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 supsicion, 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 intereference 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 mimimum
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 // _intiating_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(false),
598 _concurrent_cycles_since_last_unload(0),
599 _roots_scanning_options(0),
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 // Choose what strong roots should be scanned depending on verification options
792 if (!CMSClassUnloadingEnabled) {
793 // If class unloading is disabled we want to include all classes into the root set.
794 add_root_scanning_option(SharedHeap::SO_AllClasses);
795 } else {
796 add_root_scanning_option(SharedHeap::SO_SystemClasses);
797 }
798
799 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
800 _gc_counters = new CollectorCounters("CMS", 1);
801 _completed_initialization = true;
802 _inter_sweep_timer.start(); // start of time
803 }
804
805 const char* ConcurrentMarkSweepGeneration::name() const {
806 return "concurrent mark-sweep generation";
807 }
808 void ConcurrentMarkSweepGeneration::update_counters() {
809 if (UsePerfData) {
810 _space_counters->update_all();
811 _gen_counters->update_all();
812 }
813 }
814
815 // this is an optimized version of update_counters(). it takes the
816 // used value as a parameter rather than computing it.
817 //
818 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
819 if (UsePerfData) {
820 _space_counters->update_used(used);
821 _space_counters->update_capacity();
822 _gen_counters->update_all();
823 }
824 }
825
826 void ConcurrentMarkSweepGeneration::print() const {
827 Generation::print();
828 cmsSpace()->print();
829 }
830
831 #ifndef PRODUCT
832 void ConcurrentMarkSweepGeneration::print_statistics() {
833 cmsSpace()->printFLCensus(0);
834 }
835 #endif
836
837 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) {
838 GenCollectedHeap* gch = GenCollectedHeap::heap();
839 if (PrintGCDetails) {
840 if (Verbose) {
841 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]",
842 level(), short_name(), s, used(), capacity());
843 } else {
844 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]",
845 level(), short_name(), s, used() / K, capacity() / K);
846 }
847 }
848 if (Verbose) {
849 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")",
850 gch->used(), gch->capacity());
851 } else {
852 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)",
853 gch->used() / K, gch->capacity() / K);
854 }
855 }
856
857 size_t
858 ConcurrentMarkSweepGeneration::contiguous_available() const {
859 // dld proposes an improvement in precision here. If the committed
860 // part of the space ends in a free block we should add that to
861 // uncommitted size in the calculation below. Will make this
862 // change later, staying with the approximation below for the
863 // time being. -- ysr.
864 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
865 }
866
867 size_t
868 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
869 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
870 }
871
872 size_t ConcurrentMarkSweepGeneration::max_available() const {
873 return free() + _virtual_space.uncommitted_size();
874 }
875
876 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
877 size_t available = max_available();
878 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
879 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
880 if (Verbose && PrintGCDetails) {
881 gclog_or_tty->print_cr(
882 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT"),"
883 "max_promo("SIZE_FORMAT")",
884 res? "":" not", available, res? ">=":"<",
885 av_promo, max_promotion_in_bytes);
886 }
887 return res;
888 }
889
890 // At a promotion failure dump information on block layout in heap
891 // (cms old generation).
892 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
893 if (CMSDumpAtPromotionFailure) {
894 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty);
895 }
896 }
897
898 CompactibleSpace*
899 ConcurrentMarkSweepGeneration::first_compaction_space() const {
900 return _cmsSpace;
901 }
902
903 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
904 // Clear the promotion information. These pointers can be adjusted
905 // along with all the other pointers into the heap but
906 // compaction is expected to be a rare event with
907 // a heap using cms so don't do it without seeing the need.
908 if (CollectedHeap::use_parallel_gc_threads()) {
909 for (uint i = 0; i < ParallelGCThreads; i++) {
910 _par_gc_thread_states[i]->promo.reset();
911 }
912 }
913 }
914
915 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) {
916 blk->do_space(_cmsSpace);
917 }
918
919 void ConcurrentMarkSweepGeneration::compute_new_size() {
920 assert_locked_or_safepoint(Heap_lock);
921
922 // If incremental collection failed, we just want to expand
923 // to the limit.
924 if (incremental_collection_failed()) {
925 clear_incremental_collection_failed();
926 grow_to_reserved();
927 return;
928 }
929
930 // The heap has been compacted but not reset yet.
931 // Any metric such as free() or used() will be incorrect.
932
933 CardGeneration::compute_new_size();
934
935 // Reset again after a possible resizing
936 if (did_compact()) {
937 cmsSpace()->reset_after_compaction();
938 }
939 }
940
941 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
942 assert_locked_or_safepoint(Heap_lock);
943
944 // If incremental collection failed, we just want to expand
945 // to the limit.
946 if (incremental_collection_failed()) {
947 clear_incremental_collection_failed();
948 grow_to_reserved();
949 return;
950 }
951
952 double free_percentage = ((double) free()) / capacity();
953 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
954 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
955
956 // compute expansion delta needed for reaching desired free percentage
957 if (free_percentage < desired_free_percentage) {
958 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
959 assert(desired_capacity >= capacity(), "invalid expansion size");
960 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
961 if (PrintGCDetails && Verbose) {
962 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
963 gclog_or_tty->print_cr("\nFrom compute_new_size: ");
964 gclog_or_tty->print_cr(" Free fraction %f", free_percentage);
965 gclog_or_tty->print_cr(" Desired free fraction %f",
966 desired_free_percentage);
967 gclog_or_tty->print_cr(" Maximum free fraction %f",
968 maximum_free_percentage);
969 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000);
970 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT,
971 desired_capacity/1000);
972 int prev_level = level() - 1;
973 if (prev_level >= 0) {
974 size_t prev_size = 0;
975 GenCollectedHeap* gch = GenCollectedHeap::heap();
976 Generation* prev_gen = gch->_gens[prev_level];
977 prev_size = prev_gen->capacity();
978 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT,
979 prev_size/1000);
980 }
981 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT,
982 unsafe_max_alloc_nogc()/1000);
983 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT,
984 contiguous_available()/1000);
985 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)",
986 expand_bytes);
987 }
988 // safe if expansion fails
989 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
990 if (PrintGCDetails && Verbose) {
991 gclog_or_tty->print_cr(" Expanded free fraction %f",
992 ((double) free()) / capacity());
993 }
994 } else {
995 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
996 assert(desired_capacity <= capacity(), "invalid expansion size");
997 size_t shrink_bytes = capacity() - desired_capacity;
998 // Don't shrink unless the delta is greater than the minimum shrink we want
999 if (shrink_bytes >= MinHeapDeltaBytes) {
1000 shrink_free_list_by(shrink_bytes);
1001 }
1002 }
1003 }
1004
1005 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
1006 return cmsSpace()->freelistLock();
1007 }
1008
1009 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size,
1010 bool tlab) {
1011 CMSSynchronousYieldRequest yr;
1012 MutexLockerEx x(freelistLock(),
1013 Mutex::_no_safepoint_check_flag);
1014 return have_lock_and_allocate(size, tlab);
1015 }
1016
1017 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
1018 bool tlab /* ignored */) {
1019 assert_lock_strong(freelistLock());
1020 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
1021 HeapWord* res = cmsSpace()->allocate(adjustedSize);
1022 // Allocate the object live (grey) if the background collector has
1023 // started marking. This is necessary because the marker may
1024 // have passed this address and consequently this object will
1025 // not otherwise be greyed and would be incorrectly swept up.
1026 // Note that if this object contains references, the writing
1027 // of those references will dirty the card containing this object
1028 // allowing the object to be blackened (and its references scanned)
1029 // either during a preclean phase or at the final checkpoint.
1030 if (res != NULL) {
1031 // We may block here with an uninitialized object with
1032 // its mark-bit or P-bits not yet set. Such objects need
1033 // to be safely navigable by block_start().
1034 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
1035 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
1036 collector()->direct_allocated(res, adjustedSize);
1037 _direct_allocated_words += adjustedSize;
1038 // allocation counters
1039 NOT_PRODUCT(
1040 _numObjectsAllocated++;
1041 _numWordsAllocated += (int)adjustedSize;
1042 )
1043 }
1044 return res;
1045 }
1046
1047 // In the case of direct allocation by mutators in a generation that
1048 // is being concurrently collected, the object must be allocated
1049 // live (grey) if the background collector has started marking.
1050 // This is necessary because the marker may
1051 // have passed this address and consequently this object will
1052 // not otherwise be greyed and would be incorrectly swept up.
1053 // Note that if this object contains references, the writing
1054 // of those references will dirty the card containing this object
1055 // allowing the object to be blackened (and its references scanned)
1056 // either during a preclean phase or at the final checkpoint.
1057 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
1058 assert(_markBitMap.covers(start, size), "Out of bounds");
1059 if (_collectorState >= Marking) {
1060 MutexLockerEx y(_markBitMap.lock(),
1061 Mutex::_no_safepoint_check_flag);
1062 // [see comments preceding SweepClosure::do_blk() below for details]
1063 //
1064 // Can the P-bits be deleted now? JJJ
1065 //
1066 // 1. need to mark the object as live so it isn't collected
1067 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
1068 // 3. need to mark the end of the object so marking, precleaning or sweeping
1069 // can skip over uninitialized or unparsable objects. An allocated
1070 // object is considered uninitialized for our purposes as long as
1071 // its klass word is NULL. All old gen objects are parsable
1072 // as soon as they are initialized.)
1073 _markBitMap.mark(start); // object is live
1074 _markBitMap.mark(start + 1); // object is potentially uninitialized?
1075 _markBitMap.mark(start + size - 1);
1076 // mark end of object
1077 }
1078 // check that oop looks uninitialized
1079 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
1080 }
1081
1082 void CMSCollector::promoted(bool par, HeapWord* start,
1083 bool is_obj_array, size_t obj_size) {
1084 assert(_markBitMap.covers(start), "Out of bounds");
1085 // See comment in direct_allocated() about when objects should
1086 // be allocated live.
1087 if (_collectorState >= Marking) {
1088 // we already hold the marking bit map lock, taken in
1089 // the prologue
1090 if (par) {
1091 _markBitMap.par_mark(start);
1092 } else {
1093 _markBitMap.mark(start);
1094 }
1095 // We don't need to mark the object as uninitialized (as
1096 // in direct_allocated above) because this is being done with the
1097 // world stopped and the object will be initialized by the
1098 // time the marking, precleaning or sweeping get to look at it.
1099 // But see the code for copying objects into the CMS generation,
1100 // where we need to ensure that concurrent readers of the
1101 // block offset table are able to safely navigate a block that
1102 // is in flux from being free to being allocated (and in
1103 // transition while being copied into) and subsequently
1104 // becoming a bona-fide object when the copy/promotion is complete.
1105 assert(SafepointSynchronize::is_at_safepoint(),
1106 "expect promotion only at safepoints");
1107
1108 if (_collectorState < Sweeping) {
1109 // Mark the appropriate cards in the modUnionTable, so that
1110 // this object gets scanned before the sweep. If this is
1111 // not done, CMS generation references in the object might
1112 // not get marked.
1113 // For the case of arrays, which are otherwise precisely
1114 // marked, we need to dirty the entire array, not just its head.
1115 if (is_obj_array) {
1116 // The [par_]mark_range() method expects mr.end() below to
1117 // be aligned to the granularity of a bit's representation
1118 // in the heap. In the case of the MUT below, that's a
1119 // card size.
1120 MemRegion mr(start,
1121 (HeapWord*)round_to((intptr_t)(start + obj_size),
1122 CardTableModRefBS::card_size /* bytes */));
1123 if (par) {
1124 _modUnionTable.par_mark_range(mr);
1125 } else {
1126 _modUnionTable.mark_range(mr);
1127 }
1128 } else { // not an obj array; we can just mark the head
1129 if (par) {
1130 _modUnionTable.par_mark(start);
1131 } else {
1132 _modUnionTable.mark(start);
1133 }
1134 }
1135 }
1136 }
1137 }
1138
1139 static inline size_t percent_of_space(Space* space, HeapWord* addr)
1140 {
1141 size_t delta = pointer_delta(addr, space->bottom());
1142 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize));
1143 }
1144
1145 void CMSCollector::icms_update_allocation_limits()
1146 {
1147 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0);
1148 EdenSpace* eden = gen0->as_DefNewGeneration()->eden();
1149
1150 const unsigned int duty_cycle = stats().icms_update_duty_cycle();
1151 if (CMSTraceIncrementalPacing) {
1152 stats().print();
1153 }
1154
1155 assert(duty_cycle <= 100, "invalid duty cycle");
1156 if (duty_cycle != 0) {
1157 // The duty_cycle is a percentage between 0 and 100; convert to words and
1158 // then compute the offset from the endpoints of the space.
1159 size_t free_words = eden->free() / HeapWordSize;
1160 double free_words_dbl = (double)free_words;
1161 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0);
1162 size_t offset_words = (free_words - duty_cycle_words) / 2;
1163
1164 _icms_start_limit = eden->top() + offset_words;
1165 _icms_stop_limit = eden->end() - offset_words;
1166
1167 // The limits may be adjusted (shifted to the right) by
1168 // CMSIncrementalOffset, to allow the application more mutator time after a
1169 // young gen gc (when all mutators were stopped) and before CMS starts and
1170 // takes away one or more cpus.
1171 if (CMSIncrementalOffset != 0) {
1172 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0;
1173 size_t adjustment = (size_t)adjustment_dbl;
1174 HeapWord* tmp_stop = _icms_stop_limit + adjustment;
1175 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) {
1176 _icms_start_limit += adjustment;
1177 _icms_stop_limit = tmp_stop;
1178 }
1179 }
1180 }
1181 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) {
1182 _icms_start_limit = _icms_stop_limit = eden->end();
1183 }
1184
1185 // Install the new start limit.
1186 eden->set_soft_end(_icms_start_limit);
1187
1188 if (CMSTraceIncrementalMode) {
1189 gclog_or_tty->print(" icms alloc limits: "
1190 PTR_FORMAT "," PTR_FORMAT
1191 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ",
1192 _icms_start_limit, _icms_stop_limit,
1193 percent_of_space(eden, _icms_start_limit),
1194 percent_of_space(eden, _icms_stop_limit));
1195 if (Verbose) {
1196 gclog_or_tty->print("eden: ");
1197 eden->print_on(gclog_or_tty);
1198 }
1199 }
1200 }
1201
1202 // Any changes here should try to maintain the invariant
1203 // that if this method is called with _icms_start_limit
1204 // and _icms_stop_limit both NULL, then it should return NULL
1205 // and not notify the icms thread.
1206 HeapWord*
1207 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top,
1208 size_t word_size)
1209 {
1210 // A start_limit equal to end() means the duty cycle is 0, so treat that as a
1211 // nop.
1212 if (CMSIncrementalMode && _icms_start_limit != space->end()) {
1213 if (top <= _icms_start_limit) {
1214 if (CMSTraceIncrementalMode) {
1215 space->print_on(gclog_or_tty);
1216 gclog_or_tty->stamp();
1217 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT
1218 ", new limit=" PTR_FORMAT
1219 " (" SIZE_FORMAT "%%)",
1220 top, _icms_stop_limit,
1221 percent_of_space(space, _icms_stop_limit));
1222 }
1223 ConcurrentMarkSweepThread::start_icms();
1224 assert(top < _icms_stop_limit, "Tautology");
1225 if (word_size < pointer_delta(_icms_stop_limit, top)) {
1226 return _icms_stop_limit;
1227 }
1228
1229 // The allocation will cross both the _start and _stop limits, so do the
1230 // stop notification also and return end().
1231 if (CMSTraceIncrementalMode) {
1232 space->print_on(gclog_or_tty);
1233 gclog_or_tty->stamp();
1234 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT
1235 ", new limit=" PTR_FORMAT
1236 " (" SIZE_FORMAT "%%)",
1237 top, space->end(),
1238 percent_of_space(space, space->end()));
1239 }
1240 ConcurrentMarkSweepThread::stop_icms();
1241 return space->end();
1242 }
1243
1244 if (top <= _icms_stop_limit) {
1245 if (CMSTraceIncrementalMode) {
1246 space->print_on(gclog_or_tty);
1247 gclog_or_tty->stamp();
1248 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT
1249 ", new limit=" PTR_FORMAT
1250 " (" SIZE_FORMAT "%%)",
1251 top, space->end(),
1252 percent_of_space(space, space->end()));
1253 }
1254 ConcurrentMarkSweepThread::stop_icms();
1255 return space->end();
1256 }
1257
1258 if (CMSTraceIncrementalMode) {
1259 space->print_on(gclog_or_tty);
1260 gclog_or_tty->stamp();
1261 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT
1262 ", new limit=" PTR_FORMAT,
1263 top, NULL);
1264 }
1265 }
1266
1267 return NULL;
1268 }
1269
1270 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
1271 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
1272 // allocate, copy and if necessary update promoinfo --
1273 // delegate to underlying space.
1274 assert_lock_strong(freelistLock());
1275
1276 #ifndef PRODUCT
1277 if (Universe::heap()->promotion_should_fail()) {
1278 return NULL;
1279 }
1280 #endif // #ifndef PRODUCT
1281
1282 oop res = _cmsSpace->promote(obj, obj_size);
1283 if (res == NULL) {
1284 // expand and retry
1285 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
1286 expand(s*HeapWordSize, MinHeapDeltaBytes,
1287 CMSExpansionCause::_satisfy_promotion);
1288 // Since there's currently no next generation, we don't try to promote
1289 // into a more senior generation.
1290 assert(next_gen() == NULL, "assumption, based upon which no attempt "
1291 "is made to pass on a possibly failing "
1292 "promotion to next generation");
1293 res = _cmsSpace->promote(obj, obj_size);
1294 }
1295 if (res != NULL) {
1296 // See comment in allocate() about when objects should
1297 // be allocated live.
1298 assert(obj->is_oop(), "Will dereference klass pointer below");
1299 collector()->promoted(false, // Not parallel
1300 (HeapWord*)res, obj->is_objArray(), obj_size);
1301 // promotion counters
1302 NOT_PRODUCT(
1303 _numObjectsPromoted++;
1304 _numWordsPromoted +=
1305 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
1306 )
1307 }
1308 return res;
1309 }
1310
1311
1312 HeapWord*
1313 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space,
1314 HeapWord* top,
1315 size_t word_sz)
1316 {
1317 return collector()->allocation_limit_reached(space, top, word_sz);
1318 }
1319
1320 // IMPORTANT: Notes on object size recognition in CMS.
1321 // ---------------------------------------------------
1322 // A block of storage in the CMS generation is always in
1323 // one of three states. A free block (FREE), an allocated
1324 // object (OBJECT) whose size() method reports the correct size,
1325 // and an intermediate state (TRANSIENT) in which its size cannot
1326 // be accurately determined.
1327 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
1328 // -----------------------------------------------------
1329 // FREE: klass_word & 1 == 1; mark_word holds block size
1330 //
1331 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
1332 // obj->size() computes correct size
1333 //
1334 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1335 //
1336 // STATE IDENTIFICATION: (64 bit+COOPS)
1337 // ------------------------------------
1338 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
1339 //
1340 // OBJECT: klass_word installed; klass_word != 0;
1341 // obj->size() computes correct size
1342 //
1343 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
1344 //
1345 //
1346 // STATE TRANSITION DIAGRAM
1347 //
1348 // mut / parnew mut / parnew
1349 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
1350 // ^ |
1351 // |------------------------ DEAD <------------------------------------|
1352 // sweep mut
1353 //
1354 // While a block is in TRANSIENT state its size cannot be determined
1355 // so readers will either need to come back later or stall until
1356 // the size can be determined. Note that for the case of direct
1357 // allocation, P-bits, when available, may be used to determine the
1358 // size of an object that may not yet have been initialized.
1359
1360 // Things to support parallel young-gen collection.
1361 oop
1362 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1363 oop old, markOop m,
1364 size_t word_sz) {
1365 #ifndef PRODUCT
1366 if (Universe::heap()->promotion_should_fail()) {
1367 return NULL;
1368 }
1369 #endif // #ifndef PRODUCT
1370
1371 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1372 PromotionInfo* promoInfo = &ps->promo;
1373 // if we are tracking promotions, then first ensure space for
1374 // promotion (including spooling space for saving header if necessary).
1375 // then allocate and copy, then track promoted info if needed.
1376 // When tracking (see PromotionInfo::track()), the mark word may
1377 // be displaced and in this case restoration of the mark word
1378 // occurs in the (oop_since_save_marks_)iterate phase.
1379 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1380 // Out of space for allocating spooling buffers;
1381 // try expanding and allocating spooling buffers.
1382 if (!expand_and_ensure_spooling_space(promoInfo)) {
1383 return NULL;
1384 }
1385 }
1386 assert(promoInfo->has_spooling_space(), "Control point invariant");
1387 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1388 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1389 if (obj_ptr == NULL) {
1390 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1391 if (obj_ptr == NULL) {
1392 return NULL;
1393 }
1394 }
1395 oop obj = oop(obj_ptr);
1396 OrderAccess::storestore();
1397 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1398 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1399 // IMPORTANT: See note on object initialization for CMS above.
1400 // Otherwise, copy the object. Here we must be careful to insert the
1401 // klass pointer last, since this marks the block as an allocated object.
1402 // Except with compressed oops it's the mark word.
1403 HeapWord* old_ptr = (HeapWord*)old;
1404 // Restore the mark word copied above.
1405 obj->set_mark(m);
1406 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1407 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1408 OrderAccess::storestore();
1409
1410 if (UseCompressedClassPointers) {
1411 // Copy gap missed by (aligned) header size calculation below
1412 obj->set_klass_gap(old->klass_gap());
1413 }
1414 if (word_sz > (size_t)oopDesc::header_size()) {
1415 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1416 obj_ptr + oopDesc::header_size(),
1417 word_sz - oopDesc::header_size());
1418 }
1419
1420 // Now we can track the promoted object, if necessary. We take care
1421 // to delay the transition from uninitialized to full object
1422 // (i.e., insertion of klass pointer) until after, so that it
1423 // atomically becomes a promoted object.
1424 if (promoInfo->tracking()) {
1425 promoInfo->track((PromotedObject*)obj, old->klass());
1426 }
1427 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1428 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1429 assert(old->is_oop(), "Will use and dereference old klass ptr below");
1430
1431 // Finally, install the klass pointer (this should be volatile).
1432 OrderAccess::storestore();
1433 obj->set_klass(old->klass());
1434 // We should now be able to calculate the right size for this object
1435 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1436
1437 collector()->promoted(true, // parallel
1438 obj_ptr, old->is_objArray(), word_sz);
1439
1440 NOT_PRODUCT(
1441 Atomic::inc_ptr(&_numObjectsPromoted);
1442 Atomic::add_ptr(alloc_sz, &_numWordsPromoted);
1443 )
1444
1445 return obj;
1446 }
1447
1448 void
1449 ConcurrentMarkSweepGeneration::
1450 par_promote_alloc_undo(int thread_num,
1451 HeapWord* obj, size_t word_sz) {
1452 // CMS does not support promotion undo.
1453 ShouldNotReachHere();
1454 }
1455
1456 void
1457 ConcurrentMarkSweepGeneration::
1458 par_promote_alloc_done(int thread_num) {
1459 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1460 ps->lab.retire(thread_num);
1461 }
1462
1463 void
1464 ConcurrentMarkSweepGeneration::
1465 par_oop_since_save_marks_iterate_done(int thread_num) {
1466 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1467 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1468 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1469 }
1470
1471 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1472 size_t size,
1473 bool tlab)
1474 {
1475 // We allow a STW collection only if a full
1476 // collection was requested.
1477 return full || should_allocate(size, tlab); // FIX ME !!!
1478 // This and promotion failure handling are connected at the
1479 // hip and should be fixed by untying them.
1480 }
1481
1482 bool CMSCollector::shouldConcurrentCollect() {
1483 if (_full_gc_requested) {
1484 if (Verbose && PrintGCDetails) {
1485 gclog_or_tty->print_cr("CMSCollector: collect because of explicit "
1486 " gc request (or gc_locker)");
1487 }
1488 return true;
1489 }
1490
1491 // For debugging purposes, change the type of collection.
1492 // If the rotation is not on the concurrent collection
1493 // type, don't start a concurrent collection.
1494 NOT_PRODUCT(
1495 if (RotateCMSCollectionTypes &&
1496 (_cmsGen->debug_collection_type() !=
1497 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) {
1498 assert(_cmsGen->debug_collection_type() !=
1499 ConcurrentMarkSweepGeneration::Unknown_collection_type,
1500 "Bad cms collection type");
1501 return false;
1502 }
1503 )
1504
1505 FreelistLocker x(this);
1506 // ------------------------------------------------------------------
1507 // Print out lots of information which affects the initiation of
1508 // a collection.
1509 if (PrintCMSInitiationStatistics && stats().valid()) {
1510 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: ");
1511 gclog_or_tty->stamp();
1512 gclog_or_tty->print_cr("");
1513 stats().print_on(gclog_or_tty);
1514 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f",
1515 stats().time_until_cms_gen_full());
1516 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free());
1517 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT,
1518 _cmsGen->contiguous_available());
1519 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate());
1520 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate());
1521 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy());
1522 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1523 gclog_or_tty->print_cr("metadata initialized %d",
1524 MetaspaceGC::should_concurrent_collect());
1525 }
1526 // ------------------------------------------------------------------
1527
1528 // If the estimated time to complete a cms collection (cms_duration())
1529 // is less than the estimated time remaining until the cms generation
1530 // is full, start a collection.
1531 if (!UseCMSInitiatingOccupancyOnly) {
1532 if (stats().valid()) {
1533 if (stats().time_until_cms_start() == 0.0) {
1534 return true;
1535 }
1536 } else {
1537 // We want to conservatively collect somewhat early in order
1538 // to try and "bootstrap" our CMS/promotion statistics;
1539 // this branch will not fire after the first successful CMS
1540 // collection because the stats should then be valid.
1541 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1542 if (Verbose && PrintGCDetails) {
1543 gclog_or_tty->print_cr(
1544 " CMSCollector: collect for bootstrapping statistics:"
1545 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(),
1546 _bootstrap_occupancy);
1547 }
1548 return true;
1549 }
1550 }
1551 }
1552
1553 // Otherwise, we start a collection cycle if
1554 // old gen want a collection cycle started. Each may use
1555 // an appropriate criterion for making this decision.
1556 // XXX We need to make sure that the gen expansion
1557 // criterion dovetails well with this. XXX NEED TO FIX THIS
1558 if (_cmsGen->should_concurrent_collect()) {
1559 if (Verbose && PrintGCDetails) {
1560 gclog_or_tty->print_cr("CMS old gen initiated");
1561 }
1562 return true;
1563 }
1564
1565 // We start a collection if we believe an incremental collection may fail;
1566 // this is not likely to be productive in practice because it's probably too
1567 // late anyway.
1568 GenCollectedHeap* gch = GenCollectedHeap::heap();
1569 assert(gch->collector_policy()->is_two_generation_policy(),
1570 "You may want to check the correctness of the following");
1571 if (gch->incremental_collection_will_fail(true /* consult_young */)) {
1572 if (Verbose && PrintGCDetails) {
1573 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail ");
1574 }
1575 return true;
1576 }
1577
1578 if (MetaspaceGC::should_concurrent_collect()) {
1579 if (Verbose && PrintGCDetails) {
1580 gclog_or_tty->print("CMSCollector: collect for metadata allocation ");
1581 }
1582 return true;
1583 }
1584
1585 return false;
1586 }
1587
1588 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1589
1590 // Clear _expansion_cause fields of constituent generations
1591 void CMSCollector::clear_expansion_cause() {
1592 _cmsGen->clear_expansion_cause();
1593 }
1594
1595 // We should be conservative in starting a collection cycle. To
1596 // start too eagerly runs the risk of collecting too often in the
1597 // extreme. To collect too rarely falls back on full collections,
1598 // which works, even if not optimum in terms of concurrent work.
1599 // As a work around for too eagerly collecting, use the flag
1600 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1601 // giving the user an easily understandable way of controlling the
1602 // collections.
1603 // We want to start a new collection cycle if any of the following
1604 // conditions hold:
1605 // . our current occupancy exceeds the configured initiating occupancy
1606 // for this generation, or
1607 // . we recently needed to expand this space and have not, since that
1608 // expansion, done a collection of this generation, or
1609 // . the underlying space believes that it may be a good idea to initiate
1610 // a concurrent collection (this may be based on criteria such as the
1611 // following: the space uses linear allocation and linear allocation is
1612 // going to fail, or there is believed to be excessive fragmentation in
1613 // the generation, etc... or ...
1614 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1615 // the case of the old generation; see CR 6543076):
1616 // we may be approaching a point at which allocation requests may fail because
1617 // we will be out of sufficient free space given allocation rate estimates.]
1618 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1619
1620 assert_lock_strong(freelistLock());
1621 if (occupancy() > initiating_occupancy()) {
1622 if (PrintGCDetails && Verbose) {
1623 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ",
1624 short_name(), occupancy(), initiating_occupancy());
1625 }
1626 return true;
1627 }
1628 if (UseCMSInitiatingOccupancyOnly) {
1629 return false;
1630 }
1631 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1632 if (PrintGCDetails && Verbose) {
1633 gclog_or_tty->print(" %s: collect because expanded for allocation ",
1634 short_name());
1635 }
1636 return true;
1637 }
1638 if (_cmsSpace->should_concurrent_collect()) {
1639 if (PrintGCDetails && Verbose) {
1640 gclog_or_tty->print(" %s: collect because cmsSpace says so ",
1641 short_name());
1642 }
1643 return true;
1644 }
1645 return false;
1646 }
1647
1648 void ConcurrentMarkSweepGeneration::collect(bool full,
1649 bool clear_all_soft_refs,
1650 size_t size,
1651 bool tlab)
1652 {
1653 collector()->collect(full, clear_all_soft_refs, size, tlab);
1654 }
1655
1656 void CMSCollector::collect(bool full,
1657 bool clear_all_soft_refs,
1658 size_t size,
1659 bool tlab)
1660 {
1661 if (!UseCMSCollectionPassing && _collectorState > Idling) {
1662 // For debugging purposes skip the collection if the state
1663 // is not currently idle
1664 if (TraceCMSState) {
1665 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d",
1666 Thread::current(), full, _collectorState);
1667 }
1668 return;
1669 }
1670
1671 // The following "if" branch is present for defensive reasons.
1672 // In the current uses of this interface, it can be replaced with:
1673 // assert(!GC_locker.is_active(), "Can't be called otherwise");
1674 // But I am not placing that assert here to allow future
1675 // generality in invoking this interface.
1676 if (GC_locker::is_active()) {
1677 // A consistency test for GC_locker
1678 assert(GC_locker::needs_gc(), "Should have been set already");
1679 // Skip this foreground collection, instead
1680 // expanding the heap if necessary.
1681 // Need the free list locks for the call to free() in compute_new_size()
1682 compute_new_size();
1683 return;
1684 }
1685 acquire_control_and_collect(full, clear_all_soft_refs);
1686 _full_gcs_since_conc_gc++;
1687 }
1688
1689 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1690 GenCollectedHeap* gch = GenCollectedHeap::heap();
1691 unsigned int gc_count = gch->total_full_collections();
1692 if (gc_count == full_gc_count) {
1693 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1694 _full_gc_requested = true;
1695 _full_gc_cause = cause;
1696 CGC_lock->notify(); // nudge CMS thread
1697 } else {
1698 assert(gc_count > full_gc_count, "Error: causal loop");
1699 }
1700 }
1701
1702 bool CMSCollector::is_external_interruption() {
1703 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause();
1704 return GCCause::is_user_requested_gc(cause) ||
1705 GCCause::is_serviceability_requested_gc(cause);
1706 }
1707
1708 void CMSCollector::report_concurrent_mode_interruption() {
1709 if (is_external_interruption()) {
1710 if (PrintGCDetails) {
1711 gclog_or_tty->print(" (concurrent mode interrupted)");
1712 }
1713 } else {
1714 if (PrintGCDetails) {
1715 gclog_or_tty->print(" (concurrent mode failure)");
1716 }
1717 _gc_tracer_cm->report_concurrent_mode_failure();
1718 }
1719 }
1720
1721
1722 // The foreground and background collectors need to coordinate in order
1723 // to make sure that they do not mutually interfere with CMS collections.
1724 // When a background collection is active,
1725 // the foreground collector may need to take over (preempt) and
1726 // synchronously complete an ongoing collection. Depending on the
1727 // frequency of the background collections and the heap usage
1728 // of the application, this preemption can be seldom or frequent.
1729 // There are only certain
1730 // points in the background collection that the "collection-baton"
1731 // can be passed to the foreground collector.
1732 //
1733 // The foreground collector will wait for the baton before
1734 // starting any part of the collection. The foreground collector
1735 // will only wait at one location.
1736 //
1737 // The background collector will yield the baton before starting a new
1738 // phase of the collection (e.g., before initial marking, marking from roots,
1739 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1740 // of the loop which switches the phases. The background collector does some
1741 // of the phases (initial mark, final re-mark) with the world stopped.
1742 // Because of locking involved in stopping the world,
1743 // the foreground collector should not block waiting for the background
1744 // collector when it is doing a stop-the-world phase. The background
1745 // collector will yield the baton at an additional point just before
1746 // it enters a stop-the-world phase. Once the world is stopped, the
1747 // background collector checks the phase of the collection. If the
1748 // phase has not changed, it proceeds with the collection. If the
1749 // phase has changed, it skips that phase of the collection. See
1750 // the comments on the use of the Heap_lock in collect_in_background().
1751 //
1752 // Variable used in baton passing.
1753 // _foregroundGCIsActive - Set to true by the foreground collector when
1754 // it wants the baton. The foreground clears it when it has finished
1755 // the collection.
1756 // _foregroundGCShouldWait - Set to true by the background collector
1757 // when it is running. The foreground collector waits while
1758 // _foregroundGCShouldWait is true.
1759 // CGC_lock - monitor used to protect access to the above variables
1760 // and to notify the foreground and background collectors.
1761 // _collectorState - current state of the CMS collection.
1762 //
1763 // The foreground collector
1764 // acquires the CGC_lock
1765 // sets _foregroundGCIsActive
1766 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1767 // various locks acquired in preparation for the collection
1768 // are released so as not to block the background collector
1769 // that is in the midst of a collection
1770 // proceeds with the collection
1771 // clears _foregroundGCIsActive
1772 // returns
1773 //
1774 // The background collector in a loop iterating on the phases of the
1775 // collection
1776 // acquires the CGC_lock
1777 // sets _foregroundGCShouldWait
1778 // if _foregroundGCIsActive is set
1779 // clears _foregroundGCShouldWait, notifies _CGC_lock
1780 // waits on _CGC_lock for _foregroundGCIsActive to become false
1781 // and exits the loop.
1782 // otherwise
1783 // proceed with that phase of the collection
1784 // if the phase is a stop-the-world phase,
1785 // yield the baton once more just before enqueueing
1786 // the stop-world CMS operation (executed by the VM thread).
1787 // returns after all phases of the collection are done
1788 //
1789
1790 void CMSCollector::acquire_control_and_collect(bool full,
1791 bool clear_all_soft_refs) {
1792 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1793 assert(!Thread::current()->is_ConcurrentGC_thread(),
1794 "shouldn't try to acquire control from self!");
1795
1796 // Start the protocol for acquiring control of the
1797 // collection from the background collector (aka CMS thread).
1798 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1799 "VM thread should have CMS token");
1800 // Remember the possibly interrupted state of an ongoing
1801 // concurrent collection
1802 CollectorState first_state = _collectorState;
1803
1804 // Signal to a possibly ongoing concurrent collection that
1805 // we want to do a foreground collection.
1806 _foregroundGCIsActive = true;
1807
1808 // Disable incremental mode during a foreground collection.
1809 ICMSDisabler icms_disabler;
1810
1811 // release locks and wait for a notify from the background collector
1812 // releasing the locks in only necessary for phases which
1813 // do yields to improve the granularity of the collection.
1814 assert_lock_strong(bitMapLock());
1815 // We need to lock the Free list lock for the space that we are
1816 // currently collecting.
1817 assert(haveFreelistLocks(), "Must be holding free list locks");
1818 bitMapLock()->unlock();
1819 releaseFreelistLocks();
1820 {
1821 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1822 if (_foregroundGCShouldWait) {
1823 // We are going to be waiting for action for the CMS thread;
1824 // it had better not be gone (for instance at shutdown)!
1825 assert(ConcurrentMarkSweepThread::cmst() != NULL,
1826 "CMS thread must be running");
1827 // Wait here until the background collector gives us the go-ahead
1828 ConcurrentMarkSweepThread::clear_CMS_flag(
1829 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1830 // Get a possibly blocked CMS thread going:
1831 // Note that we set _foregroundGCIsActive true above,
1832 // without protection of the CGC_lock.
1833 CGC_lock->notify();
1834 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1835 "Possible deadlock");
1836 while (_foregroundGCShouldWait) {
1837 // wait for notification
1838 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1839 // Possibility of delay/starvation here, since CMS token does
1840 // not know to give priority to VM thread? Actually, i think
1841 // there wouldn't be any delay/starvation, but the proof of
1842 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1843 }
1844 ConcurrentMarkSweepThread::set_CMS_flag(
1845 ConcurrentMarkSweepThread::CMS_vm_has_token);
1846 }
1847 }
1848 // The CMS_token is already held. Get back the other locks.
1849 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1850 "VM thread should have CMS token");
1851 getFreelistLocks();
1852 bitMapLock()->lock_without_safepoint_check();
1853 if (TraceCMSState) {
1854 gclog_or_tty->print_cr("CMS foreground collector has asked for control "
1855 INTPTR_FORMAT " with first state %d", Thread::current(), first_state);
1856 gclog_or_tty->print_cr(" gets control with state %d", _collectorState);
1857 }
1858
1859 // Check if we need to do a compaction, or if not, whether
1860 // we need to start the mark-sweep from scratch.
1861 bool should_compact = false;
1862 bool should_start_over = false;
1863 decide_foreground_collection_type(clear_all_soft_refs,
1864 &should_compact, &should_start_over);
1865
1866 NOT_PRODUCT(
1867 if (RotateCMSCollectionTypes) {
1868 if (_cmsGen->debug_collection_type() ==
1869 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) {
1870 should_compact = true;
1871 } else if (_cmsGen->debug_collection_type() ==
1872 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) {
1873 should_compact = false;
1874 }
1875 }
1876 )
1877
1878 if (first_state > Idling) {
1879 report_concurrent_mode_interruption();
1880 }
1881
1882 set_did_compact(should_compact);
1883 if (should_compact) {
1884 // If the collection is being acquired from the background
1885 // collector, there may be references on the discovered
1886 // references lists that have NULL referents (being those
1887 // that were concurrently cleared by a mutator) or
1888 // that are no longer active (having been enqueued concurrently
1889 // by the mutator).
1890 // Scrub the list of those references because Mark-Sweep-Compact
1891 // code assumes referents are not NULL and that all discovered
1892 // Reference objects are active.
1893 ref_processor()->clean_up_discovered_references();
1894
1895 if (first_state > Idling) {
1896 save_heap_summary();
1897 }
1898
1899 do_compaction_work(clear_all_soft_refs);
1900
1901 // Has the GC time limit been exceeded?
1902 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration();
1903 size_t max_eden_size = young_gen->max_capacity() -
1904 young_gen->to()->capacity() -
1905 young_gen->from()->capacity();
1906 GenCollectedHeap* gch = GenCollectedHeap::heap();
1907 GCCause::Cause gc_cause = gch->gc_cause();
1908 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1909 young_gen->eden()->used(),
1910 _cmsGen->max_capacity(),
1911 max_eden_size,
1912 full,
1913 gc_cause,
1914 gch->collector_policy());
1915 } else {
1916 do_mark_sweep_work(clear_all_soft_refs, first_state,
1917 should_start_over);
1918 }
1919 // Reset the expansion cause, now that we just completed
1920 // a collection cycle.
1921 clear_expansion_cause();
1922 _foregroundGCIsActive = false;
1923 return;
1924 }
1925
1926 // Resize the tenured generation
1927 // after obtaining the free list locks for the
1928 // two generations.
1929 void CMSCollector::compute_new_size() {
1930 assert_locked_or_safepoint(Heap_lock);
1931 FreelistLocker z(this);
1932 MetaspaceGC::compute_new_size();
1933 _cmsGen->compute_new_size_free_list();
1934 }
1935
1936 // A work method used by foreground collection to determine
1937 // what type of collection (compacting or not, continuing or fresh)
1938 // it should do.
1939 // NOTE: the intent is to make UseCMSCompactAtFullCollection
1940 // and CMSCompactWhenClearAllSoftRefs the default in the future
1941 // and do away with the flags after a suitable period.
1942 void CMSCollector::decide_foreground_collection_type(
1943 bool clear_all_soft_refs, bool* should_compact,
1944 bool* should_start_over) {
1945 // Normally, we'll compact only if the UseCMSCompactAtFullCollection
1946 // flag is set, and we have either requested a System.gc() or
1947 // the number of full gc's since the last concurrent cycle
1948 // has exceeded the threshold set by CMSFullGCsBeforeCompaction,
1949 // or if an incremental collection has failed
1950 GenCollectedHeap* gch = GenCollectedHeap::heap();
1951 assert(gch->collector_policy()->is_two_generation_policy(),
1952 "You may want to check the correctness of the following");
1953 // Inform cms gen if this was due to partial collection failing.
1954 // The CMS gen may use this fact to determine its expansion policy.
1955 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) {
1956 assert(!_cmsGen->incremental_collection_failed(),
1957 "Should have been noticed, reacted to and cleared");
1958 _cmsGen->set_incremental_collection_failed();
1959 }
1960 *should_compact =
1961 UseCMSCompactAtFullCollection &&
1962 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) ||
1963 GCCause::is_user_requested_gc(gch->gc_cause()) ||
1964 gch->incremental_collection_will_fail(true /* consult_young */));
1965 *should_start_over = false;
1966 if (clear_all_soft_refs && !*should_compact) {
1967 // We are about to do a last ditch collection attempt
1968 // so it would normally make sense to do a compaction
1969 // to reclaim as much space as possible.
1970 if (CMSCompactWhenClearAllSoftRefs) {
1971 // Default: The rationale is that in this case either
1972 // we are past the final marking phase, in which case
1973 // we'd have to start over, or so little has been done
1974 // that there's little point in saving that work. Compaction
1975 // appears to be the sensible choice in either case.
1976 *should_compact = true;
1977 } else {
1978 // We have been asked to clear all soft refs, but not to
1979 // compact. Make sure that we aren't past the final checkpoint
1980 // phase, for that is where we process soft refs. If we are already
1981 // past that phase, we'll need to redo the refs discovery phase and
1982 // if necessary clear soft refs that weren't previously
1983 // cleared. We do so by remembering the phase in which
1984 // we came in, and if we are past the refs processing
1985 // phase, we'll choose to just redo the mark-sweep
1986 // collection from scratch.
1987 if (_collectorState > FinalMarking) {
1988 // We are past the refs processing phase;
1989 // start over and do a fresh synchronous CMS cycle
1990 _collectorState = Resetting; // skip to reset to start new cycle
1991 reset(false /* == !asynch */);
1992 *should_start_over = true;
1993 } // else we can continue a possibly ongoing current cycle
1994 }
1995 }
1996 }
1997
1998 // A work method used by the foreground collector to do
1999 // a mark-sweep-compact.
2000 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
2001 GenCollectedHeap* gch = GenCollectedHeap::heap();
2002
2003 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
2004 gc_timer->register_gc_start();
2005
2006 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
2007 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start());
2008
2009 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL);
2010 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) {
2011 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d "
2012 "collections passed to foreground collector", _full_gcs_since_conc_gc);
2013 }
2014
2015 // Sample collection interval time and reset for collection pause.
2016 if (UseAdaptiveSizePolicy) {
2017 size_policy()->msc_collection_begin();
2018 }
2019
2020 // Temporarily widen the span of the weak reference processing to
2021 // the entire heap.
2022 MemRegion new_span(GenCollectedHeap::heap()->reserved_region());
2023 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
2024 // Temporarily, clear the "is_alive_non_header" field of the
2025 // reference processor.
2026 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
2027 // Temporarily make reference _processing_ single threaded (non-MT).
2028 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
2029 // Temporarily make refs discovery atomic
2030 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
2031 // Temporarily make reference _discovery_ single threaded (non-MT)
2032 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
2033
2034 ref_processor()->set_enqueuing_is_done(false);
2035 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/);
2036 ref_processor()->setup_policy(clear_all_soft_refs);
2037 // If an asynchronous collection finishes, the _modUnionTable is
2038 // all clear. If we are assuming the collection from an asynchronous
2039 // collection, clear the _modUnionTable.
2040 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
2041 "_modUnionTable should be clear if the baton was not passed");
2042 _modUnionTable.clear_all();
2043 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(),
2044 "mod union for klasses should be clear if the baton was passed");
2045 _ct->klass_rem_set()->clear_mod_union();
2046
2047 // We must adjust the allocation statistics being maintained
2048 // in the free list space. We do so by reading and clearing
2049 // the sweep timer and updating the block flux rate estimates below.
2050 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
2051 if (_inter_sweep_timer.is_active()) {
2052 _inter_sweep_timer.stop();
2053 // Note that we do not use this sample to update the _inter_sweep_estimate.
2054 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
2055 _inter_sweep_estimate.padded_average(),
2056 _intra_sweep_estimate.padded_average());
2057 }
2058
2059 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(),
2060 ref_processor(), clear_all_soft_refs);
2061 #ifdef ASSERT
2062 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
2063 size_t free_size = cms_space->free();
2064 assert(free_size ==
2065 pointer_delta(cms_space->end(), cms_space->compaction_top())
2066 * HeapWordSize,
2067 "All the free space should be compacted into one chunk at top");
2068 assert(cms_space->dictionary()->total_chunk_size(
2069 debug_only(cms_space->freelistLock())) == 0 ||
2070 cms_space->totalSizeInIndexedFreeLists() == 0,
2071 "All the free space should be in a single chunk");
2072 size_t num = cms_space->totalCount();
2073 assert((free_size == 0 && num == 0) ||
2074 (free_size > 0 && (num == 1 || num == 2)),
2075 "There should be at most 2 free chunks after compaction");
2076 #endif // ASSERT
2077 _collectorState = Resetting;
2078 assert(_restart_addr == NULL,
2079 "Should have been NULL'd before baton was passed");
2080 reset(false /* == !asynch */);
2081 _cmsGen->reset_after_compaction();
2082 _concurrent_cycles_since_last_unload = 0;
2083
2084 // Clear any data recorded in the PLAB chunk arrays.
2085 if (_survivor_plab_array != NULL) {
2086 reset_survivor_plab_arrays();
2087 }
2088
2089 // Adjust the per-size allocation stats for the next epoch.
2090 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
2091 // Restart the "inter sweep timer" for the next epoch.
2092 _inter_sweep_timer.reset();
2093 _inter_sweep_timer.start();
2094
2095 // Sample collection pause time and reset for collection interval.
2096 if (UseAdaptiveSizePolicy) {
2097 size_policy()->msc_collection_end(gch->gc_cause());
2098 }
2099
2100 gc_timer->register_gc_end();
2101
2102 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
2103
2104 // For a mark-sweep-compact, compute_new_size() will be called
2105 // in the heap's do_collection() method.
2106 }
2107
2108 // A work method used by the foreground collector to do
2109 // a mark-sweep, after taking over from a possibly on-going
2110 // concurrent mark-sweep collection.
2111 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs,
2112 CollectorState first_state, bool should_start_over) {
2113 if (PrintGC && Verbose) {
2114 gclog_or_tty->print_cr("Pass concurrent collection to foreground "
2115 "collector with count %d",
2116 _full_gcs_since_conc_gc);
2117 }
2118 switch (_collectorState) {
2119 case Idling:
2120 if (first_state == Idling || should_start_over) {
2121 // The background GC was not active, or should
2122 // restarted from scratch; start the cycle.
2123 _collectorState = InitialMarking;
2124 }
2125 // If first_state was not Idling, then a background GC
2126 // was in progress and has now finished. No need to do it
2127 // again. Leave the state as Idling.
2128 break;
2129 case Precleaning:
2130 // In the foreground case don't do the precleaning since
2131 // it is not done concurrently and there is extra work
2132 // required.
2133 _collectorState = FinalMarking;
2134 }
2135 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause());
2136
2137 // For a mark-sweep, compute_new_size() will be called
2138 // in the heap's do_collection() method.
2139 }
2140
2141
2142 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
2143 DefNewGeneration* dng = _young_gen->as_DefNewGeneration();
2144 EdenSpace* eden_space = dng->eden();
2145 ContiguousSpace* from_space = dng->from();
2146 ContiguousSpace* to_space = dng->to();
2147 // Eden
2148 if (_eden_chunk_array != NULL) {
2149 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2150 eden_space->bottom(), eden_space->top(),
2151 eden_space->end(), eden_space->capacity());
2152 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", "
2153 "_eden_chunk_capacity=" SIZE_FORMAT,
2154 _eden_chunk_index, _eden_chunk_capacity);
2155 for (size_t i = 0; i < _eden_chunk_index; i++) {
2156 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2157 i, _eden_chunk_array[i]);
2158 }
2159 }
2160 // Survivor
2161 if (_survivor_chunk_array != NULL) {
2162 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
2163 from_space->bottom(), from_space->top(),
2164 from_space->end(), from_space->capacity());
2165 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", "
2166 "_survivor_chunk_capacity=" SIZE_FORMAT,
2167 _survivor_chunk_index, _survivor_chunk_capacity);
2168 for (size_t i = 0; i < _survivor_chunk_index; i++) {
2169 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT,
2170 i, _survivor_chunk_array[i]);
2171 }
2172 }
2173 }
2174
2175 void CMSCollector::getFreelistLocks() const {
2176 // Get locks for all free lists in all generations that this
2177 // collector is responsible for
2178 _cmsGen->freelistLock()->lock_without_safepoint_check();
2179 }
2180
2181 void CMSCollector::releaseFreelistLocks() const {
2182 // Release locks for all free lists in all generations that this
2183 // collector is responsible for
2184 _cmsGen->freelistLock()->unlock();
2185 }
2186
2187 bool CMSCollector::haveFreelistLocks() const {
2188 // Check locks for all free lists in all generations that this
2189 // collector is responsible for
2190 assert_lock_strong(_cmsGen->freelistLock());
2191 PRODUCT_ONLY(ShouldNotReachHere());
2192 return true;
2193 }
2194
2195 // A utility class that is used by the CMS collector to
2196 // temporarily "release" the foreground collector from its
2197 // usual obligation to wait for the background collector to
2198 // complete an ongoing phase before proceeding.
2199 class ReleaseForegroundGC: public StackObj {
2200 private:
2201 CMSCollector* _c;
2202 public:
2203 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
2204 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
2205 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2206 // allow a potentially blocked foreground collector to proceed
2207 _c->_foregroundGCShouldWait = false;
2208 if (_c->_foregroundGCIsActive) {
2209 CGC_lock->notify();
2210 }
2211 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2212 "Possible deadlock");
2213 }
2214
2215 ~ReleaseForegroundGC() {
2216 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
2217 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2218 _c->_foregroundGCShouldWait = true;
2219 }
2220 };
2221
2222 // There are separate collect_in_background and collect_in_foreground because of
2223 // the different locking requirements of the background collector and the
2224 // foreground collector. There was originally an attempt to share
2225 // one "collect" method between the background collector and the foreground
2226 // collector but the if-then-else required made it cleaner to have
2227 // separate methods.
2228 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) {
2229 assert(Thread::current()->is_ConcurrentGC_thread(),
2230 "A CMS asynchronous collection is only allowed on a CMS thread.");
2231
2232 GenCollectedHeap* gch = GenCollectedHeap::heap();
2233 {
2234 bool safepoint_check = Mutex::_no_safepoint_check_flag;
2235 MutexLockerEx hl(Heap_lock, safepoint_check);
2236 FreelistLocker fll(this);
2237 MutexLockerEx x(CGC_lock, safepoint_check);
2238 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) {
2239 // The foreground collector is active or we're
2240 // not using asynchronous collections. Skip this
2241 // background collection.
2242 assert(!_foregroundGCShouldWait, "Should be clear");
2243 return;
2244 } else {
2245 assert(_collectorState == Idling, "Should be idling before start.");
2246 _collectorState = InitialMarking;
2247 register_gc_start(cause);
2248 // Reset the expansion cause, now that we are about to begin
2249 // a new cycle.
2250 clear_expansion_cause();
2251
2252 // Clear the MetaspaceGC flag since a concurrent collection
2253 // is starting but also clear it after the collection.
2254 MetaspaceGC::set_should_concurrent_collect(false);
2255 }
2256 // Decide if we want to enable class unloading as part of the
2257 // ensuing concurrent GC cycle.
2258 update_should_unload_classes();
2259 _full_gc_requested = false; // acks all outstanding full gc requests
2260 _full_gc_cause = GCCause::_no_gc;
2261 // Signal that we are about to start a collection
2262 gch->increment_total_full_collections(); // ... starting a collection cycle
2263 _collection_count_start = gch->total_full_collections();
2264 }
2265
2266 // Used for PrintGC
2267 size_t prev_used;
2268 if (PrintGC && Verbose) {
2269 prev_used = _cmsGen->used(); // XXXPERM
2270 }
2271
2272 // The change of the collection state is normally done at this level;
2273 // the exceptions are phases that are executed while the world is
2274 // stopped. For those phases the change of state is done while the
2275 // world is stopped. For baton passing purposes this allows the
2276 // background collector to finish the phase and change state atomically.
2277 // The foreground collector cannot wait on a phase that is done
2278 // while the world is stopped because the foreground collector already
2279 // has the world stopped and would deadlock.
2280 while (_collectorState != Idling) {
2281 if (TraceCMSState) {
2282 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2283 Thread::current(), _collectorState);
2284 }
2285 // The foreground collector
2286 // holds the Heap_lock throughout its collection.
2287 // holds the CMS token (but not the lock)
2288 // except while it is waiting for the background collector to yield.
2289 //
2290 // The foreground collector should be blocked (not for long)
2291 // if the background collector is about to start a phase
2292 // executed with world stopped. If the background
2293 // collector has already started such a phase, the
2294 // foreground collector is blocked waiting for the
2295 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
2296 // are executed in the VM thread.
2297 //
2298 // The locking order is
2299 // PendingListLock (PLL) -- if applicable (FinalMarking)
2300 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
2301 // CMS token (claimed in
2302 // stop_world_and_do() -->
2303 // safepoint_synchronize() -->
2304 // CMSThread::synchronize())
2305
2306 {
2307 // Check if the FG collector wants us to yield.
2308 CMSTokenSync x(true); // is cms thread
2309 if (waitForForegroundGC()) {
2310 // We yielded to a foreground GC, nothing more to be
2311 // done this round.
2312 assert(_foregroundGCShouldWait == false, "We set it to false in "
2313 "waitForForegroundGC()");
2314 if (TraceCMSState) {
2315 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2316 " exiting collection CMS state %d",
2317 Thread::current(), _collectorState);
2318 }
2319 return;
2320 } else {
2321 // The background collector can run but check to see if the
2322 // foreground collector has done a collection while the
2323 // background collector was waiting to get the CGC_lock
2324 // above. If yes, break so that _foregroundGCShouldWait
2325 // is cleared before returning.
2326 if (_collectorState == Idling) {
2327 break;
2328 }
2329 }
2330 }
2331
2332 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
2333 "should be waiting");
2334
2335 switch (_collectorState) {
2336 case InitialMarking:
2337 {
2338 ReleaseForegroundGC x(this);
2339 stats().record_cms_begin();
2340 VM_CMS_Initial_Mark initial_mark_op(this);
2341 VMThread::execute(&initial_mark_op);
2342 }
2343 // The collector state may be any legal state at this point
2344 // since the background collector may have yielded to the
2345 // foreground collector.
2346 break;
2347 case Marking:
2348 // initial marking in checkpointRootsInitialWork has been completed
2349 if (markFromRoots(true)) { // we were successful
2350 assert(_collectorState == Precleaning, "Collector state should "
2351 "have changed");
2352 } else {
2353 assert(_foregroundGCIsActive, "Internal state inconsistency");
2354 }
2355 break;
2356 case Precleaning:
2357 if (UseAdaptiveSizePolicy) {
2358 size_policy()->concurrent_precleaning_begin();
2359 }
2360 // marking from roots in markFromRoots has been completed
2361 preclean();
2362 if (UseAdaptiveSizePolicy) {
2363 size_policy()->concurrent_precleaning_end();
2364 }
2365 assert(_collectorState == AbortablePreclean ||
2366 _collectorState == FinalMarking,
2367 "Collector state should have changed");
2368 break;
2369 case AbortablePreclean:
2370 if (UseAdaptiveSizePolicy) {
2371 size_policy()->concurrent_phases_resume();
2372 }
2373 abortable_preclean();
2374 if (UseAdaptiveSizePolicy) {
2375 size_policy()->concurrent_precleaning_end();
2376 }
2377 assert(_collectorState == FinalMarking, "Collector state should "
2378 "have changed");
2379 break;
2380 case FinalMarking:
2381 {
2382 ReleaseForegroundGC x(this);
2383
2384 VM_CMS_Final_Remark final_remark_op(this);
2385 VMThread::execute(&final_remark_op);
2386 }
2387 assert(_foregroundGCShouldWait, "block post-condition");
2388 break;
2389 case Sweeping:
2390 if (UseAdaptiveSizePolicy) {
2391 size_policy()->concurrent_sweeping_begin();
2392 }
2393 // final marking in checkpointRootsFinal has been completed
2394 sweep(true);
2395 assert(_collectorState == Resizing, "Collector state change "
2396 "to Resizing must be done under the free_list_lock");
2397 _full_gcs_since_conc_gc = 0;
2398
2399 // Stop the timers for adaptive size policy for the concurrent phases
2400 if (UseAdaptiveSizePolicy) {
2401 size_policy()->concurrent_sweeping_end();
2402 size_policy()->concurrent_phases_end(gch->gc_cause(),
2403 gch->prev_gen(_cmsGen)->capacity(),
2404 _cmsGen->free());
2405 }
2406
2407 case Resizing: {
2408 // Sweeping has been completed...
2409 // At this point the background collection has completed.
2410 // Don't move the call to compute_new_size() down
2411 // into code that might be executed if the background
2412 // collection was preempted.
2413 {
2414 ReleaseForegroundGC x(this); // unblock FG collection
2415 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
2416 CMSTokenSync z(true); // not strictly needed.
2417 if (_collectorState == Resizing) {
2418 compute_new_size();
2419 save_heap_summary();
2420 _collectorState = Resetting;
2421 } else {
2422 assert(_collectorState == Idling, "The state should only change"
2423 " because the foreground collector has finished the collection");
2424 }
2425 }
2426 break;
2427 }
2428 case Resetting:
2429 // CMS heap resizing has been completed
2430 reset(true);
2431 assert(_collectorState == Idling, "Collector state should "
2432 "have changed");
2433
2434 MetaspaceGC::set_should_concurrent_collect(false);
2435
2436 stats().record_cms_end();
2437 // Don't move the concurrent_phases_end() and compute_new_size()
2438 // calls to here because a preempted background collection
2439 // has it's state set to "Resetting".
2440 break;
2441 case Idling:
2442 default:
2443 ShouldNotReachHere();
2444 break;
2445 }
2446 if (TraceCMSState) {
2447 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2448 Thread::current(), _collectorState);
2449 }
2450 assert(_foregroundGCShouldWait, "block post-condition");
2451 }
2452
2453 // Should this be in gc_epilogue?
2454 collector_policy()->counters()->update_counters();
2455
2456 {
2457 // Clear _foregroundGCShouldWait and, in the event that the
2458 // foreground collector is waiting, notify it, before
2459 // returning.
2460 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2461 _foregroundGCShouldWait = false;
2462 if (_foregroundGCIsActive) {
2463 CGC_lock->notify();
2464 }
2465 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2466 "Possible deadlock");
2467 }
2468 if (TraceCMSState) {
2469 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2470 " exiting collection CMS state %d",
2471 Thread::current(), _collectorState);
2472 }
2473 if (PrintGC && Verbose) {
2474 _cmsGen->print_heap_change(prev_used);
2475 }
2476 }
2477
2478 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) {
2479 if (!_cms_start_registered) {
2480 register_gc_start(cause);
2481 }
2482 }
2483
2484 void CMSCollector::register_gc_start(GCCause::Cause cause) {
2485 _cms_start_registered = true;
2486 _gc_timer_cm->register_gc_start();
2487 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
2488 }
2489
2490 void CMSCollector::register_gc_end() {
2491 if (_cms_start_registered) {
2492 report_heap_summary(GCWhen::AfterGC);
2493
2494 _gc_timer_cm->register_gc_end();
2495 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
2496 _cms_start_registered = false;
2497 }
2498 }
2499
2500 void CMSCollector::save_heap_summary() {
2501 GenCollectedHeap* gch = GenCollectedHeap::heap();
2502 _last_heap_summary = gch->create_heap_summary();
2503 _last_metaspace_summary = gch->create_metaspace_summary();
2504 }
2505
2506 void CMSCollector::report_heap_summary(GCWhen::Type when) {
2507 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_metaspace_summary);
2508 }
2509
2510 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) {
2511 assert(_foregroundGCIsActive && !_foregroundGCShouldWait,
2512 "Foreground collector should be waiting, not executing");
2513 assert(Thread::current()->is_VM_thread(), "A foreground collection"
2514 "may only be done by the VM Thread with the world stopped");
2515 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
2516 "VM thread should have CMS token");
2517
2518 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose,
2519 true, NULL);)
2520 if (UseAdaptiveSizePolicy) {
2521 size_policy()->ms_collection_begin();
2522 }
2523 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact);
2524
2525 HandleMark hm; // Discard invalid handles created during verification
2526
2527 if (VerifyBeforeGC &&
2528 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2529 Universe::verify();
2530 }
2531
2532 // Snapshot the soft reference policy to be used in this collection cycle.
2533 ref_processor()->setup_policy(clear_all_soft_refs);
2534
2535 bool init_mark_was_synchronous = false; // until proven otherwise
2536 while (_collectorState != Idling) {
2537 if (TraceCMSState) {
2538 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d",
2539 Thread::current(), _collectorState);
2540 }
2541 switch (_collectorState) {
2542 case InitialMarking:
2543 register_foreground_gc_start(cause);
2544 init_mark_was_synchronous = true; // fact to be exploited in re-mark
2545 checkpointRootsInitial(false);
2546 assert(_collectorState == Marking, "Collector state should have changed"
2547 " within checkpointRootsInitial()");
2548 break;
2549 case Marking:
2550 // initial marking in checkpointRootsInitialWork has been completed
2551 if (VerifyDuringGC &&
2552 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2553 Universe::verify("Verify before initial mark: ");
2554 }
2555 {
2556 bool res = markFromRoots(false);
2557 assert(res && _collectorState == FinalMarking, "Collector state should "
2558 "have changed");
2559 break;
2560 }
2561 case FinalMarking:
2562 if (VerifyDuringGC &&
2563 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2564 Universe::verify("Verify before re-mark: ");
2565 }
2566 checkpointRootsFinal(false, clear_all_soft_refs,
2567 init_mark_was_synchronous);
2568 assert(_collectorState == Sweeping, "Collector state should not "
2569 "have changed within checkpointRootsFinal()");
2570 break;
2571 case Sweeping:
2572 // final marking in checkpointRootsFinal has been completed
2573 if (VerifyDuringGC &&
2574 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2575 Universe::verify("Verify before sweep: ");
2576 }
2577 sweep(false);
2578 assert(_collectorState == Resizing, "Incorrect state");
2579 break;
2580 case Resizing: {
2581 // Sweeping has been completed; the actual resize in this case
2582 // is done separately; nothing to be done in this state.
2583 _collectorState = Resetting;
2584 break;
2585 }
2586 case Resetting:
2587 // The heap has been resized.
2588 if (VerifyDuringGC &&
2589 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2590 Universe::verify("Verify before reset: ");
2591 }
2592 save_heap_summary();
2593 reset(false);
2594 assert(_collectorState == Idling, "Collector state should "
2595 "have changed");
2596 break;
2597 case Precleaning:
2598 case AbortablePreclean:
2599 // Elide the preclean phase
2600 _collectorState = FinalMarking;
2601 break;
2602 default:
2603 ShouldNotReachHere();
2604 }
2605 if (TraceCMSState) {
2606 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d",
2607 Thread::current(), _collectorState);
2608 }
2609 }
2610
2611 if (UseAdaptiveSizePolicy) {
2612 GenCollectedHeap* gch = GenCollectedHeap::heap();
2613 size_policy()->ms_collection_end(gch->gc_cause());
2614 }
2615
2616 if (VerifyAfterGC &&
2617 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
2618 Universe::verify();
2619 }
2620 if (TraceCMSState) {
2621 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT
2622 " exiting collection CMS state %d",
2623 Thread::current(), _collectorState);
2624 }
2625 }
2626
2627 bool CMSCollector::waitForForegroundGC() {
2628 bool res = false;
2629 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
2630 "CMS thread should have CMS token");
2631 // Block the foreground collector until the
2632 // background collectors decides whether to
2633 // yield.
2634 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
2635 _foregroundGCShouldWait = true;
2636 if (_foregroundGCIsActive) {
2637 // The background collector yields to the
2638 // foreground collector and returns a value
2639 // indicating that it has yielded. The foreground
2640 // collector can proceed.
2641 res = true;
2642 _foregroundGCShouldWait = false;
2643 ConcurrentMarkSweepThread::clear_CMS_flag(
2644 ConcurrentMarkSweepThread::CMS_cms_has_token);
2645 ConcurrentMarkSweepThread::set_CMS_flag(
2646 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2647 // Get a possibly blocked foreground thread going
2648 CGC_lock->notify();
2649 if (TraceCMSState) {
2650 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
2651 Thread::current(), _collectorState);
2652 }
2653 while (_foregroundGCIsActive) {
2654 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
2655 }
2656 ConcurrentMarkSweepThread::set_CMS_flag(
2657 ConcurrentMarkSweepThread::CMS_cms_has_token);
2658 ConcurrentMarkSweepThread::clear_CMS_flag(
2659 ConcurrentMarkSweepThread::CMS_cms_wants_token);
2660 }
2661 if (TraceCMSState) {
2662 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
2663 Thread::current(), _collectorState);
2664 }
2665 return res;
2666 }
2667
2668 // Because of the need to lock the free lists and other structures in
2669 // the collector, common to all the generations that the collector is
2670 // collecting, we need the gc_prologues of individual CMS generations
2671 // delegate to their collector. It may have been simpler had the
2672 // current infrastructure allowed one to call a prologue on a
2673 // collector. In the absence of that we have the generation's
2674 // prologue delegate to the collector, which delegates back
2675 // some "local" work to a worker method in the individual generations
2676 // that it's responsible for collecting, while itself doing any
2677 // work common to all generations it's responsible for. A similar
2678 // comment applies to the gc_epilogue()'s.
2679 // The role of the varaible _between_prologue_and_epilogue is to
2680 // enforce the invocation protocol.
2681 void CMSCollector::gc_prologue(bool full) {
2682 // Call gc_prologue_work() for the CMSGen
2683 // we are responsible for.
2684
2685 // The following locking discipline assumes that we are only called
2686 // when the world is stopped.
2687 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
2688
2689 // The CMSCollector prologue must call the gc_prologues for the
2690 // "generations" that it's responsible
2691 // for.
2692
2693 assert( Thread::current()->is_VM_thread()
2694 || ( CMSScavengeBeforeRemark
2695 && Thread::current()->is_ConcurrentGC_thread()),
2696 "Incorrect thread type for prologue execution");
2697
2698 if (_between_prologue_and_epilogue) {
2699 // We have already been invoked; this is a gc_prologue delegation
2700 // from yet another CMS generation that we are responsible for, just
2701 // ignore it since all relevant work has already been done.
2702 return;
2703 }
2704
2705 // set a bit saying prologue has been called; cleared in epilogue
2706 _between_prologue_and_epilogue = true;
2707 // Claim locks for common data structures, then call gc_prologue_work()
2708 // for each CMSGen.
2709
2710 getFreelistLocks(); // gets free list locks on constituent spaces
2711 bitMapLock()->lock_without_safepoint_check();
2712
2713 // Should call gc_prologue_work() for all cms gens we are responsible for
2714 bool duringMarking = _collectorState >= Marking
2715 && _collectorState < Sweeping;
2716
2717 // The young collections clear the modified oops state, which tells if
2718 // there are any modified oops in the class. The remark phase also needs
2719 // that information. Tell the young collection to save the union of all
2720 // modified klasses.
2721 if (duringMarking) {
2722 _ct->klass_rem_set()->set_accumulate_modified_oops(true);
2723 }
2724
2725 bool registerClosure = duringMarking;
2726
2727 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
2728 &_modUnionClosurePar
2729 : &_modUnionClosure;
2730 _cmsGen->gc_prologue_work(full, registerClosure, muc);
2731
2732 if (!full) {
2733 stats().record_gc0_begin();
2734 }
2735 }
2736
2737 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2738
2739 _capacity_at_prologue = capacity();
2740 _used_at_prologue = used();
2741
2742 // Delegate to CMScollector which knows how to coordinate between
2743 // this and any other CMS generations that it is responsible for
2744 // collecting.
2745 collector()->gc_prologue(full);
2746 }
2747
2748 // This is a "private" interface for use by this generation's CMSCollector.
2749 // Not to be called directly by any other entity (for instance,
2750 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2751 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2752 bool registerClosure, ModUnionClosure* modUnionClosure) {
2753 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2754 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2755 "Should be NULL");
2756 if (registerClosure) {
2757 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2758 }
2759 cmsSpace()->gc_prologue();
2760 // Clear stat counters
2761 NOT_PRODUCT(
2762 assert(_numObjectsPromoted == 0, "check");
2763 assert(_numWordsPromoted == 0, "check");
2764 if (Verbose && PrintGC) {
2765 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, "
2766 SIZE_FORMAT" bytes concurrently",
2767 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2768 }
2769 _numObjectsAllocated = 0;
2770 _numWordsAllocated = 0;
2771 )
2772 }
2773
2774 void CMSCollector::gc_epilogue(bool full) {
2775 // The following locking discipline assumes that we are only called
2776 // when the world is stopped.
2777 assert(SafepointSynchronize::is_at_safepoint(),
2778 "world is stopped assumption");
2779
2780 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2781 // if linear allocation blocks need to be appropriately marked to allow the
2782 // the blocks to be parsable. We also check here whether we need to nudge the
2783 // CMS collector thread to start a new cycle (if it's not already active).
2784 assert( Thread::current()->is_VM_thread()
2785 || ( CMSScavengeBeforeRemark
2786 && Thread::current()->is_ConcurrentGC_thread()),
2787 "Incorrect thread type for epilogue execution");
2788
2789 if (!_between_prologue_and_epilogue) {
2790 // We have already been invoked; this is a gc_epilogue delegation
2791 // from yet another CMS generation that we are responsible for, just
2792 // ignore it since all relevant work has already been done.
2793 return;
2794 }
2795 assert(haveFreelistLocks(), "must have freelist locks");
2796 assert_lock_strong(bitMapLock());
2797
2798 _ct->klass_rem_set()->set_accumulate_modified_oops(false);
2799
2800 _cmsGen->gc_epilogue_work(full);
2801
2802 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2803 // in case sampling was not already enabled, enable it
2804 _start_sampling = true;
2805 }
2806 // reset _eden_chunk_array so sampling starts afresh
2807 _eden_chunk_index = 0;
2808
2809 size_t cms_used = _cmsGen->cmsSpace()->used();
2810
2811 // update performance counters - this uses a special version of
2812 // update_counters() that allows the utilization to be passed as a
2813 // parameter, avoiding multiple calls to used().
2814 //
2815 _cmsGen->update_counters(cms_used);
2816
2817 if (CMSIncrementalMode) {
2818 icms_update_allocation_limits();
2819 }
2820
2821 bitMapLock()->unlock();
2822 releaseFreelistLocks();
2823
2824 if (!CleanChunkPoolAsync) {
2825 Chunk::clean_chunk_pool();
2826 }
2827
2828 set_did_compact(false);
2829 _between_prologue_and_epilogue = false; // ready for next cycle
2830 }
2831
2832 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2833 collector()->gc_epilogue(full);
2834
2835 // Also reset promotion tracking in par gc thread states.
2836 if (CollectedHeap::use_parallel_gc_threads()) {
2837 for (uint i = 0; i < ParallelGCThreads; i++) {
2838 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
2839 }
2840 }
2841 }
2842
2843 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2844 assert(!incremental_collection_failed(), "Should have been cleared");
2845 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2846 cmsSpace()->gc_epilogue();
2847 // Print stat counters
2848 NOT_PRODUCT(
2849 assert(_numObjectsAllocated == 0, "check");
2850 assert(_numWordsAllocated == 0, "check");
2851 if (Verbose && PrintGC) {
2852 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, "
2853 SIZE_FORMAT" bytes",
2854 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2855 }
2856 _numObjectsPromoted = 0;
2857 _numWordsPromoted = 0;
2858 )
2859
2860 if (PrintGC && Verbose) {
2861 // Call down the chain in contiguous_available needs the freelistLock
2862 // so print this out before releasing the freeListLock.
2863 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ",
2864 contiguous_available());
2865 }
2866 }
2867
2868 #ifndef PRODUCT
2869 bool CMSCollector::have_cms_token() {
2870 Thread* thr = Thread::current();
2871 if (thr->is_VM_thread()) {
2872 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2873 } else if (thr->is_ConcurrentGC_thread()) {
2874 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2875 } else if (thr->is_GC_task_thread()) {
2876 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2877 ParGCRareEvent_lock->owned_by_self();
2878 }
2879 return false;
2880 }
2881 #endif
2882
2883 // Check reachability of the given heap address in CMS generation,
2884 // treating all other generations as roots.
2885 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2886 // We could "guarantee" below, rather than assert, but i'll
2887 // leave these as "asserts" so that an adventurous debugger
2888 // could try this in the product build provided some subset of
2889 // the conditions were met, provided they were intersted in the
2890 // results and knew that the computation below wouldn't interfere
2891 // with other concurrent computations mutating the structures
2892 // being read or written.
2893 assert(SafepointSynchronize::is_at_safepoint(),
2894 "Else mutations in object graph will make answer suspect");
2895 assert(have_cms_token(), "Should hold cms token");
2896 assert(haveFreelistLocks(), "must hold free list locks");
2897 assert_lock_strong(bitMapLock());
2898
2899 // Clear the marking bit map array before starting, but, just
2900 // for kicks, first report if the given address is already marked
2901 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr,
2902 _markBitMap.isMarked(addr) ? "" : " not");
2903
2904 if (verify_after_remark()) {
2905 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2906 bool result = verification_mark_bm()->isMarked(addr);
2907 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr,
2908 result ? "IS" : "is NOT");
2909 return result;
2910 } else {
2911 gclog_or_tty->print_cr("Could not compute result");
2912 return false;
2913 }
2914 }
2915
2916
2917 void
2918 CMSCollector::print_on_error(outputStream* st) {
2919 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2920 if (collector != NULL) {
2921 CMSBitMap* bitmap = &collector->_markBitMap;
2922 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap);
2923 bitmap->print_on_error(st, " Bits: ");
2924
2925 st->cr();
2926
2927 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2928 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap);
2929 mut_bitmap->print_on_error(st, " Bits: ");
2930 }
2931 }
2932
2933 ////////////////////////////////////////////////////////
2934 // CMS Verification Support
2935 ////////////////////////////////////////////////////////
2936 // Following the remark phase, the following invariant
2937 // should hold -- each object in the CMS heap which is
2938 // marked in markBitMap() should be marked in the verification_mark_bm().
2939
2940 class VerifyMarkedClosure: public BitMapClosure {
2941 CMSBitMap* _marks;
2942 bool _failed;
2943
2944 public:
2945 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2946
2947 bool do_bit(size_t offset) {
2948 HeapWord* addr = _marks->offsetToHeapWord(offset);
2949 if (!_marks->isMarked(addr)) {
2950 oop(addr)->print_on(gclog_or_tty);
2951 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr);
2952 _failed = true;
2953 }
2954 return true;
2955 }
2956
2957 bool failed() { return _failed; }
2958 };
2959
2960 bool CMSCollector::verify_after_remark(bool silent) {
2961 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... ");
2962 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2963 static bool init = false;
2964
2965 assert(SafepointSynchronize::is_at_safepoint(),
2966 "Else mutations in object graph will make answer suspect");
2967 assert(have_cms_token(),
2968 "Else there may be mutual interference in use of "
2969 " verification data structures");
2970 assert(_collectorState > Marking && _collectorState <= Sweeping,
2971 "Else marking info checked here may be obsolete");
2972 assert(haveFreelistLocks(), "must hold free list locks");
2973 assert_lock_strong(bitMapLock());
2974
2975
2976 // Allocate marking bit map if not already allocated
2977 if (!init) { // first time
2978 if (!verification_mark_bm()->allocate(_span)) {
2979 return false;
2980 }
2981 init = true;
2982 }
2983
2984 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2985
2986 // Turn off refs discovery -- so we will be tracing through refs.
2987 // This is as intended, because by this time
2988 // GC must already have cleared any refs that need to be cleared,
2989 // and traced those that need to be marked; moreover,
2990 // the marking done here is not going to intefere in any
2991 // way with the marking information used by GC.
2992 NoRefDiscovery no_discovery(ref_processor());
2993
2994 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
2995
2996 // Clear any marks from a previous round
2997 verification_mark_bm()->clear_all();
2998 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2999 verify_work_stacks_empty();
3000
3001 GenCollectedHeap* gch = GenCollectedHeap::heap();
3002 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3003 // Update the saved marks which may affect the root scans.
3004 gch->save_marks();
3005
3006 if (CMSRemarkVerifyVariant == 1) {
3007 // In this first variant of verification, we complete
3008 // all marking, then check if the new marks-verctor is
3009 // a subset of the CMS marks-vector.
3010 verify_after_remark_work_1();
3011 } else if (CMSRemarkVerifyVariant == 2) {
3012 // In this second variant of verification, we flag an error
3013 // (i.e. an object reachable in the new marks-vector not reachable
3014 // in the CMS marks-vector) immediately, also indicating the
3015 // identify of an object (A) that references the unmarked object (B) --
3016 // presumably, a mutation to A failed to be picked up by preclean/remark?
3017 verify_after_remark_work_2();
3018 } else {
3019 warning("Unrecognized value %d for CMSRemarkVerifyVariant",
3020 CMSRemarkVerifyVariant);
3021 }
3022 if (!silent) gclog_or_tty->print(" done] ");
3023 return true;
3024 }
3025
3026 void CMSCollector::verify_after_remark_work_1() {
3027 ResourceMark rm;
3028 HandleMark hm;
3029 GenCollectedHeap* gch = GenCollectedHeap::heap();
3030
3031 // Get a clear set of claim bits for the strong roots processing to work with.
3032 ClassLoaderDataGraph::clear_claimed_marks();
3033
3034 // Mark from roots one level into CMS
3035 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
3036 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3037
3038 gch->gen_process_strong_roots(_cmsGen->level(),
3039 true, // younger gens are roots
3040 true, // activate StrongRootsScope
3041 false, // not scavenging
3042 SharedHeap::ScanningOption(roots_scanning_options()),
3043 ¬Older,
3044 true, // walk code active on stacks
3045 NULL,
3046 NULL); // SSS: Provide correct closure
3047
3048 // Now mark from the roots
3049 MarkFromRootsClosure markFromRootsClosure(this, _span,
3050 verification_mark_bm(), verification_mark_stack(),
3051 false /* don't yield */, true /* verifying */);
3052 assert(_restart_addr == NULL, "Expected pre-condition");
3053 verification_mark_bm()->iterate(&markFromRootsClosure);
3054 while (_restart_addr != NULL) {
3055 // Deal with stack overflow: by restarting at the indicated
3056 // address.
3057 HeapWord* ra = _restart_addr;
3058 markFromRootsClosure.reset(ra);
3059 _restart_addr = NULL;
3060 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3061 }
3062 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3063 verify_work_stacks_empty();
3064
3065 // Marking completed -- now verify that each bit marked in
3066 // verification_mark_bm() is also marked in markBitMap(); flag all
3067 // errors by printing corresponding objects.
3068 VerifyMarkedClosure vcl(markBitMap());
3069 verification_mark_bm()->iterate(&vcl);
3070 if (vcl.failed()) {
3071 gclog_or_tty->print("Verification failed");
3072 Universe::heap()->print_on(gclog_or_tty);
3073 fatal("CMS: failed marking verification after remark");
3074 }
3075 }
3076
3077 class VerifyKlassOopsKlassClosure : public KlassClosure {
3078 class VerifyKlassOopsClosure : public OopClosure {
3079 CMSBitMap* _bitmap;
3080 public:
3081 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
3082 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
3083 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3084 } _oop_closure;
3085 public:
3086 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
3087 void do_klass(Klass* k) {
3088 k->oops_do(&_oop_closure);
3089 }
3090 };
3091
3092 void CMSCollector::verify_after_remark_work_2() {
3093 ResourceMark rm;
3094 HandleMark hm;
3095 GenCollectedHeap* gch = GenCollectedHeap::heap();
3096
3097 // Get a clear set of claim bits for the strong roots processing to work with.
3098 ClassLoaderDataGraph::clear_claimed_marks();
3099
3100 // Mark from roots one level into CMS
3101 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
3102 markBitMap());
3103 CMKlassClosure klass_closure(¬Older);
3104
3105 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3106 gch->gen_process_strong_roots(_cmsGen->level(),
3107 true, // younger gens are roots
3108 true, // activate StrongRootsScope
3109 false, // not scavenging
3110 SharedHeap::ScanningOption(roots_scanning_options()),
3111 ¬Older,
3112 true, // walk code active on stacks
3113 NULL,
3114 &klass_closure);
3115
3116 // Now mark from the roots
3117 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
3118 verification_mark_bm(), markBitMap(), verification_mark_stack());
3119 assert(_restart_addr == NULL, "Expected pre-condition");
3120 verification_mark_bm()->iterate(&markFromRootsClosure);
3121 while (_restart_addr != NULL) {
3122 // Deal with stack overflow: by restarting at the indicated
3123 // address.
3124 HeapWord* ra = _restart_addr;
3125 markFromRootsClosure.reset(ra);
3126 _restart_addr = NULL;
3127 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
3128 }
3129 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
3130 verify_work_stacks_empty();
3131
3132 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm());
3133 ClassLoaderDataGraph::classes_do(&verify_klass_oops);
3134
3135 // Marking completed -- now verify that each bit marked in
3136 // verification_mark_bm() is also marked in markBitMap(); flag all
3137 // errors by printing corresponding objects.
3138 VerifyMarkedClosure vcl(markBitMap());
3139 verification_mark_bm()->iterate(&vcl);
3140 assert(!vcl.failed(), "Else verification above should not have succeeded");
3141 }
3142
3143 void ConcurrentMarkSweepGeneration::save_marks() {
3144 // delegate to CMS space
3145 cmsSpace()->save_marks();
3146 for (uint i = 0; i < ParallelGCThreads; i++) {
3147 _par_gc_thread_states[i]->promo.startTrackingPromotions();
3148 }
3149 }
3150
3151 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
3152 return cmsSpace()->no_allocs_since_save_marks();
3153 }
3154
3155 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
3156 \
3157 void ConcurrentMarkSweepGeneration:: \
3158 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
3159 cl->set_generation(this); \
3160 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
3161 cl->reset_generation(); \
3162 save_marks(); \
3163 }
3164
3165 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
3166
3167 void
3168 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) {
3169 cl->set_generation(this);
3170 younger_refs_in_space_iterate(_cmsSpace, cl);
3171 cl->reset_generation();
3172 }
3173
3174 void
3175 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
3176 if (freelistLock()->owned_by_self()) {
3177 Generation::oop_iterate(mr, cl);
3178 } else {
3179 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3180 Generation::oop_iterate(mr, cl);
3181 }
3182 }
3183
3184 void
3185 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
3186 if (freelistLock()->owned_by_self()) {
3187 Generation::oop_iterate(cl);
3188 } else {
3189 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3190 Generation::oop_iterate(cl);
3191 }
3192 }
3193
3194 void
3195 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
3196 if (freelistLock()->owned_by_self()) {
3197 Generation::object_iterate(cl);
3198 } else {
3199 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3200 Generation::object_iterate(cl);
3201 }
3202 }
3203
3204 void
3205 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
3206 if (freelistLock()->owned_by_self()) {
3207 Generation::safe_object_iterate(cl);
3208 } else {
3209 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3210 Generation::safe_object_iterate(cl);
3211 }
3212 }
3213
3214 void
3215 ConcurrentMarkSweepGeneration::post_compact() {
3216 }
3217
3218 void
3219 ConcurrentMarkSweepGeneration::prepare_for_verify() {
3220 // Fix the linear allocation blocks to look like free blocks.
3221
3222 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3223 // are not called when the heap is verified during universe initialization and
3224 // at vm shutdown.
3225 if (freelistLock()->owned_by_self()) {
3226 cmsSpace()->prepare_for_verify();
3227 } else {
3228 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3229 cmsSpace()->prepare_for_verify();
3230 }
3231 }
3232
3233 void
3234 ConcurrentMarkSweepGeneration::verify() {
3235 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
3236 // are not called when the heap is verified during universe initialization and
3237 // at vm shutdown.
3238 if (freelistLock()->owned_by_self()) {
3239 cmsSpace()->verify();
3240 } else {
3241 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
3242 cmsSpace()->verify();
3243 }
3244 }
3245
3246 void CMSCollector::verify() {
3247 _cmsGen->verify();
3248 }
3249
3250 #ifndef PRODUCT
3251 bool CMSCollector::overflow_list_is_empty() const {
3252 assert(_num_par_pushes >= 0, "Inconsistency");
3253 if (_overflow_list == NULL) {
3254 assert(_num_par_pushes == 0, "Inconsistency");
3255 }
3256 return _overflow_list == NULL;
3257 }
3258
3259 // The methods verify_work_stacks_empty() and verify_overflow_empty()
3260 // merely consolidate assertion checks that appear to occur together frequently.
3261 void CMSCollector::verify_work_stacks_empty() const {
3262 assert(_markStack.isEmpty(), "Marking stack should be empty");
3263 assert(overflow_list_is_empty(), "Overflow list should be empty");
3264 }
3265
3266 void CMSCollector::verify_overflow_empty() const {
3267 assert(overflow_list_is_empty(), "Overflow list should be empty");
3268 assert(no_preserved_marks(), "No preserved marks");
3269 }
3270 #endif // PRODUCT
3271
3272 // Decide if we want to enable class unloading as part of the
3273 // ensuing concurrent GC cycle. We will collect and
3274 // unload classes if it's the case that:
3275 // (1) an explicit gc request has been made and the flag
3276 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR
3277 // (2) (a) class unloading is enabled at the command line, and
3278 // (b) old gen is getting really full
3279 // NOTE: Provided there is no change in the state of the heap between
3280 // calls to this method, it should have idempotent results. Moreover,
3281 // its results should be monotonically increasing (i.e. going from 0 to 1,
3282 // but not 1 to 0) between successive calls between which the heap was
3283 // not collected. For the implementation below, it must thus rely on
3284 // the property that concurrent_cycles_since_last_unload()
3285 // will not decrease unless a collection cycle happened and that
3286 // _cmsGen->is_too_full() are
3287 // themselves also monotonic in that sense. See check_monotonicity()
3288 // below.
3289 void CMSCollector::update_should_unload_classes() {
3290 _should_unload_classes = false;
3291 // Condition 1 above
3292 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) {
3293 _should_unload_classes = true;
3294 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above
3295 // Disjuncts 2.b.(i,ii,iii) above
3296 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
3297 CMSClassUnloadingMaxInterval)
3298 || _cmsGen->is_too_full();
3299 }
3300 }
3301
3302 bool ConcurrentMarkSweepGeneration::is_too_full() const {
3303 bool res = should_concurrent_collect();
3304 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
3305 return res;
3306 }
3307
3308 void CMSCollector::setup_cms_unloading_and_verification_state() {
3309 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
3310 || VerifyBeforeExit;
3311 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
3312
3313 if (should_unload_classes()) { // Should unload classes this cycle
3314 remove_root_scanning_option(rso); // Shrink the root set appropriately
3315 set_verifying(should_verify); // Set verification state for this cycle
3316 return; // Nothing else needs to be done at this time
3317 }
3318
3319 // Not unloading classes this cycle
3320 assert(!should_unload_classes(), "Inconsitency!");
3321 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
3322 // Include symbols, strings and code cache elements to prevent their resurrection.
3323 add_root_scanning_option(rso);
3324 set_verifying(true);
3325 } else if (verifying() && !should_verify) {
3326 // We were verifying, but some verification flags got disabled.
3327 set_verifying(false);
3328 // Exclude symbols, strings and code cache elements from root scanning to
3329 // reduce IM and RM pauses.
3330 remove_root_scanning_option(rso);
3331 }
3332 }
3333
3334
3335 #ifndef PRODUCT
3336 HeapWord* CMSCollector::block_start(const void* p) const {
3337 const HeapWord* addr = (HeapWord*)p;
3338 if (_span.contains(p)) {
3339 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
3340 return _cmsGen->cmsSpace()->block_start(p);
3341 }
3342 }
3343 return NULL;
3344 }
3345 #endif
3346
3347 HeapWord*
3348 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
3349 bool tlab,
3350 bool parallel) {
3351 CMSSynchronousYieldRequest yr;
3352 assert(!tlab, "Can't deal with TLAB allocation");
3353 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
3354 expand(word_size*HeapWordSize, MinHeapDeltaBytes,
3355 CMSExpansionCause::_satisfy_allocation);
3356 if (GCExpandToAllocateDelayMillis > 0) {
3357 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3358 }
3359 return have_lock_and_allocate(word_size, tlab);
3360 }
3361
3362 // YSR: All of this generation expansion/shrinking stuff is an exact copy of
3363 // OneContigSpaceCardGeneration, which makes me wonder if we should move this
3364 // to CardGeneration and share it...
3365 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) {
3366 return CardGeneration::expand(bytes, expand_bytes);
3367 }
3368
3369 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes,
3370 CMSExpansionCause::Cause cause)
3371 {
3372
3373 bool success = expand(bytes, expand_bytes);
3374
3375 // remember why we expanded; this information is used
3376 // by shouldConcurrentCollect() when making decisions on whether to start
3377 // a new CMS cycle.
3378 if (success) {
3379 set_expansion_cause(cause);
3380 if (PrintGCDetails && Verbose) {
3381 gclog_or_tty->print_cr("Expanded CMS gen for %s",
3382 CMSExpansionCause::to_string(cause));
3383 }
3384 }
3385 }
3386
3387 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
3388 HeapWord* res = NULL;
3389 MutexLocker x(ParGCRareEvent_lock);
3390 while (true) {
3391 // Expansion by some other thread might make alloc OK now:
3392 res = ps->lab.alloc(word_sz);
3393 if (res != NULL) return res;
3394 // If there's not enough expansion space available, give up.
3395 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
3396 return NULL;
3397 }
3398 // Otherwise, we try expansion.
3399 expand(word_sz*HeapWordSize, MinHeapDeltaBytes,
3400 CMSExpansionCause::_allocate_par_lab);
3401 // Now go around the loop and try alloc again;
3402 // A competing par_promote might beat us to the expansion space,
3403 // so we may go around the loop again if promotion fails agaion.
3404 if (GCExpandToAllocateDelayMillis > 0) {
3405 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3406 }
3407 }
3408 }
3409
3410
3411 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
3412 PromotionInfo* promo) {
3413 MutexLocker x(ParGCRareEvent_lock);
3414 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
3415 while (true) {
3416 // Expansion by some other thread might make alloc OK now:
3417 if (promo->ensure_spooling_space()) {
3418 assert(promo->has_spooling_space(),
3419 "Post-condition of successful ensure_spooling_space()");
3420 return true;
3421 }
3422 // If there's not enough expansion space available, give up.
3423 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
3424 return false;
3425 }
3426 // Otherwise, we try expansion.
3427 expand(refill_size_bytes, MinHeapDeltaBytes,
3428 CMSExpansionCause::_allocate_par_spooling_space);
3429 // Now go around the loop and try alloc again;
3430 // A competing allocation might beat us to the expansion space,
3431 // so we may go around the loop again if allocation fails again.
3432 if (GCExpandToAllocateDelayMillis > 0) {
3433 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
3434 }
3435 }
3436 }
3437
3438
3439 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) {
3440 assert_locked_or_safepoint(ExpandHeap_lock);
3441 // Shrink committed space
3442 _virtual_space.shrink_by(bytes);
3443 // Shrink space; this also shrinks the space's BOT
3444 _cmsSpace->set_end((HeapWord*) _virtual_space.high());
3445 size_t new_word_size = heap_word_size(_cmsSpace->capacity());
3446 // Shrink the shared block offset array
3447 _bts->resize(new_word_size);
3448 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3449 // Shrink the card table
3450 Universe::heap()->barrier_set()->resize_covered_region(mr);
3451
3452 if (Verbose && PrintGC) {
3453 size_t new_mem_size = _virtual_space.committed_size();
3454 size_t old_mem_size = new_mem_size + bytes;
3455 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3456 name(), old_mem_size/K, new_mem_size/K);
3457 }
3458 }
3459
3460 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
3461 assert_locked_or_safepoint(Heap_lock);
3462 size_t size = ReservedSpace::page_align_size_down(bytes);
3463 // Only shrink if a compaction was done so that all the free space
3464 // in the generation is in a contiguous block at the end.
3465 if (size > 0 && did_compact()) {
3466 shrink_by(size);
3467 }
3468 }
3469
3470 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) {
3471 assert_locked_or_safepoint(Heap_lock);
3472 bool result = _virtual_space.expand_by(bytes);
3473 if (result) {
3474 size_t new_word_size =
3475 heap_word_size(_virtual_space.committed_size());
3476 MemRegion mr(_cmsSpace->bottom(), new_word_size);
3477 _bts->resize(new_word_size); // resize the block offset shared array
3478 Universe::heap()->barrier_set()->resize_covered_region(mr);
3479 // Hmmmm... why doesn't CFLS::set_end verify locking?
3480 // This is quite ugly; FIX ME XXX
3481 _cmsSpace->assert_locked(freelistLock());
3482 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
3483
3484 // update the space and generation capacity counters
3485 if (UsePerfData) {
3486 _space_counters->update_capacity();
3487 _gen_counters->update_all();
3488 }
3489
3490 if (Verbose && PrintGC) {
3491 size_t new_mem_size = _virtual_space.committed_size();
3492 size_t old_mem_size = new_mem_size - bytes;
3493 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
3494 name(), old_mem_size/K, bytes/K, new_mem_size/K);
3495 }
3496 }
3497 return result;
3498 }
3499
3500 bool ConcurrentMarkSweepGeneration::grow_to_reserved() {
3501 assert_locked_or_safepoint(Heap_lock);
3502 bool success = true;
3503 const size_t remaining_bytes = _virtual_space.uncommitted_size();
3504 if (remaining_bytes > 0) {
3505 success = grow_by(remaining_bytes);
3506 DEBUG_ONLY(if (!success) warning("grow to reserved failed");)
3507 }
3508 return success;
3509 }
3510
3511 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
3512 assert_locked_or_safepoint(Heap_lock);
3513 assert_lock_strong(freelistLock());
3514 if (PrintGCDetails && Verbose) {
3515 warning("Shrinking of CMS not yet implemented");
3516 }
3517 return;
3518 }
3519
3520
3521 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
3522 // phases.
3523 class CMSPhaseAccounting: public StackObj {
3524 public:
3525 CMSPhaseAccounting(CMSCollector *collector,
3526 const char *phase,
3527 bool print_cr = true);
3528 ~CMSPhaseAccounting();
3529
3530 private:
3531 CMSCollector *_collector;
3532 const char *_phase;
3533 elapsedTimer _wallclock;
3534 bool _print_cr;
3535
3536 public:
3537 // Not MT-safe; so do not pass around these StackObj's
3538 // where they may be accessed by other threads.
3539 jlong wallclock_millis() {
3540 assert(_wallclock.is_active(), "Wall clock should not stop");
3541 _wallclock.stop(); // to record time
3542 jlong ret = _wallclock.milliseconds();
3543 _wallclock.start(); // restart
3544 return ret;
3545 }
3546 };
3547
3548 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
3549 const char *phase,
3550 bool print_cr) :
3551 _collector(collector), _phase(phase), _print_cr(print_cr) {
3552
3553 if (PrintCMSStatistics != 0) {
3554 _collector->resetYields();
3555 }
3556 if (PrintGCDetails) {
3557 gclog_or_tty->date_stamp(PrintGCDateStamps);
3558 gclog_or_tty->stamp(PrintGCTimeStamps);
3559 gclog_or_tty->print_cr("[%s-concurrent-%s-start]",
3560 _collector->cmsGen()->short_name(), _phase);
3561 }
3562 _collector->resetTimer();
3563 _wallclock.start();
3564 _collector->startTimer();
3565 }
3566
3567 CMSPhaseAccounting::~CMSPhaseAccounting() {
3568 assert(_wallclock.is_active(), "Wall clock should not have stopped");
3569 _collector->stopTimer();
3570 _wallclock.stop();
3571 if (PrintGCDetails) {
3572 gclog_or_tty->date_stamp(PrintGCDateStamps);
3573 gclog_or_tty->stamp(PrintGCTimeStamps);
3574 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]",
3575 _collector->cmsGen()->short_name(),
3576 _phase, _collector->timerValue(), _wallclock.seconds());
3577 if (_print_cr) {
3578 gclog_or_tty->print_cr("");
3579 }
3580 if (PrintCMSStatistics != 0) {
3581 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase,
3582 _collector->yields());
3583 }
3584 }
3585 }
3586
3587 // CMS work
3588
3589 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
3590 class CMSParMarkTask : public AbstractGangTask {
3591 protected:
3592 CMSCollector* _collector;
3593 int _n_workers;
3594 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) :
3595 AbstractGangTask(name),
3596 _collector(collector),
3597 _n_workers(n_workers) {}
3598 // Work method in support of parallel rescan ... of young gen spaces
3599 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl,
3600 ContiguousSpace* space,
3601 HeapWord** chunk_array, size_t chunk_top);
3602 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl);
3603 };
3604
3605 // Parallel initial mark task
3606 class CMSParInitialMarkTask: public CMSParMarkTask {
3607 public:
3608 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) :
3609 CMSParMarkTask("Scan roots and young gen for initial mark in parallel",
3610 collector, n_workers) {}
3611 void work(uint worker_id);
3612 };
3613
3614 // Checkpoint the roots into this generation from outside
3615 // this generation. [Note this initial checkpoint need only
3616 // be approximate -- we'll do a catch up phase subsequently.]
3617 void CMSCollector::checkpointRootsInitial(bool asynch) {
3618 assert(_collectorState == InitialMarking, "Wrong collector state");
3619 check_correct_thread_executing();
3620 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
3621
3622 save_heap_summary();
3623 report_heap_summary(GCWhen::BeforeGC);
3624
3625 ReferenceProcessor* rp = ref_processor();
3626 SpecializationStats::clear();
3627 assert(_restart_addr == NULL, "Control point invariant");
3628 if (asynch) {
3629 // acquire locks for subsequent manipulations
3630 MutexLockerEx x(bitMapLock(),
3631 Mutex::_no_safepoint_check_flag);
3632 checkpointRootsInitialWork(asynch);
3633 // enable ("weak") refs discovery
3634 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/);
3635 _collectorState = Marking;
3636 } else {
3637 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection
3638 // which recognizes if we are a CMS generation, and doesn't try to turn on
3639 // discovery; verify that they aren't meddling.
3640 assert(!rp->discovery_is_atomic(),
3641 "incorrect setting of discovery predicate");
3642 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control "
3643 "ref discovery for this generation kind");
3644 // already have locks
3645 checkpointRootsInitialWork(asynch);
3646 // now enable ("weak") refs discovery
3647 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/);
3648 _collectorState = Marking;
3649 }
3650 SpecializationStats::print();
3651 }
3652
3653 void CMSCollector::checkpointRootsInitialWork(bool asynch) {
3654 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
3655 assert(_collectorState == InitialMarking, "just checking");
3656
3657 // If there has not been a GC[n-1] since last GC[n] cycle completed,
3658 // precede our marking with a collection of all
3659 // younger generations to keep floating garbage to a minimum.
3660 // XXX: we won't do this for now -- it's an optimization to be done later.
3661
3662 // already have locks
3663 assert_lock_strong(bitMapLock());
3664 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
3665
3666 // Setup the verification and class unloading state for this
3667 // CMS collection cycle.
3668 setup_cms_unloading_and_verification_state();
3669
3670 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork",
3671 PrintGCDetails && Verbose, true, _gc_timer_cm);)
3672 if (UseAdaptiveSizePolicy) {
3673 size_policy()->checkpoint_roots_initial_begin();
3674 }
3675
3676 // Reset all the PLAB chunk arrays if necessary.
3677 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
3678 reset_survivor_plab_arrays();
3679 }
3680
3681 ResourceMark rm;
3682 HandleMark hm;
3683
3684 FalseClosure falseClosure;
3685 // In the case of a synchronous collection, we will elide the
3686 // remark step, so it's important to catch all the nmethod oops
3687 // in this step.
3688 // The final 'true' flag to gen_process_strong_roots will ensure this.
3689 // If 'async' is true, we can relax the nmethod tracing.
3690 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
3691 GenCollectedHeap* gch = GenCollectedHeap::heap();
3692
3693 verify_work_stacks_empty();
3694 verify_overflow_empty();
3695
3696 gch->ensure_parsability(false); // fill TLABs, but no need to retire them
3697 // Update the saved marks which may affect the root scans.
3698 gch->save_marks();
3699
3700 // weak reference processing has not started yet.
3701 ref_processor()->set_enqueuing_is_done(false);
3702
3703 // Need to remember all newly created CLDs,
3704 // so that we can guarantee that the remark finds them.
3705 ClassLoaderDataGraph::remember_new_clds(true);
3706
3707 // Whenever a CLD is found, it will be claimed before proceeding to mark
3708 // the klasses. The claimed marks need to be cleared before marking starts.
3709 ClassLoaderDataGraph::clear_claimed_marks();
3710
3711 if (CMSPrintEdenSurvivorChunks) {
3712 print_eden_and_survivor_chunk_arrays();
3713 }
3714
3715 {
3716 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
3717 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
3718 // The parallel version.
3719 FlexibleWorkGang* workers = gch->workers();
3720 assert(workers != NULL, "Need parallel worker threads.");
3721 int n_workers = workers->active_workers();
3722 CMSParInitialMarkTask tsk(this, n_workers);
3723 gch->set_par_threads(n_workers);
3724 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
3725 if (n_workers > 1) {
3726 GenCollectedHeap::StrongRootsScope srs(gch);
3727 workers->run_task(&tsk);
3728 } else {
3729 GenCollectedHeap::StrongRootsScope srs(gch);
3730 tsk.work(0);
3731 }
3732 gch->set_par_threads(0);
3733 } else {
3734 // The serial version.
3735 CMKlassClosure klass_closure(¬Older);
3736 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
3737 gch->gen_process_strong_roots(_cmsGen->level(),
3738 true, // younger gens are roots
3739 true, // activate StrongRootsScope
3740 false, // not scavenging
3741 SharedHeap::ScanningOption(roots_scanning_options()),
3742 ¬Older,
3743 true, // walk all of code cache if (so & SO_CodeCache)
3744 NULL,
3745 &klass_closure);
3746 }
3747 }
3748
3749 // Clear mod-union table; it will be dirtied in the prologue of
3750 // CMS generation per each younger generation collection.
3751
3752 assert(_modUnionTable.isAllClear(),
3753 "Was cleared in most recent final checkpoint phase"
3754 " or no bits are set in the gc_prologue before the start of the next "
3755 "subsequent marking phase.");
3756
3757 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be");
3758
3759 // Save the end of the used_region of the constituent generations
3760 // to be used to limit the extent of sweep in each generation.
3761 save_sweep_limits();
3762 if (UseAdaptiveSizePolicy) {
3763 size_policy()->checkpoint_roots_initial_end(gch->gc_cause());
3764 }
3765 verify_overflow_empty();
3766 }
3767
3768 bool CMSCollector::markFromRoots(bool asynch) {
3769 // we might be tempted to assert that:
3770 // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
3771 // "inconsistent argument?");
3772 // However that wouldn't be right, because it's possible that
3773 // a safepoint is indeed in progress as a younger generation
3774 // stop-the-world GC happens even as we mark in this generation.
3775 assert(_collectorState == Marking, "inconsistent state?");
3776 check_correct_thread_executing();
3777 verify_overflow_empty();
3778
3779 bool res;
3780 if (asynch) {
3781
3782 // Start the timers for adaptive size policy for the concurrent phases
3783 // Do it here so that the foreground MS can use the concurrent
3784 // timer since a foreground MS might has the sweep done concurrently
3785 // or STW.
3786 if (UseAdaptiveSizePolicy) {
3787 size_policy()->concurrent_marking_begin();
3788 }
3789
3790 // Weak ref discovery note: We may be discovering weak
3791 // refs in this generation concurrent (but interleaved) with
3792 // weak ref discovery by a younger generation collector.
3793
3794 CMSTokenSyncWithLocks ts(true, bitMapLock());
3795 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
3796 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails);
3797 res = markFromRootsWork(asynch);
3798 if (res) {
3799 _collectorState = Precleaning;
3800 } else { // We failed and a foreground collection wants to take over
3801 assert(_foregroundGCIsActive, "internal state inconsistency");
3802 assert(_restart_addr == NULL, "foreground will restart from scratch");
3803 if (PrintGCDetails) {
3804 gclog_or_tty->print_cr("bailing out to foreground collection");
3805 }
3806 }
3807 if (UseAdaptiveSizePolicy) {
3808 size_policy()->concurrent_marking_end();
3809 }
3810 } else {
3811 assert(SafepointSynchronize::is_at_safepoint(),
3812 "inconsistent with asynch == false");
3813 if (UseAdaptiveSizePolicy) {
3814 size_policy()->ms_collection_marking_begin();
3815 }
3816 // already have locks
3817 res = markFromRootsWork(asynch);
3818 _collectorState = FinalMarking;
3819 if (UseAdaptiveSizePolicy) {
3820 GenCollectedHeap* gch = GenCollectedHeap::heap();
3821 size_policy()->ms_collection_marking_end(gch->gc_cause());
3822 }
3823 }
3824 verify_overflow_empty();
3825 return res;
3826 }
3827
3828 bool CMSCollector::markFromRootsWork(bool asynch) {
3829 // iterate over marked bits in bit map, doing a full scan and mark
3830 // from these roots using the following algorithm:
3831 // . if oop is to the right of the current scan pointer,
3832 // mark corresponding bit (we'll process it later)
3833 // . else (oop is to left of current scan pointer)
3834 // push oop on marking stack
3835 // . drain the marking stack
3836
3837 // Note that when we do a marking step we need to hold the
3838 // bit map lock -- recall that direct allocation (by mutators)
3839 // and promotion (by younger generation collectors) is also
3840 // marking the bit map. [the so-called allocate live policy.]
3841 // Because the implementation of bit map marking is not
3842 // robust wrt simultaneous marking of bits in the same word,
3843 // we need to make sure that there is no such interference
3844 // between concurrent such updates.
3845
3846 // already have locks
3847 assert_lock_strong(bitMapLock());
3848
3849 verify_work_stacks_empty();
3850 verify_overflow_empty();
3851 bool result = false;
3852 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
3853 result = do_marking_mt(asynch);
3854 } else {
3855 result = do_marking_st(asynch);
3856 }
3857 return result;
3858 }
3859
3860 // Forward decl
3861 class CMSConcMarkingTask;
3862
3863 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
3864 CMSCollector* _collector;
3865 CMSConcMarkingTask* _task;
3866 public:
3867 virtual void yield();
3868
3869 // "n_threads" is the number of threads to be terminated.
3870 // "queue_set" is a set of work queues of other threads.
3871 // "collector" is the CMS collector associated with this task terminator.
3872 // "yield" indicates whether we need the gang as a whole to yield.
3873 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3874 ParallelTaskTerminator(n_threads, queue_set),
3875 _collector(collector) { }
3876
3877 void set_task(CMSConcMarkingTask* task) {
3878 _task = task;
3879 }
3880 };
3881
3882 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3883 CMSConcMarkingTask* _task;
3884 public:
3885 bool should_exit_termination();
3886 void set_task(CMSConcMarkingTask* task) {
3887 _task = task;
3888 }
3889 };
3890
3891 // MT Concurrent Marking Task
3892 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3893 CMSCollector* _collector;
3894 int _n_workers; // requested/desired # workers
3895 bool _asynch;
3896 bool _result;
3897 CompactibleFreeListSpace* _cms_space;
3898 char _pad_front[64]; // padding to ...
3899 HeapWord* _global_finger; // ... avoid sharing cache line
3900 char _pad_back[64];
3901 HeapWord* _restart_addr;
3902
3903 // Exposed here for yielding support
3904 Mutex* const _bit_map_lock;
3905
3906 // The per thread work queues, available here for stealing
3907 OopTaskQueueSet* _task_queues;
3908
3909 // Termination (and yielding) support
3910 CMSConcMarkingTerminator _term;
3911 CMSConcMarkingTerminatorTerminator _term_term;
3912
3913 public:
3914 CMSConcMarkingTask(CMSCollector* collector,
3915 CompactibleFreeListSpace* cms_space,
3916 bool asynch,
3917 YieldingFlexibleWorkGang* workers,
3918 OopTaskQueueSet* task_queues):
3919 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3920 _collector(collector),
3921 _cms_space(cms_space),
3922 _asynch(asynch), _n_workers(0), _result(true),
3923 _task_queues(task_queues),
3924 _term(_n_workers, task_queues, _collector),
3925 _bit_map_lock(collector->bitMapLock())
3926 {
3927 _requested_size = _n_workers;
3928 _term.set_task(this);
3929 _term_term.set_task(this);
3930 _restart_addr = _global_finger = _cms_space->bottom();
3931 }
3932
3933
3934 OopTaskQueueSet* task_queues() { return _task_queues; }
3935
3936 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3937
3938 HeapWord** global_finger_addr() { return &_global_finger; }
3939
3940 CMSConcMarkingTerminator* terminator() { return &_term; }
3941
3942 virtual void set_for_termination(int active_workers) {
3943 terminator()->reset_for_reuse(active_workers);
3944 }
3945
3946 void work(uint worker_id);
3947 bool should_yield() {
3948 return ConcurrentMarkSweepThread::should_yield()
3949 && !_collector->foregroundGCIsActive()
3950 && _asynch;
3951 }
3952
3953 virtual void coordinator_yield(); // stuff done by coordinator
3954 bool result() { return _result; }
3955
3956 void reset(HeapWord* ra) {
3957 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3958 _restart_addr = _global_finger = ra;
3959 _term.reset_for_reuse();
3960 }
3961
3962 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3963 OopTaskQueue* work_q);
3964
3965 private:
3966 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3967 void do_work_steal(int i);
3968 void bump_global_finger(HeapWord* f);
3969 };
3970
3971 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3972 assert(_task != NULL, "Error");
3973 return _task->yielding();
3974 // Note that we do not need the disjunct || _task->should_yield() above
3975 // because we want terminating threads to yield only if the task
3976 // is already in the midst of yielding, which happens only after at least one
3977 // thread has yielded.
3978 }
3979
3980 void CMSConcMarkingTerminator::yield() {
3981 if (_task->should_yield()) {
3982 _task->yield();
3983 } else {
3984 ParallelTaskTerminator::yield();
3985 }
3986 }
3987
3988 ////////////////////////////////////////////////////////////////
3989 // Concurrent Marking Algorithm Sketch
3990 ////////////////////////////////////////////////////////////////
3991 // Until all tasks exhausted (both spaces):
3992 // -- claim next available chunk
3993 // -- bump global finger via CAS
3994 // -- find first object that starts in this chunk
3995 // and start scanning bitmap from that position
3996 // -- scan marked objects for oops
3997 // -- CAS-mark target, and if successful:
3998 // . if target oop is above global finger (volatile read)
3999 // nothing to do
4000 // . if target oop is in chunk and above local finger
4001 // then nothing to do
4002 // . else push on work-queue
4003 // -- Deal with possible overflow issues:
4004 // . local work-queue overflow causes stuff to be pushed on
4005 // global (common) overflow queue
4006 // . always first empty local work queue
4007 // . then get a batch of oops from global work queue if any
4008 // . then do work stealing
4009 // -- When all tasks claimed (both spaces)
4010 // and local work queue empty,
4011 // then in a loop do:
4012 // . check global overflow stack; steal a batch of oops and trace
4013 // . try to steal from other threads oif GOS is empty
4014 // . if neither is available, offer termination
4015 // -- Terminate and return result
4016 //
4017 void CMSConcMarkingTask::work(uint worker_id) {
4018 elapsedTimer _timer;
4019 ResourceMark rm;
4020 HandleMark hm;
4021
4022 DEBUG_ONLY(_collector->verify_overflow_empty();)
4023
4024 // Before we begin work, our work queue should be empty
4025 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
4026 // Scan the bitmap covering _cms_space, tracing through grey objects.
4027 _timer.start();
4028 do_scan_and_mark(worker_id, _cms_space);
4029 _timer.stop();
4030 if (PrintCMSStatistics != 0) {
4031 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec",
4032 worker_id, _timer.seconds());
4033 // XXX: need xxx/xxx type of notation, two timers
4034 }
4035
4036 // ... do work stealing
4037 _timer.reset();
4038 _timer.start();
4039 do_work_steal(worker_id);
4040 _timer.stop();
4041 if (PrintCMSStatistics != 0) {
4042 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec",
4043 worker_id, _timer.seconds());
4044 // XXX: need xxx/xxx type of notation, two timers
4045 }
4046 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
4047 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
4048 // Note that under the current task protocol, the
4049 // following assertion is true even of the spaces
4050 // expanded since the completion of the concurrent
4051 // marking. XXX This will likely change under a strict
4052 // ABORT semantics.
4053 // After perm removal the comparison was changed to
4054 // greater than or equal to from strictly greater than.
4055 // Before perm removal the highest address sweep would
4056 // have been at the end of perm gen but now is at the
4057 // end of the tenured gen.
4058 assert(_global_finger >= _cms_space->end(),
4059 "All tasks have been completed");
4060 DEBUG_ONLY(_collector->verify_overflow_empty();)
4061 }
4062
4063 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
4064 HeapWord* read = _global_finger;
4065 HeapWord* cur = read;
4066 while (f > read) {
4067 cur = read;
4068 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur);
4069 if (cur == read) {
4070 // our cas succeeded
4071 assert(_global_finger >= f, "protocol consistency");
4072 break;
4073 }
4074 }
4075 }
4076
4077 // This is really inefficient, and should be redone by
4078 // using (not yet available) block-read and -write interfaces to the
4079 // stack and the work_queue. XXX FIX ME !!!
4080 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
4081 OopTaskQueue* work_q) {
4082 // Fast lock-free check
4083 if (ovflw_stk->length() == 0) {
4084 return false;
4085 }
4086 assert(work_q->size() == 0, "Shouldn't steal");
4087 MutexLockerEx ml(ovflw_stk->par_lock(),
4088 Mutex::_no_safepoint_check_flag);
4089 // Grab up to 1/4 the size of the work queue
4090 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4091 (size_t)ParGCDesiredObjsFromOverflowList);
4092 num = MIN2(num, ovflw_stk->length());
4093 for (int i = (int) num; i > 0; i--) {
4094 oop cur = ovflw_stk->pop();
4095 assert(cur != NULL, "Counted wrong?");
4096 work_q->push(cur);
4097 }
4098 return num > 0;
4099 }
4100
4101 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
4102 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4103 int n_tasks = pst->n_tasks();
4104 // We allow that there may be no tasks to do here because
4105 // we are restarting after a stack overflow.
4106 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
4107 uint nth_task = 0;
4108
4109 HeapWord* aligned_start = sp->bottom();
4110 if (sp->used_region().contains(_restart_addr)) {
4111 // Align down to a card boundary for the start of 0th task
4112 // for this space.
4113 aligned_start =
4114 (HeapWord*)align_size_down((uintptr_t)_restart_addr,
4115 CardTableModRefBS::card_size);
4116 }
4117
4118 size_t chunk_size = sp->marking_task_size();
4119 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4120 // Having claimed the nth task in this space,
4121 // compute the chunk that it corresponds to:
4122 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
4123 aligned_start + (nth_task+1)*chunk_size);
4124 // Try and bump the global finger via a CAS;
4125 // note that we need to do the global finger bump
4126 // _before_ taking the intersection below, because
4127 // the task corresponding to that region will be
4128 // deemed done even if the used_region() expands
4129 // because of allocation -- as it almost certainly will
4130 // during start-up while the threads yield in the
4131 // closure below.
4132 HeapWord* finger = span.end();
4133 bump_global_finger(finger); // atomically
4134 // There are null tasks here corresponding to chunks
4135 // beyond the "top" address of the space.
4136 span = span.intersection(sp->used_region());
4137 if (!span.is_empty()) { // Non-null task
4138 HeapWord* prev_obj;
4139 assert(!span.contains(_restart_addr) || nth_task == 0,
4140 "Inconsistency");
4141 if (nth_task == 0) {
4142 // For the 0th task, we'll not need to compute a block_start.
4143 if (span.contains(_restart_addr)) {
4144 // In the case of a restart because of stack overflow,
4145 // we might additionally skip a chunk prefix.
4146 prev_obj = _restart_addr;
4147 } else {
4148 prev_obj = span.start();
4149 }
4150 } else {
4151 // We want to skip the first object because
4152 // the protocol is to scan any object in its entirety
4153 // that _starts_ in this span; a fortiori, any
4154 // object starting in an earlier span is scanned
4155 // as part of an earlier claimed task.
4156 // Below we use the "careful" version of block_start
4157 // so we do not try to navigate uninitialized objects.
4158 prev_obj = sp->block_start_careful(span.start());
4159 // Below we use a variant of block_size that uses the
4160 // Printezis bits to avoid waiting for allocated
4161 // objects to become initialized/parsable.
4162 while (prev_obj < span.start()) {
4163 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
4164 if (sz > 0) {
4165 prev_obj += sz;
4166 } else {
4167 // In this case we may end up doing a bit of redundant
4168 // scanning, but that appears unavoidable, short of
4169 // locking the free list locks; see bug 6324141.
4170 break;
4171 }
4172 }
4173 }
4174 if (prev_obj < span.end()) {
4175 MemRegion my_span = MemRegion(prev_obj, span.end());
4176 // Do the marking work within a non-empty span --
4177 // the last argument to the constructor indicates whether the
4178 // iteration should be incremental with periodic yields.
4179 Par_MarkFromRootsClosure cl(this, _collector, my_span,
4180 &_collector->_markBitMap,
4181 work_queue(i),
4182 &_collector->_markStack,
4183 _asynch);
4184 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
4185 } // else nothing to do for this task
4186 } // else nothing to do for this task
4187 }
4188 // We'd be tempted to assert here that since there are no
4189 // more tasks left to claim in this space, the global_finger
4190 // must exceed space->top() and a fortiori space->end(). However,
4191 // that would not quite be correct because the bumping of
4192 // global_finger occurs strictly after the claiming of a task,
4193 // so by the time we reach here the global finger may not yet
4194 // have been bumped up by the thread that claimed the last
4195 // task.
4196 pst->all_tasks_completed();
4197 }
4198
4199 class Par_ConcMarkingClosure: public CMSOopClosure {
4200 private:
4201 CMSCollector* _collector;
4202 CMSConcMarkingTask* _task;
4203 MemRegion _span;
4204 CMSBitMap* _bit_map;
4205 CMSMarkStack* _overflow_stack;
4206 OopTaskQueue* _work_queue;
4207 protected:
4208 DO_OOP_WORK_DEFN
4209 public:
4210 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
4211 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
4212 CMSOopClosure(collector->ref_processor()),
4213 _collector(collector),
4214 _task(task),
4215 _span(collector->_span),
4216 _work_queue(work_queue),
4217 _bit_map(bit_map),
4218 _overflow_stack(overflow_stack)
4219 { }
4220 virtual void do_oop(oop* p);
4221 virtual void do_oop(narrowOop* p);
4222
4223 void trim_queue(size_t max);
4224 void handle_stack_overflow(HeapWord* lost);
4225 void do_yield_check() {
4226 if (_task->should_yield()) {
4227 _task->yield();
4228 }
4229 }
4230 };
4231
4232 // Grey object scanning during work stealing phase --
4233 // the salient assumption here is that any references
4234 // that are in these stolen objects being scanned must
4235 // already have been initialized (else they would not have
4236 // been published), so we do not need to check for
4237 // uninitialized objects before pushing here.
4238 void Par_ConcMarkingClosure::do_oop(oop obj) {
4239 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
4240 HeapWord* addr = (HeapWord*)obj;
4241 // Check if oop points into the CMS generation
4242 // and is not marked
4243 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
4244 // a white object ...
4245 // If we manage to "claim" the object, by being the
4246 // first thread to mark it, then we push it on our
4247 // marking stack
4248 if (_bit_map->par_mark(addr)) { // ... now grey
4249 // push on work queue (grey set)
4250 bool simulate_overflow = false;
4251 NOT_PRODUCT(
4252 if (CMSMarkStackOverflowALot &&
4253 _collector->simulate_overflow()) {
4254 // simulate a stack overflow
4255 simulate_overflow = true;
4256 }
4257 )
4258 if (simulate_overflow ||
4259 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
4260 // stack overflow
4261 if (PrintCMSStatistics != 0) {
4262 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
4263 SIZE_FORMAT, _overflow_stack->capacity());
4264 }
4265 // We cannot assert that the overflow stack is full because
4266 // it may have been emptied since.
4267 assert(simulate_overflow ||
4268 _work_queue->size() == _work_queue->max_elems(),
4269 "Else push should have succeeded");
4270 handle_stack_overflow(addr);
4271 }
4272 } // Else, some other thread got there first
4273 do_yield_check();
4274 }
4275 }
4276
4277 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4278 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); }
4279
4280 void Par_ConcMarkingClosure::trim_queue(size_t max) {
4281 while (_work_queue->size() > max) {
4282 oop new_oop;
4283 if (_work_queue->pop_local(new_oop)) {
4284 assert(new_oop->is_oop(), "Should be an oop");
4285 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
4286 assert(_span.contains((HeapWord*)new_oop), "Not in span");
4287 new_oop->oop_iterate(this); // do_oop() above
4288 do_yield_check();
4289 }
4290 }
4291 }
4292
4293 // Upon stack overflow, we discard (part of) the stack,
4294 // remembering the least address amongst those discarded
4295 // in CMSCollector's _restart_address.
4296 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
4297 // We need to do this under a mutex to prevent other
4298 // workers from interfering with the work done below.
4299 MutexLockerEx ml(_overflow_stack->par_lock(),
4300 Mutex::_no_safepoint_check_flag);
4301 // Remember the least grey address discarded
4302 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
4303 _collector->lower_restart_addr(ra);
4304 _overflow_stack->reset(); // discard stack contents
4305 _overflow_stack->expand(); // expand the stack if possible
4306 }
4307
4308
4309 void CMSConcMarkingTask::do_work_steal(int i) {
4310 OopTaskQueue* work_q = work_queue(i);
4311 oop obj_to_scan;
4312 CMSBitMap* bm = &(_collector->_markBitMap);
4313 CMSMarkStack* ovflw = &(_collector->_markStack);
4314 int* seed = _collector->hash_seed(i);
4315 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
4316 while (true) {
4317 cl.trim_queue(0);
4318 assert(work_q->size() == 0, "Should have been emptied above");
4319 if (get_work_from_overflow_stack(ovflw, work_q)) {
4320 // Can't assert below because the work obtained from the
4321 // overflow stack may already have been stolen from us.
4322 // assert(work_q->size() > 0, "Work from overflow stack");
4323 continue;
4324 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4325 assert(obj_to_scan->is_oop(), "Should be an oop");
4326 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
4327 obj_to_scan->oop_iterate(&cl);
4328 } else if (terminator()->offer_termination(&_term_term)) {
4329 assert(work_q->size() == 0, "Impossible!");
4330 break;
4331 } else if (yielding() || should_yield()) {
4332 yield();
4333 }
4334 }
4335 }
4336
4337 // This is run by the CMS (coordinator) thread.
4338 void CMSConcMarkingTask::coordinator_yield() {
4339 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4340 "CMS thread should hold CMS token");
4341 // First give up the locks, then yield, then re-lock
4342 // We should probably use a constructor/destructor idiom to
4343 // do this unlock/lock or modify the MutexUnlocker class to
4344 // serve our purpose. XXX
4345 assert_lock_strong(_bit_map_lock);
4346 _bit_map_lock->unlock();
4347 ConcurrentMarkSweepThread::desynchronize(true);
4348 ConcurrentMarkSweepThread::acknowledge_yield_request();
4349 _collector->stopTimer();
4350 if (PrintCMSStatistics != 0) {
4351 _collector->incrementYields();
4352 }
4353 _collector->icms_wait();
4354
4355 // It is possible for whichever thread initiated the yield request
4356 // not to get a chance to wake up and take the bitmap lock between
4357 // this thread releasing it and reacquiring it. So, while the
4358 // should_yield() flag is on, let's sleep for a bit to give the
4359 // other thread a chance to wake up. The limit imposed on the number
4360 // of iterations is defensive, to avoid any unforseen circumstances
4361 // putting us into an infinite loop. Since it's always been this
4362 // (coordinator_yield()) method that was observed to cause the
4363 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
4364 // which is by default non-zero. For the other seven methods that
4365 // also perform the yield operation, as are using a different
4366 // parameter (CMSYieldSleepCount) which is by default zero. This way we
4367 // can enable the sleeping for those methods too, if necessary.
4368 // See 6442774.
4369 //
4370 // We really need to reconsider the synchronization between the GC
4371 // thread and the yield-requesting threads in the future and we
4372 // should really use wait/notify, which is the recommended
4373 // way of doing this type of interaction. Additionally, we should
4374 // consolidate the eight methods that do the yield operation and they
4375 // are almost identical into one for better maintenability and
4376 // readability. See 6445193.
4377 //
4378 // Tony 2006.06.29
4379 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
4380 ConcurrentMarkSweepThread::should_yield() &&
4381 !CMSCollector::foregroundGCIsActive(); ++i) {
4382 os::sleep(Thread::current(), 1, false);
4383 ConcurrentMarkSweepThread::acknowledge_yield_request();
4384 }
4385
4386 ConcurrentMarkSweepThread::synchronize(true);
4387 _bit_map_lock->lock_without_safepoint_check();
4388 _collector->startTimer();
4389 }
4390
4391 bool CMSCollector::do_marking_mt(bool asynch) {
4392 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
4393 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers(
4394 conc_workers()->total_workers(),
4395 conc_workers()->active_workers(),
4396 Threads::number_of_non_daemon_threads());
4397 conc_workers()->set_active_workers(num_workers);
4398
4399 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4400
4401 CMSConcMarkingTask tsk(this,
4402 cms_space,
4403 asynch,
4404 conc_workers(),
4405 task_queues());
4406
4407 // Since the actual number of workers we get may be different
4408 // from the number we requested above, do we need to do anything different
4409 // below? In particular, may be we need to subclass the SequantialSubTasksDone
4410 // class?? XXX
4411 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
4412
4413 // Refs discovery is already non-atomic.
4414 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
4415 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
4416 conc_workers()->start_task(&tsk);
4417 while (tsk.yielded()) {
4418 tsk.coordinator_yield();
4419 conc_workers()->continue_task(&tsk);
4420 }
4421 // If the task was aborted, _restart_addr will be non-NULL
4422 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
4423 while (_restart_addr != NULL) {
4424 // XXX For now we do not make use of ABORTED state and have not
4425 // yet implemented the right abort semantics (even in the original
4426 // single-threaded CMS case). That needs some more investigation
4427 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
4428 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
4429 // If _restart_addr is non-NULL, a marking stack overflow
4430 // occurred; we need to do a fresh marking iteration from the
4431 // indicated restart address.
4432 if (_foregroundGCIsActive && asynch) {
4433 // We may be running into repeated stack overflows, having
4434 // reached the limit of the stack size, while making very
4435 // slow forward progress. It may be best to bail out and
4436 // let the foreground collector do its job.
4437 // Clear _restart_addr, so that foreground GC
4438 // works from scratch. This avoids the headache of
4439 // a "rescan" which would otherwise be needed because
4440 // of the dirty mod union table & card table.
4441 _restart_addr = NULL;
4442 return false;
4443 }
4444 // Adjust the task to restart from _restart_addr
4445 tsk.reset(_restart_addr);
4446 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
4447 _restart_addr);
4448 _restart_addr = NULL;
4449 // Get the workers going again
4450 conc_workers()->start_task(&tsk);
4451 while (tsk.yielded()) {
4452 tsk.coordinator_yield();
4453 conc_workers()->continue_task(&tsk);
4454 }
4455 }
4456 assert(tsk.completed(), "Inconsistency");
4457 assert(tsk.result() == true, "Inconsistency");
4458 return true;
4459 }
4460
4461 bool CMSCollector::do_marking_st(bool asynch) {
4462 ResourceMark rm;
4463 HandleMark hm;
4464
4465 // Temporarily make refs discovery single threaded (non-MT)
4466 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
4467 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
4468 &_markStack, CMSYield && asynch);
4469 // the last argument to iterate indicates whether the iteration
4470 // should be incremental with periodic yields.
4471 _markBitMap.iterate(&markFromRootsClosure);
4472 // If _restart_addr is non-NULL, a marking stack overflow
4473 // occurred; we need to do a fresh iteration from the
4474 // indicated restart address.
4475 while (_restart_addr != NULL) {
4476 if (_foregroundGCIsActive && asynch) {
4477 // We may be running into repeated stack overflows, having
4478 // reached the limit of the stack size, while making very
4479 // slow forward progress. It may be best to bail out and
4480 // let the foreground collector do its job.
4481 // Clear _restart_addr, so that foreground GC
4482 // works from scratch. This avoids the headache of
4483 // a "rescan" which would otherwise be needed because
4484 // of the dirty mod union table & card table.
4485 _restart_addr = NULL;
4486 return false; // indicating failure to complete marking
4487 }
4488 // Deal with stack overflow:
4489 // we restart marking from _restart_addr
4490 HeapWord* ra = _restart_addr;
4491 markFromRootsClosure.reset(ra);
4492 _restart_addr = NULL;
4493 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
4494 }
4495 return true;
4496 }
4497
4498 void CMSCollector::preclean() {
4499 check_correct_thread_executing();
4500 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
4501 verify_work_stacks_empty();
4502 verify_overflow_empty();
4503 _abort_preclean = false;
4504 if (CMSPrecleaningEnabled) {
4505 if (!CMSEdenChunksRecordAlways) {
4506 _eden_chunk_index = 0;
4507 }
4508 size_t used = get_eden_used();
4509 size_t capacity = get_eden_capacity();
4510 // Don't start sampling unless we will get sufficiently
4511 // many samples.
4512 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100)
4513 * CMSScheduleRemarkEdenPenetration)) {
4514 _start_sampling = true;
4515 } else {
4516 _start_sampling = false;
4517 }
4518 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4519 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails);
4520 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
4521 }
4522 CMSTokenSync x(true); // is cms thread
4523 if (CMSPrecleaningEnabled) {
4524 sample_eden();
4525 _collectorState = AbortablePreclean;
4526 } else {
4527 _collectorState = FinalMarking;
4528 }
4529 verify_work_stacks_empty();
4530 verify_overflow_empty();
4531 }
4532
4533 // Try and schedule the remark such that young gen
4534 // occupancy is CMSScheduleRemarkEdenPenetration %.
4535 void CMSCollector::abortable_preclean() {
4536 check_correct_thread_executing();
4537 assert(CMSPrecleaningEnabled, "Inconsistent control state");
4538 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
4539
4540 // If Eden's current occupancy is below this threshold,
4541 // immediately schedule the remark; else preclean
4542 // past the next scavenge in an effort to
4543 // schedule the pause as described avove. By choosing
4544 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
4545 // we will never do an actual abortable preclean cycle.
4546 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
4547 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
4548 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails);
4549 // We need more smarts in the abortable preclean
4550 // loop below to deal with cases where allocation
4551 // in young gen is very very slow, and our precleaning
4552 // is running a losing race against a horde of
4553 // mutators intent on flooding us with CMS updates
4554 // (dirty cards).
4555 // One, admittedly dumb, strategy is to give up
4556 // after a certain number of abortable precleaning loops
4557 // or after a certain maximum time. We want to make
4558 // this smarter in the next iteration.
4559 // XXX FIX ME!!! YSR
4560 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
4561 while (!(should_abort_preclean() ||
4562 ConcurrentMarkSweepThread::should_terminate())) {
4563 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
4564 cumworkdone += workdone;
4565 loops++;
4566 // Voluntarily terminate abortable preclean phase if we have
4567 // been at it for too long.
4568 if ((CMSMaxAbortablePrecleanLoops != 0) &&
4569 loops >= CMSMaxAbortablePrecleanLoops) {
4570 if (PrintGCDetails) {
4571 gclog_or_tty->print(" CMS: abort preclean due to loops ");
4572 }
4573 break;
4574 }
4575 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
4576 if (PrintGCDetails) {
4577 gclog_or_tty->print(" CMS: abort preclean due to time ");
4578 }
4579 break;
4580 }
4581 // If we are doing little work each iteration, we should
4582 // take a short break.
4583 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
4584 // Sleep for some time, waiting for work to accumulate
4585 stopTimer();
4586 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
4587 startTimer();
4588 waited++;
4589 }
4590 }
4591 if (PrintCMSStatistics > 0) {
4592 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ",
4593 loops, waited, cumworkdone);
4594 }
4595 }
4596 CMSTokenSync x(true); // is cms thread
4597 if (_collectorState != Idling) {
4598 assert(_collectorState == AbortablePreclean,
4599 "Spontaneous state transition?");
4600 _collectorState = FinalMarking;
4601 } // Else, a foreground collection completed this CMS cycle.
4602 return;
4603 }
4604
4605 // Respond to an Eden sampling opportunity
4606 void CMSCollector::sample_eden() {
4607 // Make sure a young gc cannot sneak in between our
4608 // reading and recording of a sample.
4609 assert(Thread::current()->is_ConcurrentGC_thread(),
4610 "Only the cms thread may collect Eden samples");
4611 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
4612 "Should collect samples while holding CMS token");
4613 if (!_start_sampling) {
4614 return;
4615 }
4616 // When CMSEdenChunksRecordAlways is true, the eden chunk array
4617 // is populated by the young generation.
4618 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
4619 if (_eden_chunk_index < _eden_chunk_capacity) {
4620 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
4621 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4622 "Unexpected state of Eden");
4623 // We'd like to check that what we just sampled is an oop-start address;
4624 // however, we cannot do that here since the object may not yet have been
4625 // initialized. So we'll instead do the check when we _use_ this sample
4626 // later.
4627 if (_eden_chunk_index == 0 ||
4628 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4629 _eden_chunk_array[_eden_chunk_index-1])
4630 >= CMSSamplingGrain)) {
4631 _eden_chunk_index++; // commit sample
4632 }
4633 }
4634 }
4635 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
4636 size_t used = get_eden_used();
4637 size_t capacity = get_eden_capacity();
4638 assert(used <= capacity, "Unexpected state of Eden");
4639 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
4640 _abort_preclean = true;
4641 }
4642 }
4643 }
4644
4645
4646 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
4647 assert(_collectorState == Precleaning ||
4648 _collectorState == AbortablePreclean, "incorrect state");
4649 ResourceMark rm;
4650 HandleMark hm;
4651
4652 // Precleaning is currently not MT but the reference processor
4653 // may be set for MT. Disable it temporarily here.
4654 ReferenceProcessor* rp = ref_processor();
4655 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
4656
4657 // Do one pass of scrubbing the discovered reference lists
4658 // to remove any reference objects with strongly-reachable
4659 // referents.
4660 if (clean_refs) {
4661 CMSPrecleanRefsYieldClosure yield_cl(this);
4662 assert(rp->span().equals(_span), "Spans should be equal");
4663 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
4664 &_markStack, true /* preclean */);
4665 CMSDrainMarkingStackClosure complete_trace(this,
4666 _span, &_markBitMap, &_markStack,
4667 &keep_alive, true /* preclean */);
4668
4669 // We don't want this step to interfere with a young
4670 // collection because we don't want to take CPU
4671 // or memory bandwidth away from the young GC threads
4672 // (which may be as many as there are CPUs).
4673 // Note that we don't need to protect ourselves from
4674 // interference with mutators because they can't
4675 // manipulate the discovered reference lists nor affect
4676 // the computed reachability of the referents, the
4677 // only properties manipulated by the precleaning
4678 // of these reference lists.
4679 stopTimer();
4680 CMSTokenSyncWithLocks x(true /* is cms thread */,
4681 bitMapLock());
4682 startTimer();
4683 sample_eden();
4684
4685 // The following will yield to allow foreground
4686 // collection to proceed promptly. XXX YSR:
4687 // The code in this method may need further
4688 // tweaking for better performance and some restructuring
4689 // for cleaner interfaces.
4690 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
4691 rp->preclean_discovered_references(
4692 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
4693 gc_timer);
4694 }
4695
4696 if (clean_survivor) { // preclean the active survivor space(s)
4697 assert(_young_gen->kind() == Generation::DefNew ||
4698 _young_gen->kind() == Generation::ParNew ||
4699 _young_gen->kind() == Generation::ASParNew,
4700 "incorrect type for cast");
4701 DefNewGeneration* dng = (DefNewGeneration*)_young_gen;
4702 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
4703 &_markBitMap, &_modUnionTable,
4704 &_markStack, true /* precleaning phase */);
4705 stopTimer();
4706 CMSTokenSyncWithLocks ts(true /* is cms thread */,
4707 bitMapLock());
4708 startTimer();
4709 unsigned int before_count =
4710 GenCollectedHeap::heap()->total_collections();
4711 SurvivorSpacePrecleanClosure
4712 sss_cl(this, _span, &_markBitMap, &_markStack,
4713 &pam_cl, before_count, CMSYield);
4714 dng->from()->object_iterate_careful(&sss_cl);
4715 dng->to()->object_iterate_careful(&sss_cl);
4716 }
4717 MarkRefsIntoAndScanClosure
4718 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
4719 &_markStack, this, CMSYield,
4720 true /* precleaning phase */);
4721 // CAUTION: The following closure has persistent state that may need to
4722 // be reset upon a decrease in the sequence of addresses it
4723 // processes.
4724 ScanMarkedObjectsAgainCarefullyClosure
4725 smoac_cl(this, _span,
4726 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
4727
4728 // Preclean dirty cards in ModUnionTable and CardTable using
4729 // appropriate convergence criterion;
4730 // repeat CMSPrecleanIter times unless we find that
4731 // we are losing.
4732 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
4733 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
4734 "Bad convergence multiplier");
4735 assert(CMSPrecleanThreshold >= 100,
4736 "Unreasonably low CMSPrecleanThreshold");
4737
4738 size_t numIter, cumNumCards, lastNumCards, curNumCards;
4739 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
4740 numIter < CMSPrecleanIter;
4741 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
4742 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
4743 if (Verbose && PrintGCDetails) {
4744 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards);
4745 }
4746 // Either there are very few dirty cards, so re-mark
4747 // pause will be small anyway, or our pre-cleaning isn't
4748 // that much faster than the rate at which cards are being
4749 // dirtied, so we might as well stop and re-mark since
4750 // precleaning won't improve our re-mark time by much.
4751 if (curNumCards <= CMSPrecleanThreshold ||
4752 (numIter > 0 &&
4753 (curNumCards * CMSPrecleanDenominator >
4754 lastNumCards * CMSPrecleanNumerator))) {
4755 numIter++;
4756 cumNumCards += curNumCards;
4757 break;
4758 }
4759 }
4760
4761 preclean_klasses(&mrias_cl, _cmsGen->freelistLock());
4762
4763 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
4764 cumNumCards += curNumCards;
4765 if (PrintGCDetails && PrintCMSStatistics != 0) {
4766 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)",
4767 curNumCards, cumNumCards, numIter);
4768 }
4769 return cumNumCards; // as a measure of useful work done
4770 }
4771
4772 // PRECLEANING NOTES:
4773 // Precleaning involves:
4774 // . reading the bits of the modUnionTable and clearing the set bits.
4775 // . For the cards corresponding to the set bits, we scan the
4776 // objects on those cards. This means we need the free_list_lock
4777 // so that we can safely iterate over the CMS space when scanning
4778 // for oops.
4779 // . When we scan the objects, we'll be both reading and setting
4780 // marks in the marking bit map, so we'll need the marking bit map.
4781 // . For protecting _collector_state transitions, we take the CGC_lock.
4782 // Note that any races in the reading of of card table entries by the
4783 // CMS thread on the one hand and the clearing of those entries by the
4784 // VM thread or the setting of those entries by the mutator threads on the
4785 // other are quite benign. However, for efficiency it makes sense to keep
4786 // the VM thread from racing with the CMS thread while the latter is
4787 // dirty card info to the modUnionTable. We therefore also use the
4788 // CGC_lock to protect the reading of the card table and the mod union
4789 // table by the CM thread.
4790 // . We run concurrently with mutator updates, so scanning
4791 // needs to be done carefully -- we should not try to scan
4792 // potentially uninitialized objects.
4793 //
4794 // Locking strategy: While holding the CGC_lock, we scan over and
4795 // reset a maximal dirty range of the mod union / card tables, then lock
4796 // the free_list_lock and bitmap lock to do a full marking, then
4797 // release these locks; and repeat the cycle. This allows for a
4798 // certain amount of fairness in the sharing of these locks between
4799 // the CMS collector on the one hand, and the VM thread and the
4800 // mutators on the other.
4801
4802 // NOTE: preclean_mod_union_table() and preclean_card_table()
4803 // further below are largely identical; if you need to modify
4804 // one of these methods, please check the other method too.
4805
4806 size_t CMSCollector::preclean_mod_union_table(
4807 ConcurrentMarkSweepGeneration* gen,
4808 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4809 verify_work_stacks_empty();
4810 verify_overflow_empty();
4811
4812 // strategy: starting with the first card, accumulate contiguous
4813 // ranges of dirty cards; clear these cards, then scan the region
4814 // covered by these cards.
4815
4816 // Since all of the MUT is committed ahead, we can just use
4817 // that, in case the generations expand while we are precleaning.
4818 // It might also be fine to just use the committed part of the
4819 // generation, but we might potentially miss cards when the
4820 // generation is rapidly expanding while we are in the midst
4821 // of precleaning.
4822 HeapWord* startAddr = gen->reserved().start();
4823 HeapWord* endAddr = gen->reserved().end();
4824
4825 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4826
4827 size_t numDirtyCards, cumNumDirtyCards;
4828 HeapWord *nextAddr, *lastAddr;
4829 for (cumNumDirtyCards = numDirtyCards = 0,
4830 nextAddr = lastAddr = startAddr;
4831 nextAddr < endAddr;
4832 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4833
4834 ResourceMark rm;
4835 HandleMark hm;
4836
4837 MemRegion dirtyRegion;
4838 {
4839 stopTimer();
4840 // Potential yield point
4841 CMSTokenSync ts(true);
4842 startTimer();
4843 sample_eden();
4844 // Get dirty region starting at nextOffset (inclusive),
4845 // simultaneously clearing it.
4846 dirtyRegion =
4847 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
4848 assert(dirtyRegion.start() >= nextAddr,
4849 "returned region inconsistent?");
4850 }
4851 // Remember where the next search should begin.
4852 // The returned region (if non-empty) is a right open interval,
4853 // so lastOffset is obtained from the right end of that
4854 // interval.
4855 lastAddr = dirtyRegion.end();
4856 // Should do something more transparent and less hacky XXX
4857 numDirtyCards =
4858 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
4859
4860 // We'll scan the cards in the dirty region (with periodic
4861 // yields for foreground GC as needed).
4862 if (!dirtyRegion.is_empty()) {
4863 assert(numDirtyCards > 0, "consistency check");
4864 HeapWord* stop_point = NULL;
4865 stopTimer();
4866 // Potential yield point
4867 CMSTokenSyncWithLocks ts(true, gen->freelistLock(),
4868 bitMapLock());
4869 startTimer();
4870 {
4871 verify_work_stacks_empty();
4872 verify_overflow_empty();
4873 sample_eden();
4874 stop_point =
4875 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4876 }
4877 if (stop_point != NULL) {
4878 // The careful iteration stopped early either because it found an
4879 // uninitialized object, or because we were in the midst of an
4880 // "abortable preclean", which should now be aborted. Redirty
4881 // the bits corresponding to the partially-scanned or unscanned
4882 // cards. We'll either restart at the next block boundary or
4883 // abort the preclean.
4884 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4885 "Should only be AbortablePreclean.");
4886 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
4887 if (should_abort_preclean()) {
4888 break; // out of preclean loop
4889 } else {
4890 // Compute the next address at which preclean should pick up;
4891 // might need bitMapLock in order to read P-bits.
4892 lastAddr = next_card_start_after_block(stop_point);
4893 }
4894 }
4895 } else {
4896 assert(lastAddr == endAddr, "consistency check");
4897 assert(numDirtyCards == 0, "consistency check");
4898 break;
4899 }
4900 }
4901 verify_work_stacks_empty();
4902 verify_overflow_empty();
4903 return cumNumDirtyCards;
4904 }
4905
4906 // NOTE: preclean_mod_union_table() above and preclean_card_table()
4907 // below are largely identical; if you need to modify
4908 // one of these methods, please check the other method too.
4909
4910 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen,
4911 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4912 // strategy: it's similar to precleamModUnionTable above, in that
4913 // we accumulate contiguous ranges of dirty cards, mark these cards
4914 // precleaned, then scan the region covered by these cards.
4915 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high());
4916 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low());
4917
4918 cl->setFreelistLock(gen->freelistLock()); // needed for yielding
4919
4920 size_t numDirtyCards, cumNumDirtyCards;
4921 HeapWord *lastAddr, *nextAddr;
4922
4923 for (cumNumDirtyCards = numDirtyCards = 0,
4924 nextAddr = lastAddr = startAddr;
4925 nextAddr < endAddr;
4926 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4927
4928 ResourceMark rm;
4929 HandleMark hm;
4930
4931 MemRegion dirtyRegion;
4932 {
4933 // See comments in "Precleaning notes" above on why we
4934 // do this locking. XXX Could the locking overheads be
4935 // too high when dirty cards are sparse? [I don't think so.]
4936 stopTimer();
4937 CMSTokenSync x(true); // is cms thread
4938 startTimer();
4939 sample_eden();
4940 // Get and clear dirty region from card table
4941 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset(
4942 MemRegion(nextAddr, endAddr),
4943 true,
4944 CardTableModRefBS::precleaned_card_val());
4945
4946 assert(dirtyRegion.start() >= nextAddr,
4947 "returned region inconsistent?");
4948 }
4949 lastAddr = dirtyRegion.end();
4950 numDirtyCards =
4951 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words;
4952
4953 if (!dirtyRegion.is_empty()) {
4954 stopTimer();
4955 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock());
4956 startTimer();
4957 sample_eden();
4958 verify_work_stacks_empty();
4959 verify_overflow_empty();
4960 HeapWord* stop_point =
4961 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4962 if (stop_point != NULL) {
4963 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4964 "Should only be AbortablePreclean.");
4965 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4966 if (should_abort_preclean()) {
4967 break; // out of preclean loop
4968 } else {
4969 // Compute the next address at which preclean should pick up.
4970 lastAddr = next_card_start_after_block(stop_point);
4971 }
4972 }
4973 } else {
4974 break;
4975 }
4976 }
4977 verify_work_stacks_empty();
4978 verify_overflow_empty();
4979 return cumNumDirtyCards;
4980 }
4981
4982 class PrecleanKlassClosure : public KlassClosure {
4983 CMKlassClosure _cm_klass_closure;
4984 public:
4985 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
4986 void do_klass(Klass* k) {
4987 if (k->has_accumulated_modified_oops()) {
4988 k->clear_accumulated_modified_oops();
4989
4990 _cm_klass_closure.do_klass(k);
4991 }
4992 }
4993 };
4994
4995 // The freelist lock is needed to prevent asserts, is it really needed?
4996 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4997
4998 cl->set_freelistLock(freelistLock);
4999
5000 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
5001
5002 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
5003 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
5004 PrecleanKlassClosure preclean_klass_closure(cl);
5005 ClassLoaderDataGraph::classes_do(&preclean_klass_closure);
5006
5007 verify_work_stacks_empty();
5008 verify_overflow_empty();
5009 }
5010
5011 void CMSCollector::checkpointRootsFinal(bool asynch,
5012 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5013 assert(_collectorState == FinalMarking, "incorrect state transition?");
5014 check_correct_thread_executing();
5015 // world is stopped at this checkpoint
5016 assert(SafepointSynchronize::is_at_safepoint(),
5017 "world should be stopped");
5018 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
5019
5020 verify_work_stacks_empty();
5021 verify_overflow_empty();
5022
5023 SpecializationStats::clear();
5024 if (PrintGCDetails) {
5025 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]",
5026 _young_gen->used() / K,
5027 _young_gen->capacity() / K);
5028 }
5029 if (asynch) {
5030 if (CMSScavengeBeforeRemark) {
5031 GenCollectedHeap* gch = GenCollectedHeap::heap();
5032 // Temporarily set flag to false, GCH->do_collection will
5033 // expect it to be false and set to true
5034 FlagSetting fl(gch->_is_gc_active, false);
5035 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark",
5036 PrintGCDetails && Verbose, true, _gc_timer_cm);)
5037 int level = _cmsGen->level() - 1;
5038 if (level >= 0) {
5039 gch->do_collection(true, // full (i.e. force, see below)
5040 false, // !clear_all_soft_refs
5041 0, // size
5042 false, // is_tlab
5043 level // max_level
5044 );
5045 }
5046 }
5047 FreelistLocker x(this);
5048 MutexLockerEx y(bitMapLock(),
5049 Mutex::_no_safepoint_check_flag);
5050 assert(!init_mark_was_synchronous, "but that's impossible!");
5051 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false);
5052 } else {
5053 // already have all the locks
5054 checkpointRootsFinalWork(asynch, clear_all_soft_refs,
5055 init_mark_was_synchronous);
5056 }
5057 verify_work_stacks_empty();
5058 verify_overflow_empty();
5059 SpecializationStats::print();
5060 }
5061
5062 void CMSCollector::checkpointRootsFinalWork(bool asynch,
5063 bool clear_all_soft_refs, bool init_mark_was_synchronous) {
5064
5065 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);)
5066
5067 assert(haveFreelistLocks(), "must have free list locks");
5068 assert_lock_strong(bitMapLock());
5069
5070 if (UseAdaptiveSizePolicy) {
5071 size_policy()->checkpoint_roots_final_begin();
5072 }
5073
5074 ResourceMark rm;
5075 HandleMark hm;
5076
5077 GenCollectedHeap* gch = GenCollectedHeap::heap();
5078
5079 if (should_unload_classes()) {
5080 CodeCache::gc_prologue();
5081 }
5082 assert(haveFreelistLocks(), "must have free list locks");
5083 assert_lock_strong(bitMapLock());
5084
5085 if (!init_mark_was_synchronous) {
5086 // We might assume that we need not fill TLAB's when
5087 // CMSScavengeBeforeRemark is set, because we may have just done
5088 // a scavenge which would have filled all TLAB's -- and besides
5089 // Eden would be empty. This however may not always be the case --
5090 // for instance although we asked for a scavenge, it may not have
5091 // happened because of a JNI critical section. We probably need
5092 // a policy for deciding whether we can in that case wait until
5093 // the critical section releases and then do the remark following
5094 // the scavenge, and skip it here. In the absence of that policy,
5095 // or of an indication of whether the scavenge did indeed occur,
5096 // we cannot rely on TLAB's having been filled and must do
5097 // so here just in case a scavenge did not happen.
5098 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them
5099 // Update the saved marks which may affect the root scans.
5100 gch->save_marks();
5101
5102 if (CMSPrintEdenSurvivorChunks) {
5103 print_eden_and_survivor_chunk_arrays();
5104 }
5105
5106 {
5107 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;)
5108
5109 // Note on the role of the mod union table:
5110 // Since the marker in "markFromRoots" marks concurrently with
5111 // mutators, it is possible for some reachable objects not to have been
5112 // scanned. For instance, an only reference to an object A was
5113 // placed in object B after the marker scanned B. Unless B is rescanned,
5114 // A would be collected. Such updates to references in marked objects
5115 // are detected via the mod union table which is the set of all cards
5116 // dirtied since the first checkpoint in this GC cycle and prior to
5117 // the most recent young generation GC, minus those cleaned up by the
5118 // concurrent precleaning.
5119 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
5120 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm);
5121 do_remark_parallel();
5122 } else {
5123 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false,
5124 _gc_timer_cm);
5125 do_remark_non_parallel();
5126 }
5127 }
5128 } else {
5129 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode");
5130 // The initial mark was stop-world, so there's no rescanning to
5131 // do; go straight on to the next step below.
5132 }
5133 verify_work_stacks_empty();
5134 verify_overflow_empty();
5135
5136 {
5137 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);)
5138 refProcessingWork(asynch, clear_all_soft_refs);
5139 }
5140 verify_work_stacks_empty();
5141 verify_overflow_empty();
5142
5143 if (should_unload_classes()) {
5144 CodeCache::gc_epilogue();
5145 }
5146 JvmtiExport::gc_epilogue();
5147
5148 // If we encountered any (marking stack / work queue) overflow
5149 // events during the current CMS cycle, take appropriate
5150 // remedial measures, where possible, so as to try and avoid
5151 // recurrence of that condition.
5152 assert(_markStack.isEmpty(), "No grey objects");
5153 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
5154 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
5155 if (ser_ovflw > 0) {
5156 if (PrintCMSStatistics != 0) {
5157 gclog_or_tty->print_cr("Marking stack overflow (benign) "
5158 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT
5159 ", kac_preclean="SIZE_FORMAT")",
5160 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw,
5161 _ser_kac_ovflw, _ser_kac_preclean_ovflw);
5162 }
5163 _markStack.expand();
5164 _ser_pmc_remark_ovflw = 0;
5165 _ser_pmc_preclean_ovflw = 0;
5166 _ser_kac_preclean_ovflw = 0;
5167 _ser_kac_ovflw = 0;
5168 }
5169 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
5170 if (PrintCMSStatistics != 0) {
5171 gclog_or_tty->print_cr("Work queue overflow (benign) "
5172 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")",
5173 _par_pmc_remark_ovflw, _par_kac_ovflw);
5174 }
5175 _par_pmc_remark_ovflw = 0;
5176 _par_kac_ovflw = 0;
5177 }
5178 if (PrintCMSStatistics != 0) {
5179 if (_markStack._hit_limit > 0) {
5180 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")",
5181 _markStack._hit_limit);
5182 }
5183 if (_markStack._failed_double > 0) {
5184 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT"),"
5185 " current capacity "SIZE_FORMAT,
5186 _markStack._failed_double,
5187 _markStack.capacity());
5188 }
5189 }
5190 _markStack._hit_limit = 0;
5191 _markStack._failed_double = 0;
5192
5193 if ((VerifyAfterGC || VerifyDuringGC) &&
5194 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) {
5195 verify_after_remark();
5196 }
5197
5198 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
5199
5200 // Change under the freelistLocks.
5201 _collectorState = Sweeping;
5202 // Call isAllClear() under bitMapLock
5203 assert(_modUnionTable.isAllClear(),
5204 "Should be clear by end of the final marking");
5205 assert(_ct->klass_rem_set()->mod_union_is_clear(),
5206 "Should be clear by end of the final marking");
5207 if (UseAdaptiveSizePolicy) {
5208 size_policy()->checkpoint_roots_final_end(gch->gc_cause());
5209 }
5210 }
5211
5212 void CMSParInitialMarkTask::work(uint worker_id) {
5213 elapsedTimer _timer;
5214 ResourceMark rm;
5215 HandleMark hm;
5216
5217 // ---------- scan from roots --------------
5218 _timer.start();
5219 GenCollectedHeap* gch = GenCollectedHeap::heap();
5220 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
5221 CMKlassClosure klass_closure(&par_mri_cl);
5222
5223 // ---------- young gen roots --------------
5224 {
5225 work_on_young_gen_roots(worker_id, &par_mri_cl);
5226 _timer.stop();
5227 if (PrintCMSStatistics != 0) {
5228 gclog_or_tty->print_cr(
5229 "Finished young gen initial mark scan work in %dth thread: %3.3f sec",
5230 worker_id, _timer.seconds());
5231 }
5232 }
5233
5234 // ---------- remaining roots --------------
5235 _timer.reset();
5236 _timer.start();
5237 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5238 false, // yg was scanned above
5239 false, // this is parallel code
5240 false, // not scavenging
5241 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5242 &par_mri_cl,
5243 true, // walk all of code cache if (so & SO_CodeCache)
5244 NULL,
5245 &klass_closure);
5246 assert(_collector->should_unload_classes()
5247 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache),
5248 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5249 _timer.stop();
5250 if (PrintCMSStatistics != 0) {
5251 gclog_or_tty->print_cr(
5252 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec",
5253 worker_id, _timer.seconds());
5254 }
5255 }
5256
5257 // Parallel remark task
5258 class CMSParRemarkTask: public CMSParMarkTask {
5259 CompactibleFreeListSpace* _cms_space;
5260
5261 // The per-thread work queues, available here for stealing.
5262 OopTaskQueueSet* _task_queues;
5263 ParallelTaskTerminator _term;
5264
5265 public:
5266 // A value of 0 passed to n_workers will cause the number of
5267 // workers to be taken from the active workers in the work gang.
5268 CMSParRemarkTask(CMSCollector* collector,
5269 CompactibleFreeListSpace* cms_space,
5270 int n_workers, FlexibleWorkGang* workers,
5271 OopTaskQueueSet* task_queues):
5272 CMSParMarkTask("Rescan roots and grey objects in parallel",
5273 collector, n_workers),
5274 _cms_space(cms_space),
5275 _task_queues(task_queues),
5276 _term(n_workers, task_queues) { }
5277
5278 OopTaskQueueSet* task_queues() { return _task_queues; }
5279
5280 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5281
5282 ParallelTaskTerminator* terminator() { return &_term; }
5283 int n_workers() { return _n_workers; }
5284
5285 void work(uint worker_id);
5286
5287 private:
5288 // ... of dirty cards in old space
5289 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
5290 Par_MarkRefsIntoAndScanClosure* cl);
5291
5292 // ... work stealing for the above
5293 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
5294 };
5295
5296 class RemarkKlassClosure : public KlassClosure {
5297 CMKlassClosure _cm_klass_closure;
5298 public:
5299 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {}
5300 void do_klass(Klass* k) {
5301 // Check if we have modified any oops in the Klass during the concurrent marking.
5302 if (k->has_accumulated_modified_oops()) {
5303 k->clear_accumulated_modified_oops();
5304
5305 // We could have transfered the current modified marks to the accumulated marks,
5306 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
5307 } else if (k->has_modified_oops()) {
5308 // Don't clear anything, this info is needed by the next young collection.
5309 } else {
5310 // No modified oops in the Klass.
5311 return;
5312 }
5313
5314 // The klass has modified fields, need to scan the klass.
5315 _cm_klass_closure.do_klass(k);
5316 }
5317 };
5318
5319 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) {
5320 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration();
5321 EdenSpace* eden_space = dng->eden();
5322 ContiguousSpace* from_space = dng->from();
5323 ContiguousSpace* to_space = dng->to();
5324
5325 HeapWord** eca = _collector->_eden_chunk_array;
5326 size_t ect = _collector->_eden_chunk_index;
5327 HeapWord** sca = _collector->_survivor_chunk_array;
5328 size_t sct = _collector->_survivor_chunk_index;
5329
5330 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
5331 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
5332
5333 do_young_space_rescan(worker_id, cl, to_space, NULL, 0);
5334 do_young_space_rescan(worker_id, cl, from_space, sca, sct);
5335 do_young_space_rescan(worker_id, cl, eden_space, eca, ect);
5336 }
5337
5338 // work_queue(i) is passed to the closure
5339 // Par_MarkRefsIntoAndScanClosure. The "i" parameter
5340 // also is passed to do_dirty_card_rescan_tasks() and to
5341 // do_work_steal() to select the i-th task_queue.
5342
5343 void CMSParRemarkTask::work(uint worker_id) {
5344 elapsedTimer _timer;
5345 ResourceMark rm;
5346 HandleMark hm;
5347
5348 // ---------- rescan from roots --------------
5349 _timer.start();
5350 GenCollectedHeap* gch = GenCollectedHeap::heap();
5351 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector,
5352 _collector->_span, _collector->ref_processor(),
5353 &(_collector->_markBitMap),
5354 work_queue(worker_id));
5355
5356 // Rescan young gen roots first since these are likely
5357 // coarsely partitioned and may, on that account, constitute
5358 // the critical path; thus, it's best to start off that
5359 // work first.
5360 // ---------- young gen roots --------------
5361 {
5362 work_on_young_gen_roots(worker_id, &par_mrias_cl);
5363 _timer.stop();
5364 if (PrintCMSStatistics != 0) {
5365 gclog_or_tty->print_cr(
5366 "Finished young gen rescan work in %dth thread: %3.3f sec",
5367 worker_id, _timer.seconds());
5368 }
5369 }
5370
5371 // ---------- remaining roots --------------
5372 _timer.reset();
5373 _timer.start();
5374 gch->gen_process_strong_roots(_collector->_cmsGen->level(),
5375 false, // yg was scanned above
5376 false, // this is parallel code
5377 false, // not scavenging
5378 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
5379 &par_mrias_cl,
5380 true, // walk all of code cache if (so & SO_CodeCache)
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_CodeCache),
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 // imbedded in it a (subtype of) UpwardsObjectClosure. The
5540 // use of that state in the imbedded 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 // behaviour 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 bopundary).
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 false, // not scavenging
5969 SharedHeap::ScanningOption(roots_scanning_options()),
5970 &mrias_cl,
5971 true, // walk code active on stacks
5972 NULL,
5973 NULL); // The dirty klasses will be handled below
5974
5975 assert(should_unload_classes()
5976 || (roots_scanning_options() & SharedHeap::SO_CodeCache),
5977 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
5978 }
5979
5980 {
5981 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm);
5982
5983 verify_work_stacks_empty();
5984
5985 // Scan all class loader data objects that might have been introduced
5986 // during concurrent marking.
5987 ResourceMark rm;
5988 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
5989 for (int i = 0; i < array->length(); i++) {
5990 mrias_cl.do_class_loader_data(array->at(i));
5991 }
5992
5993 // We don't need to keep track of new CLDs anymore.
5994 ClassLoaderDataGraph::remember_new_clds(false);
5995
5996 verify_work_stacks_empty();
5997 }
5998
5999 {
6000 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm);
6001
6002 verify_work_stacks_empty();
6003
6004 RemarkKlassClosure remark_klass_closure(&mrias_cl);
6005 ClassLoaderDataGraph::classes_do(&remark_klass_closure);
6006
6007 verify_work_stacks_empty();
6008 }
6009
6010 // We might have added oops to ClassLoaderData::_handles during the
6011 // concurrent marking phase. These oops point to newly allocated objects
6012 // that are guaranteed to be kept alive. Either by the direct allocation
6013 // code, or when the young collector processes the strong roots. Hence,
6014 // we don't have to revisit the _handles block during the remark phase.
6015
6016 verify_work_stacks_empty();
6017 // Restore evacuated mark words, if any, used for overflow list links
6018 if (!CMSOverflowEarlyRestoration) {
6019 restore_preserved_marks_if_any();
6020 }
6021 verify_overflow_empty();
6022 }
6023
6024 ////////////////////////////////////////////////////////
6025 // Parallel Reference Processing Task Proxy Class
6026 ////////////////////////////////////////////////////////
6027 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
6028 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
6029 CMSCollector* _collector;
6030 CMSBitMap* _mark_bit_map;
6031 const MemRegion _span;
6032 ProcessTask& _task;
6033
6034 public:
6035 CMSRefProcTaskProxy(ProcessTask& task,
6036 CMSCollector* collector,
6037 const MemRegion& span,
6038 CMSBitMap* mark_bit_map,
6039 AbstractWorkGang* workers,
6040 OopTaskQueueSet* task_queues):
6041 // XXX Should superclass AGTWOQ also know about AWG since it knows
6042 // about the task_queues used by the AWG? Then it could initialize
6043 // the terminator() object. See 6984287. The set_for_termination()
6044 // below is a temporary band-aid for the regression in 6984287.
6045 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
6046 task_queues),
6047 _task(task),
6048 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
6049 {
6050 assert(_collector->_span.equals(_span) && !_span.is_empty(),
6051 "Inconsistency in _span");
6052 set_for_termination(workers->active_workers());
6053 }
6054
6055 OopTaskQueueSet* task_queues() { return queues(); }
6056
6057 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
6058
6059 void do_work_steal(int i,
6060 CMSParDrainMarkingStackClosure* drain,
6061 CMSParKeepAliveClosure* keep_alive,
6062 int* seed);
6063
6064 virtual void work(uint worker_id);
6065 };
6066
6067 void CMSRefProcTaskProxy::work(uint worker_id) {
6068 assert(_collector->_span.equals(_span), "Inconsistency in _span");
6069 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
6070 _mark_bit_map,
6071 work_queue(worker_id));
6072 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
6073 _mark_bit_map,
6074 work_queue(worker_id));
6075 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
6076 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
6077 if (_task.marks_oops_alive()) {
6078 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
6079 _collector->hash_seed(worker_id));
6080 }
6081 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
6082 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
6083 }
6084
6085 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
6086 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
6087 EnqueueTask& _task;
6088
6089 public:
6090 CMSRefEnqueueTaskProxy(EnqueueTask& task)
6091 : AbstractGangTask("Enqueue reference objects in parallel"),
6092 _task(task)
6093 { }
6094
6095 virtual void work(uint worker_id)
6096 {
6097 _task.work(worker_id);
6098 }
6099 };
6100
6101 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
6102 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
6103 _span(span),
6104 _bit_map(bit_map),
6105 _work_queue(work_queue),
6106 _mark_and_push(collector, span, bit_map, work_queue),
6107 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
6108 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads)))
6109 { }
6110
6111 // . see if we can share work_queues with ParNew? XXX
6112 void CMSRefProcTaskProxy::do_work_steal(int i,
6113 CMSParDrainMarkingStackClosure* drain,
6114 CMSParKeepAliveClosure* keep_alive,
6115 int* seed) {
6116 OopTaskQueue* work_q = work_queue(i);
6117 NOT_PRODUCT(int num_steals = 0;)
6118 oop obj_to_scan;
6119
6120 while (true) {
6121 // Completely finish any left over work from (an) earlier round(s)
6122 drain->trim_queue(0);
6123 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
6124 (size_t)ParGCDesiredObjsFromOverflowList);
6125 // Now check if there's any work in the overflow list
6126 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
6127 // only affects the number of attempts made to get work from the
6128 // overflow list and does not affect the number of workers. Just
6129 // pass ParallelGCThreads so this behavior is unchanged.
6130 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
6131 work_q,
6132 ParallelGCThreads)) {
6133 // Found something in global overflow list;
6134 // not yet ready to go stealing work from others.
6135 // We'd like to assert(work_q->size() != 0, ...)
6136 // because we just took work from the overflow list,
6137 // but of course we can't, since all of that might have
6138 // been already stolen from us.
6139 continue;
6140 }
6141 // Verify that we have no work before we resort to stealing
6142 assert(work_q->size() == 0, "Have work, shouldn't steal");
6143 // Try to steal from other queues that have work
6144 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
6145 NOT_PRODUCT(num_steals++;)
6146 assert(obj_to_scan->is_oop(), "Oops, not an oop!");
6147 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
6148 // Do scanning work
6149 obj_to_scan->oop_iterate(keep_alive);
6150 // Loop around, finish this work, and try to steal some more
6151 } else if (terminator()->offer_termination()) {
6152 break; // nirvana from the infinite cycle
6153 }
6154 }
6155 NOT_PRODUCT(
6156 if (PrintCMSStatistics != 0) {
6157 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals);
6158 }
6159 )
6160 }
6161
6162 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
6163 {
6164 GenCollectedHeap* gch = GenCollectedHeap::heap();
6165 FlexibleWorkGang* workers = gch->workers();
6166 assert(workers != NULL, "Need parallel worker threads.");
6167 CMSRefProcTaskProxy rp_task(task, &_collector,
6168 _collector.ref_processor()->span(),
6169 _collector.markBitMap(),
6170 workers, _collector.task_queues());
6171 workers->run_task(&rp_task);
6172 }
6173
6174 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
6175 {
6176
6177 GenCollectedHeap* gch = GenCollectedHeap::heap();
6178 FlexibleWorkGang* workers = gch->workers();
6179 assert(workers != NULL, "Need parallel worker threads.");
6180 CMSRefEnqueueTaskProxy enq_task(task);
6181 workers->run_task(&enq_task);
6182 }
6183
6184 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) {
6185
6186 ResourceMark rm;
6187 HandleMark hm;
6188
6189 ReferenceProcessor* rp = ref_processor();
6190 assert(rp->span().equals(_span), "Spans should be equal");
6191 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
6192 // Process weak references.
6193 rp->setup_policy(clear_all_soft_refs);
6194 verify_work_stacks_empty();
6195
6196 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
6197 &_markStack, false /* !preclean */);
6198 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
6199 _span, &_markBitMap, &_markStack,
6200 &cmsKeepAliveClosure, false /* !preclean */);
6201 {
6202 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm);
6203
6204 ReferenceProcessorStats stats;
6205 if (rp->processing_is_mt()) {
6206 // Set the degree of MT here. If the discovery is done MT, there
6207 // may have been a different number of threads doing the discovery
6208 // and a different number of discovered lists may have Ref objects.
6209 // That is OK as long as the Reference lists are balanced (see
6210 // balance_all_queues() and balance_queues()).
6211 GenCollectedHeap* gch = GenCollectedHeap::heap();
6212 int active_workers = ParallelGCThreads;
6213 FlexibleWorkGang* workers = gch->workers();
6214 if (workers != NULL) {
6215 active_workers = workers->active_workers();
6216 // The expectation is that active_workers will have already
6217 // been set to a reasonable value. If it has not been set,
6218 // investigate.
6219 assert(active_workers > 0, "Should have been set during scavenge");
6220 }
6221 rp->set_active_mt_degree(active_workers);
6222 CMSRefProcTaskExecutor task_executor(*this);
6223 stats = rp->process_discovered_references(&_is_alive_closure,
6224 &cmsKeepAliveClosure,
6225 &cmsDrainMarkingStackClosure,
6226 &task_executor,
6227 _gc_timer_cm);
6228 } else {
6229 stats = rp->process_discovered_references(&_is_alive_closure,
6230 &cmsKeepAliveClosure,
6231 &cmsDrainMarkingStackClosure,
6232 NULL,
6233 _gc_timer_cm);
6234 }
6235 _gc_tracer_cm->report_gc_reference_stats(stats);
6236
6237 }
6238
6239 // This is the point where the entire marking should have completed.
6240 verify_work_stacks_empty();
6241
6242 if (should_unload_classes()) {
6243 {
6244 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm);
6245
6246 // Unload classes and purge the SystemDictionary.
6247 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure);
6248
6249 // Unload nmethods.
6250 CodeCache::do_unloading(&_is_alive_closure, purged_class);
6251
6252 // Prune dead klasses from subklass/sibling/implementor lists.
6253 Klass::clean_weak_klass_links(&_is_alive_closure);
6254 }
6255
6256 {
6257 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm);
6258 // Clean up unreferenced symbols in symbol table.
6259 SymbolTable::unlink();
6260 }
6261 }
6262
6263 // CMS doesn't use the StringTable as hard roots when class unloading is turned off.
6264 // Need to check if we really scanned the StringTable.
6265 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) {
6266 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm);
6267 // Delete entries for dead interned strings.
6268 StringTable::unlink(&_is_alive_closure);
6269 }
6270
6271 // Restore any preserved marks as a result of mark stack or
6272 // work queue overflow
6273 restore_preserved_marks_if_any(); // done single-threaded for now
6274
6275 rp->set_enqueuing_is_done(true);
6276 if (rp->processing_is_mt()) {
6277 rp->balance_all_queues();
6278 CMSRefProcTaskExecutor task_executor(*this);
6279 rp->enqueue_discovered_references(&task_executor);
6280 } else {
6281 rp->enqueue_discovered_references(NULL);
6282 }
6283 rp->verify_no_references_recorded();
6284 assert(!rp->discovery_enabled(), "should have been disabled");
6285 }
6286
6287 #ifndef PRODUCT
6288 void CMSCollector::check_correct_thread_executing() {
6289 Thread* t = Thread::current();
6290 // Only the VM thread or the CMS thread should be here.
6291 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
6292 "Unexpected thread type");
6293 // If this is the vm thread, the foreground process
6294 // should not be waiting. Note that _foregroundGCIsActive is
6295 // true while the foreground collector is waiting.
6296 if (_foregroundGCShouldWait) {
6297 // We cannot be the VM thread
6298 assert(t->is_ConcurrentGC_thread(),
6299 "Should be CMS thread");
6300 } else {
6301 // We can be the CMS thread only if we are in a stop-world
6302 // phase of CMS collection.
6303 if (t->is_ConcurrentGC_thread()) {
6304 assert(_collectorState == InitialMarking ||
6305 _collectorState == FinalMarking,
6306 "Should be a stop-world phase");
6307 // The CMS thread should be holding the CMS_token.
6308 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6309 "Potential interference with concurrently "
6310 "executing VM thread");
6311 }
6312 }
6313 }
6314 #endif
6315
6316 void CMSCollector::sweep(bool asynch) {
6317 assert(_collectorState == Sweeping, "just checking");
6318 check_correct_thread_executing();
6319 verify_work_stacks_empty();
6320 verify_overflow_empty();
6321 increment_sweep_count();
6322 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause());
6323
6324 _inter_sweep_timer.stop();
6325 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
6326 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free());
6327
6328 assert(!_intra_sweep_timer.is_active(), "Should not be active");
6329 _intra_sweep_timer.reset();
6330 _intra_sweep_timer.start();
6331 if (asynch) {
6332 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6333 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails);
6334 // First sweep the old gen
6335 {
6336 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
6337 bitMapLock());
6338 sweepWork(_cmsGen, asynch);
6339 }
6340
6341 // Update Universe::_heap_*_at_gc figures.
6342 // We need all the free list locks to make the abstract state
6343 // transition from Sweeping to Resetting. See detailed note
6344 // further below.
6345 {
6346 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
6347 // Update heap occupancy information which is used as
6348 // input to soft ref clearing policy at the next gc.
6349 Universe::update_heap_info_at_gc();
6350 _collectorState = Resizing;
6351 }
6352 } else {
6353 // already have needed locks
6354 sweepWork(_cmsGen, asynch);
6355 // Update heap occupancy information which is used as
6356 // input to soft ref clearing policy at the next gc.
6357 Universe::update_heap_info_at_gc();
6358 _collectorState = Resizing;
6359 }
6360 verify_work_stacks_empty();
6361 verify_overflow_empty();
6362
6363 if (should_unload_classes()) {
6364 ClassLoaderDataGraph::purge();
6365 }
6366
6367 _intra_sweep_timer.stop();
6368 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
6369
6370 _inter_sweep_timer.reset();
6371 _inter_sweep_timer.start();
6372
6373 // We need to use a monotonically non-deccreasing time in ms
6374 // or we will see time-warp warnings and os::javaTimeMillis()
6375 // does not guarantee monotonicity.
6376 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
6377 update_time_of_last_gc(now);
6378
6379 // NOTE on abstract state transitions:
6380 // Mutators allocate-live and/or mark the mod-union table dirty
6381 // based on the state of the collection. The former is done in
6382 // the interval [Marking, Sweeping] and the latter in the interval
6383 // [Marking, Sweeping). Thus the transitions into the Marking state
6384 // and out of the Sweeping state must be synchronously visible
6385 // globally to the mutators.
6386 // The transition into the Marking state happens with the world
6387 // stopped so the mutators will globally see it. Sweeping is
6388 // done asynchronously by the background collector so the transition
6389 // from the Sweeping state to the Resizing state must be done
6390 // under the freelistLock (as is the check for whether to
6391 // allocate-live and whether to dirty the mod-union table).
6392 assert(_collectorState == Resizing, "Change of collector state to"
6393 " Resizing must be done under the freelistLocks (plural)");
6394
6395 // Now that sweeping has been completed, we clear
6396 // the incremental_collection_failed flag,
6397 // thus inviting a younger gen collection to promote into
6398 // this generation. If such a promotion may still fail,
6399 // the flag will be set again when a young collection is
6400 // attempted.
6401 GenCollectedHeap* gch = GenCollectedHeap::heap();
6402 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
6403 gch->update_full_collections_completed(_collection_count_start);
6404 }
6405
6406 // FIX ME!!! Looks like this belongs in CFLSpace, with
6407 // CMSGen merely delegating to it.
6408 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
6409 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
6410 HeapWord* minAddr = _cmsSpace->bottom();
6411 HeapWord* largestAddr =
6412 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
6413 if (largestAddr == NULL) {
6414 // The dictionary appears to be empty. In this case
6415 // try to coalesce at the end of the heap.
6416 largestAddr = _cmsSpace->end();
6417 }
6418 size_t largestOffset = pointer_delta(largestAddr, minAddr);
6419 size_t nearLargestOffset =
6420 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
6421 if (PrintFLSStatistics != 0) {
6422 gclog_or_tty->print_cr(
6423 "CMS: Large Block: " PTR_FORMAT ";"
6424 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
6425 largestAddr,
6426 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset);
6427 }
6428 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
6429 }
6430
6431 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
6432 return addr >= _cmsSpace->nearLargestChunk();
6433 }
6434
6435 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
6436 return _cmsSpace->find_chunk_at_end();
6437 }
6438
6439 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level,
6440 bool full) {
6441 // The next lower level has been collected. Gather any statistics
6442 // that are of interest at this point.
6443 if (!full && (current_level + 1) == level()) {
6444 // Gather statistics on the young generation collection.
6445 collector()->stats().record_gc0_end(used());
6446 }
6447 }
6448
6449 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() {
6450 GenCollectedHeap* gch = GenCollectedHeap::heap();
6451 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
6452 "Wrong type of heap");
6453 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*)
6454 gch->gen_policy()->size_policy();
6455 assert(sp->is_gc_cms_adaptive_size_policy(),
6456 "Wrong type of size policy");
6457 return sp;
6458 }
6459
6460 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() {
6461 if (PrintGCDetails && Verbose) {
6462 gclog_or_tty->print("Rotate from %d ", _debug_collection_type);
6463 }
6464 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1);
6465 _debug_collection_type =
6466 (CollectionTypes) (_debug_collection_type % Unknown_collection_type);
6467 if (PrintGCDetails && Verbose) {
6468 gclog_or_tty->print_cr("to %d ", _debug_collection_type);
6469 }
6470 }
6471
6472 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen,
6473 bool asynch) {
6474 // We iterate over the space(s) underlying this generation,
6475 // checking the mark bit map to see if the bits corresponding
6476 // to specific blocks are marked or not. Blocks that are
6477 // marked are live and are not swept up. All remaining blocks
6478 // are swept up, with coalescing on-the-fly as we sweep up
6479 // contiguous free and/or garbage blocks:
6480 // We need to ensure that the sweeper synchronizes with allocators
6481 // and stop-the-world collectors. In particular, the following
6482 // locks are used:
6483 // . CMS token: if this is held, a stop the world collection cannot occur
6484 // . freelistLock: if this is held no allocation can occur from this
6485 // generation by another thread
6486 // . bitMapLock: if this is held, no other thread can access or update
6487 //
6488
6489 // Note that we need to hold the freelistLock if we use
6490 // block iterate below; else the iterator might go awry if
6491 // a mutator (or promotion) causes block contents to change
6492 // (for instance if the allocator divvies up a block).
6493 // If we hold the free list lock, for all practical purposes
6494 // young generation GC's can't occur (they'll usually need to
6495 // promote), so we might as well prevent all young generation
6496 // GC's while we do a sweeping step. For the same reason, we might
6497 // as well take the bit map lock for the entire duration
6498
6499 // check that we hold the requisite locks
6500 assert(have_cms_token(), "Should hold cms token");
6501 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token())
6502 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()),
6503 "Should possess CMS token to sweep");
6504 assert_lock_strong(gen->freelistLock());
6505 assert_lock_strong(bitMapLock());
6506
6507 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
6508 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
6509 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
6510 _inter_sweep_estimate.padded_average(),
6511 _intra_sweep_estimate.padded_average());
6512 gen->setNearLargestChunk();
6513
6514 {
6515 SweepClosure sweepClosure(this, gen, &_markBitMap,
6516 CMSYield && asynch);
6517 gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
6518 // We need to free-up/coalesce garbage/blocks from a
6519 // co-terminal free run. This is done in the SweepClosure
6520 // destructor; so, do not remove this scope, else the
6521 // end-of-sweep-census below will be off by a little bit.
6522 }
6523 gen->cmsSpace()->sweep_completed();
6524 gen->cmsSpace()->endSweepFLCensus(sweep_count());
6525 if (should_unload_classes()) { // unloaded classes this cycle,
6526 _concurrent_cycles_since_last_unload = 0; // ... reset count
6527 } else { // did not unload classes,
6528 _concurrent_cycles_since_last_unload++; // ... increment count
6529 }
6530 }
6531
6532 // Reset CMS data structures (for now just the marking bit map)
6533 // preparatory for the next cycle.
6534 void CMSCollector::reset(bool asynch) {
6535 GenCollectedHeap* gch = GenCollectedHeap::heap();
6536 CMSAdaptiveSizePolicy* sp = size_policy();
6537 AdaptiveSizePolicyOutput(sp, gch->total_collections());
6538 if (asynch) {
6539 CMSTokenSyncWithLocks ts(true, bitMapLock());
6540
6541 // If the state is not "Resetting", the foreground thread
6542 // has done a collection and the resetting.
6543 if (_collectorState != Resetting) {
6544 assert(_collectorState == Idling, "The state should only change"
6545 " because the foreground collector has finished the collection");
6546 return;
6547 }
6548
6549 // Clear the mark bitmap (no grey objects to start with)
6550 // for the next cycle.
6551 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6552 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails);
6553
6554 HeapWord* curAddr = _markBitMap.startWord();
6555 while (curAddr < _markBitMap.endWord()) {
6556 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
6557 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
6558 _markBitMap.clear_large_range(chunk);
6559 if (ConcurrentMarkSweepThread::should_yield() &&
6560 !foregroundGCIsActive() &&
6561 CMSYield) {
6562 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6563 "CMS thread should hold CMS token");
6564 assert_lock_strong(bitMapLock());
6565 bitMapLock()->unlock();
6566 ConcurrentMarkSweepThread::desynchronize(true);
6567 ConcurrentMarkSweepThread::acknowledge_yield_request();
6568 stopTimer();
6569 if (PrintCMSStatistics != 0) {
6570 incrementYields();
6571 }
6572 icms_wait();
6573
6574 // See the comment in coordinator_yield()
6575 for (unsigned i = 0; i < CMSYieldSleepCount &&
6576 ConcurrentMarkSweepThread::should_yield() &&
6577 !CMSCollector::foregroundGCIsActive(); ++i) {
6578 os::sleep(Thread::current(), 1, false);
6579 ConcurrentMarkSweepThread::acknowledge_yield_request();
6580 }
6581
6582 ConcurrentMarkSweepThread::synchronize(true);
6583 bitMapLock()->lock_without_safepoint_check();
6584 startTimer();
6585 }
6586 curAddr = chunk.end();
6587 }
6588 // A successful mostly concurrent collection has been done.
6589 // Because only the full (i.e., concurrent mode failure) collections
6590 // are being measured for gc overhead limits, clean the "near" flag
6591 // and count.
6592 sp->reset_gc_overhead_limit_count();
6593 _collectorState = Idling;
6594 } else {
6595 // already have the lock
6596 assert(_collectorState == Resetting, "just checking");
6597 assert_lock_strong(bitMapLock());
6598 _markBitMap.clear_all();
6599 _collectorState = Idling;
6600 }
6601
6602 // Stop incremental mode after a cycle completes, so that any future cycles
6603 // are triggered by allocation.
6604 stop_icms();
6605
6606 NOT_PRODUCT(
6607 if (RotateCMSCollectionTypes) {
6608 _cmsGen->rotate_debug_collection_type();
6609 }
6610 )
6611
6612 register_gc_end();
6613 }
6614
6615 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
6616 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps);
6617 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty);
6618 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL);
6619 TraceCollectorStats tcs(counters());
6620
6621 switch (op) {
6622 case CMS_op_checkpointRootsInitial: {
6623 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6624 checkpointRootsInitial(true); // asynch
6625 if (PrintGC) {
6626 _cmsGen->printOccupancy("initial-mark");
6627 }
6628 break;
6629 }
6630 case CMS_op_checkpointRootsFinal: {
6631 SvcGCMarker sgcm(SvcGCMarker::OTHER);
6632 checkpointRootsFinal(true, // asynch
6633 false, // !clear_all_soft_refs
6634 false); // !init_mark_was_synchronous
6635 if (PrintGC) {
6636 _cmsGen->printOccupancy("remark");
6637 }
6638 break;
6639 }
6640 default:
6641 fatal("No such CMS_op");
6642 }
6643 }
6644
6645 #ifndef PRODUCT
6646 size_t const CMSCollector::skip_header_HeapWords() {
6647 return FreeChunk::header_size();
6648 }
6649
6650 // Try and collect here conditions that should hold when
6651 // CMS thread is exiting. The idea is that the foreground GC
6652 // thread should not be blocked if it wants to terminate
6653 // the CMS thread and yet continue to run the VM for a while
6654 // after that.
6655 void CMSCollector::verify_ok_to_terminate() const {
6656 assert(Thread::current()->is_ConcurrentGC_thread(),
6657 "should be called by CMS thread");
6658 assert(!_foregroundGCShouldWait, "should be false");
6659 // We could check here that all the various low-level locks
6660 // are not held by the CMS thread, but that is overkill; see
6661 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
6662 // is checked.
6663 }
6664 #endif
6665
6666 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
6667 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
6668 "missing Printezis mark?");
6669 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6670 size_t size = pointer_delta(nextOneAddr + 1, addr);
6671 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6672 "alignment problem");
6673 assert(size >= 3, "Necessary for Printezis marks to work");
6674 return size;
6675 }
6676
6677 // A variant of the above (block_size_using_printezis_bits()) except
6678 // that we return 0 if the P-bits are not yet set.
6679 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
6680 if (_markBitMap.isMarked(addr + 1)) {
6681 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
6682 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
6683 size_t size = pointer_delta(nextOneAddr + 1, addr);
6684 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6685 "alignment problem");
6686 assert(size >= 3, "Necessary for Printezis marks to work");
6687 return size;
6688 }
6689 return 0;
6690 }
6691
6692 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
6693 size_t sz = 0;
6694 oop p = (oop)addr;
6695 if (p->klass_or_null() != NULL) {
6696 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
6697 } else {
6698 sz = block_size_using_printezis_bits(addr);
6699 }
6700 assert(sz > 0, "size must be nonzero");
6701 HeapWord* next_block = addr + sz;
6702 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block,
6703 CardTableModRefBS::card_size);
6704 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) <
6705 round_down((uintptr_t)next_card, CardTableModRefBS::card_size),
6706 "must be different cards");
6707 return next_card;
6708 }
6709
6710
6711 // CMS Bit Map Wrapper /////////////////////////////////////////
6712
6713 // Construct a CMS bit map infrastructure, but don't create the
6714 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
6715 // further below.
6716 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
6717 _bm(),
6718 _shifter(shifter),
6719 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL)
6720 {
6721 _bmStartWord = 0;
6722 _bmWordSize = 0;
6723 }
6724
6725 bool CMSBitMap::allocate(MemRegion mr) {
6726 _bmStartWord = mr.start();
6727 _bmWordSize = mr.word_size();
6728 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
6729 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
6730 if (!brs.is_reserved()) {
6731 warning("CMS bit map allocation failure");
6732 return false;
6733 }
6734 // For now we'll just commit all of the bit map up fromt.
6735 // Later on we'll try to be more parsimonious with swap.
6736 if (!_virtual_space.initialize(brs, brs.size())) {
6737 warning("CMS bit map backing store failure");
6738 return false;
6739 }
6740 assert(_virtual_space.committed_size() == brs.size(),
6741 "didn't reserve backing store for all of CMS bit map?");
6742 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low());
6743 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
6744 _bmWordSize, "inconsistency in bit map sizing");
6745 _bm.set_size(_bmWordSize >> _shifter);
6746
6747 // bm.clear(); // can we rely on getting zero'd memory? verify below
6748 assert(isAllClear(),
6749 "Expected zero'd memory from ReservedSpace constructor");
6750 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
6751 "consistency check");
6752 return true;
6753 }
6754
6755 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
6756 HeapWord *next_addr, *end_addr, *last_addr;
6757 assert_locked();
6758 assert(covers(mr), "out-of-range error");
6759 // XXX assert that start and end are appropriately aligned
6760 for (next_addr = mr.start(), end_addr = mr.end();
6761 next_addr < end_addr; next_addr = last_addr) {
6762 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
6763 last_addr = dirty_region.end();
6764 if (!dirty_region.is_empty()) {
6765 cl->do_MemRegion(dirty_region);
6766 } else {
6767 assert(last_addr == end_addr, "program logic");
6768 return;
6769 }
6770 }
6771 }
6772
6773 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
6774 _bm.print_on_error(st, prefix);
6775 }
6776
6777 #ifndef PRODUCT
6778 void CMSBitMap::assert_locked() const {
6779 CMSLockVerifier::assert_locked(lock());
6780 }
6781
6782 bool CMSBitMap::covers(MemRegion mr) const {
6783 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
6784 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
6785 "size inconsistency");
6786 return (mr.start() >= _bmStartWord) &&
6787 (mr.end() <= endWord());
6788 }
6789
6790 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
6791 return (start >= _bmStartWord && (start + size) <= endWord());
6792 }
6793
6794 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
6795 // verify that there are no 1 bits in the interval [left, right)
6796 FalseBitMapClosure falseBitMapClosure;
6797 iterate(&falseBitMapClosure, left, right);
6798 }
6799
6800 void CMSBitMap::region_invariant(MemRegion mr)
6801 {
6802 assert_locked();
6803 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
6804 assert(!mr.is_empty(), "unexpected empty region");
6805 assert(covers(mr), "mr should be covered by bit map");
6806 // convert address range into offset range
6807 size_t start_ofs = heapWordToOffset(mr.start());
6808 // Make sure that end() is appropriately aligned
6809 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(),
6810 (1 << (_shifter+LogHeapWordSize))),
6811 "Misaligned mr.end()");
6812 size_t end_ofs = heapWordToOffset(mr.end());
6813 assert(end_ofs > start_ofs, "Should mark at least one bit");
6814 }
6815
6816 #endif
6817
6818 bool CMSMarkStack::allocate(size_t size) {
6819 // allocate a stack of the requisite depth
6820 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6821 size * sizeof(oop)));
6822 if (!rs.is_reserved()) {
6823 warning("CMSMarkStack allocation failure");
6824 return false;
6825 }
6826 if (!_virtual_space.initialize(rs, rs.size())) {
6827 warning("CMSMarkStack backing store failure");
6828 return false;
6829 }
6830 assert(_virtual_space.committed_size() == rs.size(),
6831 "didn't reserve backing store for all of CMS stack?");
6832 _base = (oop*)(_virtual_space.low());
6833 _index = 0;
6834 _capacity = size;
6835 NOT_PRODUCT(_max_depth = 0);
6836 return true;
6837 }
6838
6839 // XXX FIX ME !!! In the MT case we come in here holding a
6840 // leaf lock. For printing we need to take a further lock
6841 // which has lower rank. We need to recallibrate the two
6842 // lock-ranks involved in order to be able to rpint the
6843 // messages below. (Or defer the printing to the caller.
6844 // For now we take the expedient path of just disabling the
6845 // messages for the problematic case.)
6846 void CMSMarkStack::expand() {
6847 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
6848 if (_capacity == MarkStackSizeMax) {
6849 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6850 // We print a warning message only once per CMS cycle.
6851 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit");
6852 }
6853 return;
6854 }
6855 // Double capacity if possible
6856 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
6857 // Do not give up existing stack until we have managed to
6858 // get the double capacity that we desired.
6859 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
6860 new_capacity * sizeof(oop)));
6861 if (rs.is_reserved()) {
6862 // Release the backing store associated with old stack
6863 _virtual_space.release();
6864 // Reinitialize virtual space for new stack
6865 if (!_virtual_space.initialize(rs, rs.size())) {
6866 fatal("Not enough swap for expanded marking stack");
6867 }
6868 _base = (oop*)(_virtual_space.low());
6869 _index = 0;
6870 _capacity = new_capacity;
6871 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) {
6872 // Failed to double capacity, continue;
6873 // we print a detail message only once per CMS cycle.
6874 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to "
6875 SIZE_FORMAT"K",
6876 _capacity / K, new_capacity / K);
6877 }
6878 }
6879
6880
6881 // Closures
6882 // XXX: there seems to be a lot of code duplication here;
6883 // should refactor and consolidate common code.
6884
6885 // This closure is used to mark refs into the CMS generation in
6886 // the CMS bit map. Called at the first checkpoint. This closure
6887 // assumes that we do not need to re-mark dirty cards; if the CMS
6888 // generation on which this is used is not an oldest
6889 // generation then this will lose younger_gen cards!
6890
6891 MarkRefsIntoClosure::MarkRefsIntoClosure(
6892 MemRegion span, CMSBitMap* bitMap):
6893 _span(span),
6894 _bitMap(bitMap)
6895 {
6896 assert(_ref_processor == NULL, "deliberately left NULL");
6897 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6898 }
6899
6900 void MarkRefsIntoClosure::do_oop(oop obj) {
6901 // if p points into _span, then mark corresponding bit in _markBitMap
6902 assert(obj->is_oop(), "expected an oop");
6903 HeapWord* addr = (HeapWord*)obj;
6904 if (_span.contains(addr)) {
6905 // this should be made more efficient
6906 _bitMap->mark(addr);
6907 }
6908 }
6909
6910 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6911 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
6912
6913 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure(
6914 MemRegion span, CMSBitMap* bitMap):
6915 _span(span),
6916 _bitMap(bitMap)
6917 {
6918 assert(_ref_processor == NULL, "deliberately left NULL");
6919 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
6920 }
6921
6922 void Par_MarkRefsIntoClosure::do_oop(oop obj) {
6923 // if p points into _span, then mark corresponding bit in _markBitMap
6924 assert(obj->is_oop(), "expected an oop");
6925 HeapWord* addr = (HeapWord*)obj;
6926 if (_span.contains(addr)) {
6927 // this should be made more efficient
6928 _bitMap->par_mark(addr);
6929 }
6930 }
6931
6932 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6933 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); }
6934
6935 // A variant of the above, used for CMS marking verification.
6936 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
6937 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
6938 _span(span),
6939 _verification_bm(verification_bm),
6940 _cms_bm(cms_bm)
6941 {
6942 assert(_ref_processor == NULL, "deliberately left NULL");
6943 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
6944 }
6945
6946 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
6947 // if p points into _span, then mark corresponding bit in _markBitMap
6948 assert(obj->is_oop(), "expected an oop");
6949 HeapWord* addr = (HeapWord*)obj;
6950 if (_span.contains(addr)) {
6951 _verification_bm->mark(addr);
6952 if (!_cms_bm->isMarked(addr)) {
6953 oop(addr)->print();
6954 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr);
6955 fatal("... aborting");
6956 }
6957 }
6958 }
6959
6960 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6961 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
6962
6963 //////////////////////////////////////////////////
6964 // MarkRefsIntoAndScanClosure
6965 //////////////////////////////////////////////////
6966
6967 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
6968 ReferenceProcessor* rp,
6969 CMSBitMap* bit_map,
6970 CMSBitMap* mod_union_table,
6971 CMSMarkStack* mark_stack,
6972 CMSCollector* collector,
6973 bool should_yield,
6974 bool concurrent_precleaning):
6975 _collector(collector),
6976 _span(span),
6977 _bit_map(bit_map),
6978 _mark_stack(mark_stack),
6979 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table,
6980 mark_stack, concurrent_precleaning),
6981 _yield(should_yield),
6982 _concurrent_precleaning(concurrent_precleaning),
6983 _freelistLock(NULL)
6984 {
6985 _ref_processor = rp;
6986 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
6987 }
6988
6989 // This closure is used to mark refs into the CMS generation at the
6990 // second (final) checkpoint, and to scan and transitively follow
6991 // the unmarked oops. It is also used during the concurrent precleaning
6992 // phase while scanning objects on dirty cards in the CMS generation.
6993 // The marks are made in the marking bit map and the marking stack is
6994 // used for keeping the (newly) grey objects during the scan.
6995 // The parallel version (Par_...) appears further below.
6996 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
6997 if (obj != NULL) {
6998 assert(obj->is_oop(), "expected an oop");
6999 HeapWord* addr = (HeapWord*)obj;
7000 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7001 assert(_collector->overflow_list_is_empty(),
7002 "overflow list should be empty");
7003 if (_span.contains(addr) &&
7004 !_bit_map->isMarked(addr)) {
7005 // mark bit map (object is now grey)
7006 _bit_map->mark(addr);
7007 // push on marking stack (stack should be empty), and drain the
7008 // stack by applying this closure to the oops in the oops popped
7009 // from the stack (i.e. blacken the grey objects)
7010 bool res = _mark_stack->push(obj);
7011 assert(res, "Should have space to push on empty stack");
7012 do {
7013 oop new_oop = _mark_stack->pop();
7014 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7015 assert(_bit_map->isMarked((HeapWord*)new_oop),
7016 "only grey objects on this stack");
7017 // iterate over the oops in this oop, marking and pushing
7018 // the ones in CMS heap (i.e. in _span).
7019 new_oop->oop_iterate(&_pushAndMarkClosure);
7020 // check if it's time to yield
7021 do_yield_check();
7022 } while (!_mark_stack->isEmpty() ||
7023 (!_concurrent_precleaning && take_from_overflow_list()));
7024 // if marking stack is empty, and we are not doing this
7025 // during precleaning, then check the overflow list
7026 }
7027 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7028 assert(_collector->overflow_list_is_empty(),
7029 "overflow list was drained above");
7030 // We could restore evacuated mark words, if any, used for
7031 // overflow list links here because the overflow list is
7032 // provably empty here. That would reduce the maximum
7033 // size requirements for preserved_{oop,mark}_stack.
7034 // But we'll just postpone it until we are all done
7035 // so we can just stream through.
7036 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) {
7037 _collector->restore_preserved_marks_if_any();
7038 assert(_collector->no_preserved_marks(), "No preserved marks");
7039 }
7040 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(),
7041 "All preserved marks should have been restored above");
7042 }
7043 }
7044
7045 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
7046 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
7047
7048 void MarkRefsIntoAndScanClosure::do_yield_work() {
7049 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7050 "CMS thread should hold CMS token");
7051 assert_lock_strong(_freelistLock);
7052 assert_lock_strong(_bit_map->lock());
7053 // relinquish the free_list_lock and bitMaplock()
7054 _bit_map->lock()->unlock();
7055 _freelistLock->unlock();
7056 ConcurrentMarkSweepThread::desynchronize(true);
7057 ConcurrentMarkSweepThread::acknowledge_yield_request();
7058 _collector->stopTimer();
7059 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7060 if (PrintCMSStatistics != 0) {
7061 _collector->incrementYields();
7062 }
7063 _collector->icms_wait();
7064
7065 // See the comment in coordinator_yield()
7066 for (unsigned i = 0;
7067 i < CMSYieldSleepCount &&
7068 ConcurrentMarkSweepThread::should_yield() &&
7069 !CMSCollector::foregroundGCIsActive();
7070 ++i) {
7071 os::sleep(Thread::current(), 1, false);
7072 ConcurrentMarkSweepThread::acknowledge_yield_request();
7073 }
7074
7075 ConcurrentMarkSweepThread::synchronize(true);
7076 _freelistLock->lock_without_safepoint_check();
7077 _bit_map->lock()->lock_without_safepoint_check();
7078 _collector->startTimer();
7079 }
7080
7081 ///////////////////////////////////////////////////////////
7082 // Par_MarkRefsIntoAndScanClosure: a parallel version of
7083 // MarkRefsIntoAndScanClosure
7084 ///////////////////////////////////////////////////////////
7085 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure(
7086 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp,
7087 CMSBitMap* bit_map, OopTaskQueue* work_queue):
7088 _span(span),
7089 _bit_map(bit_map),
7090 _work_queue(work_queue),
7091 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4),
7092 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))),
7093 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue)
7094 {
7095 _ref_processor = rp;
7096 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7097 }
7098
7099 // This closure is used to mark refs into the CMS generation at the
7100 // second (final) checkpoint, and to scan and transitively follow
7101 // the unmarked oops. The marks are made in the marking bit map and
7102 // the work_queue is used for keeping the (newly) grey objects during
7103 // the scan phase whence they are also available for stealing by parallel
7104 // threads. Since the marking bit map is shared, updates are
7105 // synchronized (via CAS).
7106 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) {
7107 if (obj != NULL) {
7108 // Ignore mark word because this could be an already marked oop
7109 // that may be chained at the end of the overflow list.
7110 assert(obj->is_oop(true), "expected an oop");
7111 HeapWord* addr = (HeapWord*)obj;
7112 if (_span.contains(addr) &&
7113 !_bit_map->isMarked(addr)) {
7114 // mark bit map (object will become grey):
7115 // It is possible for several threads to be
7116 // trying to "claim" this object concurrently;
7117 // the unique thread that succeeds in marking the
7118 // object first will do the subsequent push on
7119 // to the work queue (or overflow list).
7120 if (_bit_map->par_mark(addr)) {
7121 // push on work_queue (which may not be empty), and trim the
7122 // queue to an appropriate length by applying this closure to
7123 // the oops in the oops popped from the stack (i.e. blacken the
7124 // grey objects)
7125 bool res = _work_queue->push(obj);
7126 assert(res, "Low water mark should be less than capacity?");
7127 trim_queue(_low_water_mark);
7128 } // Else, another thread claimed the object
7129 }
7130 }
7131 }
7132
7133 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7134 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); }
7135
7136 // This closure is used to rescan the marked objects on the dirty cards
7137 // in the mod union table and the card table proper.
7138 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
7139 oop p, MemRegion mr) {
7140
7141 size_t size = 0;
7142 HeapWord* addr = (HeapWord*)p;
7143 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7144 assert(_span.contains(addr), "we are scanning the CMS generation");
7145 // check if it's time to yield
7146 if (do_yield_check()) {
7147 // We yielded for some foreground stop-world work,
7148 // and we have been asked to abort this ongoing preclean cycle.
7149 return 0;
7150 }
7151 if (_bitMap->isMarked(addr)) {
7152 // it's marked; is it potentially uninitialized?
7153 if (p->klass_or_null() != NULL) {
7154 // an initialized object; ignore mark word in verification below
7155 // since we are running concurrent with mutators
7156 assert(p->is_oop(true), "should be an oop");
7157 if (p->is_objArray()) {
7158 // objArrays are precisely marked; restrict scanning
7159 // to dirty cards only.
7160 size = CompactibleFreeListSpace::adjustObjectSize(
7161 p->oop_iterate(_scanningClosure, mr));
7162 } else {
7163 // A non-array may have been imprecisely marked; we need
7164 // to scan object in its entirety.
7165 size = CompactibleFreeListSpace::adjustObjectSize(
7166 p->oop_iterate(_scanningClosure));
7167 }
7168 #ifdef ASSERT
7169 size_t direct_size =
7170 CompactibleFreeListSpace::adjustObjectSize(p->size());
7171 assert(size == direct_size, "Inconsistency in size");
7172 assert(size >= 3, "Necessary for Printezis marks to work");
7173 if (!_bitMap->isMarked(addr+1)) {
7174 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size);
7175 } else {
7176 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1);
7177 assert(_bitMap->isMarked(addr+size-1),
7178 "inconsistent Printezis mark");
7179 }
7180 #endif // ASSERT
7181 } else {
7182 // an unitialized object
7183 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
7184 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7185 size = pointer_delta(nextOneAddr + 1, addr);
7186 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7187 "alignment problem");
7188 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
7189 // will dirty the card when the klass pointer is installed in the
7190 // object (signalling the completion of initialization).
7191 }
7192 } else {
7193 // Either a not yet marked object or an uninitialized object
7194 if (p->klass_or_null() == NULL) {
7195 // An uninitialized object, skip to the next card, since
7196 // we may not be able to read its P-bits yet.
7197 assert(size == 0, "Initial value");
7198 } else {
7199 // An object not (yet) reached by marking: we merely need to
7200 // compute its size so as to go look at the next block.
7201 assert(p->is_oop(true), "should be an oop");
7202 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
7203 }
7204 }
7205 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7206 return size;
7207 }
7208
7209 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
7210 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7211 "CMS thread should hold CMS token");
7212 assert_lock_strong(_freelistLock);
7213 assert_lock_strong(_bitMap->lock());
7214 // relinquish the free_list_lock and bitMaplock()
7215 _bitMap->lock()->unlock();
7216 _freelistLock->unlock();
7217 ConcurrentMarkSweepThread::desynchronize(true);
7218 ConcurrentMarkSweepThread::acknowledge_yield_request();
7219 _collector->stopTimer();
7220 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7221 if (PrintCMSStatistics != 0) {
7222 _collector->incrementYields();
7223 }
7224 _collector->icms_wait();
7225
7226 // See the comment in coordinator_yield()
7227 for (unsigned i = 0; i < CMSYieldSleepCount &&
7228 ConcurrentMarkSweepThread::should_yield() &&
7229 !CMSCollector::foregroundGCIsActive(); ++i) {
7230 os::sleep(Thread::current(), 1, false);
7231 ConcurrentMarkSweepThread::acknowledge_yield_request();
7232 }
7233
7234 ConcurrentMarkSweepThread::synchronize(true);
7235 _freelistLock->lock_without_safepoint_check();
7236 _bitMap->lock()->lock_without_safepoint_check();
7237 _collector->startTimer();
7238 }
7239
7240
7241 //////////////////////////////////////////////////////////////////
7242 // SurvivorSpacePrecleanClosure
7243 //////////////////////////////////////////////////////////////////
7244 // This (single-threaded) closure is used to preclean the oops in
7245 // the survivor spaces.
7246 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
7247
7248 HeapWord* addr = (HeapWord*)p;
7249 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
7250 assert(!_span.contains(addr), "we are scanning the survivor spaces");
7251 assert(p->klass_or_null() != NULL, "object should be initializd");
7252 // an initialized object; ignore mark word in verification below
7253 // since we are running concurrent with mutators
7254 assert(p->is_oop(true), "should be an oop");
7255 // Note that we do not yield while we iterate over
7256 // the interior oops of p, pushing the relevant ones
7257 // on our marking stack.
7258 size_t size = p->oop_iterate(_scanning_closure);
7259 do_yield_check();
7260 // Observe that below, we do not abandon the preclean
7261 // phase as soon as we should; rather we empty the
7262 // marking stack before returning. This is to satisfy
7263 // some existing assertions. In general, it may be a
7264 // good idea to abort immediately and complete the marking
7265 // from the grey objects at a later time.
7266 while (!_mark_stack->isEmpty()) {
7267 oop new_oop = _mark_stack->pop();
7268 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
7269 assert(_bit_map->isMarked((HeapWord*)new_oop),
7270 "only grey objects on this stack");
7271 // iterate over the oops in this oop, marking and pushing
7272 // the ones in CMS heap (i.e. in _span).
7273 new_oop->oop_iterate(_scanning_closure);
7274 // check if it's time to yield
7275 do_yield_check();
7276 }
7277 unsigned int after_count =
7278 GenCollectedHeap::heap()->total_collections();
7279 bool abort = (_before_count != after_count) ||
7280 _collector->should_abort_preclean();
7281 return abort ? 0 : size;
7282 }
7283
7284 void SurvivorSpacePrecleanClosure::do_yield_work() {
7285 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7286 "CMS thread should hold CMS token");
7287 assert_lock_strong(_bit_map->lock());
7288 // Relinquish the bit map lock
7289 _bit_map->lock()->unlock();
7290 ConcurrentMarkSweepThread::desynchronize(true);
7291 ConcurrentMarkSweepThread::acknowledge_yield_request();
7292 _collector->stopTimer();
7293 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7294 if (PrintCMSStatistics != 0) {
7295 _collector->incrementYields();
7296 }
7297 _collector->icms_wait();
7298
7299 // See the comment in coordinator_yield()
7300 for (unsigned i = 0; i < CMSYieldSleepCount &&
7301 ConcurrentMarkSweepThread::should_yield() &&
7302 !CMSCollector::foregroundGCIsActive(); ++i) {
7303 os::sleep(Thread::current(), 1, false);
7304 ConcurrentMarkSweepThread::acknowledge_yield_request();
7305 }
7306
7307 ConcurrentMarkSweepThread::synchronize(true);
7308 _bit_map->lock()->lock_without_safepoint_check();
7309 _collector->startTimer();
7310 }
7311
7312 // This closure is used to rescan the marked objects on the dirty cards
7313 // in the mod union table and the card table proper. In the parallel
7314 // case, although the bitMap is shared, we do a single read so the
7315 // isMarked() query is "safe".
7316 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
7317 // Ignore mark word because we are running concurrent with mutators
7318 assert(p->is_oop_or_null(true), "expected an oop or null");
7319 HeapWord* addr = (HeapWord*)p;
7320 assert(_span.contains(addr), "we are scanning the CMS generation");
7321 bool is_obj_array = false;
7322 #ifdef ASSERT
7323 if (!_parallel) {
7324 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
7325 assert(_collector->overflow_list_is_empty(),
7326 "overflow list should be empty");
7327
7328 }
7329 #endif // ASSERT
7330 if (_bit_map->isMarked(addr)) {
7331 // Obj arrays are precisely marked, non-arrays are not;
7332 // so we scan objArrays precisely and non-arrays in their
7333 // entirety.
7334 if (p->is_objArray()) {
7335 is_obj_array = true;
7336 if (_parallel) {
7337 p->oop_iterate(_par_scan_closure, mr);
7338 } else {
7339 p->oop_iterate(_scan_closure, mr);
7340 }
7341 } else {
7342 if (_parallel) {
7343 p->oop_iterate(_par_scan_closure);
7344 } else {
7345 p->oop_iterate(_scan_closure);
7346 }
7347 }
7348 }
7349 #ifdef ASSERT
7350 if (!_parallel) {
7351 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
7352 assert(_collector->overflow_list_is_empty(),
7353 "overflow list should be empty");
7354
7355 }
7356 #endif // ASSERT
7357 return is_obj_array;
7358 }
7359
7360 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
7361 MemRegion span,
7362 CMSBitMap* bitMap, CMSMarkStack* markStack,
7363 bool should_yield, bool verifying):
7364 _collector(collector),
7365 _span(span),
7366 _bitMap(bitMap),
7367 _mut(&collector->_modUnionTable),
7368 _markStack(markStack),
7369 _yield(should_yield),
7370 _skipBits(0)
7371 {
7372 assert(_markStack->isEmpty(), "stack should be empty");
7373 _finger = _bitMap->startWord();
7374 _threshold = _finger;
7375 assert(_collector->_restart_addr == NULL, "Sanity check");
7376 assert(_span.contains(_finger), "Out of bounds _finger?");
7377 DEBUG_ONLY(_verifying = verifying;)
7378 }
7379
7380 void MarkFromRootsClosure::reset(HeapWord* addr) {
7381 assert(_markStack->isEmpty(), "would cause duplicates on stack");
7382 assert(_span.contains(addr), "Out of bounds _finger?");
7383 _finger = addr;
7384 _threshold = (HeapWord*)round_to(
7385 (intptr_t)_finger, CardTableModRefBS::card_size);
7386 }
7387
7388 // Should revisit to see if this should be restructured for
7389 // greater efficiency.
7390 bool MarkFromRootsClosure::do_bit(size_t offset) {
7391 if (_skipBits > 0) {
7392 _skipBits--;
7393 return true;
7394 }
7395 // convert offset into a HeapWord*
7396 HeapWord* addr = _bitMap->startWord() + offset;
7397 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
7398 "address out of range");
7399 assert(_bitMap->isMarked(addr), "tautology");
7400 if (_bitMap->isMarked(addr+1)) {
7401 // this is an allocated but not yet initialized object
7402 assert(_skipBits == 0, "tautology");
7403 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
7404 oop p = oop(addr);
7405 if (p->klass_or_null() == NULL) {
7406 DEBUG_ONLY(if (!_verifying) {)
7407 // We re-dirty the cards on which this object lies and increase
7408 // the _threshold so that we'll come back to scan this object
7409 // during the preclean or remark phase. (CMSCleanOnEnter)
7410 if (CMSCleanOnEnter) {
7411 size_t sz = _collector->block_size_using_printezis_bits(addr);
7412 HeapWord* end_card_addr = (HeapWord*)round_to(
7413 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
7414 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7415 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7416 // Bump _threshold to end_card_addr; note that
7417 // _threshold cannot possibly exceed end_card_addr, anyhow.
7418 // This prevents future clearing of the card as the scan proceeds
7419 // to the right.
7420 assert(_threshold <= end_card_addr,
7421 "Because we are just scanning into this object");
7422 if (_threshold < end_card_addr) {
7423 _threshold = end_card_addr;
7424 }
7425 if (p->klass_or_null() != NULL) {
7426 // Redirty the range of cards...
7427 _mut->mark_range(redirty_range);
7428 } // ...else the setting of klass will dirty the card anyway.
7429 }
7430 DEBUG_ONLY(})
7431 return true;
7432 }
7433 }
7434 scanOopsInOop(addr);
7435 return true;
7436 }
7437
7438 // We take a break if we've been at this for a while,
7439 // so as to avoid monopolizing the locks involved.
7440 void MarkFromRootsClosure::do_yield_work() {
7441 // First give up the locks, then yield, then re-lock
7442 // We should probably use a constructor/destructor idiom to
7443 // do this unlock/lock or modify the MutexUnlocker class to
7444 // serve our purpose. XXX
7445 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7446 "CMS thread should hold CMS token");
7447 assert_lock_strong(_bitMap->lock());
7448 _bitMap->lock()->unlock();
7449 ConcurrentMarkSweepThread::desynchronize(true);
7450 ConcurrentMarkSweepThread::acknowledge_yield_request();
7451 _collector->stopTimer();
7452 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
7453 if (PrintCMSStatistics != 0) {
7454 _collector->incrementYields();
7455 }
7456 _collector->icms_wait();
7457
7458 // See the comment in coordinator_yield()
7459 for (unsigned i = 0; i < CMSYieldSleepCount &&
7460 ConcurrentMarkSweepThread::should_yield() &&
7461 !CMSCollector::foregroundGCIsActive(); ++i) {
7462 os::sleep(Thread::current(), 1, false);
7463 ConcurrentMarkSweepThread::acknowledge_yield_request();
7464 }
7465
7466 ConcurrentMarkSweepThread::synchronize(true);
7467 _bitMap->lock()->lock_without_safepoint_check();
7468 _collector->startTimer();
7469 }
7470
7471 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
7472 assert(_bitMap->isMarked(ptr), "expected bit to be set");
7473 assert(_markStack->isEmpty(),
7474 "should drain stack to limit stack usage");
7475 // convert ptr to an oop preparatory to scanning
7476 oop obj = oop(ptr);
7477 // Ignore mark word in verification below, since we
7478 // may be running concurrent with mutators.
7479 assert(obj->is_oop(true), "should be an oop");
7480 assert(_finger <= ptr, "_finger runneth ahead");
7481 // advance the finger to right end of this object
7482 _finger = ptr + obj->size();
7483 assert(_finger > ptr, "we just incremented it above");
7484 // On large heaps, it may take us some time to get through
7485 // the marking phase (especially if running iCMS). During
7486 // this time it's possible that a lot of mutations have
7487 // accumulated in the card table and the mod union table --
7488 // these mutation records are redundant until we have
7489 // actually traced into the corresponding card.
7490 // Here, we check whether advancing the finger would make
7491 // us cross into a new card, and if so clear corresponding
7492 // cards in the MUT (preclean them in the card-table in the
7493 // future).
7494
7495 DEBUG_ONLY(if (!_verifying) {)
7496 // The clean-on-enter optimization is disabled by default,
7497 // until we fix 6178663.
7498 if (CMSCleanOnEnter && (_finger > _threshold)) {
7499 // [_threshold, _finger) represents the interval
7500 // of cards to be cleared in MUT (or precleaned in card table).
7501 // The set of cards to be cleared is all those that overlap
7502 // with the interval [_threshold, _finger); note that
7503 // _threshold is always kept card-aligned but _finger isn't
7504 // always card-aligned.
7505 HeapWord* old_threshold = _threshold;
7506 assert(old_threshold == (HeapWord*)round_to(
7507 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7508 "_threshold should always be card-aligned");
7509 _threshold = (HeapWord*)round_to(
7510 (intptr_t)_finger, CardTableModRefBS::card_size);
7511 MemRegion mr(old_threshold, _threshold);
7512 assert(!mr.is_empty(), "Control point invariant");
7513 assert(_span.contains(mr), "Should clear within span");
7514 _mut->clear_range(mr);
7515 }
7516 DEBUG_ONLY(})
7517 // Note: the finger doesn't advance while we drain
7518 // the stack below.
7519 PushOrMarkClosure pushOrMarkClosure(_collector,
7520 _span, _bitMap, _markStack,
7521 _finger, this);
7522 bool res = _markStack->push(obj);
7523 assert(res, "Empty non-zero size stack should have space for single push");
7524 while (!_markStack->isEmpty()) {
7525 oop new_oop = _markStack->pop();
7526 // Skip verifying header mark word below because we are
7527 // running concurrent with mutators.
7528 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7529 // now scan this oop's oops
7530 new_oop->oop_iterate(&pushOrMarkClosure);
7531 do_yield_check();
7532 }
7533 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
7534 }
7535
7536 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task,
7537 CMSCollector* collector, MemRegion span,
7538 CMSBitMap* bit_map,
7539 OopTaskQueue* work_queue,
7540 CMSMarkStack* overflow_stack,
7541 bool should_yield):
7542 _collector(collector),
7543 _whole_span(collector->_span),
7544 _span(span),
7545 _bit_map(bit_map),
7546 _mut(&collector->_modUnionTable),
7547 _work_queue(work_queue),
7548 _overflow_stack(overflow_stack),
7549 _yield(should_yield),
7550 _skip_bits(0),
7551 _task(task)
7552 {
7553 assert(_work_queue->size() == 0, "work_queue should be empty");
7554 _finger = span.start();
7555 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
7556 assert(_span.contains(_finger), "Out of bounds _finger?");
7557 }
7558
7559 // Should revisit to see if this should be restructured for
7560 // greater efficiency.
7561 bool Par_MarkFromRootsClosure::do_bit(size_t offset) {
7562 if (_skip_bits > 0) {
7563 _skip_bits--;
7564 return true;
7565 }
7566 // convert offset into a HeapWord*
7567 HeapWord* addr = _bit_map->startWord() + offset;
7568 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
7569 "address out of range");
7570 assert(_bit_map->isMarked(addr), "tautology");
7571 if (_bit_map->isMarked(addr+1)) {
7572 // this is an allocated object that might not yet be initialized
7573 assert(_skip_bits == 0, "tautology");
7574 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
7575 oop p = oop(addr);
7576 if (p->klass_or_null() == NULL) {
7577 // in the case of Clean-on-Enter optimization, redirty card
7578 // and avoid clearing card by increasing the threshold.
7579 return true;
7580 }
7581 }
7582 scan_oops_in_oop(addr);
7583 return true;
7584 }
7585
7586 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
7587 assert(_bit_map->isMarked(ptr), "expected bit to be set");
7588 // Should we assert that our work queue is empty or
7589 // below some drain limit?
7590 assert(_work_queue->size() == 0,
7591 "should drain stack to limit stack usage");
7592 // convert ptr to an oop preparatory to scanning
7593 oop obj = oop(ptr);
7594 // Ignore mark word in verification below, since we
7595 // may be running concurrent with mutators.
7596 assert(obj->is_oop(true), "should be an oop");
7597 assert(_finger <= ptr, "_finger runneth ahead");
7598 // advance the finger to right end of this object
7599 _finger = ptr + obj->size();
7600 assert(_finger > ptr, "we just incremented it above");
7601 // On large heaps, it may take us some time to get through
7602 // the marking phase (especially if running iCMS). During
7603 // this time it's possible that a lot of mutations have
7604 // accumulated in the card table and the mod union table --
7605 // these mutation records are redundant until we have
7606 // actually traced into the corresponding card.
7607 // Here, we check whether advancing the finger would make
7608 // us cross into a new card, and if so clear corresponding
7609 // cards in the MUT (preclean them in the card-table in the
7610 // future).
7611
7612 // The clean-on-enter optimization is disabled by default,
7613 // until we fix 6178663.
7614 if (CMSCleanOnEnter && (_finger > _threshold)) {
7615 // [_threshold, _finger) represents the interval
7616 // of cards to be cleared in MUT (or precleaned in card table).
7617 // The set of cards to be cleared is all those that overlap
7618 // with the interval [_threshold, _finger); note that
7619 // _threshold is always kept card-aligned but _finger isn't
7620 // always card-aligned.
7621 HeapWord* old_threshold = _threshold;
7622 assert(old_threshold == (HeapWord*)round_to(
7623 (intptr_t)old_threshold, CardTableModRefBS::card_size),
7624 "_threshold should always be card-aligned");
7625 _threshold = (HeapWord*)round_to(
7626 (intptr_t)_finger, CardTableModRefBS::card_size);
7627 MemRegion mr(old_threshold, _threshold);
7628 assert(!mr.is_empty(), "Control point invariant");
7629 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
7630 _mut->clear_range(mr);
7631 }
7632
7633 // Note: the local finger doesn't advance while we drain
7634 // the stack below, but the global finger sure can and will.
7635 HeapWord** gfa = _task->global_finger_addr();
7636 Par_PushOrMarkClosure pushOrMarkClosure(_collector,
7637 _span, _bit_map,
7638 _work_queue,
7639 _overflow_stack,
7640 _finger,
7641 gfa, this);
7642 bool res = _work_queue->push(obj); // overflow could occur here
7643 assert(res, "Will hold once we use workqueues");
7644 while (true) {
7645 oop new_oop;
7646 if (!_work_queue->pop_local(new_oop)) {
7647 // We emptied our work_queue; check if there's stuff that can
7648 // be gotten from the overflow stack.
7649 if (CMSConcMarkingTask::get_work_from_overflow_stack(
7650 _overflow_stack, _work_queue)) {
7651 do_yield_check();
7652 continue;
7653 } else { // done
7654 break;
7655 }
7656 }
7657 // Skip verifying header mark word below because we are
7658 // running concurrent with mutators.
7659 assert(new_oop->is_oop(true), "Oops! expected to pop an oop");
7660 // now scan this oop's oops
7661 new_oop->oop_iterate(&pushOrMarkClosure);
7662 do_yield_check();
7663 }
7664 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
7665 }
7666
7667 // Yield in response to a request from VM Thread or
7668 // from mutators.
7669 void Par_MarkFromRootsClosure::do_yield_work() {
7670 assert(_task != NULL, "sanity");
7671 _task->yield();
7672 }
7673
7674 // A variant of the above used for verifying CMS marking work.
7675 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
7676 MemRegion span,
7677 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7678 CMSMarkStack* mark_stack):
7679 _collector(collector),
7680 _span(span),
7681 _verification_bm(verification_bm),
7682 _cms_bm(cms_bm),
7683 _mark_stack(mark_stack),
7684 _pam_verify_closure(collector, span, verification_bm, cms_bm,
7685 mark_stack)
7686 {
7687 assert(_mark_stack->isEmpty(), "stack should be empty");
7688 _finger = _verification_bm->startWord();
7689 assert(_collector->_restart_addr == NULL, "Sanity check");
7690 assert(_span.contains(_finger), "Out of bounds _finger?");
7691 }
7692
7693 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
7694 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
7695 assert(_span.contains(addr), "Out of bounds _finger?");
7696 _finger = addr;
7697 }
7698
7699 // Should revisit to see if this should be restructured for
7700 // greater efficiency.
7701 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
7702 // convert offset into a HeapWord*
7703 HeapWord* addr = _verification_bm->startWord() + offset;
7704 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
7705 "address out of range");
7706 assert(_verification_bm->isMarked(addr), "tautology");
7707 assert(_cms_bm->isMarked(addr), "tautology");
7708
7709 assert(_mark_stack->isEmpty(),
7710 "should drain stack to limit stack usage");
7711 // convert addr to an oop preparatory to scanning
7712 oop obj = oop(addr);
7713 assert(obj->is_oop(), "should be an oop");
7714 assert(_finger <= addr, "_finger runneth ahead");
7715 // advance the finger to right end of this object
7716 _finger = addr + obj->size();
7717 assert(_finger > addr, "we just incremented it above");
7718 // Note: the finger doesn't advance while we drain
7719 // the stack below.
7720 bool res = _mark_stack->push(obj);
7721 assert(res, "Empty non-zero size stack should have space for single push");
7722 while (!_mark_stack->isEmpty()) {
7723 oop new_oop = _mark_stack->pop();
7724 assert(new_oop->is_oop(), "Oops! expected to pop an oop");
7725 // now scan this oop's oops
7726 new_oop->oop_iterate(&_pam_verify_closure);
7727 }
7728 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
7729 return true;
7730 }
7731
7732 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
7733 CMSCollector* collector, MemRegion span,
7734 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
7735 CMSMarkStack* mark_stack):
7736 CMSOopClosure(collector->ref_processor()),
7737 _collector(collector),
7738 _span(span),
7739 _verification_bm(verification_bm),
7740 _cms_bm(cms_bm),
7741 _mark_stack(mark_stack)
7742 { }
7743
7744 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7745 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
7746
7747 // Upon stack overflow, we discard (part of) the stack,
7748 // remembering the least address amongst those discarded
7749 // in CMSCollector's _restart_address.
7750 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
7751 // Remember the least grey address discarded
7752 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
7753 _collector->lower_restart_addr(ra);
7754 _mark_stack->reset(); // discard stack contents
7755 _mark_stack->expand(); // expand the stack if possible
7756 }
7757
7758 void PushAndMarkVerifyClosure::do_oop(oop obj) {
7759 assert(obj->is_oop_or_null(), "expected an oop or NULL");
7760 HeapWord* addr = (HeapWord*)obj;
7761 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
7762 // Oop lies in _span and isn't yet grey or black
7763 _verification_bm->mark(addr); // now grey
7764 if (!_cms_bm->isMarked(addr)) {
7765 oop(addr)->print();
7766 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)",
7767 addr);
7768 fatal("... aborting");
7769 }
7770
7771 if (!_mark_stack->push(obj)) { // stack overflow
7772 if (PrintCMSStatistics != 0) {
7773 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7774 SIZE_FORMAT, _mark_stack->capacity());
7775 }
7776 assert(_mark_stack->isFull(), "Else push should have succeeded");
7777 handle_stack_overflow(addr);
7778 }
7779 // anything including and to the right of _finger
7780 // will be scanned as we iterate over the remainder of the
7781 // bit map
7782 }
7783 }
7784
7785 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
7786 MemRegion span,
7787 CMSBitMap* bitMap, CMSMarkStack* markStack,
7788 HeapWord* finger, MarkFromRootsClosure* parent) :
7789 CMSOopClosure(collector->ref_processor()),
7790 _collector(collector),
7791 _span(span),
7792 _bitMap(bitMap),
7793 _markStack(markStack),
7794 _finger(finger),
7795 _parent(parent)
7796 { }
7797
7798 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector,
7799 MemRegion span,
7800 CMSBitMap* bit_map,
7801 OopTaskQueue* work_queue,
7802 CMSMarkStack* overflow_stack,
7803 HeapWord* finger,
7804 HeapWord** global_finger_addr,
7805 Par_MarkFromRootsClosure* parent) :
7806 CMSOopClosure(collector->ref_processor()),
7807 _collector(collector),
7808 _whole_span(collector->_span),
7809 _span(span),
7810 _bit_map(bit_map),
7811 _work_queue(work_queue),
7812 _overflow_stack(overflow_stack),
7813 _finger(finger),
7814 _global_finger_addr(global_finger_addr),
7815 _parent(parent)
7816 { }
7817
7818 // Assumes thread-safe access by callers, who are
7819 // responsible for mutual exclusion.
7820 void CMSCollector::lower_restart_addr(HeapWord* low) {
7821 assert(_span.contains(low), "Out of bounds addr");
7822 if (_restart_addr == NULL) {
7823 _restart_addr = low;
7824 } else {
7825 _restart_addr = MIN2(_restart_addr, low);
7826 }
7827 }
7828
7829 // Upon stack overflow, we discard (part of) the stack,
7830 // remembering the least address amongst those discarded
7831 // in CMSCollector's _restart_address.
7832 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7833 // Remember the least grey address discarded
7834 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
7835 _collector->lower_restart_addr(ra);
7836 _markStack->reset(); // discard stack contents
7837 _markStack->expand(); // expand the stack if possible
7838 }
7839
7840 // Upon stack overflow, we discard (part of) the stack,
7841 // remembering the least address amongst those discarded
7842 // in CMSCollector's _restart_address.
7843 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
7844 // We need to do this under a mutex to prevent other
7845 // workers from interfering with the work done below.
7846 MutexLockerEx ml(_overflow_stack->par_lock(),
7847 Mutex::_no_safepoint_check_flag);
7848 // Remember the least grey address discarded
7849 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
7850 _collector->lower_restart_addr(ra);
7851 _overflow_stack->reset(); // discard stack contents
7852 _overflow_stack->expand(); // expand the stack if possible
7853 }
7854
7855 void CMKlassClosure::do_klass(Klass* k) {
7856 assert(_oop_closure != NULL, "Not initialized?");
7857 k->oops_do(_oop_closure);
7858 }
7859
7860 void PushOrMarkClosure::do_oop(oop obj) {
7861 // Ignore mark word because we are running concurrent with mutators.
7862 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7863 HeapWord* addr = (HeapWord*)obj;
7864 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
7865 // Oop lies in _span and isn't yet grey or black
7866 _bitMap->mark(addr); // now grey
7867 if (addr < _finger) {
7868 // the bit map iteration has already either passed, or
7869 // sampled, this bit in the bit map; we'll need to
7870 // use the marking stack to scan this oop's oops.
7871 bool simulate_overflow = false;
7872 NOT_PRODUCT(
7873 if (CMSMarkStackOverflowALot &&
7874 _collector->simulate_overflow()) {
7875 // simulate a stack overflow
7876 simulate_overflow = true;
7877 }
7878 )
7879 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
7880 if (PrintCMSStatistics != 0) {
7881 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7882 SIZE_FORMAT, _markStack->capacity());
7883 }
7884 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
7885 handle_stack_overflow(addr);
7886 }
7887 }
7888 // anything including and to the right of _finger
7889 // will be scanned as we iterate over the remainder of the
7890 // bit map
7891 do_yield_check();
7892 }
7893 }
7894
7895 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
7896 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
7897
7898 void Par_PushOrMarkClosure::do_oop(oop obj) {
7899 // Ignore mark word because we are running concurrent with mutators.
7900 assert(obj->is_oop_or_null(true), "expected an oop or NULL");
7901 HeapWord* addr = (HeapWord*)obj;
7902 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
7903 // Oop lies in _span and isn't yet grey or black
7904 // We read the global_finger (volatile read) strictly after marking oop
7905 bool res = _bit_map->par_mark(addr); // now grey
7906 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
7907 // Should we push this marked oop on our stack?
7908 // -- if someone else marked it, nothing to do
7909 // -- if target oop is above global finger nothing to do
7910 // -- if target oop is in chunk and above local finger
7911 // then nothing to do
7912 // -- else push on work queue
7913 if ( !res // someone else marked it, they will deal with it
7914 || (addr >= *gfa) // will be scanned in a later task
7915 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
7916 return;
7917 }
7918 // the bit map iteration has already either passed, or
7919 // sampled, this bit in the bit map; we'll need to
7920 // use the marking stack to scan this oop's oops.
7921 bool simulate_overflow = false;
7922 NOT_PRODUCT(
7923 if (CMSMarkStackOverflowALot &&
7924 _collector->simulate_overflow()) {
7925 // simulate a stack overflow
7926 simulate_overflow = true;
7927 }
7928 )
7929 if (simulate_overflow ||
7930 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
7931 // stack overflow
7932 if (PrintCMSStatistics != 0) {
7933 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at "
7934 SIZE_FORMAT, _overflow_stack->capacity());
7935 }
7936 // We cannot assert that the overflow stack is full because
7937 // it may have been emptied since.
7938 assert(simulate_overflow ||
7939 _work_queue->size() == _work_queue->max_elems(),
7940 "Else push should have succeeded");
7941 handle_stack_overflow(addr);
7942 }
7943 do_yield_check();
7944 }
7945 }
7946
7947 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7948 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); }
7949
7950 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
7951 MemRegion span,
7952 ReferenceProcessor* rp,
7953 CMSBitMap* bit_map,
7954 CMSBitMap* mod_union_table,
7955 CMSMarkStack* mark_stack,
7956 bool concurrent_precleaning):
7957 CMSOopClosure(rp),
7958 _collector(collector),
7959 _span(span),
7960 _bit_map(bit_map),
7961 _mod_union_table(mod_union_table),
7962 _mark_stack(mark_stack),
7963 _concurrent_precleaning(concurrent_precleaning)
7964 {
7965 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
7966 }
7967
7968 // Grey object rescan during pre-cleaning and second checkpoint phases --
7969 // the non-parallel version (the parallel version appears further below.)
7970 void PushAndMarkClosure::do_oop(oop obj) {
7971 // Ignore mark word verification. If during concurrent precleaning,
7972 // the object monitor may be locked. If during the checkpoint
7973 // phases, the object may already have been reached by a different
7974 // path and may be at the end of the global overflow list (so
7975 // the mark word may be NULL).
7976 assert(obj->is_oop_or_null(true /* ignore mark word */),
7977 "expected an oop or NULL");
7978 HeapWord* addr = (HeapWord*)obj;
7979 // Check if oop points into the CMS generation
7980 // and is not marked
7981 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
7982 // a white object ...
7983 _bit_map->mark(addr); // ... now grey
7984 // push on the marking stack (grey set)
7985 bool simulate_overflow = false;
7986 NOT_PRODUCT(
7987 if (CMSMarkStackOverflowALot &&
7988 _collector->simulate_overflow()) {
7989 // simulate a stack overflow
7990 simulate_overflow = true;
7991 }
7992 )
7993 if (simulate_overflow || !_mark_stack->push(obj)) {
7994 if (_concurrent_precleaning) {
7995 // During precleaning we can just dirty the appropriate card(s)
7996 // in the mod union table, thus ensuring that the object remains
7997 // in the grey set and continue. In the case of object arrays
7998 // we need to dirty all of the cards that the object spans,
7999 // since the rescan of object arrays will be limited to the
8000 // dirty cards.
8001 // Note that no one can be intefering with us in this action
8002 // of dirtying the mod union table, so no locking or atomics
8003 // are required.
8004 if (obj->is_objArray()) {
8005 size_t sz = obj->size();
8006 HeapWord* end_card_addr = (HeapWord*)round_to(
8007 (intptr_t)(addr+sz), CardTableModRefBS::card_size);
8008 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8009 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8010 _mod_union_table->mark_range(redirty_range);
8011 } else {
8012 _mod_union_table->mark(addr);
8013 }
8014 _collector->_ser_pmc_preclean_ovflw++;
8015 } else {
8016 // During the remark phase, we need to remember this oop
8017 // in the overflow list.
8018 _collector->push_on_overflow_list(obj);
8019 _collector->_ser_pmc_remark_ovflw++;
8020 }
8021 }
8022 }
8023 }
8024
8025 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector,
8026 MemRegion span,
8027 ReferenceProcessor* rp,
8028 CMSBitMap* bit_map,
8029 OopTaskQueue* work_queue):
8030 CMSOopClosure(rp),
8031 _collector(collector),
8032 _span(span),
8033 _bit_map(bit_map),
8034 _work_queue(work_queue)
8035 {
8036 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL");
8037 }
8038
8039 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
8040 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
8041
8042 // Grey object rescan during second checkpoint phase --
8043 // the parallel version.
8044 void Par_PushAndMarkClosure::do_oop(oop obj) {
8045 // In the assert below, we ignore the mark word because
8046 // this oop may point to an already visited object that is
8047 // on the overflow stack (in which case the mark word has
8048 // been hijacked for chaining into the overflow stack --
8049 // if this is the last object in the overflow stack then
8050 // its mark word will be NULL). Because this object may
8051 // have been subsequently popped off the global overflow
8052 // stack, and the mark word possibly restored to the prototypical
8053 // value, by the time we get to examined this failing assert in
8054 // the debugger, is_oop_or_null(false) may subsequently start
8055 // to hold.
8056 assert(obj->is_oop_or_null(true),
8057 "expected an oop or NULL");
8058 HeapWord* addr = (HeapWord*)obj;
8059 // Check if oop points into the CMS generation
8060 // and is not marked
8061 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
8062 // a white object ...
8063 // If we manage to "claim" the object, by being the
8064 // first thread to mark it, then we push it on our
8065 // marking stack
8066 if (_bit_map->par_mark(addr)) { // ... now grey
8067 // push on work queue (grey set)
8068 bool simulate_overflow = false;
8069 NOT_PRODUCT(
8070 if (CMSMarkStackOverflowALot &&
8071 _collector->par_simulate_overflow()) {
8072 // simulate a stack overflow
8073 simulate_overflow = true;
8074 }
8075 )
8076 if (simulate_overflow || !_work_queue->push(obj)) {
8077 _collector->par_push_on_overflow_list(obj);
8078 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
8079 }
8080 } // Else, some other thread got there first
8081 }
8082 }
8083
8084 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8085 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); }
8086
8087 void CMSPrecleanRefsYieldClosure::do_yield_work() {
8088 Mutex* bml = _collector->bitMapLock();
8089 assert_lock_strong(bml);
8090 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8091 "CMS thread should hold CMS token");
8092
8093 bml->unlock();
8094 ConcurrentMarkSweepThread::desynchronize(true);
8095
8096 ConcurrentMarkSweepThread::acknowledge_yield_request();
8097
8098 _collector->stopTimer();
8099 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8100 if (PrintCMSStatistics != 0) {
8101 _collector->incrementYields();
8102 }
8103 _collector->icms_wait();
8104
8105 // See the comment in coordinator_yield()
8106 for (unsigned i = 0; i < CMSYieldSleepCount &&
8107 ConcurrentMarkSweepThread::should_yield() &&
8108 !CMSCollector::foregroundGCIsActive(); ++i) {
8109 os::sleep(Thread::current(), 1, false);
8110 ConcurrentMarkSweepThread::acknowledge_yield_request();
8111 }
8112
8113 ConcurrentMarkSweepThread::synchronize(true);
8114 bml->lock();
8115
8116 _collector->startTimer();
8117 }
8118
8119 bool CMSPrecleanRefsYieldClosure::should_return() {
8120 if (ConcurrentMarkSweepThread::should_yield()) {
8121 do_yield_work();
8122 }
8123 return _collector->foregroundGCIsActive();
8124 }
8125
8126 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
8127 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0,
8128 "mr should be aligned to start at a card boundary");
8129 // We'd like to assert:
8130 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0,
8131 // "mr should be a range of cards");
8132 // However, that would be too strong in one case -- the last
8133 // partition ends at _unallocated_block which, in general, can be
8134 // an arbitrary boundary, not necessarily card aligned.
8135 if (PrintCMSStatistics != 0) {
8136 _num_dirty_cards +=
8137 mr.word_size()/CardTableModRefBS::card_size_in_words;
8138 }
8139 _space->object_iterate_mem(mr, &_scan_cl);
8140 }
8141
8142 SweepClosure::SweepClosure(CMSCollector* collector,
8143 ConcurrentMarkSweepGeneration* g,
8144 CMSBitMap* bitMap, bool should_yield) :
8145 _collector(collector),
8146 _g(g),
8147 _sp(g->cmsSpace()),
8148 _limit(_sp->sweep_limit()),
8149 _freelistLock(_sp->freelistLock()),
8150 _bitMap(bitMap),
8151 _yield(should_yield),
8152 _inFreeRange(false), // No free range at beginning of sweep
8153 _freeRangeInFreeLists(false), // No free range at beginning of sweep
8154 _lastFreeRangeCoalesced(false),
8155 _freeFinger(g->used_region().start())
8156 {
8157 NOT_PRODUCT(
8158 _numObjectsFreed = 0;
8159 _numWordsFreed = 0;
8160 _numObjectsLive = 0;
8161 _numWordsLive = 0;
8162 _numObjectsAlreadyFree = 0;
8163 _numWordsAlreadyFree = 0;
8164 _last_fc = NULL;
8165
8166 _sp->initializeIndexedFreeListArrayReturnedBytes();
8167 _sp->dictionary()->initialize_dict_returned_bytes();
8168 )
8169 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8170 "sweep _limit out of bounds");
8171 if (CMSTraceSweeper) {
8172 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT,
8173 _limit);
8174 }
8175 }
8176
8177 void SweepClosure::print_on(outputStream* st) const {
8178 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
8179 _sp->bottom(), _sp->end());
8180 tty->print_cr("_limit = " PTR_FORMAT, _limit);
8181 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger);
8182 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);)
8183 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
8184 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
8185 }
8186
8187 #ifndef PRODUCT
8188 // Assertion checking only: no useful work in product mode --
8189 // however, if any of the flags below become product flags,
8190 // you may need to review this code to see if it needs to be
8191 // enabled in product mode.
8192 SweepClosure::~SweepClosure() {
8193 assert_lock_strong(_freelistLock);
8194 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8195 "sweep _limit out of bounds");
8196 if (inFreeRange()) {
8197 warning("inFreeRange() should have been reset; dumping state of SweepClosure");
8198 print();
8199 ShouldNotReachHere();
8200 }
8201 if (Verbose && PrintGC) {
8202 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes",
8203 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
8204 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, "
8205 SIZE_FORMAT" bytes "
8206 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes",
8207 _numObjectsLive, _numWordsLive*sizeof(HeapWord),
8208 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
8209 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree)
8210 * sizeof(HeapWord);
8211 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes);
8212
8213 if (PrintCMSStatistics && CMSVerifyReturnedBytes) {
8214 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
8215 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
8216 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
8217 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes);
8218 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes",
8219 indexListReturnedBytes);
8220 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes",
8221 dict_returned_bytes);
8222 }
8223 }
8224 if (CMSTraceSweeper) {
8225 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================",
8226 _limit);
8227 }
8228 }
8229 #endif // PRODUCT
8230
8231 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
8232 bool freeRangeInFreeLists) {
8233 if (CMSTraceSweeper) {
8234 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n",
8235 freeFinger, freeRangeInFreeLists);
8236 }
8237 assert(!inFreeRange(), "Trampling existing free range");
8238 set_inFreeRange(true);
8239 set_lastFreeRangeCoalesced(false);
8240
8241 set_freeFinger(freeFinger);
8242 set_freeRangeInFreeLists(freeRangeInFreeLists);
8243 if (CMSTestInFreeList) {
8244 if (freeRangeInFreeLists) {
8245 FreeChunk* fc = (FreeChunk*) freeFinger;
8246 assert(fc->is_free(), "A chunk on the free list should be free.");
8247 assert(fc->size() > 0, "Free range should have a size");
8248 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
8249 }
8250 }
8251 }
8252
8253 // Note that the sweeper runs concurrently with mutators. Thus,
8254 // it is possible for direct allocation in this generation to happen
8255 // in the middle of the sweep. Note that the sweeper also coalesces
8256 // contiguous free blocks. Thus, unless the sweeper and the allocator
8257 // synchronize appropriately freshly allocated blocks may get swept up.
8258 // This is accomplished by the sweeper locking the free lists while
8259 // it is sweeping. Thus blocks that are determined to be free are
8260 // indeed free. There is however one additional complication:
8261 // blocks that have been allocated since the final checkpoint and
8262 // mark, will not have been marked and so would be treated as
8263 // unreachable and swept up. To prevent this, the allocator marks
8264 // the bit map when allocating during the sweep phase. This leads,
8265 // however, to a further complication -- objects may have been allocated
8266 // but not yet initialized -- in the sense that the header isn't yet
8267 // installed. The sweeper can not then determine the size of the block
8268 // in order to skip over it. To deal with this case, we use a technique
8269 // (due to Printezis) to encode such uninitialized block sizes in the
8270 // bit map. Since the bit map uses a bit per every HeapWord, but the
8271 // CMS generation has a minimum object size of 3 HeapWords, it follows
8272 // that "normal marks" won't be adjacent in the bit map (there will
8273 // always be at least two 0 bits between successive 1 bits). We make use
8274 // of these "unused" bits to represent uninitialized blocks -- the bit
8275 // corresponding to the start of the uninitialized object and the next
8276 // bit are both set. Finally, a 1 bit marks the end of the object that
8277 // started with the two consecutive 1 bits to indicate its potentially
8278 // uninitialized state.
8279
8280 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
8281 FreeChunk* fc = (FreeChunk*)addr;
8282 size_t res;
8283
8284 // Check if we are done sweeping. Below we check "addr >= _limit" rather
8285 // than "addr == _limit" because although _limit was a block boundary when
8286 // we started the sweep, it may no longer be one because heap expansion
8287 // may have caused us to coalesce the block ending at the address _limit
8288 // with a newly expanded chunk (this happens when _limit was set to the
8289 // previous _end of the space), so we may have stepped past _limit:
8290 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
8291 if (addr >= _limit) { // we have swept up to or past the limit: finish up
8292 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
8293 "sweep _limit out of bounds");
8294 assert(addr < _sp->end(), "addr out of bounds");
8295 // Flush any free range we might be holding as a single
8296 // coalesced chunk to the appropriate free list.
8297 if (inFreeRange()) {
8298 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
8299 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger()));
8300 flush_cur_free_chunk(freeFinger(),
8301 pointer_delta(addr, freeFinger()));
8302 if (CMSTraceSweeper) {
8303 gclog_or_tty->print("Sweep: last chunk: ");
8304 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") "
8305 "[coalesced:"SIZE_FORMAT"]\n",
8306 freeFinger(), pointer_delta(addr, freeFinger()),
8307 lastFreeRangeCoalesced());
8308 }
8309 }
8310
8311 // help the iterator loop finish
8312 return pointer_delta(_sp->end(), addr);
8313 }
8314
8315 assert(addr < _limit, "sweep invariant");
8316 // check if we should yield
8317 do_yield_check(addr);
8318 if (fc->is_free()) {
8319 // Chunk that is already free
8320 res = fc->size();
8321 do_already_free_chunk(fc);
8322 debug_only(_sp->verifyFreeLists());
8323 // If we flush the chunk at hand in lookahead_and_flush()
8324 // and it's coalesced with a preceding chunk, then the
8325 // process of "mangling" the payload of the coalesced block
8326 // will cause erasure of the size information from the
8327 // (erstwhile) header of all the coalesced blocks but the
8328 // first, so the first disjunct in the assert will not hold
8329 // in that specific case (in which case the second disjunct
8330 // will hold).
8331 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
8332 "Otherwise the size info doesn't change at this step");
8333 NOT_PRODUCT(
8334 _numObjectsAlreadyFree++;
8335 _numWordsAlreadyFree += res;
8336 )
8337 NOT_PRODUCT(_last_fc = fc;)
8338 } else if (!_bitMap->isMarked(addr)) {
8339 // Chunk is fresh garbage
8340 res = do_garbage_chunk(fc);
8341 debug_only(_sp->verifyFreeLists());
8342 NOT_PRODUCT(
8343 _numObjectsFreed++;
8344 _numWordsFreed += res;
8345 )
8346 } else {
8347 // Chunk that is alive.
8348 res = do_live_chunk(fc);
8349 debug_only(_sp->verifyFreeLists());
8350 NOT_PRODUCT(
8351 _numObjectsLive++;
8352 _numWordsLive += res;
8353 )
8354 }
8355 return res;
8356 }
8357
8358 // For the smart allocation, record following
8359 // split deaths - a free chunk is removed from its free list because
8360 // it is being split into two or more chunks.
8361 // split birth - a free chunk is being added to its free list because
8362 // a larger free chunk has been split and resulted in this free chunk.
8363 // coal death - a free chunk is being removed from its free list because
8364 // it is being coalesced into a large free chunk.
8365 // coal birth - a free chunk is being added to its free list because
8366 // it was created when two or more free chunks where coalesced into
8367 // this free chunk.
8368 //
8369 // These statistics are used to determine the desired number of free
8370 // chunks of a given size. The desired number is chosen to be relative
8371 // to the end of a CMS sweep. The desired number at the end of a sweep
8372 // is the
8373 // count-at-end-of-previous-sweep (an amount that was enough)
8374 // - count-at-beginning-of-current-sweep (the excess)
8375 // + split-births (gains in this size during interval)
8376 // - split-deaths (demands on this size during interval)
8377 // where the interval is from the end of one sweep to the end of the
8378 // next.
8379 //
8380 // When sweeping the sweeper maintains an accumulated chunk which is
8381 // the chunk that is made up of chunks that have been coalesced. That
8382 // will be termed the left-hand chunk. A new chunk of garbage that
8383 // is being considered for coalescing will be referred to as the
8384 // right-hand chunk.
8385 //
8386 // When making a decision on whether to coalesce a right-hand chunk with
8387 // the current left-hand chunk, the current count vs. the desired count
8388 // of the left-hand chunk is considered. Also if the right-hand chunk
8389 // is near the large chunk at the end of the heap (see
8390 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
8391 // left-hand chunk is coalesced.
8392 //
8393 // When making a decision about whether to split a chunk, the desired count
8394 // vs. the current count of the candidate to be split is also considered.
8395 // If the candidate is underpopulated (currently fewer chunks than desired)
8396 // a chunk of an overpopulated (currently more chunks than desired) size may
8397 // be chosen. The "hint" associated with a free list, if non-null, points
8398 // to a free list which may be overpopulated.
8399 //
8400
8401 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
8402 const size_t size = fc->size();
8403 // Chunks that cannot be coalesced are not in the
8404 // free lists.
8405 if (CMSTestInFreeList && !fc->cantCoalesce()) {
8406 assert(_sp->verify_chunk_in_free_list(fc),
8407 "free chunk should be in free lists");
8408 }
8409 // a chunk that is already free, should not have been
8410 // marked in the bit map
8411 HeapWord* const addr = (HeapWord*) fc;
8412 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
8413 // Verify that the bit map has no bits marked between
8414 // addr and purported end of this block.
8415 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8416
8417 // Some chunks cannot be coalesced under any circumstances.
8418 // See the definition of cantCoalesce().
8419 if (!fc->cantCoalesce()) {
8420 // This chunk can potentially be coalesced.
8421 if (_sp->adaptive_freelists()) {
8422 // All the work is done in
8423 do_post_free_or_garbage_chunk(fc, size);
8424 } else { // Not adaptive free lists
8425 // this is a free chunk that can potentially be coalesced by the sweeper;
8426 if (!inFreeRange()) {
8427 // if the next chunk is a free block that can't be coalesced
8428 // it doesn't make sense to remove this chunk from the free lists
8429 FreeChunk* nextChunk = (FreeChunk*)(addr + size);
8430 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?");
8431 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ...
8432 nextChunk->is_free() && // ... which is free...
8433 nextChunk->cantCoalesce()) { // ... but can't be coalesced
8434 // nothing to do
8435 } else {
8436 // Potentially the start of a new free range:
8437 // Don't eagerly remove it from the free lists.
8438 // No need to remove it if it will just be put
8439 // back again. (Also from a pragmatic point of view
8440 // if it is a free block in a region that is beyond
8441 // any allocated blocks, an assertion will fail)
8442 // Remember the start of a free run.
8443 initialize_free_range(addr, true);
8444 // end - can coalesce with next chunk
8445 }
8446 } else {
8447 // the midst of a free range, we are coalescing
8448 print_free_block_coalesced(fc);
8449 if (CMSTraceSweeper) {
8450 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size);
8451 }
8452 // remove it from the free lists
8453 _sp->removeFreeChunkFromFreeLists(fc);
8454 set_lastFreeRangeCoalesced(true);
8455 // If the chunk is being coalesced and the current free range is
8456 // in the free lists, remove the current free range so that it
8457 // will be returned to the free lists in its entirety - all
8458 // the coalesced pieces included.
8459 if (freeRangeInFreeLists()) {
8460 FreeChunk* ffc = (FreeChunk*) freeFinger();
8461 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8462 "Size of free range is inconsistent with chunk size.");
8463 if (CMSTestInFreeList) {
8464 assert(_sp->verify_chunk_in_free_list(ffc),
8465 "free range is not in free lists");
8466 }
8467 _sp->removeFreeChunkFromFreeLists(ffc);
8468 set_freeRangeInFreeLists(false);
8469 }
8470 }
8471 }
8472 // Note that if the chunk is not coalescable (the else arm
8473 // below), we unconditionally flush, without needing to do
8474 // a "lookahead," as we do below.
8475 if (inFreeRange()) lookahead_and_flush(fc, size);
8476 } else {
8477 // Code path common to both original and adaptive free lists.
8478
8479 // cant coalesce with previous block; this should be treated
8480 // as the end of a free run if any
8481 if (inFreeRange()) {
8482 // we kicked some butt; time to pick up the garbage
8483 assert(freeFinger() < addr, "freeFinger points too high");
8484 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8485 }
8486 // else, nothing to do, just continue
8487 }
8488 }
8489
8490 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
8491 // This is a chunk of garbage. It is not in any free list.
8492 // Add it to a free list or let it possibly be coalesced into
8493 // a larger chunk.
8494 HeapWord* const addr = (HeapWord*) fc;
8495 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8496
8497 if (_sp->adaptive_freelists()) {
8498 // Verify that the bit map has no bits marked between
8499 // addr and purported end of just dead object.
8500 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8501
8502 do_post_free_or_garbage_chunk(fc, size);
8503 } else {
8504 if (!inFreeRange()) {
8505 // start of a new free range
8506 assert(size > 0, "A free range should have a size");
8507 initialize_free_range(addr, false);
8508 } else {
8509 // this will be swept up when we hit the end of the
8510 // free range
8511 if (CMSTraceSweeper) {
8512 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size);
8513 }
8514 // If the chunk is being coalesced and the current free range is
8515 // in the free lists, remove the current free range so that it
8516 // will be returned to the free lists in its entirety - all
8517 // the coalesced pieces included.
8518 if (freeRangeInFreeLists()) {
8519 FreeChunk* ffc = (FreeChunk*)freeFinger();
8520 assert(ffc->size() == pointer_delta(addr, freeFinger()),
8521 "Size of free range is inconsistent with chunk size.");
8522 if (CMSTestInFreeList) {
8523 assert(_sp->verify_chunk_in_free_list(ffc),
8524 "free range is not in free lists");
8525 }
8526 _sp->removeFreeChunkFromFreeLists(ffc);
8527 set_freeRangeInFreeLists(false);
8528 }
8529 set_lastFreeRangeCoalesced(true);
8530 }
8531 // this will be swept up when we hit the end of the free range
8532
8533 // Verify that the bit map has no bits marked between
8534 // addr and purported end of just dead object.
8535 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
8536 }
8537 assert(_limit >= addr + size,
8538 "A freshly garbage chunk can't possibly straddle over _limit");
8539 if (inFreeRange()) lookahead_and_flush(fc, size);
8540 return size;
8541 }
8542
8543 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
8544 HeapWord* addr = (HeapWord*) fc;
8545 // The sweeper has just found a live object. Return any accumulated
8546 // left hand chunk to the free lists.
8547 if (inFreeRange()) {
8548 assert(freeFinger() < addr, "freeFinger points too high");
8549 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8550 }
8551
8552 // This object is live: we'd normally expect this to be
8553 // an oop, and like to assert the following:
8554 // assert(oop(addr)->is_oop(), "live block should be an oop");
8555 // However, as we commented above, this may be an object whose
8556 // header hasn't yet been initialized.
8557 size_t size;
8558 assert(_bitMap->isMarked(addr), "Tautology for this control point");
8559 if (_bitMap->isMarked(addr + 1)) {
8560 // Determine the size from the bit map, rather than trying to
8561 // compute it from the object header.
8562 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
8563 size = pointer_delta(nextOneAddr + 1, addr);
8564 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
8565 "alignment problem");
8566
8567 #ifdef ASSERT
8568 if (oop(addr)->klass_or_null() != NULL) {
8569 // Ignore mark word because we are running concurrent with mutators
8570 assert(oop(addr)->is_oop(true), "live block should be an oop");
8571 assert(size ==
8572 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
8573 "P-mark and computed size do not agree");
8574 }
8575 #endif
8576
8577 } else {
8578 // This should be an initialized object that's alive.
8579 assert(oop(addr)->klass_or_null() != NULL,
8580 "Should be an initialized object");
8581 // Ignore mark word because we are running concurrent with mutators
8582 assert(oop(addr)->is_oop(true), "live block should be an oop");
8583 // Verify that the bit map has no bits marked between
8584 // addr and purported end of this block.
8585 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
8586 assert(size >= 3, "Necessary for Printezis marks to work");
8587 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
8588 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
8589 }
8590 return size;
8591 }
8592
8593 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
8594 size_t chunkSize) {
8595 // do_post_free_or_garbage_chunk() should only be called in the case
8596 // of the adaptive free list allocator.
8597 const bool fcInFreeLists = fc->is_free();
8598 assert(_sp->adaptive_freelists(), "Should only be used in this case.");
8599 assert((HeapWord*)fc <= _limit, "sweep invariant");
8600 if (CMSTestInFreeList && fcInFreeLists) {
8601 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
8602 }
8603
8604 if (CMSTraceSweeper) {
8605 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize);
8606 }
8607
8608 HeapWord* const fc_addr = (HeapWord*) fc;
8609
8610 bool coalesce;
8611 const size_t left = pointer_delta(fc_addr, freeFinger());
8612 const size_t right = chunkSize;
8613 switch (FLSCoalescePolicy) {
8614 // numeric value forms a coalition aggressiveness metric
8615 case 0: { // never coalesce
8616 coalesce = false;
8617 break;
8618 }
8619 case 1: { // coalesce if left & right chunks on overpopulated lists
8620 coalesce = _sp->coalOverPopulated(left) &&
8621 _sp->coalOverPopulated(right);
8622 break;
8623 }
8624 case 2: { // coalesce if left chunk on overpopulated list (default)
8625 coalesce = _sp->coalOverPopulated(left);
8626 break;
8627 }
8628 case 3: { // coalesce if left OR right chunk on overpopulated list
8629 coalesce = _sp->coalOverPopulated(left) ||
8630 _sp->coalOverPopulated(right);
8631 break;
8632 }
8633 case 4: { // always coalesce
8634 coalesce = true;
8635 break;
8636 }
8637 default:
8638 ShouldNotReachHere();
8639 }
8640
8641 // Should the current free range be coalesced?
8642 // If the chunk is in a free range and either we decided to coalesce above
8643 // or the chunk is near the large block at the end of the heap
8644 // (isNearLargestChunk() returns true), then coalesce this chunk.
8645 const bool doCoalesce = inFreeRange()
8646 && (coalesce || _g->isNearLargestChunk(fc_addr));
8647 if (doCoalesce) {
8648 // Coalesce the current free range on the left with the new
8649 // chunk on the right. If either is on a free list,
8650 // it must be removed from the list and stashed in the closure.
8651 if (freeRangeInFreeLists()) {
8652 FreeChunk* const ffc = (FreeChunk*)freeFinger();
8653 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
8654 "Size of free range is inconsistent with chunk size.");
8655 if (CMSTestInFreeList) {
8656 assert(_sp->verify_chunk_in_free_list(ffc),
8657 "Chunk is not in free lists");
8658 }
8659 _sp->coalDeath(ffc->size());
8660 _sp->removeFreeChunkFromFreeLists(ffc);
8661 set_freeRangeInFreeLists(false);
8662 }
8663 if (fcInFreeLists) {
8664 _sp->coalDeath(chunkSize);
8665 assert(fc->size() == chunkSize,
8666 "The chunk has the wrong size or is not in the free lists");
8667 _sp->removeFreeChunkFromFreeLists(fc);
8668 }
8669 set_lastFreeRangeCoalesced(true);
8670 print_free_block_coalesced(fc);
8671 } else { // not in a free range and/or should not coalesce
8672 // Return the current free range and start a new one.
8673 if (inFreeRange()) {
8674 // In a free range but cannot coalesce with the right hand chunk.
8675 // Put the current free range into the free lists.
8676 flush_cur_free_chunk(freeFinger(),
8677 pointer_delta(fc_addr, freeFinger()));
8678 }
8679 // Set up for new free range. Pass along whether the right hand
8680 // chunk is in the free lists.
8681 initialize_free_range((HeapWord*)fc, fcInFreeLists);
8682 }
8683 }
8684
8685 // Lookahead flush:
8686 // If we are tracking a free range, and this is the last chunk that
8687 // we'll look at because its end crosses past _limit, we'll preemptively
8688 // flush it along with any free range we may be holding on to. Note that
8689 // this can be the case only for an already free or freshly garbage
8690 // chunk. If this block is an object, it can never straddle
8691 // over _limit. The "straddling" occurs when _limit is set at
8692 // the previous end of the space when this cycle started, and
8693 // a subsequent heap expansion caused the previously co-terminal
8694 // free block to be coalesced with the newly expanded portion,
8695 // thus rendering _limit a non-block-boundary making it dangerous
8696 // for the sweeper to step over and examine.
8697 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
8698 assert(inFreeRange(), "Should only be called if currently in a free range.");
8699 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
8700 assert(_sp->used_region().contains(eob - 1),
8701 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
8702 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
8703 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
8704 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size));
8705 if (eob >= _limit) {
8706 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
8707 if (CMSTraceSweeper) {
8708 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block "
8709 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
8710 "[" PTR_FORMAT "," PTR_FORMAT ")",
8711 _limit, fc, eob, _sp->bottom(), _sp->end());
8712 }
8713 // Return the storage we are tracking back into the free lists.
8714 if (CMSTraceSweeper) {
8715 gclog_or_tty->print_cr("Flushing ... ");
8716 }
8717 assert(freeFinger() < eob, "Error");
8718 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
8719 }
8720 }
8721
8722 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
8723 assert(inFreeRange(), "Should only be called if currently in a free range.");
8724 assert(size > 0,
8725 "A zero sized chunk cannot be added to the free lists.");
8726 if (!freeRangeInFreeLists()) {
8727 if (CMSTestInFreeList) {
8728 FreeChunk* fc = (FreeChunk*) chunk;
8729 fc->set_size(size);
8730 assert(!_sp->verify_chunk_in_free_list(fc),
8731 "chunk should not be in free lists yet");
8732 }
8733 if (CMSTraceSweeper) {
8734 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists",
8735 chunk, size);
8736 }
8737 // A new free range is going to be starting. The current
8738 // free range has not been added to the free lists yet or
8739 // was removed so add it back.
8740 // If the current free range was coalesced, then the death
8741 // of the free range was recorded. Record a birth now.
8742 if (lastFreeRangeCoalesced()) {
8743 _sp->coalBirth(size);
8744 }
8745 _sp->addChunkAndRepairOffsetTable(chunk, size,
8746 lastFreeRangeCoalesced());
8747 } else if (CMSTraceSweeper) {
8748 gclog_or_tty->print_cr("Already in free list: nothing to flush");
8749 }
8750 set_inFreeRange(false);
8751 set_freeRangeInFreeLists(false);
8752 }
8753
8754 // We take a break if we've been at this for a while,
8755 // so as to avoid monopolizing the locks involved.
8756 void SweepClosure::do_yield_work(HeapWord* addr) {
8757 // Return current free chunk being used for coalescing (if any)
8758 // to the appropriate freelist. After yielding, the next
8759 // free block encountered will start a coalescing range of
8760 // free blocks. If the next free block is adjacent to the
8761 // chunk just flushed, they will need to wait for the next
8762 // sweep to be coalesced.
8763 if (inFreeRange()) {
8764 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
8765 }
8766
8767 // First give up the locks, then yield, then re-lock.
8768 // We should probably use a constructor/destructor idiom to
8769 // do this unlock/lock or modify the MutexUnlocker class to
8770 // serve our purpose. XXX
8771 assert_lock_strong(_bitMap->lock());
8772 assert_lock_strong(_freelistLock);
8773 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
8774 "CMS thread should hold CMS token");
8775 _bitMap->lock()->unlock();
8776 _freelistLock->unlock();
8777 ConcurrentMarkSweepThread::desynchronize(true);
8778 ConcurrentMarkSweepThread::acknowledge_yield_request();
8779 _collector->stopTimer();
8780 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr());
8781 if (PrintCMSStatistics != 0) {
8782 _collector->incrementYields();
8783 }
8784 _collector->icms_wait();
8785
8786 // See the comment in coordinator_yield()
8787 for (unsigned i = 0; i < CMSYieldSleepCount &&
8788 ConcurrentMarkSweepThread::should_yield() &&
8789 !CMSCollector::foregroundGCIsActive(); ++i) {
8790 os::sleep(Thread::current(), 1, false);
8791 ConcurrentMarkSweepThread::acknowledge_yield_request();
8792 }
8793
8794 ConcurrentMarkSweepThread::synchronize(true);
8795 _freelistLock->lock();
8796 _bitMap->lock()->lock_without_safepoint_check();
8797 _collector->startTimer();
8798 }
8799
8800 #ifndef PRODUCT
8801 // This is actually very useful in a product build if it can
8802 // be called from the debugger. Compile it into the product
8803 // as needed.
8804 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
8805 return debug_cms_space->verify_chunk_in_free_list(fc);
8806 }
8807 #endif
8808
8809 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
8810 if (CMSTraceSweeper) {
8811 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
8812 fc, fc->size());
8813 }
8814 }
8815
8816 // CMSIsAliveClosure
8817 bool CMSIsAliveClosure::do_object_b(oop obj) {
8818 HeapWord* addr = (HeapWord*)obj;
8819 return addr != NULL &&
8820 (!_span.contains(addr) || _bit_map->isMarked(addr));
8821 }
8822
8823
8824 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
8825 MemRegion span,
8826 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
8827 bool cpc):
8828 _collector(collector),
8829 _span(span),
8830 _bit_map(bit_map),
8831 _mark_stack(mark_stack),
8832 _concurrent_precleaning(cpc) {
8833 assert(!_span.is_empty(), "Empty span could spell trouble");
8834 }
8835
8836
8837 // CMSKeepAliveClosure: the serial version
8838 void CMSKeepAliveClosure::do_oop(oop obj) {
8839 HeapWord* addr = (HeapWord*)obj;
8840 if (_span.contains(addr) &&
8841 !_bit_map->isMarked(addr)) {
8842 _bit_map->mark(addr);
8843 bool simulate_overflow = false;
8844 NOT_PRODUCT(
8845 if (CMSMarkStackOverflowALot &&
8846 _collector->simulate_overflow()) {
8847 // simulate a stack overflow
8848 simulate_overflow = true;
8849 }
8850 )
8851 if (simulate_overflow || !_mark_stack->push(obj)) {
8852 if (_concurrent_precleaning) {
8853 // We dirty the overflown object and let the remark
8854 // phase deal with it.
8855 assert(_collector->overflow_list_is_empty(), "Error");
8856 // In the case of object arrays, we need to dirty all of
8857 // the cards that the object spans. No locking or atomics
8858 // are needed since no one else can be mutating the mod union
8859 // table.
8860 if (obj->is_objArray()) {
8861 size_t sz = obj->size();
8862 HeapWord* end_card_addr =
8863 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size);
8864 MemRegion redirty_range = MemRegion(addr, end_card_addr);
8865 assert(!redirty_range.is_empty(), "Arithmetical tautology");
8866 _collector->_modUnionTable.mark_range(redirty_range);
8867 } else {
8868 _collector->_modUnionTable.mark(addr);
8869 }
8870 _collector->_ser_kac_preclean_ovflw++;
8871 } else {
8872 _collector->push_on_overflow_list(obj);
8873 _collector->_ser_kac_ovflw++;
8874 }
8875 }
8876 }
8877 }
8878
8879 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8880 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
8881
8882 // CMSParKeepAliveClosure: a parallel version of the above.
8883 // The work queues are private to each closure (thread),
8884 // but (may be) available for stealing by other threads.
8885 void CMSParKeepAliveClosure::do_oop(oop obj) {
8886 HeapWord* addr = (HeapWord*)obj;
8887 if (_span.contains(addr) &&
8888 !_bit_map->isMarked(addr)) {
8889 // In general, during recursive tracing, several threads
8890 // may be concurrently getting here; the first one to
8891 // "tag" it, claims it.
8892 if (_bit_map->par_mark(addr)) {
8893 bool res = _work_queue->push(obj);
8894 assert(res, "Low water mark should be much less than capacity");
8895 // Do a recursive trim in the hope that this will keep
8896 // stack usage lower, but leave some oops for potential stealers
8897 trim_queue(_low_water_mark);
8898 } // Else, another thread got there first
8899 }
8900 }
8901
8902 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8903 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
8904
8905 void CMSParKeepAliveClosure::trim_queue(uint max) {
8906 while (_work_queue->size() > max) {
8907 oop new_oop;
8908 if (_work_queue->pop_local(new_oop)) {
8909 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop");
8910 assert(_bit_map->isMarked((HeapWord*)new_oop),
8911 "no white objects on this stack!");
8912 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
8913 // iterate over the oops in this oop, marking and pushing
8914 // the ones in CMS heap (i.e. in _span).
8915 new_oop->oop_iterate(&_mark_and_push);
8916 }
8917 }
8918 }
8919
8920 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
8921 CMSCollector* collector,
8922 MemRegion span, CMSBitMap* bit_map,
8923 OopTaskQueue* work_queue):
8924 _collector(collector),
8925 _span(span),
8926 _bit_map(bit_map),
8927 _work_queue(work_queue) { }
8928
8929 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
8930 HeapWord* addr = (HeapWord*)obj;
8931 if (_span.contains(addr) &&
8932 !_bit_map->isMarked(addr)) {
8933 if (_bit_map->par_mark(addr)) {
8934 bool simulate_overflow = false;
8935 NOT_PRODUCT(
8936 if (CMSMarkStackOverflowALot &&
8937 _collector->par_simulate_overflow()) {
8938 // simulate a stack overflow
8939 simulate_overflow = true;
8940 }
8941 )
8942 if (simulate_overflow || !_work_queue->push(obj)) {
8943 _collector->par_push_on_overflow_list(obj);
8944 _collector->_par_kac_ovflw++;
8945 }
8946 } // Else another thread got there already
8947 }
8948 }
8949
8950 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8951 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
8952
8953 //////////////////////////////////////////////////////////////////
8954 // CMSExpansionCause /////////////////////////////
8955 //////////////////////////////////////////////////////////////////
8956 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
8957 switch (cause) {
8958 case _no_expansion:
8959 return "No expansion";
8960 case _satisfy_free_ratio:
8961 return "Free ratio";
8962 case _satisfy_promotion:
8963 return "Satisfy promotion";
8964 case _satisfy_allocation:
8965 return "allocation";
8966 case _allocate_par_lab:
8967 return "Par LAB";
8968 case _allocate_par_spooling_space:
8969 return "Par Spooling Space";
8970 case _adaptive_size_policy:
8971 return "Ergonomics";
8972 default:
8973 return "unknown";
8974 }
8975 }
8976
8977 void CMSDrainMarkingStackClosure::do_void() {
8978 // the max number to take from overflow list at a time
8979 const size_t num = _mark_stack->capacity()/4;
8980 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
8981 "Overflow list should be NULL during concurrent phases");
8982 while (!_mark_stack->isEmpty() ||
8983 // if stack is empty, check the overflow list
8984 _collector->take_from_overflow_list(num, _mark_stack)) {
8985 oop obj = _mark_stack->pop();
8986 HeapWord* addr = (HeapWord*)obj;
8987 assert(_span.contains(addr), "Should be within span");
8988 assert(_bit_map->isMarked(addr), "Should be marked");
8989 assert(obj->is_oop(), "Should be an oop");
8990 obj->oop_iterate(_keep_alive);
8991 }
8992 }
8993
8994 void CMSParDrainMarkingStackClosure::do_void() {
8995 // drain queue
8996 trim_queue(0);
8997 }
8998
8999 // Trim our work_queue so its length is below max at return
9000 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
9001 while (_work_queue->size() > max) {
9002 oop new_oop;
9003 if (_work_queue->pop_local(new_oop)) {
9004 assert(new_oop->is_oop(), "Expected an oop");
9005 assert(_bit_map->isMarked((HeapWord*)new_oop),
9006 "no white objects on this stack!");
9007 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
9008 // iterate over the oops in this oop, marking and pushing
9009 // the ones in CMS heap (i.e. in _span).
9010 new_oop->oop_iterate(&_mark_and_push);
9011 }
9012 }
9013 }
9014
9015 ////////////////////////////////////////////////////////////////////
9016 // Support for Marking Stack Overflow list handling and related code
9017 ////////////////////////////////////////////////////////////////////
9018 // Much of the following code is similar in shape and spirit to the
9019 // code used in ParNewGC. We should try and share that code
9020 // as much as possible in the future.
9021
9022 #ifndef PRODUCT
9023 // Debugging support for CMSStackOverflowALot
9024
9025 // It's OK to call this multi-threaded; the worst thing
9026 // that can happen is that we'll get a bunch of closely
9027 // spaced simulated oveflows, but that's OK, in fact
9028 // probably good as it would exercise the overflow code
9029 // under contention.
9030 bool CMSCollector::simulate_overflow() {
9031 if (_overflow_counter-- <= 0) { // just being defensive
9032 _overflow_counter = CMSMarkStackOverflowInterval;
9033 return true;
9034 } else {
9035 return false;
9036 }
9037 }
9038
9039 bool CMSCollector::par_simulate_overflow() {
9040 return simulate_overflow();
9041 }
9042 #endif
9043
9044 // Single-threaded
9045 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
9046 assert(stack->isEmpty(), "Expected precondition");
9047 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
9048 size_t i = num;
9049 oop cur = _overflow_list;
9050 const markOop proto = markOopDesc::prototype();
9051 NOT_PRODUCT(ssize_t n = 0;)
9052 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
9053 next = oop(cur->mark());
9054 cur->set_mark(proto); // until proven otherwise
9055 assert(cur->is_oop(), "Should be an oop");
9056 bool res = stack->push(cur);
9057 assert(res, "Bit off more than can chew?");
9058 NOT_PRODUCT(n++;)
9059 }
9060 _overflow_list = cur;
9061 #ifndef PRODUCT
9062 assert(_num_par_pushes >= n, "Too many pops?");
9063 _num_par_pushes -=n;
9064 #endif
9065 return !stack->isEmpty();
9066 }
9067
9068 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
9069 // (MT-safe) Get a prefix of at most "num" from the list.
9070 // The overflow list is chained through the mark word of
9071 // each object in the list. We fetch the entire list,
9072 // break off a prefix of the right size and return the
9073 // remainder. If other threads try to take objects from
9074 // the overflow list at that time, they will wait for
9075 // some time to see if data becomes available. If (and
9076 // only if) another thread places one or more object(s)
9077 // on the global list before we have returned the suffix
9078 // to the global list, we will walk down our local list
9079 // to find its end and append the global list to
9080 // our suffix before returning it. This suffix walk can
9081 // prove to be expensive (quadratic in the amount of traffic)
9082 // when there are many objects in the overflow list and
9083 // there is much producer-consumer contention on the list.
9084 // *NOTE*: The overflow list manipulation code here and
9085 // in ParNewGeneration:: are very similar in shape,
9086 // except that in the ParNew case we use the old (from/eden)
9087 // copy of the object to thread the list via its klass word.
9088 // Because of the common code, if you make any changes in
9089 // the code below, please check the ParNew version to see if
9090 // similar changes might be needed.
9091 // CR 6797058 has been filed to consolidate the common code.
9092 bool CMSCollector::par_take_from_overflow_list(size_t num,
9093 OopTaskQueue* work_q,
9094 int no_of_gc_threads) {
9095 assert(work_q->size() == 0, "First empty local work queue");
9096 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
9097 if (_overflow_list == NULL) {
9098 return false;
9099 }
9100 // Grab the entire list; we'll put back a suffix
9101 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9102 Thread* tid = Thread::current();
9103 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
9104 // set to ParallelGCThreads.
9105 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
9106 size_t sleep_time_millis = MAX2((size_t)1, num/100);
9107 // If the list is busy, we spin for a short while,
9108 // sleeping between attempts to get the list.
9109 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
9110 os::sleep(tid, sleep_time_millis, false);
9111 if (_overflow_list == NULL) {
9112 // Nothing left to take
9113 return false;
9114 } else if (_overflow_list != BUSY) {
9115 // Try and grab the prefix
9116 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
9117 }
9118 }
9119 // If the list was found to be empty, or we spun long
9120 // enough, we give up and return empty-handed. If we leave
9121 // the list in the BUSY state below, it must be the case that
9122 // some other thread holds the overflow list and will set it
9123 // to a non-BUSY state in the future.
9124 if (prefix == NULL || prefix == BUSY) {
9125 // Nothing to take or waited long enough
9126 if (prefix == NULL) {
9127 // Write back the NULL in case we overwrote it with BUSY above
9128 // and it is still the same value.
9129 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9130 }
9131 return false;
9132 }
9133 assert(prefix != NULL && prefix != BUSY, "Error");
9134 size_t i = num;
9135 oop cur = prefix;
9136 // Walk down the first "num" objects, unless we reach the end.
9137 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
9138 if (cur->mark() == NULL) {
9139 // We have "num" or fewer elements in the list, so there
9140 // is nothing to return to the global list.
9141 // Write back the NULL in lieu of the BUSY we wrote
9142 // above, if it is still the same value.
9143 if (_overflow_list == BUSY) {
9144 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
9145 }
9146 } else {
9147 // Chop off the suffix and rerturn it to the global list.
9148 assert(cur->mark() != BUSY, "Error");
9149 oop suffix_head = cur->mark(); // suffix will be put back on global list
9150 cur->set_mark(NULL); // break off suffix
9151 // It's possible that the list is still in the empty(busy) state
9152 // we left it in a short while ago; in that case we may be
9153 // able to place back the suffix without incurring the cost
9154 // of a walk down the list.
9155 oop observed_overflow_list = _overflow_list;
9156 oop cur_overflow_list = observed_overflow_list;
9157 bool attached = false;
9158 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
9159 observed_overflow_list =
9160 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9161 if (cur_overflow_list == observed_overflow_list) {
9162 attached = true;
9163 break;
9164 } else cur_overflow_list = observed_overflow_list;
9165 }
9166 if (!attached) {
9167 // Too bad, someone else sneaked in (at least) an element; we'll need
9168 // to do a splice. Find tail of suffix so we can prepend suffix to global
9169 // list.
9170 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
9171 oop suffix_tail = cur;
9172 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
9173 "Tautology");
9174 observed_overflow_list = _overflow_list;
9175 do {
9176 cur_overflow_list = observed_overflow_list;
9177 if (cur_overflow_list != BUSY) {
9178 // Do the splice ...
9179 suffix_tail->set_mark(markOop(cur_overflow_list));
9180 } else { // cur_overflow_list == BUSY
9181 suffix_tail->set_mark(NULL);
9182 }
9183 // ... and try to place spliced list back on overflow_list ...
9184 observed_overflow_list =
9185 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
9186 } while (cur_overflow_list != observed_overflow_list);
9187 // ... until we have succeeded in doing so.
9188 }
9189 }
9190
9191 // Push the prefix elements on work_q
9192 assert(prefix != NULL, "control point invariant");
9193 const markOop proto = markOopDesc::prototype();
9194 oop next;
9195 NOT_PRODUCT(ssize_t n = 0;)
9196 for (cur = prefix; cur != NULL; cur = next) {
9197 next = oop(cur->mark());
9198 cur->set_mark(proto); // until proven otherwise
9199 assert(cur->is_oop(), "Should be an oop");
9200 bool res = work_q->push(cur);
9201 assert(res, "Bit off more than we can chew?");
9202 NOT_PRODUCT(n++;)
9203 }
9204 #ifndef PRODUCT
9205 assert(_num_par_pushes >= n, "Too many pops?");
9206 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
9207 #endif
9208 return true;
9209 }
9210
9211 // Single-threaded
9212 void CMSCollector::push_on_overflow_list(oop p) {
9213 NOT_PRODUCT(_num_par_pushes++;)
9214 assert(p->is_oop(), "Not an oop");
9215 preserve_mark_if_necessary(p);
9216 p->set_mark((markOop)_overflow_list);
9217 _overflow_list = p;
9218 }
9219
9220 // Multi-threaded; use CAS to prepend to overflow list
9221 void CMSCollector::par_push_on_overflow_list(oop p) {
9222 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);)
9223 assert(p->is_oop(), "Not an oop");
9224 par_preserve_mark_if_necessary(p);
9225 oop observed_overflow_list = _overflow_list;
9226 oop cur_overflow_list;
9227 do {
9228 cur_overflow_list = observed_overflow_list;
9229 if (cur_overflow_list != BUSY) {
9230 p->set_mark(markOop(cur_overflow_list));
9231 } else {
9232 p->set_mark(NULL);
9233 }
9234 observed_overflow_list =
9235 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
9236 } while (cur_overflow_list != observed_overflow_list);
9237 }
9238 #undef BUSY
9239
9240 // Single threaded
9241 // General Note on GrowableArray: pushes may silently fail
9242 // because we are (temporarily) out of C-heap for expanding
9243 // the stack. The problem is quite ubiquitous and affects
9244 // a lot of code in the JVM. The prudent thing for GrowableArray
9245 // to do (for now) is to exit with an error. However, that may
9246 // be too draconian in some cases because the caller may be
9247 // able to recover without much harm. For such cases, we
9248 // should probably introduce a "soft_push" method which returns
9249 // an indication of success or failure with the assumption that
9250 // the caller may be able to recover from a failure; code in
9251 // the VM can then be changed, incrementally, to deal with such
9252 // failures where possible, thus, incrementally hardening the VM
9253 // in such low resource situations.
9254 void CMSCollector::preserve_mark_work(oop p, markOop m) {
9255 _preserved_oop_stack.push(p);
9256 _preserved_mark_stack.push(m);
9257 assert(m == p->mark(), "Mark word changed");
9258 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9259 "bijection");
9260 }
9261
9262 // Single threaded
9263 void CMSCollector::preserve_mark_if_necessary(oop p) {
9264 markOop m = p->mark();
9265 if (m->must_be_preserved(p)) {
9266 preserve_mark_work(p, m);
9267 }
9268 }
9269
9270 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
9271 markOop m = p->mark();
9272 if (m->must_be_preserved(p)) {
9273 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
9274 // Even though we read the mark word without holding
9275 // the lock, we are assured that it will not change
9276 // because we "own" this oop, so no other thread can
9277 // be trying to push it on the overflow list; see
9278 // the assertion in preserve_mark_work() that checks
9279 // that m == p->mark().
9280 preserve_mark_work(p, m);
9281 }
9282 }
9283
9284 // We should be able to do this multi-threaded,
9285 // a chunk of stack being a task (this is
9286 // correct because each oop only ever appears
9287 // once in the overflow list. However, it's
9288 // not very easy to completely overlap this with
9289 // other operations, so will generally not be done
9290 // until all work's been completed. Because we
9291 // expect the preserved oop stack (set) to be small,
9292 // it's probably fine to do this single-threaded.
9293 // We can explore cleverer concurrent/overlapped/parallel
9294 // processing of preserved marks if we feel the
9295 // need for this in the future. Stack overflow should
9296 // be so rare in practice and, when it happens, its
9297 // effect on performance so great that this will
9298 // likely just be in the noise anyway.
9299 void CMSCollector::restore_preserved_marks_if_any() {
9300 assert(SafepointSynchronize::is_at_safepoint(),
9301 "world should be stopped");
9302 assert(Thread::current()->is_ConcurrentGC_thread() ||
9303 Thread::current()->is_VM_thread(),
9304 "should be single-threaded");
9305 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
9306 "bijection");
9307
9308 while (!_preserved_oop_stack.is_empty()) {
9309 oop p = _preserved_oop_stack.pop();
9310 assert(p->is_oop(), "Should be an oop");
9311 assert(_span.contains(p), "oop should be in _span");
9312 assert(p->mark() == markOopDesc::prototype(),
9313 "Set when taken from overflow list");
9314 markOop m = _preserved_mark_stack.pop();
9315 p->set_mark(m);
9316 }
9317 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
9318 "stacks were cleared above");
9319 }
9320
9321 #ifndef PRODUCT
9322 bool CMSCollector::no_preserved_marks() const {
9323 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
9324 }
9325 #endif
9326
9327 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const
9328 {
9329 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap();
9330 CMSAdaptiveSizePolicy* size_policy =
9331 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy();
9332 assert(size_policy->is_gc_cms_adaptive_size_policy(),
9333 "Wrong type for size policy");
9334 return size_policy;
9335 }
9336
9337 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size,
9338 size_t desired_promo_size) {
9339 if (cur_promo_size < desired_promo_size) {
9340 size_t expand_bytes = desired_promo_size - cur_promo_size;
9341 if (PrintAdaptiveSizePolicy && Verbose) {
9342 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9343 "Expanding tenured generation by " SIZE_FORMAT " (bytes)",
9344 expand_bytes);
9345 }
9346 expand(expand_bytes,
9347 MinHeapDeltaBytes,
9348 CMSExpansionCause::_adaptive_size_policy);
9349 } else if (desired_promo_size < cur_promo_size) {
9350 size_t shrink_bytes = cur_promo_size - desired_promo_size;
9351 if (PrintAdaptiveSizePolicy && Verbose) {
9352 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize "
9353 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)",
9354 shrink_bytes);
9355 }
9356 shrink(shrink_bytes);
9357 }
9358 }
9359
9360 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() {
9361 GenCollectedHeap* gch = GenCollectedHeap::heap();
9362 CMSGCAdaptivePolicyCounters* counters =
9363 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
9364 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
9365 "Wrong kind of counters");
9366 return counters;
9367 }
9368
9369
9370 void ASConcurrentMarkSweepGeneration::update_counters() {
9371 if (UsePerfData) {
9372 _space_counters->update_all();
9373 _gen_counters->update_all();
9374 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9375 GenCollectedHeap* gch = GenCollectedHeap::heap();
9376 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9377 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9378 "Wrong gc statistics type");
9379 counters->update_counters(gc_stats_l);
9380 }
9381 }
9382
9383 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) {
9384 if (UsePerfData) {
9385 _space_counters->update_used(used);
9386 _space_counters->update_capacity();
9387 _gen_counters->update_all();
9388
9389 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters();
9390 GenCollectedHeap* gch = GenCollectedHeap::heap();
9391 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats();
9392 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind,
9393 "Wrong gc statistics type");
9394 counters->update_counters(gc_stats_l);
9395 }
9396 }
9397
9398 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) {
9399 assert_locked_or_safepoint(Heap_lock);
9400 assert_lock_strong(freelistLock());
9401 HeapWord* old_end = _cmsSpace->end();
9402 HeapWord* unallocated_start = _cmsSpace->unallocated_block();
9403 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start");
9404 FreeChunk* chunk_at_end = find_chunk_at_end();
9405 if (chunk_at_end == NULL) {
9406 // No room to shrink
9407 if (PrintGCDetails && Verbose) {
9408 gclog_or_tty->print_cr("No room to shrink: old_end "
9409 PTR_FORMAT " unallocated_start " PTR_FORMAT
9410 " chunk_at_end " PTR_FORMAT,
9411 old_end, unallocated_start, chunk_at_end);
9412 }
9413 return;
9414 } else {
9415
9416 // Find the chunk at the end of the space and determine
9417 // how much it can be shrunk.
9418 size_t shrinkable_size_in_bytes = chunk_at_end->size();
9419 size_t aligned_shrinkable_size_in_bytes =
9420 align_size_down(shrinkable_size_in_bytes, os::vm_page_size());
9421 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(),
9422 "Inconsistent chunk at end of space");
9423 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes);
9424 size_t word_size_before = heap_word_size(_virtual_space.committed_size());
9425
9426 // Shrink the underlying space
9427 _virtual_space.shrink_by(bytes);
9428 if (PrintGCDetails && Verbose) {
9429 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:"
9430 " desired_bytes " SIZE_FORMAT
9431 " shrinkable_size_in_bytes " SIZE_FORMAT
9432 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT
9433 " bytes " SIZE_FORMAT,
9434 desired_bytes, shrinkable_size_in_bytes,
9435 aligned_shrinkable_size_in_bytes, bytes);
9436 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT
9437 " unallocated_start " SIZE_FORMAT,
9438 old_end, unallocated_start);
9439 }
9440
9441 // If the space did shrink (shrinking is not guaranteed),
9442 // shrink the chunk at the end by the appropriate amount.
9443 if (((HeapWord*)_virtual_space.high()) < old_end) {
9444 size_t new_word_size =
9445 heap_word_size(_virtual_space.committed_size());
9446
9447 // Have to remove the chunk from the dictionary because it is changing
9448 // size and might be someplace elsewhere in the dictionary.
9449
9450 // Get the chunk at end, shrink it, and put it
9451 // back.
9452 _cmsSpace->removeChunkFromDictionary(chunk_at_end);
9453 size_t word_size_change = word_size_before - new_word_size;
9454 size_t chunk_at_end_old_size = chunk_at_end->size();
9455 assert(chunk_at_end_old_size >= word_size_change,
9456 "Shrink is too large");
9457 chunk_at_end->set_size(chunk_at_end_old_size -
9458 word_size_change);
9459 _cmsSpace->freed((HeapWord*) chunk_at_end->end(),
9460 word_size_change);
9461
9462 _cmsSpace->returnChunkToDictionary(chunk_at_end);
9463
9464 MemRegion mr(_cmsSpace->bottom(), new_word_size);
9465 _bts->resize(new_word_size); // resize the block offset shared array
9466 Universe::heap()->barrier_set()->resize_covered_region(mr);
9467 _cmsSpace->assert_locked();
9468 _cmsSpace->set_end((HeapWord*)_virtual_space.high());
9469
9470 NOT_PRODUCT(_cmsSpace->dictionary()->verify());
9471
9472 // update the space and generation capacity counters
9473 if (UsePerfData) {
9474 _space_counters->update_capacity();
9475 _gen_counters->update_all();
9476 }
9477
9478 if (Verbose && PrintGCDetails) {
9479 size_t new_mem_size = _virtual_space.committed_size();
9480 size_t old_mem_size = new_mem_size + bytes;
9481 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K",
9482 name(), old_mem_size/K, bytes/K, new_mem_size/K);
9483 }
9484 }
9485
9486 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(),
9487 "Inconsistency at end of space");
9488 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(),
9489 "Shrinking is inconsistent");
9490 return;
9491 }
9492 }
9493 // Transfer some number of overflown objects to usual marking
9494 // stack. Return true if some objects were transferred.
9495 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
9496 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
9497 (size_t)ParGCDesiredObjsFromOverflowList);
9498
9499 bool res = _collector->take_from_overflow_list(num, _mark_stack);
9500 assert(_collector->overflow_list_is_empty() || res,
9501 "If list is not empty, we should have taken something");
9502 assert(!res || !_mark_stack->isEmpty(),
9503 "If we took something, it should now be on our stack");
9504 return res;
9505 }
9506
9507 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
9508 size_t res = _sp->block_size_no_stall(addr, _collector);
9509 if (_sp->block_is_obj(addr)) {
9510 if (_live_bit_map->isMarked(addr)) {
9511 // It can't have been dead in a previous cycle
9512 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
9513 } else {
9514 _dead_bit_map->mark(addr); // mark the dead object
9515 }
9516 }
9517 // Could be 0, if the block size could not be computed without stalling.
9518 return res;
9519 }
9520
9521 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
9522
9523 switch (phase) {
9524 case CMSCollector::InitialMarking:
9525 initialize(true /* fullGC */ ,
9526 cause /* cause of the GC */,
9527 true /* recordGCBeginTime */,
9528 true /* recordPreGCUsage */,
9529 false /* recordPeakUsage */,
9530 false /* recordPostGCusage */,
9531 true /* recordAccumulatedGCTime */,
9532 false /* recordGCEndTime */,
9533 false /* countCollection */ );
9534 break;
9535
9536 case CMSCollector::FinalMarking:
9537 initialize(true /* fullGC */ ,
9538 cause /* cause of the GC */,
9539 false /* recordGCBeginTime */,
9540 false /* recordPreGCUsage */,
9541 false /* recordPeakUsage */,
9542 false /* recordPostGCusage */,
9543 true /* recordAccumulatedGCTime */,
9544 false /* recordGCEndTime */,
9545 false /* countCollection */ );
9546 break;
9547
9548 case CMSCollector::Sweeping:
9549 initialize(true /* fullGC */ ,
9550 cause /* cause of the GC */,
9551 false /* recordGCBeginTime */,
9552 false /* recordPreGCUsage */,
9553 true /* recordPeakUsage */,
9554 true /* recordPostGCusage */,
9555 false /* recordAccumulatedGCTime */,
9556 true /* recordGCEndTime */,
9557 true /* countCollection */ );
9558 break;
9559
9560 default:
9561 ShouldNotReachHere();
9562 }
9563 }