class ICMSDisabler: public StackObj {
169 public:
170 // The ctor disables icms and wakes up the thread so it notices the change;
171 // the dtor re-enables icms. Note that the CMSCollector methods will check
172 // CMSIncrementalMode.
173 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); }
174 ~ICMSDisabler() { CMSCollector::enable_icms(); }
175 };
176
177 //////////////////////////////////////////////////////////////////
178 // Concurrent Mark-Sweep Generation /////////////////////////////
179 //////////////////////////////////////////////////////////////////
180
181 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
182
183 // This struct contains per-thread things necessary to support parallel
184 // young-gen collection.
185 class CMSParGCThreadState: public CHeapObj<mtGC> {
186 public:
187 CFLS_LAB lab;
188 PromotionInfo promo;
189
190 // Constructor.
191 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
192 promo.setSpace(cfls);
193 }
194 };
195
196 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
197 ReservedSpace rs, size_t initial_byte_size, int level,
198 CardTableRS* ct, bool use_adaptive_freelists,
199 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) :
200 CardGeneration(rs, initial_byte_size, level, ct),
201 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
202 _debug_collection_type(Concurrent_collection_type),
203 _did_compact(false)
204 {
205 HeapWord* bottom = (HeapWord*) _virtual_space.low();
206 HeapWord* end = (HeapWord*) _virtual_space.high();
207
208 _direct_allocated_words = 0;
209 NOT_PRODUCT(
210 _numObjectsPromoted = 0;
211 _numWordsPromoted = 0;
212 _numObjectsAllocated = 0;
213 _numWordsAllocated = 0;
214 )
215
216 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end),
217 use_adaptive_freelists,
218 dictionaryChoice);
219 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
220 if (_cmsSpace == NULL) {
221 vm_exit_during_initialization(
222 "CompactibleFreeListSpace allocation failure");
223 }
224 _cmsSpace->_gen = this;
225
226 _gc_stats = new CMSGCStats();
227
228 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
229 // offsets match. The ability to tell free chunks from objects
230 // depends on this property.
231 debug_only(
232 FreeChunk* junk = NULL;
233 assert(UseCompressedClassPointers ||
234 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
235 "Offset of FreeChunk::_prev within FreeChunk must match"
236 " that of OopDesc::_klass within OopDesc");
237 )
238 if (CollectedHeap::use_parallel_gc_threads()) {
239 typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
240 _par_gc_thread_states =
241 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC);
242 if (_par_gc_thread_states == NULL) {
243 vm_exit_during_initialization("Could not allocate par gc structs");
244 }
245 for (uint i = 0; i < ParallelGCThreads; i++) {
246 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
247 if (_par_gc_thread_states[i] == NULL) {
248 vm_exit_during_initialization("Could not allocate par gc structs");
249 }
250 }
251 } else {
252 _par_gc_thread_states = NULL;
253 }
254 _incremental_collection_failed = false;
255 // The "dilatation_factor" is the expansion that can occur on
256 // account of the fact that the minimum object size in the CMS
257 // generation may be larger than that in, say, a contiguous young
258 // generation.
259 // Ideally, in the calculation below, we'd compute the dilatation
260 // factor as: MinChunkSize/(promoting_gen's min object size)
261 // Since we do not have such a general query interface for the
262 // promoting generation, we'll instead just use the minimum
263 // object size (which today is a header's worth of space);
264 // note that all arithmetic is in units of HeapWords.
265 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
266 assert(_dilatation_factor >= 1.0, "from previous assert");
267 }
268
269
270 // The field "_initiating_occupancy" represents the occupancy percentage
271 // at which we trigger a new collection cycle. Unless explicitly specified
272 // via CMSInitiatingOccupancyFraction (argument "io" below), it
273 // is calculated by:
274 //
275 // Let "f" be MinHeapFreeRatio in
276 //
277 // _initiating_occupancy = 100-f +
278 // f * (CMSTriggerRatio/100)
279 // where CMSTriggerRatio is the argument "tr" below.
280 //
281 // That is, if we assume the heap is at its desired maximum occupancy at the
282 // end of a collection, we let CMSTriggerRatio of the (purported) free
283 // space be allocated before initiating a new collection cycle.
284 //
285 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
286 assert(io <= 100 && tr <= 100, "Check the arguments");
287 if (io >= 0) {
288 _initiating_occupancy = (double)io / 100.0;
289 } else {
290 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
291 (double)(tr * MinHeapFreeRatio) / 100.0)
292 / 100.0;
293 }
294 }
295
296 void ConcurrentMarkSweepGeneration::ref_processor_init() {
297 assert(collector() != NULL, "no collector");
298 collector()->ref_processor_init();
299 }
300
301 void CMSCollector::ref_processor_init() {
302 if (_ref_processor == NULL) {
303 // Allocate and initialize a reference processor
304 _ref_processor =
305 new ReferenceProcessor(_span, // span
306 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
307 (int) ParallelGCThreads, // mt processing degree
308 _cmsGen->refs_discovery_is_mt(), // mt discovery
309 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
310 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
311 &_is_alive_closure, // closure for liveness info
312 false); // next field updates do not need write barrier
313 // Initialize the _ref_processor field of CMSGen
314 _cmsGen->set_ref_processor(_ref_processor);
315
316 }
317 }
318
319 CMSAdaptiveSizePolicy* CMSCollector::size_policy() {
320 GenCollectedHeap* gch = GenCollectedHeap::heap();
321 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
322 "Wrong type of heap");
323 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
324 gch->gen_policy()->size_policy();
325 assert(sp->is_gc_cms_adaptive_size_policy(),
326 "Wrong type of size policy");
327 return sp;
328 }
329
330 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() {
331 CMSGCAdaptivePolicyCounters* results =
332 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters();
333 assert(
334 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
335 "Wrong gc policy counter kind");
336 return results;
337 }
338
339
340 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
341
342 const char* gen_name = "old";
343
344 // Generation Counters - generation 1, 1 subspace
345 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space);
346
347 _space_counters = new GSpaceCounters(gen_name, 0,
348 _virtual_space.reserved_size(),
349 this, _gen_counters);
350 }
351
352 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
353 _cms_gen(cms_gen)
354 {
355 assert(alpha <= 100, "bad value");
356 _saved_alpha = alpha;
357
358 // Initialize the alphas to the bootstrap value of 100.
359 _gc0_alpha = _cms_alpha = 100;
360
361 _cms_begin_time.update();
362 _cms_end_time.update();
363
364 _gc0_duration = 0.0;
365 _gc0_period = 0.0;
366 _gc0_promoted = 0;
367
368 _cms_duration = 0.0;
369 _cms_period = 0.0;
370 _cms_allocated = 0;
371
372 _cms_used_at_gc0_begin = 0;
373 _cms_used_at_gc0_end = 0;
374 _allow_duty_cycle_reduction = false;
375 _valid_bits = 0;
376 _icms_duty_cycle = CMSIncrementalDutyCycle;
377 }
378
379 double CMSStats::cms_free_adjustment_factor(size_t free) const {
380 // TBD: CR 6909490
381 return 1.0;
382 }
383
384 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
385 }
386
387 // If promotion failure handling is on use
388 // the padded average size of the promotion for each
389 // young generation collection.
390 double CMSStats::time_until_cms_gen_full() const {
391 size_t cms_free = _cms_gen->cmsSpace()->free();
392 GenCollectedHeap* gch = GenCollectedHeap::heap();
393 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(),
394 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
395 if (cms_free > expected_promotion) {
396 // Start a cms collection if there isn't enough space to promote
397 // for the next minor collection. Use the padded average as
398 // a safety factor.
399 cms_free -= expected_promotion;
400
401 // Adjust by the safety factor.
402 double cms_free_dbl = (double)cms_free;
403 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0;
404 // Apply a further correction factor which tries to adjust
405 // for recent occurance of concurrent mode failures.
406 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
407 cms_free_dbl = cms_free_dbl * cms_adjustment;
408
409 if (PrintGCDetails && Verbose) {
410 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free "
411 SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
412 cms_free, expected_promotion);
413 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f",
414 cms_free_dbl, cms_consumption_rate() + 1.0);
415 }
416 // Add 1 in case the consumption rate goes to zero.
417 return cms_free_dbl / (cms_consumption_rate() + 1.0);
418 }
419 return 0.0;
420 }
421
422 // Compare the duration of the cms collection to the
423 // time remaining before the cms generation is empty.
424 // Note that the time from the start of the cms collection
425 // to the start of the cms sweep (less than the total
426 // duration of the cms collection) can be used. This
427 // has been tried and some applications experienced
428 // promotion failures early in execution. This was
429 // possibly because the averages were not accurate
430 // enough at the beginning.
431 double CMSStats::time_until_cms_start() const {
432 // We add "gc0_period" to the "work" calculation
433 // below because this query is done (mostly) at the
434 // end of a scavenge, so we need to conservatively
435 // account for that much possible delay
436 // in the query so as to avoid concurrent mode failures
437 // due to starting the collection just a wee bit too
438 // late.
439 double work = cms_duration() + gc0_period();
440 double deadline = time_until_cms_gen_full();
441 // If a concurrent mode failure occurred recently, we want to be
442 // more conservative and halve our expected time_until_cms_gen_full()
443 if (work > deadline) {
444 if (Verbose && PrintGCDetails) {
445 gclog_or_tty->print(
446 " CMSCollector: collect because of anticipated promotion "
447 "before full %3.7f + %3.7f > %3.7f ", cms_duration(),
448 gc0_period(), time_until_cms_gen_full());
449 }
450 return 0.0;
451 }
452 return work - deadline;
453 }
454
455 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the
456 // amount of change to prevent wild oscillation.
457 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle,
458 unsigned int new_duty_cycle) {
459 assert(old_duty_cycle <= 100, "bad input value");
460 assert(new_duty_cycle <= 100, "bad input value");
461
462 // Note: use subtraction with caution since it may underflow (values are
463 // unsigned). Addition is safe since we're in the range 0-100.
464 unsigned int damped_duty_cycle = new_duty_cycle;
465 if (new_duty_cycle < old_duty_cycle) {
466 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U);
467 if (new_duty_cycle + largest_delta < old_duty_cycle) {
468 damped_duty_cycle = old_duty_cycle - largest_delta;
469 }
470 } else if (new_duty_cycle > old_duty_cycle) {
471 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U);
472 if (new_duty_cycle > old_duty_cycle + largest_delta) {
473 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U);
474 }
475 }
476 assert(damped_duty_cycle <= 100, "invalid duty cycle computed");
477
478 if (CMSTraceIncrementalPacing) {
479 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ",
480 old_duty_cycle, new_duty_cycle, damped_duty_cycle);
481 }
482 return damped_duty_cycle;
483 }
484
485 unsigned int CMSStats::icms_update_duty_cycle_impl() {
486 assert(CMSIncrementalPacing && valid(),
487 "should be handled in icms_update_duty_cycle()");
488
489 double cms_time_so_far = cms_timer().seconds();
490 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M;
491 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far);
492
493 // Avoid division by 0.
494 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01);
495 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full;
496
497 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U);
498 if (new_duty_cycle > _icms_duty_cycle) {
499 // Avoid very small duty cycles (1 or 2); 0 is allowed.
500 if (new_duty_cycle > 2) {
501 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle,
502 new_duty_cycle);
503 }
504 } else if (_allow_duty_cycle_reduction) {
505 // The duty cycle is reduced only once per cms cycle (see record_cms_end()).
506 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle);
507 // Respect the minimum duty cycle.
508 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin;
509 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle);
510 }
511
512 if (PrintGCDetails || CMSTraceIncrementalPacing) {
513 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle);
514 }
515
516 _allow_duty_cycle_reduction = false;
517 return _icms_duty_cycle;
518 }
519
520 #ifndef PRODUCT
521 void CMSStats::print_on(outputStream *st) const {
522 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
523 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
524 gc0_duration(), gc0_period(), gc0_promoted());
525 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
526 cms_duration(), cms_duration_per_mb(),
527 cms_period(), cms_allocated());
528 st->print(",cms_since_beg=%g,cms_since_end=%g",
529 cms_time_since_begin(), cms_time_since_end());
530 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
531 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
532 if (CMSIncrementalMode) {
533 st->print(",dc=%d", icms_duty_cycle());
534 }
535
536 if (valid()) {
537 st->print(",promo_rate=%g,cms_alloc_rate=%g",
538 promotion_rate(), cms_allocation_rate());
539 st->print(",cms_consumption_rate=%g,time_until_full=%g",
540 cms_consumption_rate(), time_until_cms_gen_full());
541 }
542 st->print(" ");
543 }
544 #endif // #ifndef PRODUCT
545
546 CMSCollector::CollectorState CMSCollector::_collectorState =
547 CMSCollector::Idling;
548 bool CMSCollector::_foregroundGCIsActive = false;
549 bool CMSCollector::_foregroundGCShouldWait = false;
550
551 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
552 CardTableRS* ct,
553 ConcurrentMarkSweepPolicy* cp):
554 _cmsGen(cmsGen),
555 _ct(ct),
556 _ref_processor(NULL), // will be set later
557 _conc_workers(NULL), // may be set later
558 _abort_preclean(false),
559 _start_sampling(false),
560 _between_prologue_and_epilogue(false),
561 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
562 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize),
563 -1 /* lock-free */, "No_lock" /* dummy */),
564 _modUnionClosure(&_modUnionTable),
565 _modUnionClosurePar(&_modUnionTable),
566 // Adjust my span to cover old (cms) gen
567 _span(cmsGen->reserved()),
568 // Construct the is_alive_closure with _span & markBitMap
569 _is_alive_closure(_span, &_markBitMap),
570 _restart_addr(NULL),
571 _overflow_list(NULL),
572 _stats(cmsGen),
573 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)),
574 _eden_chunk_array(NULL), // may be set in ctor body
575 _eden_chunk_capacity(0), // -- ditto --
576 _eden_chunk_index(0), // -- ditto --
577 _survivor_plab_array(NULL), // -- ditto --
578 _survivor_chunk_array(NULL), // -- ditto --
579 _survivor_chunk_capacity(0), // -- ditto --
580 _survivor_chunk_index(0), // -- ditto --
581 _ser_pmc_preclean_ovflw(0),
582 _ser_kac_preclean_ovflw(0),
583 _ser_pmc_remark_ovflw(0),
584 _par_pmc_remark_ovflw(0),
585 _ser_kac_ovflw(0),
586 _par_kac_ovflw(0),
587 #ifndef PRODUCT
588 _num_par_pushes(0),
589 #endif
590 _collection_count_start(0),
591 _verifying(false),
592 _icms_start_limit(NULL),
593 _icms_stop_limit(NULL),
594 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
595 _completed_initialization(false),
596 _collector_policy(cp),
597 _should_unload_classes(CMSClassUnloadingEnabled),
598 _concurrent_cycles_since_last_unload(0),
599 _roots_scanning_options(SharedHeap::SO_None),
600 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
601 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
602 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
603 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
604 _cms_start_registered(false)
605 {
606 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) {
607 ExplicitGCInvokesConcurrent = true;
608 }
609 // Now expand the span and allocate the collection support structures
610 // (MUT, marking bit map etc.) to cover both generations subject to
611 // collection.
612
613 // For use by dirty card to oop closures.
614 _cmsGen->cmsSpace()->set_collector(this);
615
616 // Allocate MUT and marking bit map
617 {
618 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
619 if (!_markBitMap.allocate(_span)) {
620 warning("Failed to allocate CMS Bit Map");
621 return;
622 }
623 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
624 }
625 {
626 _modUnionTable.allocate(_span);
627 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
628 }
629
630 if (!_markStack.allocate(MarkStackSize)) {
631 warning("Failed to allocate CMS Marking Stack");
632 return;
633 }
634
635 // Support for multi-threaded concurrent phases
636 if (CMSConcurrentMTEnabled) {
637 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
638 // just for now
639 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
640 }
641 if (ConcGCThreads > 1) {
642 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads",
643 ConcGCThreads, true);
644 if (_conc_workers == NULL) {
645 warning("GC/CMS: _conc_workers allocation failure: "
646 "forcing -CMSConcurrentMTEnabled");
647 CMSConcurrentMTEnabled = false;
648 } else {
649 _conc_workers->initialize_workers();
650 }
651 } else {
652 CMSConcurrentMTEnabled = false;
653 }
654 }
655 if (!CMSConcurrentMTEnabled) {
656 ConcGCThreads = 0;
657 } else {
658 // Turn off CMSCleanOnEnter optimization temporarily for
659 // the MT case where it's not fixed yet; see 6178663.
660 CMSCleanOnEnter = false;
661 }
662 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
663 "Inconsistency");
664
665 // Parallel task queues; these are shared for the
666 // concurrent and stop-world phases of CMS, but
667 // are not shared with parallel scavenge (ParNew).
668 {
669 uint i;
670 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads);
671
672 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
673 || ParallelRefProcEnabled)
674 && num_queues > 0) {
675 _task_queues = new OopTaskQueueSet(num_queues);
676 if (_task_queues == NULL) {
677 warning("task_queues allocation failure.");
678 return;
679 }
680 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
681 if (_hash_seed == NULL) {
682 warning("_hash_seed array allocation failure");
683 return;
684 }
685
686 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
687 for (i = 0; i < num_queues; i++) {
688 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
689 if (q == NULL) {
690 warning("work_queue allocation failure.");
691 return;
692 }
693 _task_queues->register_queue(i, q);
694 }
695 for (i = 0; i < num_queues; i++) {
696 _task_queues->queue(i)->initialize();
697 _hash_seed[i] = 17; // copied from ParNew
698 }
699 }
700 }
701
702 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
703
704 // Clip CMSBootstrapOccupancy between 0 and 100.
705 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100;
706
707 _full_gcs_since_conc_gc = 0;
708
709 // Now tell CMS generations the identity of their collector
710 ConcurrentMarkSweepGeneration::set_collector(this);
711
712 // Create & start a CMS thread for this CMS collector
713 _cmsThread = ConcurrentMarkSweepThread::start(this);
714 assert(cmsThread() != NULL, "CMS Thread should have been created");
715 assert(cmsThread()->collector() == this,
716 "CMS Thread should refer to this gen");
717 assert(CGC_lock != NULL, "Where's the CGC_lock?");
718
719 // Support for parallelizing young gen rescan
720 GenCollectedHeap* gch = GenCollectedHeap::heap();
721 _young_gen = gch->prev_gen(_cmsGen);
722 if (gch->supports_inline_contig_alloc()) {
723 _top_addr = gch->top_addr();
724 _end_addr = gch->end_addr();
725 assert(_young_gen != NULL, "no _young_gen");
726 _eden_chunk_index = 0;
727 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain;
728 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
729 if (_eden_chunk_array == NULL) {
730 _eden_chunk_capacity = 0;
731 warning("GC/CMS: _eden_chunk_array allocation failure");
732 }
733 }
734 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error");
735
736 // Support for parallelizing survivor space rescan
737 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
738 const size_t max_plab_samples =
739 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize;
740
741 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
742 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC);
743 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
744 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL
745 || _cursor == NULL) {
746 warning("Failed to allocate survivor plab/chunk array");
747 if (_survivor_plab_array != NULL) {
748 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
749 _survivor_plab_array = NULL;
750 }
751 if (_survivor_chunk_array != NULL) {
752 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
753 _survivor_chunk_array = NULL;
754 }
755 if (_cursor != NULL) {
756 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC);
757 _cursor = NULL;
758 }
759 } else {
760 _survivor_chunk_capacity = 2*max_plab_samples;
761 for (uint i = 0; i < ParallelGCThreads; i++) {
762 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
763 if (vec == NULL) {
764 warning("Failed to allocate survivor plab array");
765 for (int j = i; j > 0; j--) {
766 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC);
767 }
768 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC);
769 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC);
770 _survivor_plab_array = NULL;
771 _survivor_chunk_array = NULL;
772 _survivor_chunk_capacity = 0;
773 break;
774 } else {
775 ChunkArray* cur =
776 ::new (&_survivor_plab_array[i]) ChunkArray(vec,
777 max_plab_samples);
778 assert(cur->end() == 0, "Should be 0");
779 assert(cur->array() == vec, "Should be vec");
780 assert(cur->capacity() == max_plab_samples, "Error");
781 }
782 }
783 }
784 }
785 assert( ( _survivor_plab_array != NULL
786 && _survivor_chunk_array != NULL)
787 || ( _survivor_chunk_capacity == 0
788 && _survivor_chunk_index == 0),
789 "Error");
790
791 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
792 _gc_counters = new CollectorCounters("CMS", 1);
793 _completed_initialization = true;
794 _inter_sweep_timer.start(); // start of time
795 }
796
797 const char* ConcurrentMarkSweepGeneration::name() const {
798 return "concurrent mark-sweep generation";
799 }
800 void ConcurrentMarkSweepGeneration::update_counters() {
801 if (UsePerfData) {
802 _space_counters->update_all();
803 _gen_counters->update_all();
804 }
805 }
806
807 // this is an optimized version of update_counters(). it takes the
808 // used value as a parameter rather than computing it.
809 //
810 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
811 if (UsePerfData) {
812 _space_counters->update_used(used);
813 _space_counters->update_capacity();
814 _gen_counters->update_all();
815 }
816 }
817
818 void ConcurrentMarkSweepGeneration::print() const {
819 Generation::print();
820 cmsSpace()->print();
821 }
822
823 #ifndef PRODUCT
824 void ConcurrentMarkSweepGeneration::print_statistics() {
825 cmsSpace()->printFLCensus(0);
826 }
827 #endif
828
829 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
830 GenCollectedHeap* gch = GenCollectedHeap::heap();
831 if (PrintGCDetails) {
832 if (Verbose) {
833 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
834 level(), short_name(), s, used(), capacity());
835 } else {
836 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
837 level(), short_name(), s, used() / K, capacity() / K);
838 }
839 }
840 if (Verbose) {
841 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
842 gch->used(), gch->capacity());
843 } else {
844 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
845 gch->used() / K, gch->capacity() / K);
846 }
847 }
848
849 size_t
850 ConcurrentMarkSweepGeneration::contiguous_available() const {
851 // dld proposes an improvement in precision here. If the committed
852 // part of the space ends in a free block we should add that to
853 // uncommitted size in the calculation below. Will make this
854 // change later, staying with the approximation below for the
855 // time being. -- ysr.
856 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
857 }
858
859 size_t
860 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
861 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
862 }
863
864 size_t ConcurrentMarkSweepGeneration::max_available() const {
865 return free() + _virtual_space.uncommitted_size();
866 }
867
868 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
869 size_t available = max_available();
870 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
871 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
872 if (Verbose && PrintGCDetails) {
873 gclog_or_tty->print_cr(
874 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
875 "max_promo("SIZE_FORMAT")",
876 res? "":" not", available, res? ">=":"<",
877 av_promo, max_promotion_in_bytes);
878 }
879 return res;
880 }
881
882 // At a promotion failure dump information on block layout in heap
883 // (cms old generation).
884 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
885 if (CMSDumpAtPromotionFailure) {
886 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
887 }
888 }
889
890 CompactibleSpace*
891 ConcurrentMarkSweepGeneration::first_compaction_space() const {
892 return _cmsSpace;
893 }
894
895 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
896 // Clear the promotion information. These pointers can be adjusted
897 // along with all the other pointers into the heap but
898 // compaction is expected to be a rare event with
899 // a heap using cms so don't do it without seeing the need.
900 if (CollectedHeap::use_parallel_gc_threads()) {
901 for (uint i = 0; i < ParallelGCThreads; i++) {
902 _par_gc_thread_states[i]->promo.reset();
903 }
904 }
905 }
906
907 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
908 blk->do_space(_cmsSpace);
909 }
910
911 void ConcurrentMarkSweepGeneration::compute_new_size() {
912 assert_locked_or_safepoint(Heap_lock);
913
914 // If incremental collection failed, we just want to expand
915 // to the limit.
916 if (incremental_collection_failed()) {
917 clear_incremental_collection_failed();
918 grow_to_reserved();
919 return;
920 }
921
922 // The heap has been compacted but not reset yet.
923 // Any metric such as free() or used() will be incorrect.
924
925 CardGeneration::compute_new_size();
926
927 // Reset again after a possible resizing
928 if (did_compact()) {
929 cmsSpace()->reset_after_compaction();
930 }
931 }
932
933 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
934 assert_locked_or_safepoint(Heap_lock);
935
936 // If incremental collection failed, we just want to expand
937 // to the limit.
938 if (incremental_collection_failed()) {
939 clear_incremental_collection_failed();
940 grow_to_reserved();
941 return;
942 }
943
944 double free_percentage = ((double) free()) / capacity();
945 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
946 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
947
948 // compute expansion delta needed for reaching desired free percentage
949 if (free_percentage < desired_free_percentage) {
950 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
951 assert(desired_capacity >= capacity(), "invalid expansion size");
952 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
953 if (PrintGCDetails && Verbose) {
954 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
955 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
956 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
957 gclog_or_tty->print_cr(" Desired free fraction %f",
958 desired_free_percentage);
959 gclog_or_tty->print_cr(" Maximum free fraction %f",
960 maximum_free_percentage);
961 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000);
962 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
963 desired_capacity/1000);
964 int prev_level = level() - 1;
965 if (prev_level >= 0) {
966 size_t prev_size = 0;
967 GenCollectedHeap* gch = GenCollectedHeap::heap();
968 Generation* prev_gen = gch->_gens[prev_level];
969 prev_size = prev_gen->capacity();
970 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
971 prev_size/1000);
972 }
973 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
974 unsafe_max_alloc_nogc()/1000);
975 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
976 contiguous_available()/1000);
977 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
978 expand_bytes);
979 }
980 // safe if expansion fails
981 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
982 if (PrintGCDetails && Verbose) {
983 gclog_or_tty->print_cr(" Expanded free fraction %f",
984 ((double) free()) / capacity());
985 }
986 } else {
987 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
988 assert(desired_capacity <= capacity(), "invalid expansion size");
989 size_t shrink_bytes = capacity() - desired_capacity;
990 // Don't shrink unless the delta is greater than the minimum shrink we want
991 if (shrink_bytes >= MinHeapDeltaBytes) {
992 shrink_free_list_by(shrink_bytes);
993 }
994 }
995 }
996
997 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
998 return cmsSpace()->freelistLock();
999 }
1000
1001 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1002 bool tlab) {
1003 CMSSynchronousYieldRequest yr;
1004 MutexLockerEx x(freelistLock(),
1005 Mutex::_no_safepoint_check_flag);
1006 return have_lock_and_allocate(size, tlab);
1007 }
1008
1009 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1010 bool tlab /* ignored */) {
1011 assert_lock_strong(freelistLock());
1012 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1013 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1014 // Allocate the object live (grey) if the background collector has
1015 // started marking. This is necessary because the marker may
1016 // have passed this address and consequently this object will
1017 // not otherwise be greyed and would be incorrectly swept up.
1018 // Note that if this object contains references, the writing
1019 // of those references will dirty the card containing this object
1020 // allowing the object to be blackened (and its references scanned)
1021 // either during a preclean phase or at the final checkpoint.
1022 if (res != NULL) {
1023 // We may block here with an uninitialized object with
1024 // its mark-bit or P-bits not yet set. Such objects need
1025 // to be safely navigable by block_start().
1026 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1027 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
1028 collector()->direct_allocated(res, adjustedSize);
1029 _direct_allocated_words += adjustedSize;
1030 // allocation counters
1031 NOT_PRODUCT(
1032 _numObjectsAllocated++;
1033 _numWordsAllocated += (int)adjustedSize;
1034 )
1035 }
1036 return res;
1037 }
1038
1039 // In the case of direct allocation by mutators in a generation that
1040 // is being concurrently collected, the object must be allocated
1041 // live (grey) if the background collector has started marking.
1042 // This is necessary because the marker may
1043 // have passed this address and consequently this object will
1044 // not otherwise be greyed and would be incorrectly swept up.
1045 // Note that if this object contains references, the writing
1046 // of those references will dirty the card containing this object
1047 // allowing the object to be blackened (and its references scanned)
1048 // either during a preclean phase or at the final checkpoint.
1049 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1050 assert(_markBitMap.covers(start, size), "Out of bounds");
1051 if (_collectorState >= Marking) {
1052 MutexLockerEx y(_markBitMap.lock(),
1053 Mutex::_no_safepoint_check_flag);
1054 // [see comments preceding SweepClosure::do_blk() below for details]
1055 //
1056 // Can the P-bits be deleted now? JJJ
1057 //
1058 // 1. need to mark the object as live so it isn't collected
1059 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1060 // 3. need to mark the end of the object so marking, precleaning or sweeping
1061 // can skip over uninitialized or unparsable objects. An allocated
1062 // object is considered uninitialized for our purposes as long as
1063 // its klass word is NULL. All old gen objects are parsable
1064 // as soon as they are initialized.)
1065 _markBitMap.mark(start); // object is live
1066 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1067 _markBitMap.mark(start + size - 1);
1068 // mark end of object
1069 }
1070 // check that oop looks uninitialized
1071 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1072 }
1073
1074 void CMSCollector::promoted(bool par, HeapWord* start,
1075 bool is_obj_array, size_t obj_size) {
1076 assert(_markBitMap.covers(start), "Out of bounds");
1077 // See comment in direct_allocated() about when objects should
1078 // be allocated live.
1079 if (_collectorState >= Marking) {
1080 // we already hold the marking bit map lock, taken in
1081 // the prologue
1082 if (par) {
1083 _markBitMap.par_mark(start);
1084 } else {
1085 _markBitMap.mark(start);
1086 }
1087 // We don't need to mark the object as uninitialized (as
1088 // in direct_allocated above) because this is being done with the
1089 // world stopped and the object will be initialized by the
1090 // time the marking, precleaning or sweeping get to look at it.
1091 // But see the code for copying objects into the CMS generation,
1092 // where we need to ensure that concurrent readers of the
1093 // block offset table are able to safely navigate a block that
1094 // is in flux from being free to being allocated (and in
1095 // transition while being copied into) and subsequently
1096 // becoming a bona-fide object when the copy/promotion is complete.
1097 assert(SafepointSynchronize::is_at_safepoint(),
1098 "expect promotion only at safepoints");
1099
1100 if (_collectorState < Sweeping) {
1101 // Mark the appropriate cards in the modUnionTable, so that
1102 // this object gets scanned before the sweep. If this is
1103 // not done, CMS generation references in the object might
1104 // not get marked.
1105 // For the case of arrays, which are otherwise precisely
1106 // marked, we need to dirty the entire array, not just its head.
1107 if (is_obj_array) {
1108 // The [par_]mark_range() method expects mr.end() below to
1109 // be aligned to the granularity of a bit's representation
1110 // in the heap. In the case of the MUT below, that's a
1111 // card size.
1112 MemRegion mr(start,
1113 (HeapWord*)round_to((intptr_t)(start + obj_size),
1114 CardTableModRefBS::card_size /* bytes */));
1115 if (par) {
1116 _modUnionTable.par_mark_range(mr);
1117 } else {
1118 _modUnionTable.mark_range(mr);
1119 }
1120 } else { // not an obj array; we can just mark the head
1121 if (par) {
1122 _modUnionTable.par_mark(start);
1123 } else {
1124 _modUnionTable.mark(start);
1125 }
1126 }
1127 }
1128 }
1129 }
1130
1131 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1132 {
1133 size_t delta = pointer_delta(addr, space->bottom());
1134 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1135 }
1136
1137 void CMSCollector::icms_update_allocation_limits()
1138 {
1139 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1140 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1141
1142 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1143 if (CMSTraceIncrementalPacing) {
1144 stats().print();
1145 }
1146
1147 assert(duty_cycle <= 100, "invalid duty cycle");
1148 if (duty_cycle != 0) {
1149 // The duty_cycle is a percentage between 0 and 100; convert to words and
1150 // then compute the offset from the endpoints of the space.
1151 size_t free_words = eden->free() / HeapWordSize;
1152 double free_words_dbl = (double)free_words;
1153 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1154 size_t offset_words = (free_words - duty_cycle_words) / 2;
1155
1156 _icms_start_limit = eden->top() + offset_words;
1157 _icms_stop_limit = eden->end() - offset_words;
1158
1159 // The limits may be adjusted (shifted to the right) by
1160 // CMSIncrementalOffset, to allow the application more mutator time after a
1161 // young gen gc (when all mutators were stopped) and before CMS starts and
1162 // takes away one or more cpus.
1163 if (CMSIncrementalOffset != 0) {
1164 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1165 size_t adjustment = (size_t)adjustment_dbl;
1166 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1167 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1168 _icms_start_limit += adjustment;
1169 _icms_stop_limit = tmp_stop;
1170 }
1171 }
1172 }
1173 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1174 _icms_start_limit = _icms_stop_limit = eden->end();
1175 }
1176
1177 // Install the new start limit.
1178 eden->set_soft_end(_icms_start_limit);
1179
1180 if (CMSTraceIncrementalMode) {
1181 gclog_or_tty->print(" icms alloc limits: "
1182 PTR_FORMAT "," PTR_FORMAT
1183 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1184 _icms_start_limit, _icms_stop_limit,
1185 percent_of_space(eden, _icms_start_limit),
1186 percent_of_space(eden, _icms_stop_limit));
1187 if (Verbose) {
1188 gclog_or_tty->print("eden: ");
1189 eden->print_on(gclog_or_tty);
1190 }
1191 }
1192 }
1193
1194 // Any changes here should try to maintain the invariant
1195 // that if this method is called with _icms_start_limit
1196 // and _icms_stop_limit both NULL, then it should return NULL
1197 // and not notify the icms thread.
1198 HeapWord*
1199 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1200 size_t word_size)
1201 {
1202 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1203 // nop.
1204 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1205 if (top <= _icms_start_limit) {
1206 if (CMSTraceIncrementalMode) {
1207 space->print_on(gclog_or_tty);
1208 gclog_or_tty->stamp();
1209 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1210 ", new limit=" PTR_FORMAT
1211 " (" SIZE_FORMAT "%%)",
1212 top, _icms_stop_limit,
1213 percent_of_space(space, _icms_stop_limit));
1214 }
1215 ConcurrentMarkSweepThread::start_icms();
1216 assert(top < _icms_stop_limit, "Tautology");
1217 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1218 return _icms_stop_limit;
1219 }
1220
1221 // The allocation will cross both the _start and _stop limits, so do the
1222 // stop notification also and return end().
1223 if (CMSTraceIncrementalMode) {
1224 space->print_on(gclog_or_tty);
1225 gclog_or_tty->stamp();
1226 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1227 ", new limit=" PTR_FORMAT
1228 " (" SIZE_FORMAT "%%)",
1229 top, space->end(),
1230 percent_of_space(space, space->end()));
1231 }
1232 ConcurrentMarkSweepThread::stop_icms();
1233 return space->end();
1234 }
1235
1236 if (top <= _icms_stop_limit) {
1237 if (CMSTraceIncrementalMode) {
1238 space->print_on(gclog_or_tty);
1239 gclog_or_tty->stamp();
1240 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1241 ", new limit=" PTR_FORMAT
1242 " (" SIZE_FORMAT "%%)",
1243 top, space->end(),
1244 percent_of_space(space, space->end()));
1245 }
1246 ConcurrentMarkSweepThread::stop_icms();
1247 return space->end();
1248 }
1249
1250 if (CMSTraceIncrementalMode) {
1251 space->print_on(gclog_or_tty);
1252 gclog_or_tty->stamp();
1253 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1254 ", new limit=" PTR_FORMAT,
1255 top, NULL);
1256 }
1257 }
1258
1259 return NULL;
1260 }
1261
1262 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1263 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1264 // allocate, copy and if necessary update promoinfo --
1265 // delegate to underlying space.
1266 assert_lock_strong(freelistLock());
1267
1268 #ifndef PRODUCT
1269 if (Universe::heap()->promotion_should_fail()) {
1270 return NULL;
1271 }
1272 #endif // #ifndef PRODUCT
1273
1274 oop res = _cmsSpace->promote(obj, obj_size);
1275 if (res == NULL) {
1276 // expand and retry
1277 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1278 expand(s*HeapWordSize, MinHeapDeltaBytes,
1279 CMSExpansionCause::_satisfy_promotion);
1280 // Since there's currently no next generation, we don't try to promote
1281 // into a more senior generation.
1282 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1283 "is made to pass on a possibly failing "
1284 "promotion to next generation");
1285 res = _cmsSpace->promote(obj, obj_size);
1286 }
1287 if (res != NULL) {
1288 // See comment in allocate() about when objects should
1289 // be allocated live.
1290 assert(obj->is_oop(), "Will dereference klass pointer below");
1291 collector()->promoted(false, // Not parallel
1292 (HeapWord*)res, obj->is_objArray(), obj_size);
1293 // promotion counters
1294 NOT_PRODUCT(
1295 _numObjectsPromoted++;
1296 _numWordsPromoted +=
1297 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1298 )
1299 }
1300 return res;
1301 }
1302
1303
1304 HeapWord*
1305 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1306 HeapWord* top,
1307 size_t word_sz)
1308 {
1309 return collector()->allocation_limit_reached(space, top, word_sz);
1310 }
1311
1312 // IMPORTANT: Notes on object size recognition in CMS.
1313 // ---------------------------------------------------
1314 // A block of storage in the CMS generation is always in
1315 // one of three states. A free block (FREE), an allocated
1316 // object (OBJECT) whose size() method reports the correct size,
1317 // and an intermediate state (TRANSIENT) in which its size cannot
1318 // be accurately determined.
1319 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1320 // -----------------------------------------------------
1321 // FREE: klass_word & 1 == 1; mark_word holds block size
1322 //
1323 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1324 // obj->size() computes correct size
1325 //
1326 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1327 //
1328 // STATE IDENTIFICATION: (64 bit+COOPS)
1329 // ------------------------------------
1330 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1331 //
1332 // OBJECT: klass_word installed; klass_word != 0;
1333 // obj->size() computes correct size
1334 //
1335 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1336 //
1337 //
1338 // STATE TRANSITION DIAGRAM
1339 //
1340 // mut / parnew mut / parnew
1341 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1342 // ^ |
1343 // |------------------------ DEAD <------------------------------------|
1344 // sweep mut
1345 //
1346 // While a block is in TRANSIENT state its size cannot be determined
1347 // so readers will either need to come back later or stall until
1348 // the size can be determined. Note that for the case of direct
1349 // allocation, P-bits, when available, may be used to determine the
1350 // size of an object that may not yet have been initialized.
1351
1352 // Things to support parallel young-gen collection.
1353 oop
1354 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1355 oop old, markOop m,
1356 size_t word_sz) {
1357 #ifndef PRODUCT
1358 if (Universe::heap()->promotion_should_fail()) {
1359 return NULL;
1360 }
1361 #endif // #ifndef PRODUCT
1362
1363 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1364 PromotionInfo* promoInfo = &ps->promo;
1365 // if we are tracking promotions, then first ensure space for
1366 // promotion (including spooling space for saving header if necessary).
1367 // then allocate and copy, then track promoted info if needed.
1368 // When tracking (see PromotionInfo::track()), the mark word may
1369 // be displaced and in this case restoration of the mark word
1370 // occurs in the (oop_since_save_marks_)iterate phase.
1371 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1372 // Out of space for allocating spooling buffers;
1373 // try expanding and allocating spooling buffers.
1374 if (!expand_and_ensure_spooling_space(promoInfo)) {
1375 return NULL;
1376 }
1377 }
1378 assert(promoInfo->has_spooling_space(), "Control point invariant");
1379 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1380 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1381 if (obj_ptr == NULL) {
1382 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1383 if (obj_ptr == NULL) {
1384 return NULL;
1385 }
1386 }
1387 oop obj = oop(obj_ptr);
1388 OrderAccess::storestore();
1389 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1390 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1391 // IMPORTANT: See note on object initialization for CMS above.
1392 // Otherwise, copy the object. Here we must be careful to insert the
1393 // klass pointer last, since this marks the block as an allocated object.
1394 // Except with compressed oops it's the mark word.
1395 HeapWord* old_ptr = (HeapWord*)old;
1396 // Restore the mark word copied above.
1397 obj->set_mark(m);
1398 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1399 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1400 OrderAccess::storestore();
1401
1402 if (UseCompressedClassPointers) {
1403 // Copy gap missed by (aligned) header size calculation below
1404 obj->set_klass_gap(old->klass_gap());
1405 }
1406 if (word_sz > (size_t)oopDesc::header_size()) {
1407 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1408 obj_ptr + oopDesc::header_size(),
1409 word_sz - oopDesc::header_size());
1410 }
1411
1412 // Now we can track the promoted object, if necessary. We take care
1413 // to delay the transition from uninitialized to full object
1414 // (i.e., insertion of klass pointer) until after, so that it
1415 // atomically becomes a promoted object.
1416 if (promoInfo->tracking()) {
1417 promoInfo->track((PromotedObject*)obj, old->klass());
1418 }
1419 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1420 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1421 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1422
1423 // Finally, install the klass pointer (this should be volatile).
1424 OrderAccess::storestore();
1425 obj->set_klass(old->klass());
1426 // We should now be able to calculate the right size for this object
1427 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1428
1429 collector()->promoted(true, // parallel
1430 obj_ptr, old->is_objArray(), word_sz);
1431
1432 NOT_PRODUCT(
1433 Atomic::inc_ptr(&_numObjectsPromoted);
1434 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1435 )
1436
1437 return obj;
1438 }
1439
1440 void
1441 ConcurrentMarkSweepGeneration::
1442 par_promote_alloc_undo(int thread_num,
1443 HeapWord* obj, size_t word_sz) {
1444 // CMS does not support promotion undo.
1445 ShouldNotReachHere();
1446 }
1447
1448 void
1449 ConcurrentMarkSweepGeneration::
1450 par_promote_alloc_done(int thread_num) {
1451 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1452 ps->lab.retire(thread_num);
1453 }
1454
1455 void
1456 ConcurrentMarkSweepGeneration::
1457 par_oop_since_save_marks_iterate_done(int thread_num) {
1458 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1459 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1460 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1461 }
1462
1463 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1464 size_t size,
1465 bool tlab)
1466 {
1467 // We allow a STW collection only if a full
1468 // collection was requested.
1469 return full || should_allocate(size, tlab); // FIX ME !!!
1470 // This and promotion failure handling are connected at the
1471 // hip and should be fixed by untying them.
1472 }
1473
1474 bool CMSCollector::shouldConcurrentCollect() {
1475 if (_full_gc_requested) {
1476 if (Verbose && PrintGCDetails) {
1477 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1478 " gc request (or gc_locker)");
1479 }
1480 return true;
1481 }
1482
1483 // For debugging purposes, change the type of collection.
1484 // If the rotation is not on the concurrent collection
1485 // type, don't start a concurrent collection.
1486 NOT_PRODUCT(
1487 if (RotateCMSCollectionTypes &&
1488 (_cmsGen->debug_collection_type() !=
1489 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1490 assert(_cmsGen->debug_collection_type() !=
1491 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1492 "Bad cms collection type");
1493 return false;
1494 }
1495 )
1496
1497 FreelistLocker x(this);
1498 // ------------------------------------------------------------------
1499 // Print out lots of information which affects the initiation of
1500 // a collection.
1501 if (PrintCMSInitiationStatistics && stats().valid()) {
1502 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1503 gclog_or_tty->stamp();
1504 gclog_or_tty->print_cr("");
1505 stats().print_on(gclog_or_tty);
1506 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1507 stats().time_until_cms_gen_full());
1508 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1509 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1510 _cmsGen->contiguous_available());
1511 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1512 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1513 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1514 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1515 gclog_or_tty->print_cr("metadata initialized %d",
1516 MetaspaceGC::should_concurrent_collect());
1517 }
1518 // ------------------------------------------------------------------
1519
1520 // If the estimated time to complete a cms collection (cms_duration())
1521 // is less than the estimated time remaining until the cms generation
1522 // is full, start a collection.
1523 if (!UseCMSInitiatingOccupancyOnly) {
1524 if (stats().valid()) {
1525 if (stats().time_until_cms_start() == 0.0) {
1526 return true;
1527 }
1528 } else {
1529 // We want to conservatively collect somewhat early in order
1530 // to try and "bootstrap" our CMS/promotion statistics;
1531 // this branch will not fire after the first successful CMS
1532 // collection because the stats should then be valid.
1533 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1534 if (Verbose && PrintGCDetails) {
1535 gclog_or_tty->print_cr(
1536 " CMSCollector: collect for bootstrapping statistics:"
1537 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1538 _bootstrap_occupancy);
1539 }
1540 return true;
1541 }
1542 }
1543 }
1544
1545 // Otherwise, we start a collection cycle if
1546 // old gen want a collection cycle started. Each may use
1547 // an appropriate criterion for making this decision.
1548 // XXX We need to make sure that the gen expansion
1549 // criterion dovetails well with this. XXX NEED TO FIX THIS
1550 if (_cmsGen->should_concurrent_collect()) {
1551 if (Verbose && PrintGCDetails) {
1552 gclog_or_tty->print_cr("CMS old gen initiated");
1553 }
1554 return true;
1555 }
1556
1557 // We start a collection if we believe an incremental collection may fail;
1558 // this is not likely to be productive in practice because it's probably too
1559 // late anyway.
1560 GenCollectedHeap* gch = GenCollectedHeap::heap();
1561 assert(gch->collector_policy()->is_two_generation_policy(),
1562 "You may want to check the correctness of the following");
1563 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1564 if (Verbose && PrintGCDetails) {
1565 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1566 }
1567 return true;
1568 }
1569
1570 if (MetaspaceGC::should_concurrent_collect()) {
1571 if (Verbose && PrintGCDetails) {
1572 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1573 }
1574 return true;
1575 }
1576
1577 return false;
1578 }
1579
1580 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1581
1582 // Clear _expansion_cause fields of constituent generations
1583 void CMSCollector::clear_expansion_cause() {
1584 _cmsGen->clear_expansion_cause();
1585 }
1586
1587 // We should be conservative in starting a collection cycle. To
1588 // start too eagerly runs the risk of collecting too often in the
1589 // extreme. To collect too rarely falls back on full collections,
1590 // which works, even if not optimum in terms of concurrent work.
1591 // As a work around for too eagerly collecting, use the flag
1592 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1593 // giving the user an easily understandable way of controlling the
1594 // collections.
1595 // We want to start a new collection cycle if any of the following
1596 // conditions hold:
1597 // . our current occupancy exceeds the configured initiating occupancy
1598 // for this generation, or
1599 // . we recently needed to expand this space and have not, since that
1600 // expansion, done a collection of this generation, or
1601 // . the underlying space believes that it may be a good idea to initiate
1602 // a concurrent collection (this may be based on criteria such as the
1603 // following: the space uses linear allocation and linear allocation is
1604 // going to fail, or there is believed to be excessive fragmentation in
1605 // the generation, etc... or ...
1606 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1607 // the case of the old generation; see CR 6543076):
1608 // we may be approaching a point at which allocation requests may fail because
1609 // we will be out of sufficient free space given allocation rate estimates.]
1610 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1611
1612 assert_lock_strong(freelistLock());
1613 if (occupancy() > initiating_occupancy()) {
1614 if (PrintGCDetails && Verbose) {
1615 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1616 short_name(), occupancy(), initiating_occupancy());
1617 }
1618 return true;
1619 }
1620 if (UseCMSInitiatingOccupancyOnly) {
1621 return false;
1622 }
1623 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1624 if (PrintGCDetails && Verbose) {
1625 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1626 short_name());
1627 }
1628 return true;
1629 }
1630 if (_cmsSpace->should_concurrent_collect()) {
1631 if (PrintGCDetails && Verbose) {
1632 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1633 short_name());
1634 }
1635 return true;
1636 }
1637 return false;
1638 }
1639
1640 void ConcurrentMarkSweepGeneration::collect(bool full,
1641 bool clear_all_soft_refs,
1642 size_t size,
1643 bool tlab)
1644 {
1645 collector()->collect(full, clear_all_soft_refs, size, tlab);
1646 }
1647
1648 void CMSCollector::collect(bool full,
1649 bool clear_all_soft_refs,
1650 size_t size,
1651 bool tlab)
1652 {
1653 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1654 // For debugging purposes skip the collection if the state
1655 // is not currently idle
1656 if (TraceCMSState) {
1657 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1658 Thread::current(), full, _collectorState);
1659 }
1660 return;
1661 }
1662
1663 // The following "if" branch is present for defensive reasons.
1664 // In the current uses of this interface, it can be replaced with:
1665 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1666 // But I am not placing that assert here to allow future
1667 // generality in invoking this interface.
1668 if (GC_locker::is_active()) {
1669 // A consistency test for GC_locker
1670 assert(GC_locker::needs_gc(), "Should have been set already");
1671 // Skip this foreground collection, instead
1672 // expanding the heap if necessary.
1673 // Need the free list locks for the call to free() in compute_new_size()
1674 compute_new_size();
1675 return;
1676 }
1677 acquire_control_and_collect(full, clear_all_soft_refs);
1678 _full_gcs_since_conc_gc++;
1679 }
1680
1681 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1682 GenCollectedHeap* gch = GenCollectedHeap::heap();
1683 unsigned int gc_count = gch->total_full_collections();
1684 if (gc_count == full_gc_count) {
1685 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1686 _full_gc_requested = true;
1687 _full_gc_cause = cause;
1688 CGC_lock->notify(); // nudge CMS thread
1689 } else {
1690 assert(gc_count > full_gc_count, "Error: causal loop");
1691 }
1692 }
1693
1694 bool CMSCollector::is_external_interruption() {
1695 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1696 return GCCause::is_user_requested_gc(cause) ||
1697 GCCause::is_serviceability_requested_gc(cause);
1698 }
1699
1700 void CMSCollector::report_concurrent_mode_interruption() {
1701 if (is_external_interruption()) {
1702 if (PrintGCDetails) {
1703 gclog_or_tty->print(" (concurrent mode interrupted)");
1704 }
1705 } else {
1706 if (PrintGCDetails) {
1707 gclog_or_tty->print(" (concurrent mode failure)");
1708 }
1709 _gc_tracer_cm->report_concurrent_mode_failure();
1710 }
1711 }
1712
1713
1714 // The foreground and background collectors need to coordinate in order
1715 // to make sure that they do not mutually interfere with CMS collections.
1716 // When a background collection is active,
1717 // the foreground collector may need to take over (preempt) and
1718 // synchronously complete an ongoing collection. Depending on the
1719 // frequency of the background collections and the heap usage
1720 // of the application, this preemption can be seldom or frequent.
1721 // There are only certain
1722 // points in the background collection that the "collection-baton"
1723 // can be passed to the foreground collector.
1724 //
1725 // The foreground collector will wait for the baton before
1726 // starting any part of the collection. The foreground collector
1727 // will only wait at one location.
1728 //
1729 // The background collector will yield the baton before starting a new
1730 // phase of the collection (e.g., before initial marking, marking from roots,
1731 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1732 // of the loop which switches the phases. The background collector does some
1733 // of the phases (initial mark, final re-mark) with the world stopped.
1734 // Because of locking involved in stopping the world,
1735 // the foreground collector should not block waiting for the background
1736 // collector when it is doing a stop-the-world phase. The background
1737 // collector will yield the baton at an additional point just before
1738 // it enters a stop-the-world phase. Once the world is stopped, the
1739 // background collector checks the phase of the collection. If the
1740 // phase has not changed, it proceeds with the collection. If the
1741 // phase has changed, it skips that phase of the collection. See
1742 // the comments on the use of the Heap_lock in collect_in_background().
1743 //
1744 // Variable used in baton passing.
1745 // _foregroundGCIsActive - Set to true by the foreground collector when
1746 // it wants the baton. The foreground clears it when it has finished
1747 // the collection.
1748 // _foregroundGCShouldWait - Set to true by the background collector
1749 // when it is running. The foreground collector waits while
1750 // _foregroundGCShouldWait is true.
1751 // CGC_lock - monitor used to protect access to the above variables
1752 // and to notify the foreground and background collectors.
1753 // _collectorState - current state of the CMS collection.
1754 //
1755 // The foreground collector
1756 // acquires the CGC_lock
1757 // sets _foregroundGCIsActive
1758 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1759 // various locks acquired in preparation for the collection
1760 // are released so as not to block the background collector
1761 // that is in the midst of a collection
1762 // proceeds with the collection
1763 // clears _foregroundGCIsActive
1764 // returns
1765 //
1766 // The background collector in a loop iterating on the phases of the
1767 // collection
1768 // acquires the CGC_lock
1769 // sets _foregroundGCShouldWait
1770 // if _foregroundGCIsActive is set
1771 // clears _foregroundGCShouldWait, notifies _CGC_lock
1772 // waits on _CGC_lock for _foregroundGCIsActive to become false
1773 // and exits the loop.
1774 // otherwise
1775 // proceed with that phase of the collection
1776 // if the phase is a stop-the-world phase,
1777 // yield the baton once more just before enqueueing
1778 // the stop-world CMS operation (executed by the VM thread).
1779 // returns after all phases of the collection are done
1780 //
1781
1782 void CMSCollector::acquire_control_and_collect(bool full,
1783 bool clear_all_soft_refs) {
1784 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1785 assert(!Thread::current()->is_ConcurrentGC_thread(),
1786 "shouldn't try to acquire control from self!");
1787
1788 // Start the protocol for acquiring control of the
1789 // collection from the background collector (aka CMS thread).
1790 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1791 "VM thread should have CMS token");
1792 // Remember the possibly interrupted state of an ongoing
1793 // concurrent collection
1794 CollectorState first_state = _collectorState;
1795
1796 // Signal to a possibly ongoing concurrent collection that
1797 // we want to do a foreground collection.
1798 _foregroundGCIsActive = true;
1799
1800 // Disable incremental mode during a foreground collection.
1801 ICMSDisabler icms_disabler;
1802
1803 // release locks and wait for a notify from the background collector
1804 // releasing the locks in only necessary for phases which
1805 // do yields to improve the granularity of the collection.
1806 assert_lock_strong(bitMapLock());
1807 // We need to lock the Free list lock for the space that we are
1808 // currently collecting.
1809 assert(haveFreelistLocks(), "Must be holding free list locks");
1810 bitMapLock()->unlock();
1811 releaseFreelistLocks();
1812 {
1813 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1814 if (_foregroundGCShouldWait) {
1815 // We are going to be waiting for action for the CMS thread;
1816 // it had better not be gone (for instance at shutdown)!
1817 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1818 "CMS thread must be running");
1819 // Wait here until the background collector gives us the go-ahead
1820 ConcurrentMarkSweepThread::clear_CMS_flag(
1821 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1822 // Get a possibly blocked CMS thread going:
1823 // Note that we set _foregroundGCIsActive true above,
1824 // without protection of the CGC_lock.
1825 CGC_lock->notify();
1826 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1827 "Possible deadlock");
1828 while (_foregroundGCShouldWait) {
1829 // wait for notification
1830 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1831 // Possibility of delay/starvation here, since CMS token does
1832 // not know to give priority to VM thread? Actually, i think
1833 // there wouldn't be any delay/starvation, but the proof of
1834 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1835 }
1836 ConcurrentMarkSweepThread::set_CMS_flag(
1837 ConcurrentMarkSweepThread::CMS_vm_has_token);
1838 }
1839 }
1840 // The CMS_token is already held. Get back the other locks.
1841 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1842 "VM thread should have CMS token");
1843 getFreelistLocks();
1844 bitMapLock()->lock_without_safepoint_check();
1845 if (TraceCMSState) {
1846 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1847 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1848 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1849 }
1850
1851 // Check if we need to do a compaction, or if not, whether
1852 // we need to start the mark-sweep from scratch.
1853 bool should_compact = false;
1854 bool should_start_over = false;
1855 decide_foreground_collection_type(clear_all_soft_refs,
1856 &should_compact, &should_start_over);
1857
1858 NOT_PRODUCT(
1859 if (RotateCMSCollectionTypes) {
1860 if (_cmsGen->debug_collection_type() ==
1861 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1862 should_compact = true;
1863 } else if (_cmsGen->debug_collection_type() ==
1864 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1865 should_compact = false;
1866 }
1867 }
1868 )
1869
1870 if (first_state > Idling) {
1871 report_concurrent_mode_interruption();
1872 }
1873
1874 set_did_compact(should_compact);
1875 if (should_compact) {
1876 // If the collection is being acquired from the background
1877 // collector, there may be references on the discovered
1878 // references lists that have NULL referents (being those
1879 // that were concurrently cleared by a mutator) or
1880 // that are no longer active (having been enqueued concurrently
1881 // by the mutator).
1882 // Scrub the list of those references because Mark-Sweep-Compact
1883 // code assumes referents are not NULL and that all discovered
1884 // Reference objects are active.
1885 ref_processor()->clean_up_discovered_references();
1886
1887 if (first_state > Idling) {
1888 save_heap_summary();
1889 }
1890
1891 do_compaction_work(clear_all_soft_refs);
1892
1893 // Has the GC time limit been exceeded?
1894 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1895 size_t max_eden_size = young_gen->max_capacity() -
1896 young_gen->to()->capacity() -
1897 young_gen->from()->capacity();
1898 GenCollectedHeap* gch = GenCollectedHeap::heap();
1899 GCCause::Cause gc_cause = gch->gc_cause();
1900 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1901 young_gen->eden()->used(),
1902 _cmsGen->max_capacity(),
1903 max_eden_size,
1904 full,
1905 gc_cause,
1906 gch->collector_policy());
1907 } else {
1908 do_mark_sweep_work(clear_all_soft_refs, first_state,
1909 should_start_over);
1910 }
1911 // Reset the expansion cause, now that we just completed
1912 // a collection cycle.
1913 clear_expansion_cause();
1914 _foregroundGCIsActive = false;
1915 return;
1916 }
1917
1918 // Resize the tenured generation
1919 // after obtaining the free list locks for the
1920 // two generations.
1921 void CMSCollector::compute_new_size() {
1922 assert_locked_or_safepoint(Heap_lock);
1923 FreelistLocker z(this);
1924 MetaspaceGC::compute_new_size();
1925 _cmsGen->compute_new_size_free_list();
1926 }
1927
1928 // A work method used by foreground collection to determine
1929 // what type of collection (compacting or not, continuing or fresh)
1930 // it should do.
1931 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1932 // and CMSCompactWhenClearAllSoftRefs the default in the future
1933 // and do away with the flags after a suitable period.
1934 void CMSCollector::decide_foreground_collection_type(
1935 bool clear_all_soft_refs, bool* should_compact,
1936 bool* should_start_over) {
1937 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1938 // flag is set, and we have either requested a System.gc() or
1939 // the number of full gc's since the last concurrent cycle
1940 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1941 // or if an incremental collection has failed
1942 GenCollectedHeap* gch = GenCollectedHeap::heap();
1943 assert(gch->collector_policy()->is_two_generation_policy(),
1944 "You may want to check the correctness of the following");
1945 // Inform cms gen if this was due to partial collection failing.
1946 // The CMS gen may use this fact to determine its expansion policy.
1947 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1948 assert(!_cmsGen->incremental_collection_failed(),
1949 "Should have been noticed, reacted to and cleared");
1950 _cmsGen->set_incremental_collection_failed();
1951 }
1952 *should_compact =
1953 UseCMSCompactAtFullCollection &&
1954 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1955 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1956 gch->incremental_collection_will_fail(true /* consult_young */));
1957 *should_start_over = false;
1958 if (clear_all_soft_refs && !*should_compact) {
1959 // We are about to do a last ditch collection attempt
1960 // so it would normally make sense to do a compaction
1961 // to reclaim as much space as possible.
1962 if (CMSCompactWhenClearAllSoftRefs) {
1963 // Default: The rationale is that in this case either
1964 // we are past the final marking phase, in which case
1965 // we'd have to start over, or so little has been done
1966 // that there's little point in saving that work. Compaction
1967 // appears to be the sensible choice in either case.
1968 *should_compact = true;
1969 } else {
1970 // We have been asked to clear all soft refs, but not to
1971 // compact. Make sure that we aren't past the final checkpoint
1972 // phase, for that is where we process soft refs. If we are already
1973 // past that phase, we'll need to redo the refs discovery phase and
1974 // if necessary clear soft refs that weren't previously
1975 // cleared. We do so by remembering the phase in which
1976 // we came in, and if we are past the refs processing
1977 // phase, we'll choose to just redo the mark-sweep
1978 // collection from scratch.
1979 if (_collectorState > FinalMarking) {
1980 // We are past the refs processing phase;
1981 // start over and do a fresh synchronous CMS cycle
1982 _collectorState = Resetting; // skip to reset to start new cycle
1983 reset(false /* == !asynch */);
1984 *should_start_over = true;
1985 } // else we can continue a possibly ongoing current cycle
1986 }
1987 }
1988 }
1989
1990 // A work method used by the foreground collector to do
1991 // a mark-sweep-compact.
1992 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1993 GenCollectedHeap* gch = GenCollectedHeap::heap();
1994
1995 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1996 gc_timer->register_gc_start();
1997
1998 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1999 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
2000
2001 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
2002 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
2003 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
2004 "collections passed to foreground collector", _full_gcs_since_conc_gc);
2005 }
2006
2007 // Sample collection interval time and reset for collection pause.
2008 if (UseAdaptiveSizePolicy) {
2009 size_policy()->msc_collection_begin();
2010 }
2011
2012 // Temporarily widen the span of the weak reference processing to
2013 // the entire heap.
2014 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2015 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2016 // Temporarily, clear the "is_alive_non_header" field of the
2017 // reference processor.
2018 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2019 // Temporarily make reference _processing_ single threaded (non-MT).
2020 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2021 // Temporarily make refs discovery atomic
2022 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2023 // Temporarily make reference _discovery_ single threaded (non-MT)
2024 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2025
2026 ref_processor()->set_enqueuing_is_done(false);
2027 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2028 ref_processor()->setup_policy(clear_all_soft_refs);
2029 // If an asynchronous collection finishes, the _modUnionTable is
2030 // all clear. If we are assuming the collection from an asynchronous
2031 // collection, clear the _modUnionTable.
2032 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2033 "_modUnionTable should be clear if the baton was not passed");
2034 _modUnionTable.clear_all();
2035 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
2036 "mod union for klasses should be clear if the baton was passed");
2037 _ct->klass_rem_set()->clear_mod_union();
2038
2039 // We must adjust the allocation statistics being maintained
2040 // in the free list space. We do so by reading and clearing
2041 // the sweep timer and updating the block flux rate estimates below.
2042 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2043 if (_inter_sweep_timer.is_active()) {
2044 _inter_sweep_timer.stop();
2045 // Note that we do not use this sample to update the _inter_sweep_estimate.
2046 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2047 _inter_sweep_estimate.padded_average(),
2048 _intra_sweep_estimate.padded_average());
2049 }
2050
2051 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2052 ref_processor(), clear_all_soft_refs);
2053 #ifdef ASSERT
2054 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2055 size_t free_size = cms_space->free();
2056 assert(free_size ==
2057 pointer_delta(cms_space->end(), cms_space->compaction_top())
2058 * HeapWordSize,
2059 "All the free space should be compacted into one chunk at top");
2060 assert(cms_space->dictionary()->total_chunk_size(
2061 debug_only(cms_space->freelistLock())) == 0 ||
2062 cms_space->totalSizeInIndexedFreeLists() == 0,
2063 "All the free space should be in a single chunk");
2064 size_t num = cms_space->totalCount();
2065 assert((free_size == 0 && num == 0) ||
2066 (free_size > 0 && (num == 1 || num == 2)),
2067 "There should be at most 2 free chunks after compaction");
2068 #endif // ASSERT
2069 _collectorState = Resetting;
2070 assert(_restart_addr == NULL,
2071 "Should have been NULL'd before baton was passed");
2072 reset(false /* == !asynch */);
2073 _cmsGen->reset_after_compaction();
2074 _concurrent_cycles_since_last_unload = 0;
2075
2076 // Clear any data recorded in the PLAB chunk arrays.
2077 if (_survivor_plab_array != NULL) {
2078 reset_survivor_plab_arrays();
2079 }
2080
2081 // Adjust the per-size allocation stats for the next epoch.
2082 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2083 // Restart the "inter sweep timer" for the next epoch.
2084 _inter_sweep_timer.reset();
2085 _inter_sweep_timer.start();
2086
2087 // Sample collection pause time and reset for collection interval.
2088 if (UseAdaptiveSizePolicy) {
2089 size_policy()->msc_collection_end(gch->gc_cause());
2090 }
2091
2092 gc_timer->register_gc_end();
2093
2094 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2095
2096 // For a mark-sweep-compact, compute_new_size() will be called
2097 // in the heap's do_collection() method.
2098 }
2099
2100 // A work method used by the foreground collector to do
2101 // a mark-sweep, after taking over from a possibly on-going
2102 // concurrent mark-sweep collection.
2103 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2104 CollectorState first_state, bool should_start_over) {
2105 if (PrintGC && Verbose) {
2106 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2107 "collector with count %d",
2108 _full_gcs_since_conc_gc);
2109 }
2110 switch (_collectorState) {
2111 case Idling:
2112 if (first_state == Idling || should_start_over) {
2113 // The background GC was not active, or should
2114 // restarted from scratch; start the cycle.
2115 _collectorState = InitialMarking;
2116 }
2117 // If first_state was not Idling, then a background GC
2118 // was in progress and has now finished. No need to do it
2119 // again. Leave the state as Idling.
2120 break;
2121 case Precleaning:
2122 // In the foreground case don't do the precleaning since
2123 // it is not done concurrently and there is extra work
2124 // required.
2125 _collectorState = FinalMarking;
2126 }
2127 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause());
2128
2129 // For a mark-sweep, compute_new_size() will be called
2130 // in the heap's do_collection() method.
2131 }
2132
2133
2134 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
2135 DefNewGeneration* dng = _young_gen->as_DefNewGeneration();
2136 EdenSpace* eden_space = dng->eden();
2137 ContiguousSpace* from_space = dng->from();
2138 ContiguousSpace* to_space = dng->to();
2139 // Eden
2140 if (_eden_chunk_array != NULL) {
2141 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2142 eden_space->bottom(), eden_space->top(),
2143 eden_space->end(), eden_space->capacity());
2144 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
2145 "_eden_chunk_capacity=" SIZE_FORMAT,
2146 _eden_chunk_index, _eden_chunk_capacity);
2147 for (size_t i = 0; i < _eden_chunk_index; i++) {
2148 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2149 i, _eden_chunk_array[i]);
2150 }
2151 }
2152 // Survivor
2153 if (_survivor_chunk_array != NULL) {
2154 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2155 from_space->bottom(), from_space->top(),
2156 from_space->end(), from_space->capacity());
2157 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
2158 "_survivor_chunk_capacity=" SIZE_FORMAT,
2159 _survivor_chunk_index, _survivor_chunk_capacity);
2160 for (size_t i = 0; i < _survivor_chunk_index; i++) {
2161 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2162 i, _survivor_chunk_array[i]);
2163 }
2164 }
2165 }
2166
2167 void CMSCollector::getFreelistLocks() const {
2168 // Get locks for all free lists in all generations that this
2169 // collector is responsible for
2170 _cmsGen->freelistLock()->lock_without_safepoint_check();
2171 }
2172
2173 void CMSCollector::releaseFreelistLocks() const {
2174 // Release locks for all free lists in all generations that this
2175 // collector is responsible for
2176 _cmsGen->freelistLock()->unlock();
2177 }
2178
2179 bool CMSCollector::haveFreelistLocks() const {
2180 // Check locks for all free lists in all generations that this
2181 // collector is responsible for
2182 assert_lock_strong(_cmsGen->freelistLock());
2183 PRODUCT_ONLY(ShouldNotReachHere());
2184 return true;
2185 }
2186
2187 // A utility class that is used by the CMS collector to
2188 // temporarily "release" the foreground collector from its
2189 // usual obligation to wait for the background collector to
2190 // complete an ongoing phase before proceeding.
2191 class ReleaseForegroundGC: public StackObj {
2192 private:
2193 CMSCollector* _c;
2194 public:
2195 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2196 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2197 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2198 // allow a potentially blocked foreground collector to proceed
2199 _c->_foregroundGCShouldWait = false;
2200 if (_c->_foregroundGCIsActive) {
2201 CGC_lock->notify();
2202 }
2203 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2204 "Possible deadlock");
2205 }
2206
2207 ~ReleaseForegroundGC() {
2208 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2209 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2210 _c->_foregroundGCShouldWait = true;
2211 }
2212 };
2213
2214 // There are separate collect_in_background and collect_in_foreground because of
2215 // the different locking requirements of the background collector and the
2216 // foreground collector. There was originally an attempt to share
2217 // one "collect" method between the background collector and the foreground
2218 // collector but the if-then-else required made it cleaner to have
2219 // separate methods.
2220 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) {
2221 assert(Thread::current()->is_ConcurrentGC_thread(),
2222 "A CMS asynchronous collection is only allowed on a CMS thread.");
2223
2224 GenCollectedHeap* gch = GenCollectedHeap::heap();
2225 {
2226 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2227 MutexLockerEx hl(Heap_lock, safepoint_check);
2228 FreelistLocker fll(this);
2229 MutexLockerEx x(CGC_lock, safepoint_check);
2230 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2231 // The foreground collector is active or we're
2232 // not using asynchronous collections. Skip this
2233 // background collection.
2234 assert(!_foregroundGCShouldWait, "Should be clear");
2235 return;
2236 } else {
2237 assert(_collectorState == Idling, "Should be idling before start.");
2238 _collectorState = InitialMarking;
2239 register_gc_start(cause);
2240 // Reset the expansion cause, now that we are about to begin
2241 // a new cycle.
2242 clear_expansion_cause();
2243
2244 // Clear the MetaspaceGC flag since a concurrent collection
2245 // is starting but also clear it after the collection.
2246 MetaspaceGC::set_should_concurrent_collect(false);
2247 }
2248 // Decide if we want to enable class unloading as part of the
2249 // ensuing concurrent GC cycle.
2250 update_should_unload_classes();
2251 _full_gc_requested = false; // acks all outstanding full gc requests
2252 _full_gc_cause = GCCause::_no_gc;
2253 // Signal that we are about to start a collection
2254 gch->increment_total_full_collections(); // ... starting a collection cycle
2255 _collection_count_start = gch->total_full_collections();
2256 }
2257
2258 // Used for PrintGC
2259 size_t prev_used;
2260 if (PrintGC && Verbose) {
2261 prev_used = _cmsGen->used(); // XXXPERM
2262 }
2263
2264 // The change of the collection state is normally done at this level;
2265 // the exceptions are phases that are executed while the world is
2266 // stopped. For those phases the change of state is done while the
2267 // world is stopped. For baton passing purposes this allows the
2268 // background collector to finish the phase and change state atomically.
2269 // The foreground collector cannot wait on a phase that is done
2270 // while the world is stopped because the foreground collector already
2271 // has the world stopped and would deadlock.
2272 while (_collectorState != Idling) {
2273 if (TraceCMSState) {
2274 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2275 Thread::current(), _collectorState);
2276 }
2277 // The foreground collector
2278 // holds the Heap_lock throughout its collection.
2279 // holds the CMS token (but not the lock)
2280 // except while it is waiting for the background collector to yield.
2281 //
2282 // The foreground collector should be blocked (not for long)
2283 // if the background collector is about to start a phase
2284 // executed with world stopped. If the background
2285 // collector has already started such a phase, the
2286 // foreground collector is blocked waiting for the
2287 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2288 // are executed in the VM thread.
2289 //
2290 // The locking order is
2291 // PendingListLock (PLL) -- if applicable (FinalMarking)
2292 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2293 // CMS token (claimed in
2294 // stop_world_and_do() -->
2295 // safepoint_synchronize() -->
2296 // CMSThread::synchronize())
2297
2298 {
2299 // Check if the FG collector wants us to yield.
2300 CMSTokenSync x(true); // is cms thread
2301 if (waitForForegroundGC()) {
2302 // We yielded to a foreground GC, nothing more to be
2303 // done this round.
2304 assert(_foregroundGCShouldWait == false, "We set it to false in "
2305 "waitForForegroundGC()");
2306 if (TraceCMSState) {
2307 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2308 " exiting collection CMS state %d",
2309 Thread::current(), _collectorState);
2310 }
2311 return;
2312 } else {
2313 // The background collector can run but check to see if the
2314 // foreground collector has done a collection while the
2315 // background collector was waiting to get the CGC_lock
2316 // above. If yes, break so that _foregroundGCShouldWait
2317 // is cleared before returning.
2318 if (_collectorState == Idling) {
2319 break;
2320 }
2321 }
2322 }
2323
2324 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2325 "should be waiting");
2326
2327 switch (_collectorState) {
2328 case InitialMarking:
2329 {
2330 ReleaseForegroundGC x(this);
2331 stats().record_cms_begin();
2332 VM_CMS_Initial_Mark initial_mark_op(this);
2333 VMThread::execute(&initial_mark_op);
2334 }
2335 // The collector state may be any legal state at this point
2336 // since the background collector may have yielded to the
2337 // foreground collector.
2338 break;
2339 case Marking:
2340 // initial marking in checkpointRootsInitialWork has been completed
2341 if (markFromRoots(true)) { // we were successful
2342 assert(_collectorState == Precleaning, "Collector state should "
2343 "have changed");
2344 } else {
2345 assert(_foregroundGCIsActive, "Internal state inconsistency");
2346 }
2347 break;
2348 case Precleaning:
2349 if (UseAdaptiveSizePolicy) {
2350 size_policy()->concurrent_precleaning_begin();
2351 }
2352 // marking from roots in markFromRoots has been completed
2353 preclean();
2354 if (UseAdaptiveSizePolicy) {
2355 size_policy()->concurrent_precleaning_end();
2356 }
2357 assert(_collectorState == AbortablePreclean ||
2358 _collectorState == FinalMarking,
2359 "Collector state should have changed");
2360 break;
2361 case AbortablePreclean:
2362 if (UseAdaptiveSizePolicy) {
2363 size_policy()->concurrent_phases_resume();
2364 }
2365 abortable_preclean();
2366 if (UseAdaptiveSizePolicy) {
2367 size_policy()->concurrent_precleaning_end();
2368 }
2369 assert(_collectorState == FinalMarking, "Collector state should "
2370 "have changed");
2371 break;
2372 case FinalMarking:
2373 {
2374 ReleaseForegroundGC x(this);
2375
2376 VM_CMS_Final_Remark final_remark_op(this);
2377 VMThread::execute(&final_remark_op);
2378 }
2379 assert(_foregroundGCShouldWait, "block post-condition");
2380 break;
2381 case Sweeping:
2382 if (UseAdaptiveSizePolicy) {
2383 size_policy()->concurrent_sweeping_begin();
2384 }
2385 // final marking in checkpointRootsFinal has been completed
2386 sweep(true);
2387 assert(_collectorState == Resizing, "Collector state change "
2388 "to Resizing must be done under the free_list_lock");
2389 _full_gcs_since_conc_gc = 0;
2390
2391 // Stop the timers for adaptive size policy for the concurrent phases
2392 if (UseAdaptiveSizePolicy) {
2393 size_policy()->concurrent_sweeping_end();
2394 size_policy()->concurrent_phases_end(gch->gc_cause(),
2395 gch->prev_gen(_cmsGen)->capacity(),
2396 _cmsGen->free());
2397 }
2398
2399 case Resizing: {
2400 // Sweeping has been completed...
2401 // At this point the background collection has completed.
2402 // Don't move the call to compute_new_size() down
2403 // into code that might be executed if the background
2404 // collection was preempted.
2405 {
2406 ReleaseForegroundGC x(this); // unblock FG collection
2407 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2408 CMSTokenSync z(true); // not strictly needed.
2409 if (_collectorState == Resizing) {
2410 compute_new_size();
2411 save_heap_summary();
2412 _collectorState = Resetting;
2413 } else {
2414 assert(_collectorState == Idling, "The state should only change"
2415 " because the foreground collector has finished the collection");
2416 }
2417 }
2418 break;
2419 }
2420 case Resetting:
2421 // CMS heap resizing has been completed
2422 reset(true);
2423 assert(_collectorState == Idling, "Collector state should "
2424 "have changed");
2425
2426 MetaspaceGC::set_should_concurrent_collect(false);
2427
2428 stats().record_cms_end();
2429 // Don't move the concurrent_phases_end() and compute_new_size()
2430 // calls to here because a preempted background collection
2431 // has it's state set to "Resetting".
2432 break;
2433 case Idling:
2434 default:
2435 ShouldNotReachHere();
2436 break;
2437 }
2438 if (TraceCMSState) {
2439 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2440 Thread::current(), _collectorState);
2441 }
2442 assert(_foregroundGCShouldWait, "block post-condition");
2443 }
2444
2445 // Should this be in gc_epilogue?
2446 collector_policy()->counters()->update_counters();
2447
2448 {
2449 // Clear _foregroundGCShouldWait and, in the event that the
2450 // foreground collector is waiting, notify it, before
2451 // returning.
2452 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2453 _foregroundGCShouldWait = false;
2454 if (_foregroundGCIsActive) {
2455 CGC_lock->notify();
2456 }
2457 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2458 "Possible deadlock");
2459 }
2460 if (TraceCMSState) {
2461 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2462 " exiting collection CMS state %d",
2463 Thread::current(), _collectorState);
2464 }
2465 if (PrintGC && Verbose) {
2466 _cmsGen->print_heap_change(prev_used);
2467 }
2468 }
2469
2470 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
2471 if (!_cms_start_registered) {
2472 register_gc_start(cause);
2473 }
2474 }
2475
2476 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2477 _cms_start_registered = true;
2478 _gc_timer_cm->register_gc_start();
2479 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2480 }
2481
2482 void CMSCollector::register_gc_end() {
2483 if (_cms_start_registered) {
2484 report_heap_summary(GCWhen::AfterGC);
2485
2486 _gc_timer_cm->register_gc_end();
2487 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2488 _cms_start_registered = false;
2489 }
2490 }
2491
2492 void CMSCollector::save_heap_summary() {
2493 GenCollectedHeap* gch = GenCollectedHeap::heap();
2494 _last_heap_summary = gch->create_heap_summary();
2495 _last_metaspace_summary = gch->create_metaspace_summary();
2496
2497 for (int i = 0; i < Metaspace::MetadataTypeCount; i++) {
2498 Metaspace::MetadataType mdtype = (Metaspace::MetadataType) i;
2499 if (MetaspaceAux::has_chunk_free_list(mdtype)) {
2500 _last_metaspace_free_list_summaries[i] = MetaspaceAux::chunk_free_list_summary(mdtype);
2501 }
2502 }
2503 }
2504
2505 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2506 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_metaspace_summary);
2507
2508 for (int i = 0; i < Metaspace::MetadataTypeCount; i++) {
2509 Metaspace::MetadataType mdtype = (Metaspace::MetadataType) i;
2510 if (MetaspaceAux::has_chunk_free_list(mdtype)) {
2511 _gc_tracer_cm->report_metaspace_chunk_free_list_summary(when, mdtype,
2512 _last_metaspace_free_list_summaries[i]);
2513 }
2514 }
2515 }
2516
2517 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) {
2518 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2519 "Foreground collector should be waiting, not executing");
2520 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2521 "may only be done by the VM Thread with the world stopped");
2522 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2523 "VM thread should have CMS token");
2524
2525 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2526 true, NULL);)
2527 if (UseAdaptiveSizePolicy) {
2528 size_policy()->ms_collection_begin();
2529 }
2530 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2531
2532 HandleMark hm; // Discard invalid handles created during verification
2533
2534 if (VerifyBeforeGC &&
2535 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2536 Universe::verify();
2537 }
2538
2539 // Snapshot the soft reference policy to be used in this collection cycle.
2540 ref_processor()->setup_policy(clear_all_soft_refs);
2541
2542 // Decide if class unloading should be done
2543 update_should_unload_classes();
2544
2545 bool init_mark_was_synchronous = false; // until proven otherwise
2546 while (_collectorState != Idling) {
2547 if (TraceCMSState) {
2548 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2549 Thread::current(), _collectorState);
2550 }
2551 switch (_collectorState) {
2552 case InitialMarking:
2553 register_foreground_gc_start(cause);
2554 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2555 checkpointRootsInitial(false);
2556 assert(_collectorState == Marking, "Collector state should have changed"
2557 " within checkpointRootsInitial()");
2558 break;
2559 case Marking:
2560 // initial marking in checkpointRootsInitialWork has been completed
2561 if (VerifyDuringGC &&
2562 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2563 Universe::verify("Verify before initial mark: ");
2564 }
2565 {
2566 bool res = markFromRoots(false);
2567 assert(res && _collectorState == FinalMarking, "Collector state should "
2568 "have changed");
2569 break;
2570 }
2571 case FinalMarking:
2572 if (VerifyDuringGC &&
2573 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2574 Universe::verify("Verify before re-mark: ");
2575 }
2576 checkpointRootsFinal(false, clear_all_soft_refs,
2577 init_mark_was_synchronous);
2578 assert(_collectorState == Sweeping, "Collector state should not "
2579 "have changed within checkpointRootsFinal()");
2580 break;
2581 case Sweeping:
2582 // final marking in checkpointRootsFinal has been completed
2583 if (VerifyDuringGC &&
2584 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2585 Universe::verify("Verify before sweep: ");
2586 }
2587 sweep(false);
2588 assert(_collectorState == Resizing, "Incorrect state");
2589 break;
2590 case Resizing: {
2591 // Sweeping has been completed; the actual resize in this case
2592 // is done separately; nothing to be done in this state.
2593 _collectorState = Resetting;
2594 break;
2595 }
2596 case Resetting:
2597 // The heap has been resized.
2598 if (VerifyDuringGC &&
2599 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2600 Universe::verify("Verify before reset: ");
2601 }
2602 save_heap_summary();
2603 reset(false);
2604 assert(_collectorState == Idling, "Collector state should "
2605 "have changed");
2606 break;
2607 case Precleaning:
2608 case AbortablePreclean:
2609 // Elide the preclean phase
2610 _collectorState = FinalMarking;
2611 break;
2612 default:
2613 ShouldNotReachHere();
2614 }
2615 if (TraceCMSState) {
2616 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2617 Thread::current(), _collectorState);
2618 }
2619 }
2620
2621 if (UseAdaptiveSizePolicy) {
2622 GenCollectedHeap* gch = GenCollectedHeap::heap();
2623 size_policy()->ms_collection_end(gch->gc_cause());
2624 }
2625
2626 if (VerifyAfterGC &&
2627 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2628 Universe::verify();
2629 }
2630 if (TraceCMSState) {
2631 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2632 " exiting collection CMS state %d",
2633 Thread::current(), _collectorState);
2634 }
2635 }
2636
2637 bool CMSCollector::waitForForegroundGC() {
2638 bool res = false;
2639 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2640 "CMS thread should have CMS token");
2641 // Block the foreground collector until the
2642 // background collectors decides whether to
2643 // yield.
2644 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2645 _foregroundGCShouldWait = true;
2646 if (_foregroundGCIsActive) {
2647 // The background collector yields to the
2648 // foreground collector and returns a value
2649 // indicating that it has yielded. The foreground
2650 // collector can proceed.
2651 res = true;
2652 _foregroundGCShouldWait = false;
2653 ConcurrentMarkSweepThread::clear_CMS_flag(
2654 ConcurrentMarkSweepThread::CMS_cms_has_token);
2655 ConcurrentMarkSweepThread::set_CMS_flag(
2656 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2657 // Get a possibly blocked foreground thread going
2658 CGC_lock->notify();
2659 if (TraceCMSState) {
2660 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2661 Thread::current(), _collectorState);
2662 }
2663 while (_foregroundGCIsActive) {
2664 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2665 }
2666 ConcurrentMarkSweepThread::set_CMS_flag(
2667 ConcurrentMarkSweepThread::CMS_cms_has_token);
2668 ConcurrentMarkSweepThread::clear_CMS_flag(
2669 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2670 }
2671 if (TraceCMSState) {
2672 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2673 Thread::current(), _collectorState);
2674 }
2675 return res;
2676 }
2677
2678 // Because of the need to lock the free lists and other structures in
2679 // the collector, common to all the generations that the collector is
2680 // collecting, we need the gc_prologues of individual CMS generations
2681 // delegate to their collector. It may have been simpler had the
2682 // current infrastructure allowed one to call a prologue on a
2683 // collector. In the absence of that we have the generation's
2684 // prologue delegate to the collector, which delegates back
2685 // some "local" work to a worker method in the individual generations
2686 // that it's responsible for collecting, while itself doing any
2687 // work common to all generations it's responsible for. A similar
2688 // comment applies to the gc_epilogue()'s.
2689 // The role of the variable _between_prologue_and_epilogue is to
2690 // enforce the invocation protocol.
2691 void CMSCollector::gc_prologue(bool full) {
2692 // Call gc_prologue_work() for the CMSGen
2693 // we are responsible for.
2694
2695 // The following locking discipline assumes that we are only called
2696 // when the world is stopped.
2697 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2698
2699 // The CMSCollector prologue must call the gc_prologues for the
2700 // "generations" that it's responsible
2701 // for.
2702
2703 assert( Thread::current()->is_VM_thread()
2704 || ( CMSScavengeBeforeRemark
2705 && Thread::current()->is_ConcurrentGC_thread()),
2706 "Incorrect thread type for prologue execution");
2707
2708 if (_between_prologue_and_epilogue) {
2709 // We have already been invoked; this is a gc_prologue delegation
2710 // from yet another CMS generation that we are responsible for, just
2711 // ignore it since all relevant work has already been done.
2712 return;
2713 }
2714
2715 // set a bit saying prologue has been called; cleared in epilogue
2716 _between_prologue_and_epilogue = true;
2717 // Claim locks for common data structures, then call gc_prologue_work()
2718 // for each CMSGen.
2719
2720 getFreelistLocks(); // gets free list locks on constituent spaces
2721 bitMapLock()->lock_without_safepoint_check();
2722
2723 // Should call gc_prologue_work() for all cms gens we are responsible for
2724 bool duringMarking = _collectorState >= Marking
2725 && _collectorState < Sweeping;
2726
2727 // The young collections clear the modified oops state, which tells if
2728 // there are any modified oops in the class. The remark phase also needs
2729 // that information. Tell the young collection to save the union of all
2730 // modified klasses.
2731 if (duringMarking) {
2732 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2733 }
2734
2735 bool registerClosure = duringMarking;
2736
2737 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2738 &_modUnionClosurePar
2739 : &_modUnionClosure;
2740 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2741
2742 if (!full) {
2743 stats().record_gc0_begin();
2744 }
2745 }
2746
2747 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2748
2749 _capacity_at_prologue = capacity();
2750 _used_at_prologue = used();
2751
2752 // Delegate to CMScollector which knows how to coordinate between
2753 // this and any other CMS generations that it is responsible for
2754 // collecting.
2755 collector()->gc_prologue(full);
2756 }
2757
2758 // This is a "private" interface for use by this generation's CMSCollector.
2759 // Not to be called directly by any other entity (for instance,
2760 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2761 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2762 bool registerClosure, ModUnionClosure* modUnionClosure) {
2763 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2764 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2765 "Should be NULL");
2766 if (registerClosure) {
2767 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2768 }
2769 cmsSpace()->gc_prologue();
2770 // Clear stat counters
2771 NOT_PRODUCT(
2772 assert(_numObjectsPromoted == 0, "check");
2773 assert(_numWordsPromoted == 0, "check");
2774 if (Verbose && PrintGC) {
2775 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2776 SIZE_FORMAT" bytes concurrently",
2777 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2778 }
2779 _numObjectsAllocated = 0;
2780 _numWordsAllocated = 0;
2781 )
2782 }
2783
2784 void CMSCollector::gc_epilogue(bool full) {
2785 // The following locking discipline assumes that we are only called
2786 // when the world is stopped.
2787 assert(SafepointSynchronize::is_at_safepoint(),
2788 "world is stopped assumption");
2789
2790 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2791 // if linear allocation blocks need to be appropriately marked to allow the
2792 // the blocks to be parsable. We also check here whether we need to nudge the
2793 // CMS collector thread to start a new cycle (if it's not already active).
2794 assert( Thread::current()->is_VM_thread()
2795 || ( CMSScavengeBeforeRemark
2796 && Thread::current()->is_ConcurrentGC_thread()),
2797 "Incorrect thread type for epilogue execution");
2798
2799 if (!_between_prologue_and_epilogue) {
2800 // We have already been invoked; this is a gc_epilogue delegation
2801 // from yet another CMS generation that we are responsible for, just
2802 // ignore it since all relevant work has already been done.
2803 return;
2804 }
2805 assert(haveFreelistLocks(), "must have freelist locks");
2806 assert_lock_strong(bitMapLock());
2807
2808 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2809
2810 _cmsGen->gc_epilogue_work(full);
2811
2812 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2813 // in case sampling was not already enabled, enable it
2814 _start_sampling = true;
2815 }
2816 // reset _eden_chunk_array so sampling starts afresh
2817 _eden_chunk_index = 0;
2818
2819 size_t cms_used = _cmsGen->cmsSpace()->used();
2820
2821 // update performance counters - this uses a special version of
2822 // update_counters() that allows the utilization to be passed as a
2823 // parameter, avoiding multiple calls to used().
2824 //
2825 _cmsGen->update_counters(cms_used);
2826
2827 if (CMSIncrementalMode) {
2828 icms_update_allocation_limits();
2829 }
2830
2831 bitMapLock()->unlock();
2832 releaseFreelistLocks();
2833
2834 if (!CleanChunkPoolAsync) {
2835 Chunk::clean_chunk_pool();
2836 }
2837
2838 set_did_compact(false);
2839 _between_prologue_and_epilogue = false; // ready for next cycle
2840 }
2841
2842 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2843 collector()->gc_epilogue(full);
2844
2845 // Also reset promotion tracking in par gc thread states.
2846 if (CollectedHeap::use_parallel_gc_threads()) {
2847 for (uint i = 0; i < ParallelGCThreads; i++) {
2848 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2849 }
2850 }
2851 }
2852
2853 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2854 assert(!incremental_collection_failed(), "Should have been cleared");
2855 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2856 cmsSpace()->gc_epilogue();
2857 // Print stat counters
2858 NOT_PRODUCT(
2859 assert(_numObjectsAllocated == 0, "check");
2860 assert(_numWordsAllocated == 0, "check");
2861 if (Verbose && PrintGC) {
2862 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2863 SIZE_FORMAT" bytes",
2864 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2865 }
2866 _numObjectsPromoted = 0;
2867 _numWordsPromoted = 0;
2868 )
2869
2870 if (PrintGC && Verbose) {
2871 // Call down the chain in contiguous_available needs the freelistLock
2872 // so print this out before releasing the freeListLock.
2873 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2874 contiguous_available());
2875 }
2876 }
2877
2878 #ifndef PRODUCT
2879 bool CMSCollector::have_cms_token() {
2880 Thread* thr = Thread::current();
2881 if (thr->is_VM_thread()) {
2882 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2883 } else if (thr->is_ConcurrentGC_thread()) {
2884 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2885 } else if (thr->is_GC_task_thread()) {
2886 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2887 ParGCRareEvent_lock->owned_by_self();
2888 }
2889 return false;
2890 }
2891 #endif
2892
2893 // Check reachability of the given heap address in CMS generation,
2894 // treating all other generations as roots.
2895 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2896 // We could "guarantee" below, rather than assert, but I'll
2897 // leave these as "asserts" so that an adventurous debugger
2898 // could try this in the product build provided some subset of
2899 // the conditions were met, provided they were interested in the
2900 // results and knew that the computation below wouldn't interfere
2901 // with other concurrent computations mutating the structures
2902 // being read or written.
2903 assert(SafepointSynchronize::is_at_safepoint(),
2904 "Else mutations in object graph will make answer suspect");
2905 assert(have_cms_token(), "Should hold cms token");
2906 assert(haveFreelistLocks(), "must hold free list locks");
2907 assert_lock_strong(bitMapLock());
2908
2909 // Clear the marking bit map array before starting, but, just
2910 // for kicks, first report if the given address is already marked
2911 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2912 _markBitMap.isMarked(addr) ? "" : " not");
2913
2914 if (verify_after_remark()) {
2915 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2916 bool result = verification_mark_bm()->isMarked(addr);
2917 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2918 result ? "IS" : "is NOT");
2919 return result;
2920 } else {
2921 gclog_or_tty->print_cr("Could not compute result");
2922 return false;
2923 }
2924 }
2925
2926
2927 void
2928 CMSCollector::print_on_error(outputStream* st) {
2929 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2930 if (collector != NULL) {
2931 CMSBitMap* bitmap = &collector->_markBitMap;
2932 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap);
2933 bitmap->print_on_error(st, " Bits: ");
2934
2935 st->cr();
2936
2937 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2938 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap);
2939 mut_bitmap->print_on_error(st, " Bits: ");
2940 }
2941 }
2942
2943 ////////////////////////////////////////////////////////
2944 // CMS Verification Support
2945 ////////////////////////////////////////////////////////
2946 // Following the remark phase, the following invariant
2947 // should hold -- each object in the CMS heap which is
2948 // marked in markBitMap() should be marked in the verification_mark_bm().
2949
2950 class VerifyMarkedClosure: public BitMapClosure {
2951 CMSBitMap* _marks;
2952 bool _failed;
2953
2954 public:
2955 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2956
2957 bool do_bit(size_t offset) {
2958 HeapWord* addr = _marks->offsetToHeapWord(offset);
2959 if (!_marks->isMarked(addr)) {
2960 oop(addr)->print_on(gclog_or_tty);
2961 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2962 _failed = true;
2963 }
2964 return true;
2965 }
2966
2967 bool failed() { return _failed; }
2968 };
2969
2970 bool CMSCollector::verify_after_remark(bool silent) {
2971 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2972 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2973 static bool init = false;
2974
2975 assert(SafepointSynchronize::is_at_safepoint(),
2976 "Else mutations in object graph will make answer suspect");
2977 assert(have_cms_token(),
2978 "Else there may be mutual interference in use of "
2979 " verification data structures");
2980 assert(_collectorState > Marking && _collectorState <= Sweeping,
2981 "Else marking info checked here may be obsolete");
2982 assert(haveFreelistLocks(), "must hold free list locks");
2983 assert_lock_strong(bitMapLock());
2984
2985
2986 // Allocate marking bit map if not already allocated
2987 if (!init) { // first time
2988 if (!verification_mark_bm()->allocate(_span)) {
2989 return false;
2990 }
2991 init = true;
2992 }
2993
2994 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2995
2996 // Turn off refs discovery -- so we will be tracing through refs.
2997 // This is as intended, because by this time
2998 // GC must already have cleared any refs that need to be cleared,
2999 // and traced those that need to be marked; moreover,
3000 // the marking done here is not going to interfere in any
3001 // way with the marking information used by GC.
3002 NoRefDiscovery no_discovery(ref_processor());
3003
3004 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3005
3006 // Clear any marks from a previous round
3007 verification_mark_bm()->clear_all();
3008 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
3009 verify_work_stacks_empty();
3010
3011 GenCollectedHeap* gch = GenCollectedHeap::heap();
3012 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3013 // Update the saved marks which may affect the root scans.
3014 gch->save_marks();
3015
3016 if (CMSRemarkVerifyVariant == 1) {
3017 // In this first variant of verification, we complete
3018 // all marking, then check if the new marks-vector is
3019 // a subset of the CMS marks-vector.
3020 verify_after_remark_work_1();
3021 } else if (CMSRemarkVerifyVariant == 2) {
3022 // In this second variant of verification, we flag an error
3023 // (i.e. an object reachable in the new marks-vector not reachable
3024 // in the CMS marks-vector) immediately, also indicating the
3025 // identify of an object (A) that references the unmarked object (B) --
3026 // presumably, a mutation to A failed to be picked up by preclean/remark?
3027 verify_after_remark_work_2();
3028 } else {
3029 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
3030 CMSRemarkVerifyVariant);
3031 }
3032 if (!silent) gclog_or_tty->print(" done] ");
3033 return true;
3034 }
3035
3036 void CMSCollector::verify_after_remark_work_1() {
3037 ResourceMark rm;
3038 HandleMark hm;
3039 GenCollectedHeap* gch = GenCollectedHeap::heap();
3040
3041 // Get a clear set of claim bits for the strong roots processing to work with.
3042 ClassLoaderDataGraph::clear_claimed_marks();
3043
3044 // Mark from roots one level into CMS
3045 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
3046 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3047
3048 gch->gen_process_strong_roots(_cmsGen->level(),
3049 true, // younger gens are roots
3050 true, // activate StrongRootsScope
3051 SharedHeap::ScanningOption(roots_scanning_options()),
3052 ¬Older,
3053 NULL,
3054 NULL); // SSS: Provide correct closure
3055
3056 // Now mark from the roots
3057 MarkFromRootsClosure markFromRootsClosure(this, _span,
3058 verification_mark_bm(), verification_mark_stack(),
3059 false /* don't yield */, true /* verifying */);
3060 assert(_restart_addr == NULL, "Expected pre-condition");
3061 verification_mark_bm()->iterate(&markFromRootsClosure);
3062 while (_restart_addr != NULL) {
3063 // Deal with stack overflow: by restarting at the indicated
3064 // address.
3065 HeapWord* ra = _restart_addr;
3066 markFromRootsClosure.reset(ra);
3067 _restart_addr = NULL;
3068 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3069 }
3070 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3071 verify_work_stacks_empty();
3072
3073 // Marking completed -- now verify that each bit marked in
3074 // verification_mark_bm() is also marked in markBitMap(); flag all
3075 // errors by printing corresponding objects.
3076 VerifyMarkedClosure vcl(markBitMap());
3077 verification_mark_bm()->iterate(&vcl);
3078 if (vcl.failed()) {
3079 gclog_or_tty->print("Verification failed");
3080 Universe::heap()->print_on(gclog_or_tty);
3081 fatal("CMS: failed marking verification after remark");
3082 }
3083 }
3084
3085 class VerifyKlassOopsKlassClosure : public KlassClosure {
3086 class VerifyKlassOopsClosure : public OopClosure {
3087 CMSBitMap* _bitmap;
3088 public:
3089 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
3090 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
3091 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3092 } _oop_closure;
3093 public:
3094 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
3095 void do_klass(Klass* k) {
3096 k->oops_do(&_oop_closure);
3097 }
3098 };
3099
3100 void CMSCollector::verify_after_remark_work_2() {
3101 ResourceMark rm;
3102 HandleMark hm;
3103 GenCollectedHeap* gch = GenCollectedHeap::heap();
3104
3105 // Get a clear set of claim bits for the strong roots processing to work with.
3106 ClassLoaderDataGraph::clear_claimed_marks();
3107
3108 // Mark from roots one level into CMS
3109 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
3110 markBitMap());
3111 CMKlassClosure klass_closure(¬Older);
3112
3113 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3114 gch->gen_process_strong_roots(_cmsGen->level(),
3115 true, // younger gens are roots
3116 true, // activate StrongRootsScope
3117 SharedHeap::ScanningOption(roots_scanning_options()),
3118 ¬Older,
3119 NULL,
3120 &klass_closure);
3121
3122 // Now mark from the roots
3123 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
3124 verification_mark_bm(), markBitMap(), verification_mark_stack());
3125 assert(_restart_addr == NULL, "Expected pre-condition");
3126 verification_mark_bm()->iterate(&markFromRootsClosure);
3127 while (_restart_addr != NULL) {
3128 // Deal with stack overflow: by restarting at the indicated
3129 // address.
3130 HeapWord* ra = _restart_addr;
3131 markFromRootsClosure.reset(ra);
3132 _restart_addr = NULL;
3133 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3134 }
3135 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3136 verify_work_stacks_empty();
3137
3138 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
3139 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
3140
3141 // Marking completed -- now verify that each bit marked in
3142 // verification_mark_bm() is also marked in markBitMap(); flag all
3143 // errors by printing corresponding objects.
3144 VerifyMarkedClosure vcl(markBitMap());
3145 verification_mark_bm()->iterate(&vcl);
3146 assert(!vcl.failed(), "Else verification above should not have succeeded");
3147 }
3148
3149 void ConcurrentMarkSweepGeneration::save_marks() {
3150 // delegate to CMS space
3151 cmsSpace()->save_marks();
3152 for (uint i = 0; i < ParallelGCThreads; i++) {
3153 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3154 }
3155 }
3156
3157 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3158 return cmsSpace()->no_allocs_since_save_marks();
3159 }
3160
3161 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3162 \
3163 void ConcurrentMarkSweepGeneration:: \
3164 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3165 cl->set_generation(this); \
3166 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3167 cl->reset_generation(); \
3168 save_marks(); \
3169 }
3170
3171 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3172
3173 void
3174 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3175 cl->set_generation(this);
3176 younger_refs_in_space_iterate(_cmsSpace, cl);
3177 cl->reset_generation();
3178 }
3179
3180 void
3181 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
3182 if (freelistLock()->owned_by_self()) {
3183 Generation::oop_iterate(mr, cl);
3184 } else {
3185 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3186 Generation::oop_iterate(mr, cl);
3187 }
3188 }
3189
3190 void
3191 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
3192 if (freelistLock()->owned_by_self()) {
3193 Generation::oop_iterate(cl);
3194 } else {
3195 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3196 Generation::oop_iterate(cl);
3197 }
3198 }
3199
3200 void
3201 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3202 if (freelistLock()->owned_by_self()) {
3203 Generation::object_iterate(cl);
3204 } else {
3205 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3206 Generation::object_iterate(cl);
3207 }
3208 }
3209
3210 void
3211 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3212 if (freelistLock()->owned_by_self()) {
3213 Generation::safe_object_iterate(cl);
3214 } else {
3215 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3216 Generation::safe_object_iterate(cl);
3217 }
3218 }
3219
3220 void
3221 ConcurrentMarkSweepGeneration::post_compact() {
3222 }
3223
3224 void
3225 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3226 // Fix the linear allocation blocks to look like free blocks.
3227
3228 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3229 // are not called when the heap is verified during universe initialization and
3230 // at vm shutdown.
3231 if (freelistLock()->owned_by_self()) {
3232 cmsSpace()->prepare_for_verify();
3233 } else {
3234 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3235 cmsSpace()->prepare_for_verify();
3236 }
3237 }
3238
3239 void
3240 ConcurrentMarkSweepGeneration::verify() {
3241 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3242 // are not called when the heap is verified during universe initialization and
3243 // at vm shutdown.
3244 if (freelistLock()->owned_by_self()) {
3245 cmsSpace()->verify();
3246 } else {
3247 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3248 cmsSpace()->verify();
3249 }
3250 }
3251
3252 void CMSCollector::verify() {
3253 _cmsGen->verify();
3254 }
3255
3256 #ifndef PRODUCT
3257 bool CMSCollector::overflow_list_is_empty() const {
3258 assert(_num_par_pushes >= 0, "Inconsistency");
3259 if (_overflow_list == NULL) {
3260 assert(_num_par_pushes == 0, "Inconsistency");
3261 }
3262 return _overflow_list == NULL;
3263 }
3264
3265 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3266 // merely consolidate assertion checks that appear to occur together frequently.
3267 void CMSCollector::verify_work_stacks_empty() const {
3268 assert(_markStack.isEmpty(), "Marking stack should be empty");
3269 assert(overflow_list_is_empty(), "Overflow list should be empty");
3270 }
3271
3272 void CMSCollector::verify_overflow_empty() const {
3273 assert(overflow_list_is_empty(), "Overflow list should be empty");
3274 assert(no_preserved_marks(), "No preserved marks");
3275 }
3276 #endif // PRODUCT
3277
3278 // Decide if we want to enable class unloading as part of the
3279 // ensuing concurrent GC cycle. We will collect and
3280 // unload classes if it's the case that:
3281 // (1) an explicit gc request has been made and the flag
3282 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3283 // (2) (a) class unloading is enabled at the command line, and
3284 // (b) old gen is getting really full
3285 // NOTE: Provided there is no change in the state of the heap between
3286 // calls to this method, it should have idempotent results. Moreover,
3287 // its results should be monotonically increasing (i.e. going from 0 to 1,
3288 // but not 1 to 0) between successive calls between which the heap was
3289 // not collected. For the implementation below, it must thus rely on
3290 // the property that concurrent_cycles_since_last_unload()
3291 // will not decrease unless a collection cycle happened and that
3292 // _cmsGen->is_too_full() are
3293 // themselves also monotonic in that sense. See check_monotonicity()
3294 // below.
3295 void CMSCollector::update_should_unload_classes() {
3296 _should_unload_classes = false;
3297 // Condition 1 above
3298 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3299 _should_unload_classes = true;
3300 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3301 // Disjuncts 2.b.(i,ii,iii) above
3302 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3303 CMSClassUnloadingMaxInterval)
3304 || _cmsGen->is_too_full();
3305 }
3306 }
3307
3308 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3309 bool res = should_concurrent_collect();
3310 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3311 return res;
3312 }
3313
3314 void CMSCollector::setup_cms_unloading_and_verification_state() {
3315 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3316 || VerifyBeforeExit;
3317 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_AllCodeCache;
3318
3319 // We set the proper root for this CMS cycle here.
3320 if (should_unload_classes()) { // Should unload classes this cycle
3321 remove_root_scanning_option(SharedHeap::SO_AllClasses);
3322 add_root_scanning_option(SharedHeap::SO_SystemClasses);
3323 remove_root_scanning_option(rso); // Shrink the root set appropriately
3324 set_verifying(should_verify); // Set verification state for this cycle
3325 return; // Nothing else needs to be done at this time
3326 }
3327
3328 // Not unloading classes this cycle
3329 assert(!should_unload_classes(), "Inconsistency!");
3330 remove_root_scanning_option(SharedHeap::SO_SystemClasses);
3331 add_root_scanning_option(SharedHeap::SO_AllClasses);
3332
3333 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3334 // Include symbols, strings and code cache elements to prevent their resurrection.
3335 add_root_scanning_option(rso);
3336 set_verifying(true);
3337 } else if (verifying() && !should_verify) {
3338 // We were verifying, but some verification flags got disabled.
3339 set_verifying(false);
3340 // Exclude symbols, strings and code cache elements from root scanning to
3341 // reduce IM and RM pauses.
3342 remove_root_scanning_option(rso);
3343 }
3344 }
3345
3346
3347 #ifndef PRODUCT
3348 HeapWord* CMSCollector::block_start(const void* p) const {
3349 const HeapWord* addr = (HeapWord*)p;
3350 if (_span.contains(p)) {
3351 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3352 return _cmsGen->cmsSpace()->block_start(p);
3353 }
3354 }
3355 return NULL;
3356 }
3357 #endif
3358
3359 HeapWord*
3360 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3361 bool tlab,
3362 bool parallel) {
3363 CMSSynchronousYieldRequest yr;
3364 assert(!tlab, "Can't deal with TLAB allocation");
3365 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3366 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3367 CMSExpansionCause::_satisfy_allocation);
3368 if (GCExpandToAllocateDelayMillis > 0) {
3369 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3370 }
3371 return have_lock_and_allocate(word_size, tlab);
3372 }
3373
3374 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3375 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3376 // to CardGeneration and share it...
3377 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3378 return CardGeneration::expand(bytes, expand_bytes);
3379 }
3380
3381 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3382 CMSExpansionCause::Cause cause)
3383 {
3384
3385 bool success = expand(bytes, expand_bytes);
3386
3387 // remember why we expanded; this information is used
3388 // by shouldConcurrentCollect() when making decisions on whether to start
3389 // a new CMS cycle.
3390 if (success) {
3391 set_expansion_cause(cause);
3392 if (PrintGCDetails && Verbose) {
3393 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3394 CMSExpansionCause::to_string(cause));
3395 }
3396 }
3397 }
3398
3399 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3400 HeapWord* res = NULL;
3401 MutexLocker x(ParGCRareEvent_lock);
3402 while (true) {
3403 // Expansion by some other thread might make alloc OK now:
3404 res = ps->lab.alloc(word_sz);
3405 if (res != NULL) return res;
3406 // If there's not enough expansion space available, give up.
3407 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3408 return NULL;
3409 }
3410 // Otherwise, we try expansion.
3411 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3412 CMSExpansionCause::_allocate_par_lab);
3413 // Now go around the loop and try alloc again;
3414 // A competing par_promote might beat us to the expansion space,
3415 // so we may go around the loop again if promotion fails again.
3416 if (GCExpandToAllocateDelayMillis > 0) {
3417 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3418 }
3419 }
3420 }
3421
3422
3423 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3424 PromotionInfo* promo) {
3425 MutexLocker x(ParGCRareEvent_lock);
3426 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3427 while (true) {
3428 // Expansion by some other thread might make alloc OK now:
3429 if (promo->ensure_spooling_space()) {
3430 assert(promo->has_spooling_space(),
3431 "Post-condition of successful ensure_spooling_space()");
3432 return true;
3433 }
3434 // If there's not enough expansion space available, give up.
3435 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3436 return false;
3437 }
3438 // Otherwise, we try expansion.
3439 expand(refill_size_bytes, MinHeapDeltaBytes,
3440 CMSExpansionCause::_allocate_par_spooling_space);
3441 // Now go around the loop and try alloc again;
3442 // A competing allocation might beat us to the expansion space,
3443 // so we may go around the loop again if allocation fails again.
3444 if (GCExpandToAllocateDelayMillis > 0) {
3445 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3446 }
3447 }
3448 }
3449
3450
3451 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3452 assert_locked_or_safepoint(ExpandHeap_lock);
3453 // Shrink committed space
3454 _virtual_space.shrink_by(bytes);
3455 // Shrink space; this also shrinks the space's BOT
3456 _cmsSpace->set_end((HeapWord*) _virtual_space.high());
3457 size_t new_word_size = heap_word_size(_cmsSpace->capacity());
3458 // Shrink the shared block offset array
3459 _bts->resize(new_word_size);
3460 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3461 // Shrink the card table
3462 Universe::heap()->barrier_set()->resize_covered_region(mr);
3463
3464 if (Verbose && PrintGC) {
3465 size_t new_mem_size = _virtual_space.committed_size();
3466 size_t old_mem_size = new_mem_size + bytes;
3467 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3468 name(), old_mem_size/K, new_mem_size/K);
3469 }
3470 }
3471
3472 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3473 assert_locked_or_safepoint(Heap_lock);
3474 size_t size = ReservedSpace::page_align_size_down(bytes);
3475 // Only shrink if a compaction was done so that all the free space
3476 // in the generation is in a contiguous block at the end.
3477 if (size > 0 && did_compact()) {
3478 shrink_by(size);
3479 }
3480 }
3481
3482 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3483 assert_locked_or_safepoint(Heap_lock);
3484 bool result = _virtual_space.expand_by(bytes);
3485 if (result) {
3486 size_t new_word_size =
3487 heap_word_size(_virtual_space.committed_size());
3488 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3489 _bts->resize(new_word_size); // resize the block offset shared array
3490 Universe::heap()->barrier_set()->resize_covered_region(mr);
3491 // Hmmmm... why doesn't CFLS::set_end verify locking?
3492 // This is quite ugly; FIX ME XXX
3493 _cmsSpace->assert_locked(freelistLock());
3494 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3495
3496 // update the space and generation capacity counters
3497 if (UsePerfData) {
3498 _space_counters->update_capacity();
3499 _gen_counters->update_all();
3500 }
3501
3502 if (Verbose && PrintGC) {
3503 size_t new_mem_size = _virtual_space.committed_size();
3504 size_t old_mem_size = new_mem_size - bytes;
3505 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3506 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3507 }
3508 }
3509 return result;
3510 }
3511
3512 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3513 assert_locked_or_safepoint(Heap_lock);
3514 bool success = true;
3515 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3516 if (remaining_bytes > 0) {
3517 success = grow_by(remaining_bytes);
3518 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3519 }
3520 return success;
3521 }
3522
3523 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
3524 assert_locked_or_safepoint(Heap_lock);
3525 assert_lock_strong(freelistLock());
3526 if (PrintGCDetails && Verbose) {
3527 warning("Shrinking of CMS not yet implemented");
3528 }
3529 return;
3530 }
3531
3532
3533 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3534 // phases.
3535 class CMSPhaseAccounting: public StackObj {
3536 public:
3537 CMSPhaseAccounting(CMSCollector *collector,
3538 const char *phase,
3539 bool print_cr = true);
3540 ~CMSPhaseAccounting();
3541
3542 private:
3543 CMSCollector *_collector;
3544 const char *_phase;
3545 elapsedTimer _wallclock;
3546 bool _print_cr;
3547
3548 public:
3549 // Not MT-safe; so do not pass around these StackObj's
3550 // where they may be accessed by other threads.
3551 jlong wallclock_millis() {
3552 assert(_wallclock.is_active(), "Wall clock should not stop");
3553 _wallclock.stop(); // to record time
3554 jlong ret = _wallclock.milliseconds();
3555 _wallclock.start(); // restart
3556 return ret;
3557 }
3558 };
3559
3560 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3561 const char *phase,
3562 bool print_cr) :
3563 _collector(collector), _phase(phase), _print_cr(print_cr) {
3564
3565 if (PrintCMSStatistics != 0) {
3566 _collector->resetYields();
3567 }
3568 if (PrintGCDetails) {
3569 gclog_or_tty->date_stamp(PrintGCDateStamps);
3570 gclog_or_tty->stamp(PrintGCTimeStamps);
3571 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
3572 _collector->cmsGen()->short_name(), _phase);
3573 }
3574 _collector->resetTimer();
3575 _wallclock.start();
3576 _collector->startTimer();
3577 }
3578
3579 CMSPhaseAccounting::~CMSPhaseAccounting() {
3580 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3581 _collector->stopTimer();
3582 _wallclock.stop();
3583 if (PrintGCDetails) {
3584 gclog_or_tty->date_stamp(PrintGCDateStamps);
3585 gclog_or_tty->stamp(PrintGCTimeStamps);
3586 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3587 _collector->cmsGen()->short_name(),
3588 _phase, _collector->timerValue(), _wallclock.seconds());
3589 if (_print_cr) {
3590 gclog_or_tty->print_cr("");
3591 }
3592 if (PrintCMSStatistics != 0) {
3593 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3594 _collector->yields());
3595 }
3596 }
3597 }
3598
3599 // CMS work
3600
3601 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
3602 class CMSParMarkTask : public AbstractGangTask {
3603 protected:
3604 CMSCollector* _collector;
3605 int _n_workers;
3606 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
3607 AbstractGangTask(name),
3608 _collector(collector),
3609 _n_workers(n_workers) {}
3610 // Work method in support of parallel rescan ... of young gen spaces
3611 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
3612 ContiguousSpace* space,
3613 HeapWord** chunk_array, size_t chunk_top);
3614 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
3615 };
3616
3617 // Parallel initial mark task
3618 class CMSParInitialMarkTask: public CMSParMarkTask {
3619 public:
3620 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
3621 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
3622 collector, n_workers) {}
3623 void work(uint worker_id);
3624 };
3625
3626 // Checkpoint the roots into this generation from outside
3627 // this generation. [Note this initial checkpoint need only
3628 // be approximate -- we'll do a catch up phase subsequently.]
3629 void CMSCollector::checkpointRootsInitial(bool asynch) {
3630 assert(_collectorState == InitialMarking, "Wrong collector state");
3631 check_correct_thread_executing();
3632 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3633
3634 save_heap_summary();
3635 report_heap_summary(GCWhen::BeforeGC);
3636
3637 ReferenceProcessor* rp = ref_processor();
3638 SpecializationStats::clear();
3639 assert(_restart_addr == NULL, "Control point invariant");
3640 if (asynch) {
3641 // acquire locks for subsequent manipulations
3642 MutexLockerEx x(bitMapLock(),
3643 Mutex::_no_safepoint_check_flag);
3644 checkpointRootsInitialWork(asynch);
3645 // enable ("weak") refs discovery
3646 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3647 _collectorState = Marking;
3648 } else {
3649 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3650 // which recognizes if we are a CMS generation, and doesn't try to turn on
3651 // discovery; verify that they aren't meddling.
3652 assert(!rp->discovery_is_atomic(),
3653 "incorrect setting of discovery predicate");
3654 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3655 "ref discovery for this generation kind");
3656 // already have locks
3657 checkpointRootsInitialWork(asynch);
3658 // now enable ("weak") refs discovery
3659 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3660 _collectorState = Marking;
3661 }
3662 SpecializationStats::print();
3663 }
3664
3665 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3666 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3667 assert(_collectorState == InitialMarking, "just checking");
3668
3669 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3670 // precede our marking with a collection of all
3671 // younger generations to keep floating garbage to a minimum.
3672 // XXX: we won't do this for now -- it's an optimization to be done later.
3673
3674 // already have locks
3675 assert_lock_strong(bitMapLock());
3676 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3677
3678 // Setup the verification and class unloading state for this
3679 // CMS collection cycle.
3680 setup_cms_unloading_and_verification_state();
3681
3682 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3683 PrintGCDetails && Verbose, true, _gc_timer_cm);)
3684 if (UseAdaptiveSizePolicy) {
3685 size_policy()->checkpoint_roots_initial_begin();
3686 }
3687
3688 // Reset all the PLAB chunk arrays if necessary.
3689 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3690 reset_survivor_plab_arrays();
3691 }
3692
3693 ResourceMark rm;
3694 HandleMark hm;
3695
3696 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3697 GenCollectedHeap* gch = GenCollectedHeap::heap();
3698
3699 verify_work_stacks_empty();
3700 verify_overflow_empty();
3701
3702 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3703 // Update the saved marks which may affect the root scans.
3704 gch->save_marks();
3705
3706 // weak reference processing has not started yet.
3707 ref_processor()->set_enqueuing_is_done(false);
3708
3709 // Need to remember all newly created CLDs,
3710 // so that we can guarantee that the remark finds them.
3711 ClassLoaderDataGraph::remember_new_clds(true);
3712
3713 // Whenever a CLD is found, it will be claimed before proceeding to mark
3714 // the klasses. The claimed marks need to be cleared before marking starts.
3715 ClassLoaderDataGraph::clear_claimed_marks();
3716
3717 if (CMSPrintEdenSurvivorChunks) {
3718 print_eden_and_survivor_chunk_arrays();
3719 }
3720
3721 {
3722 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3723 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
3724 // The parallel version.
3725 FlexibleWorkGang* workers = gch->workers();
3726 assert(workers != NULL, "Need parallel worker threads.");
3727 int n_workers = workers->active_workers();
3728 CMSParInitialMarkTask tsk(this, n_workers);
3729 gch->set_par_threads(n_workers);
3730 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3731 if (n_workers > 1) {
3732 GenCollectedHeap::StrongRootsScope srs(gch);
3733 workers->run_task(&tsk);
3734 } else {
3735 GenCollectedHeap::StrongRootsScope srs(gch);
3736 tsk.work(0);
3737 }
3738 gch->set_par_threads(0);
3739 } else {
3740 // The serial version.
3741 CMKlassClosure klass_closure(¬Older);
3742 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3743 gch->gen_process_strong_roots(_cmsGen->level(),
3744 true, // younger gens are roots
3745 true, // activate StrongRootsScope
3746 SharedHeap::ScanningOption(roots_scanning_options()),
3747 ¬Older,
3748 NULL,
3749 &klass_closure);
3750 }
3751 }
3752
3753 // Clear mod-union table; it will be dirtied in the prologue of
3754 // CMS generation per each younger generation collection.
3755
3756 assert(_modUnionTable.isAllClear(),
3757 "Was cleared in most recent final checkpoint phase"
3758 " or no bits are set in the gc_prologue before the start of the next "
3759 "subsequent marking phase.");
3760
3761 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3762
3763 // Save the end of the used_region of the constituent generations
3764 // to be used to limit the extent of sweep in each generation.
3765 save_sweep_limits();
3766 if (UseAdaptiveSizePolicy) {
3767 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3768 }
3769 verify_overflow_empty();
3770 }
3771
3772 bool CMSCollector::markFromRoots(bool asynch) {
3773 // we might be tempted to assert that:
3774 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3775 // "inconsistent argument?");
3776 // However that wouldn't be right, because it's possible that
3777 // a safepoint is indeed in progress as a younger generation
3778 // stop-the-world GC happens even as we mark in this generation.
3779 assert(_collectorState == Marking, "inconsistent state?");
3780 check_correct_thread_executing();
3781 verify_overflow_empty();
3782
3783 bool res;
3784 if (asynch) {
3785
3786 // Start the timers for adaptive size policy for the concurrent phases
3787 // Do it here so that the foreground MS can use the concurrent
3788 // timer since a foreground MS might has the sweep done concurrently
3789 // or STW.
3790 if (UseAdaptiveSizePolicy) {
3791 size_policy()->concurrent_marking_begin();
3792 }
3793
3794 // Weak ref discovery note: We may be discovering weak
3795 // refs in this generation concurrent (but interleaved) with
3796 // weak ref discovery by a younger generation collector.
3797
3798 CMSTokenSyncWithLocks ts(true, bitMapLock());
3799 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3800 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3801 res = markFromRootsWork(asynch);
3802 if (res) {
3803 _collectorState = Precleaning;
3804 } else { // We failed and a foreground collection wants to take over
3805 assert(_foregroundGCIsActive, "internal state inconsistency");
3806 assert(_restart_addr == NULL, "foreground will restart from scratch");
3807 if (PrintGCDetails) {
3808 gclog_or_tty->print_cr("bailing out to foreground collection");
3809 }
3810 }
3811 if (UseAdaptiveSizePolicy) {
3812 size_policy()->concurrent_marking_end();
3813 }
3814 } else {
3815 assert(SafepointSynchronize::is_at_safepoint(),
3816 "inconsistent with asynch == false");
3817 if (UseAdaptiveSizePolicy) {
3818 size_policy()->ms_collection_marking_begin();
3819 }
3820 // already have locks
3821 res = markFromRootsWork(asynch);
3822 _collectorState = FinalMarking;
3823 if (UseAdaptiveSizePolicy) {
3824 GenCollectedHeap* gch = GenCollectedHeap::heap();
3825 size_policy()->ms_collection_marking_end(gch->gc_cause());
3826 }
3827 }
3828 verify_overflow_empty();
3829 return res;
3830 }
3831
3832 bool CMSCollector::markFromRootsWork(bool asynch) {
3833 // iterate over marked bits in bit map, doing a full scan and mark
3834 // from these roots using the following algorithm:
3835 // . if oop is to the right of the current scan pointer,
3836 // mark corresponding bit (we'll process it later)
3837 // . else (oop is to left of current scan pointer)
3838 // push oop on marking stack
3839 // . drain the marking stack
3840
3841 // Note that when we do a marking step we need to hold the
3842 // bit map lock -- recall that direct allocation (by mutators)
3843 // and promotion (by younger generation collectors) is also
3844 // marking the bit map. [the so-called allocate live policy.]
3845 // Because the implementation of bit map marking is not
3846 // robust wrt simultaneous marking of bits in the same word,
3847 // we need to make sure that there is no such interference
3848 // between concurrent such updates.
3849
3850 // already have locks
3851 assert_lock_strong(bitMapLock());
3852
3853 verify_work_stacks_empty();
3854 verify_overflow_empty();
3855 bool result = false;
3856 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3857 result = do_marking_mt(asynch);
3858 } else {
3859 result = do_marking_st(asynch);
3860 }
3861 return result;
3862 }
3863
3864 // Forward decl
3865 class CMSConcMarkingTask;
3866
3867 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3868 CMSCollector* _collector;
3869 CMSConcMarkingTask* _task;
3870 public:
3871 virtual void yield();
3872
3873 // "n_threads" is the number of threads to be terminated.
3874 // "queue_set" is a set of work queues of other threads.
3875 // "collector" is the CMS collector associated with this task terminator.
3876 // "yield" indicates whether we need the gang as a whole to yield.
3877 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3878 ParallelTaskTerminator(n_threads, queue_set),
3879 _collector(collector) { }
3880
3881 void set_task(CMSConcMarkingTask* task) {
3882 _task = task;
3883 }
3884 };
3885
3886 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3887 CMSConcMarkingTask* _task;
3888 public:
3889 bool should_exit_termination();
3890 void set_task(CMSConcMarkingTask* task) {
3891 _task = task;
3892 }
3893 };
3894
3895 // MT Concurrent Marking Task
3896 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3897 CMSCollector* _collector;
3898 int _n_workers; // requested/desired # workers
3899 bool _asynch;
3900 bool _result;
3901 CompactibleFreeListSpace* _cms_space;
3902 char _pad_front[64]; // padding to ...
3903 HeapWord* _global_finger; // ... avoid sharing cache line
3904 char _pad_back[64];
3905 HeapWord* _restart_addr;
3906
3907 // Exposed here for yielding support
3908 Mutex* const _bit_map_lock;
3909
3910 // The per thread work queues, available here for stealing
3911 OopTaskQueueSet* _task_queues;
3912
3913 // Termination (and yielding) support
3914 CMSConcMarkingTerminator _term;
3915 CMSConcMarkingTerminatorTerminator _term_term;
3916
3917 public:
3918 CMSConcMarkingTask(CMSCollector* collector,
3919 CompactibleFreeListSpace* cms_space,
3920 bool asynch,
3921 YieldingFlexibleWorkGang* workers,
3922 OopTaskQueueSet* task_queues):
3923 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3924 _collector(collector),
3925 _cms_space(cms_space),
3926 _asynch(asynch), _n_workers(0), _result(true),
3927 _task_queues(task_queues),
3928 _term(_n_workers, task_queues, _collector),
3929 _bit_map_lock(collector->bitMapLock())
3930 {
3931 _requested_size = _n_workers;
3932 _term.set_task(this);
3933 _term_term.set_task(this);
3934 _restart_addr = _global_finger = _cms_space->bottom();
3935 }
3936
3937
3938 OopTaskQueueSet* task_queues() { return _task_queues; }
3939
3940 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3941
3942 HeapWord** global_finger_addr() { return &_global_finger; }
3943
3944 CMSConcMarkingTerminator* terminator() { return &_term; }
3945
3946 virtual void set_for_termination(int active_workers) {
3947 terminator()->reset_for_reuse(active_workers);
3948 }
3949
3950 void work(uint worker_id);
3951 bool should_yield() {
3952 return ConcurrentMarkSweepThread::should_yield()
3953 && !_collector->foregroundGCIsActive()
3954 && _asynch;
3955 }
3956
3957 virtual void coordinator_yield(); // stuff done by coordinator
3958 bool result() { return _result; }
3959
3960 void reset(HeapWord* ra) {
3961 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3962 _restart_addr = _global_finger = ra;
3963 _term.reset_for_reuse();
3964 }
3965
3966 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3967 OopTaskQueue* work_q);
3968
3969 private:
3970 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3971 void do_work_steal(int i);
3972 void bump_global_finger(HeapWord* f);
3973 };
3974
3975 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3976 assert(_task != NULL, "Error");
3977 return _task->yielding();
3978 // Note that we do not need the disjunct || _task->should_yield() above
3979 // because we want terminating threads to yield only if the task
3980 // is already in the midst of yielding, which happens only after at least one
3981 // thread has yielded.
3982 }
3983
3984 void CMSConcMarkingTerminator::yield() {
3985 if (_task->should_yield()) {
3986 _task->yield();
3987 } else {
3988 ParallelTaskTerminator::yield();
3989 }
3990 }
3991
3992 ////////////////////////////////////////////////////////////////
3993 // Concurrent Marking Algorithm Sketch
3994 ////////////////////////////////////////////////////////////////
3995 // Until all tasks exhausted (both spaces):
3996 // -- claim next available chunk
3997 // -- bump global finger via CAS
3998 // -- find first object that starts in this chunk
3999 // and start scanning bitmap from that position
4000 // -- scan marked objects for oops
4001 // -- CAS-mark target, and if successful:
4002 // . if target oop is above global finger (volatile read)
4003 // nothing to do
4004 // . if target oop is in chunk and above local finger
4005 // then nothing to do
4006 // . else push on work-queue
4007 // -- Deal with possible overflow issues:
4008 // . local work-queue overflow causes stuff to be pushed on
4009 // global (common) overflow queue
4010 // . always first empty local work queue
4011 // . then get a batch of oops from global work queue if any
4012 // . then do work stealing
4013 // -- When all tasks claimed (both spaces)
4014 // and local work queue empty,
4015 // then in a loop do:
4016 // . check global overflow stack; steal a batch of oops and trace
4017 // . try to steal from other threads oif GOS is empty
4018 // . if neither is available, offer termination
4019 // -- Terminate and return result
4020 //
4021 void CMSConcMarkingTask::work(uint worker_id) {
4022 elapsedTimer _timer;
4023 ResourceMark rm;
4024 HandleMark hm;
4025
4026 DEBUG_ONLY(_collector->verify_overflow_empty();)
4027
4028 // Before we begin work, our work queue should be empty
4029 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
4030 // Scan the bitmap covering _cms_space, tracing through grey objects.
4031 _timer.start();
4032 do_scan_and_mark(worker_id, _cms_space);
4033 _timer.stop();
4034 if (PrintCMSStatistics != 0) {
4035 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
4036 worker_id, _timer.seconds());
4037 // XXX: need xxx/xxx type of notation, two timers
4038 }
4039
4040 // ... do work stealing
4041 _timer.reset();
4042 _timer.start();
4043 do_work_steal(worker_id);
4044 _timer.stop();
4045 if (PrintCMSStatistics != 0) {
4046 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
4047 worker_id, _timer.seconds());
4048 // XXX: need xxx/xxx type of notation, two timers
4049 }
4050 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
4051 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
4052 // Note that under the current task protocol, the
4053 // following assertion is true even of the spaces
4054 // expanded since the completion of the concurrent
4055 // marking. XXX This will likely change under a strict
4056 // ABORT semantics.
4057 // After perm removal the comparison was changed to
4058 // greater than or equal to from strictly greater than.
4059 // Before perm removal the highest address sweep would
4060 // have been at the end of perm gen but now is at the
4061 // end of the tenured gen.
4062 assert(_global_finger >= _cms_space->end(),
4063 "All tasks have been completed");
4064 DEBUG_ONLY(_collector->verify_overflow_empty();)
4065 }
4066
4067 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
4068 HeapWord* read = _global_finger;
4069 HeapWord* cur = read;
4070 while (f > read) {
4071 cur = read;
4072 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
4073 if (cur == read) {
4074 // our cas succeeded
4075 assert(_global_finger >= f, "protocol consistency");
4076 break;
4077 }
4078 }
4079 }
4080
4081 // This is really inefficient, and should be redone by
4082 // using (not yet available) block-read and -write interfaces to the
4083 // stack and the work_queue. XXX FIX ME !!!
4084 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
4085 OopTaskQueue* work_q) {
4086 // Fast lock-free check
4087 if (ovflw_stk->length() == 0) {
4088 return false;
4089 }
4090 assert(work_q->size() == 0, "Shouldn't steal");
4091 MutexLockerEx ml(ovflw_stk->par_lock(),
4092 Mutex::_no_safepoint_check_flag);
4093 // Grab up to 1/4 the size of the work queue
4094 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4095 (size_t)ParGCDesiredObjsFromOverflowList);
4096 num = MIN2(num, ovflw_stk->length());
4097 for (int i = (int) num; i > 0; i--) {
4098 oop cur = ovflw_stk->pop();
4099 assert(cur != NULL, "Counted wrong?");
4100 work_q->push(cur);
4101 }
4102 return num > 0;
4103 }
4104
4105 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
4106 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4107 int n_tasks = pst->n_tasks();
4108 // We allow that there may be no tasks to do here because
4109 // we are restarting after a stack overflow.
4110 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
4111 uint nth_task = 0;
4112
4113 HeapWord* aligned_start = sp->bottom();
4114 if (sp->used_region().contains(_restart_addr)) {
4115 // Align down to a card boundary for the start of 0th task
4116 // for this space.
4117 aligned_start =
4118 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
4119 CardTableModRefBS::card_size);
4120 }
4121
4122 size_t chunk_size = sp->marking_task_size();
4123 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4124 // Having claimed the nth task in this space,
4125 // compute the chunk that it corresponds to:
4126 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
4127 aligned_start + (nth_task+1)*chunk_size);
4128 // Try and bump the global finger via a CAS;
4129 // note that we need to do the global finger bump
4130 // _before_ taking the intersection below, because
4131 // the task corresponding to that region will be
4132 // deemed done even if the used_region() expands
4133 // because of allocation -- as it almost certainly will
4134 // during start-up while the threads yield in the
4135 // closure below.
4136 HeapWord* finger = span.end();
4137 bump_global_finger(finger); // atomically
4138 // There are null tasks here corresponding to chunks
4139 // beyond the "top" address of the space.
4140 span = span.intersection(sp->used_region());
4141 if (!span.is_empty()) { // Non-null task
4142 HeapWord* prev_obj;
4143 assert(!span.contains(_restart_addr) || nth_task == 0,
4144 "Inconsistency");
4145 if (nth_task == 0) {
4146 // For the 0th task, we'll not need to compute a block_start.
4147 if (span.contains(_restart_addr)) {
4148 // In the case of a restart because of stack overflow,
4149 // we might additionally skip a chunk prefix.
4150 prev_obj = _restart_addr;
4151 } else {
4152 prev_obj = span.start();
4153 }
4154 } else {
4155 // We want to skip the first object because
4156 // the protocol is to scan any object in its entirety
4157 // that _starts_ in this span; a fortiori, any
4158 // object starting in an earlier span is scanned
4159 // as part of an earlier claimed task.
4160 // Below we use the "careful" version of block_start
4161 // so we do not try to navigate uninitialized objects.
4162 prev_obj = sp->block_start_careful(span.start());
4163 // Below we use a variant of block_size that uses the
4164 // Printezis bits to avoid waiting for allocated
4165 // objects to become initialized/parsable.
4166 while (prev_obj < span.start()) {
4167 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4168 if (sz > 0) {
4169 prev_obj += sz;
4170 } else {
4171 // In this case we may end up doing a bit of redundant
4172 // scanning, but that appears unavoidable, short of
4173 // locking the free list locks; see bug 6324141.
4174 break;
4175 }
4176 }
4177 }
4178 if (prev_obj < span.end()) {
4179 MemRegion my_span = MemRegion(prev_obj, span.end());
4180 // Do the marking work within a non-empty span --
4181 // the last argument to the constructor indicates whether the
4182 // iteration should be incremental with periodic yields.
4183 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4184 &_collector->_markBitMap,
4185 work_queue(i),
4186 &_collector->_markStack,
4187 _asynch);
4188 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4189 } // else nothing to do for this task
4190 } // else nothing to do for this task
4191 }
4192 // We'd be tempted to assert here that since there are no
4193 // more tasks left to claim in this space, the global_finger
4194 // must exceed space->top() and a fortiori space->end(). However,
4195 // that would not quite be correct because the bumping of
4196 // global_finger occurs strictly after the claiming of a task,
4197 // so by the time we reach here the global finger may not yet
4198 // have been bumped up by the thread that claimed the last
4199 // task.
4200 pst->all_tasks_completed();
4201 }
4202
4203 class Par_ConcMarkingClosure: public CMSOopClosure {
4204 private:
4205 CMSCollector* _collector;
4206 CMSConcMarkingTask* _task;
4207 MemRegion _span;
4208 CMSBitMap* _bit_map;
4209 CMSMarkStack* _overflow_stack;
4210 OopTaskQueue* _work_queue;
4211 protected:
4212 DO_OOP_WORK_DEFN
4213 public:
4214 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4215 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
4216 CMSOopClosure(collector->ref_processor()),
4217 _collector(collector),
4218 _task(task),
4219 _span(collector->_span),
4220 _work_queue(work_queue),
4221 _bit_map(bit_map),
4222 _overflow_stack(overflow_stack)
4223 { }
4224 virtual void do_oop(oop* p);
4225 virtual void do_oop(narrowOop* p);
4226
4227 void trim_queue(size_t max);
4228 void handle_stack_overflow(HeapWord* lost);
4229 void do_yield_check() {
4230 if (_task->should_yield()) {
4231 _task->yield();
4232 }
4233 }
4234 };
4235
4236 // Grey object scanning during work stealing phase --
4237 // the salient assumption here is that any references
4238 // that are in these stolen objects being scanned must
4239 // already have been initialized (else they would not have
4240 // been published), so we do not need to check for
4241 // uninitialized objects before pushing here.
4242 void Par_ConcMarkingClosure::do_oop(oop obj) {
4243 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4244 HeapWord* addr = (HeapWord*)obj;
4245 // Check if oop points into the CMS generation
4246 // and is not marked
4247 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4248 // a white object ...
4249 // If we manage to "claim" the object, by being the
4250 // first thread to mark it, then we push it on our
4251 // marking stack
4252 if (_bit_map->par_mark(addr)) { // ... now grey
4253 // push on work queue (grey set)
4254 bool simulate_overflow = false;
4255 NOT_PRODUCT(
4256 if (CMSMarkStackOverflowALot &&
4257 _collector->simulate_overflow()) {
4258 // simulate a stack overflow
4259 simulate_overflow = true;
4260 }
4261 )
4262 if (simulate_overflow ||
4263 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4264 // stack overflow
4265 if (PrintCMSStatistics != 0) {
4266 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4267 SIZE_FORMAT, _overflow_stack->capacity());
4268 }
4269 // We cannot assert that the overflow stack is full because
4270 // it may have been emptied since.
4271 assert(simulate_overflow ||
4272 _work_queue->size() == _work_queue->max_elems(),
4273 "Else push should have succeeded");
4274 handle_stack_overflow(addr);
4275 }
4276 } // Else, some other thread got there first
4277 do_yield_check();
4278 }
4279 }
4280
4281 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4282 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4283
4284 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4285 while (_work_queue->size() > max) {
4286 oop new_oop;
4287 if (_work_queue->pop_local(new_oop)) {
4288 assert(new_oop->is_oop(), "Should be an oop");
4289 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4290 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4291 new_oop->oop_iterate(this); // do_oop() above
4292 do_yield_check();
4293 }
4294 }
4295 }
4296
4297 // Upon stack overflow, we discard (part of) the stack,
4298 // remembering the least address amongst those discarded
4299 // in CMSCollector's _restart_address.
4300 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4301 // We need to do this under a mutex to prevent other
4302 // workers from interfering with the work done below.
4303 MutexLockerEx ml(_overflow_stack->par_lock(),
4304 Mutex::_no_safepoint_check_flag);
4305 // Remember the least grey address discarded
4306 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4307 _collector->lower_restart_addr(ra);
4308 _overflow_stack->reset(); // discard stack contents
4309 _overflow_stack->expand(); // expand the stack if possible
4310 }
4311
4312
4313 void CMSConcMarkingTask::do_work_steal(int i) {
4314 OopTaskQueue* work_q = work_queue(i);
4315 oop obj_to_scan;
4316 CMSBitMap* bm = &(_collector->_markBitMap);
4317 CMSMarkStack* ovflw = &(_collector->_markStack);
4318 int* seed = _collector->hash_seed(i);
4319 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
4320 while (true) {
4321 cl.trim_queue(0);
4322 assert(work_q->size() == 0, "Should have been emptied above");
4323 if (get_work_from_overflow_stack(ovflw, work_q)) {
4324 // Can't assert below because the work obtained from the
4325 // overflow stack may already have been stolen from us.
4326 // assert(work_q->size() > 0, "Work from overflow stack");
4327 continue;
4328 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4329 assert(obj_to_scan->is_oop(), "Should be an oop");
4330 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4331 obj_to_scan->oop_iterate(&cl);
4332 } else if (terminator()->offer_termination(&_term_term)) {
4333 assert(work_q->size() == 0, "Impossible!");
4334 break;
4335 } else if (yielding() || should_yield()) {
4336 yield();
4337 }
4338 }
4339 }
4340
4341 // This is run by the CMS (coordinator) thread.
4342 void CMSConcMarkingTask::coordinator_yield() {
4343 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4344 "CMS thread should hold CMS token");
4345 // First give up the locks, then yield, then re-lock
4346 // We should probably use a constructor/destructor idiom to
4347 // do this unlock/lock or modify the MutexUnlocker class to
4348 // serve our purpose. XXX
4349 assert_lock_strong(_bit_map_lock);
4350 _bit_map_lock->unlock();
4351 ConcurrentMarkSweepThread::desynchronize(true);
4352 ConcurrentMarkSweepThread::acknowledge_yield_request();
4353 _collector->stopTimer();
4354 if (PrintCMSStatistics != 0) {
4355 _collector->incrementYields();
4356 }
4357 _collector->icms_wait();
4358
4359 // It is possible for whichever thread initiated the yield request
4360 // not to get a chance to wake up and take the bitmap lock between
4361 // this thread releasing it and reacquiring it. So, while the
4362 // should_yield() flag is on, let's sleep for a bit to give the
4363 // other thread a chance to wake up. The limit imposed on the number
4364 // of iterations is defensive, to avoid any unforseen circumstances
4365 // putting us into an infinite loop. Since it's always been this
4366 // (coordinator_yield()) method that was observed to cause the
4367 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4368 // which is by default non-zero. For the other seven methods that
4369 // also perform the yield operation, as are using a different
4370 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4371 // can enable the sleeping for those methods too, if necessary.
4372 // See 6442774.
4373 //
4374 // We really need to reconsider the synchronization between the GC
4375 // thread and the yield-requesting threads in the future and we
4376 // should really use wait/notify, which is the recommended
4377 // way of doing this type of interaction. Additionally, we should
4378 // consolidate the eight methods that do the yield operation and they
4379 // are almost identical into one for better maintainability and
4380 // readability. See 6445193.
4381 //
4382 // Tony 2006.06.29
4383 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4384 ConcurrentMarkSweepThread::should_yield() &&
4385 !CMSCollector::foregroundGCIsActive(); ++i) {
4386 os::sleep(Thread::current(), 1, false);
4387 ConcurrentMarkSweepThread::acknowledge_yield_request();
4388 }
4389
4390 ConcurrentMarkSweepThread::synchronize(true);
4391 _bit_map_lock->lock_without_safepoint_check();
4392 _collector->startTimer();
4393 }
4394
4395 bool CMSCollector::do_marking_mt(bool asynch) {
4396 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4397 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4398 conc_workers()->total_workers(),
4399 conc_workers()->active_workers(),
4400 Threads::number_of_non_daemon_threads());
4401 conc_workers()->set_active_workers(num_workers);
4402
4403 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4404
4405 CMSConcMarkingTask tsk(this,
4406 cms_space,
4407 asynch,
4408 conc_workers(),
4409 task_queues());
4410
4411 // Since the actual number of workers we get may be different
4412 // from the number we requested above, do we need to do anything different
4413 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4414 // class?? XXX
4415 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4416
4417 // Refs discovery is already non-atomic.
4418 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4419 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4420 conc_workers()->start_task(&tsk);
4421 while (tsk.yielded()) {
4422 tsk.coordinator_yield();
4423 conc_workers()->continue_task(&tsk);
4424 }
4425 // If the task was aborted, _restart_addr will be non-NULL
4426 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4427 while (_restart_addr != NULL) {
4428 // XXX For now we do not make use of ABORTED state and have not
4429 // yet implemented the right abort semantics (even in the original
4430 // single-threaded CMS case). That needs some more investigation
4431 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4432 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4433 // If _restart_addr is non-NULL, a marking stack overflow
4434 // occurred; we need to do a fresh marking iteration from the
4435 // indicated restart address.
4436 if (_foregroundGCIsActive && asynch) {
4437 // We may be running into repeated stack overflows, having
4438 // reached the limit of the stack size, while making very
4439 // slow forward progress. It may be best to bail out and
4440 // let the foreground collector do its job.
4441 // Clear _restart_addr, so that foreground GC
4442 // works from scratch. This avoids the headache of
4443 // a "rescan" which would otherwise be needed because
4444 // of the dirty mod union table & card table.
4445 _restart_addr = NULL;
4446 return false;
4447 }
4448 // Adjust the task to restart from _restart_addr
4449 tsk.reset(_restart_addr);
4450 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4451 _restart_addr);
4452 _restart_addr = NULL;
4453 // Get the workers going again
4454 conc_workers()->start_task(&tsk);
4455 while (tsk.yielded()) {
4456 tsk.coordinator_yield();
4457 conc_workers()->continue_task(&tsk);
4458 }
4459 }
4460 assert(tsk.completed(), "Inconsistency");
4461 assert(tsk.result() == true, "Inconsistency");
4462 return true;
4463 }
4464
4465 bool CMSCollector::do_marking_st(bool asynch) {
4466 ResourceMark rm;
4467 HandleMark hm;
4468
4469 // Temporarily make refs discovery single threaded (non-MT)
4470 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4471 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4472 &_markStack, CMSYield && asynch);
4473 // the last argument to iterate indicates whether the iteration
4474 // should be incremental with periodic yields.
4475 _markBitMap.iterate(&markFromRootsClosure);
4476 // If _restart_addr is non-NULL, a marking stack overflow
4477 // occurred; we need to do a fresh iteration from the
4478 // indicated restart address.
4479 while (_restart_addr != NULL) {
4480 if (_foregroundGCIsActive && asynch) {
4481 // We may be running into repeated stack overflows, having
4482 // reached the limit of the stack size, while making very
4483 // slow forward progress. It may be best to bail out and
4484 // let the foreground collector do its job.
4485 // Clear _restart_addr, so that foreground GC
4486 // works from scratch. This avoids the headache of
4487 // a "rescan" which would otherwise be needed because
4488 // of the dirty mod union table & card table.
4489 _restart_addr = NULL;
4490 return false; // indicating failure to complete marking
4491 }
4492 // Deal with stack overflow:
4493 // we restart marking from _restart_addr
4494 HeapWord* ra = _restart_addr;
4495 markFromRootsClosure.reset(ra);
4496 _restart_addr = NULL;
4497 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4498 }
4499 return true;
4500 }
4501
4502 void CMSCollector::preclean() {
4503 check_correct_thread_executing();
4504 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4505 verify_work_stacks_empty();
4506 verify_overflow_empty();
4507 _abort_preclean = false;
4508 if (CMSPrecleaningEnabled) {
4509 if (!CMSEdenChunksRecordAlways) {
4510 _eden_chunk_index = 0;
4511 }
4512 size_t used = get_eden_used();
4513 size_t capacity = get_eden_capacity();
4514 // Don't start sampling unless we will get sufficiently
4515 // many samples.
4516 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4517 * CMSScheduleRemarkEdenPenetration)) {
4518 _start_sampling = true;
4519 } else {
4520 _start_sampling = false;
4521 }
4522 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4523 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4524 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4525 }
4526 CMSTokenSync x(true); // is cms thread
4527 if (CMSPrecleaningEnabled) {
4528 sample_eden();
4529 _collectorState = AbortablePreclean;
4530 } else {
4531 _collectorState = FinalMarking;
4532 }
4533 verify_work_stacks_empty();
4534 verify_overflow_empty();
4535 }
4536
4537 // Try and schedule the remark such that young gen
4538 // occupancy is CMSScheduleRemarkEdenPenetration %.
4539 void CMSCollector::abortable_preclean() {
4540 check_correct_thread_executing();
4541 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4542 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4543
4544 // If Eden's current occupancy is below this threshold,
4545 // immediately schedule the remark; else preclean
4546 // past the next scavenge in an effort to
4547 // schedule the pause as described above. By choosing
4548 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4549 // we will never do an actual abortable preclean cycle.
4550 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4551 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4552 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4553 // We need more smarts in the abortable preclean
4554 // loop below to deal with cases where allocation
4555 // in young gen is very very slow, and our precleaning
4556 // is running a losing race against a horde of
4557 // mutators intent on flooding us with CMS updates
4558 // (dirty cards).
4559 // One, admittedly dumb, strategy is to give up
4560 // after a certain number of abortable precleaning loops
4561 // or after a certain maximum time. We want to make
4562 // this smarter in the next iteration.
4563 // XXX FIX ME!!! YSR
4564 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4565 while (!(should_abort_preclean() ||
4566 ConcurrentMarkSweepThread::should_terminate())) {
4567 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4568 cumworkdone += workdone;
4569 loops++;
4570 // Voluntarily terminate abortable preclean phase if we have
4571 // been at it for too long.
4572 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4573 loops >= CMSMaxAbortablePrecleanLoops) {
4574 if (PrintGCDetails) {
4575 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4576 }
4577 break;
4578 }
4579 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4580 if (PrintGCDetails) {
4581 gclog_or_tty->print(" CMS: abort preclean due to time ");
4582 }
4583 break;
4584 }
4585 // If we are doing little work each iteration, we should
4586 // take a short break.
4587 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4588 // Sleep for some time, waiting for work to accumulate
4589 stopTimer();
4590 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4591 startTimer();
4592 waited++;
4593 }
4594 }
4595 if (PrintCMSStatistics > 0) {
4596 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4597 loops, waited, cumworkdone);
4598 }
4599 }
4600 CMSTokenSync x(true); // is cms thread
4601 if (_collectorState != Idling) {
4602 assert(_collectorState == AbortablePreclean,
4603 "Spontaneous state transition?");
4604 _collectorState = FinalMarking;
4605 } // Else, a foreground collection completed this CMS cycle.
4606 return;
4607 }
4608
4609 // Respond to an Eden sampling opportunity
4610 void CMSCollector::sample_eden() {
4611 // Make sure a young gc cannot sneak in between our
4612 // reading and recording of a sample.
4613 assert(Thread::current()->is_ConcurrentGC_thread(),
4614 "Only the cms thread may collect Eden samples");
4615 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4616 "Should collect samples while holding CMS token");
4617 if (!_start_sampling) {
4618 return;
4619 }
4620 // When CMSEdenChunksRecordAlways is true, the eden chunk array
4621 // is populated by the young generation.
4622 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
4623 if (_eden_chunk_index < _eden_chunk_capacity) {
4624 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4625 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4626 "Unexpected state of Eden");
4627 // We'd like to check that what we just sampled is an oop-start address;
4628 // however, we cannot do that here since the object may not yet have been
4629 // initialized. So we'll instead do the check when we _use_ this sample
4630 // later.
4631 if (_eden_chunk_index == 0 ||
4632 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4633 _eden_chunk_array[_eden_chunk_index-1])
4634 >= CMSSamplingGrain)) {
4635 _eden_chunk_index++; // commit sample
4636 }
4637 }
4638 }
4639 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4640 size_t used = get_eden_used();
4641 size_t capacity = get_eden_capacity();
4642 assert(used <= capacity, "Unexpected state of Eden");
4643 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4644 _abort_preclean = true;
4645 }
4646 }
4647 }
4648
4649
4650 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4651 assert(_collectorState == Precleaning ||
4652 _collectorState == AbortablePreclean, "incorrect state");
4653 ResourceMark rm;
4654 HandleMark hm;
4655
4656 // Precleaning is currently not MT but the reference processor
4657 // may be set for MT. Disable it temporarily here.
4658 ReferenceProcessor* rp = ref_processor();
4659 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4660
4661 // Do one pass of scrubbing the discovered reference lists
4662 // to remove any reference objects with strongly-reachable
4663 // referents.
4664 if (clean_refs) {
4665 CMSPrecleanRefsYieldClosure yield_cl(this);
4666 assert(rp->span().equals(_span), "Spans should be equal");
4667 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4668 &_markStack, true /* preclean */);
4669 CMSDrainMarkingStackClosure complete_trace(this,
4670 _span, &_markBitMap, &_markStack,
4671 &keep_alive, true /* preclean */);
4672
4673 // We don't want this step to interfere with a young
4674 // collection because we don't want to take CPU
4675 // or memory bandwidth away from the young GC threads
4676 // (which may be as many as there are CPUs).
4677 // Note that we don't need to protect ourselves from
4678 // interference with mutators because they can't
4679 // manipulate the discovered reference lists nor affect
4680 // the computed reachability of the referents, the
4681 // only properties manipulated by the precleaning
4682 // of these reference lists.
4683 stopTimer();
4684 CMSTokenSyncWithLocks x(true /* is cms thread */,
4685 bitMapLock());
4686 startTimer();
4687 sample_eden();
4688
4689 // The following will yield to allow foreground
4690 // collection to proceed promptly. XXX YSR:
4691 // The code in this method may need further
4692 // tweaking for better performance and some restructuring
4693 // for cleaner interfaces.
4694 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4695 rp->preclean_discovered_references(
4696 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
4697 gc_timer);
4698 }
4699
4700 if (clean_survivor) { // preclean the active survivor space(s)
4701 assert(_young_gen->kind() == Generation::DefNew ||
4702 _young_gen->kind() == Generation::ParNew ||
4703 _young_gen->kind() == Generation::ASParNew,
4704 "incorrect type for cast");
4705 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4706 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4707 &_markBitMap, &_modUnionTable,
4708 &_markStack, true /* precleaning phase */);
4709 stopTimer();
4710 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4711 bitMapLock());
4712 startTimer();
4713 unsigned int before_count =
4714 GenCollectedHeap::heap()->total_collections();
4715 SurvivorSpacePrecleanClosure
4716 sss_cl(this, _span, &_markBitMap, &_markStack,
4717 &pam_cl, before_count, CMSYield);
4718 dng->from()->object_iterate_careful(&sss_cl);
4719 dng->to()->object_iterate_careful(&sss_cl);
4720 }
4721 MarkRefsIntoAndScanClosure
4722 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4723 &_markStack, this, CMSYield,
4724 true /* precleaning phase */);
4725 // CAUTION: The following closure has persistent state that may need to
4726 // be reset upon a decrease in the sequence of addresses it
4727 // processes.
4728 ScanMarkedObjectsAgainCarefullyClosure
4729 smoac_cl(this, _span,
4730 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
4731
4732 // Preclean dirty cards in ModUnionTable and CardTable using
4733 // appropriate convergence criterion;
4734 // repeat CMSPrecleanIter times unless we find that
4735 // we are losing.
4736 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4737 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4738 "Bad convergence multiplier");
4739 assert(CMSPrecleanThreshold >= 100,
4740 "Unreasonably low CMSPrecleanThreshold");
4741
4742 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4743 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4744 numIter < CMSPrecleanIter;
4745 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4746 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4747 if (Verbose && PrintGCDetails) {
4748 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4749 }
4750 // Either there are very few dirty cards, so re-mark
4751 // pause will be small anyway, or our pre-cleaning isn't
4752 // that much faster than the rate at which cards are being
4753 // dirtied, so we might as well stop and re-mark since
4754 // precleaning won't improve our re-mark time by much.
4755 if (curNumCards <= CMSPrecleanThreshold ||
4756 (numIter > 0 &&
4757 (curNumCards * CMSPrecleanDenominator >
4758 lastNumCards * CMSPrecleanNumerator))) {
4759 numIter++;
4760 cumNumCards += curNumCards;
4761 break;
4762 }
4763 }
4764
4765 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4766
4767 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4768 cumNumCards += curNumCards;
4769 if (PrintGCDetails && PrintCMSStatistics != 0) {
4770 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4771 curNumCards, cumNumCards, numIter);
4772 }
4773 return cumNumCards; // as a measure of useful work done
4774 }
4775
4776 // PRECLEANING NOTES:
4777 // Precleaning involves:
4778 // . reading the bits of the modUnionTable and clearing the set bits.
4779 // . For the cards corresponding to the set bits, we scan the
4780 // objects on those cards. This means we need the free_list_lock
4781 // so that we can safely iterate over the CMS space when scanning
4782 // for oops.
4783 // . When we scan the objects, we'll be both reading and setting
4784 // marks in the marking bit map, so we'll need the marking bit map.
4785 // . For protecting _collector_state transitions, we take the CGC_lock.
4786 // Note that any races in the reading of of card table entries by the
4787 // CMS thread on the one hand and the clearing of those entries by the
4788 // VM thread or the setting of those entries by the mutator threads on the
4789 // other are quite benign. However, for efficiency it makes sense to keep
4790 // the VM thread from racing with the CMS thread while the latter is
4791 // dirty card info to the modUnionTable. We therefore also use the
4792 // CGC_lock to protect the reading of the card table and the mod union
4793 // table by the CM thread.
4794 // . We run concurrently with mutator updates, so scanning
4795 // needs to be done carefully -- we should not try to scan
4796 // potentially uninitialized objects.
4797 //
4798 // Locking strategy: While holding the CGC_lock, we scan over and
4799 // reset a maximal dirty range of the mod union / card tables, then lock
4800 // the free_list_lock and bitmap lock to do a full marking, then
4801 // release these locks; and repeat the cycle. This allows for a
4802 // certain amount of fairness in the sharing of these locks between
4803 // the CMS collector on the one hand, and the VM thread and the
4804 // mutators on the other.
4805
4806 // NOTE: preclean_mod_union_table() and preclean_card_table()
4807 // further below are largely identical; if you need to modify
4808 // one of these methods, please check the other method too.
4809
4810 size_t CMSCollector::preclean_mod_union_table(
4811 ConcurrentMarkSweepGeneration* gen,
4812 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4813 verify_work_stacks_empty();
4814 verify_overflow_empty();
4815
4816 // strategy: starting with the first card, accumulate contiguous
4817 // ranges of dirty cards; clear these cards, then scan the region
4818 // covered by these cards.
4819
4820 // Since all of the MUT is committed ahead, we can just use
4821 // that, in case the generations expand while we are precleaning.
4822 // It might also be fine to just use the committed part of the
4823 // generation, but we might potentially miss cards when the
4824 // generation is rapidly expanding while we are in the midst
4825 // of precleaning.
4826 HeapWord* startAddr = gen->reserved().start();
4827 HeapWord* endAddr = gen->reserved().end();
4828
4829 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4830
4831 size_t numDirtyCards, cumNumDirtyCards;
4832 HeapWord *nextAddr, *lastAddr;
4833 for (cumNumDirtyCards = numDirtyCards = 0,
4834 nextAddr = lastAddr = startAddr;
4835 nextAddr < endAddr;
4836 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4837
4838 ResourceMark rm;
4839 HandleMark hm;
4840
4841 MemRegion dirtyRegion;
4842 {
4843 stopTimer();
4844 // Potential yield point
4845 CMSTokenSync ts(true);
4846 startTimer();
4847 sample_eden();
4848 // Get dirty region starting at nextOffset (inclusive),
4849 // simultaneously clearing it.
4850 dirtyRegion =
4851 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4852 assert(dirtyRegion.start() >= nextAddr,
4853 "returned region inconsistent?");
4854 }
4855 // Remember where the next search should begin.
4856 // The returned region (if non-empty) is a right open interval,
4857 // so lastOffset is obtained from the right end of that
4858 // interval.
4859 lastAddr = dirtyRegion.end();
4860 // Should do something more transparent and less hacky XXX
4861 numDirtyCards =
4862 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4863
4864 // We'll scan the cards in the dirty region (with periodic
4865 // yields for foreground GC as needed).
4866 if (!dirtyRegion.is_empty()) {
4867 assert(numDirtyCards > 0, "consistency check");
4868 HeapWord* stop_point = NULL;
4869 stopTimer();
4870 // Potential yield point
4871 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4872 bitMapLock());
4873 startTimer();
4874 {
4875 verify_work_stacks_empty();
4876 verify_overflow_empty();
4877 sample_eden();
4878 stop_point =
4879 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4880 }
4881 if (stop_point != NULL) {
4882 // The careful iteration stopped early either because it found an
4883 // uninitialized object, or because we were in the midst of an
4884 // "abortable preclean", which should now be aborted. Redirty
4885 // the bits corresponding to the partially-scanned or unscanned
4886 // cards. We'll either restart at the next block boundary or
4887 // abort the preclean.
4888 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4889 "Should only be AbortablePreclean.");
4890 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4891 if (should_abort_preclean()) {
4892 break; // out of preclean loop
4893 } else {
4894 // Compute the next address at which preclean should pick up;
4895 // might need bitMapLock in order to read P-bits.
4896 lastAddr = next_card_start_after_block(stop_point);
4897 }
4898 }
4899 } else {
4900 assert(lastAddr == endAddr, "consistency check");
4901 assert(numDirtyCards == 0, "consistency check");
4902 break;
4903 }
4904 }
4905 verify_work_stacks_empty();
4906 verify_overflow_empty();
4907 return cumNumDirtyCards;
4908 }
4909
4910 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4911 // below are largely identical; if you need to modify
4912 // one of these methods, please check the other method too.
4913
4914 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4915 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4916 // strategy: it's similar to precleamModUnionTable above, in that
4917 // we accumulate contiguous ranges of dirty cards, mark these cards
4918 // precleaned, then scan the region covered by these cards.
4919 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4920 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4921
4922 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4923
4924 size_t numDirtyCards, cumNumDirtyCards;
4925 HeapWord *lastAddr, *nextAddr;
4926
4927 for (cumNumDirtyCards = numDirtyCards = 0,
4928 nextAddr = lastAddr = startAddr;
4929 nextAddr < endAddr;
4930 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4931
4932 ResourceMark rm;
4933 HandleMark hm;
4934
4935 MemRegion dirtyRegion;
4936 {
4937 // See comments in "Precleaning notes" above on why we
4938 // do this locking. XXX Could the locking overheads be
4939 // too high when dirty cards are sparse? [I don't think so.]
4940 stopTimer();
4941 CMSTokenSync x(true); // is cms thread
4942 startTimer();
4943 sample_eden();
4944 // Get and clear dirty region from card table
4945 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4946 MemRegion(nextAddr, endAddr),
4947 true,
4948 CardTableModRefBS::precleaned_card_val());
4949
4950 assert(dirtyRegion.start() >= nextAddr,
4951 "returned region inconsistent?");
4952 }
4953 lastAddr = dirtyRegion.end();
4954 numDirtyCards =
4955 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4956
4957 if (!dirtyRegion.is_empty()) {
4958 stopTimer();
4959 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4960 startTimer();
4961 sample_eden();
4962 verify_work_stacks_empty();
4963 verify_overflow_empty();
4964 HeapWord* stop_point =
4965 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4966 if (stop_point != NULL) {
4967 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4968 "Should only be AbortablePreclean.");
4969 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4970 if (should_abort_preclean()) {
4971 break; // out of preclean loop
4972 } else {
4973 // Compute the next address at which preclean should pick up.
4974 lastAddr = next_card_start_after_block(stop_point);
4975 }
4976 }
4977 } else {
4978 break;
4979 }
4980 }
4981 verify_work_stacks_empty();
4982 verify_overflow_empty();
4983 return cumNumDirtyCards;
4984 }
4985
4986 class PrecleanKlassClosure : public KlassClosure {
4987 CMKlassClosure _cm_klass_closure;
4988 public:
4989 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4990 void do_klass(Klass* k) {
4991 if (k->has_accumulated_modified_oops()) {
4992 k->clear_accumulated_modified_oops();
4993
4994 _cm_klass_closure.do_klass(k);
4995 }
4996 }
4997 };
4998
4999 // The freelist lock is needed to prevent asserts, is it really needed?
5000 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
5001
5002 cl->set_freelistLock(freelistLock);
5003
5004 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
5005
5006 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
5007 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
5008 PrecleanKlassClosure preclean_klass_closure(cl);
5009 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
5010
5011 verify_work_stacks_empty();
5012 verify_overflow_empty();
5013 }
5014
5015 void CMSCollector::checkpointRootsFinal(bool asynch,
5016 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5017 assert(_collectorState == FinalMarking, "incorrect state transition?");
5018 check_correct_thread_executing();
5019 // world is stopped at this checkpoint
5020 assert(SafepointSynchronize::is_at_safepoint(),
5021 "world should be stopped");
5022 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5023
5024 verify_work_stacks_empty();
5025 verify_overflow_empty();
5026
5027 SpecializationStats::clear();
5028 if (PrintGCDetails) {
5029 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
5030 _young_gen->used() / K,
5031 _young_gen->capacity() / K);
5032 }
5033 if (asynch) {
5034 if (CMSScavengeBeforeRemark) {
5035 GenCollectedHeap* gch = GenCollectedHeap::heap();
5036 // Temporarily set flag to false, GCH->do_collection will
5037 // expect it to be false and set to true
5038 FlagSetting fl(gch->_is_gc_active, false);
5039 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
5040 PrintGCDetails && Verbose, true, _gc_timer_cm);)
5041 int level = _cmsGen->level() - 1;
5042 if (level >= 0) {
5043 gch->do_collection(true, // full (i.e. force, see below)
5044 false, // !clear_all_soft_refs
5045 0, // size
5046 false, // is_tlab
5047 level // max_level
5048 );
5049 }
5050 }
5051 FreelistLocker x(this);
5052 MutexLockerEx y(bitMapLock(),
5053 Mutex::_no_safepoint_check_flag);
5054 assert(!init_mark_was_synchronous, "but that's impossible!");
5055 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
5056 } else {
5057 // already have all the locks
5058 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
5059 init_mark_was_synchronous);
5060 }
5061 verify_work_stacks_empty();
5062 verify_overflow_empty();
5063 SpecializationStats::print();
5064 }
5065
5066 void CMSCollector::checkpointRootsFinalWork(bool asynch,
5067 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5068
5069 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
5070
5071 assert(haveFreelistLocks(), "must have free list locks");
5072 assert_lock_strong(bitMapLock());
5073
5074 if (UseAdaptiveSizePolicy) {
5075 size_policy()->checkpoint_roots_final_begin();
5076 }
5077
5078 ResourceMark rm;
5079 HandleMark hm;
5080
5081 GenCollectedHeap* gch = GenCollectedHeap::heap();
5082
5083 if (should_unload_classes()) {
5084 CodeCache::gc_prologue();
5085 }
5086 assert(haveFreelistLocks(), "must have free list locks");
5087 assert_lock_strong(bitMapLock());
5088
5089 if (!init_mark_was_synchronous) {
5090 // We might assume that we need not fill TLAB's when
5091 // CMSScavengeBeforeRemark is set, because we may have just done
5092 // a scavenge which would have filled all TLAB's -- and besides
5093 // Eden would be empty. This however may not always be the case --
5094 // for instance although we asked for a scavenge, it may not have
5095 // happened because of a JNI critical section. We probably need
5096 // a policy for deciding whether we can in that case wait until
5097 // the critical section releases and then do the remark following
5098 // the scavenge, and skip it here. In the absence of that policy,
5099 // or of an indication of whether the scavenge did indeed occur,
5100 // we cannot rely on TLAB's having been filled and must do
5101 // so here just in case a scavenge did not happen.
5102 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
5103 // Update the saved marks which may affect the root scans.
5104 gch->save_marks();
5105
5106 if (CMSPrintEdenSurvivorChunks) {
5107 print_eden_and_survivor_chunk_arrays();
5108 }
5109
5110 {
5111 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
5112
5113 // Note on the role of the mod union table:
5114 // Since the marker in "markFromRoots" marks concurrently with
5115 // mutators, it is possible for some reachable objects not to have been
5116 // scanned. For instance, an only reference to an object A was
5117 // placed in object B after the marker scanned B. Unless B is rescanned,
5118 // A would be collected. Such updates to references in marked objects
5119 // are detected via the mod union table which is the set of all cards
5120 // dirtied since the first checkpoint in this GC cycle and prior to
5121 // the most recent young generation GC, minus those cleaned up by the
5122 // concurrent precleaning.
5123 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
5124 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
5125 do_remark_parallel();
5126 } else {
5127 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
5128 _gc_timer_cm);
5129 do_remark_non_parallel();
5130 }
5131 }
5132 } else {
5133 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
5134 // The initial mark was stop-world, so there's no rescanning to
5135 // do; go straight on to the next step below.
5136 }
5137 verify_work_stacks_empty();
5138 verify_overflow_empty();
5139
5140 {
5141 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
5142 refProcessingWork(asynch, clear_all_soft_refs);
5143 }
5144 verify_work_stacks_empty();
5145 verify_overflow_empty();
5146
5147 if (should_unload_classes()) {
5148 CodeCache::gc_epilogue();
5149 }
5150 JvmtiExport::gc_epilogue();
5151
5152 // If we encountered any (marking stack / work queue) overflow
5153 // events during the current CMS cycle, take appropriate
5154 // remedial measures, where possible, so as to try and avoid
5155 // recurrence of that condition.
5156 assert(_markStack.isEmpty(), "No grey objects");
5157 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5158 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5159 if (ser_ovflw > 0) {
5160 if (PrintCMSStatistics != 0) {
5161 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5162 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5163 ", kac_preclean="SIZE_FORMAT")",
5164 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5165 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5166 }
5167 _markStack.expand();
5168 _ser_pmc_remark_ovflw = 0;
5169 _ser_pmc_preclean_ovflw = 0;
5170 _ser_kac_preclean_ovflw = 0;
5171 _ser_kac_ovflw = 0;
5172 }
5173 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5174 if (PrintCMSStatistics != 0) {
5175 gclog_or_tty->print_cr("Work queue overflow (benign) "
5176 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5177 _par_pmc_remark_ovflw, _par_kac_ovflw);
5178 }
5179 _par_pmc_remark_ovflw = 0;
5180 _par_kac_ovflw = 0;
5181 }
5182 if (PrintCMSStatistics != 0) {
5183 if (_markStack._hit_limit > 0) {
5184 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5185 _markStack._hit_limit);
5186 }
5187 if (_markStack._failed_double > 0) {
5188 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5189 " current capacity "SIZE_FORMAT,
5190 _markStack._failed_double,
5191 _markStack.capacity());
5192 }
5193 }
5194 _markStack._hit_limit = 0;
5195 _markStack._failed_double = 0;
5196
5197 if ((VerifyAfterGC || VerifyDuringGC) &&
5198 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5199 verify_after_remark();
5200 }
5201
5202 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5203
5204 // Change under the freelistLocks.
5205 _collectorState = Sweeping;
5206 // Call isAllClear() under bitMapLock
5207 assert(_modUnionTable.isAllClear(),
5208 "Should be clear by end of the final marking");
5209 assert(_ct->klass_rem_set()->mod_union_is_clear(),
5210 "Should be clear by end of the final marking");
5211 if (UseAdaptiveSizePolicy) {
5212 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5213 }
5214 }
5215
5216 void CMSParInitialMarkTask::work(uint worker_id) {
5217 elapsedTimer _timer;
5218 ResourceMark rm;
5219 HandleMark hm;
5220
5221 // ---------- scan from roots --------------
5222 _timer.start();
5223 GenCollectedHeap* gch = GenCollectedHeap::heap();
5224 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
5225 CMKlassClosure klass_closure(&par_mri_cl);
5226
5227 // ---------- young gen roots --------------
5228 {
5229 work_on_young_gen_roots(worker_id, &par_mri_cl);
5230 _timer.stop();
5231 if (PrintCMSStatistics != 0) {
5232 gclog_or_tty->print_cr(
5233 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
5234 worker_id, _timer.seconds());
5235 }
5236 }
5237
5238 // ---------- remaining roots --------------
5239 _timer.reset();
5240 _timer.start();
5241 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5242 false, // yg was scanned above
5243 false, // this is parallel code
5244 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5245 &par_mri_cl,
5246 NULL,
5247 &klass_closure);
5248 assert(_collector->should_unload_classes()
5249 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache),
5250 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5251 _timer.stop();
5252 if (PrintCMSStatistics != 0) {
5253 gclog_or_tty->print_cr(
5254 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
5255 worker_id, _timer.seconds());
5256 }
5257 }
5258
5259 // Parallel remark task
5260 class CMSParRemarkTask: public CMSParMarkTask {
5261 CompactibleFreeListSpace* _cms_space;
5262
5263 // The per-thread work queues, available here for stealing.
5264 OopTaskQueueSet* _task_queues;
5265 ParallelTaskTerminator _term;
5266
5267 public:
5268 // A value of 0 passed to n_workers will cause the number of
5269 // workers to be taken from the active workers in the work gang.
5270 CMSParRemarkTask(CMSCollector* collector,
5271 CompactibleFreeListSpace* cms_space,
5272 int n_workers, FlexibleWorkGang* workers,
5273 OopTaskQueueSet* task_queues):
5274 CMSParMarkTask("Rescan roots and grey objects in parallel",
5275 collector, n_workers),
5276 _cms_space(cms_space),
5277 _task_queues(task_queues),
5278 _term(n_workers, task_queues) { }
5279
5280 OopTaskQueueSet* task_queues() { return _task_queues; }
5281
5282 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5283
5284 ParallelTaskTerminator* terminator() { return &_term; }
5285 int n_workers() { return _n_workers; }
5286
5287 void work(uint worker_id);
5288
5289 private:
5290 // ... of dirty cards in old space
5291 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5292 Par_MarkRefsIntoAndScanClosure* cl);
5293
5294 // ... work stealing for the above
5295 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5296 };
5297
5298 class RemarkKlassClosure : public KlassClosure {
5299 CMKlassClosure _cm_klass_closure;
5300 public:
5301 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
5302 void do_klass(Klass* k) {
5303 // Check if we have modified any oops in the Klass during the concurrent marking.
5304 if (k->has_accumulated_modified_oops()) {
5305 k->clear_accumulated_modified_oops();
5306
5307 // We could have transfered the current modified marks to the accumulated marks,
5308 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
5309 } else if (k->has_modified_oops()) {
5310 // Don't clear anything, this info is needed by the next young collection.
5311 } else {
5312 // No modified oops in the Klass.
5313 return;
5314 }
5315
5316 // The klass has modified fields, need to scan the klass.
5317 _cm_klass_closure.do_klass(k);
5318 }
5319 };
5320
5321 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
5322 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5323 EdenSpace* eden_space = dng->eden();
5324 ContiguousSpace* from_space = dng->from();
5325 ContiguousSpace* to_space = dng->to();
5326
5327 HeapWord** eca = _collector->_eden_chunk_array;
5328 size_t ect = _collector->_eden_chunk_index;
5329 HeapWord** sca = _collector->_survivor_chunk_array;
5330 size_t sct = _collector->_survivor_chunk_index;
5331
5332 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5333 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5334
5335 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
5336 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
5337 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
5338 }
5339
5340 // work_queue(i) is passed to the closure
5341 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5342 // also is passed to do_dirty_card_rescan_tasks() and to
5343 // do_work_steal() to select the i-th task_queue.
5344
5345 void CMSParRemarkTask::work(uint worker_id) {
5346 elapsedTimer _timer;
5347 ResourceMark rm;
5348 HandleMark hm;
5349
5350 // ---------- rescan from roots --------------
5351 _timer.start();
5352 GenCollectedHeap* gch = GenCollectedHeap::heap();
5353 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5354 _collector->_span, _collector->ref_processor(),
5355 &(_collector->_markBitMap),
5356 work_queue(worker_id));
5357
5358 // Rescan young gen roots first since these are likely
5359 // coarsely partitioned and may, on that account, constitute
5360 // the critical path; thus, it's best to start off that
5361 // work first.
5362 // ---------- young gen roots --------------
5363 {
5364 work_on_young_gen_roots(worker_id, &par_mrias_cl);
5365 _timer.stop();
5366 if (PrintCMSStatistics != 0) {
5367 gclog_or_tty->print_cr(
5368 "Finished young gen rescan work in %dth thread: %3.3f sec",
5369 worker_id, _timer.seconds());
5370 }
5371 }
5372
5373 // ---------- remaining roots --------------
5374 _timer.reset();
5375 _timer.start();
5376 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5377 false, // yg was scanned above
5378 false, // this is parallel code
5379 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5380 &par_mrias_cl,
5381 NULL,
5382 NULL); // The dirty klasses will be handled below
5383 assert(_collector->should_unload_classes()
5384 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache),
5385 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5386 _timer.stop();
5387 if (PrintCMSStatistics != 0) {
5388 gclog_or_tty->print_cr(
5389 "Finished remaining root rescan work in %dth thread: %3.3f sec",
5390 worker_id, _timer.seconds());
5391 }
5392
5393 // ---------- unhandled CLD scanning ----------
5394 if (worker_id == 0) { // Single threaded at the moment.
5395 _timer.reset();
5396 _timer.start();
5397
5398 // Scan all new class loader data objects and new dependencies that were
5399 // introduced during concurrent marking.
5400 ResourceMark rm;
5401 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5402 for (int i = 0; i < array->length(); i++) {
5403 par_mrias_cl.do_class_loader_data(array->at(i));
5404 }
5405
5406 // We don't need to keep track of new CLDs anymore.
5407 ClassLoaderDataGraph::remember_new_clds(false);
5408
5409 _timer.stop();
5410 if (PrintCMSStatistics != 0) {
5411 gclog_or_tty->print_cr(
5412 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec",
5413 worker_id, _timer.seconds());
5414 }
5415 }
5416
5417 // ---------- dirty klass scanning ----------
5418 if (worker_id == 0) { // Single threaded at the moment.
5419 _timer.reset();
5420 _timer.start();
5421
5422 // Scan all classes that was dirtied during the concurrent marking phase.
5423 RemarkKlassClosure remark_klass_closure(&par_mrias_cl);
5424 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
5425
5426 _timer.stop();
5427 if (PrintCMSStatistics != 0) {
5428 gclog_or_tty->print_cr(
5429 "Finished dirty klass scanning work in %dth thread: %3.3f sec",
5430 worker_id, _timer.seconds());
5431 }
5432 }
5433
5434 // We might have added oops to ClassLoaderData::_handles during the
5435 // concurrent marking phase. These oops point to newly allocated objects
5436 // that are guaranteed to be kept alive. Either by the direct allocation
5437 // code, or when the young collector processes the strong roots. Hence,
5438 // we don't have to revisit the _handles block during the remark phase.
5439
5440 // ---------- rescan dirty cards ------------
5441 _timer.reset();
5442 _timer.start();
5443
5444 // Do the rescan tasks for each of the two spaces
5445 // (cms_space) in turn.
5446 // "worker_id" is passed to select the task_queue for "worker_id"
5447 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
5448 _timer.stop();
5449 if (PrintCMSStatistics != 0) {
5450 gclog_or_tty->print_cr(
5451 "Finished dirty card rescan work in %dth thread: %3.3f sec",
5452 worker_id, _timer.seconds());
5453 }
5454
5455 // ---------- steal work from other threads ...
5456 // ---------- ... and drain overflow list.
5457 _timer.reset();
5458 _timer.start();
5459 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
5460 _timer.stop();
5461 if (PrintCMSStatistics != 0) {
5462 gclog_or_tty->print_cr(
5463 "Finished work stealing in %dth thread: %3.3f sec",
5464 worker_id, _timer.seconds());
5465 }
5466 }
5467
5468 // Note that parameter "i" is not used.
5469 void
5470 CMSParMarkTask::do_young_space_rescan(uint worker_id,
5471 OopsInGenClosure* cl, ContiguousSpace* space,
5472 HeapWord** chunk_array, size_t chunk_top) {
5473 // Until all tasks completed:
5474 // . claim an unclaimed task
5475 // . compute region boundaries corresponding to task claimed
5476 // using chunk_array
5477 // . par_oop_iterate(cl) over that region
5478
5479 ResourceMark rm;
5480 HandleMark hm;
5481
5482 SequentialSubTasksDone* pst = space->par_seq_tasks();
5483
5484 uint nth_task = 0;
5485 uint n_tasks = pst->n_tasks();
5486
5487 if (n_tasks > 0) {
5488 assert(pst->valid(), "Uninitialized use?");
5489 HeapWord *start, *end;
5490 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5491 // We claimed task # nth_task; compute its boundaries.
5492 if (chunk_top == 0) { // no samples were taken
5493 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task");
5494 start = space->bottom();
5495 end = space->top();
5496 } else if (nth_task == 0) {
5497 start = space->bottom();
5498 end = chunk_array[nth_task];
5499 } else if (nth_task < (uint)chunk_top) {
5500 assert(nth_task >= 1, "Control point invariant");
5501 start = chunk_array[nth_task - 1];
5502 end = chunk_array[nth_task];
5503 } else {
5504 assert(nth_task == (uint)chunk_top, "Control point invariant");
5505 start = chunk_array[chunk_top - 1];
5506 end = space->top();
5507 }
5508 MemRegion mr(start, end);
5509 // Verify that mr is in space
5510 assert(mr.is_empty() || space->used_region().contains(mr),
5511 "Should be in space");
5512 // Verify that "start" is an object boundary
5513 assert(mr.is_empty() || oop(mr.start())->is_oop(),
5514 "Should be an oop");
5515 space->par_oop_iterate(mr, cl);
5516 }
5517 pst->all_tasks_completed();
5518 }
5519 }
5520
5521 void
5522 CMSParRemarkTask::do_dirty_card_rescan_tasks(
5523 CompactibleFreeListSpace* sp, int i,
5524 Par_MarkRefsIntoAndScanClosure* cl) {
5525 // Until all tasks completed:
5526 // . claim an unclaimed task
5527 // . compute region boundaries corresponding to task claimed
5528 // . transfer dirty bits ct->mut for that region
5529 // . apply rescanclosure to dirty mut bits for that region
5530
5531 ResourceMark rm;
5532 HandleMark hm;
5533
5534 OopTaskQueue* work_q = work_queue(i);
5535 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
5536 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
5537 // CAUTION: This closure has state that persists across calls to
5538 // the work method dirty_range_iterate_clear() in that it has
5539 // embedded in it a (subtype of) UpwardsObjectClosure. The
5540 // use of that state in the embedded UpwardsObjectClosure instance
5541 // assumes that the cards are always iterated (even if in parallel
5542 // by several threads) in monotonically increasing order per each
5543 // thread. This is true of the implementation below which picks
5544 // card ranges (chunks) in monotonically increasing order globally
5545 // and, a-fortiori, in monotonically increasing order per thread
5546 // (the latter order being a subsequence of the former).
5547 // If the work code below is ever reorganized into a more chaotic
5548 // work-partitioning form than the current "sequential tasks"
5549 // paradigm, the use of that persistent state will have to be
5550 // revisited and modified appropriately. See also related
5551 // bug 4756801 work on which should examine this code to make
5552 // sure that the changes there do not run counter to the
5553 // assumptions made here and necessary for correctness and
5554 // efficiency. Note also that this code might yield inefficient
5555 // behavior in the case of very large objects that span one or
5556 // more work chunks. Such objects would potentially be scanned
5557 // several times redundantly. Work on 4756801 should try and
5558 // address that performance anomaly if at all possible. XXX
5559 MemRegion full_span = _collector->_span;
5560 CMSBitMap* bm = &(_collector->_markBitMap); // shared
5561 MarkFromDirtyCardsClosure
5562 greyRescanClosure(_collector, full_span, // entire span of interest
5563 sp, bm, work_q, cl);
5564
5565 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
5566 assert(pst->valid(), "Uninitialized use?");
5567 uint nth_task = 0;
5568 const int alignment = CardTableModRefBS::card_size * BitsPerWord;
5569 MemRegion span = sp->used_region();
5570 HeapWord* start_addr = span.start();
5571 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(),
5572 alignment);
5573 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
5574 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) ==
5575 start_addr, "Check alignment");
5576 assert((size_t)round_to((intptr_t)chunk_size, alignment) ==
5577 chunk_size, "Check alignment");
5578
5579 while (!pst->is_task_claimed(/* reference */ nth_task)) {
5580 // Having claimed the nth_task, compute corresponding mem-region,
5581 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
5582 // The alignment restriction ensures that we do not need any
5583 // synchronization with other gang-workers while setting or
5584 // clearing bits in thus chunk of the MUT.
5585 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
5586 start_addr + (nth_task+1)*chunk_size);
5587 // The last chunk's end might be way beyond end of the
5588 // used region. In that case pull back appropriately.
5589 if (this_span.end() > end_addr) {
5590 this_span.set_end(end_addr);
5591 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
5592 }
5593 // Iterate over the dirty cards covering this chunk, marking them
5594 // precleaned, and setting the corresponding bits in the mod union
5595 // table. Since we have been careful to partition at Card and MUT-word
5596 // boundaries no synchronization is needed between parallel threads.
5597 _collector->_ct->ct_bs()->dirty_card_iterate(this_span,
5598 &modUnionClosure);
5599
5600 // Having transferred these marks into the modUnionTable,
5601 // rescan the marked objects on the dirty cards in the modUnionTable.
5602 // Even if this is at a synchronous collection, the initial marking
5603 // may have been done during an asynchronous collection so there
5604 // may be dirty bits in the mod-union table.
5605 _collector->_modUnionTable.dirty_range_iterate_clear(
5606 this_span, &greyRescanClosure);
5607 _collector->_modUnionTable.verifyNoOneBitsInRange(
5608 this_span.start(),
5609 this_span.end());
5610 }
5611 pst->all_tasks_completed(); // declare that i am done
5612 }
5613
5614 // . see if we can share work_queues with ParNew? XXX
5615 void
5616 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
5617 int* seed) {
5618 OopTaskQueue* work_q = work_queue(i);
5619 NOT_PRODUCT(int num_steals = 0;)
5620 oop obj_to_scan;
5621 CMSBitMap* bm = &(_collector->_markBitMap);
5622
5623 while (true) {
5624 // Completely finish any left over work from (an) earlier round(s)
5625 cl->trim_queue(0);
5626 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5627 (size_t)ParGCDesiredObjsFromOverflowList);
5628 // Now check if there's any work in the overflow list
5629 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5630 // only affects the number of attempts made to get work from the
5631 // overflow list and does not affect the number of workers. Just
5632 // pass ParallelGCThreads so this behavior is unchanged.
5633 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5634 work_q,
5635 ParallelGCThreads)) {
5636 // found something in global overflow list;
5637 // not yet ready to go stealing work from others.
5638 // We'd like to assert(work_q->size() != 0, ...)
5639 // because we just took work from the overflow list,
5640 // but of course we can't since all of that could have
5641 // been already stolen from us.
5642 // "He giveth and He taketh away."
5643 continue;
5644 }
5645 // Verify that we have no work before we resort to stealing
5646 assert(work_q->size() == 0, "Have work, shouldn't steal");
5647 // Try to steal from other queues that have work
5648 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5649 NOT_PRODUCT(num_steals++;)
5650 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
5651 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5652 // Do scanning work
5653 obj_to_scan->oop_iterate(cl);
5654 // Loop around, finish this work, and try to steal some more
5655 } else if (terminator()->offer_termination()) {
5656 break; // nirvana from the infinite cycle
5657 }
5658 }
5659 NOT_PRODUCT(
5660 if (PrintCMSStatistics != 0) {
5661 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
5662 }
5663 )
5664 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
5665 "Else our work is not yet done");
5666 }
5667
5668 // Record object boundaries in _eden_chunk_array by sampling the eden
5669 // top in the slow-path eden object allocation code path and record
5670 // the boundaries, if CMSEdenChunksRecordAlways is true. If
5671 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
5672 // sampling in sample_eden() that activates during the part of the
5673 // preclean phase.
5674 void CMSCollector::sample_eden_chunk() {
5675 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
5676 if (_eden_chunk_lock->try_lock()) {
5677 // Record a sample. This is the critical section. The contents
5678 // of the _eden_chunk_array have to be non-decreasing in the
5679 // address order.
5680 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
5681 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
5682 "Unexpected state of Eden");
5683 if (_eden_chunk_index == 0 ||
5684 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
5685 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
5686 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
5687 _eden_chunk_index++; // commit sample
5688 }
5689 _eden_chunk_lock->unlock();
5690 }
5691 }
5692 }
5693
5694 // Return a thread-local PLAB recording array, as appropriate.
5695 void* CMSCollector::get_data_recorder(int thr_num) {
5696 if (_survivor_plab_array != NULL &&
5697 (CMSPLABRecordAlways ||
5698 (_collectorState > Marking && _collectorState < FinalMarking))) {
5699 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
5700 ChunkArray* ca = &_survivor_plab_array[thr_num];
5701 ca->reset(); // clear it so that fresh data is recorded
5702 return (void*) ca;
5703 } else {
5704 return NULL;
5705 }
5706 }
5707
5708 // Reset all the thread-local PLAB recording arrays
5709 void CMSCollector::reset_survivor_plab_arrays() {
5710 for (uint i = 0; i < ParallelGCThreads; i++) {
5711 _survivor_plab_array[i].reset();
5712 }
5713 }
5714
5715 // Merge the per-thread plab arrays into the global survivor chunk
5716 // array which will provide the partitioning of the survivor space
5717 // for CMS initial scan and rescan.
5718 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
5719 int no_of_gc_threads) {
5720 assert(_survivor_plab_array != NULL, "Error");
5721 assert(_survivor_chunk_array != NULL, "Error");
5722 assert(_collectorState == FinalMarking ||
5723 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
5724 for (int j = 0; j < no_of_gc_threads; j++) {
5725 _cursor[j] = 0;
5726 }
5727 HeapWord* top = surv->top();
5728 size_t i;
5729 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
5730 HeapWord* min_val = top; // Higher than any PLAB address
5731 uint min_tid = 0; // position of min_val this round
5732 for (int j = 0; j < no_of_gc_threads; j++) {
5733 ChunkArray* cur_sca = &_survivor_plab_array[j];
5734 if (_cursor[j] == cur_sca->end()) {
5735 continue;
5736 }
5737 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
5738 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
5739 assert(surv->used_region().contains(cur_val), "Out of bounds value");
5740 if (cur_val < min_val) {
5741 min_tid = j;
5742 min_val = cur_val;
5743 } else {
5744 assert(cur_val < top, "All recorded addresses should be less");
5745 }
5746 }
5747 // At this point min_val and min_tid are respectively
5748 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
5749 // and the thread (j) that witnesses that address.
5750 // We record this address in the _survivor_chunk_array[i]
5751 // and increment _cursor[min_tid] prior to the next round i.
5752 if (min_val == top) {
5753 break;
5754 }
5755 _survivor_chunk_array[i] = min_val;
5756 _cursor[min_tid]++;
5757 }
5758 // We are all done; record the size of the _survivor_chunk_array
5759 _survivor_chunk_index = i; // exclusive: [0, i)
5760 if (PrintCMSStatistics > 0) {
5761 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i);
5762 }
5763 // Verify that we used up all the recorded entries
5764 #ifdef ASSERT
5765 size_t total = 0;
5766 for (int j = 0; j < no_of_gc_threads; j++) {
5767 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
5768 total += _cursor[j];
5769 }
5770 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
5771 // Check that the merged array is in sorted order
5772 if (total > 0) {
5773 for (size_t i = 0; i < total - 1; i++) {
5774 if (PrintCMSStatistics > 0) {
5775 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
5776 i, _survivor_chunk_array[i]);
5777 }
5778 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
5779 "Not sorted");
5780 }
5781 }
5782 #endif // ASSERT
5783 }
5784
5785 // Set up the space's par_seq_tasks structure for work claiming
5786 // for parallel initial scan and rescan of young gen.
5787 // See ParRescanTask where this is currently used.
5788 void
5789 CMSCollector::
5790 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
5791 assert(n_threads > 0, "Unexpected n_threads argument");
5792 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
5793
5794 // Eden space
5795 if (!dng->eden()->is_empty()) {
5796 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks();
5797 assert(!pst->valid(), "Clobbering existing data?");
5798 // Each valid entry in [0, _eden_chunk_index) represents a task.
5799 size_t n_tasks = _eden_chunk_index + 1;
5800 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
5801 // Sets the condition for completion of the subtask (how many threads
5802 // need to finish in order to be done).
5803 pst->set_n_threads(n_threads);
5804 pst->set_n_tasks((int)n_tasks);
5805 }
5806
5807 // Merge the survivor plab arrays into _survivor_chunk_array
5808 if (_survivor_plab_array != NULL) {
5809 merge_survivor_plab_arrays(dng->from(), n_threads);
5810 } else {
5811 assert(_survivor_chunk_index == 0, "Error");
5812 }
5813
5814 // To space
5815 {
5816 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
5817 assert(!pst->valid(), "Clobbering existing data?");
5818 // Sets the condition for completion of the subtask (how many threads
5819 // need to finish in order to be done).
5820 pst->set_n_threads(n_threads);
5821 pst->set_n_tasks(1);
5822 assert(pst->valid(), "Error");
5823 }
5824
5825 // From space
5826 {
5827 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks();
5828 assert(!pst->valid(), "Clobbering existing data?");
5829 size_t n_tasks = _survivor_chunk_index + 1;
5830 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
5831 // Sets the condition for completion of the subtask (how many threads
5832 // need to finish in order to be done).
5833 pst->set_n_threads(n_threads);
5834 pst->set_n_tasks((int)n_tasks);
5835 assert(pst->valid(), "Error");
5836 }
5837 }
5838
5839 // Parallel version of remark
5840 void CMSCollector::do_remark_parallel() {
5841 GenCollectedHeap* gch = GenCollectedHeap::heap();
5842 FlexibleWorkGang* workers = gch->workers();
5843 assert(workers != NULL, "Need parallel worker threads.");
5844 // Choose to use the number of GC workers most recently set
5845 // into "active_workers". If active_workers is not set, set it
5846 // to ParallelGCThreads.
5847 int n_workers = workers->active_workers();
5848 if (n_workers == 0) {
5849 assert(n_workers > 0, "Should have been set during scavenge");
5850 n_workers = ParallelGCThreads;
5851 workers->set_active_workers(n_workers);
5852 }
5853 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
5854
5855 CMSParRemarkTask tsk(this,
5856 cms_space,
5857 n_workers, workers, task_queues());
5858
5859 // Set up for parallel process_strong_roots work.
5860 gch->set_par_threads(n_workers);
5861 // We won't be iterating over the cards in the card table updating
5862 // the younger_gen cards, so we shouldn't call the following else
5863 // the verification code as well as subsequent younger_refs_iterate
5864 // code would get confused. XXX
5865 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
5866
5867 // The young gen rescan work will not be done as part of
5868 // process_strong_roots (which currently doesn't knw how to
5869 // parallelize such a scan), but rather will be broken up into
5870 // a set of parallel tasks (via the sampling that the [abortable]
5871 // preclean phase did of EdenSpace, plus the [two] tasks of
5872 // scanning the [two] survivor spaces. Further fine-grain
5873 // parallelization of the scanning of the survivor spaces
5874 // themselves, and of precleaning of the younger gen itself
5875 // is deferred to the future.
5876 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
5877
5878 // The dirty card rescan work is broken up into a "sequence"
5879 // of parallel tasks (per constituent space) that are dynamically
5880 // claimed by the parallel threads.
5881 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
5882
5883 // It turns out that even when we're using 1 thread, doing the work in a
5884 // separate thread causes wide variance in run times. We can't help this
5885 // in the multi-threaded case, but we special-case n=1 here to get
5886 // repeatable measurements of the 1-thread overhead of the parallel code.
5887 if (n_workers > 1) {
5888 // Make refs discovery MT-safe, if it isn't already: it may not
5889 // necessarily be so, since it's possible that we are doing
5890 // ST marking.
5891 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
5892 GenCollectedHeap::StrongRootsScope srs(gch);
5893 workers->run_task(&tsk);
5894 } else {
5895 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5896 GenCollectedHeap::StrongRootsScope srs(gch);
5897 tsk.work(0);
5898 }
5899
5900 gch->set_par_threads(0); // 0 ==> non-parallel.
5901 // restore, single-threaded for now, any preserved marks
5902 // as a result of work_q overflow
5903 restore_preserved_marks_if_any();
5904 }
5905
5906 // Non-parallel version of remark
5907 void CMSCollector::do_remark_non_parallel() {
5908 ResourceMark rm;
5909 HandleMark hm;
5910 GenCollectedHeap* gch = GenCollectedHeap::heap();
5911 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
5912
5913 MarkRefsIntoAndScanClosure
5914 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
5915 &_markStack, this,
5916 false /* should_yield */, false /* not precleaning */);
5917 MarkFromDirtyCardsClosure
5918 markFromDirtyCardsClosure(this, _span,
5919 NULL, // space is set further below
5920 &_markBitMap, &_markStack, &mrias_cl);
5921 {
5922 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm);
5923 // Iterate over the dirty cards, setting the corresponding bits in the
5924 // mod union table.
5925 {
5926 ModUnionClosure modUnionClosure(&_modUnionTable);
5927 _ct->ct_bs()->dirty_card_iterate(
5928 _cmsGen->used_region(),
5929 &modUnionClosure);
5930 }
5931 // Having transferred these marks into the modUnionTable, we just need
5932 // to rescan the marked objects on the dirty cards in the modUnionTable.
5933 // The initial marking may have been done during an asynchronous
5934 // collection so there may be dirty bits in the mod-union table.
5935 const int alignment =
5936 CardTableModRefBS::card_size * BitsPerWord;
5937 {
5938 // ... First handle dirty cards in CMS gen
5939 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
5940 MemRegion ur = _cmsGen->used_region();
5941 HeapWord* lb = ur.start();
5942 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment);
5943 MemRegion cms_span(lb, ub);
5944 _modUnionTable.dirty_range_iterate_clear(cms_span,
5945 &markFromDirtyCardsClosure);
5946 verify_work_stacks_empty();
5947 if (PrintCMSStatistics != 0) {
5948 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ",
5949 markFromDirtyCardsClosure.num_dirty_cards());
5950 }
5951 }
5952 }
5953 if (VerifyDuringGC &&
5954 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5955 HandleMark hm; // Discard invalid handles created during verification
5956 Universe::verify();
5957 }
5958 {
5959 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm);
5960
5961 verify_work_stacks_empty();
5962
5963 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
5964 GenCollectedHeap::StrongRootsScope srs(gch);
5965 gch->gen_process_strong_roots(_cmsGen->level(),
5966 true, // younger gens as roots
5967 false, // use the local StrongRootsScope
5968 SharedHeap::ScanningOption(roots_scanning_options()),
5969 &mrias_cl,
5970 NULL,
5971 NULL); // The dirty klasses will be handled below
5972
5973 assert(should_unload_classes()
5974 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache),
5975 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5976 }
5977
5978 {
5979 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm);
5980
5981 verify_work_stacks_empty();
5982
5983 // Scan all class loader data objects that might have been introduced
5984 // during concurrent marking.
5985 ResourceMark rm;
5986 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5987 for (int i = 0; i < array->length(); i++) {
5988 mrias_cl.do_class_loader_data(array->at(i));
5989 }
5990
5991 // We don't need to keep track of new CLDs anymore.
5992 ClassLoaderDataGraph::remember_new_clds(false);
5993
5994 verify_work_stacks_empty();
5995 }
5996
5997 {
5998 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm);
5999
6000 verify_work_stacks_empty();
6001
6002 RemarkKlassClosure remark_klass_closure(&mrias_cl);
6003 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
6004
6005 verify_work_stacks_empty();
6006 }
6007
6008 // We might have added oops to ClassLoaderData::_handles during the
6009 // concurrent marking phase. These oops point to newly allocated objects
6010 // that are guaranteed to be kept alive. Either by the direct allocation
6011 // code, or when the young collector processes the strong roots. Hence,
6012 // we don't have to revisit the _handles block during the remark phase.
6013
6014 verify_work_stacks_empty();
6015 // Restore evacuated mark words, if any, used for overflow list links
6016 if (!CMSOverflowEarlyRestoration) {
6017 restore_preserved_marks_if_any();
6018 }
6019 verify_overflow_empty();
6020 }
6021
6022 ////////////////////////////////////////////////////////
6023 // Parallel Reference Processing Task Proxy Class
6024 ////////////////////////////////////////////////////////
6025 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
6026 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
6027 CMSCollector* _collector;
6028 CMSBitMap* _mark_bit_map;
6029 const MemRegion _span;
6030 ProcessTask& _task;
6031
6032 public:
6033 CMSRefProcTaskProxy(ProcessTask& task,
6034 CMSCollector* collector,
6035 const MemRegion& span,
6036 CMSBitMap* mark_bit_map,
6037 AbstractWorkGang* workers,
6038 OopTaskQueueSet* task_queues):
6039 // XXX Should superclass AGTWOQ also know about AWG since it knows
6040 // about the task_queues used by the AWG? Then it could initialize
6041 // the terminator() object. See 6984287. The set_for_termination()
6042 // below is a temporary band-aid for the regression in 6984287.
6043 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
6044 task_queues),
6045 _task(task),
6046 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
6047 {
6048 assert(_collector->_span.equals(_span) && !_span.is_empty(),
6049 "Inconsistency in _span");
6050 set_for_termination(workers->active_workers());
6051 }
6052
6053 OopTaskQueueSet* task_queues() { return queues(); }
6054
6055 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
6056
6057 void do_work_steal(int i,
6058 CMSParDrainMarkingStackClosure* drain,
6059 CMSParKeepAliveClosure* keep_alive,
6060 int* seed);
6061
6062 virtual void work(uint worker_id);
6063 };
6064
6065 void CMSRefProcTaskProxy::work(uint worker_id) {
6066 assert(_collector->_span.equals(_span), "Inconsistency in _span");
6067 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
6068 _mark_bit_map,
6069 work_queue(worker_id));
6070 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
6071 _mark_bit_map,
6072 work_queue(worker_id));
6073 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
6074 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
6075 if (_task.marks_oops_alive()) {
6076 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
6077 _collector->hash_seed(worker_id));
6078 }
6079 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
6080 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
6081 }
6082
6083 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
6084 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
6085 EnqueueTask& _task;
6086
6087 public:
6088 CMSRefEnqueueTaskProxy(EnqueueTask& task)
6089 : AbstractGangTask("Enqueue reference objects in parallel"),
6090 _task(task)
6091 { }
6092
6093 virtual void work(uint worker_id)
6094 {
6095 _task.work(worker_id);
6096 }
6097 };
6098
6099 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
6100 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
6101 _span(span),
6102 _bit_map(bit_map),
6103 _work_queue(work_queue),
6104 _mark_and_push(collector, span, bit_map, work_queue),
6105 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6106 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
6107 { }
6108
6109 // . see if we can share work_queues with ParNew? XXX
6110 void CMSRefProcTaskProxy::do_work_steal(int i,
6111 CMSParDrainMarkingStackClosure* drain,
6112 CMSParKeepAliveClosure* keep_alive,
6113 int* seed) {
6114 OopTaskQueue* work_q = work_queue(i);
6115 NOT_PRODUCT(int num_steals = 0;)
6116 oop obj_to_scan;
6117
6118 while (true) {
6119 // Completely finish any left over work from (an) earlier round(s)
6120 drain->trim_queue(0);
6121 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
6122 (size_t)ParGCDesiredObjsFromOverflowList);
6123 // Now check if there's any work in the overflow list
6124 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
6125 // only affects the number of attempts made to get work from the
6126 // overflow list and does not affect the number of workers. Just
6127 // pass ParallelGCThreads so this behavior is unchanged.
6128 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
6129 work_q,
6130 ParallelGCThreads)) {
6131 // Found something in global overflow list;
6132 // not yet ready to go stealing work from others.
6133 // We'd like to assert(work_q->size() != 0, ...)
6134 // because we just took work from the overflow list,
6135 // but of course we can't, since all of that might have
6136 // been already stolen from us.
6137 continue;
6138 }
6139 // Verify that we have no work before we resort to stealing
6140 assert(work_q->size() == 0, "Have work, shouldn't steal");
6141 // Try to steal from other queues that have work
6142 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
6143 NOT_PRODUCT(num_steals++;)
6144 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
6145 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
6146 // Do scanning work
6147 obj_to_scan->oop_iterate(keep_alive);
6148 // Loop around, finish this work, and try to steal some more
6149 } else if (terminator()->offer_termination()) {
6150 break; // nirvana from the infinite cycle
6151 }
6152 }
6153 NOT_PRODUCT(
6154 if (PrintCMSStatistics != 0) {
6155 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
6156 }
6157 )
6158 }
6159
6160 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
6161 {
6162 GenCollectedHeap* gch = GenCollectedHeap::heap();
6163 FlexibleWorkGang* workers = gch->workers();
6164 assert(workers != NULL, "Need parallel worker threads.");
6165 CMSRefProcTaskProxy rp_task(task, &_collector,
6166 _collector.ref_processor()->span(),
6167 _collector.markBitMap(),
6168 workers, _collector.task_queues());
6169 workers->run_task(&rp_task);
6170 }
6171
6172 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
6173 {
6174
6175 GenCollectedHeap* gch = GenCollectedHeap::heap();
6176 FlexibleWorkGang* workers = gch->workers();
6177 assert(workers != NULL, "Need parallel worker threads.");
6178 CMSRefEnqueueTaskProxy enq_task(task);
6179 workers->run_task(&enq_task);
6180 }
6181
6182 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
6183
6184 ResourceMark rm;
6185 HandleMark hm;
6186
6187 ReferenceProcessor* rp = ref_processor();
6188 assert(rp->span().equals(_span), "Spans should be equal");
6189 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
6190 // Process weak references.
6191 rp->setup_policy(clear_all_soft_refs);
6192 verify_work_stacks_empty();
6193
6194 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
6195 &_markStack, false /* !preclean */);
6196 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
6197 _span, &_markBitMap, &_markStack,
6198 &cmsKeepAliveClosure, false /* !preclean */);
6199 {
6200 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
6201
6202 ReferenceProcessorStats stats;
6203 if (rp->processing_is_mt()) {
6204 // Set the degree of MT here. If the discovery is done MT, there
6205 // may have been a different number of threads doing the discovery
6206 // and a different number of discovered lists may have Ref objects.
6207 // That is OK as long as the Reference lists are balanced (see
6208 // balance_all_queues() and balance_queues()).
6209 GenCollectedHeap* gch = GenCollectedHeap::heap();
6210 int active_workers = ParallelGCThreads;
6211 FlexibleWorkGang* workers = gch->workers();
6212 if (workers != NULL) {
6213 active_workers = workers->active_workers();
6214 // The expectation is that active_workers will have already
6215 // been set to a reasonable value. If it has not been set,
6216 // investigate.
6217 assert(active_workers > 0, "Should have been set during scavenge");
6218 }
6219 rp->set_active_mt_degree(active_workers);
6220 CMSRefProcTaskExecutor task_executor(*this);
6221 stats = rp->process_discovered_references(&_is_alive_closure,
6222 &cmsKeepAliveClosure,
6223 &cmsDrainMarkingStackClosure,
6224 &task_executor,
6225 _gc_timer_cm);
6226 } else {
6227 stats = rp->process_discovered_references(&_is_alive_closure,
6228 &cmsKeepAliveClosure,
6229 &cmsDrainMarkingStackClosure,
6230 NULL,
6231 _gc_timer_cm);
6232 }
6233 _gc_tracer_cm->report_gc_reference_stats(stats);
6234
6235 }
6236
6237 // This is the point where the entire marking should have completed.
6238 verify_work_stacks_empty();
6239
6240 if (should_unload_classes()) {
6241 {
6242 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
6243
6244 // Unload classes and purge the SystemDictionary.
6245 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
6246
6247 // Unload nmethods.
6248 CodeCache::do_unloading(&_is_alive_closure, purged_class);
6249
6250 // Prune dead klasses from subklass/sibling/implementor lists.
6251 Klass::clean_weak_klass_links(&_is_alive_closure);
6252 }
6253
6254 {
6255 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
6256 // Clean up unreferenced symbols in symbol table.
6257 SymbolTable::unlink();
6258 }
6259 }
6260
6261 // CMS doesn't use the StringTable as hard roots when class unloading is turned off.
6262 // Need to check if we really scanned the StringTable.
6263 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) {
6264 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
6265 // Delete entries for dead interned strings.
6266 StringTable::unlink(&_is_alive_closure);
6267 }
6268
6269 // Restore any preserved marks as a result of mark stack or
6270 // work queue overflow
6271 restore_preserved_marks_if_any(); // done single-threaded for now
6272
6273 rp->set_enqueuing_is_done(true);
6274 if (rp->processing_is_mt()) {
6275 rp->balance_all_queues();
6276 CMSRefProcTaskExecutor task_executor(*this);
6277 rp->enqueue_discovered_references(&task_executor);
6278 } else {
6279 rp->enqueue_discovered_references(NULL);
6280 }
6281 rp->verify_no_references_recorded();
6282 assert(!rp->discovery_enabled(), "should have been disabled");
6283 }
6284
6285 #ifndef PRODUCT
6286 void CMSCollector::check_correct_thread_executing() {
6287 Thread* t = Thread::current();
6288 // Only the VM thread or the CMS thread should be here.
6289 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6290 "Unexpected thread type");
6291 // If this is the vm thread, the foreground process
6292 // should not be waiting. Note that _foregroundGCIsActive is
6293 // true while the foreground collector is waiting.
6294 if (_foregroundGCShouldWait) {
6295 // We cannot be the VM thread
6296 assert(t->is_ConcurrentGC_thread(),
6297 "Should be CMS thread");
6298 } else {
6299 // We can be the CMS thread only if we are in a stop-world
6300 // phase of CMS collection.
6301 if (t->is_ConcurrentGC_thread()) {
6302 assert(_collectorState == InitialMarking ||
6303 _collectorState == FinalMarking,
6304 "Should be a stop-world phase");
6305 // The CMS thread should be holding the CMS_token.
6306 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6307 "Potential interference with concurrently "
6308 "executing VM thread");
6309 }
6310 }
6311 }
6312 #endif
6313
6314 void CMSCollector::sweep(bool asynch) {
6315 assert(_collectorState == Sweeping, "just checking");
6316 check_correct_thread_executing();
6317 verify_work_stacks_empty();
6318 verify_overflow_empty();
6319 increment_sweep_count();
6320 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6321
6322 _inter_sweep_timer.stop();
6323 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6324 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6325
6326 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6327 _intra_sweep_timer.reset();
6328 _intra_sweep_timer.start();
6329 if (asynch) {
6330 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6331 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6332 // First sweep the old gen
6333 {
6334 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6335 bitMapLock());
6336 sweepWork(_cmsGen, asynch);
6337 }
6338
6339 // Update Universe::_heap_*_at_gc figures.
6340 // We need all the free list locks to make the abstract state
6341 // transition from Sweeping to Resetting. See detailed note
6342 // further below.
6343 {
6344 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
6345 // Update heap occupancy information which is used as
6346 // input to soft ref clearing policy at the next gc.
6347 Universe::update_heap_info_at_gc();
6348 _collectorState = Resizing;
6349 }
6350 } else {
6351 // already have needed locks
6352 sweepWork(_cmsGen, asynch);
6353 // Update heap occupancy information which is used as
6354 // input to soft ref clearing policy at the next gc.
6355 Universe::update_heap_info_at_gc();
6356 _collectorState = Resizing;
6357 }
6358 verify_work_stacks_empty();
6359 verify_overflow_empty();
6360
6361 if (should_unload_classes()) {
6362 ClassLoaderDataGraph::purge();
6363 }
6364
6365 _intra_sweep_timer.stop();
6366 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6367
6368 _inter_sweep_timer.reset();
6369 _inter_sweep_timer.start();
6370
6371 // We need to use a monotonically non-decreasing time in ms
6372 // or we will see time-warp warnings and os::javaTimeMillis()
6373 // does not guarantee monotonicity.
6374 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6375 update_time_of_last_gc(now);
6376
6377 // NOTE on abstract state transitions:
6378 // Mutators allocate-live and/or mark the mod-union table dirty
6379 // based on the state of the collection. The former is done in
6380 // the interval [Marking, Sweeping] and the latter in the interval
6381 // [Marking, Sweeping). Thus the transitions into the Marking state
6382 // and out of the Sweeping state must be synchronously visible
6383 // globally to the mutators.
6384 // The transition into the Marking state happens with the world
6385 // stopped so the mutators will globally see it. Sweeping is
6386 // done asynchronously by the background collector so the transition
6387 // from the Sweeping state to the Resizing state must be done
6388 // under the freelistLock (as is the check for whether to
6389 // allocate-live and whether to dirty the mod-union table).
6390 assert(_collectorState == Resizing, "Change of collector state to"
6391 " Resizing must be done under the freelistLocks (plural)");
6392
6393 // Now that sweeping has been completed, we clear
6394 // the incremental_collection_failed flag,
6395 // thus inviting a younger gen collection to promote into
6396 // this generation. If such a promotion may still fail,
6397 // the flag will be set again when a young collection is
6398 // attempted.
6399 GenCollectedHeap* gch = GenCollectedHeap::heap();
6400 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6401 gch->update_full_collections_completed(_collection_count_start);
6402 }
6403
6404 // FIX ME!!! Looks like this belongs in CFLSpace, with
6405 // CMSGen merely delegating to it.
6406 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6407 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6408 HeapWord* minAddr = _cmsSpace->bottom();
6409 HeapWord* largestAddr =
6410 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6411 if (largestAddr == NULL) {
6412 // The dictionary appears to be empty. In this case
6413 // try to coalesce at the end of the heap.
6414 largestAddr = _cmsSpace->end();
6415 }
6416 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6417 size_t nearLargestOffset =
6418 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6419 if (PrintFLSStatistics != 0) {
6420 gclog_or_tty->print_cr(
6421 "CMS: Large Block: " PTR_FORMAT ";"
6422 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6423 largestAddr,
6424 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6425 }
6426 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6427 }
6428
6429 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6430 return addr >= _cmsSpace->nearLargestChunk();
6431 }
6432
6433 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6434 return _cmsSpace->find_chunk_at_end();
6435 }
6436
6437 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6438 bool full) {
6439 // The next lower level has been collected. Gather any statistics
6440 // that are of interest at this point.
6441 if (!full && (current_level + 1) == level()) {
6442 // Gather statistics on the young generation collection.
6443 collector()->stats().record_gc0_end(used());
6444 }
6445 }
6446
6447 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6448 GenCollectedHeap* gch = GenCollectedHeap::heap();
6449 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6450 "Wrong type of heap");
6451 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6452 gch->gen_policy()->size_policy();
6453 assert(sp->is_gc_cms_adaptive_size_policy(),
6454 "Wrong type of size policy");
6455 return sp;
6456 }
6457
6458 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6459 if (PrintGCDetails && Verbose) {
6460 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6461 }
6462 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6463 _debug_collection_type =
6464 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6465 if (PrintGCDetails && Verbose) {
6466 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6467 }
6468 }
6469
6470 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6471 bool asynch) {
6472 // We iterate over the space(s) underlying this generation,
6473 // checking the mark bit map to see if the bits corresponding
6474 // to specific blocks are marked or not. Blocks that are
6475 // marked are live and are not swept up. All remaining blocks
6476 // are swept up, with coalescing on-the-fly as we sweep up
6477 // contiguous free and/or garbage blocks:
6478 // We need to ensure that the sweeper synchronizes with allocators
6479 // and stop-the-world collectors. In particular, the following
6480 // locks are used:
6481 // . CMS token: if this is held, a stop the world collection cannot occur
6482 // . freelistLock: if this is held no allocation can occur from this
6483 // generation by another thread
6484 // . bitMapLock: if this is held, no other thread can access or update
6485 //
6486
6487 // Note that we need to hold the freelistLock if we use
6488 // block iterate below; else the iterator might go awry if
6489 // a mutator (or promotion) causes block contents to change
6490 // (for instance if the allocator divvies up a block).
6491 // If we hold the free list lock, for all practical purposes
6492 // young generation GC's can't occur (they'll usually need to
6493 // promote), so we might as well prevent all young generation
6494 // GC's while we do a sweeping step. For the same reason, we might
6495 // as well take the bit map lock for the entire duration
6496
6497 // check that we hold the requisite locks
6498 assert(have_cms_token(), "Should hold cms token");
6499 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6500 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6501 "Should possess CMS token to sweep");
6502 assert_lock_strong(gen->freelistLock());
6503 assert_lock_strong(bitMapLock());
6504
6505 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6506 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6507 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6508 _inter_sweep_estimate.padded_average(),
6509 _intra_sweep_estimate.padded_average());
6510 gen->setNearLargestChunk();
6511
6512 {
6513 SweepClosure sweepClosure(this, gen, &_markBitMap,
6514 CMSYield && asynch);
6515 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6516 // We need to free-up/coalesce garbage/blocks from a
6517 // co-terminal free run. This is done in the SweepClosure
6518 // destructor; so, do not remove this scope, else the
6519 // end-of-sweep-census below will be off by a little bit.
6520 }
6521 gen->cmsSpace()->sweep_completed();
6522 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6523 if (should_unload_classes()) { // unloaded classes this cycle,
6524 _concurrent_cycles_since_last_unload = 0; // ... reset count
6525 } else { // did not unload classes,
6526 _concurrent_cycles_since_last_unload++; // ... increment count
6527 }
6528 }
6529
6530 // Reset CMS data structures (for now just the marking bit map)
6531 // preparatory for the next cycle.
6532 void CMSCollector::reset(bool asynch) {
6533 GenCollectedHeap* gch = GenCollectedHeap::heap();
6534 CMSAdaptiveSizePolicy* sp = size_policy();
6535 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6536 if (asynch) {
6537 CMSTokenSyncWithLocks ts(true, bitMapLock());
6538
6539 // If the state is not "Resetting", the foreground thread
6540 // has done a collection and the resetting.
6541 if (_collectorState != Resetting) {
6542 assert(_collectorState == Idling, "The state should only change"
6543 " because the foreground collector has finished the collection");
6544 return;
6545 }
6546
6547 // Clear the mark bitmap (no grey objects to start with)
6548 // for the next cycle.
6549 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6550 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6551
6552 HeapWord* curAddr = _markBitMap.startWord();
6553 while (curAddr < _markBitMap.endWord()) {
6554 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6555 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6556 _markBitMap.clear_large_range(chunk);
6557 if (ConcurrentMarkSweepThread::should_yield() &&
6558 !foregroundGCIsActive() &&
6559 CMSYield) {
6560 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6561 "CMS thread should hold CMS token");
6562 assert_lock_strong(bitMapLock());
6563 bitMapLock()->unlock();
6564 ConcurrentMarkSweepThread::desynchronize(true);
6565 ConcurrentMarkSweepThread::acknowledge_yield_request();
6566 stopTimer();
6567 if (PrintCMSStatistics != 0) {
6568 incrementYields();
6569 }
6570 icms_wait();
6571
6572 // See the comment in coordinator_yield()
6573 for (unsigned i = 0; i < CMSYieldSleepCount &&
6574 ConcurrentMarkSweepThread::should_yield() &&
6575 !CMSCollector::foregroundGCIsActive(); ++i) {
6576 os::sleep(Thread::current(), 1, false);
6577 ConcurrentMarkSweepThread::acknowledge_yield_request();
6578 }
6579
6580 ConcurrentMarkSweepThread::synchronize(true);
6581 bitMapLock()->lock_without_safepoint_check();
6582 startTimer();
6583 }
6584 curAddr = chunk.end();
6585 }
6586 // A successful mostly concurrent collection has been done.
6587 // Because only the full (i.e., concurrent mode failure) collections
6588 // are being measured for gc overhead limits, clean the "near" flag
6589 // and count.
6590 sp->reset_gc_overhead_limit_count();
6591 _collectorState = Idling;
6592 } else {
6593 // already have the lock
6594 assert(_collectorState == Resetting, "just checking");
6595 assert_lock_strong(bitMapLock());
6596 _markBitMap.clear_all();
6597 _collectorState = Idling;
6598 }
6599
6600 // Stop incremental mode after a cycle completes, so that any future cycles
6601 // are triggered by allocation.
6602 stop_icms();
6603
6604 NOT_PRODUCT(
6605 if (RotateCMSCollectionTypes) {
6606 _cmsGen->rotate_debug_collection_type();
6607 }
6608 )
6609
6610 register_gc_end();
6611 }
6612
6613 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6614 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6615 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6616 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
6617 TraceCollectorStats tcs(counters());
6618
6619 switch (op) {
6620 case CMS_op_checkpointRootsInitial: {
6621 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6622 checkpointRootsInitial(true); // asynch
6623 if (PrintGC) {
6624 _cmsGen->printOccupancy("initial-mark");
6625 }
6626 break;
6627 }
6628 case CMS_op_checkpointRootsFinal: {
6629 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6630 checkpointRootsFinal(true, // asynch
6631 false, // !clear_all_soft_refs
6632 false); // !init_mark_was_synchronous
6633 if (PrintGC) {
6634 _cmsGen->printOccupancy("remark");
6635 }
6636 break;
6637 }
6638 default:
6639 fatal("No such CMS_op");
6640 }
6641 }
6642
6643 #ifndef PRODUCT
6644 size_t const CMSCollector::skip_header_HeapWords() {
6645 return FreeChunk::header_size();
6646 }
6647
6648 // Try and collect here conditions that should hold when
6649 // CMS thread is exiting. The idea is that the foreground GC
6650 // thread should not be blocked if it wants to terminate
6651 // the CMS thread and yet continue to run the VM for a while
6652 // after that.
6653 void CMSCollector::verify_ok_to_terminate() const {
6654 assert(Thread::current()->is_ConcurrentGC_thread(),
6655 "should be called by CMS thread");
6656 assert(!_foregroundGCShouldWait, "should be false");
6657 // We could check here that all the various low-level locks
6658 // are not held by the CMS thread, but that is overkill; see
6659 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6660 // is checked.
6661 }
6662 #endif
6663
6664 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6665 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6666 "missing Printezis mark?");
6667 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6668 size_t size = pointer_delta(nextOneAddr + 1, addr);
6669 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6670 "alignment problem");
6671 assert(size >= 3, "Necessary for Printezis marks to work");
6672 return size;
6673 }
6674
6675 // A variant of the above (block_size_using_printezis_bits()) except
6676 // that we return 0 if the P-bits are not yet set.
6677 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6678 if (_markBitMap.isMarked(addr + 1)) {
6679 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6680 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6681 size_t size = pointer_delta(nextOneAddr + 1, addr);
6682 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6683 "alignment problem");
6684 assert(size >= 3, "Necessary for Printezis marks to work");
6685 return size;
6686 }
6687 return 0;
6688 }
6689
6690 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6691 size_t sz = 0;
6692 oop p = (oop)addr;
6693 if (p->klass_or_null() != NULL) {
6694 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6695 } else {
6696 sz = block_size_using_printezis_bits(addr);
6697 }
6698 assert(sz > 0, "size must be nonzero");
6699 HeapWord* next_block = addr + sz;
6700 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6701 CardTableModRefBS::card_size);
6702 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6703 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6704 "must be different cards");
6705 return next_card;
6706 }
6707
6708
6709 // CMS Bit Map Wrapper /////////////////////////////////////////
6710
6711 // Construct a CMS bit map infrastructure, but don't create the
6712 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6713 // further below.
6714 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6715 _bm(),
6716 _shifter(shifter),
6717 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6718 {
6719 _bmStartWord = 0;
6720 _bmWordSize = 0;
6721 }
6722
6723 bool CMSBitMap::allocate(MemRegion mr) {
6724 _bmStartWord = mr.start();
6725 _bmWordSize = mr.word_size();
6726 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6727 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6728 if (!brs.is_reserved()) {
6729 warning("CMS bit map allocation failure");
6730 return false;
6731 }
6732 // For now we'll just commit all of the bit map up front.
6733 // Later on we'll try to be more parsimonious with swap.
6734 if (!_virtual_space.initialize(brs, brs.size())) {
6735 warning("CMS bit map backing store failure");
6736 return false;
6737 }
6738 assert(_virtual_space.committed_size() == brs.size(),
6739 "didn't reserve backing store for all of CMS bit map?");
6740 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6741 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6742 _bmWordSize, "inconsistency in bit map sizing");
6743 _bm.set_size(_bmWordSize >> _shifter);
6744
6745 // bm.clear(); // can we rely on getting zero'd memory? verify below
6746 assert(isAllClear(),
6747 "Expected zero'd memory from ReservedSpace constructor");
6748 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6749 "consistency check");
6750 return true;
6751 }
6752
6753 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6754 HeapWord *next_addr, *end_addr, *last_addr;
6755 assert_locked();
6756 assert(covers(mr), "out-of-range error");
6757 // XXX assert that start and end are appropriately aligned
6758 for (next_addr = mr.start(), end_addr = mr.end();
6759 next_addr < end_addr; next_addr = last_addr) {
6760 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6761 last_addr = dirty_region.end();
6762 if (!dirty_region.is_empty()) {
6763 cl->do_MemRegion(dirty_region);
6764 } else {
6765 assert(last_addr == end_addr, "program logic");
6766 return;
6767 }
6768 }
6769 }
6770
6771 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
6772 _bm.print_on_error(st, prefix);
6773 }
6774
6775 #ifndef PRODUCT
6776 void CMSBitMap::assert_locked() const {
6777 CMSLockVerifier::assert_locked(lock());
6778 }
6779
6780 bool CMSBitMap::covers(MemRegion mr) const {
6781 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6782 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6783 "size inconsistency");
6784 return (mr.start() >= _bmStartWord) &&
6785 (mr.end() <= endWord());
6786 }
6787
6788 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6789 return (start >= _bmStartWord && (start + size) <= endWord());
6790 }
6791
6792 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6793 // verify that there are no 1 bits in the interval [left, right)
6794 FalseBitMapClosure falseBitMapClosure;
6795 iterate(&falseBitMapClosure, left, right);
6796 }
6797
6798 void CMSBitMap::region_invariant(MemRegion mr)
6799 {
6800 assert_locked();
6801 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6802 assert(!mr.is_empty(), "unexpected empty region");
6803 assert(covers(mr), "mr should be covered by bit map");
6804 // convert address range into offset range
6805 size_t start_ofs = heapWordToOffset(mr.start());
6806 // Make sure that end() is appropriately aligned
6807 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6808 (1 << (_shifter+LogHeapWordSize))),
6809 "Misaligned mr.end()");
6810 size_t end_ofs = heapWordToOffset(mr.end());
6811 assert(end_ofs > start_ofs, "Should mark at least one bit");
6812 }
6813
6814 #endif
6815
6816 bool CMSMarkStack::allocate(size_t size) {
6817 // allocate a stack of the requisite depth
6818 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6819 size * sizeof(oop)));
6820 if (!rs.is_reserved()) {
6821 warning("CMSMarkStack allocation failure");
6822 return false;
6823 }
6824 if (!_virtual_space.initialize(rs, rs.size())) {
6825 warning("CMSMarkStack backing store failure");
6826 return false;
6827 }
6828 assert(_virtual_space.committed_size() == rs.size(),
6829 "didn't reserve backing store for all of CMS stack?");
6830 _base = (oop*)(_virtual_space.low());
6831 _index = 0;
6832 _capacity = size;
6833 NOT_PRODUCT(_max_depth = 0);
6834 return true;
6835 }
6836
6837 // XXX FIX ME !!! In the MT case we come in here holding a
6838 // leaf lock. For printing we need to take a further lock
6839 // which has lower rank. We need to recalibrate the two
6840 // lock-ranks involved in order to be able to print the
6841 // messages below. (Or defer the printing to the caller.
6842 // For now we take the expedient path of just disabling the
6843 // messages for the problematic case.)
6844 void CMSMarkStack::expand() {
6845 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6846 if (_capacity == MarkStackSizeMax) {
6847 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6848 // We print a warning message only once per CMS cycle.
6849 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6850 }
6851 return;
6852 }
6853 // Double capacity if possible
6854 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6855 // Do not give up existing stack until we have managed to
6856 // get the double capacity that we desired.
6857 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6858 new_capacity * sizeof(oop)));
6859 if (rs.is_reserved()) {
6860 // Release the backing store associated with old stack
6861 _virtual_space.release();
6862 // Reinitialize virtual space for new stack
6863 if (!_virtual_space.initialize(rs, rs.size())) {
6864 fatal("Not enough swap for expanded marking stack");
6865 }
6866 _base = (oop*)(_virtual_space.low());
6867 _index = 0;
6868 _capacity = new_capacity;
6869 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6870 // Failed to double capacity, continue;
6871 // we print a detail message only once per CMS cycle.
6872 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6873 SIZE_FORMAT"K",
6874 _capacity / K, new_capacity / K);
6875 }
6876 }
6877
6878
6879 // Closures
6880 // XXX: there seems to be a lot of code duplication here;
6881 // should refactor and consolidate common code.
6882
6883 // This closure is used to mark refs into the CMS generation in
6884 // the CMS bit map. Called at the first checkpoint. This closure
6885 // assumes that we do not need to re-mark dirty cards; if the CMS
6886 // generation on which this is used is not an oldest
6887 // generation then this will lose younger_gen cards!
6888
6889 MarkRefsIntoClosure::MarkRefsIntoClosure(
6890 MemRegion span, CMSBitMap* bitMap):
6891 _span(span),
6892 _bitMap(bitMap)
6893 {
6894 assert(_ref_processor == NULL, "deliberately left NULL");
6895 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6896 }
6897
6898 void MarkRefsIntoClosure::do_oop(oop obj) {
6899 // if p points into _span, then mark corresponding bit in _markBitMap
6900 assert(obj->is_oop(), "expected an oop");
6901 HeapWord* addr = (HeapWord*)obj;
6902 if (_span.contains(addr)) {
6903 // this should be made more efficient
6904 _bitMap->mark(addr);
6905 }
6906 }
6907
6908 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6909 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6910
6911 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6912 MemRegion span, CMSBitMap* bitMap):
6913 _span(span),
6914 _bitMap(bitMap)
6915 {
6916 assert(_ref_processor == NULL, "deliberately left NULL");
6917 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6918 }
6919
6920 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6921 // if p points into _span, then mark corresponding bit in _markBitMap
6922 assert(obj->is_oop(), "expected an oop");
6923 HeapWord* addr = (HeapWord*)obj;
6924 if (_span.contains(addr)) {
6925 // this should be made more efficient
6926 _bitMap->par_mark(addr);
6927 }
6928 }
6929
6930 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6931 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6932
6933 // A variant of the above, used for CMS marking verification.
6934 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6935 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6936 _span(span),
6937 _verification_bm(verification_bm),
6938 _cms_bm(cms_bm)
6939 {
6940 assert(_ref_processor == NULL, "deliberately left NULL");
6941 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6942 }
6943
6944 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6945 // if p points into _span, then mark corresponding bit in _markBitMap
6946 assert(obj->is_oop(), "expected an oop");
6947 HeapWord* addr = (HeapWord*)obj;
6948 if (_span.contains(addr)) {
6949 _verification_bm->mark(addr);
6950 if (!_cms_bm->isMarked(addr)) {
6951 oop(addr)->print();
6952 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6953 fatal("... aborting");
6954 }
6955 }
6956 }
6957
6958 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6959 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6960
6961 //////////////////////////////////////////////////
6962 // MarkRefsIntoAndScanClosure
6963 //////////////////////////////////////////////////
6964
6965 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6966 ReferenceProcessor* rp,
6967 CMSBitMap* bit_map,
6968 CMSBitMap* mod_union_table,
6969 CMSMarkStack* mark_stack,
6970 CMSCollector* collector,
6971 bool should_yield,
6972 bool concurrent_precleaning):
6973 _collector(collector),
6974 _span(span),
6975 _bit_map(bit_map),
6976 _mark_stack(mark_stack),
6977 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6978 mark_stack, concurrent_precleaning),
6979 _yield(should_yield),
6980 _concurrent_precleaning(concurrent_precleaning),
6981 _freelistLock(NULL)
6982 {
6983 _ref_processor = rp;
6984 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6985 }
6986
6987 // This closure is used to mark refs into the CMS generation at the
6988 // second (final) checkpoint, and to scan and transitively follow
6989 // the unmarked oops. It is also used during the concurrent precleaning
6990 // phase while scanning objects on dirty cards in the CMS generation.
6991 // The marks are made in the marking bit map and the marking stack is
6992 // used for keeping the (newly) grey objects during the scan.
6993 // The parallel version (Par_...) appears further below.
6994 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6995 if (obj != NULL) {
6996 assert(obj->is_oop(), "expected an oop");
6997 HeapWord* addr = (HeapWord*)obj;
6998 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6999 assert(_collector->overflow_list_is_empty(),
7000 "overflow list should be empty");
7001 if (_span.contains(addr) &&
7002 !_bit_map->isMarked(addr)) {
7003 // mark bit map (object is now grey)
7004 _bit_map->mark(addr);
7005 // push on marking stack (stack should be empty), and drain the
7006 // stack by applying this closure to the oops in the oops popped
7007 // from the stack (i.e. blacken the grey objects)
7008 bool res = _mark_stack->push(obj);
7009 assert(res, "Should have space to push on empty stack");
7010 do {
7011 oop new_oop = _mark_stack->pop();
7012 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7013 assert(_bit_map->isMarked((HeapWord*)new_oop),
7014 "only grey objects on this stack");
7015 // iterate over the oops in this oop, marking and pushing
7016 // the ones in CMS heap (i.e. in _span).
7017 new_oop->oop_iterate(&_pushAndMarkClosure);
7018 // check if it's time to yield
7019 do_yield_check();
7020 } while (!_mark_stack->isEmpty() ||
7021 (!_concurrent_precleaning && take_from_overflow_list()));
7022 // if marking stack is empty, and we are not doing this
7023 // during precleaning, then check the overflow list
7024 }
7025 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7026 assert(_collector->overflow_list_is_empty(),
7027 "overflow list was drained above");
7028 // We could restore evacuated mark words, if any, used for
7029 // overflow list links here because the overflow list is
7030 // provably empty here. That would reduce the maximum
7031 // size requirements for preserved_{oop,mark}_stack.
7032 // But we'll just postpone it until we are all done
7033 // so we can just stream through.
7034 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
7035 _collector->restore_preserved_marks_if_any();
7036 assert(_collector->no_preserved_marks(), "No preserved marks");
7037 }
7038 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
7039 "All preserved marks should have been restored above");
7040 }
7041 }
7042
7043 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
7044 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
7045
7046 void MarkRefsIntoAndScanClosure::do_yield_work() {
7047 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7048 "CMS thread should hold CMS token");
7049 assert_lock_strong(_freelistLock);
7050 assert_lock_strong(_bit_map->lock());
7051 // relinquish the free_list_lock and bitMaplock()
7052 _bit_map->lock()->unlock();
7053 _freelistLock->unlock();
7054 ConcurrentMarkSweepThread::desynchronize(true);
7055 ConcurrentMarkSweepThread::acknowledge_yield_request();
7056 _collector->stopTimer();
7057 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7058 if (PrintCMSStatistics != 0) {
7059 _collector->incrementYields();
7060 }
7061 _collector->icms_wait();
7062
7063 // See the comment in coordinator_yield()
7064 for (unsigned i = 0;
7065 i < CMSYieldSleepCount &&
7066 ConcurrentMarkSweepThread::should_yield() &&
7067 !CMSCollector::foregroundGCIsActive();
7068 ++i) {
7069 os::sleep(Thread::current(), 1, false);
7070 ConcurrentMarkSweepThread::acknowledge_yield_request();
7071 }
7072
7073 ConcurrentMarkSweepThread::synchronize(true);
7074 _freelistLock->lock_without_safepoint_check();
7075 _bit_map->lock()->lock_without_safepoint_check();
7076 _collector->startTimer();
7077 }
7078
7079 ///////////////////////////////////////////////////////////
7080 // Par_MarkRefsIntoAndScanClosure: a parallel version of
7081 // MarkRefsIntoAndScanClosure
7082 ///////////////////////////////////////////////////////////
7083 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
7084 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
7085 CMSBitMap* bit_map, OopTaskQueue* work_queue):
7086 _span(span),
7087 _bit_map(bit_map),
7088 _work_queue(work_queue),
7089 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
7090 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
7091 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
7092 {
7093 _ref_processor = rp;
7094 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7095 }
7096
7097 // This closure is used to mark refs into the CMS generation at the
7098 // second (final) checkpoint, and to scan and transitively follow
7099 // the unmarked oops. The marks are made in the marking bit map and
7100 // the work_queue is used for keeping the (newly) grey objects during
7101 // the scan phase whence they are also available for stealing by parallel
7102 // threads. Since the marking bit map is shared, updates are
7103 // synchronized (via CAS).
7104 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
7105 if (obj != NULL) {
7106 // Ignore mark word because this could be an already marked oop
7107 // that may be chained at the end of the overflow list.
7108 assert(obj->is_oop(true), "expected an oop");
7109 HeapWord* addr = (HeapWord*)obj;
7110 if (_span.contains(addr) &&
7111 !_bit_map->isMarked(addr)) {
7112 // mark bit map (object will become grey):
7113 // It is possible for several threads to be
7114 // trying to "claim" this object concurrently;
7115 // the unique thread that succeeds in marking the
7116 // object first will do the subsequent push on
7117 // to the work queue (or overflow list).
7118 if (_bit_map->par_mark(addr)) {
7119 // push on work_queue (which may not be empty), and trim the
7120 // queue to an appropriate length by applying this closure to
7121 // the oops in the oops popped from the stack (i.e. blacken the
7122 // grey objects)
7123 bool res = _work_queue->push(obj);
7124 assert(res, "Low water mark should be less than capacity?");
7125 trim_queue(_low_water_mark);
7126 } // Else, another thread claimed the object
7127 }
7128 }
7129 }
7130
7131 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7132 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7133
7134 // This closure is used to rescan the marked objects on the dirty cards
7135 // in the mod union table and the card table proper.
7136 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
7137 oop p, MemRegion mr) {
7138
7139 size_t size = 0;
7140 HeapWord* addr = (HeapWord*)p;
7141 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7142 assert(_span.contains(addr), "we are scanning the CMS generation");
7143 // check if it's time to yield
7144 if (do_yield_check()) {
7145 // We yielded for some foreground stop-world work,
7146 // and we have been asked to abort this ongoing preclean cycle.
7147 return 0;
7148 }
7149 if (_bitMap->isMarked(addr)) {
7150 // it's marked; is it potentially uninitialized?
7151 if (p->klass_or_null() != NULL) {
7152 // an initialized object; ignore mark word in verification below
7153 // since we are running concurrent with mutators
7154 assert(p->is_oop(true), "should be an oop");
7155 if (p->is_objArray()) {
7156 // objArrays are precisely marked; restrict scanning
7157 // to dirty cards only.
7158 size = CompactibleFreeListSpace::adjustObjectSize(
7159 p->oop_iterate(_scanningClosure, mr));
7160 } else {
7161 // A non-array may have been imprecisely marked; we need
7162 // to scan object in its entirety.
7163 size = CompactibleFreeListSpace::adjustObjectSize(
7164 p->oop_iterate(_scanningClosure));
7165 }
7166 #ifdef ASSERT
7167 size_t direct_size =
7168 CompactibleFreeListSpace::adjustObjectSize(p->size());
7169 assert(size == direct_size, "Inconsistency in size");
7170 assert(size >= 3, "Necessary for Printezis marks to work");
7171 if (!_bitMap->isMarked(addr+1)) {
7172 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
7173 } else {
7174 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
7175 assert(_bitMap->isMarked(addr+size-1),
7176 "inconsistent Printezis mark");
7177 }
7178 #endif // ASSERT
7179 } else {
7180 // An uninitialized object.
7181 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
7182 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7183 size = pointer_delta(nextOneAddr + 1, addr);
7184 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7185 "alignment problem");
7186 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
7187 // will dirty the card when the klass pointer is installed in the
7188 // object (signaling the completion of initialization).
7189 }
7190 } else {
7191 // Either a not yet marked object or an uninitialized object
7192 if (p->klass_or_null() == NULL) {
7193 // An uninitialized object, skip to the next card, since
7194 // we may not be able to read its P-bits yet.
7195 assert(size == 0, "Initial value");
7196 } else {
7197 // An object not (yet) reached by marking: we merely need to
7198 // compute its size so as to go look at the next block.
7199 assert(p->is_oop(true), "should be an oop");
7200 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
7201 }
7202 }
7203 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7204 return size;
7205 }
7206
7207 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
7208 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7209 "CMS thread should hold CMS token");
7210 assert_lock_strong(_freelistLock);
7211 assert_lock_strong(_bitMap->lock());
7212 // relinquish the free_list_lock and bitMaplock()
7213 _bitMap->lock()->unlock();
7214 _freelistLock->unlock();
7215 ConcurrentMarkSweepThread::desynchronize(true);
7216 ConcurrentMarkSweepThread::acknowledge_yield_request();
7217 _collector->stopTimer();
7218 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7219 if (PrintCMSStatistics != 0) {
7220 _collector->incrementYields();
7221 }
7222 _collector->icms_wait();
7223
7224 // See the comment in coordinator_yield()
7225 for (unsigned i = 0; i < CMSYieldSleepCount &&
7226 ConcurrentMarkSweepThread::should_yield() &&
7227 !CMSCollector::foregroundGCIsActive(); ++i) {
7228 os::sleep(Thread::current(), 1, false);
7229 ConcurrentMarkSweepThread::acknowledge_yield_request();
7230 }
7231
7232 ConcurrentMarkSweepThread::synchronize(true);
7233 _freelistLock->lock_without_safepoint_check();
7234 _bitMap->lock()->lock_without_safepoint_check();
7235 _collector->startTimer();
7236 }
7237
7238
7239 //////////////////////////////////////////////////////////////////
7240 // SurvivorSpacePrecleanClosure
7241 //////////////////////////////////////////////////////////////////
7242 // This (single-threaded) closure is used to preclean the oops in
7243 // the survivor spaces.
7244 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
7245
7246 HeapWord* addr = (HeapWord*)p;
7247 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7248 assert(!_span.contains(addr), "we are scanning the survivor spaces");
7249 assert(p->klass_or_null() != NULL, "object should be initialized");
7250 // an initialized object; ignore mark word in verification below
7251 // since we are running concurrent with mutators
7252 assert(p->is_oop(true), "should be an oop");
7253 // Note that we do not yield while we iterate over
7254 // the interior oops of p, pushing the relevant ones
7255 // on our marking stack.
7256 size_t size = p->oop_iterate(_scanning_closure);
7257 do_yield_check();
7258 // Observe that below, we do not abandon the preclean
7259 // phase as soon as we should; rather we empty the
7260 // marking stack before returning. This is to satisfy
7261 // some existing assertions. In general, it may be a
7262 // good idea to abort immediately and complete the marking
7263 // from the grey objects at a later time.
7264 while (!_mark_stack->isEmpty()) {
7265 oop new_oop = _mark_stack->pop();
7266 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7267 assert(_bit_map->isMarked((HeapWord*)new_oop),
7268 "only grey objects on this stack");
7269 // iterate over the oops in this oop, marking and pushing
7270 // the ones in CMS heap (i.e. in _span).
7271 new_oop->oop_iterate(_scanning_closure);
7272 // check if it's time to yield
7273 do_yield_check();
7274 }
7275 unsigned int after_count =
7276 GenCollectedHeap::heap()->total_collections();
7277 bool abort = (_before_count != after_count) ||
7278 _collector->should_abort_preclean();
7279 return abort ? 0 : size;
7280 }
7281
7282 void SurvivorSpacePrecleanClosure::do_yield_work() {
7283 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7284 "CMS thread should hold CMS token");
7285 assert_lock_strong(_bit_map->lock());
7286 // Relinquish the bit map lock
7287 _bit_map->lock()->unlock();
7288 ConcurrentMarkSweepThread::desynchronize(true);
7289 ConcurrentMarkSweepThread::acknowledge_yield_request();
7290 _collector->stopTimer();
7291 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7292 if (PrintCMSStatistics != 0) {
7293 _collector->incrementYields();
7294 }
7295 _collector->icms_wait();
7296
7297 // See the comment in coordinator_yield()
7298 for (unsigned i = 0; i < CMSYieldSleepCount &&
7299 ConcurrentMarkSweepThread::should_yield() &&
7300 !CMSCollector::foregroundGCIsActive(); ++i) {
7301 os::sleep(Thread::current(), 1, false);
7302 ConcurrentMarkSweepThread::acknowledge_yield_request();
7303 }
7304
7305 ConcurrentMarkSweepThread::synchronize(true);
7306 _bit_map->lock()->lock_without_safepoint_check();
7307 _collector->startTimer();
7308 }
7309
7310 // This closure is used to rescan the marked objects on the dirty cards
7311 // in the mod union table and the card table proper. In the parallel
7312 // case, although the bitMap is shared, we do a single read so the
7313 // isMarked() query is "safe".
7314 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7315 // Ignore mark word because we are running concurrent with mutators
7316 assert(p->is_oop_or_null(true), "expected an oop or null");
7317 HeapWord* addr = (HeapWord*)p;
7318 assert(_span.contains(addr), "we are scanning the CMS generation");
7319 bool is_obj_array = false;
7320 #ifdef ASSERT
7321 if (!_parallel) {
7322 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7323 assert(_collector->overflow_list_is_empty(),
7324 "overflow list should be empty");
7325
7326 }
7327 #endif // ASSERT
7328 if (_bit_map->isMarked(addr)) {
7329 // Obj arrays are precisely marked, non-arrays are not;
7330 // so we scan objArrays precisely and non-arrays in their
7331 // entirety.
7332 if (p->is_objArray()) {
7333 is_obj_array = true;
7334 if (_parallel) {
7335 p->oop_iterate(_par_scan_closure, mr);
7336 } else {
7337 p->oop_iterate(_scan_closure, mr);
7338 }
7339 } else {
7340 if (_parallel) {
7341 p->oop_iterate(_par_scan_closure);
7342 } else {
7343 p->oop_iterate(_scan_closure);
7344 }
7345 }
7346 }
7347 #ifdef ASSERT
7348 if (!_parallel) {
7349 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7350 assert(_collector->overflow_list_is_empty(),
7351 "overflow list should be empty");
7352
7353 }
7354 #endif // ASSERT
7355 return is_obj_array;
7356 }
7357
7358 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7359 MemRegion span,
7360 CMSBitMap* bitMap, CMSMarkStack* markStack,
7361 bool should_yield, bool verifying):
7362 _collector(collector),
7363 _span(span),
7364 _bitMap(bitMap),
7365 _mut(&collector->_modUnionTable),
7366 _markStack(markStack),
7367 _yield(should_yield),
7368 _skipBits(0)
7369 {
7370 assert(_markStack->isEmpty(), "stack should be empty");
7371 _finger = _bitMap->startWord();
7372 _threshold = _finger;
7373 assert(_collector->_restart_addr == NULL, "Sanity check");
7374 assert(_span.contains(_finger), "Out of bounds _finger?");
7375 DEBUG_ONLY(_verifying = verifying;)
7376 }
7377
7378 void MarkFromRootsClosure::reset(HeapWord* addr) {
7379 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7380 assert(_span.contains(addr), "Out of bounds _finger?");
7381 _finger = addr;
7382 _threshold = (HeapWord*)round_to(
7383 (intptr_t)_finger, CardTableModRefBS::card_size);
7384 }
7385
7386 // Should revisit to see if this should be restructured for
7387 // greater efficiency.
7388 bool MarkFromRootsClosure::do_bit(size_t offset) {
7389 if (_skipBits > 0) {
7390 _skipBits--;
7391 return true;
7392 }
7393 // convert offset into a HeapWord*
7394 HeapWord* addr = _bitMap->startWord() + offset;
7395 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7396 "address out of range");
7397 assert(_bitMap->isMarked(addr), "tautology");
7398 if (_bitMap->isMarked(addr+1)) {
7399 // this is an allocated but not yet initialized object
7400 assert(_skipBits == 0, "tautology");
7401 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7402 oop p = oop(addr);
7403 if (p->klass_or_null() == NULL) {
7404 DEBUG_ONLY(if (!_verifying) {)
7405 // We re-dirty the cards on which this object lies and increase
7406 // the _threshold so that we'll come back to scan this object
7407 // during the preclean or remark phase. (CMSCleanOnEnter)
7408 if (CMSCleanOnEnter) {
7409 size_t sz = _collector->block_size_using_printezis_bits(addr);
7410 HeapWord* end_card_addr = (HeapWord*)round_to(
7411 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7412 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7413 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7414 // Bump _threshold to end_card_addr; note that
7415 // _threshold cannot possibly exceed end_card_addr, anyhow.
7416 // This prevents future clearing of the card as the scan proceeds
7417 // to the right.
7418 assert(_threshold <= end_card_addr,
7419 "Because we are just scanning into this object");
7420 if (_threshold < end_card_addr) {
7421 _threshold = end_card_addr;
7422 }
7423 if (p->klass_or_null() != NULL) {
7424 // Redirty the range of cards...
7425 _mut->mark_range(redirty_range);
7426 } // ...else the setting of klass will dirty the card anyway.
7427 }
7428 DEBUG_ONLY(})
7429 return true;
7430 }
7431 }
7432 scanOopsInOop(addr);
7433 return true;
7434 }
7435
7436 // We take a break if we've been at this for a while,
7437 // so as to avoid monopolizing the locks involved.
7438 void MarkFromRootsClosure::do_yield_work() {
7439 // First give up the locks, then yield, then re-lock
7440 // We should probably use a constructor/destructor idiom to
7441 // do this unlock/lock or modify the MutexUnlocker class to
7442 // serve our purpose. XXX
7443 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7444 "CMS thread should hold CMS token");
7445 assert_lock_strong(_bitMap->lock());
7446 _bitMap->lock()->unlock();
7447 ConcurrentMarkSweepThread::desynchronize(true);
7448 ConcurrentMarkSweepThread::acknowledge_yield_request();
7449 _collector->stopTimer();
7450 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7451 if (PrintCMSStatistics != 0) {
7452 _collector->incrementYields();
7453 }
7454 _collector->icms_wait();
7455
7456 // See the comment in coordinator_yield()
7457 for (unsigned i = 0; i < CMSYieldSleepCount &&
7458 ConcurrentMarkSweepThread::should_yield() &&
7459 !CMSCollector::foregroundGCIsActive(); ++i) {
7460 os::sleep(Thread::current(), 1, false);
7461 ConcurrentMarkSweepThread::acknowledge_yield_request();
7462 }
7463
7464 ConcurrentMarkSweepThread::synchronize(true);
7465 _bitMap->lock()->lock_without_safepoint_check();
7466 _collector->startTimer();
7467 }
7468
7469 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7470 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7471 assert(_markStack->isEmpty(),
7472 "should drain stack to limit stack usage");
7473 // convert ptr to an oop preparatory to scanning
7474 oop obj = oop(ptr);
7475 // Ignore mark word in verification below, since we
7476 // may be running concurrent with mutators.
7477 assert(obj->is_oop(true), "should be an oop");
7478 assert(_finger <= ptr, "_finger runneth ahead");
7479 // advance the finger to right end of this object
7480 _finger = ptr + obj->size();
7481 assert(_finger > ptr, "we just incremented it above");
7482 // On large heaps, it may take us some time to get through
7483 // the marking phase (especially if running iCMS). During
7484 // this time it's possible that a lot of mutations have
7485 // accumulated in the card table and the mod union table --
7486 // these mutation records are redundant until we have
7487 // actually traced into the corresponding card.
7488 // Here, we check whether advancing the finger would make
7489 // us cross into a new card, and if so clear corresponding
7490 // cards in the MUT (preclean them in the card-table in the
7491 // future).
7492
7493 DEBUG_ONLY(if (!_verifying) {)
7494 // The clean-on-enter optimization is disabled by default,
7495 // until we fix 6178663.
7496 if (CMSCleanOnEnter && (_finger > _threshold)) {
7497 // [_threshold, _finger) represents the interval
7498 // of cards to be cleared in MUT (or precleaned in card table).
7499 // The set of cards to be cleared is all those that overlap
7500 // with the interval [_threshold, _finger); note that
7501 // _threshold is always kept card-aligned but _finger isn't
7502 // always card-aligned.
7503 HeapWord* old_threshold = _threshold;
7504 assert(old_threshold == (HeapWord*)round_to(
7505 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7506 "_threshold should always be card-aligned");
7507 _threshold = (HeapWord*)round_to(
7508 (intptr_t)_finger, CardTableModRefBS::card_size);
7509 MemRegion mr(old_threshold, _threshold);
7510 assert(!mr.is_empty(), "Control point invariant");
7511 assert(_span.contains(mr), "Should clear within span");
7512 _mut->clear_range(mr);
7513 }
7514 DEBUG_ONLY(})
7515 // Note: the finger doesn't advance while we drain
7516 // the stack below.
7517 PushOrMarkClosure pushOrMarkClosure(_collector,
7518 _span, _bitMap, _markStack,
7519 _finger, this);
7520 bool res = _markStack->push(obj);
7521 assert(res, "Empty non-zero size stack should have space for single push");
7522 while (!_markStack->isEmpty()) {
7523 oop new_oop = _markStack->pop();
7524 // Skip verifying header mark word below because we are
7525 // running concurrent with mutators.
7526 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7527 // now scan this oop's oops
7528 new_oop->oop_iterate(&pushOrMarkClosure);
7529 do_yield_check();
7530 }
7531 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7532 }
7533
7534 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7535 CMSCollector* collector, MemRegion span,
7536 CMSBitMap* bit_map,
7537 OopTaskQueue* work_queue,
7538 CMSMarkStack* overflow_stack,
7539 bool should_yield):
7540 _collector(collector),
7541 _whole_span(collector->_span),
7542 _span(span),
7543 _bit_map(bit_map),
7544 _mut(&collector->_modUnionTable),
7545 _work_queue(work_queue),
7546 _overflow_stack(overflow_stack),
7547 _yield(should_yield),
7548 _skip_bits(0),
7549 _task(task)
7550 {
7551 assert(_work_queue->size() == 0, "work_queue should be empty");
7552 _finger = span.start();
7553 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7554 assert(_span.contains(_finger), "Out of bounds _finger?");
7555 }
7556
7557 // Should revisit to see if this should be restructured for
7558 // greater efficiency.
7559 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7560 if (_skip_bits > 0) {
7561 _skip_bits--;
7562 return true;
7563 }
7564 // convert offset into a HeapWord*
7565 HeapWord* addr = _bit_map->startWord() + offset;
7566 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7567 "address out of range");
7568 assert(_bit_map->isMarked(addr), "tautology");
7569 if (_bit_map->isMarked(addr+1)) {
7570 // this is an allocated object that might not yet be initialized
7571 assert(_skip_bits == 0, "tautology");
7572 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7573 oop p = oop(addr);
7574 if (p->klass_or_null() == NULL) {
7575 // in the case of Clean-on-Enter optimization, redirty card
7576 // and avoid clearing card by increasing the threshold.
7577 return true;
7578 }
7579 }
7580 scan_oops_in_oop(addr);
7581 return true;
7582 }
7583
7584 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7585 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7586 // Should we assert that our work queue is empty or
7587 // below some drain limit?
7588 assert(_work_queue->size() == 0,
7589 "should drain stack to limit stack usage");
7590 // convert ptr to an oop preparatory to scanning
7591 oop obj = oop(ptr);
7592 // Ignore mark word in verification below, since we
7593 // may be running concurrent with mutators.
7594 assert(obj->is_oop(true), "should be an oop");
7595 assert(_finger <= ptr, "_finger runneth ahead");
7596 // advance the finger to right end of this object
7597 _finger = ptr + obj->size();
7598 assert(_finger > ptr, "we just incremented it above");
7599 // On large heaps, it may take us some time to get through
7600 // the marking phase (especially if running iCMS). During
7601 // this time it's possible that a lot of mutations have
7602 // accumulated in the card table and the mod union table --
7603 // these mutation records are redundant until we have
7604 // actually traced into the corresponding card.
7605 // Here, we check whether advancing the finger would make
7606 // us cross into a new card, and if so clear corresponding
7607 // cards in the MUT (preclean them in the card-table in the
7608 // future).
7609
7610 // The clean-on-enter optimization is disabled by default,
7611 // until we fix 6178663.
7612 if (CMSCleanOnEnter && (_finger > _threshold)) {
7613 // [_threshold, _finger) represents the interval
7614 // of cards to be cleared in MUT (or precleaned in card table).
7615 // The set of cards to be cleared is all those that overlap
7616 // with the interval [_threshold, _finger); note that
7617 // _threshold is always kept card-aligned but _finger isn't
7618 // always card-aligned.
7619 HeapWord* old_threshold = _threshold;
7620 assert(old_threshold == (HeapWord*)round_to(
7621 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7622 "_threshold should always be card-aligned");
7623 _threshold = (HeapWord*)round_to(
7624 (intptr_t)_finger, CardTableModRefBS::card_size);
7625 MemRegion mr(old_threshold, _threshold);
7626 assert(!mr.is_empty(), "Control point invariant");
7627 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7628 _mut->clear_range(mr);
7629 }
7630
7631 // Note: the local finger doesn't advance while we drain
7632 // the stack below, but the global finger sure can and will.
7633 HeapWord** gfa = _task->global_finger_addr();
7634 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7635 _span, _bit_map,
7636 _work_queue,
7637 _overflow_stack,
7638 _finger,
7639 gfa, this);
7640 bool res = _work_queue->push(obj); // overflow could occur here
7641 assert(res, "Will hold once we use workqueues");
7642 while (true) {
7643 oop new_oop;
7644 if (!_work_queue->pop_local(new_oop)) {
7645 // We emptied our work_queue; check if there's stuff that can
7646 // be gotten from the overflow stack.
7647 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7648 _overflow_stack, _work_queue)) {
7649 do_yield_check();
7650 continue;
7651 } else { // done
7652 break;
7653 }
7654 }
7655 // Skip verifying header mark word below because we are
7656 // running concurrent with mutators.
7657 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7658 // now scan this oop's oops
7659 new_oop->oop_iterate(&pushOrMarkClosure);
7660 do_yield_check();
7661 }
7662 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7663 }
7664
7665 // Yield in response to a request from VM Thread or
7666 // from mutators.
7667 void Par_MarkFromRootsClosure::do_yield_work() {
7668 assert(_task != NULL, "sanity");
7669 _task->yield();
7670 }
7671
7672 // A variant of the above used for verifying CMS marking work.
7673 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7674 MemRegion span,
7675 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7676 CMSMarkStack* mark_stack):
7677 _collector(collector),
7678 _span(span),
7679 _verification_bm(verification_bm),
7680 _cms_bm(cms_bm),
7681 _mark_stack(mark_stack),
7682 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7683 mark_stack)
7684 {
7685 assert(_mark_stack->isEmpty(), "stack should be empty");
7686 _finger = _verification_bm->startWord();
7687 assert(_collector->_restart_addr == NULL, "Sanity check");
7688 assert(_span.contains(_finger), "Out of bounds _finger?");
7689 }
7690
7691 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7692 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7693 assert(_span.contains(addr), "Out of bounds _finger?");
7694 _finger = addr;
7695 }
7696
7697 // Should revisit to see if this should be restructured for
7698 // greater efficiency.
7699 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7700 // convert offset into a HeapWord*
7701 HeapWord* addr = _verification_bm->startWord() + offset;
7702 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7703 "address out of range");
7704 assert(_verification_bm->isMarked(addr), "tautology");
7705 assert(_cms_bm->isMarked(addr), "tautology");
7706
7707 assert(_mark_stack->isEmpty(),
7708 "should drain stack to limit stack usage");
7709 // convert addr to an oop preparatory to scanning
7710 oop obj = oop(addr);
7711 assert(obj->is_oop(), "should be an oop");
7712 assert(_finger <= addr, "_finger runneth ahead");
7713 // advance the finger to right end of this object
7714 _finger = addr + obj->size();
7715 assert(_finger > addr, "we just incremented it above");
7716 // Note: the finger doesn't advance while we drain
7717 // the stack below.
7718 bool res = _mark_stack->push(obj);
7719 assert(res, "Empty non-zero size stack should have space for single push");
7720 while (!_mark_stack->isEmpty()) {
7721 oop new_oop = _mark_stack->pop();
7722 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7723 // now scan this oop's oops
7724 new_oop->oop_iterate(&_pam_verify_closure);
7725 }
7726 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7727 return true;
7728 }
7729
7730 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7731 CMSCollector* collector, MemRegion span,
7732 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7733 CMSMarkStack* mark_stack):
7734 CMSOopClosure(collector->ref_processor()),
7735 _collector(collector),
7736 _span(span),
7737 _verification_bm(verification_bm),
7738 _cms_bm(cms_bm),
7739 _mark_stack(mark_stack)
7740 { }
7741
7742 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7743 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7744
7745 // Upon stack overflow, we discard (part of) the stack,
7746 // remembering the least address amongst those discarded
7747 // in CMSCollector's _restart_address.
7748 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7749 // Remember the least grey address discarded
7750 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7751 _collector->lower_restart_addr(ra);
7752 _mark_stack->reset(); // discard stack contents
7753 _mark_stack->expand(); // expand the stack if possible
7754 }
7755
7756 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7757 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7758 HeapWord* addr = (HeapWord*)obj;
7759 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7760 // Oop lies in _span and isn't yet grey or black
7761 _verification_bm->mark(addr); // now grey
7762 if (!_cms_bm->isMarked(addr)) {
7763 oop(addr)->print();
7764 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7765 addr);
7766 fatal("... aborting");
7767 }
7768
7769 if (!_mark_stack->push(obj)) { // stack overflow
7770 if (PrintCMSStatistics != 0) {
7771 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7772 SIZE_FORMAT, _mark_stack->capacity());
7773 }
7774 assert(_mark_stack->isFull(), "Else push should have succeeded");
7775 handle_stack_overflow(addr);
7776 }
7777 // anything including and to the right of _finger
7778 // will be scanned as we iterate over the remainder of the
7779 // bit map
7780 }
7781 }
7782
7783 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7784 MemRegion span,
7785 CMSBitMap* bitMap, CMSMarkStack* markStack,
7786 HeapWord* finger, MarkFromRootsClosure* parent) :
7787 CMSOopClosure(collector->ref_processor()),
7788 _collector(collector),
7789 _span(span),
7790 _bitMap(bitMap),
7791 _markStack(markStack),
7792 _finger(finger),
7793 _parent(parent)
7794 { }
7795
7796 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7797 MemRegion span,
7798 CMSBitMap* bit_map,
7799 OopTaskQueue* work_queue,
7800 CMSMarkStack* overflow_stack,
7801 HeapWord* finger,
7802 HeapWord** global_finger_addr,
7803 Par_MarkFromRootsClosure* parent) :
7804 CMSOopClosure(collector->ref_processor()),
7805 _collector(collector),
7806 _whole_span(collector->_span),
7807 _span(span),
7808 _bit_map(bit_map),
7809 _work_queue(work_queue),
7810 _overflow_stack(overflow_stack),
7811 _finger(finger),
7812 _global_finger_addr(global_finger_addr),
7813 _parent(parent)
7814 { }
7815
7816 // Assumes thread-safe access by callers, who are
7817 // responsible for mutual exclusion.
7818 void CMSCollector::lower_restart_addr(HeapWord* low) {
7819 assert(_span.contains(low), "Out of bounds addr");
7820 if (_restart_addr == NULL) {
7821 _restart_addr = low;
7822 } else {
7823 _restart_addr = MIN2(_restart_addr, low);
7824 }
7825 }
7826
7827 // Upon stack overflow, we discard (part of) the stack,
7828 // remembering the least address amongst those discarded
7829 // in CMSCollector's _restart_address.
7830 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7831 // Remember the least grey address discarded
7832 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7833 _collector->lower_restart_addr(ra);
7834 _markStack->reset(); // discard stack contents
7835 _markStack->expand(); // expand the stack if possible
7836 }
7837
7838 // Upon stack overflow, we discard (part of) the stack,
7839 // remembering the least address amongst those discarded
7840 // in CMSCollector's _restart_address.
7841 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7842 // We need to do this under a mutex to prevent other
7843 // workers from interfering with the work done below.
7844 MutexLockerEx ml(_overflow_stack->par_lock(),
7845 Mutex::_no_safepoint_check_flag);
7846 // Remember the least grey address discarded
7847 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7848 _collector->lower_restart_addr(ra);
7849 _overflow_stack->reset(); // discard stack contents
7850 _overflow_stack->expand(); // expand the stack if possible
7851 }
7852
7853 void CMKlassClosure::do_klass(Klass* k) {
7854 assert(_oop_closure != NULL, "Not initialized?");
7855 k->oops_do(_oop_closure);
7856 }
7857
7858 void PushOrMarkClosure::do_oop(oop obj) {
7859 // Ignore mark word because we are running concurrent with mutators.
7860 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7861 HeapWord* addr = (HeapWord*)obj;
7862 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7863 // Oop lies in _span and isn't yet grey or black
7864 _bitMap->mark(addr); // now grey
7865 if (addr < _finger) {
7866 // the bit map iteration has already either passed, or
7867 // sampled, this bit in the bit map; we'll need to
7868 // use the marking stack to scan this oop's oops.
7869 bool simulate_overflow = false;
7870 NOT_PRODUCT(
7871 if (CMSMarkStackOverflowALot &&
7872 _collector->simulate_overflow()) {
7873 // simulate a stack overflow
7874 simulate_overflow = true;
7875 }
7876 )
7877 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7878 if (PrintCMSStatistics != 0) {
7879 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7880 SIZE_FORMAT, _markStack->capacity());
7881 }
7882 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7883 handle_stack_overflow(addr);
7884 }
7885 }
7886 // anything including and to the right of _finger
7887 // will be scanned as we iterate over the remainder of the
7888 // bit map
7889 do_yield_check();
7890 }
7891 }
7892
7893 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7894 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7895
7896 void Par_PushOrMarkClosure::do_oop(oop obj) {
7897 // Ignore mark word because we are running concurrent with mutators.
7898 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7899 HeapWord* addr = (HeapWord*)obj;
7900 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7901 // Oop lies in _span and isn't yet grey or black
7902 // We read the global_finger (volatile read) strictly after marking oop
7903 bool res = _bit_map->par_mark(addr); // now grey
7904 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7905 // Should we push this marked oop on our stack?
7906 // -- if someone else marked it, nothing to do
7907 // -- if target oop is above global finger nothing to do
7908 // -- if target oop is in chunk and above local finger
7909 // then nothing to do
7910 // -- else push on work queue
7911 if ( !res // someone else marked it, they will deal with it
7912 || (addr >= *gfa) // will be scanned in a later task
7913 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7914 return;
7915 }
7916 // the bit map iteration has already either passed, or
7917 // sampled, this bit in the bit map; we'll need to
7918 // use the marking stack to scan this oop's oops.
7919 bool simulate_overflow = false;
7920 NOT_PRODUCT(
7921 if (CMSMarkStackOverflowALot &&
7922 _collector->simulate_overflow()) {
7923 // simulate a stack overflow
7924 simulate_overflow = true;
7925 }
7926 )
7927 if (simulate_overflow ||
7928 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7929 // stack overflow
7930 if (PrintCMSStatistics != 0) {
7931 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7932 SIZE_FORMAT, _overflow_stack->capacity());
7933 }
7934 // We cannot assert that the overflow stack is full because
7935 // it may have been emptied since.
7936 assert(simulate_overflow ||
7937 _work_queue->size() == _work_queue->max_elems(),
7938 "Else push should have succeeded");
7939 handle_stack_overflow(addr);
7940 }
7941 do_yield_check();
7942 }
7943 }
7944
7945 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7946 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7947
7948 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7949 MemRegion span,
7950 ReferenceProcessor* rp,
7951 CMSBitMap* bit_map,
7952 CMSBitMap* mod_union_table,
7953 CMSMarkStack* mark_stack,
7954 bool concurrent_precleaning):
7955 CMSOopClosure(rp),
7956 _collector(collector),
7957 _span(span),
7958 _bit_map(bit_map),
7959 _mod_union_table(mod_union_table),
7960 _mark_stack(mark_stack),
7961 _concurrent_precleaning(concurrent_precleaning)
7962 {
7963 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7964 }
7965
7966 // Grey object rescan during pre-cleaning and second checkpoint phases --
7967 // the non-parallel version (the parallel version appears further below.)
7968 void PushAndMarkClosure::do_oop(oop obj) {
7969 // Ignore mark word verification. If during concurrent precleaning,
7970 // the object monitor may be locked. If during the checkpoint
7971 // phases, the object may already have been reached by a different
7972 // path and may be at the end of the global overflow list (so
7973 // the mark word may be NULL).
7974 assert(obj->is_oop_or_null(true /* ignore mark word */),
7975 "expected an oop or NULL");
7976 HeapWord* addr = (HeapWord*)obj;
7977 // Check if oop points into the CMS generation
7978 // and is not marked
7979 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7980 // a white object ...
7981 _bit_map->mark(addr); // ... now grey
7982 // push on the marking stack (grey set)
7983 bool simulate_overflow = false;
7984 NOT_PRODUCT(
7985 if (CMSMarkStackOverflowALot &&
7986 _collector->simulate_overflow()) {
7987 // simulate a stack overflow
7988 simulate_overflow = true;
7989 }
7990 )
7991 if (simulate_overflow || !_mark_stack->push(obj)) {
7992 if (_concurrent_precleaning) {
7993 // During precleaning we can just dirty the appropriate card(s)
7994 // in the mod union table, thus ensuring that the object remains
7995 // in the grey set and continue. In the case of object arrays
7996 // we need to dirty all of the cards that the object spans,
7997 // since the rescan of object arrays will be limited to the
7998 // dirty cards.
7999 // Note that no one can be interfering with us in this action
8000 // of dirtying the mod union table, so no locking or atomics
8001 // are required.
8002 if (obj->is_objArray()) {
8003 size_t sz = obj->size();
8004 HeapWord* end_card_addr = (HeapWord*)round_to(
8005 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
8006 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8007 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8008 _mod_union_table->mark_range(redirty_range);
8009 } else {
8010 _mod_union_table->mark(addr);
8011 }
8012 _collector->_ser_pmc_preclean_ovflw++;
8013 } else {
8014 // During the remark phase, we need to remember this oop
8015 // in the overflow list.
8016 _collector->push_on_overflow_list(obj);
8017 _collector->_ser_pmc_remark_ovflw++;
8018 }
8019 }
8020 }
8021 }
8022
8023 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
8024 MemRegion span,
8025 ReferenceProcessor* rp,
8026 CMSBitMap* bit_map,
8027 OopTaskQueue* work_queue):
8028 CMSOopClosure(rp),
8029 _collector(collector),
8030 _span(span),
8031 _bit_map(bit_map),
8032 _work_queue(work_queue)
8033 {
8034 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
8035 }
8036
8037 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
8038 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
8039
8040 // Grey object rescan during second checkpoint phase --
8041 // the parallel version.
8042 void Par_PushAndMarkClosure::do_oop(oop obj) {
8043 // In the assert below, we ignore the mark word because
8044 // this oop may point to an already visited object that is
8045 // on the overflow stack (in which case the mark word has
8046 // been hijacked for chaining into the overflow stack --
8047 // if this is the last object in the overflow stack then
8048 // its mark word will be NULL). Because this object may
8049 // have been subsequently popped off the global overflow
8050 // stack, and the mark word possibly restored to the prototypical
8051 // value, by the time we get to examined this failing assert in
8052 // the debugger, is_oop_or_null(false) may subsequently start
8053 // to hold.
8054 assert(obj->is_oop_or_null(true),
8055 "expected an oop or NULL");
8056 HeapWord* addr = (HeapWord*)obj;
8057 // Check if oop points into the CMS generation
8058 // and is not marked
8059 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
8060 // a white object ...
8061 // If we manage to "claim" the object, by being the
8062 // first thread to mark it, then we push it on our
8063 // marking stack
8064 if (_bit_map->par_mark(addr)) { // ... now grey
8065 // push on work queue (grey set)
8066 bool simulate_overflow = false;
8067 NOT_PRODUCT(
8068 if (CMSMarkStackOverflowALot &&
8069 _collector->par_simulate_overflow()) {
8070 // simulate a stack overflow
8071 simulate_overflow = true;
8072 }
8073 )
8074 if (simulate_overflow || !_work_queue->push(obj)) {
8075 _collector->par_push_on_overflow_list(obj);
8076 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
8077 }
8078 } // Else, some other thread got there first
8079 }
8080 }
8081
8082 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8083 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8084
8085 void CMSPrecleanRefsYieldClosure::do_yield_work() {
8086 Mutex* bml = _collector->bitMapLock();
8087 assert_lock_strong(bml);
8088 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8089 "CMS thread should hold CMS token");
8090
8091 bml->unlock();
8092 ConcurrentMarkSweepThread::desynchronize(true);
8093
8094 ConcurrentMarkSweepThread::acknowledge_yield_request();
8095
8096 _collector->stopTimer();
8097 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8098 if (PrintCMSStatistics != 0) {
8099 _collector->incrementYields();
8100 }
8101 _collector->icms_wait();
8102
8103 // See the comment in coordinator_yield()
8104 for (unsigned i = 0; i < CMSYieldSleepCount &&
8105 ConcurrentMarkSweepThread::should_yield() &&
8106 !CMSCollector::foregroundGCIsActive(); ++i) {
8107 os::sleep(Thread::current(), 1, false);
8108 ConcurrentMarkSweepThread::acknowledge_yield_request();
8109 }
8110
8111 ConcurrentMarkSweepThread::synchronize(true);
8112 bml->lock();
8113
8114 _collector->startTimer();
8115 }
8116
8117 bool CMSPrecleanRefsYieldClosure::should_return() {
8118 if (ConcurrentMarkSweepThread::should_yield()) {
8119 do_yield_work();
8120 }
8121 return _collector->foregroundGCIsActive();
8122 }
8123
8124 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
8125 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
8126 "mr should be aligned to start at a card boundary");
8127 // We'd like to assert:
8128 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
8129 // "mr should be a range of cards");
8130 // However, that would be too strong in one case -- the last
8131 // partition ends at _unallocated_block which, in general, can be
8132 // an arbitrary boundary, not necessarily card aligned.
8133 if (PrintCMSStatistics != 0) {
8134 _num_dirty_cards +=
8135 mr.word_size()/CardTableModRefBS::card_size_in_words;
8136 }
8137 _space->object_iterate_mem(mr, &_scan_cl);
8138 }
8139
8140 SweepClosure::SweepClosure(CMSCollector* collector,
8141 ConcurrentMarkSweepGeneration* g,
8142 CMSBitMap* bitMap, bool should_yield) :
8143 _collector(collector),
8144 _g(g),
8145 _sp(g->cmsSpace()),
8146 _limit(_sp->sweep_limit()),
8147 _freelistLock(_sp->freelistLock()),
8148 _bitMap(bitMap),
8149 _yield(should_yield),
8150 _inFreeRange(false), // No free range at beginning of sweep
8151 _freeRangeInFreeLists(false), // No free range at beginning of sweep
8152 _lastFreeRangeCoalesced(false),
8153 _freeFinger(g->used_region().start())
8154 {
8155 NOT_PRODUCT(
8156 _numObjectsFreed = 0;
8157 _numWordsFreed = 0;
8158 _numObjectsLive = 0;
8159 _numWordsLive = 0;
8160 _numObjectsAlreadyFree = 0;
8161 _numWordsAlreadyFree = 0;
8162 _last_fc = NULL;
8163
8164 _sp->initializeIndexedFreeListArrayReturnedBytes();
8165 _sp->dictionary()->initialize_dict_returned_bytes();
8166 )
8167 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8168 "sweep _limit out of bounds");
8169 if (CMSTraceSweeper) {
8170 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
8171 _limit);
8172 }
8173 }
8174
8175 void SweepClosure::print_on(outputStream* st) const {
8176 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
8177 _sp->bottom(), _sp->end());
8178 tty->print_cr("_limit = " PTR_FORMAT, _limit);
8179 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
8180 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
8181 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
8182 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
8183 }
8184
8185 #ifndef PRODUCT
8186 // Assertion checking only: no useful work in product mode --
8187 // however, if any of the flags below become product flags,
8188 // you may need to review this code to see if it needs to be
8189 // enabled in product mode.
8190 SweepClosure::~SweepClosure() {
8191 assert_lock_strong(_freelistLock);
8192 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8193 "sweep _limit out of bounds");
8194 if (inFreeRange()) {
8195 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
8196 print();
8197 ShouldNotReachHere();
8198 }
8199 if (Verbose && PrintGC) {
8200 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
8201 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
8202 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
8203 SIZE_FORMAT" bytes "
8204 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
8205 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
8206 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
8207 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
8208 * sizeof(HeapWord);
8209 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
8210
8211 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
8212 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
8213 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
8214 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
8215 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
8216 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
8217 indexListReturnedBytes);
8218 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
8219 dict_returned_bytes);
8220 }
8221 }
8222 if (CMSTraceSweeper) {
8223 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
8224 _limit);
8225 }
8226 }
8227 #endif // PRODUCT
8228
8229 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8230 bool freeRangeInFreeLists) {
8231 if (CMSTraceSweeper) {
8232 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
8233 freeFinger, freeRangeInFreeLists);
8234 }
8235 assert(!inFreeRange(), "Trampling existing free range");
8236 set_inFreeRange(true);
8237 set_lastFreeRangeCoalesced(false);
8238
8239 set_freeFinger(freeFinger);
8240 set_freeRangeInFreeLists(freeRangeInFreeLists);
8241 if (CMSTestInFreeList) {
8242 if (freeRangeInFreeLists) {
8243 FreeChunk* fc = (FreeChunk*) freeFinger;
8244 assert(fc->is_free(), "A chunk on the free list should be free.");
8245 assert(fc->size() > 0, "Free range should have a size");
8246 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8247 }
8248 }
8249 }
8250
8251 // Note that the sweeper runs concurrently with mutators. Thus,
8252 // it is possible for direct allocation in this generation to happen
8253 // in the middle of the sweep. Note that the sweeper also coalesces
8254 // contiguous free blocks. Thus, unless the sweeper and the allocator
8255 // synchronize appropriately freshly allocated blocks may get swept up.
8256 // This is accomplished by the sweeper locking the free lists while
8257 // it is sweeping. Thus blocks that are determined to be free are
8258 // indeed free. There is however one additional complication:
8259 // blocks that have been allocated since the final checkpoint and
8260 // mark, will not have been marked and so would be treated as
8261 // unreachable and swept up. To prevent this, the allocator marks
8262 // the bit map when allocating during the sweep phase. This leads,
8263 // however, to a further complication -- objects may have been allocated
8264 // but not yet initialized -- in the sense that the header isn't yet
8265 // installed. The sweeper can not then determine the size of the block
8266 // in order to skip over it. To deal with this case, we use a technique
8267 // (due to Printezis) to encode such uninitialized block sizes in the
8268 // bit map. Since the bit map uses a bit per every HeapWord, but the
8269 // CMS generation has a minimum object size of 3 HeapWords, it follows
8270 // that "normal marks" won't be adjacent in the bit map (there will
8271 // always be at least two 0 bits between successive 1 bits). We make use
8272 // of these "unused" bits to represent uninitialized blocks -- the bit
8273 // corresponding to the start of the uninitialized object and the next
8274 // bit are both set. Finally, a 1 bit marks the end of the object that
8275 // started with the two consecutive 1 bits to indicate its potentially
8276 // uninitialized state.
8277
8278 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8279 FreeChunk* fc = (FreeChunk*)addr;
8280 size_t res;
8281
8282 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8283 // than "addr == _limit" because although _limit was a block boundary when
8284 // we started the sweep, it may no longer be one because heap expansion
8285 // may have caused us to coalesce the block ending at the address _limit
8286 // with a newly expanded chunk (this happens when _limit was set to the
8287 // previous _end of the space), so we may have stepped past _limit:
8288 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8289 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8290 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8291 "sweep _limit out of bounds");
8292 assert(addr < _sp->end(), "addr out of bounds");
8293 // Flush any free range we might be holding as a single
8294 // coalesced chunk to the appropriate free list.
8295 if (inFreeRange()) {
8296 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8297 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8298 flush_cur_free_chunk(freeFinger(),
8299 pointer_delta(addr, freeFinger()));
8300 if (CMSTraceSweeper) {
8301 gclog_or_tty->print("Sweep: last chunk: ");
8302 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8303 "[coalesced:"SIZE_FORMAT"]\n",
8304 freeFinger(), pointer_delta(addr, freeFinger()),
8305 lastFreeRangeCoalesced());
8306 }
8307 }
8308
8309 // help the iterator loop finish
8310 return pointer_delta(_sp->end(), addr);
8311 }
8312
8313 assert(addr < _limit, "sweep invariant");
8314 // check if we should yield
8315 do_yield_check(addr);
8316 if (fc->is_free()) {
8317 // Chunk that is already free
8318 res = fc->size();
8319 do_already_free_chunk(fc);
8320 debug_only(_sp->verifyFreeLists());
8321 // If we flush the chunk at hand in lookahead_and_flush()
8322 // and it's coalesced with a preceding chunk, then the
8323 // process of "mangling" the payload of the coalesced block
8324 // will cause erasure of the size information from the
8325 // (erstwhile) header of all the coalesced blocks but the
8326 // first, so the first disjunct in the assert will not hold
8327 // in that specific case (in which case the second disjunct
8328 // will hold).
8329 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8330 "Otherwise the size info doesn't change at this step");
8331 NOT_PRODUCT(
8332 _numObjectsAlreadyFree++;
8333 _numWordsAlreadyFree += res;
8334 )
8335 NOT_PRODUCT(_last_fc = fc;)
8336 } else if (!_bitMap->isMarked(addr)) {
8337 // Chunk is fresh garbage
8338 res = do_garbage_chunk(fc);
8339 debug_only(_sp->verifyFreeLists());
8340 NOT_PRODUCT(
8341 _numObjectsFreed++;
8342 _numWordsFreed += res;
8343 )
8344 } else {
8345 // Chunk that is alive.
8346 res = do_live_chunk(fc);
8347 debug_only(_sp->verifyFreeLists());
8348 NOT_PRODUCT(
8349 _numObjectsLive++;
8350 _numWordsLive += res;
8351 )
8352 }
8353 return res;
8354 }
8355
8356 // For the smart allocation, record following
8357 // split deaths - a free chunk is removed from its free list because
8358 // it is being split into two or more chunks.
8359 // split birth - a free chunk is being added to its free list because
8360 // a larger free chunk has been split and resulted in this free chunk.
8361 // coal death - a free chunk is being removed from its free list because
8362 // it is being coalesced into a large free chunk.
8363 // coal birth - a free chunk is being added to its free list because
8364 // it was created when two or more free chunks where coalesced into
8365 // this free chunk.
8366 //
8367 // These statistics are used to determine the desired number of free
8368 // chunks of a given size. The desired number is chosen to be relative
8369 // to the end of a CMS sweep. The desired number at the end of a sweep
8370 // is the
8371 // count-at-end-of-previous-sweep (an amount that was enough)
8372 // - count-at-beginning-of-current-sweep (the excess)
8373 // + split-births (gains in this size during interval)
8374 // - split-deaths (demands on this size during interval)
8375 // where the interval is from the end of one sweep to the end of the
8376 // next.
8377 //
8378 // When sweeping the sweeper maintains an accumulated chunk which is
8379 // the chunk that is made up of chunks that have been coalesced. That
8380 // will be termed the left-hand chunk. A new chunk of garbage that
8381 // is being considered for coalescing will be referred to as the
8382 // right-hand chunk.
8383 //
8384 // When making a decision on whether to coalesce a right-hand chunk with
8385 // the current left-hand chunk, the current count vs. the desired count
8386 // of the left-hand chunk is considered. Also if the right-hand chunk
8387 // is near the large chunk at the end of the heap (see
8388 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8389 // left-hand chunk is coalesced.
8390 //
8391 // When making a decision about whether to split a chunk, the desired count
8392 // vs. the current count of the candidate to be split is also considered.
8393 // If the candidate is underpopulated (currently fewer chunks than desired)
8394 // a chunk of an overpopulated (currently more chunks than desired) size may
8395 // be chosen. The "hint" associated with a free list, if non-null, points
8396 // to a free list which may be overpopulated.
8397 //
8398
8399 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8400 const size_t size = fc->size();
8401 // Chunks that cannot be coalesced are not in the
8402 // free lists.
8403 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8404 assert(_sp->verify_chunk_in_free_list(fc),
8405 "free chunk should be in free lists");
8406 }
8407 // a chunk that is already free, should not have been
8408 // marked in the bit map
8409 HeapWord* const addr = (HeapWord*) fc;
8410 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8411 // Verify that the bit map has no bits marked between
8412 // addr and purported end of this block.
8413 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8414
8415 // Some chunks cannot be coalesced under any circumstances.
8416 // See the definition of cantCoalesce().
8417 if (!fc->cantCoalesce()) {
8418 // This chunk can potentially be coalesced.
8419 if (_sp->adaptive_freelists()) {
8420 // All the work is done in
8421 do_post_free_or_garbage_chunk(fc, size);
8422 } else { // Not adaptive free lists
8423 // this is a free chunk that can potentially be coalesced by the sweeper;
8424 if (!inFreeRange()) {
8425 // if the next chunk is a free block that can't be coalesced
8426 // it doesn't make sense to remove this chunk from the free lists
8427 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8428 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8429 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8430 nextChunk->is_free() && // ... which is free...
8431 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8432 // nothing to do
8433 } else {
8434 // Potentially the start of a new free range:
8435 // Don't eagerly remove it from the free lists.
8436 // No need to remove it if it will just be put
8437 // back again. (Also from a pragmatic point of view
8438 // if it is a free block in a region that is beyond
8439 // any allocated blocks, an assertion will fail)
8440 // Remember the start of a free run.
8441 initialize_free_range(addr, true);
8442 // end - can coalesce with next chunk
8443 }
8444 } else {
8445 // the midst of a free range, we are coalescing
8446 print_free_block_coalesced(fc);
8447 if (CMSTraceSweeper) {
8448 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8449 }
8450 // remove it from the free lists
8451 _sp->removeFreeChunkFromFreeLists(fc);
8452 set_lastFreeRangeCoalesced(true);
8453 // If the chunk is being coalesced and the current free range is
8454 // in the free lists, remove the current free range so that it
8455 // will be returned to the free lists in its entirety - all
8456 // the coalesced pieces included.
8457 if (freeRangeInFreeLists()) {
8458 FreeChunk* ffc = (FreeChunk*) freeFinger();
8459 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8460 "Size of free range is inconsistent with chunk size.");
8461 if (CMSTestInFreeList) {
8462 assert(_sp->verify_chunk_in_free_list(ffc),
8463 "free range is not in free lists");
8464 }
8465 _sp->removeFreeChunkFromFreeLists(ffc);
8466 set_freeRangeInFreeLists(false);
8467 }
8468 }
8469 }
8470 // Note that if the chunk is not coalescable (the else arm
8471 // below), we unconditionally flush, without needing to do
8472 // a "lookahead," as we do below.
8473 if (inFreeRange()) lookahead_and_flush(fc, size);
8474 } else {
8475 // Code path common to both original and adaptive free lists.
8476
8477 // cant coalesce with previous block; this should be treated
8478 // as the end of a free run if any
8479 if (inFreeRange()) {
8480 // we kicked some butt; time to pick up the garbage
8481 assert(freeFinger() < addr, "freeFinger points too high");
8482 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8483 }
8484 // else, nothing to do, just continue
8485 }
8486 }
8487
8488 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8489 // This is a chunk of garbage. It is not in any free list.
8490 // Add it to a free list or let it possibly be coalesced into
8491 // a larger chunk.
8492 HeapWord* const addr = (HeapWord*) fc;
8493 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8494
8495 if (_sp->adaptive_freelists()) {
8496 // Verify that the bit map has no bits marked between
8497 // addr and purported end of just dead object.
8498 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8499
8500 do_post_free_or_garbage_chunk(fc, size);
8501 } else {
8502 if (!inFreeRange()) {
8503 // start of a new free range
8504 assert(size > 0, "A free range should have a size");
8505 initialize_free_range(addr, false);
8506 } else {
8507 // this will be swept up when we hit the end of the
8508 // free range
8509 if (CMSTraceSweeper) {
8510 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8511 }
8512 // If the chunk is being coalesced and the current free range is
8513 // in the free lists, remove the current free range so that it
8514 // will be returned to the free lists in its entirety - all
8515 // the coalesced pieces included.
8516 if (freeRangeInFreeLists()) {
8517 FreeChunk* ffc = (FreeChunk*)freeFinger();
8518 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8519 "Size of free range is inconsistent with chunk size.");
8520 if (CMSTestInFreeList) {
8521 assert(_sp->verify_chunk_in_free_list(ffc),
8522 "free range is not in free lists");
8523 }
8524 _sp->removeFreeChunkFromFreeLists(ffc);
8525 set_freeRangeInFreeLists(false);
8526 }
8527 set_lastFreeRangeCoalesced(true);
8528 }
8529 // this will be swept up when we hit the end of the free range
8530
8531 // Verify that the bit map has no bits marked between
8532 // addr and purported end of just dead object.
8533 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8534 }
8535 assert(_limit >= addr + size,
8536 "A freshly garbage chunk can't possibly straddle over _limit");
8537 if (inFreeRange()) lookahead_and_flush(fc, size);
8538 return size;
8539 }
8540
8541 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8542 HeapWord* addr = (HeapWord*) fc;
8543 // The sweeper has just found a live object. Return any accumulated
8544 // left hand chunk to the free lists.
8545 if (inFreeRange()) {
8546 assert(freeFinger() < addr, "freeFinger points too high");
8547 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8548 }
8549
8550 // This object is live: we'd normally expect this to be
8551 // an oop, and like to assert the following:
8552 // assert(oop(addr)->is_oop(), "live block should be an oop");
8553 // However, as we commented above, this may be an object whose
8554 // header hasn't yet been initialized.
8555 size_t size;
8556 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8557 if (_bitMap->isMarked(addr + 1)) {
8558 // Determine the size from the bit map, rather than trying to
8559 // compute it from the object header.
8560 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8561 size = pointer_delta(nextOneAddr + 1, addr);
8562 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8563 "alignment problem");
8564
8565 #ifdef ASSERT
8566 if (oop(addr)->klass_or_null() != NULL) {
8567 // Ignore mark word because we are running concurrent with mutators
8568 assert(oop(addr)->is_oop(true), "live block should be an oop");
8569 assert(size ==
8570 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8571 "P-mark and computed size do not agree");
8572 }
8573 #endif
8574
8575 } else {
8576 // This should be an initialized object that's alive.
8577 assert(oop(addr)->klass_or_null() != NULL,
8578 "Should be an initialized object");
8579 // Ignore mark word because we are running concurrent with mutators
8580 assert(oop(addr)->is_oop(true), "live block should be an oop");
8581 // Verify that the bit map has no bits marked between
8582 // addr and purported end of this block.
8583 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8584 assert(size >= 3, "Necessary for Printezis marks to work");
8585 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8586 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8587 }
8588 return size;
8589 }
8590
8591 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8592 size_t chunkSize) {
8593 // do_post_free_or_garbage_chunk() should only be called in the case
8594 // of the adaptive free list allocator.
8595 const bool fcInFreeLists = fc->is_free();
8596 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8597 assert((HeapWord*)fc <= _limit, "sweep invariant");
8598 if (CMSTestInFreeList && fcInFreeLists) {
8599 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8600 }
8601
8602 if (CMSTraceSweeper) {
8603 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8604 }
8605
8606 HeapWord* const fc_addr = (HeapWord*) fc;
8607
8608 bool coalesce;
8609 const size_t left = pointer_delta(fc_addr, freeFinger());
8610 const size_t right = chunkSize;
8611 switch (FLSCoalescePolicy) {
8612 // numeric value forms a coalition aggressiveness metric
8613 case 0: { // never coalesce
8614 coalesce = false;
8615 break;
8616 }
8617 case 1: { // coalesce if left & right chunks on overpopulated lists
8618 coalesce = _sp->coalOverPopulated(left) &&
8619 _sp->coalOverPopulated(right);
8620 break;
8621 }
8622 case 2: { // coalesce if left chunk on overpopulated list (default)
8623 coalesce = _sp->coalOverPopulated(left);
8624 break;
8625 }
8626 case 3: { // coalesce if left OR right chunk on overpopulated list
8627 coalesce = _sp->coalOverPopulated(left) ||
8628 _sp->coalOverPopulated(right);
8629 break;
8630 }
8631 case 4: { // always coalesce
8632 coalesce = true;
8633 break;
8634 }
8635 default:
8636 ShouldNotReachHere();
8637 }
8638
8639 // Should the current free range be coalesced?
8640 // If the chunk is in a free range and either we decided to coalesce above
8641 // or the chunk is near the large block at the end of the heap
8642 // (isNearLargestChunk() returns true), then coalesce this chunk.
8643 const bool doCoalesce = inFreeRange()
8644 && (coalesce || _g->isNearLargestChunk(fc_addr));
8645 if (doCoalesce) {
8646 // Coalesce the current free range on the left with the new
8647 // chunk on the right. If either is on a free list,
8648 // it must be removed from the list and stashed in the closure.
8649 if (freeRangeInFreeLists()) {
8650 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8651 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8652 "Size of free range is inconsistent with chunk size.");
8653 if (CMSTestInFreeList) {
8654 assert(_sp->verify_chunk_in_free_list(ffc),
8655 "Chunk is not in free lists");
8656 }
8657 _sp->coalDeath(ffc->size());
8658 _sp->removeFreeChunkFromFreeLists(ffc);
8659 set_freeRangeInFreeLists(false);
8660 }
8661 if (fcInFreeLists) {
8662 _sp->coalDeath(chunkSize);
8663 assert(fc->size() == chunkSize,
8664 "The chunk has the wrong size or is not in the free lists");
8665 _sp->removeFreeChunkFromFreeLists(fc);
8666 }
8667 set_lastFreeRangeCoalesced(true);
8668 print_free_block_coalesced(fc);
8669 } else { // not in a free range and/or should not coalesce
8670 // Return the current free range and start a new one.
8671 if (inFreeRange()) {
8672 // In a free range but cannot coalesce with the right hand chunk.
8673 // Put the current free range into the free lists.
8674 flush_cur_free_chunk(freeFinger(),
8675 pointer_delta(fc_addr, freeFinger()));
8676 }
8677 // Set up for new free range. Pass along whether the right hand
8678 // chunk is in the free lists.
8679 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8680 }
8681 }
8682
8683 // Lookahead flush:
8684 // If we are tracking a free range, and this is the last chunk that
8685 // we'll look at because its end crosses past _limit, we'll preemptively
8686 // flush it along with any free range we may be holding on to. Note that
8687 // this can be the case only for an already free or freshly garbage
8688 // chunk. If this block is an object, it can never straddle
8689 // over _limit. The "straddling" occurs when _limit is set at
8690 // the previous end of the space when this cycle started, and
8691 // a subsequent heap expansion caused the previously co-terminal
8692 // free block to be coalesced with the newly expanded portion,
8693 // thus rendering _limit a non-block-boundary making it dangerous
8694 // for the sweeper to step over and examine.
8695 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8696 assert(inFreeRange(), "Should only be called if currently in a free range.");
8697 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8698 assert(_sp->used_region().contains(eob - 1),
8699 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
8700 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8701 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8702 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8703 if (eob >= _limit) {
8704 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8705 if (CMSTraceSweeper) {
8706 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8707 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8708 "[" PTR_FORMAT "," PTR_FORMAT ")",
8709 _limit, fc, eob, _sp->bottom(), _sp->end());
8710 }
8711 // Return the storage we are tracking back into the free lists.
8712 if (CMSTraceSweeper) {
8713 gclog_or_tty->print_cr("Flushing ... ");
8714 }
8715 assert(freeFinger() < eob, "Error");
8716 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8717 }
8718 }
8719
8720 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8721 assert(inFreeRange(), "Should only be called if currently in a free range.");
8722 assert(size > 0,
8723 "A zero sized chunk cannot be added to the free lists.");
8724 if (!freeRangeInFreeLists()) {
8725 if (CMSTestInFreeList) {
8726 FreeChunk* fc = (FreeChunk*) chunk;
8727 fc->set_size(size);
8728 assert(!_sp->verify_chunk_in_free_list(fc),
8729 "chunk should not be in free lists yet");
8730 }
8731 if (CMSTraceSweeper) {
8732 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8733 chunk, size);
8734 }
8735 // A new free range is going to be starting. The current
8736 // free range has not been added to the free lists yet or
8737 // was removed so add it back.
8738 // If the current free range was coalesced, then the death
8739 // of the free range was recorded. Record a birth now.
8740 if (lastFreeRangeCoalesced()) {
8741 _sp->coalBirth(size);
8742 }
8743 _sp->addChunkAndRepairOffsetTable(chunk, size,
8744 lastFreeRangeCoalesced());
8745 } else if (CMSTraceSweeper) {
8746 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8747 }
8748 set_inFreeRange(false);
8749 set_freeRangeInFreeLists(false);
8750 }
8751
8752 // We take a break if we've been at this for a while,
8753 // so as to avoid monopolizing the locks involved.
8754 void SweepClosure::do_yield_work(HeapWord* addr) {
8755 // Return current free chunk being used for coalescing (if any)
8756 // to the appropriate freelist. After yielding, the next
8757 // free block encountered will start a coalescing range of
8758 // free blocks. If the next free block is adjacent to the
8759 // chunk just flushed, they will need to wait for the next
8760 // sweep to be coalesced.
8761 if (inFreeRange()) {
8762 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8763 }
8764
8765 // First give up the locks, then yield, then re-lock.
8766 // We should probably use a constructor/destructor idiom to
8767 // do this unlock/lock or modify the MutexUnlocker class to
8768 // serve our purpose. XXX
8769 assert_lock_strong(_bitMap->lock());
8770 assert_lock_strong(_freelistLock);
8771 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8772 "CMS thread should hold CMS token");
8773 _bitMap->lock()->unlock();
8774 _freelistLock->unlock();
8775 ConcurrentMarkSweepThread::desynchronize(true);
8776 ConcurrentMarkSweepThread::acknowledge_yield_request();
8777 _collector->stopTimer();
8778 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8779 if (PrintCMSStatistics != 0) {
8780 _collector->incrementYields();
8781 }
8782 _collector->icms_wait();
8783
8784 // See the comment in coordinator_yield()
8785 for (unsigned i = 0; i < CMSYieldSleepCount &&
8786 ConcurrentMarkSweepThread::should_yield() &&
8787 !CMSCollector::foregroundGCIsActive(); ++i) {
8788 os::sleep(Thread::current(), 1, false);
8789 ConcurrentMarkSweepThread::acknowledge_yield_request();
8790 }
8791
8792 ConcurrentMarkSweepThread::synchronize(true);
8793 _freelistLock->lock();
8794 _bitMap->lock()->lock_without_safepoint_check();
8795 _collector->startTimer();
8796 }
8797
8798 #ifndef PRODUCT
8799 // This is actually very useful in a product build if it can
8800 // be called from the debugger. Compile it into the product
8801 // as needed.
8802 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8803 return debug_cms_space->verify_chunk_in_free_list(fc);
8804 }
8805 #endif
8806
8807 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8808 if (CMSTraceSweeper) {
8809 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8810 fc, fc->size());
8811 }
8812 }
8813
8814 // CMSIsAliveClosure
8815 bool CMSIsAliveClosure::do_object_b(oop obj) {
8816 HeapWord* addr = (HeapWord*)obj;
8817 return addr != NULL &&
8818 (!_span.contains(addr) || _bit_map->isMarked(addr));
8819 }
8820
8821
8822 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8823 MemRegion span,
8824 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8825 bool cpc):
8826 _collector(collector),
8827 _span(span),
8828 _bit_map(bit_map),
8829 _mark_stack(mark_stack),
8830 _concurrent_precleaning(cpc) {
8831 assert(!_span.is_empty(), "Empty span could spell trouble");
8832 }
8833
8834
8835 // CMSKeepAliveClosure: the serial version
8836 void CMSKeepAliveClosure::do_oop(oop obj) {
8837 HeapWord* addr = (HeapWord*)obj;
8838 if (_span.contains(addr) &&
8839 !_bit_map->isMarked(addr)) {
8840 _bit_map->mark(addr);
8841 bool simulate_overflow = false;
8842 NOT_PRODUCT(
8843 if (CMSMarkStackOverflowALot &&
8844 _collector->simulate_overflow()) {
8845 // simulate a stack overflow
8846 simulate_overflow = true;
8847 }
8848 )
8849 if (simulate_overflow || !_mark_stack->push(obj)) {
8850 if (_concurrent_precleaning) {
8851 // We dirty the overflown object and let the remark
8852 // phase deal with it.
8853 assert(_collector->overflow_list_is_empty(), "Error");
8854 // In the case of object arrays, we need to dirty all of
8855 // the cards that the object spans. No locking or atomics
8856 // are needed since no one else can be mutating the mod union
8857 // table.
8858 if (obj->is_objArray()) {
8859 size_t sz = obj->size();
8860 HeapWord* end_card_addr =
8861 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8862 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8863 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8864 _collector->_modUnionTable.mark_range(redirty_range);
8865 } else {
8866 _collector->_modUnionTable.mark(addr);
8867 }
8868 _collector->_ser_kac_preclean_ovflw++;
8869 } else {
8870 _collector->push_on_overflow_list(obj);
8871 _collector->_ser_kac_ovflw++;
8872 }
8873 }
8874 }
8875 }
8876
8877 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8878 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8879
8880 // CMSParKeepAliveClosure: a parallel version of the above.
8881 // The work queues are private to each closure (thread),
8882 // but (may be) available for stealing by other threads.
8883 void CMSParKeepAliveClosure::do_oop(oop obj) {
8884 HeapWord* addr = (HeapWord*)obj;
8885 if (_span.contains(addr) &&
8886 !_bit_map->isMarked(addr)) {
8887 // In general, during recursive tracing, several threads
8888 // may be concurrently getting here; the first one to
8889 // "tag" it, claims it.
8890 if (_bit_map->par_mark(addr)) {
8891 bool res = _work_queue->push(obj);
8892 assert(res, "Low water mark should be much less than capacity");
8893 // Do a recursive trim in the hope that this will keep
8894 // stack usage lower, but leave some oops for potential stealers
8895 trim_queue(_low_water_mark);
8896 } // Else, another thread got there first
8897 }
8898 }
8899
8900 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8901 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8902
8903 void CMSParKeepAliveClosure::trim_queue(uint max) {
8904 while (_work_queue->size() > max) {
8905 oop new_oop;
8906 if (_work_queue->pop_local(new_oop)) {
8907 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8908 assert(_bit_map->isMarked((HeapWord*)new_oop),
8909 "no white objects on this stack!");
8910 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8911 // iterate over the oops in this oop, marking and pushing
8912 // the ones in CMS heap (i.e. in _span).
8913 new_oop->oop_iterate(&_mark_and_push);
8914 }
8915 }
8916 }
8917
8918 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8919 CMSCollector* collector,
8920 MemRegion span, CMSBitMap* bit_map,
8921 OopTaskQueue* work_queue):
8922 _collector(collector),
8923 _span(span),
8924 _bit_map(bit_map),
8925 _work_queue(work_queue) { }
8926
8927 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8928 HeapWord* addr = (HeapWord*)obj;
8929 if (_span.contains(addr) &&
8930 !_bit_map->isMarked(addr)) {
8931 if (_bit_map->par_mark(addr)) {
8932 bool simulate_overflow = false;
8933 NOT_PRODUCT(
8934 if (CMSMarkStackOverflowALot &&
8935 _collector->par_simulate_overflow()) {
8936 // simulate a stack overflow
8937 simulate_overflow = true;
8938 }
8939 )
8940 if (simulate_overflow || !_work_queue->push(obj)) {
8941 _collector->par_push_on_overflow_list(obj);
8942 _collector->_par_kac_ovflw++;
8943 }
8944 } // Else another thread got there already
8945 }
8946 }
8947
8948 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8949 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8950
8951 //////////////////////////////////////////////////////////////////
8952 // CMSExpansionCause /////////////////////////////
8953 //////////////////////////////////////////////////////////////////
8954 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8955 switch (cause) {
8956 case _no_expansion:
8957 return "No expansion";
8958 case _satisfy_free_ratio:
8959 return "Free ratio";
8960 case _satisfy_promotion:
8961 return "Satisfy promotion";
8962 case _satisfy_allocation:
8963 return "allocation";
8964 case _allocate_par_lab:
8965 return "Par LAB";
8966 case _allocate_par_spooling_space:
8967 return "Par Spooling Space";
8968 case _adaptive_size_policy:
8969 return "Ergonomics";
8970 default:
8971 return "unknown";
8972 }
8973 }
8974
8975 void CMSDrainMarkingStackClosure::do_void() {
8976 // the max number to take from overflow list at a time
8977 const size_t num = _mark_stack->capacity()/4;
8978 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8979 "Overflow list should be NULL during concurrent phases");
8980 while (!_mark_stack->isEmpty() ||
8981 // if stack is empty, check the overflow list
8982 _collector->take_from_overflow_list(num, _mark_stack)) {
8983 oop obj = _mark_stack->pop();
8984 HeapWord* addr = (HeapWord*)obj;
8985 assert(_span.contains(addr), "Should be within span");
8986 assert(_bit_map->isMarked(addr), "Should be marked");
8987 assert(obj->is_oop(), "Should be an oop");
8988 obj->oop_iterate(_keep_alive);
8989 }
8990 }
8991
8992 void CMSParDrainMarkingStackClosure::do_void() {
8993 // drain queue
8994 trim_queue(0);
8995 }
8996
8997 // Trim our work_queue so its length is below max at return
8998 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
8999 while (_work_queue->size() > max) {
9000 oop new_oop;
9001 if (_work_queue->pop_local(new_oop)) {
9002 assert(new_oop->is_oop(), "Expected an oop");
9003 assert(_bit_map->isMarked((HeapWord*)new_oop),
9004 "no white objects on this stack!");
9005 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
9006 // iterate over the oops in this oop, marking and pushing
9007 // the ones in CMS heap (i.e. in _span).
9008 new_oop->oop_iterate(&_mark_and_push);
9009 }
9010 }
9011 }
9012
9013 ////////////////////////////////////////////////////////////////////
9014 // Support for Marking Stack Overflow list handling and related code
9015 ////////////////////////////////////////////////////////////////////
9016 // Much of the following code is similar in shape and spirit to the
9017 // code used in ParNewGC. We should try and share that code
9018 // as much as possible in the future.
9019
9020 #ifndef PRODUCT
9021 // Debugging support for CMSStackOverflowALot
9022
9023 // It's OK to call this multi-threaded; the worst thing
9024 // that can happen is that we'll get a bunch of closely
9025 // spaced simulated overflows, but that's OK, in fact
9026 // probably good as it would exercise the overflow code
9027 // under contention.
9028 bool CMSCollector::simulate_overflow() {
9029 if (_overflow_counter-- <= 0) { // just being defensive
9030 _overflow_counter = CMSMarkStackOverflowInterval;
9031 return true;
9032 } else {
9033 return false;
9034 }
9035 }
9036
9037 bool CMSCollector::par_simulate_overflow() {
9038 return simulate_overflow();
9039 }
9040 #endif
9041
9042 // Single-threaded
9043 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
9044 assert(stack->isEmpty(), "Expected precondition");
9045 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
9046 size_t i = num;
9047 oop cur = _overflow_list;
9048 const markOop proto = markOopDesc::prototype();
9049 NOT_PRODUCT(ssize_t n = 0;)
9050 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
9051 next = oop(cur->mark());
9052 cur->set_mark(proto); // until proven otherwise
9053 assert(cur->is_oop(), "Should be an oop");
9054 bool res = stack->push(cur);
9055 assert(res, "Bit off more than can chew?");
9056 NOT_PRODUCT(n++;)
9057 }
9058 _overflow_list = cur;
9059 #ifndef PRODUCT
9060 assert(_num_par_pushes >= n, "Too many pops?");
9061 _num_par_pushes -=n;
9062 #endif
9063 return !stack->isEmpty();
9064 }
9065
9066 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
9067 // (MT-safe) Get a prefix of at most "num" from the list.
9068 // The overflow list is chained through the mark word of
9069 // each object in the list. We fetch the entire list,
9070 // break off a prefix of the right size and return the
9071 // remainder. If other threads try to take objects from
9072 // the overflow list at that time, they will wait for
9073 // some time to see if data becomes available. If (and
9074 // only if) another thread places one or more object(s)
9075 // on the global list before we have returned the suffix
9076 // to the global list, we will walk down our local list
9077 // to find its end and append the global list to
9078 // our suffix before returning it. This suffix walk can
9079 // prove to be expensive (quadratic in the amount of traffic)
9080 // when there are many objects in the overflow list and
9081 // there is much producer-consumer contention on the list.
9082 // *NOTE*: The overflow list manipulation code here and
9083 // in ParNewGeneration:: are very similar in shape,
9084 // except that in the ParNew case we use the old (from/eden)
9085 // copy of the object to thread the list via its klass word.
9086 // Because of the common code, if you make any changes in
9087 // the code below, please check the ParNew version to see if
9088 // similar changes might be needed.
9089 // CR 6797058 has been filed to consolidate the common code.
9090 bool CMSCollector::par_take_from_overflow_list(size_t num,
9091 OopTaskQueue* work_q,
9092 int no_of_gc_threads) {
9093 assert(work_q->size() == 0, "First empty local work queue");
9094 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
9095 if (_overflow_list == NULL) {
9096 return false;
9097 }
9098 // Grab the entire list; we'll put back a suffix
9099 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9100 Thread* tid = Thread::current();
9101 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
9102 // set to ParallelGCThreads.
9103 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
9104 size_t sleep_time_millis = MAX2((size_t)1, num/100);
9105 // If the list is busy, we spin for a short while,
9106 // sleeping between attempts to get the list.
9107 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
9108 os::sleep(tid, sleep_time_millis, false);
9109 if (_overflow_list == NULL) {
9110 // Nothing left to take
9111 return false;
9112 } else if (_overflow_list != BUSY) {
9113 // Try and grab the prefix
9114 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9115 }
9116 }
9117 // If the list was found to be empty, or we spun long
9118 // enough, we give up and return empty-handed. If we leave
9119 // the list in the BUSY state below, it must be the case that
9120 // some other thread holds the overflow list and will set it
9121 // to a non-BUSY state in the future.
9122 if (prefix == NULL || prefix == BUSY) {
9123 // Nothing to take or waited long enough
9124 if (prefix == NULL) {
9125 // Write back the NULL in case we overwrote it with BUSY above
9126 // and it is still the same value.
9127 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9128 }
9129 return false;
9130 }
9131 assert(prefix != NULL && prefix != BUSY, "Error");
9132 size_t i = num;
9133 oop cur = prefix;
9134 // Walk down the first "num" objects, unless we reach the end.
9135 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
9136 if (cur->mark() == NULL) {
9137 // We have "num" or fewer elements in the list, so there
9138 // is nothing to return to the global list.
9139 // Write back the NULL in lieu of the BUSY we wrote
9140 // above, if it is still the same value.
9141 if (_overflow_list == BUSY) {
9142 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9143 }
9144 } else {
9145 // Chop off the suffix and return it to the global list.
9146 assert(cur->mark() != BUSY, "Error");
9147 oop suffix_head = cur->mark(); // suffix will be put back on global list
9148 cur->set_mark(NULL); // break off suffix
9149 // It's possible that the list is still in the empty(busy) state
9150 // we left it in a short while ago; in that case we may be
9151 // able to place back the suffix without incurring the cost
9152 // of a walk down the list.
9153 oop observed_overflow_list = _overflow_list;
9154 oop cur_overflow_list = observed_overflow_list;
9155 bool attached = false;
9156 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
9157 observed_overflow_list =
9158 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9159 if (cur_overflow_list == observed_overflow_list) {
9160 attached = true;
9161 break;
9162 } else cur_overflow_list = observed_overflow_list;
9163 }
9164 if (!attached) {
9165 // Too bad, someone else sneaked in (at least) an element; we'll need
9166 // to do a splice. Find tail of suffix so we can prepend suffix to global
9167 // list.
9168 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
9169 oop suffix_tail = cur;
9170 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
9171 "Tautology");
9172 observed_overflow_list = _overflow_list;
9173 do {
9174 cur_overflow_list = observed_overflow_list;
9175 if (cur_overflow_list != BUSY) {
9176 // Do the splice ...
9177 suffix_tail->set_mark(markOop(cur_overflow_list));
9178 } else { // cur_overflow_list == BUSY
9179 suffix_tail->set_mark(NULL);
9180 }
9181 // ... and try to place spliced list back on overflow_list ...
9182 observed_overflow_list =
9183 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9184 } while (cur_overflow_list != observed_overflow_list);
9185 // ... until we have succeeded in doing so.
9186 }
9187 }
9188
9189 // Push the prefix elements on work_q
9190 assert(prefix != NULL, "control point invariant");
9191 const markOop proto = markOopDesc::prototype();
9192 oop next;
9193 NOT_PRODUCT(ssize_t n = 0;)
9194 for (cur = prefix; cur != NULL; cur = next) {
9195 next = oop(cur->mark());
9196 cur->set_mark(proto); // until proven otherwise
9197 assert(cur->is_oop(), "Should be an oop");
9198 bool res = work_q->push(cur);
9199 assert(res, "Bit off more than we can chew?");
9200 NOT_PRODUCT(n++;)
9201 }
9202 #ifndef PRODUCT
9203 assert(_num_par_pushes >= n, "Too many pops?");
9204 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
9205 #endif
9206 return true;
9207 }
9208
9209 // Single-threaded
9210 void CMSCollector::push_on_overflow_list(oop p) {
9211 NOT_PRODUCT(_num_par_pushes++;)
9212 assert(p->is_oop(), "Not an oop");
9213 preserve_mark_if_necessary(p);
9214 p->set_mark((markOop)_overflow_list);
9215 _overflow_list = p;
9216 }
9217
9218 // Multi-threaded; use CAS to prepend to overflow list
9219 void CMSCollector::par_push_on_overflow_list(oop p) {
9220 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9221 assert(p->is_oop(), "Not an oop");
9222 par_preserve_mark_if_necessary(p);
9223 oop observed_overflow_list = _overflow_list;
9224 oop cur_overflow_list;
9225 do {
9226 cur_overflow_list = observed_overflow_list;
9227 if (cur_overflow_list != BUSY) {
9228 p->set_mark(markOop(cur_overflow_list));
9229 } else {
9230 p->set_mark(NULL);
9231 }
9232 observed_overflow_list =
9233 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9234 } while (cur_overflow_list != observed_overflow_list);
9235 }
9236 #undef BUSY
9237
9238 // Single threaded
9239 // General Note on GrowableArray: pushes may silently fail
9240 // because we are (temporarily) out of C-heap for expanding
9241 // the stack. The problem is quite ubiquitous and affects
9242 // a lot of code in the JVM. The prudent thing for GrowableArray
9243 // to do (for now) is to exit with an error. However, that may
9244 // be too draconian in some cases because the caller may be
9245 // able to recover without much harm. For such cases, we
9246 // should probably introduce a "soft_push" method which returns
9247 // an indication of success or failure with the assumption that
9248 // the caller may be able to recover from a failure; code in
9249 // the VM can then be changed, incrementally, to deal with such
9250 // failures where possible, thus, incrementally hardening the VM
9251 // in such low resource situations.
9252 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9253 _preserved_oop_stack.push(p);
9254 _preserved_mark_stack.push(m);
9255 assert(m == p->mark(), "Mark word changed");
9256 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9257 "bijection");
9258 }
9259
9260 // Single threaded
9261 void CMSCollector::preserve_mark_if_necessary(oop p) {
9262 markOop m = p->mark();
9263 if (m->must_be_preserved(p)) {
9264 preserve_mark_work(p, m);
9265 }
9266 }
9267
9268 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9269 markOop m = p->mark();
9270 if (m->must_be_preserved(p)) {
9271 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9272 // Even though we read the mark word without holding
9273 // the lock, we are assured that it will not change
9274 // because we "own" this oop, so no other thread can
9275 // be trying to push it on the overflow list; see
9276 // the assertion in preserve_mark_work() that checks
9277 // that m == p->mark().
9278 preserve_mark_work(p, m);
9279 }
9280 }
9281
9282 // We should be able to do this multi-threaded,
9283 // a chunk of stack being a task (this is
9284 // correct because each oop only ever appears
9285 // once in the overflow list. However, it's
9286 // not very easy to completely overlap this with
9287 // other operations, so will generally not be done
9288 // until all work's been completed. Because we
9289 // expect the preserved oop stack (set) to be small,
9290 // it's probably fine to do this single-threaded.
9291 // We can explore cleverer concurrent/overlapped/parallel
9292 // processing of preserved marks if we feel the
9293 // need for this in the future. Stack overflow should
9294 // be so rare in practice and, when it happens, its
9295 // effect on performance so great that this will
9296 // likely just be in the noise anyway.
9297 void CMSCollector::restore_preserved_marks_if_any() {
9298 assert(SafepointSynchronize::is_at_safepoint(),
9299 "world should be stopped");
9300 assert(Thread::current()->is_ConcurrentGC_thread() ||
9301 Thread::current()->is_VM_thread(),
9302 "should be single-threaded");
9303 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9304 "bijection");
9305
9306 while (!_preserved_oop_stack.is_empty()) {
9307 oop p = _preserved_oop_stack.pop();
9308 assert(p->is_oop(), "Should be an oop");
9309 assert(_span.contains(p), "oop should be in _span");
9310 assert(p->mark() == markOopDesc::prototype(),
9311 "Set when taken from overflow list");
9312 markOop m = _preserved_mark_stack.pop();
9313 p->set_mark(m);
9314 }
9315 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9316 "stacks were cleared above");
9317 }
9318
9319 #ifndef PRODUCT
9320 bool CMSCollector::no_preserved_marks() const {
9321 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9322 }
9323 #endif
9324
9325 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9326 {
9327 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9328 CMSAdaptiveSizePolicy* size_policy =
9329 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9330 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9331 "Wrong type for size policy");
9332 return size_policy;
9333 }
9334
9335 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9336 size_t desired_promo_size) {
9337 if (cur_promo_size < desired_promo_size) {
9338 size_t expand_bytes = desired_promo_size - cur_promo_size;
9339 if (PrintAdaptiveSizePolicy && Verbose) {
9340 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9341 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9342 expand_bytes);
9343 }
9344 expand(expand_bytes,
9345 MinHeapDeltaBytes,
9346 CMSExpansionCause::_adaptive_size_policy);
9347 } else if (desired_promo_size < cur_promo_size) {
9348 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9349 if (PrintAdaptiveSizePolicy && Verbose) {
9350 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9351 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9352 shrink_bytes);
9353 }
9354 shrink(shrink_bytes);
9355 }
9356 }
9357
9358 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9359 GenCollectedHeap* gch = GenCollectedHeap::heap();
9360 CMSGCAdaptivePolicyCounters* counters =
9361 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9362 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9363 "Wrong kind of counters");
9364 return counters;
9365 }
9366
9367
9368 void ASConcurrentMarkSweepGeneration::update_counters() {
9369 if (UsePerfData) {
9370 _space_counters->update_all();
9371 _gen_counters->update_all();
9372 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9373 GenCollectedHeap* gch = GenCollectedHeap::heap();
9374 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9375 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9376 "Wrong gc statistics type");
9377 counters->update_counters(gc_stats_l);
9378 }
9379 }
9380
9381 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9382 if (UsePerfData) {
9383 _space_counters->update_used(used);
9384 _space_counters->update_capacity();
9385 _gen_counters->update_all();
9386
9387 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9388 GenCollectedHeap* gch = GenCollectedHeap::heap();
9389 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9390 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9391 "Wrong gc statistics type");
9392 counters->update_counters(gc_stats_l);
9393 }
9394 }
9395
9396 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9397 assert_locked_or_safepoint(Heap_lock);
9398 assert_lock_strong(freelistLock());
9399 HeapWord* old_end = _cmsSpace->end();
9400 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9401 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9402 FreeChunk* chunk_at_end = find_chunk_at_end();
9403 if (chunk_at_end == NULL) {
9404 // No room to shrink
9405 if (PrintGCDetails && Verbose) {
9406 gclog_or_tty->print_cr("No room to shrink: old_end "
9407 PTR_FORMAT " unallocated_start " PTR_FORMAT
9408 " chunk_at_end " PTR_FORMAT,
9409 old_end, unallocated_start, chunk_at_end);
9410 }
9411 return;
9412 } else {
9413
9414 // Find the chunk at the end of the space and determine
9415 // how much it can be shrunk.
9416 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9417 size_t aligned_shrinkable_size_in_bytes =
9418 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9419 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(),
9420 "Inconsistent chunk at end of space");
9421 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9422 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9423
9424 // Shrink the underlying space
9425 _virtual_space.shrink_by(bytes);
9426 if (PrintGCDetails && Verbose) {
9427 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9428 " desired_bytes " SIZE_FORMAT
9429 " shrinkable_size_in_bytes " SIZE_FORMAT
9430 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9431 " bytes " SIZE_FORMAT,
9432 desired_bytes, shrinkable_size_in_bytes,
9433 aligned_shrinkable_size_in_bytes, bytes);
9434 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9435 " unallocated_start " SIZE_FORMAT,
9436 old_end, unallocated_start);
9437 }
9438
9439 // If the space did shrink (shrinking is not guaranteed),
9440 // shrink the chunk at the end by the appropriate amount.
9441 if (((HeapWord*)_virtual_space.high()) < old_end) {
9442 size_t new_word_size =
9443 heap_word_size(_virtual_space.committed_size());
9444
9445 // Have to remove the chunk from the dictionary because it is changing
9446 // size and might be someplace elsewhere in the dictionary.
9447
9448 // Get the chunk at end, shrink it, and put it
9449 // back.
9450 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9451 size_t word_size_change = word_size_before - new_word_size;
9452 size_t chunk_at_end_old_size = chunk_at_end->size();
9453 assert(chunk_at_end_old_size >= word_size_change,
9454 "Shrink is too large");
9455 chunk_at_end->set_size(chunk_at_end_old_size -
9456 word_size_change);
9457 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9458 word_size_change);
9459
9460 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9461
9462 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9463 _bts->resize(new_word_size); // resize the block offset shared array
9464 Universe::heap()->barrier_set()->resize_covered_region(mr);
9465 _cmsSpace->assert_locked();
9466 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9467
9468 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9469
9470 // update the space and generation capacity counters
9471 if (UsePerfData) {
9472 _space_counters->update_capacity();
9473 _gen_counters->update_all();
9474 }
9475
9476 if (Verbose && PrintGCDetails) {
9477 size_t new_mem_size = _virtual_space.committed_size();
9478 size_t old_mem_size = new_mem_size + bytes;
9479 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
9480 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9481 }
9482 }
9483
9484 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9485 "Inconsistency at end of space");
9486 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(),
9487 "Shrinking is inconsistent");
9488 return;
9489 }
9490 }
9491 // Transfer some number of overflown objects to usual marking
9492 // stack. Return true if some objects were transferred.
9493 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9494 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9495 (size_t)ParGCDesiredObjsFromOverflowList);
9496
9497 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9498 assert(_collector->overflow_list_is_empty() || res,
9499 "If list is not empty, we should have taken something");
9500 assert(!res || !_mark_stack->isEmpty(),
9501 "If we took something, it should now be on our stack");
9502 return res;
9503 }
9504
9505 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9506 size_t res = _sp->block_size_no_stall(addr, _collector);
9507 if (_sp->block_is_obj(addr)) {
9508 if (_live_bit_map->isMarked(addr)) {
9509 // It can't have been dead in a previous cycle
9510 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9511 } else {
9512 _dead_bit_map->mark(addr); // mark the dead object
9513 }
9514 }
9515 // Could be 0, if the block size could not be computed without stalling.
9516 return res;
9517 }
9518
9519 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9520
9521 switch (phase) {
9522 case CMSCollector::InitialMarking:
9523 initialize(true /* fullGC */ ,
9524 cause /* cause of the GC */,
9525 true /* recordGCBeginTime */,
9526 true /* recordPreGCUsage */,
9527 false /* recordPeakUsage */,
9528 false /* recordPostGCusage */,
9529 true /* recordAccumulatedGCTime */,
9530 false /* recordGCEndTime */,
9531 false /* countCollection */ );
9532 break;
9533
9534 case CMSCollector::FinalMarking:
9535 initialize(true /* fullGC */ ,
9536 cause /* cause of the GC */,
9537 false /* recordGCBeginTime */,
9538 false /* recordPreGCUsage */,
9539 false /* recordPeakUsage */,
9540 false /* recordPostGCusage */,
9541 true /* recordAccumulatedGCTime */,
9542 false /* recordGCEndTime */,
9543 false /* countCollection */ );
9544 break;
9545
9546 case CMSCollector::Sweeping:
9547 initialize(true /* fullGC */ ,
9548 cause /* cause of the GC */,
9549 false /* recordGCBeginTime */,
9550 false /* recordPreGCUsage */,
9551 true /* recordPeakUsage */,
9552 true /* recordPostGCusage */,
9553 false /* recordAccumulatedGCTime */,
9554 true /* recordGCEndTime */,
9555 true /* countCollection */ );
9556 break;
9557
9558 default:
9559 ShouldNotReachHere();
9560 }
9561 }